revert pageserver parts of "feat(walredo): use posix_spawn by moving close_fds() work to walredo C code (#6574)"

the addition of close_range() to the C code remains, it doesn't matter
for the purposes of this reproducer.

This reverts parts of commit 1be5e564ce.

.github/workflows/actionlint.yml

@@ -17,6 +17,7 @@ concurrency:
 
 jobs:
   actionlint:
+    if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4

.github/workflows/approved-for-ci-run.yml

@@ -93,6 +93,7 @@ jobs:
                                                        --body-file "body.md" \
                                                        --head "${BRANCH}" \
                                                        --base "main" \
+                                                       --label "run-e2e-tests-in-draft" \
                                                        --draft
           fi
 

.github/workflows/build_and_test.yml

@@ -22,12 +22,12 @@ env:
   AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
   AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
   # A concurrency group that we use for e2e-tests runs, matches `concurrency.group` above with `github.repository` as a prefix
-  E2E_CONCURRENCY_GROUP: ${{ github.repository }}-${{ github.workflow }}-${{ github.ref_name }}-${{ github.ref_name == 'main' && github.sha || 'anysha' }}
+  E2E_CONCURRENCY_GROUP: ${{ github.repository }}-e2e-tests-${{ github.ref_name }}-${{ github.ref_name == 'main' && github.sha || 'anysha' }}
 
 jobs:
   check-permissions:
+    if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
     runs-on: ubuntu-latest
-
     steps:
     - name: Disallow PRs from forks
       if: |
@@ -253,7 +253,7 @@ jobs:
           done
 
           if [ "${FAILED}" = "true" ]; then
-            echo >&2 "Please update vendors/revisions.json if these changes are intentional"
+            echo >&2 "Please update vendor/revisions.json if these changes are intentional"
             exit 1
           fi
 
@@ -692,50 +692,10 @@ jobs:
             })
 
   trigger-e2e-tests:
+    if: ${{ !github.event.pull_request.draft || contains( github.event.pull_request.labels.*.name, 'run-e2e-tests-in-draft') || github.ref_name == 'main' || github.ref_name == 'release' }}
     needs: [ check-permissions, promote-images, tag ]
-    runs-on: [ self-hosted, gen3, small ]
-    container:
-      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
-      options: --init
-    steps:
-      - name: Set PR's status to pending and request a remote CI test
-        run: |
-          # For pull requests, GH Actions set "github.sha" variable to point at a fake merge commit
-          # but we need to use a real sha of a latest commit in the PR's branch for the e2e job,
-          # to place a job run status update later.
-          COMMIT_SHA=${{ github.event.pull_request.head.sha }}
-          # For non-PR kinds of runs, the above will produce an empty variable, pick the original sha value for those
-          COMMIT_SHA=${COMMIT_SHA:-${{ github.sha }}}
-
-          REMOTE_REPO="${{ github.repository_owner }}/cloud"
-
-          curl -f -X POST \
-          https://api.github.com/repos/${{ github.repository }}/statuses/$COMMIT_SHA \
-          -H "Accept: application/vnd.github.v3+json" \
-          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
-          --data \
-            "{
-              \"state\": \"pending\",
-              \"context\": \"neon-cloud-e2e\",
-              \"description\": \"[$REMOTE_REPO] Remote CI job is about to start\"
-            }"
-
-          curl -f -X POST \
-          https://api.github.com/repos/$REMOTE_REPO/actions/workflows/testing.yml/dispatches \
-          -H "Accept: application/vnd.github.v3+json" \
-          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
-          --data \
-            "{
-              \"ref\": \"main\",
-              \"inputs\": {
-                \"ci_job_name\": \"neon-cloud-e2e\",
-                \"commit_hash\": \"$COMMIT_SHA\",
-                \"remote_repo\": \"${{ github.repository }}\",
-                \"storage_image_tag\": \"${{ needs.tag.outputs.build-tag }}\",
-                \"compute_image_tag\": \"${{ needs.tag.outputs.build-tag }}\",
-                \"concurrency_group\": \"${{ env.E2E_CONCURRENCY_GROUP }}\"
-              }
-            }"
+    uses: ./.github/workflows/trigger-e2e-tests.yml
+    secrets: inherit
 
   neon-image:
     needs: [ check-permissions, build-buildtools-image, tag ]

.github/workflows/neon_extra_builds.yml

@@ -117,6 +117,7 @@ jobs:
 
   check-linux-arm-build:
     timeout-minutes: 90
+    if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
     runs-on: [ self-hosted, dev, arm64 ]
 
     env:
@@ -237,6 +238,7 @@ jobs:
 
   check-codestyle-rust-arm:
     timeout-minutes: 90
+    if: ${{ !contains(github.event.pull_request.labels.*.name, 'run-no-ci') }}
     runs-on: [ self-hosted, dev, arm64 ]
 
     container:

.github/workflows/trigger-e2e-tests.yml

@@ -0,0 +1,118 @@
+name: Trigger E2E Tests
+
+on:
+  pull_request:
+    types:
+      - ready_for_review
+  workflow_call:
+
+defaults:
+  run:
+    shell: bash -euxo pipefail {0}
+    
+env:
+  # A concurrency group that we use for e2e-tests runs, matches `concurrency.group` above with `github.repository` as a prefix
+  E2E_CONCURRENCY_GROUP: ${{ github.repository }}-e2e-tests-${{ github.ref_name }}-${{ github.ref_name == 'main' && github.sha || 'anysha' }}
+  AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_DEV }}
+  AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_KEY_DEV }}
+
+jobs:
+  cancel-previous-e2e-tests:
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Cancel previous e2e-tests runs for this PR
+        env:
+          GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
+        run: |
+          gh workflow --repo neondatabase/cloud \
+            run cancel-previous-in-concurrency-group.yml \
+              --field concurrency_group="${{ env.E2E_CONCURRENCY_GROUP }}"
+
+  tag:
+    runs-on: [ ubuntu-latest ]
+    outputs:
+      build-tag: ${{ steps.build-tag.outputs.tag }}
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
+      - name: Get build tag
+        env:
+          GH_TOKEN: ${{ secrets.CI_ACCESS_TOKEN }}
+          CURRENT_BRANCH: ${{ github.head_ref || github.ref_name }}
+          CURRENT_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
+        run: |
+          if [[ "$GITHUB_REF_NAME" == "main" ]]; then
+            echo "tag=$(git rev-list --count HEAD)" | tee -a $GITHUB_OUTPUT
+          elif [[ "$GITHUB_REF_NAME" == "release" ]]; then
+            echo "tag=release-$(git rev-list --count HEAD)" | tee -a $GITHUB_OUTPUT
+          else
+            echo "GITHUB_REF_NAME (value '$GITHUB_REF_NAME') is not set to either 'main' or 'release'"
+            BUILD_AND_TEST_RUN_ID=$(gh run list -b $CURRENT_BRANCH -c $CURRENT_SHA -w 'Build and Test' -L 1 --json databaseId --jq '.[].databaseId')
+            echo "tag=$BUILD_AND_TEST_RUN_ID" | tee -a $GITHUB_OUTPUT
+          fi
+        id: build-tag
+
+  trigger-e2e-tests:
+    needs: [ tag ]
+    runs-on: [ self-hosted, gen3, small ]
+    env:
+      TAG: ${{ needs.tag.outputs.build-tag }}
+    container:
+      image: 369495373322.dkr.ecr.eu-central-1.amazonaws.com/base:pinned
+      options: --init
+    steps:
+      - name: check if ecr image are present
+        run: |
+          for REPO in neon compute-tools compute-node-v14 vm-compute-node-v14 compute-node-v15 vm-compute-node-v15 compute-node-v16 vm-compute-node-v16; do
+            OUTPUT=$(aws ecr describe-images --repository-name ${REPO} --region eu-central-1 --query "imageDetails[?imageTags[?contains(@, '${TAG}')]]" --output text)
+            if [ "$OUTPUT" == "" ]; then
+              echo "$REPO with image tag $TAG not found" >> $GITHUB_OUTPUT
+              exit 1
+            fi
+          done
+
+      - name: Set PR's status to pending and request a remote CI test
+        run: |
+          # For pull requests, GH Actions set "github.sha" variable to point at a fake merge commit
+          # but we need to use a real sha of a latest commit in the PR's branch for the e2e job,
+          # to place a job run status update later.
+          COMMIT_SHA=${{ github.event.pull_request.head.sha }}
+          # For non-PR kinds of runs, the above will produce an empty variable, pick the original sha value for those
+          COMMIT_SHA=${COMMIT_SHA:-${{ github.sha }}}
+
+          REMOTE_REPO="${{ github.repository_owner }}/cloud"
+
+          curl -f -X POST \
+          https://api.github.com/repos/${{ github.repository }}/statuses/$COMMIT_SHA \
+          -H "Accept: application/vnd.github.v3+json" \
+          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
+          --data \
+            "{
+              \"state\": \"pending\",
+              \"context\": \"neon-cloud-e2e\",
+              \"description\": \"[$REMOTE_REPO] Remote CI job is about to start\"
+            }"
+
+          curl -f -X POST \
+          https://api.github.com/repos/$REMOTE_REPO/actions/workflows/testing.yml/dispatches \
+          -H "Accept: application/vnd.github.v3+json" \
+          --user "${{ secrets.CI_ACCESS_TOKEN }}" \
+          --data \
+            "{
+              \"ref\": \"main\",
+              \"inputs\": {
+                \"ci_job_name\": \"neon-cloud-e2e\",
+                \"commit_hash\": \"$COMMIT_SHA\",
+                \"remote_repo\": \"${{ github.repository }}\",
+                \"storage_image_tag\": \"${TAG}\",
+                \"compute_image_tag\": \"${TAG}\",
+                \"concurrency_group\": \"${{ env.E2E_CONCURRENCY_GROUP }}\"
+              }
+            }"
+ 

CONTRIBUTING.md

@@ -54,6 +54,9 @@ _An instruction for maintainers_
 - If and only if it looks **safe** (i.e. it doesn't contain any malicious code which could expose secrets or harm the CI), then:
     - Press the "Approve and run" button in GitHub UI
     - Add the `approved-for-ci-run` label to the PR
+    - Currently draft PR will skip e2e test (only for internal contributors). After turning the PR 'Ready to Review' CI will trigger e2e test
+      - Add `run-e2e-tests-in-draft` label to run e2e test in draft PR (override above behaviour)
+      - The `approved-for-ci-run` workflow will add `run-e2e-tests-in-draft` automatically to run e2e test for external contributors
 
 Repeat all steps after any change to the PR.
 - When the changes are ready to get merged — merge the original PR (not the internal one)

Cargo.lock

@@ -289,6 +289,7 @@ dependencies = [
  "pageserver_api",
  "pageserver_client",
  "postgres_connection",
+ "r2d2",
  "reqwest",
  "serde",
  "serde_json",
@@ -1155,6 +1156,16 @@ version = "0.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2da6da31387c7e4ef160ffab6d5e7f00c42626fe39aea70a7b0f1773f7dd6c1b"
 
+[[package]]
+name = "close_fds"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3bc416f33de9d59e79e57560f450d21ff8393adcf1cdfc3e6d8fb93d5f88a2ed"
+dependencies = [
+ "cfg-if",
+ "libc",
+]
+
 [[package]]
 name = "colorchoice"
 version = "1.0.0"
@@ -1328,8 +1339,6 @@ dependencies = [
  "clap",
  "comfy-table",
  "compute_api",
- "diesel",
- "diesel_migrations",
  "futures",
  "git-version",
  "hex",
@@ -1640,6 +1649,22 @@ dependencies = [
  "rusticata-macros",
 ]
 
+[[package]]
+name = "desim"
+version = "0.1.0"
+dependencies = [
+ "anyhow",
+ "bytes",
+ "hex",
+ "parking_lot 0.12.1",
+ "rand 0.8.5",
+ "scopeguard",
+ "smallvec",
+ "tracing",
+ "utils",
+ "workspace_hack",
+]
+
 [[package]]
 name = "diesel"
 version = "2.1.4"
@@ -1651,6 +1676,7 @@ dependencies = [
  "diesel_derives",
  "itoa",
  "pq-sys",
+ "r2d2",
  "serde_json",
 ]
 
@@ -2247,11 +2273,11 @@ dependencies = [
 
 [[package]]
 name = "hashlink"
-version = "0.8.2"
+version = "0.8.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0761a1b9491c4f2e3d66aa0f62d0fba0af9a0e2852e4d48ea506632a4b56e6aa"
+checksum = "e8094feaf31ff591f651a2664fb9cfd92bba7a60ce3197265e9482ebe753c8f7"
 dependencies = [
- "hashbrown 0.13.2",
+ "hashbrown 0.14.0",
 ]
 
 [[package]]
@@ -2867,6 +2893,7 @@ dependencies = [
  "chrono",
  "libc",
  "once_cell",
+ "procfs",
  "prometheus",
  "rand 0.8.5",
  "rand_distr",
@@ -3440,6 +3467,7 @@ dependencies = [
  "camino-tempfile",
  "chrono",
  "clap",
+ "close_fds",
  "const_format",
  "consumption_metrics",
  "crc32c",
@@ -3935,6 +3963,7 @@ dependencies = [
  "pin-project-lite",
  "postgres-protocol",
  "rand 0.8.5",
+ "serde",
  "thiserror",
  "tokio",
  "tracing",
@@ -3984,6 +4013,8 @@ checksum = "b1de8dacb0873f77e6aefc6d71e044761fcc68060290f5b1089fcdf84626bb69"
 dependencies = [
  "bitflags 1.3.2",
  "byteorder",
+ "chrono",
+ "flate2",
  "hex",
  "lazy_static",
  "rustix 0.36.16",
@@ -4074,6 +4105,7 @@ dependencies = [
  "clap",
  "consumption_metrics",
  "dashmap",
+ "env_logger",
  "futures",
  "git-version",
  "hashbrown 0.13.2",
@@ -4121,6 +4153,7 @@ dependencies = [
  "serde",
  "serde_json",
  "sha2",
+ "smallvec",
  "smol_str",
  "socket2 0.5.5",
  "sync_wrapper",
@@ -4139,6 +4172,7 @@ dependencies = [
  "tracing-subscriber",
  "tracing-utils",
  "url",
+ "urlencoding",
  "utils",
  "uuid",
  "walkdir",
@@ -4166,6 +4200,17 @@ dependencies = [
  "proc-macro2",
 ]
 
+[[package]]
+name = "r2d2"
+version = "0.8.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "51de85fb3fb6524929c8a2eb85e6b6d363de4e8c48f9e2c2eac4944abc181c93"
+dependencies = [
+ "log",
+ "parking_lot 0.12.1",
+ "scheduled-thread-pool",
+]
+
 [[package]]
 name = "rand"
 version = "0.7.3"
@@ -4402,6 +4447,7 @@ dependencies = [
  "futures",
  "futures-util",
  "http-types",
+ "humantime",
  "hyper",
  "itertools",
  "metrics",
@@ -4413,6 +4459,7 @@ dependencies = [
  "serde_json",
  "test-context",
  "tokio",
+ "tokio-stream",
  "tokio-util",
  "toml_edit",
  "tracing",
@@ -4810,6 +4857,7 @@ dependencies = [
  "clap",
  "const_format",
  "crc32c",
+ "desim",
  "fail",
  "fs2",
  "futures",
@@ -4825,6 +4873,7 @@ dependencies = [
  "postgres_backend",
  "postgres_ffi",
  "pq_proto",
+ "rand 0.8.5",
  "regex",
  "remote_storage",
  "reqwest",
@@ -4845,8 +4894,10 @@ dependencies = [
  "tokio-util",
  "toml_edit",
  "tracing",
+ "tracing-subscriber",
  "url",
  "utils",
+ "walproposer",
  "workspace_hack",
 ]
 
@@ -4879,6 +4930,15 @@ dependencies = [
  "windows-sys 0.42.0",
 ]
 
+[[package]]
+name = "scheduled-thread-pool"
+version = "0.2.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3cbc66816425a074528352f5789333ecff06ca41b36b0b0efdfbb29edc391a19"
+dependencies = [
+ "parking_lot 0.12.1",
+]
+
 [[package]]
 name = "scopeguard"
 version = "1.1.0"
@@ -5714,7 +5774,7 @@ dependencies = [
 [[package]]
 name = "tokio-epoll-uring"
 version = "0.1.0"
-source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#0e1af4ccddf2f01805cfc9eaefa97ee13c04b52d"
+source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#868d2c42b5d54ca82fead6e8f2f233b69a540d3e"
 dependencies = [
  "futures",
  "nix 0.26.4",
@@ -6239,8 +6299,9 @@ dependencies = [
 [[package]]
 name = "uring-common"
 version = "0.1.0"
-source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#0e1af4ccddf2f01805cfc9eaefa97ee13c04b52d"
+source = "git+https://github.com/neondatabase/tokio-epoll-uring.git?branch=main#868d2c42b5d54ca82fead6e8f2f233b69a540d3e"
 dependencies = [
+ "bytes",
  "io-uring",
  "libc",
 ]
@@ -6806,7 +6867,6 @@ dependencies = [
  "clap",
  "clap_builder",
  "crossbeam-utils",
- "diesel",
  "either",
  "fail",
  "futures-channel",

Cargo.toml

@@ -18,6 +18,7 @@ members = [
     "libs/pageserver_api",
     "libs/postgres_ffi",
     "libs/safekeeper_api",
+    "libs/desim",
     "libs/utils",
     "libs/consumption_metrics",
     "libs/postgres_backend",
@@ -65,6 +66,7 @@ camino = "1.1.6"
 cfg-if = "1.0.0"
 chrono = { version = "0.4", default-features = false, features = ["clock"] }
 clap = { version = "4.0", features = ["derive"] }
+close_fds = "0.3.2"
 comfy-table = "6.1"
 const_format = "0.2"
 crc32c = "0.6"
@@ -80,7 +82,7 @@ futures-core = "0.3"
 futures-util = "0.3"
 git-version = "0.3"
 hashbrown = "0.13"
-hashlink = "0.8.1"
+hashlink = "0.8.4"
 hdrhistogram = "7.5.2"
 hex = "0.4"
 hex-literal = "0.4"
@@ -113,6 +115,7 @@ parquet = { version = "49.0.0", default-features = false, features = ["zstd"] }
 parquet_derive = "49.0.0"
 pbkdf2 = { version = "0.12.1", features = ["simple", "std"] }
 pin-project-lite = "0.2"
+procfs = "0.14"
 prometheus = {version = "0.13", default_features=false, features = ["process"]} # removes protobuf dependency
 prost = "0.11"
 rand = "0.8"
@@ -170,6 +173,7 @@ tracing-opentelemetry = "0.20.0"
 tracing-subscriber = { version = "0.3", default_features = false, features = ["smallvec", "fmt", "tracing-log", "std", "env-filter", "json"] }
 twox-hash = { version = "1.6.3", default-features = false }
 url = "2.2"
+urlencoding = "2.1"
 uuid = { version = "1.6.1", features = ["v4", "v7", "serde"] }
 walkdir = "2.3.2"
 webpki-roots = "0.25"
@@ -201,6 +205,7 @@ postgres_ffi = { version = "0.1", path = "./libs/postgres_ffi/" }
 pq_proto = { version = "0.1", path = "./libs/pq_proto/" }
 remote_storage = { version = "0.1", path = "./libs/remote_storage/" }
 safekeeper_api = { version = "0.1", path = "./libs/safekeeper_api" }
+desim = { version = "0.1", path = "./libs/desim" }
 storage_broker = { version = "0.1", path = "./storage_broker/" } # Note: main broker code is inside the binary crate, so linking with the library shouldn't be heavy.
 tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }

Dockerfile

@@ -100,6 +100,11 @@ RUN mkdir -p /data/.neon/ && chown -R neon:neon /data/.neon/ \
        -c "listen_pg_addr='0.0.0.0:6400'" \
        -c "listen_http_addr='0.0.0.0:9898'"
 
+# When running a binary that links with libpq, default to using our most recent postgres version.  Binaries
+# that want a particular postgres version will select it explicitly: this is just a default.
+ENV LD_LIBRARY_PATH /usr/local/v16/lib
+
+
 VOLUME ["/data"]
 USER neon
 EXPOSE 6400

Dockerfile.buildtools

@@ -135,7 +135,7 @@ WORKDIR /home/nonroot
 
 # Rust
 # Please keep the version of llvm (installed above) in sync with rust llvm (`rustc --version --verbose | grep LLVM`)
-ENV RUSTC_VERSION=1.75.0
+ENV RUSTC_VERSION=1.76.0
 ENV RUSTUP_HOME="/home/nonroot/.rustup"
 ENV PATH="/home/nonroot/.cargo/bin:${PATH}"
 RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux-gnu/rustup-init && whoami && \

Dockerfile.compute-node

@@ -639,8 +639,8 @@ FROM build-deps AS pg-anon-pg-build
 COPY --from=pg-build /usr/local/pgsql/ /usr/local/pgsql/
 
 ENV PATH "/usr/local/pgsql/bin/:$PATH"
-RUN wget https://gitlab.com/dalibo/postgresql_anonymizer/-/archive/1.1.0/postgresql_anonymizer-1.1.0.tar.gz -O pg_anon.tar.gz && \
-    echo "08b09d2ff9b962f96c60db7e6f8e79cf7253eb8772516998fc35ece08633d3ad pg_anon.tar.gz" | sha256sum --check && \
+RUN wget  https://github.com/neondatabase/postgresql_anonymizer/archive/refs/tags/neon_1.1.1.tar.gz -O pg_anon.tar.gz && \
+    echo "321ea8d5c1648880aafde850a2c576e4a9e7b9933a34ce272efc839328999fa9  pg_anon.tar.gz" | sha256sum --check && \
     mkdir pg_anon-src && cd pg_anon-src && tar xvzf ../pg_anon.tar.gz --strip-components=1 -C . && \
     find /usr/local/pgsql -type f | sed 's|^/usr/local/pgsql/||' > /before.txt &&\
     make -j $(getconf _NPROCESSORS_ONLN) install PG_CONFIG=/usr/local/pgsql/bin/pg_config && \
@@ -809,6 +809,7 @@ COPY --from=pg-roaringbitmap-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-semver-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=pg-embedding-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY --from=wal2json-pg-build /usr/local/pgsql /usr/local/pgsql
+COPY --from=pg-anon-pg-build /usr/local/pgsql/ /usr/local/pgsql/
 COPY pgxn/ pgxn/
 
 RUN make -j $(getconf _NPROCESSORS_ONLN) \
@@ -819,6 +820,10 @@ RUN make -j $(getconf _NPROCESSORS_ONLN) \
         PG_CONFIG=/usr/local/pgsql/bin/pg_config \
         -C pgxn/neon_utils \
         -s install && \
+    make -j $(getconf _NPROCESSORS_ONLN) \
+        PG_CONFIG=/usr/local/pgsql/bin/pg_config \
+        -C pgxn/neon_test_utils \
+        -s install && \
     make -j $(getconf _NPROCESSORS_ONLN) \
         PG_CONFIG=/usr/local/pgsql/bin/pg_config \
         -C pgxn/neon_rmgr \

NOTICE

@@ -1,5 +1,5 @@
 Neon
-Copyright 2022 Neon Inc.
+Copyright 2022 - 2024 Neon Inc.
 
 The PostgreSQL submodules in vendor/ are licensed under the PostgreSQL license.
 See vendor/postgres-vX/COPYRIGHT for details.

compute_tools/src/compute.rs

@@ -2,7 +2,7 @@ use std::collections::HashMap;
 use std::env;
 use std::fs;
 use std::io::BufRead;
-use std::os::unix::fs::PermissionsExt;
+use std::os::unix::fs::{symlink, PermissionsExt};
 use std::path::Path;
 use std::process::{Command, Stdio};
 use std::str::FromStr;
@@ -634,6 +634,48 @@ impl ComputeNode {
         // Update pg_hba.conf received with basebackup.
         update_pg_hba(pgdata_path)?;
 
+        // Place pg_dynshmem under /dev/shm. This allows us to use
+        // 'dynamic_shared_memory_type = mmap' so that the files are placed in
+        // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
+        //
+        // Why on earth don't we just stick to the 'posix' default, you might
+        // ask.  It turns out that making large allocations with 'posix' doesn't
+        // work very well with autoscaling. The behavior we want is that:
+        //
+        // 1. You can make large DSM allocations, larger than the current RAM
+        //    size of the VM, without errors
+        //
+        // 2. If the allocated memory is really used, the VM is scaled up
+        //    automatically to accommodate that
+        //
+        // We try to make that possible by having swap in the VM. But with the
+        // default 'posix' DSM implementation, we fail step 1, even when there's
+        // plenty of swap available. PostgreSQL uses posix_fallocate() to create
+        // the shmem segment, which is really just a file in /dev/shm in Linux,
+        // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
+        // than available RAM.
+        //
+        // Using 'dynamic_shared_memory_type = mmap' works around that, because
+        // the Postgres 'mmap' DSM implementation doesn't use
+        // posix_fallocate(). Instead, it uses repeated calls to write(2) to
+        // fill the file with zeros. It's weird that that differs between
+        // 'posix' and 'mmap', but we take advantage of it. When the file is
+        // filled slowly with write(2), the kernel allows it to grow larger, as
+        // long as there's swap available.
+        //
+        // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
+        // segment to be larger than currently available RAM. But because we
+        // don't want to store it on a real file, which the kernel would try to
+        // flush to disk, so symlink pg_dynshm to /dev/shm.
+        //
+        // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
+        // control plane control that option. If 'mmap' is not used, this
+        // symlink doesn't affect anything.
+        //
+        // See https://github.com/neondatabase/autoscaling/issues/800
+        std::fs::remove_dir(pgdata_path.join("pg_dynshmem"))?;
+        symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
+
         match spec.mode {
             ComputeMode::Primary => {}
             ComputeMode::Replica | ComputeMode::Static(..) => {
@@ -765,7 +807,12 @@ impl ComputeNode {
         handle_roles(spec, &mut client)?;
         handle_databases(spec, &mut client)?;
         handle_role_deletions(spec, connstr.as_str(), &mut client)?;
-        handle_grants(spec, &mut client, connstr.as_str())?;
+        handle_grants(
+            spec,
+            &mut client,
+            connstr.as_str(),
+            self.has_feature(ComputeFeature::AnonExtension),
+        )?;
         handle_extensions(spec, &mut client)?;
         handle_extension_neon(&mut client)?;
         create_availability_check_data(&mut client)?;
@@ -773,12 +820,11 @@ impl ComputeNode {
         // 'Close' connection
         drop(client);
 
-        if self.has_feature(ComputeFeature::Migrations) {
-            thread::spawn(move || {
-                let mut client = Client::connect(connstr.as_str(), NoTls)?;
-                handle_migrations(&mut client)
-            });
-        }
+        // Run migrations separately to not hold up cold starts
+        thread::spawn(move || {
+            let mut client = Client::connect(connstr.as_str(), NoTls)?;
+            handle_migrations(&mut client)
+        });
         Ok(())
     }
 
@@ -840,7 +886,12 @@ impl ComputeNode {
             handle_roles(&spec, &mut client)?;
             handle_databases(&spec, &mut client)?;
             handle_role_deletions(&spec, self.connstr.as_str(), &mut client)?;
-            handle_grants(&spec, &mut client, self.connstr.as_str())?;
+            handle_grants(
+                &spec,
+                &mut client,
+                self.connstr.as_str(),
+                self.has_feature(ComputeFeature::AnonExtension),
+            )?;
             handle_extensions(&spec, &mut client)?;
             handle_extension_neon(&mut client)?;
             // We can skip handle_migrations here because a new migration can only appear

compute_tools/src/pg_helpers.rs

@@ -264,9 +264,10 @@ pub fn wait_for_postgres(pg: &mut Child, pgdata: &Path) -> Result<()> {
     // case we miss some events for some reason. Not strictly necessary, but
     // better safe than sorry.
     let (tx, rx) = std::sync::mpsc::channel();
-    let (mut watcher, rx): (Box<dyn Watcher>, _) = match notify::recommended_watcher(move |res| {
+    let watcher_res = notify::recommended_watcher(move |res| {
         let _ = tx.send(res);
-    }) {
+    });
+    let (mut watcher, rx): (Box<dyn Watcher>, _) = match watcher_res {
         Ok(watcher) => (Box::new(watcher), rx),
         Err(e) => {
             match e.kind {

compute_tools/src/spec.rs

@@ -581,7 +581,12 @@ pub fn handle_databases(spec: &ComputeSpec, client: &mut Client) -> Result<()> {
 /// Grant CREATE ON DATABASE to the database owner and do some other alters and grants
 /// to allow users creating trusted extensions and re-creating `public` schema, for example.
 #[instrument(skip_all)]
-pub fn handle_grants(spec: &ComputeSpec, client: &mut Client, connstr: &str) -> Result<()> {
+pub fn handle_grants(
+    spec: &ComputeSpec,
+    client: &mut Client,
+    connstr: &str,
+    enable_anon_extension: bool,
+) -> Result<()> {
     info!("modifying database permissions");
     let existing_dbs = get_existing_dbs(client)?;
 
@@ -678,6 +683,11 @@ pub fn handle_grants(spec: &ComputeSpec, client: &mut Client, connstr: &str) ->
             inlinify(&grant_query)
         );
         db_client.simple_query(&grant_query)?;
+
+        // it is important to run this after all grants
+        if enable_anon_extension {
+            handle_extension_anon(spec, &db.owner, &mut db_client, false)?;
+        }
     }
 
     Ok(())
@@ -766,6 +776,7 @@ BEGIN
     END IF;
 END
 $$;"#,
+        "GRANT pg_monitor TO neon_superuser WITH ADMIN OPTION",
     ];
 
     let mut query = "CREATE SCHEMA IF NOT EXISTS neon_migration";
@@ -809,5 +820,125 @@ $$;"#,
         "Ran {} migrations",
         (migrations.len() - starting_migration_id)
     );
+
+    Ok(())
+}
+
+/// Connect to the database as superuser and pre-create anon extension
+/// if it is present in shared_preload_libraries
+#[instrument(skip_all)]
+pub fn handle_extension_anon(
+    spec: &ComputeSpec,
+    db_owner: &str,
+    db_client: &mut Client,
+    grants_only: bool,
+) -> Result<()> {
+    info!("handle extension anon");
+
+    if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
+        if libs.contains("anon") {
+            if !grants_only {
+                // check if extension is already initialized using anon.is_initialized()
+                let query = "SELECT anon.is_initialized()";
+                match db_client.query(query, &[]) {
+                    Ok(rows) => {
+                        if !rows.is_empty() {
+                            let is_initialized: bool = rows[0].get(0);
+                            if is_initialized {
+                                info!("anon extension is already initialized");
+                                return Ok(());
+                            }
+                        }
+                    }
+                    Err(e) => {
+                        warn!(
+                            "anon extension is_installed check failed with expected error: {}",
+                            e
+                        );
+                    }
+                };
+
+                // Create anon extension if this compute needs it
+                // Users cannot create it themselves, because superuser is required.
+                let mut query = "CREATE EXTENSION IF NOT EXISTS anon CASCADE";
+                info!("creating anon extension with query: {}", query);
+                match db_client.query(query, &[]) {
+                    Ok(_) => {}
+                    Err(e) => {
+                        error!("anon extension creation failed with error: {}", e);
+                        return Ok(());
+                    }
+                }
+
+                // check that extension is installed
+                query = "SELECT extname FROM pg_extension WHERE extname = 'anon'";
+                let rows = db_client.query(query, &[])?;
+                if rows.is_empty() {
+                    error!("anon extension is not installed");
+                    return Ok(());
+                }
+
+                // Initialize anon extension
+                // This also requires superuser privileges, so users cannot do it themselves.
+                query = "SELECT anon.init()";
+                match db_client.query(query, &[]) {
+                    Ok(_) => {}
+                    Err(e) => {
+                        error!("anon.init() failed with error: {}", e);
+                        return Ok(());
+                    }
+                }
+            }
+
+            // check that extension is installed, if not bail early
+            let query = "SELECT extname FROM pg_extension WHERE extname = 'anon'";
+            match db_client.query(query, &[]) {
+                Ok(rows) => {
+                    if rows.is_empty() {
+                        error!("anon extension is not installed");
+                        return Ok(());
+                    }
+                }
+                Err(e) => {
+                    error!("anon extension check failed with error: {}", e);
+                    return Ok(());
+                }
+            };
+
+            let query = format!("GRANT ALL ON SCHEMA anon TO {}", db_owner);
+            info!("granting anon extension permissions with query: {}", query);
+            db_client.simple_query(&query)?;
+
+            // Grant permissions to db_owner to use anon extension functions
+            let query = format!("GRANT ALL ON ALL FUNCTIONS IN SCHEMA anon TO {}", db_owner);
+            info!("granting anon extension permissions with query: {}", query);
+            db_client.simple_query(&query)?;
+
+            // This is needed, because some functions are defined as SECURITY DEFINER.
+            // In Postgres SECURITY DEFINER functions are executed with the privileges
+            // of the owner.
+            // In anon extension this it is needed to access some GUCs, which are only accessible to
+            // superuser. But we've patched postgres to allow db_owner to access them as well.
+            // So we need to change owner of these functions to db_owner.
+            let query = format!("
+                SELECT 'ALTER FUNCTION '||nsp.nspname||'.'||p.proname||'('||pg_get_function_identity_arguments(p.oid)||') OWNER TO {};'
+                from pg_proc p
+                join pg_namespace nsp ON p.pronamespace = nsp.oid
+                where nsp.nspname = 'anon';", db_owner);
+
+            info!("change anon extension functions owner to db owner");
+            db_client.simple_query(&query)?;
+
+            //  affects views as well
+            let query = format!("GRANT ALL ON ALL TABLES IN SCHEMA anon TO {}", db_owner);
+            info!("granting anon extension permissions with query: {}", query);
+            db_client.simple_query(&query)?;
+
+            let query = format!("GRANT ALL ON ALL SEQUENCES IN SCHEMA anon TO {}", db_owner);
+            info!("granting anon extension permissions with query: {}", query);
+            db_client.simple_query(&query)?;
+        }
+    }
+
     Ok(())
 }

control_plane/Cargo.toml

@@ -10,8 +10,6 @@ async-trait.workspace = true
 camino.workspace = true
 clap.workspace = true
 comfy-table.workspace = true
-diesel = { version = "2.1.4", features = ["postgres"]}
-diesel_migrations = { version = "2.1.0", features = ["postgres"]}
 futures.workspace = true
 git-version.workspace = true
 nix.workspace = true

control_plane/attachment_service/Cargo.toml

@@ -24,8 +24,9 @@ tokio.workspace = true
 tokio-util.workspace = true
 tracing.workspace = true
 
-diesel = { version = "2.1.4", features = ["serde_json", "postgres"] }
+diesel = { version = "2.1.4", features = ["serde_json", "postgres", "r2d2"] }
 diesel_migrations = { version = "2.1.0" }
+r2d2 = { version = "0.8.10" }
 
 utils = { path = "../../libs/utils/" }
 metrics = { path = "../../libs/metrics/" }

control_plane/attachment_service/migrations/2024-01-07-211257_create_tenant_shards/up.sql

@@ -7,6 +7,7 @@ CREATE TABLE tenant_shards (
   generation INTEGER NOT NULL,
   generation_pageserver BIGINT NOT NULL,
   placement_policy VARCHAR NOT NULL,
+  splitting SMALLINT NOT NULL,
   -- config is JSON encoded, opaque to the database.
   config TEXT NOT NULL
 );

control_plane/attachment_service/src/compute_hook.rs

@@ -170,7 +170,7 @@ impl ComputeHook {
         reconfigure_request: &ComputeHookNotifyRequest,
         cancel: &CancellationToken,
     ) -> Result<(), NotifyError> {
-        let req = client.request(Method::POST, url);
+        let req = client.request(Method::PUT, url);
         let req = if let Some(value) = &self.authorization_header {
             req.header(reqwest::header::AUTHORIZATION, value)
         } else {
@@ -240,7 +240,7 @@ impl ComputeHook {
         let client = reqwest::Client::new();
         backoff::retry(
             || self.do_notify_iteration(&client, url, &reconfigure_request, cancel),
-            |e| matches!(e, NotifyError::Fatal(_)),
+            |e| matches!(e, NotifyError::Fatal(_) | NotifyError::Unexpected(_)),
             3,
             10,
             "Send compute notification",

control_plane/attachment_service/src/http.rs

@@ -3,7 +3,8 @@ use crate::service::{Service, STARTUP_RECONCILE_TIMEOUT};
 use hyper::{Body, Request, Response};
 use hyper::{StatusCode, Uri};
 use pageserver_api::models::{
-    TenantCreateRequest, TenantLocationConfigRequest, TimelineCreateRequest,
+    TenantCreateRequest, TenantLocationConfigRequest, TenantShardSplitRequest,
+    TimelineCreateRequest,
 };
 use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api;
@@ -41,7 +42,7 @@ pub struct HttpState {
 
 impl HttpState {
     pub fn new(service: Arc<crate::service::Service>, auth: Option<Arc<SwappableJwtAuth>>) -> Self {
-        let allowlist_routes = ["/status"]
+        let allowlist_routes = ["/status", "/ready", "/metrics"]
             .iter()
             .map(|v| v.parse().unwrap())
             .collect::<Vec<_>>();
@@ -279,6 +280,12 @@ async fn handle_node_list(req: Request<Body>) -> Result<Response<Body>, ApiError
     json_response(StatusCode::OK, state.service.node_list().await?)
 }
 
+async fn handle_node_drop(req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let state = get_state(&req);
+    let node_id: NodeId = parse_request_param(&req, "node_id")?;
+    json_response(StatusCode::OK, state.service.node_drop(node_id).await?)
+}
+
 async fn handle_node_configure(mut req: Request<Body>) -> Result<Response<Body>, ApiError> {
     let node_id: NodeId = parse_request_param(&req, "node_id")?;
     let config_req = json_request::<NodeConfigureRequest>(&mut req).await?;
@@ -292,6 +299,19 @@ async fn handle_node_configure(mut req: Request<Body>) -> Result<Response<Body>,
     json_response(StatusCode::OK, state.service.node_configure(config_req)?)
 }
 
+async fn handle_tenant_shard_split(
+    service: Arc<Service>,
+    mut req: Request<Body>,
+) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    let split_req = json_request::<TenantShardSplitRequest>(&mut req).await?;
+
+    json_response(
+        StatusCode::OK,
+        service.tenant_shard_split(tenant_id, split_req).await?,
+    )
+}
+
 async fn handle_tenant_shard_migrate(
     service: Arc<Service>,
     mut req: Request<Body>,
@@ -306,11 +326,29 @@ async fn handle_tenant_shard_migrate(
     )
 }
 
+async fn handle_tenant_drop(req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let tenant_id: TenantId = parse_request_param(&req, "tenant_id")?;
+    let state = get_state(&req);
+
+    json_response(StatusCode::OK, state.service.tenant_drop(tenant_id).await?)
+}
+
 /// Status endpoint is just used for checking that our HTTP listener is up
 async fn handle_status(_req: Request<Body>) -> Result<Response<Body>, ApiError> {
     json_response(StatusCode::OK, ())
 }
 
+/// Readiness endpoint indicates when we're done doing startup I/O (e.g. reconciling
+/// with remote pageserver nodes).  This is intended for use as a kubernetes readiness probe.
+async fn handle_ready(req: Request<Body>) -> Result<Response<Body>, ApiError> {
+    let state = get_state(&req);
+    if state.service.startup_complete.is_ready() {
+        json_response(StatusCode::OK, ())
+    } else {
+        json_response(StatusCode::SERVICE_UNAVAILABLE, ())
+    }
+}
+
 impl From<ReconcileError> for ApiError {
     fn from(value: ReconcileError) -> Self {
         ApiError::Conflict(format!("Reconciliation error: {}", value))
@@ -366,6 +404,7 @@ pub fn make_router(
         .data(Arc::new(HttpState::new(service, auth)))
         // Non-prefixed generic endpoints (status, metrics)
         .get("/status", |r| request_span(r, handle_status))
+        .get("/ready", |r| request_span(r, handle_ready))
         // Upcalls for the pageserver: point the pageserver's `control_plane_api` config to this prefix
         .post("/upcall/v1/re-attach", |r| {
             request_span(r, handle_re_attach)
@@ -376,6 +415,12 @@ pub fn make_router(
             request_span(r, handle_attach_hook)
         })
         .post("/debug/v1/inspect", |r| request_span(r, handle_inspect))
+        .post("/debug/v1/tenant/:tenant_id/drop", |r| {
+            request_span(r, handle_tenant_drop)
+        })
+        .post("/debug/v1/node/:node_id/drop", |r| {
+            request_span(r, handle_node_drop)
+        })
         .get("/control/v1/tenant/:tenant_id/locate", |r| {
             tenant_service_handler(r, handle_tenant_locate)
         })
@@ -391,6 +436,9 @@ pub fn make_router(
         .put("/control/v1/tenant/:tenant_shard_id/migrate", |r| {
             tenant_service_handler(r, handle_tenant_shard_migrate)
         })
+        .put("/control/v1/tenant/:tenant_id/shard_split", |r| {
+            tenant_service_handler(r, handle_tenant_shard_split)
+        })
         // Tenant operations
         // The ^/v1/ endpoints act as a "Virtual Pageserver", enabling shard-naive clients to call into
         // this service to manage tenants that actually consist of many tenant shards, as if they are a single entity.

control_plane/attachment_service/src/main.rs

@@ -170,6 +170,7 @@ impl Secrets {
     }
 }
 
+/// Execute the diesel migrations that are built into this binary
 async fn migration_run(database_url: &str) -> anyhow::Result<()> {
     use diesel::PgConnection;
     use diesel_migrations::{HarnessWithOutput, MigrationHarness};
@@ -183,8 +184,18 @@ async fn migration_run(database_url: &str) -> anyhow::Result<()> {
     Ok(())
 }
 
-#[tokio::main]
-async fn main() -> anyhow::Result<()> {
+fn main() -> anyhow::Result<()> {
+    tokio::runtime::Builder::new_current_thread()
+        // We use spawn_blocking for database operations, so require approximately
+        // as many blocking threads as we will open database connections.
+        .max_blocking_threads(Persistence::MAX_CONNECTIONS as usize)
+        .enable_all()
+        .build()
+        .unwrap()
+        .block_on(async_main())
+}
+
+async fn async_main() -> anyhow::Result<()> {
     let launch_ts = Box::leak(Box::new(LaunchTimestamp::generate()));
 
     logging::init(

control_plane/attachment_service/src/persistence.rs

@@ -1,6 +1,9 @@
+pub(crate) mod split_state;
 use std::collections::HashMap;
 use std::str::FromStr;
+use std::time::Duration;
 
+use self::split_state::SplitState;
 use camino::Utf8Path;
 use camino::Utf8PathBuf;
 use control_plane::attachment_service::{NodeAvailability, NodeSchedulingPolicy};
@@ -44,7 +47,7 @@ use crate::PlacementPolicy;
 /// updated, and reads of nodes are always from memory, not the database.  We only require that
 /// we can UPDATE a node's scheduling mode reasonably quickly to mark a bad node offline.
 pub struct Persistence {
-    database_url: String,
+    connection_pool: diesel::r2d2::Pool<diesel::r2d2::ConnectionManager<PgConnection>>,
 
     // In test environments, we support loading+saving a JSON file.  This is temporary, for the benefit of
     // test_compatibility.py, so that we don't have to commit to making the database contents fully backward/forward
@@ -64,6 +67,8 @@ pub(crate) enum DatabaseError {
     Query(#[from] diesel::result::Error),
     #[error(transparent)]
     Connection(#[from] diesel::result::ConnectionError),
+    #[error(transparent)]
+    ConnectionPool(#[from] r2d2::Error),
     #[error("Logical error: {0}")]
     Logical(String),
 }
@@ -71,9 +76,31 @@ pub(crate) enum DatabaseError {
 pub(crate) type DatabaseResult<T> = Result<T, DatabaseError>;
 
 impl Persistence {
+    // The default postgres connection limit is 100.  We use up to 99, to leave one free for a human admin under
+    // normal circumstances.  This assumes we have exclusive use of the database cluster to which we connect.
+    pub const MAX_CONNECTIONS: u32 = 99;
+
+    // We don't want to keep a lot of connections alive: close them down promptly if they aren't being used.
+    const IDLE_CONNECTION_TIMEOUT: Duration = Duration::from_secs(10);
+    const MAX_CONNECTION_LIFETIME: Duration = Duration::from_secs(60);
+
     pub fn new(database_url: String, json_path: Option<Utf8PathBuf>) -> Self {
+        let manager = diesel::r2d2::ConnectionManager::<PgConnection>::new(database_url);
+
+        // We will use a connection pool: this is primarily to _limit_ our connection count, rather than to optimize time
+        // to execute queries (database queries are not generally on latency-sensitive paths).
+        let connection_pool = diesel::r2d2::Pool::builder()
+            .max_size(Self::MAX_CONNECTIONS)
+            .max_lifetime(Some(Self::MAX_CONNECTION_LIFETIME))
+            .idle_timeout(Some(Self::IDLE_CONNECTION_TIMEOUT))
+            // Always keep at least one connection ready to go
+            .min_idle(Some(1))
+            .test_on_check_out(true)
+            .build(manager)
+            .expect("Could not build connection pool");
+
         Self {
-            database_url,
+            connection_pool,
             json_path,
         }
     }
@@ -84,14 +111,10 @@ impl Persistence {
         F: Fn(&mut PgConnection) -> DatabaseResult<R> + Send + 'static,
         R: Send + 'static,
     {
-        let database_url = self.database_url.clone();
-        tokio::task::spawn_blocking(move || -> DatabaseResult<R> {
-            // TODO: connection pooling, such as via diesel::r2d2
-            let mut conn = PgConnection::establish(&database_url)?;
-            func(&mut conn)
-        })
-        .await
-        .expect("Task panic")
+        let mut conn = self.connection_pool.get()?;
+        tokio::task::spawn_blocking(move || -> DatabaseResult<R> { func(&mut conn) })
+            .await
+            .expect("Task panic")
     }
 
     /// When a node is first registered, persist it before using it for anything
@@ -237,7 +260,6 @@ impl Persistence {
 
     /// Ordering: call this _after_ deleting the tenant on pageservers, but _before_ dropping state for
     /// the tenant from memory on this server.
-    #[allow(unused)]
     pub(crate) async fn delete_tenant(&self, del_tenant_id: TenantId) -> DatabaseResult<()> {
         use crate::schema::tenant_shards::dsl::*;
         self.with_conn(move |conn| -> DatabaseResult<()> {
@@ -250,6 +272,18 @@ impl Persistence {
         .await
     }
 
+    pub(crate) async fn delete_node(&self, del_node_id: NodeId) -> DatabaseResult<()> {
+        use crate::schema::nodes::dsl::*;
+        self.with_conn(move |conn| -> DatabaseResult<()> {
+            diesel::delete(nodes)
+                .filter(node_id.eq(del_node_id.0 as i64))
+                .execute(conn)?;
+
+            Ok(())
+        })
+        .await
+    }
+
     /// When a tenant invokes the /re-attach API, this function is responsible for doing an efficient
     /// batched increment of the generations of all tenants whose generation_pageserver is equal to
     /// the node that called /re-attach.
@@ -342,19 +376,105 @@ impl Persistence {
         Ok(())
     }
 
-    // TODO: when we start shard splitting, we must durably mark the tenant so that
-    // on restart, we know that we must go through recovery (list shards that exist
-    // and pick up where we left off and/or revert to parent shards).
-    #[allow(dead_code)]
-    pub(crate) async fn begin_shard_split(&self, _tenant_id: TenantId) -> anyhow::Result<()> {
-        todo!();
+    // When we start shard splitting, we must durably mark the tenant so that
+    // on restart, we know that we must go through recovery.
+    //
+    // We create the child shards here, so that they will be available for increment_generation calls
+    // if some pageserver holding a child shard needs to restart before the overall tenant split is complete.
+    pub(crate) async fn begin_shard_split(
+        &self,
+        old_shard_count: ShardCount,
+        split_tenant_id: TenantId,
+        parent_to_children: Vec<(TenantShardId, Vec<TenantShardPersistence>)>,
+    ) -> DatabaseResult<()> {
+        use crate::schema::tenant_shards::dsl::*;
+        self.with_conn(move |conn| -> DatabaseResult<()> {
+            conn.transaction(|conn| -> DatabaseResult<()> {
+                // Mark parent shards as splitting
+
+                let expect_parent_records = std::cmp::max(1, old_shard_count.0);
+
+                let updated = diesel::update(tenant_shards)
+                    .filter(tenant_id.eq(split_tenant_id.to_string()))
+                    .filter(shard_count.eq(old_shard_count.0 as i32))
+                    .set((splitting.eq(1),))
+                    .execute(conn)?;
+                if u8::try_from(updated)
+                    .map_err(|_| DatabaseError::Logical(
+                        format!("Overflow existing shard count {} while splitting", updated))
+                    )? != expect_parent_records {
+                    // Perhaps a deletion or another split raced with this attempt to split, mutating
+                    // the parent shards that we intend to split. In this case the split request should fail.
+                    return Err(DatabaseError::Logical(
+                        format!("Unexpected existing shard count {updated} when preparing tenant for split (expected {expect_parent_records})")
+                    ));
+                }
+
+                // FIXME: spurious clone to sidestep closure move rules
+                let parent_to_children = parent_to_children.clone();
+
+                // Insert child shards
+                for (parent_shard_id, children) in parent_to_children {
+                    let mut parent = crate::schema::tenant_shards::table
+                        .filter(tenant_id.eq(parent_shard_id.tenant_id.to_string()))
+                        .filter(shard_number.eq(parent_shard_id.shard_number.0 as i32))
+                        .filter(shard_count.eq(parent_shard_id.shard_count.0 as i32))
+                        .load::<TenantShardPersistence>(conn)?;
+                    let parent = if parent.len() != 1 {
+                        return Err(DatabaseError::Logical(format!(
+                            "Parent shard {parent_shard_id} not found"
+                        )));
+                    } else {
+                        parent.pop().unwrap()
+                    };
+                    for mut shard in children {
+                        // Carry the parent's generation into the child
+                        shard.generation = parent.generation;
+
+                        debug_assert!(shard.splitting == SplitState::Splitting);
+                        diesel::insert_into(tenant_shards)
+                            .values(shard)
+                            .execute(conn)?;
+                    }
+                }
+
+                Ok(())
+            })?;
+
+            Ok(())
+        })
+        .await
     }
 
-    // TODO: when we finish shard splitting, we must atomically clean up the old shards
+    // When we finish shard splitting, we must atomically clean up the old shards
     // and insert the new shards, and clear the splitting marker.
-    #[allow(dead_code)]
-    pub(crate) async fn complete_shard_split(&self, _tenant_id: TenantId) -> anyhow::Result<()> {
-        todo!();
+    pub(crate) async fn complete_shard_split(
+        &self,
+        split_tenant_id: TenantId,
+        old_shard_count: ShardCount,
+    ) -> DatabaseResult<()> {
+        use crate::schema::tenant_shards::dsl::*;
+        self.with_conn(move |conn| -> DatabaseResult<()> {
+            conn.transaction(|conn| -> QueryResult<()> {
+                // Drop parent shards
+                diesel::delete(tenant_shards)
+                    .filter(tenant_id.eq(split_tenant_id.to_string()))
+                    .filter(shard_count.eq(old_shard_count.0 as i32))
+                    .execute(conn)?;
+
+                // Clear sharding flag
+                let updated = diesel::update(tenant_shards)
+                    .filter(tenant_id.eq(split_tenant_id.to_string()))
+                    .set((splitting.eq(0),))
+                    .execute(conn)?;
+                debug_assert!(updated > 0);
+
+                Ok(())
+            })?;
+
+            Ok(())
+        })
+        .await
     }
 }
 
@@ -382,6 +502,8 @@ pub(crate) struct TenantShardPersistence {
     #[serde(default)]
     pub(crate) placement_policy: String,
     #[serde(default)]
+    pub(crate) splitting: SplitState,
+    #[serde(default)]
     pub(crate) config: String,
 }
 

control_plane/attachment_service/src/persistence/split_state.rs

@@ -0,0 +1,46 @@
+use diesel::pg::{Pg, PgValue};
+use diesel::{
+    deserialize::FromSql, deserialize::FromSqlRow, expression::AsExpression, serialize::ToSql,
+    sql_types::Int2,
+};
+use serde::{Deserialize, Serialize};
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, PartialOrd, Ord, FromSqlRow, AsExpression)]
+#[diesel(sql_type = SplitStateSQLRepr)]
+#[derive(Deserialize, Serialize)]
+pub enum SplitState {
+    Idle = 0,
+    Splitting = 1,
+}
+
+impl Default for SplitState {
+    fn default() -> Self {
+        Self::Idle
+    }
+}
+
+type SplitStateSQLRepr = Int2;
+
+impl ToSql<SplitStateSQLRepr, Pg> for SplitState {
+    fn to_sql<'a>(
+        &'a self,
+        out: &'a mut diesel::serialize::Output<Pg>,
+    ) -> diesel::serialize::Result {
+        let raw_value: i16 = *self as i16;
+        let mut new_out = out.reborrow();
+        ToSql::<SplitStateSQLRepr, Pg>::to_sql(&raw_value, &mut new_out)
+    }
+}
+
+impl FromSql<SplitStateSQLRepr, Pg> for SplitState {
+    fn from_sql(pg_value: PgValue) -> diesel::deserialize::Result<Self> {
+        match FromSql::<SplitStateSQLRepr, Pg>::from_sql(pg_value).map(|v| match v {
+            0 => Some(Self::Idle),
+            1 => Some(Self::Splitting),
+            _ => None,
+        })? {
+            Some(v) => Ok(v),
+            None => Err(format!("Invalid SplitState value, was: {:?}", pg_value.as_bytes()).into()),
+        }
+    }
+}

control_plane/attachment_service/src/schema.rs

@@ -20,6 +20,7 @@ diesel::table! {
         generation -> Int4,
         generation_pageserver -> Int8,
         placement_policy -> Varchar,
+        splitting -> Int2,
         config -> Text,
     }
 }

control_plane/attachment_service/src/service.rs

@@ -1,5 +1,6 @@
 use std::{
-    collections::{BTreeMap, HashMap},
+    cmp::Ordering,
+    collections::{BTreeMap, HashMap, HashSet},
     str::FromStr,
     sync::Arc,
     time::{Duration, Instant},
@@ -23,13 +24,14 @@ use pageserver_api::{
     models::{
         LocationConfig, LocationConfigMode, ShardParameters, TenantConfig, TenantCreateRequest,
         TenantLocationConfigRequest, TenantLocationConfigResponse, TenantShardLocation,
-        TimelineCreateRequest, TimelineInfo,
+        TenantShardSplitRequest, TenantShardSplitResponse, TimelineCreateRequest, TimelineInfo,
     },
     shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize, TenantShardId},
 };
 use pageserver_client::mgmt_api;
 use tokio_util::sync::CancellationToken;
 use utils::{
+    backoff,
     completion::Barrier,
     generation::Generation,
     http::error::ApiError,
@@ -40,7 +42,11 @@ use utils::{
 use crate::{
     compute_hook::{self, ComputeHook},
     node::Node,
-    persistence::{DatabaseError, NodePersistence, Persistence, TenantShardPersistence},
+    persistence::{
+        split_state::SplitState, DatabaseError, NodePersistence, Persistence,
+        TenantShardPersistence,
+    },
+    reconciler::attached_location_conf,
     scheduler::Scheduler,
     tenant_state::{
         IntentState, ObservedState, ObservedStateLocation, ReconcileResult, ReconcileWaitError,
@@ -103,7 +109,9 @@ impl From<DatabaseError> for ApiError {
         match err {
             DatabaseError::Query(e) => ApiError::InternalServerError(e.into()),
             // FIXME: ApiError doesn't have an Unavailable variant, but ShuttingDown maps to 503.
-            DatabaseError::Connection(_e) => ApiError::ShuttingDown,
+            DatabaseError::Connection(_) | DatabaseError::ConnectionPool(_) => {
+                ApiError::ShuttingDown
+            }
             DatabaseError::Logical(reason) => {
                 ApiError::InternalServerError(anyhow::anyhow!(reason))
             }
@@ -143,31 +151,71 @@ impl Service {
         // indeterminate, same as in [`ObservedStateLocation`])
         let mut observed = HashMap::new();
 
-        let nodes = {
-            let locked = self.inner.read().unwrap();
-            locked.nodes.clone()
-        };
+        let mut nodes_online = HashSet::new();
+
+        // TODO: give Service a cancellation token for clean shutdown
+        let cancel = CancellationToken::new();
 
         // TODO: issue these requests concurrently
-        for node in nodes.values() {
-            let client = mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+        {
+            let nodes = {
+                let locked = self.inner.read().unwrap();
+                locked.nodes.clone()
+            };
+            for node in nodes.values() {
+                let http_client = reqwest::ClientBuilder::new()
+                    .timeout(Duration::from_secs(5))
+                    .build()
+                    .expect("Failed to construct HTTP client");
+                let client = mgmt_api::Client::from_client(
+                    http_client,
+                    node.base_url(),
+                    self.config.jwt_token.as_deref(),
+                );
 
-            tracing::info!("Scanning shards on node {}...", node.id);
-            match client.list_location_config().await {
-                Err(e) => {
-                    tracing::warn!("Could not contact pageserver {} ({e})", node.id);
-                    // TODO: be more tolerant, apply a generous 5-10 second timeout with retries, in case
-                    // pageserver is being restarted at the same time as we are
+                fn is_fatal(e: &mgmt_api::Error) -> bool {
+                    use mgmt_api::Error::*;
+                    match e {
+                        ReceiveBody(_) | ReceiveErrorBody(_) => false,
+                        ApiError(StatusCode::SERVICE_UNAVAILABLE, _)
+                        | ApiError(StatusCode::GATEWAY_TIMEOUT, _)
+                        | ApiError(StatusCode::REQUEST_TIMEOUT, _) => false,
+                        ApiError(_, _) => true,
+                    }
                 }
-                Ok(listing) => {
-                    tracing::info!(
-                        "Received {} shard statuses from pageserver {}, setting it to Active",
-                        listing.tenant_shards.len(),
-                        node.id
-                    );
 
-                    for (tenant_shard_id, conf_opt) in listing.tenant_shards {
-                        observed.insert(tenant_shard_id, (node.id, conf_opt));
+                let list_response = backoff::retry(
+                    || client.list_location_config(),
+                    is_fatal,
+                    1,
+                    5,
+                    "Location config listing",
+                    &cancel,
+                )
+                .await;
+                let Some(list_response) = list_response else {
+                    tracing::info!("Shutdown during startup_reconcile");
+                    return;
+                };
+
+                tracing::info!("Scanning shards on node {}...", node.id);
+                match list_response {
+                    Err(e) => {
+                        tracing::warn!("Could not contact pageserver {} ({e})", node.id);
+                        // TODO: be more tolerant, do some retries, in case
+                        // pageserver is being restarted at the same time as we are
+                    }
+                    Ok(listing) => {
+                        tracing::info!(
+                            "Received {} shard statuses from pageserver {}, setting it to Active",
+                            listing.tenant_shards.len(),
+                            node.id
+                        );
+                        nodes_online.insert(node.id);
+
+                        for (tenant_shard_id, conf_opt) in listing.tenant_shards {
+                            observed.insert(tenant_shard_id, (node.id, conf_opt));
+                        }
                     }
                 }
             }
@@ -178,8 +226,19 @@ impl Service {
         let mut compute_notifications = Vec::new();
 
         // Populate intent and observed states for all tenants, based on reported state on pageservers
-        let shard_count = {
+        let (shard_count, nodes) = {
             let mut locked = self.inner.write().unwrap();
+
+            // Mark nodes online if they responded to us: nodes are offline by default after a restart.
+            let mut nodes = (*locked.nodes).clone();
+            for (node_id, node) in nodes.iter_mut() {
+                if nodes_online.contains(node_id) {
+                    node.availability = NodeAvailability::Active;
+                }
+            }
+            locked.nodes = Arc::new(nodes);
+            let nodes = locked.nodes.clone();
+
             for (tenant_shard_id, (node_id, observed_loc)) in observed {
                 let Some(tenant_state) = locked.tenants.get_mut(&tenant_shard_id) else {
                     cleanup.push((tenant_shard_id, node_id));
@@ -211,7 +270,7 @@ impl Service {
                 }
             }
 
-            locked.tenants.len()
+            (locked.tenants.len(), nodes)
         };
 
         // TODO: if any tenant's intent now differs from its loaded generation_pageserver, we should clear that
@@ -272,9 +331,8 @@ impl Service {
         let stream = futures::stream::iter(compute_notifications.into_iter())
             .map(|(tenant_shard_id, node_id)| {
                 let compute_hook = compute_hook.clone();
+                let cancel = cancel.clone();
                 async move {
-                    // TODO: give Service a cancellation token for clean shutdown
-                    let cancel = CancellationToken::new();
                     if let Err(e) = compute_hook.notify(tenant_shard_id, node_id, &cancel).await {
                         tracing::error!(
                             tenant_shard_id=%tenant_shard_id,
@@ -380,7 +438,7 @@ impl Service {
             ))),
             config,
             persistence,
-            startup_complete,
+            startup_complete: startup_complete.clone(),
         });
 
         let result_task_this = this.clone();
@@ -474,6 +532,7 @@ impl Service {
                 generation_pageserver: i64::MAX,
                 placement_policy: serde_json::to_string(&PlacementPolicy::default()).unwrap(),
                 config: serde_json::to_string(&TenantConfig::default()).unwrap(),
+                splitting: SplitState::default(),
             };
 
             match self.persistence.insert_tenant_shards(vec![tsp]).await {
@@ -716,6 +775,7 @@ impl Service {
                 generation_pageserver: i64::MAX,
                 placement_policy: serde_json::to_string(&placement_policy).unwrap(),
                 config: serde_json::to_string(&create_req.config).unwrap(),
+                splitting: SplitState::default(),
             })
             .collect();
         self.persistence
@@ -975,6 +1035,10 @@ impl Service {
             }
         };
 
+        // TODO: if we timeout/fail on reconcile, we should still succeed this request,
+        // because otherwise a broken compute hook causes a feedback loop where
+        // location_config returns 500 and gets retried forever.
+
         if let Some(create_req) = maybe_create {
             let create_resp = self.tenant_create(create_req).await?;
             result.shards = create_resp
@@ -987,7 +1051,15 @@ impl Service {
                 .collect();
         } else {
             // This was an update, wait for reconciliation
-            self.await_waiters(waiters).await?;
+            if let Err(e) = self.await_waiters(waiters).await {
+                // Do not treat a reconcile error as fatal: we have already applied any requested
+                // Intent changes, and the reconcile can fail for external reasons like unavailable
+                // compute notification API.  In these cases, it is important that we do not
+                // cause the cloud control plane to retry forever on this API.
+                tracing::warn!(
+                    "Failed to reconcile after /location_config: {e}, returning success anyway"
+                );
+            }
         }
 
         Ok(result)
@@ -1090,6 +1162,7 @@ impl Service {
         self.ensure_attached_wait(tenant_id).await?;
 
         // TODO: refuse to do this if shard splitting is in progress
+        // (https://github.com/neondatabase/neon/issues/6676)
         let targets = {
             let locked = self.inner.read().unwrap();
             let mut targets = Vec::new();
@@ -1170,6 +1243,7 @@ impl Service {
         self.ensure_attached_wait(tenant_id).await?;
 
         // TODO: refuse to do this if shard splitting is in progress
+        // (https://github.com/neondatabase/neon/issues/6676)
         let targets = {
             let locked = self.inner.read().unwrap();
             let mut targets = Vec::new();
@@ -1342,6 +1416,326 @@ impl Service {
         })
     }
 
+    pub(crate) async fn tenant_shard_split(
+        &self,
+        tenant_id: TenantId,
+        split_req: TenantShardSplitRequest,
+    ) -> Result<TenantShardSplitResponse, ApiError> {
+        let mut policy = None;
+        let mut shard_ident = None;
+
+        // TODO: put a cancellation token on Service for clean shutdown
+        let cancel = CancellationToken::new();
+
+        // A parent shard which will be split
+        struct SplitTarget {
+            parent_id: TenantShardId,
+            node: Node,
+            child_ids: Vec<TenantShardId>,
+        }
+
+        // Validate input, and calculate which shards we will create
+        let (old_shard_count, targets, compute_hook) = {
+            let locked = self.inner.read().unwrap();
+
+            let pageservers = locked.nodes.clone();
+
+            let mut targets = Vec::new();
+
+            // In case this is a retry, count how many already-split shards we found
+            let mut children_found = Vec::new();
+            let mut old_shard_count = None;
+
+            for (tenant_shard_id, shard) in
+                locked.tenants.range(TenantShardId::tenant_range(tenant_id))
+            {
+                match shard.shard.count.0.cmp(&split_req.new_shard_count) {
+                    Ordering::Equal => {
+                        //  Already split this
+                        children_found.push(*tenant_shard_id);
+                        continue;
+                    }
+                    Ordering::Greater => {
+                        return Err(ApiError::BadRequest(anyhow::anyhow!(
+                            "Requested count {} but already have shards at count {}",
+                            split_req.new_shard_count,
+                            shard.shard.count.0
+                        )));
+                    }
+                    Ordering::Less => {
+                        // Fall through: this shard has lower count than requested,
+                        // is a candidate for splitting.
+                    }
+                }
+
+                match old_shard_count {
+                    None => old_shard_count = Some(shard.shard.count),
+                    Some(old_shard_count) => {
+                        if old_shard_count != shard.shard.count {
+                            // We may hit this case if a caller asked for two splits to
+                            // different sizes, before the first one is complete.
+                            // e.g. 1->2, 2->4, where the 4 call comes while we have a mixture
+                            // of shard_count=1 and shard_count=2 shards in the map.
+                            return Err(ApiError::Conflict(
+                                "Cannot split, currently mid-split".to_string(),
+                            ));
+                        }
+                    }
+                }
+                if policy.is_none() {
+                    policy = Some(shard.policy.clone());
+                }
+                if shard_ident.is_none() {
+                    shard_ident = Some(shard.shard);
+                }
+
+                if tenant_shard_id.shard_count == ShardCount(split_req.new_shard_count) {
+                    tracing::info!(
+                        "Tenant shard {} already has shard count {}",
+                        tenant_shard_id,
+                        split_req.new_shard_count
+                    );
+                    continue;
+                }
+
+                let node_id =
+                    shard
+                        .intent
+                        .attached
+                        .ok_or(ApiError::BadRequest(anyhow::anyhow!(
+                            "Cannot split a tenant that is not attached"
+                        )))?;
+
+                let node = pageservers
+                    .get(&node_id)
+                    .expect("Pageservers may not be deleted while referenced");
+
+                // TODO: if any reconciliation is currently in progress for this shard, wait for it.
+
+                targets.push(SplitTarget {
+                    parent_id: *tenant_shard_id,
+                    node: node.clone(),
+                    child_ids: tenant_shard_id.split(ShardCount(split_req.new_shard_count)),
+                });
+            }
+
+            if targets.is_empty() {
+                if children_found.len() == split_req.new_shard_count as usize {
+                    return Ok(TenantShardSplitResponse {
+                        new_shards: children_found,
+                    });
+                } else {
+                    // No shards found to split, and no existing children found: the
+                    // tenant doesn't exist at all.
+                    return Err(ApiError::NotFound(
+                        anyhow::anyhow!("Tenant {} not found", tenant_id).into(),
+                    ));
+                }
+            }
+
+            (old_shard_count, targets, locked.compute_hook.clone())
+        };
+
+        // unwrap safety: we would have returned above if we didn't find at least one shard to split
+        let old_shard_count = old_shard_count.unwrap();
+        let shard_ident = shard_ident.unwrap();
+        let policy = policy.unwrap();
+
+        // FIXME: we have dropped self.inner lock, and not yet written anything to the database: another
+        // request could occur here, deleting or mutating the tenant.  begin_shard_split checks that the
+        // parent shards exist as expected, but it would be neater to do the above pre-checks within the
+        // same database transaction rather than pre-check in-memory and then maybe-fail the database write.
+        // (https://github.com/neondatabase/neon/issues/6676)
+
+        // Before creating any new child shards in memory or on the pageservers, persist them: this
+        // enables us to ensure that we will always be able to clean up if something goes wrong.  This also
+        // acts as the protection against two concurrent attempts to split: one of them will get a database
+        // error trying to insert the child shards.
+        let mut child_tsps = Vec::new();
+        for target in &targets {
+            let mut this_child_tsps = Vec::new();
+            for child in &target.child_ids {
+                let mut child_shard = shard_ident;
+                child_shard.number = child.shard_number;
+                child_shard.count = child.shard_count;
+
+                this_child_tsps.push(TenantShardPersistence {
+                    tenant_id: child.tenant_id.to_string(),
+                    shard_number: child.shard_number.0 as i32,
+                    shard_count: child.shard_count.0 as i32,
+                    shard_stripe_size: shard_ident.stripe_size.0 as i32,
+                    // Note: this generation is a placeholder, [`Persistence::begin_shard_split`] will
+                    // populate the correct generation as part of its transaction, to protect us
+                    // against racing with changes in the state of the parent.
+                    generation: 0,
+                    generation_pageserver: target.node.id.0 as i64,
+                    placement_policy: serde_json::to_string(&policy).unwrap(),
+                    // TODO: get the config out of the map
+                    config: serde_json::to_string(&TenantConfig::default()).unwrap(),
+                    splitting: SplitState::Splitting,
+                });
+            }
+
+            child_tsps.push((target.parent_id, this_child_tsps));
+        }
+
+        if let Err(e) = self
+            .persistence
+            .begin_shard_split(old_shard_count, tenant_id, child_tsps)
+            .await
+        {
+            match e {
+                DatabaseError::Query(diesel::result::Error::DatabaseError(
+                    DatabaseErrorKind::UniqueViolation,
+                    _,
+                )) => {
+                    // Inserting a child shard violated a unique constraint: we raced with another call to
+                    // this function
+                    tracing::warn!("Conflicting attempt to split {tenant_id}: {e}");
+                    return Err(ApiError::Conflict("Tenant is already splitting".into()));
+                }
+                _ => return Err(ApiError::InternalServerError(e.into())),
+            }
+        }
+
+        // FIXME: we have now committed the shard split state to the database, so any subsequent
+        // failure needs to roll it back.  We will later wrap this function in logic to roll back
+        // the split if it fails.
+        // (https://github.com/neondatabase/neon/issues/6676)
+
+        // TODO: issue split calls concurrently (this only matters once we're splitting
+        // N>1 shards into M shards -- initially we're usually splitting 1 shard into N).
+
+        for target in &targets {
+            let SplitTarget {
+                parent_id,
+                node,
+                child_ids,
+            } = target;
+            let client = mgmt_api::Client::new(node.base_url(), self.config.jwt_token.as_deref());
+            let response = client
+                .tenant_shard_split(
+                    *parent_id,
+                    TenantShardSplitRequest {
+                        new_shard_count: split_req.new_shard_count,
+                    },
+                )
+                .await
+                .map_err(|e| ApiError::Conflict(format!("Failed to split {}: {}", parent_id, e)))?;
+
+            tracing::info!(
+                "Split {} into {}",
+                parent_id,
+                response
+                    .new_shards
+                    .iter()
+                    .map(|s| format!("{:?}", s))
+                    .collect::<Vec<_>>()
+                    .join(",")
+            );
+
+            if &response.new_shards != child_ids {
+                // This should never happen: the pageserver should agree with us on how shard splits work.
+                return Err(ApiError::InternalServerError(anyhow::anyhow!(
+                    "Splitting shard {} resulted in unexpected IDs: {:?} (expected {:?})",
+                    parent_id,
+                    response.new_shards,
+                    child_ids
+                )));
+            }
+        }
+
+        // TODO: if the pageserver restarted concurrently with our split API call,
+        // the actual generation of the child shard might differ from the generation
+        // we expect it to have.  In order for our in-database generation to end up
+        // correct, we should carry the child generation back in the response and apply it here
+        // in complete_shard_split (and apply the correct generation in memory)
+        // (or, we can carry generation in the request and reject the request if
+        //  it doesn't match, but that requires more retry logic on this side)
+
+        self.persistence
+            .complete_shard_split(tenant_id, old_shard_count)
+            .await?;
+
+        // Replace all the shards we just split with their children
+        let mut response = TenantShardSplitResponse {
+            new_shards: Vec::new(),
+        };
+        let mut child_locations = Vec::new();
+        {
+            let mut locked = self.inner.write().unwrap();
+            for target in targets {
+                let SplitTarget {
+                    parent_id,
+                    node: _node,
+                    child_ids,
+                } = target;
+                let (pageserver, generation, config) = {
+                    let old_state = locked
+                        .tenants
+                        .remove(&parent_id)
+                        .expect("It was present, we just split it");
+                    (
+                        old_state.intent.attached.unwrap(),
+                        old_state.generation,
+                        old_state.config.clone(),
+                    )
+                };
+
+                locked.tenants.remove(&parent_id);
+
+                for child in child_ids {
+                    let mut child_shard = shard_ident;
+                    child_shard.number = child.shard_number;
+                    child_shard.count = child.shard_count;
+
+                    let mut child_observed: HashMap<NodeId, ObservedStateLocation> = HashMap::new();
+                    child_observed.insert(
+                        pageserver,
+                        ObservedStateLocation {
+                            conf: Some(attached_location_conf(generation, &child_shard, &config)),
+                        },
+                    );
+
+                    let mut child_state = TenantState::new(child, child_shard, policy.clone());
+                    child_state.intent = IntentState::single(Some(pageserver));
+                    child_state.observed = ObservedState {
+                        locations: child_observed,
+                    };
+                    child_state.generation = generation;
+                    child_state.config = config.clone();
+
+                    child_locations.push((child, pageserver));
+
+                    locked.tenants.insert(child, child_state);
+                    response.new_shards.push(child);
+                }
+            }
+        }
+
+        // Send compute notifications for all the new shards
+        let mut failed_notifications = Vec::new();
+        for (child_id, child_ps) in child_locations {
+            if let Err(e) = compute_hook.notify(child_id, child_ps, &cancel).await {
+                tracing::warn!("Failed to update compute of {}->{} during split, proceeding anyway to complete split ({e})",
+                        child_id, child_ps);
+                failed_notifications.push(child_id);
+            }
+        }
+
+        // If we failed any compute notifications, make a note to retry later.
+        if !failed_notifications.is_empty() {
+            let mut locked = self.inner.write().unwrap();
+            for failed in failed_notifications {
+                if let Some(shard) = locked.tenants.get_mut(&failed) {
+                    shard.pending_compute_notification = true;
+                }
+            }
+        }
+
+        Ok(response)
+    }
+
     pub(crate) async fn tenant_shard_migrate(
         &self,
         tenant_shard_id: TenantShardId,
@@ -1410,6 +1804,45 @@ impl Service {
         Ok(TenantShardMigrateResponse {})
     }
 
+    /// This is for debug/support only: we simply drop all state for a tenant, without
+    /// detaching or deleting it on pageservers.
+    pub(crate) async fn tenant_drop(&self, tenant_id: TenantId) -> Result<(), ApiError> {
+        self.persistence.delete_tenant(tenant_id).await?;
+
+        let mut locked = self.inner.write().unwrap();
+        let mut shards = Vec::new();
+        for (tenant_shard_id, _) in locked.tenants.range(TenantShardId::tenant_range(tenant_id)) {
+            shards.push(*tenant_shard_id);
+        }
+
+        for shard in shards {
+            locked.tenants.remove(&shard);
+        }
+
+        Ok(())
+    }
+
+    /// This is for debug/support only: we simply drop all state for a tenant, without
+    /// detaching or deleting it on pageservers.  We do not try and re-schedule any
+    /// tenants that were on this node.
+    ///
+    /// TODO: proper node deletion API that unhooks things more gracefully
+    pub(crate) async fn node_drop(&self, node_id: NodeId) -> Result<(), ApiError> {
+        self.persistence.delete_node(node_id).await?;
+
+        let mut locked = self.inner.write().unwrap();
+
+        for shard in locked.tenants.values_mut() {
+            shard.deref_node(node_id);
+        }
+
+        let mut nodes = (*locked.nodes).clone();
+        nodes.remove(&node_id);
+        locked.nodes = Arc::new(nodes);
+
+        Ok(())
+    }
+
     pub(crate) async fn node_list(&self) -> Result<Vec<NodePersistence>, ApiError> {
         // It is convenient to avoid taking the big lock and converting Node to a serializable
         // structure, by fetching from storage instead of reading in-memory state.

control_plane/attachment_service/src/tenant_state.rs

@@ -193,6 +193,13 @@ impl IntentState {
         result
     }
 
+    pub(crate) fn single(node_id: Option<NodeId>) -> Self {
+        Self {
+            attached: node_id,
+            secondary: vec![],
+        }
+    }
+
     /// When a node goes offline, we update intents to avoid using it
     /// as their attached pageserver.
     ///
@@ -286,6 +293,9 @@ impl TenantState {
         // self.intent refers to pageservers that are offline, and pick other
         // pageservers if so.
 
+        // TODO: respect the splitting bit on tenants: if they are currently splitting then we may not
+        // change their attach location.
+
         // Build the set of pageservers already in use by this tenant, to avoid scheduling
         // more work on the same pageservers we're already using.
         let mut used_pageservers = self.intent.all_pageservers();
@@ -524,4 +534,18 @@ impl TenantState {
             seq: self.sequence,
         })
     }
+
+    // If we had any state at all referring to this node ID, drop it.  Does not
+    // attempt to reschedule.
+    pub(crate) fn deref_node(&mut self, node_id: NodeId) {
+        if self.intent.attached == Some(node_id) {
+            self.intent.attached = None;
+        }
+
+        self.intent.secondary.retain(|n| n != &node_id);
+
+        self.observed.locations.remove(&node_id);
+
+        debug_assert!(!self.intent.all_pageservers().contains(&node_id));
+    }
 }

control_plane/src/attachment_service.rs

@@ -1,20 +1,17 @@
 use crate::{background_process, local_env::LocalEnv};
 use camino::{Utf8Path, Utf8PathBuf};
-use diesel::{
-    backend::Backend,
-    query_builder::{AstPass, QueryFragment, QueryId},
-    Connection, PgConnection, QueryResult, RunQueryDsl,
-};
-use diesel_migrations::{HarnessWithOutput, MigrationHarness};
 use hyper::Method;
 use pageserver_api::{
-    models::{ShardParameters, TenantCreateRequest, TimelineCreateRequest, TimelineInfo},
+    models::{
+        ShardParameters, TenantCreateRequest, TenantShardSplitRequest, TenantShardSplitResponse,
+        TimelineCreateRequest, TimelineInfo,
+    },
     shard::TenantShardId,
 };
 use pageserver_client::mgmt_api::ResponseErrorMessageExt;
 use postgres_backend::AuthType;
 use serde::{de::DeserializeOwned, Deserialize, Serialize};
-use std::{env, str::FromStr};
+use std::str::FromStr;
 use tokio::process::Command;
 use tracing::instrument;
 use url::Url;
@@ -270,37 +267,6 @@ impl AttachmentService {
         .expect("non-Unicode path")
     }
 
-    /// In order to access database migrations, we need to find the Neon source tree
-    async fn find_source_root(&self) -> anyhow::Result<Utf8PathBuf> {
-        // We assume that either prd or our binary is in the source tree. The former is usually
-        // true for automated test runners, the latter is usually true for developer workstations. Often
-        // both are true, which is fine.
-        let candidate_start_points = [
-            // Current working directory
-            Utf8PathBuf::from_path_buf(std::env::current_dir()?).unwrap(),
-            // Directory containing the binary we're running inside
-            Utf8PathBuf::from_path_buf(env::current_exe()?.parent().unwrap().to_owned()).unwrap(),
-        ];
-
-        // For each candidate start point, search through ancestors looking for a neon.git source tree root
-        for start_point in &candidate_start_points {
-            // Start from the build dir: assumes we are running out of a built neon source tree
-            for path in start_point.ancestors() {
-                // A crude approximation: the root of the source tree is whatever contains a "control_plane"
-                // subdirectory.
-                let control_plane = path.join("control_plane");
-                if tokio::fs::try_exists(&control_plane).await? {
-                    return Ok(path.to_owned());
-                }
-            }
-        }
-
-        // Fall-through
-        Err(anyhow::anyhow!(
-            "Could not find control_plane src dir, after searching ancestors of {candidate_start_points:?}"
-        ))
-    }
-
     /// Find the directory containing postgres binaries, such as `initdb` and `pg_ctl`
     ///
     /// This usually uses ATTACHMENT_SERVICE_POSTGRES_VERSION of postgres, but will fall back
@@ -340,69 +306,32 @@ impl AttachmentService {
     ///
     /// Returns the database url
     pub async fn setup_database(&self) -> anyhow::Result<String> {
-        let database_url = format!(
-            "postgresql://localhost:{}/attachment_service",
-            self.postgres_port
-        );
-        println!("Running attachment service database setup...");
-        fn change_database_of_url(database_url: &str, default_database: &str) -> (String, String) {
-            let base = ::url::Url::parse(database_url).unwrap();
-            let database = base.path_segments().unwrap().last().unwrap().to_owned();
-            let mut new_url = base.join(default_database).unwrap();
-            new_url.set_query(base.query());
-            (database, new_url.into())
-        }
+        const DB_NAME: &str = "attachment_service";
+        let database_url = format!("postgresql://localhost:{}/{DB_NAME}", self.postgres_port);
 
-        #[derive(Debug, Clone)]
-        pub struct CreateDatabaseStatement {
-            db_name: String,
-        }
+        let pg_bin_dir = self.get_pg_bin_dir().await?;
+        let createdb_path = pg_bin_dir.join("createdb");
+        let output = Command::new(&createdb_path)
+            .args([
+                "-h",
+                "localhost",
+                "-p",
+                &format!("{}", self.postgres_port),
+                &DB_NAME,
+            ])
+            .output()
+            .await
+            .expect("Failed to spawn createdb");
 
-        impl CreateDatabaseStatement {
-            pub fn new(db_name: &str) -> Self {
-                CreateDatabaseStatement {
-                    db_name: db_name.to_owned(),
-                }
+        if !output.status.success() {
+            let stderr = String::from_utf8(output.stderr).expect("Non-UTF8 output from createdb");
+            if stderr.contains("already exists") {
+                tracing::info!("Database {DB_NAME} already exists");
+            } else {
+                anyhow::bail!("createdb failed with status {}: {stderr}", output.status);
             }
         }
 
-        impl<DB: Backend> QueryFragment<DB> for CreateDatabaseStatement {
-            fn walk_ast<'b>(&'b self, mut out: AstPass<'_, 'b, DB>) -> QueryResult<()> {
-                out.push_sql("CREATE DATABASE ");
-                out.push_identifier(&self.db_name)?;
-                Ok(())
-            }
-        }
-
-        impl<Conn> RunQueryDsl<Conn> for CreateDatabaseStatement {}
-
-        impl QueryId for CreateDatabaseStatement {
-            type QueryId = ();
-
-            const HAS_STATIC_QUERY_ID: bool = false;
-        }
-        if PgConnection::establish(&database_url).is_err() {
-            let (database, postgres_url) = change_database_of_url(&database_url, "postgres");
-            println!("Creating database: {database}");
-            let mut conn = PgConnection::establish(&postgres_url)?;
-            CreateDatabaseStatement::new(&database).execute(&mut conn)?;
-        }
-        let mut conn = PgConnection::establish(&database_url)?;
-
-        let migrations_dir = self
-            .find_source_root()
-            .await?
-            .join("control_plane/attachment_service/migrations");
-
-        let migrations = diesel_migrations::FileBasedMigrations::from_path(migrations_dir)?;
-        println!("Running migrations in {}", migrations.path().display());
-        HarnessWithOutput::write_to_stdout(&mut conn)
-            .run_pending_migrations(migrations)
-            .map(|_| ())
-            .map_err(|e| anyhow::anyhow!(e))?;
-
-        println!("Migrations complete");
-
         Ok(database_url)
     }
 
@@ -648,7 +577,7 @@ impl AttachmentService {
     ) -> anyhow::Result<TenantShardMigrateResponse> {
         self.dispatch(
             Method::PUT,
-            format!("tenant/{tenant_shard_id}/migrate"),
+            format!("control/v1/tenant/{tenant_shard_id}/migrate"),
             Some(TenantShardMigrateRequest {
                 tenant_shard_id,
                 node_id,
@@ -657,6 +586,20 @@ impl AttachmentService {
         .await
     }
 
+    #[instrument(skip(self), fields(%tenant_id, %new_shard_count))]
+    pub async fn tenant_split(
+        &self,
+        tenant_id: TenantId,
+        new_shard_count: u8,
+    ) -> anyhow::Result<TenantShardSplitResponse> {
+        self.dispatch(
+            Method::PUT,
+            format!("control/v1/tenant/{tenant_id}/shard_split"),
+            Some(TenantShardSplitRequest { new_shard_count }),
+        )
+        .await
+    }
+
     #[instrument(skip_all, fields(node_id=%req.node_id))]
     pub async fn node_register(&self, req: NodeRegisterRequest) -> anyhow::Result<()> {
         self.dispatch::<_, ()>(Method::POST, "control/v1/node".to_string(), Some(req))

control_plane/src/background_process.rs

@@ -72,7 +72,6 @@ where
     let log_path = datadir.join(format!("{process_name}.log"));
     let process_log_file = fs::OpenOptions::new()
         .create(true)
-        .write(true)
         .append(true)
         .open(&log_path)
         .with_context(|| {

control_plane/src/bin/neon_local.rs

@@ -575,6 +575,26 @@ async fn handle_tenant(
             println!("{tenant_table}");
             println!("{shard_table}");
         }
+        Some(("shard-split", matches)) => {
+            let tenant_id = get_tenant_id(matches, env)?;
+            let shard_count: u8 = matches.get_one::<u8>("shard-count").cloned().unwrap_or(0);
+
+            let attachment_service = AttachmentService::from_env(env);
+            let result = attachment_service
+                .tenant_split(tenant_id, shard_count)
+                .await?;
+            println!(
+                "Split tenant {} into shards {}",
+                tenant_id,
+                result
+                    .new_shards
+                    .iter()
+                    .map(|s| format!("{:?}", s))
+                    .collect::<Vec<_>>()
+                    .join(",")
+            );
+        }
+
         Some((sub_name, _)) => bail!("Unexpected tenant subcommand '{}'", sub_name),
         None => bail!("no tenant subcommand provided"),
     }
@@ -994,12 +1014,13 @@ async fn handle_endpoint(ep_match: &ArgMatches, env: &local_env::LocalEnv) -> Re
                 .get_one::<String>("endpoint_id")
                 .ok_or_else(|| anyhow!("No endpoint ID was provided to stop"))?;
             let destroy = sub_args.get_flag("destroy");
+            let mode = sub_args.get_one::<String>("mode").expect("has a default");
 
             let endpoint = cplane
                 .endpoints
                 .get(endpoint_id.as_str())
                 .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
-            endpoint.stop(destroy)?;
+            endpoint.stop(mode, destroy)?;
         }
 
         _ => bail!("Unexpected endpoint subcommand '{sub_name}'"),
@@ -1283,7 +1304,7 @@ async fn try_stop_all(env: &local_env::LocalEnv, immediate: bool) {
     match ComputeControlPlane::load(env.clone()) {
         Ok(cplane) => {
             for (_k, node) in cplane.endpoints {
-                if let Err(e) = node.stop(false) {
+                if let Err(e) = node.stop(if immediate { "immediate" } else { "fast " }, false) {
                     eprintln!("postgres stop failed: {e:#}");
                 }
             }
@@ -1524,6 +1545,11 @@ fn cli() -> Command {
             .subcommand(Command::new("status")
                 .about("Human readable summary of the tenant's shards and attachment locations")
                 .arg(tenant_id_arg.clone()))
+            .subcommand(Command::new("shard-split")
+                .about("Increase the number of shards in the tenant")
+                .arg(tenant_id_arg.clone())
+                .arg(Arg::new("shard-count").value_parser(value_parser!(u8)).long("shard-count").action(ArgAction::Set).help("Number of shards in the new tenant (default 1)"))
+                )
         )
         .subcommand(
             Command::new("pageserver")
@@ -1627,7 +1653,16 @@ fn cli() -> Command {
                             .long("destroy")
                             .action(ArgAction::SetTrue)
                             .required(false)
-                        )
+                    )
+                    .arg(
+                        Arg::new("mode")
+                            .help("Postgres shutdown mode, passed to \"pg_ctl -m <mode>\"")
+                            .long("mode")
+                            .action(ArgAction::Set)
+                            .required(false)
+                            .value_parser(["smart", "fast", "immediate"])
+                            .default_value("fast")
+                    )
                 )
 
         )

control_plane/src/endpoint.rs

@@ -290,7 +290,6 @@ impl Endpoint {
         conf.append("wal_log_hints", "off");
         conf.append("max_replication_slots", "10");
         conf.append("hot_standby", "on");
-        conf.append("hot_standby_feedback", "on");
         conf.append("shared_buffers", "1MB");
         conf.append("fsync", "off");
         conf.append("max_connections", "100");
@@ -307,9 +306,6 @@ impl Endpoint {
         // Load the 'neon' extension
         conf.append("shared_preload_libraries", "neon");
 
-		// It is only for testing, in real environment this GUC is passed by control plane
-		conf.append("neon.primary_is_running", "on");
-
         conf.append_line("");
         // Replication-related configurations, such as WAL sending
         match &self.mode {
@@ -765,22 +761,8 @@ impl Endpoint {
         }
     }
 
-    pub fn stop(&self, destroy: bool) -> Result<()> {
-        // If we are going to destroy data directory,
-        // use immediate shutdown mode, otherwise,
-        // shutdown gracefully to leave the data directory sane.
-        //
-        // Postgres is always started from scratch, so stop
-        // without destroy only used for testing and debugging.
-        //
-        self.pg_ctl(
-            if destroy {
-                &["-m", "immediate", "stop"]
-            } else {
-                &["stop"]
-            },
-            &None,
-        )?;
+    pub fn stop(&self, mode: &str, destroy: bool) -> Result<()> {
+        self.pg_ctl(&["-m", mode, "stop"], &None)?;
 
         // Also wait for the compute_ctl process to die. It might have some
         // cleanup work to do after postgres stops, like syncing safekeepers,

libs/compute_api/src/spec.rs

@@ -90,8 +90,8 @@ pub enum ComputeFeature {
     /// track short-lived connections as user activity.
     ActivityMonitorExperimental,
 
-    /// Enable running migrations
-    Migrations,
+    /// Pre-install and initialize anon extension for every database in the cluster
+    AnonExtension,
 
     /// This is a special feature flag that is used to represent unknown feature flags.
     /// Basically all unknown to enum flags are represented as this one. See unit test

libs/desim/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "desim"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+anyhow.workspace = true
+rand.workspace = true
+tracing.workspace = true
+bytes.workspace = true
+utils.workspace = true
+parking_lot.workspace = true
+hex.workspace = true
+scopeguard.workspace = true
+smallvec = { workspace = true, features = ["write"] }
+
+workspace_hack.workspace = true

libs/desim/README.md

@@ -0,0 +1,7 @@
+# Discrete Event SIMulator
+
+This is a library for running simulations of distributed systems. The main idea is borrowed from [FoundationDB](https://www.youtube.com/watch?v=4fFDFbi3toc).
+
+Each node runs as a separate thread. This library was not optimized for speed yet, but it's already much faster than running usual intergration tests in real time, because it uses virtual simulation time and can fast-forward time to skip intervals where all nodes are doing nothing but sleeping or waiting for something.
+
+The original purpose for this library is to test walproposer and safekeeper implementation working together, in a scenarios close to the real world environment. This simulator is determenistic and can inject failures in networking without waiting minutes of wall-time to trigger timeout, which makes it easier to find bugs in our consensus implementation compared to using integration tests.

libs/desim/src/chan.rs

@@ -0,0 +1,108 @@
+use std::{collections::VecDeque, sync::Arc};
+
+use parking_lot::{Mutex, MutexGuard};
+
+use crate::executor::{self, PollSome, Waker};
+
+/// FIFO channel with blocking send and receive. Can be cloned and shared between threads.
+/// Blocking functions should be used only from threads that are managed by the executor.
+pub struct Chan<T> {
+    shared: Arc<State<T>>,
+}
+
+impl<T> Clone for Chan<T> {
+    fn clone(&self) -> Self {
+        Chan {
+            shared: self.shared.clone(),
+        }
+    }
+}
+
+impl<T> Default for Chan<T> {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl<T> Chan<T> {
+    pub fn new() -> Chan<T> {
+        Chan {
+            shared: Arc::new(State {
+                queue: Mutex::new(VecDeque::new()),
+                waker: Waker::new(),
+            }),
+        }
+    }
+
+    /// Get a message from the front of the queue, block if the queue is empty.
+    /// If not called from the executor thread, it can block forever.
+    pub fn recv(&self) -> T {
+        self.shared.recv()
+    }
+
+    /// Panic if the queue is empty.
+    pub fn must_recv(&self) -> T {
+        self.shared
+            .try_recv()
+            .expect("message should've been ready")
+    }
+
+    /// Get a message from the front of the queue, return None if the queue is empty.
+    /// Never blocks.
+    pub fn try_recv(&self) -> Option<T> {
+        self.shared.try_recv()
+    }
+
+    /// Send a message to the back of the queue.
+    pub fn send(&self, t: T) {
+        self.shared.send(t);
+    }
+}
+
+struct State<T> {
+    queue: Mutex<VecDeque<T>>,
+    waker: Waker,
+}
+
+impl<T> State<T> {
+    fn send(&self, t: T) {
+        self.queue.lock().push_back(t);
+        self.waker.wake_all();
+    }
+
+    fn try_recv(&self) -> Option<T> {
+        let mut q = self.queue.lock();
+        q.pop_front()
+    }
+
+    fn recv(&self) -> T {
+        // interrupt the receiver to prevent consuming everything at once
+        executor::yield_me(0);
+
+        let mut queue = self.queue.lock();
+        if let Some(t) = queue.pop_front() {
+            return t;
+        }
+        loop {
+            self.waker.wake_me_later();
+            if let Some(t) = queue.pop_front() {
+                return t;
+            }
+            MutexGuard::unlocked(&mut queue, || {
+                executor::yield_me(-1);
+            });
+        }
+    }
+}
+
+impl<T> PollSome for Chan<T> {
+    /// Schedules a wakeup for the current thread.
+    fn wake_me(&self) {
+        self.shared.waker.wake_me_later();
+    }
+
+    /// Checks if chan has any pending messages.
+    fn has_some(&self) -> bool {
+        !self.shared.queue.lock().is_empty()
+    }
+}

libs/desim/src/executor.rs

@@ -0,0 +1,483 @@
+use std::{
+    panic::AssertUnwindSafe,
+    sync::{
+        atomic::{AtomicBool, AtomicU32, AtomicU8, Ordering},
+        mpsc, Arc, OnceLock,
+    },
+    thread::JoinHandle,
+};
+
+use tracing::{debug, error, trace};
+
+use crate::time::Timing;
+
+/// Stores status of the running threads. Threads are registered in the runtime upon creation
+/// and deregistered upon termination.
+pub struct Runtime {
+    // stores handles to all threads that are currently running
+    threads: Vec<ThreadHandle>,
+    // stores current time and pending wakeups
+    clock: Arc<Timing>,
+    // thread counter
+    thread_counter: AtomicU32,
+    // Thread step counter -- how many times all threads has been actually
+    // stepped (note that all world/time/executor/thread have slightly different
+    // meaning of steps). For observability.
+    pub step_counter: u64,
+}
+
+impl Runtime {
+    /// Init new runtime, no running threads.
+    pub fn new(clock: Arc<Timing>) -> Self {
+        Self {
+            threads: Vec::new(),
+            clock,
+            thread_counter: AtomicU32::new(0),
+            step_counter: 0,
+        }
+    }
+
+    /// Spawn a new thread and register it in the runtime.
+    pub fn spawn<F>(&mut self, f: F) -> ExternalHandle
+    where
+        F: FnOnce() + Send + 'static,
+    {
+        let (tx, rx) = mpsc::channel();
+
+        let clock = self.clock.clone();
+        let tid = self.thread_counter.fetch_add(1, Ordering::SeqCst);
+        debug!("spawning thread-{}", tid);
+
+        let join = std::thread::spawn(move || {
+            let _guard = tracing::info_span!("", tid).entered();
+
+            let res = std::panic::catch_unwind(AssertUnwindSafe(|| {
+                with_thread_context(|ctx| {
+                    assert!(ctx.clock.set(clock).is_ok());
+                    ctx.id.store(tid, Ordering::SeqCst);
+                    tx.send(ctx.clone()).expect("failed to send thread context");
+                    // suspend thread to put it to `threads` in sleeping state
+                    ctx.yield_me(0);
+                });
+
+                // start user-provided function
+                f();
+            }));
+            debug!("thread finished");
+
+            if let Err(e) = res {
+                with_thread_context(|ctx| {
+                    if !ctx.allow_panic.load(std::sync::atomic::Ordering::SeqCst) {
+                        error!("thread panicked, terminating the process: {:?}", e);
+                        std::process::exit(1);
+                    }
+
+                    debug!("thread panicked: {:?}", e);
+                    let mut result = ctx.result.lock();
+                    if result.0 == -1 {
+                        *result = (256, format!("thread panicked: {:?}", e));
+                    }
+                });
+            }
+
+            with_thread_context(|ctx| {
+                ctx.finish_me();
+            });
+        });
+
+        let ctx = rx.recv().expect("failed to receive thread context");
+        let handle = ThreadHandle::new(ctx.clone(), join);
+
+        self.threads.push(handle);
+
+        ExternalHandle { ctx }
+    }
+
+    /// Returns true if there are any unfinished activity, such as running thread or pending events.
+    /// Otherwise returns false, which means all threads are blocked forever.
+    pub fn step(&mut self) -> bool {
+        trace!("runtime step");
+
+        // have we run any thread?
+        let mut ran = false;
+
+        self.threads.retain(|thread: &ThreadHandle| {
+            let res = thread.ctx.wakeup.compare_exchange(
+                PENDING_WAKEUP,
+                NO_WAKEUP,
+                Ordering::SeqCst,
+                Ordering::SeqCst,
+            );
+            if res.is_err() {
+                // thread has no pending wakeups, leaving as is
+                return true;
+            }
+            ran = true;
+
+            trace!("entering thread-{}", thread.ctx.tid());
+            let status = thread.step();
+            self.step_counter += 1;
+            trace!(
+                "out of thread-{} with status {:?}",
+                thread.ctx.tid(),
+                status
+            );
+
+            if status == Status::Sleep {
+                true
+            } else {
+                trace!("thread has finished");
+                // removing the thread from the list
+                false
+            }
+        });
+
+        if !ran {
+            trace!("no threads were run, stepping clock");
+            if let Some(ctx_to_wake) = self.clock.step() {
+                trace!("waking up thread-{}", ctx_to_wake.tid());
+                ctx_to_wake.inc_wake();
+            } else {
+                return false;
+            }
+        }
+
+        true
+    }
+
+    /// Kill all threads. This is done by setting a flag in each thread context and waking it up.
+    pub fn crash_all_threads(&mut self) {
+        for thread in self.threads.iter() {
+            thread.ctx.crash_stop();
+        }
+
+        // all threads should be finished after a few steps
+        while !self.threads.is_empty() {
+            self.step();
+        }
+    }
+}
+
+impl Drop for Runtime {
+    fn drop(&mut self) {
+        debug!("dropping the runtime");
+        self.crash_all_threads();
+    }
+}
+
+#[derive(Clone)]
+pub struct ExternalHandle {
+    ctx: Arc<ThreadContext>,
+}
+
+impl ExternalHandle {
+    /// Returns true if thread has finished execution.
+    pub fn is_finished(&self) -> bool {
+        let status = self.ctx.mutex.lock();
+        *status == Status::Finished
+    }
+
+    /// Returns exitcode and message, which is available after thread has finished execution.
+    pub fn result(&self) -> (i32, String) {
+        let result = self.ctx.result.lock();
+        result.clone()
+    }
+
+    /// Returns thread id.
+    pub fn id(&self) -> u32 {
+        self.ctx.id.load(Ordering::SeqCst)
+    }
+
+    /// Sets a flag to crash thread on the next wakeup.
+    pub fn crash_stop(&self) {
+        self.ctx.crash_stop();
+    }
+}
+
+struct ThreadHandle {
+    ctx: Arc<ThreadContext>,
+    _join: JoinHandle<()>,
+}
+
+impl ThreadHandle {
+    /// Create a new [`ThreadHandle`] and wait until thread will enter [`Status::Sleep`] state.
+    fn new(ctx: Arc<ThreadContext>, join: JoinHandle<()>) -> Self {
+        let mut status = ctx.mutex.lock();
+        // wait until thread will go into the first yield
+        while *status != Status::Sleep {
+            ctx.condvar.wait(&mut status);
+        }
+        drop(status);
+
+        Self { ctx, _join: join }
+    }
+
+    /// Allows thread to execute one step of its execution.
+    /// Returns [`Status`] of the thread after the step.
+    fn step(&self) -> Status {
+        let mut status = self.ctx.mutex.lock();
+        assert!(matches!(*status, Status::Sleep));
+
+        *status = Status::Running;
+        self.ctx.condvar.notify_all();
+
+        while *status == Status::Running {
+            self.ctx.condvar.wait(&mut status);
+        }
+
+        *status
+    }
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum Status {
+    /// Thread is running.
+    Running,
+    /// Waiting for event to complete, will be resumed by the executor step, once wakeup flag is set.
+    Sleep,
+    /// Thread finished execution.
+    Finished,
+}
+
+const NO_WAKEUP: u8 = 0;
+const PENDING_WAKEUP: u8 = 1;
+
+pub struct ThreadContext {
+    id: AtomicU32,
+    // used to block thread until it is woken up
+    mutex: parking_lot::Mutex<Status>,
+    condvar: parking_lot::Condvar,
+    // used as a flag to indicate runtime that thread is ready to be woken up
+    wakeup: AtomicU8,
+    clock: OnceLock<Arc<Timing>>,
+    // execution result, set by exit() call
+    result: parking_lot::Mutex<(i32, String)>,
+    // determines if process should be killed on receiving panic
+    allow_panic: AtomicBool,
+    // acts as a signal that thread should crash itself on the next wakeup
+    crash_request: AtomicBool,
+}
+
+impl ThreadContext {
+    pub(crate) fn new() -> Self {
+        Self {
+            id: AtomicU32::new(0),
+            mutex: parking_lot::Mutex::new(Status::Running),
+            condvar: parking_lot::Condvar::new(),
+            wakeup: AtomicU8::new(NO_WAKEUP),
+            clock: OnceLock::new(),
+            result: parking_lot::Mutex::new((-1, String::new())),
+            allow_panic: AtomicBool::new(false),
+            crash_request: AtomicBool::new(false),
+        }
+    }
+}
+
+// Functions for executor to control thread execution.
+impl ThreadContext {
+    /// Set atomic flag to indicate that thread is ready to be woken up.
+    fn inc_wake(&self) {
+        self.wakeup.store(PENDING_WAKEUP, Ordering::SeqCst);
+    }
+
+    /// Internal function used for event queues.
+    pub(crate) fn schedule_wakeup(self: &Arc<Self>, after_ms: u64) {
+        self.clock
+            .get()
+            .unwrap()
+            .schedule_wakeup(after_ms, self.clone());
+    }
+
+    fn tid(&self) -> u32 {
+        self.id.load(Ordering::SeqCst)
+    }
+
+    fn crash_stop(&self) {
+        let status = self.mutex.lock();
+        if *status == Status::Finished {
+            debug!(
+                "trying to crash thread-{}, which is already finished",
+                self.tid()
+            );
+            return;
+        }
+        assert!(matches!(*status, Status::Sleep));
+        drop(status);
+
+        self.allow_panic.store(true, Ordering::SeqCst);
+        self.crash_request.store(true, Ordering::SeqCst);
+        // set a wakeup
+        self.inc_wake();
+        // it will panic on the next wakeup
+    }
+}
+
+// Internal functions.
+impl ThreadContext {
+    /// Blocks thread until it's woken up by the executor. If `after_ms` is 0, is will be
+    /// woken on the next step. If `after_ms` > 0, wakeup is scheduled after that time.
+    /// Otherwise wakeup is not scheduled inside `yield_me`, and should be arranged before
+    /// calling this function.
+    fn yield_me(self: &Arc<Self>, after_ms: i64) {
+        let mut status = self.mutex.lock();
+        assert!(matches!(*status, Status::Running));
+
+        match after_ms.cmp(&0) {
+            std::cmp::Ordering::Less => {
+                // block until something wakes us up
+            }
+            std::cmp::Ordering::Equal => {
+                // tell executor that we are ready to be woken up
+                self.inc_wake();
+            }
+            std::cmp::Ordering::Greater => {
+                // schedule wakeup
+                self.clock
+                    .get()
+                    .unwrap()
+                    .schedule_wakeup(after_ms as u64, self.clone());
+            }
+        }
+
+        *status = Status::Sleep;
+        self.condvar.notify_all();
+
+        // wait until executor wakes us up
+        while *status != Status::Running {
+            self.condvar.wait(&mut status);
+        }
+
+        if self.crash_request.load(Ordering::SeqCst) {
+            panic!("crashed by request");
+        }
+    }
+
+    /// Called only once, exactly before thread finishes execution.
+    fn finish_me(&self) {
+        let mut status = self.mutex.lock();
+        assert!(matches!(*status, Status::Running));
+
+        *status = Status::Finished;
+        {
+            let mut result = self.result.lock();
+            if result.0 == -1 {
+                *result = (0, "finished normally".to_owned());
+            }
+        }
+        self.condvar.notify_all();
+    }
+}
+
+/// Invokes the given closure with a reference to the current thread [`ThreadContext`].
+#[inline(always)]
+fn with_thread_context<T>(f: impl FnOnce(&Arc<ThreadContext>) -> T) -> T {
+    thread_local!(static THREAD_DATA: Arc<ThreadContext> = Arc::new(ThreadContext::new()));
+    THREAD_DATA.with(f)
+}
+
+/// Waker is used to wake up threads that are blocked on condition.
+/// It keeps track of contexts [`Arc<ThreadContext>`] and can increment the counter
+/// of several contexts to send a notification.
+pub struct Waker {
+    // contexts that are waiting for a notification
+    contexts: parking_lot::Mutex<smallvec::SmallVec<[Arc<ThreadContext>; 8]>>,
+}
+
+impl Default for Waker {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl Waker {
+    pub fn new() -> Self {
+        Self {
+            contexts: parking_lot::Mutex::new(smallvec::SmallVec::new()),
+        }
+    }
+
+    /// Subscribe current thread to receive a wake notification later.
+    pub fn wake_me_later(&self) {
+        with_thread_context(|ctx| {
+            self.contexts.lock().push(ctx.clone());
+        });
+    }
+
+    /// Wake up all threads that are waiting for a notification and clear the list.
+    pub fn wake_all(&self) {
+        let mut v = self.contexts.lock();
+        for ctx in v.iter() {
+            ctx.inc_wake();
+        }
+        v.clear();
+    }
+}
+
+/// See [`ThreadContext::yield_me`].
+pub fn yield_me(after_ms: i64) {
+    with_thread_context(|ctx| ctx.yield_me(after_ms))
+}
+
+/// Get current time.
+pub fn now() -> u64 {
+    with_thread_context(|ctx| ctx.clock.get().unwrap().now())
+}
+
+pub fn exit(code: i32, msg: String) {
+    with_thread_context(|ctx| {
+        ctx.allow_panic.store(true, Ordering::SeqCst);
+        let mut result = ctx.result.lock();
+        *result = (code, msg);
+        panic!("exit");
+    });
+}
+
+pub(crate) fn get_thread_ctx() -> Arc<ThreadContext> {
+    with_thread_context(|ctx| ctx.clone())
+}
+
+/// Trait for polling channels until they have something.
+pub trait PollSome {
+    /// Schedule wakeup for message arrival.
+    fn wake_me(&self);
+
+    /// Check if channel has a ready message.
+    fn has_some(&self) -> bool;
+}
+
+/// Blocks current thread until one of the channels has a ready message. Returns
+/// index of the channel that has a message. If timeout is reached, returns None.
+///
+/// Negative timeout means block forever. Zero timeout means check channels and return
+/// immediately. Positive timeout means block until timeout is reached.
+pub fn epoll_chans(chans: &[Box<dyn PollSome>], timeout: i64) -> Option<usize> {
+    let deadline = if timeout < 0 {
+        0
+    } else {
+        now() + timeout as u64
+    };
+
+    loop {
+        for chan in chans {
+            chan.wake_me()
+        }
+
+        for (i, chan) in chans.iter().enumerate() {
+            if chan.has_some() {
+                return Some(i);
+            }
+        }
+
+        if timeout < 0 {
+            // block until wakeup
+            yield_me(-1);
+        } else {
+            let current_time = now();
+            if current_time >= deadline {
+                return None;
+            }
+
+            yield_me((deadline - current_time) as i64);
+        }
+    }
+}

libs/desim/src/lib.rs

@@ -0,0 +1,8 @@
+pub mod chan;
+pub mod executor;
+pub mod network;
+pub mod node_os;
+pub mod options;
+pub mod proto;
+pub mod time;
+pub mod world;

libs/desim/src/network.rs

@@ -0,0 +1,451 @@
+use std::{
+    cmp::Ordering,
+    collections::{BinaryHeap, VecDeque},
+    fmt::{self, Debug},
+    ops::DerefMut,
+    sync::{mpsc, Arc},
+};
+
+use parking_lot::{
+    lock_api::{MappedMutexGuard, MutexGuard},
+    Mutex, RawMutex,
+};
+use rand::rngs::StdRng;
+use tracing::debug;
+
+use crate::{
+    executor::{self, ThreadContext},
+    options::NetworkOptions,
+    proto::NetEvent,
+    proto::NodeEvent,
+};
+
+use super::{chan::Chan, proto::AnyMessage};
+
+pub struct NetworkTask {
+    options: Arc<NetworkOptions>,
+    connections: Mutex<Vec<VirtualConnection>>,
+    /// min-heap of connections having something to deliver.
+    events: Mutex<BinaryHeap<Event>>,
+    task_context: Arc<ThreadContext>,
+}
+
+impl NetworkTask {
+    pub fn start_new(options: Arc<NetworkOptions>, tx: mpsc::Sender<Arc<NetworkTask>>) {
+        let ctx = executor::get_thread_ctx();
+        let task = Arc::new(Self {
+            options,
+            connections: Mutex::new(Vec::new()),
+            events: Mutex::new(BinaryHeap::new()),
+            task_context: ctx,
+        });
+
+        // send the task upstream
+        tx.send(task.clone()).unwrap();
+
+        // start the task
+        task.start();
+    }
+
+    pub fn start_new_connection(self: &Arc<Self>, rng: StdRng, dst_accept: Chan<NodeEvent>) -> TCP {
+        let now = executor::now();
+        let connection_id = self.connections.lock().len();
+
+        let vc = VirtualConnection {
+            connection_id,
+            dst_accept,
+            dst_sockets: [Chan::new(), Chan::new()],
+            state: Mutex::new(ConnectionState {
+                buffers: [NetworkBuffer::new(None), NetworkBuffer::new(Some(now))],
+                rng,
+            }),
+        };
+        vc.schedule_timeout(self);
+        vc.send_connect(self);
+
+        let recv_chan = vc.dst_sockets[0].clone();
+        self.connections.lock().push(vc);
+
+        TCP {
+            net: self.clone(),
+            conn_id: connection_id,
+            dir: 0,
+            recv_chan,
+        }
+    }
+}
+
+// private functions
+impl NetworkTask {
+    /// Schedule to wakeup network task (self) `after_ms` later to deliver
+    /// messages of connection `id`.
+    fn schedule(&self, id: usize, after_ms: u64) {
+        self.events.lock().push(Event {
+            time: executor::now() + after_ms,
+            conn_id: id,
+        });
+        self.task_context.schedule_wakeup(after_ms);
+    }
+
+    /// Get locked connection `id`.
+    fn get(&self, id: usize) -> MappedMutexGuard<'_, RawMutex, VirtualConnection> {
+        MutexGuard::map(self.connections.lock(), |connections| {
+            connections.get_mut(id).unwrap()
+        })
+    }
+
+    fn collect_pending_events(&self, now: u64, vec: &mut Vec<Event>) {
+        vec.clear();
+        let mut events = self.events.lock();
+        while let Some(event) = events.peek() {
+            if event.time > now {
+                break;
+            }
+            let event = events.pop().unwrap();
+            vec.push(event);
+        }
+    }
+
+    fn start(self: &Arc<Self>) {
+        debug!("started network task");
+
+        let mut events = Vec::new();
+        loop {
+            let now = executor::now();
+            self.collect_pending_events(now, &mut events);
+
+            for event in events.drain(..) {
+                let conn = self.get(event.conn_id);
+                conn.process(self);
+            }
+
+            // block until wakeup
+            executor::yield_me(-1);
+        }
+    }
+}
+
+// 0 - from node(0) to node(1)
+// 1 - from node(1) to node(0)
+type MessageDirection = u8;
+
+fn sender_str(dir: MessageDirection) -> &'static str {
+    match dir {
+        0 => "client",
+        1 => "server",
+        _ => unreachable!(),
+    }
+}
+
+fn receiver_str(dir: MessageDirection) -> &'static str {
+    match dir {
+        0 => "server",
+        1 => "client",
+        _ => unreachable!(),
+    }
+}
+
+/// Virtual connection between two nodes.
+/// Node 0 is the creator of the connection (client),
+/// and node 1 is the acceptor (server).
+struct VirtualConnection {
+    connection_id: usize,
+    /// one-off chan, used to deliver Accept message to dst
+    dst_accept: Chan<NodeEvent>,
+    /// message sinks
+    dst_sockets: [Chan<NetEvent>; 2],
+    state: Mutex<ConnectionState>,
+}
+
+struct ConnectionState {
+    buffers: [NetworkBuffer; 2],
+    rng: StdRng,
+}
+
+impl VirtualConnection {
+    /// Notify the future about the possible timeout.
+    fn schedule_timeout(&self, net: &NetworkTask) {
+        if let Some(timeout) = net.options.keepalive_timeout {
+            net.schedule(self.connection_id, timeout);
+        }
+    }
+
+    /// Send the handshake (Accept) to the server.
+    fn send_connect(&self, net: &NetworkTask) {
+        let now = executor::now();
+        let mut state = self.state.lock();
+        let delay = net.options.connect_delay.delay(&mut state.rng);
+        let buffer = &mut state.buffers[0];
+        assert!(buffer.buf.is_empty());
+        assert!(!buffer.recv_closed);
+        assert!(!buffer.send_closed);
+        assert!(buffer.last_recv.is_none());
+
+        let delay = if let Some(ms) = delay {
+            ms
+        } else {
+            debug!("NET: TCP #{} dropped connect", self.connection_id);
+            buffer.send_closed = true;
+            return;
+        };
+
+        // Send a message into the future.
+        buffer
+            .buf
+            .push_back((now + delay, AnyMessage::InternalConnect));
+        net.schedule(self.connection_id, delay);
+    }
+
+    /// Transmit some of the messages from the buffer to the nodes.
+    fn process(&self, net: &Arc<NetworkTask>) {
+        let now = executor::now();
+
+        let mut state = self.state.lock();
+
+        for direction in 0..2 {
+            self.process_direction(
+                net,
+                state.deref_mut(),
+                now,
+                direction as MessageDirection,
+                &self.dst_sockets[direction ^ 1],
+            );
+        }
+
+        // Close the one side of the connection by timeout if the node
+        // has not received any messages for a long time.
+        if let Some(timeout) = net.options.keepalive_timeout {
+            let mut to_close = [false, false];
+            for direction in 0..2 {
+                let buffer = &mut state.buffers[direction];
+                if buffer.recv_closed {
+                    continue;
+                }
+                if let Some(last_recv) = buffer.last_recv {
+                    if now - last_recv >= timeout {
+                        debug!(
+                            "NET: connection {} timed out at {}",
+                            self.connection_id,
+                            receiver_str(direction as MessageDirection)
+                        );
+                        let node_idx = direction ^ 1;
+                        to_close[node_idx] = true;
+                    }
+                }
+            }
+            drop(state);
+
+            for (node_idx, should_close) in to_close.iter().enumerate() {
+                if *should_close {
+                    self.close(node_idx);
+                }
+            }
+        }
+    }
+
+    /// Process messages in the buffer in the given direction.
+    fn process_direction(
+        &self,
+        net: &Arc<NetworkTask>,
+        state: &mut ConnectionState,
+        now: u64,
+        direction: MessageDirection,
+        to_socket: &Chan<NetEvent>,
+    ) {
+        let buffer = &mut state.buffers[direction as usize];
+        if buffer.recv_closed {
+            assert!(buffer.buf.is_empty());
+        }
+
+        while !buffer.buf.is_empty() && buffer.buf.front().unwrap().0 <= now {
+            let msg = buffer.buf.pop_front().unwrap().1;
+
+            buffer.last_recv = Some(now);
+            self.schedule_timeout(net);
+
+            if let AnyMessage::InternalConnect = msg {
+                // TODO: assert to_socket is the server
+                let server_to_client = TCP {
+                    net: net.clone(),
+                    conn_id: self.connection_id,
+                    dir: direction ^ 1,
+                    recv_chan: to_socket.clone(),
+                };
+                // special case, we need to deliver new connection to a separate channel
+                self.dst_accept.send(NodeEvent::Accept(server_to_client));
+            } else {
+                to_socket.send(NetEvent::Message(msg));
+            }
+        }
+    }
+
+    /// Try to send a message to the buffer, optionally dropping it and
+    /// determining delivery timestamp.
+    fn send(&self, net: &NetworkTask, direction: MessageDirection, msg: AnyMessage) {
+        let now = executor::now();
+        let mut state = self.state.lock();
+
+        let (delay, close) = if let Some(ms) = net.options.send_delay.delay(&mut state.rng) {
+            (ms, false)
+        } else {
+            (0, true)
+        };
+
+        let buffer = &mut state.buffers[direction as usize];
+        if buffer.send_closed {
+            debug!(
+                "NET: TCP #{} dropped message {:?} (broken pipe)",
+                self.connection_id, msg
+            );
+            return;
+        }
+
+        if close {
+            debug!(
+                "NET: TCP #{} dropped message {:?} (pipe just broke)",
+                self.connection_id, msg
+            );
+            buffer.send_closed = true;
+            return;
+        }
+
+        if buffer.recv_closed {
+            debug!(
+                "NET: TCP #{} dropped message {:?} (recv closed)",
+                self.connection_id, msg
+            );
+            return;
+        }
+
+        // Send a message into the future.
+        buffer.buf.push_back((now + delay, msg));
+        net.schedule(self.connection_id, delay);
+    }
+
+    /// Close the connection. Only one side of the connection will be closed,
+    /// and no further messages will be delivered. The other side will not be notified.
+    fn close(&self, node_idx: usize) {
+        let mut state = self.state.lock();
+        let recv_buffer = &mut state.buffers[1 ^ node_idx];
+        if recv_buffer.recv_closed {
+            debug!(
+                "NET: TCP #{} closed twice at {}",
+                self.connection_id,
+                sender_str(node_idx as MessageDirection),
+            );
+            return;
+        }
+
+        debug!(
+            "NET: TCP #{} closed at {}",
+            self.connection_id,
+            sender_str(node_idx as MessageDirection),
+        );
+        recv_buffer.recv_closed = true;
+        for msg in recv_buffer.buf.drain(..) {
+            debug!(
+                "NET: TCP #{} dropped message {:?} (closed)",
+                self.connection_id, msg
+            );
+        }
+
+        let send_buffer = &mut state.buffers[node_idx];
+        send_buffer.send_closed = true;
+        drop(state);
+
+        // TODO: notify the other side?
+
+        self.dst_sockets[node_idx].send(NetEvent::Closed);
+    }
+}
+
+struct NetworkBuffer {
+    /// Messages paired with time of delivery
+    buf: VecDeque<(u64, AnyMessage)>,
+    /// True if the connection is closed on the receiving side,
+    /// i.e. no more messages from the buffer will be delivered.
+    recv_closed: bool,
+    /// True if the connection is closed on the sending side,
+    /// i.e. no more messages will be added to the buffer.
+    send_closed: bool,
+    /// Last time a message was delivered from the buffer.
+    /// If None, it means that the server is the receiver and
+    /// it has not yet aware of this connection (i.e. has not
+    /// received the Accept).
+    last_recv: Option<u64>,
+}
+
+impl NetworkBuffer {
+    fn new(last_recv: Option<u64>) -> Self {
+        Self {
+            buf: VecDeque::new(),
+            recv_closed: false,
+            send_closed: false,
+            last_recv,
+        }
+    }
+}
+
+/// Single end of a bidirectional network stream without reordering (TCP-like).
+/// Reads are implemented using channels, writes go to the buffer inside VirtualConnection.
+pub struct TCP {
+    net: Arc<NetworkTask>,
+    conn_id: usize,
+    dir: MessageDirection,
+    recv_chan: Chan<NetEvent>,
+}
+
+impl Debug for TCP {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "TCP #{} ({})", self.conn_id, sender_str(self.dir),)
+    }
+}
+
+impl TCP {
+    /// Send a message to the other side. It's guaranteed that it will not arrive
+    /// before the arrival of all messages sent earlier.
+    pub fn send(&self, msg: AnyMessage) {
+        let conn = self.net.get(self.conn_id);
+        conn.send(&self.net, self.dir, msg);
+    }
+
+    /// Get a channel to receive incoming messages.
+    pub fn recv_chan(&self) -> Chan<NetEvent> {
+        self.recv_chan.clone()
+    }
+
+    pub fn connection_id(&self) -> usize {
+        self.conn_id
+    }
+
+    pub fn close(&self) {
+        let conn = self.net.get(self.conn_id);
+        conn.close(self.dir as usize);
+    }
+}
+struct Event {
+    time: u64,
+    conn_id: usize,
+}
+
+// BinaryHeap is a max-heap, and we want a min-heap. Reverse the ordering here
+// to get that.
+impl PartialOrd for Event {
+    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl Ord for Event {
+    fn cmp(&self, other: &Self) -> Ordering {
+        (other.time, other.conn_id).cmp(&(self.time, self.conn_id))
+    }
+}
+
+impl PartialEq for Event {
+    fn eq(&self, other: &Self) -> bool {
+        (other.time, other.conn_id) == (self.time, self.conn_id)
+    }
+}
+
+impl Eq for Event {}

libs/desim/src/node_os.rs

@@ -0,0 +1,54 @@
+use std::sync::Arc;
+
+use rand::Rng;
+
+use crate::proto::NodeEvent;
+
+use super::{
+    chan::Chan,
+    network::TCP,
+    world::{Node, NodeId, World},
+};
+
+/// Abstraction with all functions (aka syscalls) available to the node.
+#[derive(Clone)]
+pub struct NodeOs {
+    world: Arc<World>,
+    internal: Arc<Node>,
+}
+
+impl NodeOs {
+    pub fn new(world: Arc<World>, internal: Arc<Node>) -> NodeOs {
+        NodeOs { world, internal }
+    }
+
+    /// Get the node id.
+    pub fn id(&self) -> NodeId {
+        self.internal.id
+    }
+
+    /// Opens a bidirectional connection with the other node. Always successful.
+    pub fn open_tcp(&self, dst: NodeId) -> TCP {
+        self.world.open_tcp(dst)
+    }
+
+    /// Returns a channel to receive node events (socket Accept and internal messages).
+    pub fn node_events(&self) -> Chan<NodeEvent> {
+        self.internal.node_events()
+    }
+
+    /// Get current time.
+    pub fn now(&self) -> u64 {
+        self.world.now()
+    }
+
+    /// Generate a random number in range [0, max).
+    pub fn random(&self, max: u64) -> u64 {
+        self.internal.rng.lock().gen_range(0..max)
+    }
+
+    /// Append a new event to the world event log.
+    pub fn log_event(&self, data: String) {
+        self.internal.log_event(data)
+    }
+}

libs/desim/src/options.rs

@@ -0,0 +1,50 @@
+use rand::{rngs::StdRng, Rng};
+
+/// Describes random delays and failures. Delay will be uniformly distributed in [min, max].
+/// Connection failure will occur with the probablity fail_prob.
+#[derive(Clone, Debug)]
+pub struct Delay {
+    pub min: u64,
+    pub max: u64,
+    pub fail_prob: f64, // [0; 1]
+}
+
+impl Delay {
+    /// Create a struct with no delay, no failures.
+    pub fn empty() -> Delay {
+        Delay {
+            min: 0,
+            max: 0,
+            fail_prob: 0.0,
+        }
+    }
+
+    /// Create a struct with a fixed delay.
+    pub fn fixed(ms: u64) -> Delay {
+        Delay {
+            min: ms,
+            max: ms,
+            fail_prob: 0.0,
+        }
+    }
+
+    /// Generate a random delay in range [min, max]. Return None if the
+    /// message should be dropped.
+    pub fn delay(&self, rng: &mut StdRng) -> Option<u64> {
+        if rng.gen_bool(self.fail_prob) {
+            return None;
+        }
+        Some(rng.gen_range(self.min..=self.max))
+    }
+}
+
+/// Describes network settings. All network packets will be subjected to the same delays and failures.
+#[derive(Clone, Debug)]
+pub struct NetworkOptions {
+    /// Connection will be automatically closed after this timeout if no data is received.
+    pub keepalive_timeout: Option<u64>,
+    /// New connections will be delayed by this amount of time.
+    pub connect_delay: Delay,
+    /// Each message will be delayed by this amount of time.
+    pub send_delay: Delay,
+}

libs/desim/src/proto.rs

@@ -0,0 +1,63 @@
+use std::fmt::Debug;
+
+use bytes::Bytes;
+use utils::lsn::Lsn;
+
+use crate::{network::TCP, world::NodeId};
+
+/// Internal node events.
+#[derive(Debug)]
+pub enum NodeEvent {
+    Accept(TCP),
+    Internal(AnyMessage),
+}
+
+/// Events that are coming from a network socket.
+#[derive(Clone, Debug)]
+pub enum NetEvent {
+    Message(AnyMessage),
+    Closed,
+}
+
+/// Custom events generated throughout the simulation. Can be used by the test to verify the correctness.
+#[derive(Debug)]
+pub struct SimEvent {
+    pub time: u64,
+    pub node: NodeId,
+    pub data: String,
+}
+
+/// Umbrella type for all possible flavours of messages. These events can be sent over network
+/// or to an internal node events channel.
+#[derive(Clone)]
+pub enum AnyMessage {
+    /// Not used, empty placeholder.
+    None,
+    /// Used internally for notifying node about new incoming connection.
+    InternalConnect,
+    Just32(u32),
+    ReplCell(ReplCell),
+    Bytes(Bytes),
+    LSN(u64),
+}
+
+impl Debug for AnyMessage {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            AnyMessage::None => write!(f, "None"),
+            AnyMessage::InternalConnect => write!(f, "InternalConnect"),
+            AnyMessage::Just32(v) => write!(f, "Just32({})", v),
+            AnyMessage::ReplCell(v) => write!(f, "ReplCell({:?})", v),
+            AnyMessage::Bytes(v) => write!(f, "Bytes({})", hex::encode(v)),
+            AnyMessage::LSN(v) => write!(f, "LSN({})", Lsn(*v)),
+        }
+    }
+}
+
+/// Used in reliable_copy_test.rs
+#[derive(Clone, Debug)]
+pub struct ReplCell {
+    pub value: u32,
+    pub client_id: u32,
+    pub seqno: u32,
+}

libs/desim/src/time.rs

@@ -0,0 +1,129 @@
+use std::{
+    cmp::Ordering,
+    collections::BinaryHeap,
+    ops::DerefMut,
+    sync::{
+        atomic::{AtomicU32, AtomicU64},
+        Arc,
+    },
+};
+
+use parking_lot::Mutex;
+use tracing::trace;
+
+use crate::executor::ThreadContext;
+
+/// Holds current time and all pending wakeup events.
+pub struct Timing {
+    /// Current world's time.
+    current_time: AtomicU64,
+    /// Pending timers.
+    queue: Mutex<BinaryHeap<Pending>>,
+    /// Global nonce. Makes picking events from binary heap queue deterministic
+    /// by appending a number to events with the same timestamp.
+    nonce: AtomicU32,
+    /// Used to schedule fake events.
+    fake_context: Arc<ThreadContext>,
+}
+
+impl Default for Timing {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl Timing {
+    /// Create a new empty clock with time set to 0.
+    pub fn new() -> Timing {
+        Timing {
+            current_time: AtomicU64::new(0),
+            queue: Mutex::new(BinaryHeap::new()),
+            nonce: AtomicU32::new(0),
+            fake_context: Arc::new(ThreadContext::new()),
+        }
+    }
+
+    /// Return the current world's time.
+    pub fn now(&self) -> u64 {
+        self.current_time.load(std::sync::atomic::Ordering::SeqCst)
+    }
+
+    /// Tick-tock the global clock. Return the event ready to be processed
+    /// or move the clock forward and then return the event.
+    pub(crate) fn step(&self) -> Option<Arc<ThreadContext>> {
+        let mut queue = self.queue.lock();
+
+        if queue.is_empty() {
+            // no future events
+            return None;
+        }
+
+        if !self.is_event_ready(queue.deref_mut()) {
+            let next_time = queue.peek().unwrap().time;
+            self.current_time
+                .store(next_time, std::sync::atomic::Ordering::SeqCst);
+            trace!("rewind time to {}", next_time);
+            assert!(self.is_event_ready(queue.deref_mut()));
+        }
+
+        Some(queue.pop().unwrap().wake_context)
+    }
+
+    /// Append an event to the queue, to wakeup the thread in `ms` milliseconds.
+    pub(crate) fn schedule_wakeup(&self, ms: u64, wake_context: Arc<ThreadContext>) {
+        self.nonce.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+        let nonce = self.nonce.load(std::sync::atomic::Ordering::SeqCst);
+        self.queue.lock().push(Pending {
+            time: self.now() + ms,
+            nonce,
+            wake_context,
+        })
+    }
+
+    /// Append a fake event to the queue, to prevent clocks from skipping this time.
+    pub fn schedule_fake(&self, ms: u64) {
+        self.queue.lock().push(Pending {
+            time: self.now() + ms,
+            nonce: 0,
+            wake_context: self.fake_context.clone(),
+        });
+    }
+
+    /// Return true if there is a ready event.
+    fn is_event_ready(&self, queue: &mut BinaryHeap<Pending>) -> bool {
+        queue.peek().map_or(false, |x| x.time <= self.now())
+    }
+
+    /// Clear all pending events.
+    pub(crate) fn clear(&self) {
+        self.queue.lock().clear();
+    }
+}
+
+struct Pending {
+    time: u64,
+    nonce: u32,
+    wake_context: Arc<ThreadContext>,
+}
+
+// BinaryHeap is a max-heap, and we want a min-heap. Reverse the ordering here
+// to get that.
+impl PartialOrd for Pending {
+    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
+        Some(self.cmp(other))
+    }
+}
+
+impl Ord for Pending {
+    fn cmp(&self, other: &Self) -> Ordering {
+        (other.time, other.nonce).cmp(&(self.time, self.nonce))
+    }
+}
+
+impl PartialEq for Pending {
+    fn eq(&self, other: &Self) -> bool {
+        (other.time, other.nonce) == (self.time, self.nonce)
+    }
+}
+
+impl Eq for Pending {}

libs/desim/src/world.rs

@@ -0,0 +1,180 @@
+use parking_lot::Mutex;
+use rand::{rngs::StdRng, SeedableRng};
+use std::{
+    ops::DerefMut,
+    sync::{mpsc, Arc},
+};
+
+use crate::{
+    executor::{ExternalHandle, Runtime},
+    network::NetworkTask,
+    options::NetworkOptions,
+    proto::{NodeEvent, SimEvent},
+    time::Timing,
+};
+
+use super::{chan::Chan, network::TCP, node_os::NodeOs};
+
+pub type NodeId = u32;
+
+/// World contains simulation state.
+pub struct World {
+    nodes: Mutex<Vec<Arc<Node>>>,
+    /// Random number generator.
+    rng: Mutex<StdRng>,
+    /// Internal event log.
+    events: Mutex<Vec<SimEvent>>,
+    /// Separate task that processes all network messages.
+    network_task: Arc<NetworkTask>,
+    /// Runtime for running threads and moving time.
+    runtime: Mutex<Runtime>,
+    /// To get current time.
+    timing: Arc<Timing>,
+}
+
+impl World {
+    pub fn new(seed: u64, options: Arc<NetworkOptions>) -> World {
+        let timing = Arc::new(Timing::new());
+        let mut runtime = Runtime::new(timing.clone());
+
+        let (tx, rx) = mpsc::channel();
+
+        runtime.spawn(move || {
+            // create and start network background thread, and send it back via the channel
+            NetworkTask::start_new(options, tx)
+        });
+
+        // wait for the network task to start
+        while runtime.step() {}
+
+        let network_task = rx.recv().unwrap();
+
+        World {
+            nodes: Mutex::new(Vec::new()),
+            rng: Mutex::new(StdRng::seed_from_u64(seed)),
+            events: Mutex::new(Vec::new()),
+            network_task,
+            runtime: Mutex::new(runtime),
+            timing,
+        }
+    }
+
+    pub fn step(&self) -> bool {
+        self.runtime.lock().step()
+    }
+
+    pub fn get_thread_step_count(&self) -> u64 {
+        self.runtime.lock().step_counter
+    }
+
+    /// Create a new random number generator.
+    pub fn new_rng(&self) -> StdRng {
+        let mut rng = self.rng.lock();
+        StdRng::from_rng(rng.deref_mut()).unwrap()
+    }
+
+    /// Create a new node.
+    pub fn new_node(self: &Arc<Self>) -> Arc<Node> {
+        let mut nodes = self.nodes.lock();
+        let id = nodes.len() as NodeId;
+        let node = Arc::new(Node::new(id, self.clone(), self.new_rng()));
+        nodes.push(node.clone());
+        node
+    }
+
+    /// Get an internal node state by id.
+    fn get_node(&self, id: NodeId) -> Option<Arc<Node>> {
+        let nodes = self.nodes.lock();
+        let num = id as usize;
+        if num < nodes.len() {
+            Some(nodes[num].clone())
+        } else {
+            None
+        }
+    }
+
+    pub fn stop_all(&self) {
+        self.runtime.lock().crash_all_threads();
+    }
+
+    /// Returns a writable end of a TCP connection, to send src->dst messages.
+    pub fn open_tcp(self: &Arc<World>, dst: NodeId) -> TCP {
+        // TODO: replace unwrap() with /dev/null socket.
+        let dst = self.get_node(dst).unwrap();
+        let dst_accept = dst.node_events.lock().clone();
+
+        let rng = self.new_rng();
+        self.network_task.start_new_connection(rng, dst_accept)
+    }
+
+    /// Get current time.
+    pub fn now(&self) -> u64 {
+        self.timing.now()
+    }
+
+    /// Get a copy of the internal clock.
+    pub fn clock(&self) -> Arc<Timing> {
+        self.timing.clone()
+    }
+
+    pub fn add_event(&self, node: NodeId, data: String) {
+        let time = self.now();
+        self.events.lock().push(SimEvent { time, node, data });
+    }
+
+    pub fn take_events(&self) -> Vec<SimEvent> {
+        let mut events = self.events.lock();
+        let mut res = Vec::new();
+        std::mem::swap(&mut res, &mut events);
+        res
+    }
+
+    pub fn deallocate(&self) {
+        self.stop_all();
+        self.timing.clear();
+        self.nodes.lock().clear();
+    }
+}
+
+/// Internal node state.
+pub struct Node {
+    pub id: NodeId,
+    node_events: Mutex<Chan<NodeEvent>>,
+    world: Arc<World>,
+    pub(crate) rng: Mutex<StdRng>,
+}
+
+impl Node {
+    pub fn new(id: NodeId, world: Arc<World>, rng: StdRng) -> Node {
+        Node {
+            id,
+            node_events: Mutex::new(Chan::new()),
+            world,
+            rng: Mutex::new(rng),
+        }
+    }
+
+    /// Spawn a new thread with this node context.
+    pub fn launch(self: &Arc<Self>, f: impl FnOnce(NodeOs) + Send + 'static) -> ExternalHandle {
+        let node = self.clone();
+        let world = self.world.clone();
+        self.world.runtime.lock().spawn(move || {
+            f(NodeOs::new(world, node.clone()));
+        })
+    }
+
+    /// Returns a channel to receive Accepts and internal messages.
+    pub fn node_events(&self) -> Chan<NodeEvent> {
+        self.node_events.lock().clone()
+    }
+
+    /// This will drop all in-flight Accept messages.
+    pub fn replug_node_events(&self, chan: Chan<NodeEvent>) {
+        *self.node_events.lock() = chan;
+    }
+
+    /// Append event to the world's log.
+    pub fn log_event(&self, data: String) {
+        self.world.add_event(self.id, data)
+    }
+}

libs/desim/tests/reliable_copy_test.rs

@@ -0,0 +1,244 @@
+//! Simple test to verify that simulator is working.
+#[cfg(test)]
+mod reliable_copy_test {
+    use anyhow::Result;
+    use desim::executor::{self, PollSome};
+    use desim::options::{Delay, NetworkOptions};
+    use desim::proto::{NetEvent, NodeEvent, ReplCell};
+    use desim::world::{NodeId, World};
+    use desim::{node_os::NodeOs, proto::AnyMessage};
+    use parking_lot::Mutex;
+    use std::sync::Arc;
+    use tracing::info;
+
+    /// Disk storage trait and implementation.
+    pub trait Storage<T> {
+        fn flush_pos(&self) -> u32;
+        fn flush(&mut self) -> Result<()>;
+        fn write(&mut self, t: T);
+    }
+
+    #[derive(Clone)]
+    pub struct SharedStorage<T> {
+        pub state: Arc<Mutex<InMemoryStorage<T>>>,
+    }
+
+    impl<T> SharedStorage<T> {
+        pub fn new() -> Self {
+            Self {
+                state: Arc::new(Mutex::new(InMemoryStorage::new())),
+            }
+        }
+    }
+
+    impl<T> Storage<T> for SharedStorage<T> {
+        fn flush_pos(&self) -> u32 {
+            self.state.lock().flush_pos
+        }
+
+        fn flush(&mut self) -> Result<()> {
+            executor::yield_me(0);
+            self.state.lock().flush()
+        }
+
+        fn write(&mut self, t: T) {
+            executor::yield_me(0);
+            self.state.lock().write(t);
+        }
+    }
+
+    pub struct InMemoryStorage<T> {
+        pub data: Vec<T>,
+        pub flush_pos: u32,
+    }
+
+    impl<T> InMemoryStorage<T> {
+        pub fn new() -> Self {
+            Self {
+                data: Vec::new(),
+                flush_pos: 0,
+            }
+        }
+
+        pub fn flush(&mut self) -> Result<()> {
+            self.flush_pos = self.data.len() as u32;
+            Ok(())
+        }
+
+        pub fn write(&mut self, t: T) {
+            self.data.push(t);
+        }
+    }
+
+    /// Server implementation.
+    pub fn run_server(os: NodeOs, mut storage: Box<dyn Storage<u32>>) {
+        info!("started server");
+
+        let node_events = os.node_events();
+        let mut epoll_vec: Vec<Box<dyn PollSome>> = vec![Box::new(node_events.clone())];
+        let mut sockets = vec![];
+
+        loop {
+            let index = executor::epoll_chans(&epoll_vec, -1).unwrap();
+
+            if index == 0 {
+                let node_event = node_events.must_recv();
+                info!("got node event: {:?}", node_event);
+                if let NodeEvent::Accept(tcp) = node_event {
+                    tcp.send(AnyMessage::Just32(storage.flush_pos()));
+                    epoll_vec.push(Box::new(tcp.recv_chan()));
+                    sockets.push(tcp);
+                }
+                continue;
+            }
+
+            let recv_chan = sockets[index - 1].recv_chan();
+            let socket = &sockets[index - 1];
+
+            let event = recv_chan.must_recv();
+            info!("got event: {:?}", event);
+            if let NetEvent::Message(AnyMessage::ReplCell(cell)) = event {
+                if cell.seqno != storage.flush_pos() {
+                    info!("got out of order data: {:?}", cell);
+                    continue;
+                }
+                storage.write(cell.value);
+                storage.flush().unwrap();
+                socket.send(AnyMessage::Just32(storage.flush_pos()));
+            }
+        }
+    }
+
+    /// Client copies all data from array to the remote node.
+    pub fn run_client(os: NodeOs, data: &[ReplCell], dst: NodeId) {
+        info!("started client");
+
+        let mut delivered = 0;
+
+        let mut sock = os.open_tcp(dst);
+        let mut recv_chan = sock.recv_chan();
+
+        while delivered < data.len() {
+            let num = &data[delivered];
+            info!("sending data: {:?}", num.clone());
+            sock.send(AnyMessage::ReplCell(num.clone()));
+
+            // loop {
+            let event = recv_chan.recv();
+            match event {
+                NetEvent::Message(AnyMessage::Just32(flush_pos)) => {
+                    if flush_pos == 1 + delivered as u32 {
+                        delivered += 1;
+                    }
+                }
+                NetEvent::Closed => {
+                    info!("connection closed, reestablishing");
+                    sock = os.open_tcp(dst);
+                    recv_chan = sock.recv_chan();
+                }
+                _ => {}
+            }
+
+            // }
+        }
+
+        let sock = os.open_tcp(dst);
+        for num in data {
+            info!("sending data: {:?}", num.clone());
+            sock.send(AnyMessage::ReplCell(num.clone()));
+        }
+
+        info!("sent all data and finished client");
+    }
+
+    /// Run test simulations.
+    #[test]
+    fn sim_example_reliable_copy() {
+        utils::logging::init(
+            utils::logging::LogFormat::Test,
+            utils::logging::TracingErrorLayerEnablement::Disabled,
+            utils::logging::Output::Stdout,
+        )
+        .expect("logging init failed");
+
+        let delay = Delay {
+            min: 1,
+            max: 60,
+            fail_prob: 0.4,
+        };
+
+        let network = NetworkOptions {
+            keepalive_timeout: Some(50),
+            connect_delay: delay.clone(),
+            send_delay: delay.clone(),
+        };
+
+        for seed in 0..20 {
+            let u32_data: [u32; 5] = [1, 2, 3, 4, 5];
+            let data = u32_to_cells(&u32_data, 1);
+            let world = Arc::new(World::new(seed, Arc::new(network.clone())));
+
+            start_simulation(Options {
+                world,
+                time_limit: 1_000_000,
+                client_fn: Box::new(move |os, server_id| run_client(os, &data, server_id)),
+                u32_data,
+            });
+        }
+    }
+
+    pub struct Options {
+        pub world: Arc<World>,
+        pub time_limit: u64,
+        pub u32_data: [u32; 5],
+        pub client_fn: Box<dyn FnOnce(NodeOs, u32) + Send + 'static>,
+    }
+
+    pub fn start_simulation(options: Options) {
+        let world = options.world;
+
+        let client_node = world.new_node();
+        let server_node = world.new_node();
+        let server_id = server_node.id;
+
+        // start the client thread
+        client_node.launch(move |os| {
+            let client_fn = options.client_fn;
+            client_fn(os, server_id);
+        });
+
+        // start the server thread
+        let shared_storage = SharedStorage::new();
+        let server_storage = shared_storage.clone();
+        server_node.launch(move |os| run_server(os, Box::new(server_storage)));
+
+        while world.step() && world.now() < options.time_limit {}
+
+        let disk_data = shared_storage.state.lock().data.clone();
+        assert!(verify_data(&disk_data, &options.u32_data[..]));
+    }
+
+    pub fn u32_to_cells(data: &[u32], client_id: u32) -> Vec<ReplCell> {
+        let mut res = Vec::new();
+        for (i, _) in data.iter().enumerate() {
+            res.push(ReplCell {
+                client_id,
+                seqno: i as u32,
+                value: data[i],
+            });
+        }
+        res
+    }
+
+    fn verify_data(disk_data: &[u32], data: &[u32]) -> bool {
+        if disk_data.len() != data.len() {
+            return false;
+        }
+        for i in 0..data.len() {
+            if disk_data[i] != data[i] {
+                return false;
+            }
+        }
+        true
+    }
+}

libs/metrics/Cargo.toml

@@ -13,6 +13,9 @@ twox-hash.workspace = true
 
 workspace_hack.workspace = true
 
+[target.'cfg(target_os = "linux")'.dependencies]
+procfs.workspace = true
+
 [dev-dependencies]
 rand = "0.8"
 rand_distr = "0.4.3"

libs/metrics/src/lib.rs

@@ -31,6 +31,8 @@ pub use wrappers::{CountedReader, CountedWriter};
 mod hll;
 pub mod metric_vec_duration;
 pub use hll::{HyperLogLog, HyperLogLogVec};
+#[cfg(target_os = "linux")]
+pub mod more_process_metrics;
 
 pub type UIntGauge = GenericGauge<AtomicU64>;
 pub type UIntGaugeVec = GenericGaugeVec<AtomicU64>;
@@ -113,7 +115,6 @@ pub fn set_build_info_metric(revision: &str, build_tag: &str) {
 // performed by the process.
 // We know the size of the block, so we can determine the I/O bytes out of it.
 // The value might be not 100% exact, but should be fine for Prometheus metrics in this case.
-#[allow(clippy::unnecessary_cast)]
 fn update_rusage_metrics() {
     let rusage_stats = get_rusage_stats();
 

libs/metrics/src/more_process_metrics.rs

@@ -0,0 +1,54 @@
+//! process metrics that the [`::prometheus`] crate doesn't provide.
+
+// This module has heavy inspiration from the prometheus crate's `process_collector.rs`.
+
+use crate::UIntGauge;
+
+pub struct Collector {
+    descs: Vec<prometheus::core::Desc>,
+    vmlck: crate::UIntGauge,
+}
+
+const NMETRICS: usize = 1;
+
+impl prometheus::core::Collector for Collector {
+    fn desc(&self) -> Vec<&prometheus::core::Desc> {
+        self.descs.iter().collect()
+    }
+
+    fn collect(&self) -> Vec<prometheus::proto::MetricFamily> {
+        let Ok(myself) = procfs::process::Process::myself() else {
+            return vec![];
+        };
+        let mut mfs = Vec::with_capacity(NMETRICS);
+        if let Ok(status) = myself.status() {
+            if let Some(vmlck) = status.vmlck {
+                self.vmlck.set(vmlck);
+                mfs.extend(self.vmlck.collect())
+            }
+        }
+        mfs
+    }
+}
+
+impl Collector {
+    pub fn new() -> Self {
+        let mut descs = Vec::new();
+
+        let vmlck =
+            UIntGauge::new("libmetrics_process_status_vmlck", "/proc/self/status vmlck").unwrap();
+        descs.extend(
+            prometheus::core::Collector::desc(&vmlck)
+                .into_iter()
+                .cloned(),
+        );
+
+        Self { descs, vmlck }
+    }
+}
+
+impl Default for Collector {
+    fn default() -> Self {
+        Self::new()
+    }
+}

libs/pageserver_api/src/models.rs

@@ -192,6 +192,16 @@ pub struct TimelineCreateRequest {
     pub pg_version: Option<u32>,
 }
 
+#[derive(Serialize, Deserialize)]
+pub struct TenantShardSplitRequest {
+    pub new_shard_count: u8,
+}
+
+#[derive(Serialize, Deserialize)]
+pub struct TenantShardSplitResponse {
+    pub new_shards: Vec<TenantShardId>,
+}
+
 /// Parameters that apply to all shards in a tenant.  Used during tenant creation.
 #[derive(Serialize, Deserialize, Debug)]
 #[serde(deny_unknown_fields)]
@@ -484,6 +494,8 @@ pub struct TimelineInfo {
     pub current_logical_size: u64,
     pub current_logical_size_is_accurate: bool,
 
+    pub directory_entries_counts: Vec<u64>,
+
     /// Sum of the size of all layer files.
     /// If a layer is present in both local FS and S3, it counts only once.
     pub current_physical_size: Option<u64>, // is None when timeline is Unloaded
@@ -649,15 +661,35 @@ pub struct WalRedoManagerStatus {
     pub pid: Option<u32>,
 }
 
+pub mod virtual_file {
+    #[derive(
+        Copy,
+        Clone,
+        PartialEq,
+        Eq,
+        Hash,
+        strum_macros::EnumString,
+        strum_macros::Display,
+        serde_with::DeserializeFromStr,
+        serde_with::SerializeDisplay,
+        Debug,
+    )]
+    #[strum(serialize_all = "kebab-case")]
+    pub enum IoEngineKind {
+        StdFs,
+        #[cfg(target_os = "linux")]
+        TokioEpollUring,
+    }
+}
+
 // Wrapped in libpq CopyData
 #[derive(PartialEq, Eq, Debug)]
 pub enum PagestreamFeMessage {
     Exists(PagestreamExistsRequest),
     Nblocks(PagestreamNblocksRequest),
-    GetLatestPage(PagestreamGetLatestPageRequest), // for compatinility with old clients
+    GetPage(PagestreamGetPageRequest),
     DbSize(PagestreamDbSizeRequest),
     GetSlruSegment(PagestreamGetSlruSegmentRequest),
-    GetPage(PagestreamGetPageRequest),
 }
 
 // Wrapped in libpq CopyData
@@ -698,44 +730,36 @@ impl TryFrom<u8> for PagestreamBeMessageTag {
 
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamExistsRequest {
-    pub horizon: Lsn,
+    pub latest: bool,
     pub lsn: Lsn,
     pub rel: RelTag,
 }
 
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamNblocksRequest {
-    pub horizon: Lsn,
+    pub latest: bool,
     pub lsn: Lsn,
     pub rel: RelTag,
 }
 
 #[derive(Debug, PartialEq, Eq)]
-pub struct PagestreamGetLatestPageRequest {
+pub struct PagestreamGetPageRequest {
     pub latest: bool,
     pub lsn: Lsn,
     pub rel: RelTag,
     pub blkno: u32,
 }
 
-#[derive(Debug, PartialEq, Eq)]
-pub struct PagestreamGetPageRequest {
-    pub horizon: Lsn,
-    pub lsn: Lsn,
-    pub rel: RelTag,
-    pub blkno: u32,
-}
-
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamDbSizeRequest {
-    pub horizon: Lsn,
+    pub latest: bool,
     pub lsn: Lsn,
     pub dbnode: u32,
 }
 
 #[derive(Debug, PartialEq, Eq)]
 pub struct PagestreamGetSlruSegmentRequest {
-    pub horizon: Lsn,
+    pub latest: bool,
     pub lsn: Lsn,
     pub kind: u8,
     pub segno: u32,
@@ -789,7 +813,7 @@ impl PagestreamFeMessage {
         match self {
             Self::Exists(req) => {
                 bytes.put_u8(0);
-                bytes.put_u64(req.horizon.0);
+                bytes.put_u8(u8::from(req.latest));
                 bytes.put_u64(req.lsn.0);
                 bytes.put_u32(req.rel.spcnode);
                 bytes.put_u32(req.rel.dbnode);
@@ -799,7 +823,7 @@ impl PagestreamFeMessage {
 
             Self::Nblocks(req) => {
                 bytes.put_u8(1);
-                bytes.put_u64(req.horizon.0);
+                bytes.put_u8(u8::from(req.latest));
                 bytes.put_u64(req.lsn.0);
                 bytes.put_u32(req.rel.spcnode);
                 bytes.put_u32(req.rel.dbnode);
@@ -807,7 +831,7 @@ impl PagestreamFeMessage {
                 bytes.put_u8(req.rel.forknum);
             }
 
-            Self::GetLatestPage(req) => {
+            Self::GetPage(req) => {
                 bytes.put_u8(2);
                 bytes.put_u8(u8::from(req.latest));
                 bytes.put_u64(req.lsn.0);
@@ -820,29 +844,18 @@ impl PagestreamFeMessage {
 
             Self::DbSize(req) => {
                 bytes.put_u8(3);
-                bytes.put_u64(req.horizon.0);
+                bytes.put_u8(u8::from(req.latest));
                 bytes.put_u64(req.lsn.0);
                 bytes.put_u32(req.dbnode);
             }
 
             Self::GetSlruSegment(req) => {
                 bytes.put_u8(4);
-                bytes.put_u64(req.horizon.0);
+                bytes.put_u8(u8::from(req.latest));
                 bytes.put_u64(req.lsn.0);
                 bytes.put_u8(req.kind);
                 bytes.put_u32(req.segno);
             }
-
-            Self::GetPage(req) => {
-                bytes.put_u8(5);
-                bytes.put_u64(req.horizon.0);
-                bytes.put_u64(req.lsn.0);
-                bytes.put_u32(req.rel.spcnode);
-                bytes.put_u32(req.rel.dbnode);
-                bytes.put_u32(req.rel.relnode);
-                bytes.put_u8(req.rel.forknum);
-                bytes.put_u32(req.blkno);
-            }
         }
 
         bytes.into()
@@ -858,7 +871,7 @@ impl PagestreamFeMessage {
         let msg_tag = body.read_u8()?;
         match msg_tag {
             0 => Ok(PagestreamFeMessage::Exists(PagestreamExistsRequest {
-                horizon: Lsn::from(body.read_u64::<BigEndian>()?),
+                latest: body.read_u8()? != 0,
                 lsn: Lsn::from(body.read_u64::<BigEndian>()?),
                 rel: RelTag {
                     spcnode: body.read_u32::<BigEndian>()?,
@@ -868,7 +881,7 @@ impl PagestreamFeMessage {
                 },
             })),
             1 => Ok(PagestreamFeMessage::Nblocks(PagestreamNblocksRequest {
-                horizon: Lsn::from(body.read_u64::<BigEndian>()?),
+                latest: body.read_u8()? != 0,
                 lsn: Lsn::from(body.read_u64::<BigEndian>()?),
                 rel: RelTag {
                     spcnode: body.read_u32::<BigEndian>()?,
@@ -877,34 +890,8 @@ impl PagestreamFeMessage {
                     forknum: body.read_u8()?,
                 },
             })),
-            2 => Ok(PagestreamFeMessage::GetLatestPage(
-                PagestreamGetLatestPageRequest {
-                    latest: body.read_u8()? != 0,
-                    lsn: Lsn::from(body.read_u64::<BigEndian>()?),
-                    rel: RelTag {
-                        spcnode: body.read_u32::<BigEndian>()?,
-                        dbnode: body.read_u32::<BigEndian>()?,
-                        relnode: body.read_u32::<BigEndian>()?,
-                        forknum: body.read_u8()?,
-                    },
-                    blkno: body.read_u32::<BigEndian>()?,
-                },
-            )),
-            3 => Ok(PagestreamFeMessage::DbSize(PagestreamDbSizeRequest {
-                horizon: Lsn::from(body.read_u64::<BigEndian>()?),
-                lsn: Lsn::from(body.read_u64::<BigEndian>()?),
-                dbnode: body.read_u32::<BigEndian>()?,
-            })),
-            4 => Ok(PagestreamFeMessage::GetSlruSegment(
-                PagestreamGetSlruSegmentRequest {
-                    horizon: Lsn::from(body.read_u64::<BigEndian>()?),
-                    lsn: Lsn::from(body.read_u64::<BigEndian>()?),
-                    kind: body.read_u8()?,
-                    segno: body.read_u32::<BigEndian>()?,
-                },
-            )),
-            5 => Ok(PagestreamFeMessage::GetPage(PagestreamGetPageRequest {
-                horizon: Lsn::from(body.read_u64::<BigEndian>()?),
+            2 => Ok(PagestreamFeMessage::GetPage(PagestreamGetPageRequest {
+                latest: body.read_u8()? != 0,
                 lsn: Lsn::from(body.read_u64::<BigEndian>()?),
                 rel: RelTag {
                     spcnode: body.read_u32::<BigEndian>()?,
@@ -914,6 +901,19 @@ impl PagestreamFeMessage {
                 },
                 blkno: body.read_u32::<BigEndian>()?,
             })),
+            3 => Ok(PagestreamFeMessage::DbSize(PagestreamDbSizeRequest {
+                latest: body.read_u8()? != 0,
+                lsn: Lsn::from(body.read_u64::<BigEndian>()?),
+                dbnode: body.read_u32::<BigEndian>()?,
+            })),
+            4 => Ok(PagestreamFeMessage::GetSlruSegment(
+                PagestreamGetSlruSegmentRequest {
+                    latest: body.read_u8()? != 0,
+                    lsn: Lsn::from(body.read_u64::<BigEndian>()?),
+                    kind: body.read_u8()?,
+                    segno: body.read_u32::<BigEndian>()?,
+                },
+            )),
             _ => bail!("unknown smgr message tag: {:?}", msg_tag),
         }
     }
@@ -1038,7 +1038,7 @@ mod tests {
         // Test serialization/deserialization of PagestreamFeMessage
         let messages = vec![
             PagestreamFeMessage::Exists(PagestreamExistsRequest {
-                horizon: Lsn::MAX,
+                latest: true,
                 lsn: Lsn(4),
                 rel: RelTag {
                     forknum: 1,
@@ -1048,7 +1048,7 @@ mod tests {
                 },
             }),
             PagestreamFeMessage::Nblocks(PagestreamNblocksRequest {
-                horizon: Lsn::INVALID,
+                latest: false,
                 lsn: Lsn(4),
                 rel: RelTag {
                     forknum: 1,
@@ -1058,8 +1058,8 @@ mod tests {
                 },
             }),
             PagestreamFeMessage::GetPage(PagestreamGetPageRequest {
-                horizon: Lsn::MAX,
-                lsn: Lsn::INVALID,
+                latest: true,
+                lsn: Lsn(4),
                 rel: RelTag {
                     forknum: 1,
                     spcnode: 2,
@@ -1069,7 +1069,7 @@ mod tests {
                 blkno: 7,
             }),
             PagestreamFeMessage::DbSize(PagestreamDbSizeRequest {
-                horizon: Lsn::MAX,
+                latest: true,
                 lsn: Lsn(4),
                 dbnode: 7,
             }),

libs/pageserver_api/src/reltag.rs

@@ -124,6 +124,7 @@ impl RelTag {
     Ord,
     strum_macros::EnumIter,
     strum_macros::FromRepr,
+    enum_map::Enum,
 )]
 #[repr(u8)]
 pub enum SlruKind {

libs/pageserver_api/src/shard.rs

@@ -88,12 +88,36 @@ impl TenantShardId {
     pub fn is_unsharded(&self) -> bool {
         self.shard_number == ShardNumber(0) && self.shard_count == ShardCount(0)
     }
+
+    /// Convenience for dropping the tenant_id and just getting the ShardIndex: this
+    /// is useful when logging from code that is already in a span that includes tenant ID, to
+    /// keep messages reasonably terse.
     pub fn to_index(&self) -> ShardIndex {
         ShardIndex {
             shard_number: self.shard_number,
             shard_count: self.shard_count,
         }
     }
+
+    /// Calculate the children of this TenantShardId when splitting the overall tenant into
+    /// the given number of shards.
+    pub fn split(&self, new_shard_count: ShardCount) -> Vec<TenantShardId> {
+        let effective_old_shard_count = std::cmp::max(self.shard_count.0, 1);
+        let mut child_shards = Vec::new();
+        for shard_number in 0..ShardNumber(new_shard_count.0).0 {
+            // Key mapping is based on a round robin mapping of key hash modulo shard count,
+            // so our child shards are the ones which the same keys would map to.
+            if shard_number % effective_old_shard_count == self.shard_number.0 {
+                child_shards.push(TenantShardId {
+                    tenant_id: self.tenant_id,
+                    shard_number: ShardNumber(shard_number),
+                    shard_count: new_shard_count,
+                })
+            }
+        }
+
+        child_shards
+    }
 }
 
 /// Formatting helper
@@ -793,4 +817,108 @@ mod tests {
         let shard = key_to_shard_number(ShardCount(10), DEFAULT_STRIPE_SIZE, &key);
         assert_eq!(shard, ShardNumber(8));
     }
+
+    #[test]
+    fn shard_id_split() {
+        let tenant_id = TenantId::generate();
+        let parent = TenantShardId::unsharded(tenant_id);
+
+        // Unsharded into 2
+        assert_eq!(
+            parent.split(ShardCount(2)),
+            vec![
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(2),
+                    shard_number: ShardNumber(0)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(2),
+                    shard_number: ShardNumber(1)
+                }
+            ]
+        );
+
+        // Unsharded into 4
+        assert_eq!(
+            parent.split(ShardCount(4)),
+            vec![
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(4),
+                    shard_number: ShardNumber(0)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(4),
+                    shard_number: ShardNumber(1)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(4),
+                    shard_number: ShardNumber(2)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(4),
+                    shard_number: ShardNumber(3)
+                }
+            ]
+        );
+
+        // count=1 into 2 (check this works the same as unsharded.)
+        let parent = TenantShardId {
+            tenant_id,
+            shard_count: ShardCount(1),
+            shard_number: ShardNumber(0),
+        };
+        assert_eq!(
+            parent.split(ShardCount(2)),
+            vec![
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(2),
+                    shard_number: ShardNumber(0)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(2),
+                    shard_number: ShardNumber(1)
+                }
+            ]
+        );
+
+        // count=2 into count=8
+        let parent = TenantShardId {
+            tenant_id,
+            shard_count: ShardCount(2),
+            shard_number: ShardNumber(1),
+        };
+        assert_eq!(
+            parent.split(ShardCount(8)),
+            vec![
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(8),
+                    shard_number: ShardNumber(1)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(8),
+                    shard_number: ShardNumber(3)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(8),
+                    shard_number: ShardNumber(5)
+                },
+                TenantShardId {
+                    tenant_id,
+                    shard_count: ShardCount(8),
+                    shard_number: ShardNumber(7)
+                },
+            ]
+        );
+    }
 }

libs/postgres_ffi/src/lib.rs

@@ -3,7 +3,7 @@
 #![allow(non_snake_case)]
 // bindgen creates some unsafe code with no doc comments.
 #![allow(clippy::missing_safety_doc)]
-// noted at 1.63 that in many cases there's a u32 -> u32 transmutes in bindgen code.
+// noted at 1.63 that in many cases there's u32 -> u32 transmutes in bindgen code.
 #![allow(clippy::useless_transmute)]
 // modules included with the postgres_ffi macro depend on the types of the specific version's
 // types, and trigger a too eager lint.

libs/postgres_ffi/src/xlog_utils.rs

@@ -119,6 +119,11 @@ pub fn generate_pg_control(
     // Generate new pg_control needed for bootstrap
     checkpoint.redo = normalize_lsn(lsn, WAL_SEGMENT_SIZE).0;
 
+    //reset some fields we don't want to preserve
+    //TODO Check this.
+    //We may need to determine the value from twophase data.
+    checkpoint.oldestActiveXid = 0;
+
     //save new values in pg_control
     pg_control.checkPoint = 0;
     pg_control.checkPointCopy = checkpoint;
@@ -426,11 +431,11 @@ pub fn generate_wal_segment(segno: u64, system_id: u64, lsn: Lsn) -> Result<Byte
 
 #[repr(C)]
 #[derive(Serialize)]
-struct XlLogicalMessage {
-    db_id: Oid,
-    transactional: uint32, // bool, takes 4 bytes due to alignment in C structures
-    prefix_size: uint64,
-    message_size: uint64,
+pub struct XlLogicalMessage {
+    pub db_id: Oid,
+    pub transactional: uint32, // bool, takes 4 bytes due to alignment in C structures
+    pub prefix_size: uint64,
+    pub message_size: uint64,
 }
 
 impl XlLogicalMessage {

libs/pq_proto/Cargo.toml

@@ -13,5 +13,6 @@ rand.workspace = true
 tokio.workspace = true
 tracing.workspace = true
 thiserror.workspace = true
+serde.workspace = true
 
 workspace_hack.workspace = true

libs/pq_proto/src/lib.rs

@@ -7,6 +7,7 @@ pub mod framed;
 
 use byteorder::{BigEndian, ReadBytesExt};
 use bytes::{Buf, BufMut, Bytes, BytesMut};
+use serde::{Deserialize, Serialize};
 use std::{borrow::Cow, collections::HashMap, fmt, io, str};
 
 // re-export for use in utils pageserver_feedback.rs
@@ -123,7 +124,7 @@ impl StartupMessageParams {
     }
 }
 
-#[derive(Debug, Hash, PartialEq, Eq, Clone, Copy)]
+#[derive(Debug, Hash, PartialEq, Eq, Clone, Copy, Serialize, Deserialize)]
 pub struct CancelKeyData {
     pub backend_pid: i32,
     pub cancel_key: i32,

libs/remote_storage/Cargo.toml

@@ -15,11 +15,13 @@ aws-sdk-s3.workspace = true
 aws-credential-types.workspace = true
 bytes.workspace = true
 camino.workspace = true
+humantime.workspace = true
 hyper = { workspace = true, features = ["stream"] }
 futures.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 tokio = { workspace = true, features = ["sync", "fs", "io-util"] }
+tokio-stream.workspace = true
 tokio-util = { workspace = true, features = ["compat"] }
 toml_edit.workspace = true
 tracing.workspace = true

libs/remote_storage/src/azure_blob.rs

@@ -22,16 +22,15 @@ use azure_storage_blobs::{blob::operations::GetBlobBuilder, prelude::ContainerCl
 use bytes::Bytes;
 use futures::stream::Stream;
 use futures_util::StreamExt;
+use futures_util::TryStreamExt;
 use http_types::{StatusCode, Url};
-use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
 use tracing::debug;
 
-use crate::s3_bucket::RequestKind;
-use crate::TimeTravelError;
 use crate::{
-    AzureConfig, ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, RemotePath,
-    RemoteStorage, StorageMetadata,
+    error::Cancelled, s3_bucket::RequestKind, AzureConfig, ConcurrencyLimiter, Download,
+    DownloadError, Listing, ListingMode, RemotePath, RemoteStorage, StorageMetadata,
+    TimeTravelError, TimeoutOrCancel,
 };
 
 pub struct AzureBlobStorage {
@@ -39,10 +38,12 @@ pub struct AzureBlobStorage {
     prefix_in_container: Option<String>,
     max_keys_per_list_response: Option<NonZeroU32>,
     concurrency_limiter: ConcurrencyLimiter,
+    // Per-request timeout. Accessible for tests.
+    pub timeout: Duration,
 }
 
 impl AzureBlobStorage {
-    pub fn new(azure_config: &AzureConfig) -> Result<Self> {
+    pub fn new(azure_config: &AzureConfig, timeout: Duration) -> Result<Self> {
         debug!(
             "Creating azure remote storage for azure container {}",
             azure_config.container_name
@@ -79,6 +80,7 @@ impl AzureBlobStorage {
             prefix_in_container: azure_config.prefix_in_container.to_owned(),
             max_keys_per_list_response,
             concurrency_limiter: ConcurrencyLimiter::new(azure_config.concurrency_limit.get()),
+            timeout,
         })
     }
 
@@ -121,8 +123,11 @@ impl AzureBlobStorage {
     async fn download_for_builder(
         &self,
         builder: GetBlobBuilder,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
-        let mut response = builder.into_stream();
+        let kind = RequestKind::Get;
+
+        let _permit = self.permit(kind, cancel).await?;
 
         let mut etag = None;
         let mut last_modified = None;
@@ -130,39 +135,70 @@ impl AzureBlobStorage {
         // TODO give proper streaming response instead of buffering into RAM
         // https://github.com/neondatabase/neon/issues/5563
 
-        let mut bufs = Vec::new();
-        while let Some(part) = response.next().await {
-            let part = part.map_err(to_download_error)?;
-            let etag_str: &str = part.blob.properties.etag.as_ref();
-            if etag.is_none() {
-                etag = Some(etag.unwrap_or_else(|| etag_str.to_owned()));
+        let download = async {
+            let response = builder
+                // convert to concrete Pageable
+                .into_stream()
+                // convert to TryStream
+                .into_stream()
+                .map_err(to_download_error);
+
+            // apply per request timeout
+            let response = tokio_stream::StreamExt::timeout(response, self.timeout);
+
+            // flatten
+            let response = response.map(|res| match res {
+                Ok(res) => res,
+                Err(_elapsed) => Err(DownloadError::Timeout),
+            });
+
+            let mut response = std::pin::pin!(response);
+
+            let mut bufs = Vec::new();
+            while let Some(part) = response.next().await {
+                let part = part?;
+                let etag_str: &str = part.blob.properties.etag.as_ref();
+                if etag.is_none() {
+                    etag = Some(etag.unwrap_or_else(|| etag_str.to_owned()));
+                }
+                if last_modified.is_none() {
+                    last_modified = Some(part.blob.properties.last_modified.into());
+                }
+                if let Some(blob_meta) = part.blob.metadata {
+                    metadata.extend(blob_meta.iter().map(|(k, v)| (k.to_owned(), v.to_owned())));
+                }
+                let data = part
+                    .data
+                    .collect()
+                    .await
+                    .map_err(|e| DownloadError::Other(e.into()))?;
+                bufs.push(data);
             }
-            if last_modified.is_none() {
-                last_modified = Some(part.blob.properties.last_modified.into());
-            }
-            if let Some(blob_meta) = part.blob.metadata {
-                metadata.extend(blob_meta.iter().map(|(k, v)| (k.to_owned(), v.to_owned())));
-            }
-            let data = part
-                .data
-                .collect()
-                .await
-                .map_err(|e| DownloadError::Other(e.into()))?;
-            bufs.push(data);
+            Ok(Download {
+                download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
+                etag,
+                last_modified,
+                metadata: Some(StorageMetadata(metadata)),
+            })
+        };
+
+        tokio::select! {
+            bufs = download => bufs,
+            _ = cancel.cancelled() => Err(DownloadError::Cancelled),
         }
-        Ok(Download {
-            download_stream: Box::pin(futures::stream::iter(bufs.into_iter().map(Ok))),
-            etag,
-            last_modified,
-            metadata: Some(StorageMetadata(metadata)),
-        })
     }
 
-    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
-        self.concurrency_limiter
-            .acquire(kind)
-            .await
-            .expect("semaphore is never closed")
+    async fn permit(
+        &self,
+        kind: RequestKind,
+        cancel: &CancellationToken,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, Cancelled> {
+        let acquire = self.concurrency_limiter.acquire(kind);
+
+        tokio::select! {
+            permit = acquire => Ok(permit.expect("never closed")),
+            _ = cancel.cancelled() => Err(Cancelled),
+        }
     }
 }
 
@@ -191,53 +227,88 @@ impl RemoteStorage for AzureBlobStorage {
         &self,
         prefix: Option<&RemotePath>,
         mode: ListingMode,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<Listing, DownloadError> {
-        // get the passed prefix or if it is not set use prefix_in_bucket value
-        let list_prefix = prefix
-            .map(|p| self.relative_path_to_name(p))
-            .or_else(|| self.prefix_in_container.clone())
-            .map(|mut p| {
-                // required to end with a separator
-                // otherwise request will return only the entry of a prefix
-                if matches!(mode, ListingMode::WithDelimiter)
-                    && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
-                {
-                    p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
-                }
-                p
+        let _permit = self.permit(RequestKind::List, cancel).await?;
+
+        let op = async {
+            // get the passed prefix or if it is not set use prefix_in_bucket value
+            let list_prefix = prefix
+                .map(|p| self.relative_path_to_name(p))
+                .or_else(|| self.prefix_in_container.clone())
+                .map(|mut p| {
+                    // required to end with a separator
+                    // otherwise request will return only the entry of a prefix
+                    if matches!(mode, ListingMode::WithDelimiter)
+                        && !p.ends_with(REMOTE_STORAGE_PREFIX_SEPARATOR)
+                    {
+                        p.push(REMOTE_STORAGE_PREFIX_SEPARATOR);
+                    }
+                    p
+                });
+
+            let mut builder = self.client.list_blobs();
+
+            if let ListingMode::WithDelimiter = mode {
+                builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+            }
+
+            if let Some(prefix) = list_prefix {
+                builder = builder.prefix(Cow::from(prefix.to_owned()));
+            }
+
+            if let Some(limit) = self.max_keys_per_list_response {
+                builder = builder.max_results(MaxResults::new(limit));
+            }
+
+            let response = builder.into_stream();
+            let response = response.into_stream().map_err(to_download_error);
+            let response = tokio_stream::StreamExt::timeout(response, self.timeout);
+            let response = response.map(|res| match res {
+                Ok(res) => res,
+                Err(_elapsed) => Err(DownloadError::Timeout),
             });
 
-        let mut builder = self.client.list_blobs();
+            let mut response = std::pin::pin!(response);
 
-        if let ListingMode::WithDelimiter = mode {
-            builder = builder.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
+            let mut res = Listing::default();
+
+            let mut max_keys = max_keys.map(|mk| mk.get());
+            while let Some(entry) = response.next().await {
+                let entry = entry?;
+                let prefix_iter = entry
+                    .blobs
+                    .prefixes()
+                    .map(|prefix| self.name_to_relative_path(&prefix.name));
+                res.prefixes.extend(prefix_iter);
+
+                let blob_iter = entry
+                    .blobs
+                    .blobs()
+                    .map(|k| self.name_to_relative_path(&k.name));
+
+                for key in blob_iter {
+                    res.keys.push(key);
+
+                    if let Some(mut mk) = max_keys {
+                        assert!(mk > 0);
+                        mk -= 1;
+                        if mk == 0 {
+                            return Ok(res); // limit reached
+                        }
+                        max_keys = Some(mk);
+                    }
+                }
+            }
+
+            Ok(res)
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(DownloadError::Cancelled),
         }
-
-        if let Some(prefix) = list_prefix {
-            builder = builder.prefix(Cow::from(prefix.to_owned()));
-        }
-
-        if let Some(limit) = self.max_keys_per_list_response {
-            builder = builder.max_results(MaxResults::new(limit));
-        }
-
-        let mut response = builder.into_stream();
-        let mut res = Listing::default();
-        while let Some(l) = response.next().await {
-            let entry = l.map_err(to_download_error)?;
-            let prefix_iter = entry
-                .blobs
-                .prefixes()
-                .map(|prefix| self.name_to_relative_path(&prefix.name));
-            res.prefixes.extend(prefix_iter);
-
-            let blob_iter = entry
-                .blobs
-                .blobs()
-                .map(|k| self.name_to_relative_path(&k.name));
-            res.keys.extend(blob_iter);
-        }
-        Ok(res)
     }
 
     async fn upload(
@@ -246,35 +317,52 @@ impl RemoteStorage for AzureBlobStorage {
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
-        let _permit = self.permit(RequestKind::Put).await;
-        let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+        let _permit = self.permit(RequestKind::Put, cancel).await?;
 
-        let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
-            Box::pin(from);
+        let op = async {
+            let blob_client = self.client.blob_client(self.relative_path_to_name(to));
 
-        let from = NonSeekableStream::new(from, data_size_bytes);
+            let from: Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>> =
+                Box::pin(from);
 
-        let body = azure_core::Body::SeekableStream(Box::new(from));
+            let from = NonSeekableStream::new(from, data_size_bytes);
 
-        let mut builder = blob_client.put_block_blob(body);
+            let body = azure_core::Body::SeekableStream(Box::new(from));
 
-        if let Some(metadata) = metadata {
-            builder = builder.metadata(to_azure_metadata(metadata));
+            let mut builder = blob_client.put_block_blob(body);
+
+            if let Some(metadata) = metadata {
+                builder = builder.metadata(to_azure_metadata(metadata));
+            }
+
+            let fut = builder.into_future();
+            let fut = tokio::time::timeout(self.timeout, fut);
+
+            match fut.await {
+                Ok(Ok(_response)) => Ok(()),
+                Ok(Err(azure)) => Err(azure.into()),
+                Err(_timeout) => Err(TimeoutOrCancel::Cancel.into()),
+            }
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(TimeoutOrCancel::Cancel.into()),
         }
-
-        let _response = builder.into_future().await?;
-
-        Ok(())
     }
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
-        let _permit = self.permit(RequestKind::Get).await;
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         let blob_client = self.client.blob_client(self.relative_path_to_name(from));
 
         let builder = blob_client.get();
 
-        self.download_for_builder(builder).await
+        self.download_for_builder(builder, cancel).await
     }
 
     async fn download_byte_range(
@@ -282,8 +370,8 @@ impl RemoteStorage for AzureBlobStorage {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
-        let _permit = self.permit(RequestKind::Get).await;
         let blob_client = self.client.blob_client(self.relative_path_to_name(from));
 
         let mut builder = blob_client.get();
@@ -295,82 +383,113 @@ impl RemoteStorage for AzureBlobStorage {
         };
         builder = builder.range(range);
 
-        self.download_for_builder(builder).await
+        self.download_for_builder(builder, cancel).await
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
-        let _permit = self.permit(RequestKind::Delete).await;
-        let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
+        self.delete_objects(std::array::from_ref(path), cancel)
+            .await
+    }
 
-        let builder = blob_client.delete();
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Delete, cancel).await?;
 
-        match builder.into_future().await {
-            Ok(_response) => Ok(()),
-            Err(e) => {
-                if let Some(http_err) = e.as_http_error() {
-                    if http_err.status() == StatusCode::NotFound {
-                        return Ok(());
+        let op = async {
+            // TODO batch requests are also not supported by the SDK
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1068
+            // https://github.com/Azure/azure-sdk-for-rust/issues/1249
+            for path in paths {
+                let blob_client = self.client.blob_client(self.relative_path_to_name(path));
+
+                let request = blob_client.delete().into_future();
+
+                let res = tokio::time::timeout(self.timeout, request).await;
+
+                match res {
+                    Ok(Ok(_response)) => continue,
+                    Ok(Err(e)) => {
+                        if let Some(http_err) = e.as_http_error() {
+                            if http_err.status() == StatusCode::NotFound {
+                                continue;
+                            }
+                        }
+                        return Err(e.into());
                     }
+                    Err(_elapsed) => return Err(TimeoutOrCancel::Timeout.into()),
                 }
-                Err(anyhow::Error::new(e))
             }
+
+            Ok(())
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(TimeoutOrCancel::Cancel.into()),
         }
     }
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
-        // Permit is already obtained by inner delete function
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let _permit = self.permit(RequestKind::Copy, cancel).await?;
 
-        // TODO batch requests are also not supported by the SDK
-        // https://github.com/Azure/azure-sdk-for-rust/issues/1068
-        // https://github.com/Azure/azure-sdk-for-rust/issues/1249
-        for path in paths {
-            self.delete(path).await?;
-        }
-        Ok(())
-    }
+        let timeout = tokio::time::sleep(self.timeout);
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
-        let _permit = self.permit(RequestKind::Copy).await;
-        let blob_client = self.client.blob_client(self.relative_path_to_name(to));
+        let mut copy_status = None;
 
-        let source_url = format!(
-            "{}/{}",
-            self.client.url()?,
-            self.relative_path_to_name(from)
-        );
-        let builder = blob_client.copy(Url::from_str(&source_url)?);
+        let op = async {
+            let blob_client = self.client.blob_client(self.relative_path_to_name(to));
 
-        let result = builder.into_future().await?;
+            let source_url = format!(
+                "{}/{}",
+                self.client.url()?,
+                self.relative_path_to_name(from)
+            );
 
-        let mut copy_status = result.copy_status;
-        let start_time = Instant::now();
-        const MAX_WAIT_TIME: Duration = Duration::from_secs(60);
-        loop {
-            match copy_status {
-                CopyStatus::Aborted => {
-                    anyhow::bail!("Received abort for copy from {from} to {to}.");
+            let builder = blob_client.copy(Url::from_str(&source_url)?);
+            let copy = builder.into_future();
+
+            let result = copy.await?;
+
+            copy_status = Some(result.copy_status);
+            loop {
+                match copy_status.as_ref().expect("we always set it to Some") {
+                    CopyStatus::Aborted => {
+                        anyhow::bail!("Received abort for copy from {from} to {to}.");
+                    }
+                    CopyStatus::Failed => {
+                        anyhow::bail!("Received failure response for copy from {from} to {to}.");
+                    }
+                    CopyStatus::Success => return Ok(()),
+                    CopyStatus::Pending => (),
                 }
-                CopyStatus::Failed => {
-                    anyhow::bail!("Received failure response for copy from {from} to {to}.");
-                }
-                CopyStatus::Success => return Ok(()),
-                CopyStatus::Pending => (),
+                // The copy is taking longer. Waiting a second and then re-trying.
+                // TODO estimate time based on copy_progress and adjust time based on that
+                tokio::time::sleep(Duration::from_millis(1000)).await;
+                let properties = blob_client.get_properties().into_future().await?;
+                let Some(status) = properties.blob.properties.copy_status else {
+                    tracing::warn!("copy_status for copy is None!, from={from}, to={to}");
+                    return Ok(());
+                };
+                copy_status = Some(status);
             }
-            // The copy is taking longer. Waiting a second and then re-trying.
-            // TODO estimate time based on copy_progress and adjust time based on that
-            tokio::time::sleep(Duration::from_millis(1000)).await;
-            let properties = blob_client.get_properties().into_future().await?;
-            let Some(status) = properties.blob.properties.copy_status else {
-                tracing::warn!("copy_status for copy is None!, from={from}, to={to}");
-                return Ok(());
-            };
-            if start_time.elapsed() > MAX_WAIT_TIME {
-                anyhow::bail!("Copy from from {from} to {to} took longer than limit MAX_WAIT_TIME={}s. copy_pogress={:?}.",
-                    MAX_WAIT_TIME.as_secs_f32(),
-                    properties.blob.properties.copy_progress,
-                );
-            }
-            copy_status = status;
+        };
+
+        tokio::select! {
+            res = op => res,
+            _ = cancel.cancelled() => Err(anyhow::Error::new(TimeoutOrCancel::Cancel)),
+            _ = timeout => {
+                let e = anyhow::Error::new(TimeoutOrCancel::Timeout);
+                let e = e.context(format!("Timeout, last status: {copy_status:?}"));
+                Err(e)
+            },
         }
     }
 

libs/remote_storage/src/error.rs

@@ -0,0 +1,181 @@
+/// Reasons for downloads or listings to fail.
+#[derive(Debug)]
+pub enum DownloadError {
+    /// Validation or other error happened due to user input.
+    BadInput(anyhow::Error),
+    /// The file was not found in the remote storage.
+    NotFound,
+    /// A cancellation token aborted the download, typically during
+    /// tenant detach or process shutdown.
+    Cancelled,
+    /// A timeout happened while executing the request. Possible reasons:
+    /// - stuck tcp connection
+    ///
+    /// Concurrency control is not timed within timeout.
+    Timeout,
+    /// The file was found in the remote storage, but the download failed.
+    Other(anyhow::Error),
+}
+
+impl std::fmt::Display for DownloadError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            DownloadError::BadInput(e) => {
+                write!(f, "Failed to download a remote file due to user input: {e}")
+            }
+            DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
+            DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
+            DownloadError::Timeout => write!(f, "timeout"),
+            DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
+        }
+    }
+}
+
+impl std::error::Error for DownloadError {}
+
+impl DownloadError {
+    /// Returns true if the error should not be retried with backoff
+    pub fn is_permanent(&self) -> bool {
+        use DownloadError::*;
+        match self {
+            BadInput(_) | NotFound | Cancelled => true,
+            Timeout | Other(_) => false,
+        }
+    }
+}
+
+#[derive(Debug)]
+pub enum TimeTravelError {
+    /// Validation or other error happened due to user input.
+    BadInput(anyhow::Error),
+    /// The used remote storage does not have time travel recovery implemented
+    Unimplemented,
+    /// The number of versions/deletion markers is above our limit.
+    TooManyVersions,
+    /// A cancellation token aborted the process, typically during
+    /// request closure or process shutdown.
+    Cancelled,
+    /// Other errors
+    Other(anyhow::Error),
+}
+
+impl std::fmt::Display for TimeTravelError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            TimeTravelError::BadInput(e) => {
+                write!(
+                    f,
+                    "Failed to time travel recover a prefix due to user input: {e}"
+                )
+            }
+            TimeTravelError::Unimplemented => write!(
+                f,
+                "time travel recovery is not implemented for the current storage backend"
+            ),
+            TimeTravelError::Cancelled => write!(f, "Cancelled, shutting down"),
+            TimeTravelError::TooManyVersions => {
+                write!(f, "Number of versions/delete markers above limit")
+            }
+            TimeTravelError::Other(e) => write!(f, "Failed to time travel recover a prefix: {e:?}"),
+        }
+    }
+}
+
+impl std::error::Error for TimeTravelError {}
+
+/// Plain cancelled error.
+///
+/// By design this type does not not implement `std::error::Error` so it cannot be put as the root
+/// cause of `std::io::Error` or `anyhow::Error`. It should never need to be exposed out of this
+/// crate.
+///
+/// It exists to implement permit acquiring in `{Download,TimeTravel}Error` and `anyhow::Error` returning
+/// operations and ensuring that those get converted to proper versions with just `?`.
+#[derive(Debug)]
+pub(crate) struct Cancelled;
+
+impl From<Cancelled> for anyhow::Error {
+    fn from(_: Cancelled) -> Self {
+        anyhow::Error::new(TimeoutOrCancel::Cancel)
+    }
+}
+
+impl From<Cancelled> for TimeTravelError {
+    fn from(_: Cancelled) -> Self {
+        TimeTravelError::Cancelled
+    }
+}
+
+impl From<Cancelled> for TimeoutOrCancel {
+    fn from(_: Cancelled) -> Self {
+        TimeoutOrCancel::Cancel
+    }
+}
+
+impl From<Cancelled> for DownloadError {
+    fn from(_: Cancelled) -> Self {
+        DownloadError::Cancelled
+    }
+}
+
+/// This type is used at as the root cause for timeouts and cancellations with `anyhow::Error` returning
+/// RemoteStorage methods.
+///
+/// For use with `utils::backoff::retry` and `anyhow::Error` returning operations there is
+/// `TimeoutOrCancel::caused_by_cancel` method to query "proper form" errors.
+#[derive(Debug)]
+pub enum TimeoutOrCancel {
+    Timeout,
+    Cancel,
+}
+
+impl std::fmt::Display for TimeoutOrCancel {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        use TimeoutOrCancel::*;
+        match self {
+            Timeout => write!(f, "timeout"),
+            Cancel => write!(f, "cancel"),
+        }
+    }
+}
+
+impl std::error::Error for TimeoutOrCancel {}
+
+impl TimeoutOrCancel {
+    pub fn caused(error: &anyhow::Error) -> Option<&Self> {
+        error.root_cause().downcast_ref()
+    }
+
+    /// Returns true if the error was caused by [`TimeoutOrCancel::Cancel`].
+    pub fn caused_by_cancel(error: &anyhow::Error) -> bool {
+        Self::caused(error).is_some_and(Self::is_cancel)
+    }
+
+    pub fn is_cancel(&self) -> bool {
+        matches!(self, TimeoutOrCancel::Cancel)
+    }
+
+    pub fn is_timeout(&self) -> bool {
+        matches!(self, TimeoutOrCancel::Timeout)
+    }
+}
+
+/// This conversion is used when [`crate::support::DownloadStream`] notices a cancellation or
+/// timeout to wrap it in an `std::io::Error`.
+impl From<TimeoutOrCancel> for std::io::Error {
+    fn from(value: TimeoutOrCancel) -> Self {
+        let e = DownloadError::from(value);
+        std::io::Error::other(e)
+    }
+}
+
+impl From<TimeoutOrCancel> for DownloadError {
+    fn from(value: TimeoutOrCancel) -> Self {
+        use TimeoutOrCancel::*;
+
+        match value {
+            Timeout => DownloadError::Timeout,
+            Cancel => DownloadError::Cancelled,
+        }
+    }
+}

libs/remote_storage/src/lib.rs

@@ -10,12 +10,19 @@
 #![deny(clippy::undocumented_unsafe_blocks)]
 
 mod azure_blob;
+mod error;
 mod local_fs;
 mod s3_bucket;
 mod simulate_failures;
+mod support;
 
 use std::{
-    collections::HashMap, fmt::Debug, num::NonZeroUsize, pin::Pin, sync::Arc, time::SystemTime,
+    collections::HashMap,
+    fmt::Debug,
+    num::{NonZeroU32, NonZeroUsize},
+    pin::Pin,
+    sync::Arc,
+    time::{Duration, SystemTime},
 };
 
 use anyhow::{bail, Context};
@@ -35,6 +42,8 @@ pub use self::{
 };
 use s3_bucket::RequestKind;
 
+pub use error::{DownloadError, TimeTravelError, TimeoutOrCancel};
+
 /// Currently, sync happens with AWS S3, that has two limits on requests per second:
 /// ~200 RPS for IAM services
 /// <https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/UsingWithRDS.IAMDBAuth.html>
@@ -152,9 +161,10 @@ pub trait RemoteStorage: Send + Sync + 'static {
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         let result = self
-            .list(prefix, ListingMode::WithDelimiter)
+            .list(prefix, ListingMode::WithDelimiter, None, cancel)
             .await?
             .prefixes;
         Ok(result)
@@ -170,8 +180,18 @@ pub trait RemoteStorage: Send + Sync + 'static {
     /// whereas,
     /// list_prefixes("foo/bar/") = ["cat", "dog"]
     /// See `test_real_s3.rs` for more details.
-    async fn list_files(&self, prefix: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
-        let result = self.list(prefix, ListingMode::NoDelimiter).await?.keys;
+    ///
+    /// max_keys limits max number of keys returned; None means unlimited.
+    async fn list_files(
+        &self,
+        prefix: Option<&RemotePath>,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
+    ) -> Result<Vec<RemotePath>, DownloadError> {
+        let result = self
+            .list(prefix, ListingMode::NoDelimiter, max_keys, cancel)
+            .await?
+            .keys;
         Ok(result)
     }
 
@@ -179,9 +199,14 @@ pub trait RemoteStorage: Send + Sync + 'static {
         &self,
         prefix: Option<&RemotePath>,
         _mode: ListingMode,
-    ) -> anyhow::Result<Listing, DownloadError>;
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
+    ) -> Result<Listing, DownloadError>;
 
     /// Streams the local file contents into remote into the remote storage entry.
+    ///
+    /// If the operation fails because of timeout or cancellation, the root cause of the error will be
+    /// set to `TimeoutOrCancel`.
     async fn upload(
         &self,
         from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
@@ -190,27 +215,61 @@ pub trait RemoteStorage: Send + Sync + 'static {
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()>;
 
-    /// Streams the remote storage entry contents into the buffered writer given, returns the filled writer.
+    /// Streams the remote storage entry contents.
+    ///
+    /// The returned download stream will obey initial timeout and cancellation signal by erroring
+    /// on whichever happens first. Only one of the reasons will fail the stream, which is usually
+    /// enough for `tokio::io::copy_buf` usage. If needed the error can be filtered out.
+    ///
     /// Returns the metadata, if any was stored with the file previously.
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError>;
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError>;
 
-    /// Streams a given byte range of the remote storage entry contents into the buffered writer given, returns the filled writer.
+    /// Streams a given byte range of the remote storage entry contents.
+    ///
+    /// The returned download stream will obey initial timeout and cancellation signal by erroring
+    /// on whichever happens first. Only one of the reasons will fail the stream, which is usually
+    /// enough for `tokio::io::copy_buf` usage. If needed the error can be filtered out.
+    ///
     /// Returns the metadata, if any was stored with the file previously.
     async fn download_byte_range(
         &self,
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError>;
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()>;
+    /// Delete a single path from remote storage.
+    ///
+    /// If the operation fails because of timeout or cancellation, the root cause of the error will be
+    /// set to `TimeoutOrCancel`. In such situation it is unknown if the deletion went through.
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()>;
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()>;
+    /// Delete a multiple paths from remote storage.
+    ///
+    /// If the operation fails because of timeout or cancellation, the root cause of the error will be
+    /// set to `TimeoutOrCancel`. In such situation it is unknown which deletions, if any, went
+    /// through.
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()>;
 
     /// Copy a remote object inside a bucket from one path to another.
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()>;
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()>;
 
     /// Resets the content of everything with the given prefix to the given state
     async fn time_travel_recover(
@@ -222,7 +281,13 @@ pub trait RemoteStorage: Send + Sync + 'static {
     ) -> Result<(), TimeTravelError>;
 }
 
-pub type DownloadStream = Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Unpin + Send + Sync>>;
+/// DownloadStream is sensitive to the timeout and cancellation used with the original
+/// [`RemoteStorage::download`] request. The type yields `std::io::Result<Bytes>` to be compatible
+/// with `tokio::io::copy_buf`.
+// This has 'static because safekeepers do not use cancellation tokens (yet)
+pub type DownloadStream =
+    Pin<Box<dyn Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static>>;
+
 pub struct Download {
     pub download_stream: DownloadStream,
     /// The last time the file was modified (`last-modified` HTTP header)
@@ -241,73 +306,6 @@ impl Debug for Download {
     }
 }
 
-#[derive(Debug)]
-pub enum DownloadError {
-    /// Validation or other error happened due to user input.
-    BadInput(anyhow::Error),
-    /// The file was not found in the remote storage.
-    NotFound,
-    /// A cancellation token aborted the download, typically during
-    /// tenant detach or process shutdown.
-    Cancelled,
-    /// The file was found in the remote storage, but the download failed.
-    Other(anyhow::Error),
-}
-
-impl std::fmt::Display for DownloadError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            DownloadError::BadInput(e) => {
-                write!(f, "Failed to download a remote file due to user input: {e}")
-            }
-            DownloadError::Cancelled => write!(f, "Cancelled, shutting down"),
-            DownloadError::NotFound => write!(f, "No file found for the remote object id given"),
-            DownloadError::Other(e) => write!(f, "Failed to download a remote file: {e:?}"),
-        }
-    }
-}
-
-impl std::error::Error for DownloadError {}
-
-#[derive(Debug)]
-pub enum TimeTravelError {
-    /// Validation or other error happened due to user input.
-    BadInput(anyhow::Error),
-    /// The used remote storage does not have time travel recovery implemented
-    Unimplemented,
-    /// The number of versions/deletion markers is above our limit.
-    TooManyVersions,
-    /// A cancellation token aborted the process, typically during
-    /// request closure or process shutdown.
-    Cancelled,
-    /// Other errors
-    Other(anyhow::Error),
-}
-
-impl std::fmt::Display for TimeTravelError {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        match self {
-            TimeTravelError::BadInput(e) => {
-                write!(
-                    f,
-                    "Failed to time travel recover a prefix due to user input: {e}"
-                )
-            }
-            TimeTravelError::Unimplemented => write!(
-                f,
-                "time travel recovery is not implemented for the current storage backend"
-            ),
-            TimeTravelError::Cancelled => write!(f, "Cancelled, shutting down"),
-            TimeTravelError::TooManyVersions => {
-                write!(f, "Number of versions/delete markers above limit")
-            }
-            TimeTravelError::Other(e) => write!(f, "Failed to time travel recover a prefix: {e:?}"),
-        }
-    }
-}
-
-impl std::error::Error for TimeTravelError {}
-
 /// Every storage, currently supported.
 /// Serves as a simple way to pass around the [`RemoteStorage`] without dealing with generics.
 #[derive(Clone)]
@@ -324,24 +322,33 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         &self,
         prefix: Option<&RemotePath>,
         mode: ListingMode,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<Listing, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.list(prefix, mode).await,
-            Self::AwsS3(s) => s.list(prefix, mode).await,
-            Self::AzureBlob(s) => s.list(prefix, mode).await,
-            Self::Unreliable(s) => s.list(prefix, mode).await,
+            Self::LocalFs(s) => s.list(prefix, mode, max_keys, cancel).await,
+            Self::AwsS3(s) => s.list(prefix, mode, max_keys, cancel).await,
+            Self::AzureBlob(s) => s.list(prefix, mode, max_keys, cancel).await,
+            Self::Unreliable(s) => s.list(prefix, mode, max_keys, cancel).await,
         }
     }
 
     // A function for listing all the files in a "directory"
     // Example:
     // list_files("foo/bar") = ["foo/bar/a.txt", "foo/bar/b.txt"]
-    pub async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
+    //
+    // max_keys limits max number of keys returned; None means unlimited.
+    pub async fn list_files(
+        &self,
+        folder: Option<&RemotePath>,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
+    ) -> Result<Vec<RemotePath>, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.list_files(folder).await,
-            Self::AwsS3(s) => s.list_files(folder).await,
-            Self::AzureBlob(s) => s.list_files(folder).await,
-            Self::Unreliable(s) => s.list_files(folder).await,
+            Self::LocalFs(s) => s.list_files(folder, max_keys, cancel).await,
+            Self::AwsS3(s) => s.list_files(folder, max_keys, cancel).await,
+            Self::AzureBlob(s) => s.list_files(folder, max_keys, cancel).await,
+            Self::Unreliable(s) => s.list_files(folder, max_keys, cancel).await,
         }
     }
 
@@ -351,36 +358,43 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
     pub async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.list_prefixes(prefix).await,
-            Self::AwsS3(s) => s.list_prefixes(prefix).await,
-            Self::AzureBlob(s) => s.list_prefixes(prefix).await,
-            Self::Unreliable(s) => s.list_prefixes(prefix).await,
+            Self::LocalFs(s) => s.list_prefixes(prefix, cancel).await,
+            Self::AwsS3(s) => s.list_prefixes(prefix, cancel).await,
+            Self::AzureBlob(s) => s.list_prefixes(prefix, cancel).await,
+            Self::Unreliable(s) => s.list_prefixes(prefix, cancel).await,
         }
     }
 
+    /// See [`RemoteStorage::upload`]
     pub async fn upload(
         &self,
         from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata).await,
-            Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata).await,
-            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata).await,
-            Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata).await,
+            Self::LocalFs(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
+            Self::AwsS3(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
+            Self::AzureBlob(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
+            Self::Unreliable(s) => s.upload(from, data_size_bytes, to, metadata, cancel).await,
         }
     }
 
-    pub async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+    pub async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         match self {
-            Self::LocalFs(s) => s.download(from).await,
-            Self::AwsS3(s) => s.download(from).await,
-            Self::AzureBlob(s) => s.download(from).await,
-            Self::Unreliable(s) => s.download(from).await,
+            Self::LocalFs(s) => s.download(from, cancel).await,
+            Self::AwsS3(s) => s.download(from, cancel).await,
+            Self::AzureBlob(s) => s.download(from, cancel).await,
+            Self::Unreliable(s) => s.download(from, cancel).await,
         }
     }
 
@@ -389,54 +403,72 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         match self {
             Self::LocalFs(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
             Self::AwsS3(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
             Self::AzureBlob(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
             Self::Unreliable(s) => {
-                s.download_byte_range(from, start_inclusive, end_exclusive)
+                s.download_byte_range(from, start_inclusive, end_exclusive, cancel)
                     .await
             }
         }
     }
 
-    pub async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+    /// See [`RemoteStorage::delete`]
+    pub async fn delete(
+        &self,
+        path: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.delete(path).await,
-            Self::AwsS3(s) => s.delete(path).await,
-            Self::AzureBlob(s) => s.delete(path).await,
-            Self::Unreliable(s) => s.delete(path).await,
+            Self::LocalFs(s) => s.delete(path, cancel).await,
+            Self::AwsS3(s) => s.delete(path, cancel).await,
+            Self::AzureBlob(s) => s.delete(path, cancel).await,
+            Self::Unreliable(s) => s.delete(path, cancel).await,
         }
     }
 
-    pub async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+    /// See [`RemoteStorage::delete_objects`]
+    pub async fn delete_objects(
+        &self,
+        paths: &[RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.delete_objects(paths).await,
-            Self::AwsS3(s) => s.delete_objects(paths).await,
-            Self::AzureBlob(s) => s.delete_objects(paths).await,
-            Self::Unreliable(s) => s.delete_objects(paths).await,
+            Self::LocalFs(s) => s.delete_objects(paths, cancel).await,
+            Self::AwsS3(s) => s.delete_objects(paths, cancel).await,
+            Self::AzureBlob(s) => s.delete_objects(paths, cancel).await,
+            Self::Unreliable(s) => s.delete_objects(paths, cancel).await,
         }
     }
 
-    pub async fn copy_object(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    /// See [`RemoteStorage::copy`]
+    pub async fn copy_object(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         match self {
-            Self::LocalFs(s) => s.copy(from, to).await,
-            Self::AwsS3(s) => s.copy(from, to).await,
-            Self::AzureBlob(s) => s.copy(from, to).await,
-            Self::Unreliable(s) => s.copy(from, to).await,
+            Self::LocalFs(s) => s.copy(from, to, cancel).await,
+            Self::AwsS3(s) => s.copy(from, to, cancel).await,
+            Self::AzureBlob(s) => s.copy(from, to, cancel).await,
+            Self::Unreliable(s) => s.copy(from, to, cancel).await,
         }
     }
 
+    /// See [`RemoteStorage::time_travel_recover`].
     pub async fn time_travel_recover(
         &self,
         prefix: Option<&RemotePath>,
@@ -467,10 +499,11 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
 
 impl GenericRemoteStorage {
     pub fn from_config(storage_config: &RemoteStorageConfig) -> anyhow::Result<Self> {
+        let timeout = storage_config.timeout;
         Ok(match &storage_config.storage {
-            RemoteStorageKind::LocalFs(root) => {
-                info!("Using fs root '{root}' as a remote storage");
-                Self::LocalFs(LocalFs::new(root.clone())?)
+            RemoteStorageKind::LocalFs(path) => {
+                info!("Using fs root '{path}' as a remote storage");
+                Self::LocalFs(LocalFs::new(path.clone(), timeout)?)
             }
             RemoteStorageKind::AwsS3(s3_config) => {
                 // The profile and access key id are only printed here for debugging purposes,
@@ -480,12 +513,12 @@ impl GenericRemoteStorage {
                     std::env::var("AWS_ACCESS_KEY_ID").unwrap_or_else(|_| "<none>".into());
                 info!("Using s3 bucket '{}' in region '{}' as a remote storage, prefix in bucket: '{:?}', bucket endpoint: '{:?}', profile: {profile}, access_key_id: {access_key_id}",
                       s3_config.bucket_name, s3_config.bucket_region, s3_config.prefix_in_bucket, s3_config.endpoint);
-                Self::AwsS3(Arc::new(S3Bucket::new(s3_config)?))
+                Self::AwsS3(Arc::new(S3Bucket::new(s3_config, timeout)?))
             }
             RemoteStorageKind::AzureContainer(azure_config) => {
                 info!("Using azure container '{}' in region '{}' as a remote storage, prefix in container: '{:?}'",
                       azure_config.container_name, azure_config.container_region, azure_config.prefix_in_container);
-                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config)?))
+                Self::AzureBlob(Arc::new(AzureBlobStorage::new(azure_config, timeout)?))
             }
         })
     }
@@ -494,18 +527,15 @@ impl GenericRemoteStorage {
         Self::Unreliable(Arc::new(UnreliableWrapper::new(s, fail_first)))
     }
 
-    /// Takes storage object contents and its size and uploads to remote storage,
-    /// mapping `from_path` to the corresponding remote object id in the storage.
-    ///
-    /// The storage object does not have to be present on the `from_path`,
-    /// this path is used for the remote object id conversion only.
+    /// See [`RemoteStorage::upload`], which this method calls with `None` as metadata.
     pub async fn upload_storage_object(
         &self,
         from: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync + 'static,
         from_size_bytes: usize,
         to: &RemotePath,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
-        self.upload(from, from_size_bytes, to, None)
+        self.upload(from, from_size_bytes, to, None, cancel)
             .await
             .with_context(|| {
                 format!("Failed to upload data of length {from_size_bytes} to storage path {to:?}")
@@ -518,10 +548,11 @@ impl GenericRemoteStorage {
         &self,
         byte_range: Option<(u64, Option<u64>)>,
         from: &RemotePath,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         match byte_range {
-            Some((start, end)) => self.download_byte_range(from, start, end).await,
-            None => self.download(from).await,
+            Some((start, end)) => self.download_byte_range(from, start, end, cancel).await,
+            None => self.download(from, cancel).await,
         }
     }
 }
@@ -536,6 +567,9 @@ pub struct StorageMetadata(HashMap<String, String>);
 pub struct RemoteStorageConfig {
     /// The storage connection configuration.
     pub storage: RemoteStorageKind,
+    /// A common timeout enforced for all requests after concurrency limiter permit has been
+    /// acquired.
+    pub timeout: Duration,
 }
 
 /// A kind of a remote storage to connect to, with its connection configuration.
@@ -620,6 +654,8 @@ impl Debug for AzureConfig {
 }
 
 impl RemoteStorageConfig {
+    pub const DEFAULT_TIMEOUT: Duration = std::time::Duration::from_secs(120);
+
     pub fn from_toml(toml: &toml_edit::Item) -> anyhow::Result<Option<RemoteStorageConfig>> {
         let local_path = toml.get("local_path");
         let bucket_name = toml.get("bucket_name");
@@ -649,6 +685,27 @@ impl RemoteStorageConfig {
             .map(|endpoint| parse_toml_string("endpoint", endpoint))
             .transpose()?;
 
+        let timeout = toml
+            .get("timeout")
+            .map(|timeout| {
+                timeout
+                    .as_str()
+                    .ok_or_else(|| anyhow::Error::msg("timeout was not a string"))
+            })
+            .transpose()
+            .and_then(|timeout| {
+                timeout
+                    .map(humantime::parse_duration)
+                    .transpose()
+                    .map_err(anyhow::Error::new)
+            })
+            .context("parse timeout")?
+            .unwrap_or(Self::DEFAULT_TIMEOUT);
+
+        if timeout < Duration::from_secs(1) {
+            bail!("timeout was specified as {timeout:?} which is too low");
+        }
+
         let storage = match (
             local_path,
             bucket_name,
@@ -710,7 +767,7 @@ impl RemoteStorageConfig {
             }
         };
 
-        Ok(Some(RemoteStorageConfig { storage }))
+        Ok(Some(RemoteStorageConfig { storage, timeout }))
     }
 }
 
@@ -806,4 +863,24 @@ mod tests {
         let err = RemotePath::new(Utf8Path::new("/")).expect_err("Should fail on absolute paths");
         assert_eq!(err.to_string(), "Path \"/\" is not relative");
     }
+
+    #[test]
+    fn parse_localfs_config_with_timeout() {
+        let input = "local_path = '.'
+timeout = '5s'";
+
+        let toml = input.parse::<toml_edit::Document>().unwrap();
+
+        let config = RemoteStorageConfig::from_toml(toml.as_item())
+            .unwrap()
+            .expect("it exists");
+
+        assert_eq!(
+            config,
+            RemoteStorageConfig {
+                storage: RemoteStorageKind::LocalFs(Utf8PathBuf::from(".")),
+                timeout: Duration::from_secs(5)
+            }
+        );
+    }
 }

libs/remote_storage/src/local_fs.rs

@@ -4,7 +4,14 @@
 //! This storage used in tests, but can also be used in cases when a certain persistent
 //! volume is mounted to the local FS.
 
-use std::{borrow::Cow, future::Future, io::ErrorKind, pin::Pin, time::SystemTime};
+use std::{
+    borrow::Cow,
+    future::Future,
+    io::ErrorKind,
+    num::NonZeroU32,
+    pin::Pin,
+    time::{Duration, SystemTime},
+};
 
 use anyhow::{bail, ensure, Context};
 use bytes::Bytes;
@@ -19,7 +26,7 @@ use tracing::*;
 use utils::{crashsafe::path_with_suffix_extension, fs_ext::is_directory_empty};
 
 use crate::{
-    Download, DownloadError, DownloadStream, Listing, ListingMode, RemotePath, TimeTravelError,
+    Download, DownloadError, Listing, ListingMode, RemotePath, TimeTravelError, TimeoutOrCancel,
 };
 
 use super::{RemoteStorage, StorageMetadata};
@@ -29,12 +36,13 @@ const LOCAL_FS_TEMP_FILE_SUFFIX: &str = "___temp";
 #[derive(Debug, Clone)]
 pub struct LocalFs {
     storage_root: Utf8PathBuf,
+    timeout: Duration,
 }
 
 impl LocalFs {
     /// Attempts to create local FS storage, along with its root directory.
     /// Storage root will be created (if does not exist) and transformed into an absolute path (if passed as relative).
-    pub fn new(mut storage_root: Utf8PathBuf) -> anyhow::Result<Self> {
+    pub fn new(mut storage_root: Utf8PathBuf, timeout: Duration) -> anyhow::Result<Self> {
         if !storage_root.exists() {
             std::fs::create_dir_all(&storage_root).with_context(|| {
                 format!("Failed to create all directories in the given root path {storage_root:?}")
@@ -46,7 +54,10 @@ impl LocalFs {
             })?;
         }
 
-        Ok(Self { storage_root })
+        Ok(Self {
+            storage_root,
+            timeout,
+        })
     }
 
     // mirrors S3Bucket::s3_object_to_relative_path
@@ -157,76 +168,14 @@ impl LocalFs {
 
         Ok(files)
     }
-}
 
-impl RemoteStorage for LocalFs {
-    async fn list(
-        &self,
-        prefix: Option<&RemotePath>,
-        mode: ListingMode,
-    ) -> Result<Listing, DownloadError> {
-        let mut result = Listing::default();
-
-        if let ListingMode::NoDelimiter = mode {
-            let keys = self
-                .list_recursive(prefix)
-                .await
-                .map_err(DownloadError::Other)?;
-
-            result.keys = keys
-                .into_iter()
-                .filter(|k| {
-                    let path = k.with_base(&self.storage_root);
-                    !path.is_dir()
-                })
-                .collect();
-
-            return Ok(result);
-        }
-
-        let path = match prefix {
-            Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
-            None => Cow::Borrowed(&self.storage_root),
-        };
-
-        let prefixes_to_filter = get_all_files(path.as_ref(), false)
-            .await
-            .map_err(DownloadError::Other)?;
-
-        // filter out empty directories to mirror s3 behavior.
-        for prefix in prefixes_to_filter {
-            if prefix.is_dir()
-                && is_directory_empty(&prefix)
-                    .await
-                    .map_err(DownloadError::Other)?
-            {
-                continue;
-            }
-
-            let stripped = prefix
-                .strip_prefix(&self.storage_root)
-                .context("Failed to strip prefix")
-                .and_then(RemotePath::new)
-                .expect(
-                    "We list files for storage root, hence should be able to remote the prefix",
-                );
-
-            if prefix.is_dir() {
-                result.prefixes.push(stripped);
-            } else {
-                result.keys.push(stripped);
-            }
-        }
-
-        Ok(result)
-    }
-
-    async fn upload(
+    async fn upload0(
         &self,
         data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         let target_file_path = to.with_base(&self.storage_root);
         create_target_directory(&target_file_path).await?;
@@ -261,9 +210,26 @@ impl RemoteStorage for LocalFs {
         let mut buffer_to_read = data.take(from_size_bytes);
 
         // alternatively we could just write the bytes to a file, but local_fs is a testing utility
-        let bytes_read = io::copy_buf(&mut buffer_to_read, &mut destination)
-            .await
-            .with_context(|| {
+        let copy = io::copy_buf(&mut buffer_to_read, &mut destination);
+
+        let bytes_read = tokio::select! {
+            biased;
+            _ = cancel.cancelled() => {
+                let file = destination.into_inner();
+                // wait for the inflight operation(s) to complete so that there could be a next
+                // attempt right away and our writes are not directed to their file.
+                file.into_std().await;
+
+                // TODO: leave the temp or not? leaving is probably less racy. enabled truncate at
+                // least.
+                fs::remove_file(temp_file_path).await.context("remove temp_file_path after cancellation or timeout")?;
+                return Err(TimeoutOrCancel::Cancel.into());
+            }
+            read = copy => read,
+        };
+
+        let bytes_read =
+            bytes_read.with_context(|| {
                 format!(
                     "Failed to upload file (write temp) to the local storage at '{temp_file_path}'",
                 )
@@ -295,6 +261,9 @@ impl RemoteStorage for LocalFs {
             })?;
 
         if let Some(storage_metadata) = metadata {
+            // FIXME: we must not be using metadata much, since this would forget the old metadata
+            // for new writes? or perhaps metadata is sticky; could consider removing if it's never
+            // used.
             let storage_metadata_path = storage_metadata_path(&target_file_path);
             fs::write(
                 &storage_metadata_path,
@@ -311,8 +280,131 @@ impl RemoteStorage for LocalFs {
 
         Ok(())
     }
+}
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+impl RemoteStorage for LocalFs {
+    async fn list(
+        &self,
+        prefix: Option<&RemotePath>,
+        mode: ListingMode,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
+    ) -> Result<Listing, DownloadError> {
+        let op = async {
+            let mut result = Listing::default();
+
+            if let ListingMode::NoDelimiter = mode {
+                let keys = self
+                    .list_recursive(prefix)
+                    .await
+                    .map_err(DownloadError::Other)?;
+
+                result.keys = keys
+                    .into_iter()
+                    .filter(|k| {
+                        let path = k.with_base(&self.storage_root);
+                        !path.is_dir()
+                    })
+                    .collect();
+
+                if let Some(max_keys) = max_keys {
+                    result.keys.truncate(max_keys.get() as usize);
+                }
+
+                return Ok(result);
+            }
+
+            let path = match prefix {
+                Some(prefix) => Cow::Owned(prefix.with_base(&self.storage_root)),
+                None => Cow::Borrowed(&self.storage_root),
+            };
+
+            let prefixes_to_filter = get_all_files(path.as_ref(), false)
+                .await
+                .map_err(DownloadError::Other)?;
+
+            // filter out empty directories to mirror s3 behavior.
+            for prefix in prefixes_to_filter {
+                if prefix.is_dir()
+                    && is_directory_empty(&prefix)
+                        .await
+                        .map_err(DownloadError::Other)?
+                {
+                    continue;
+                }
+
+                let stripped = prefix
+                    .strip_prefix(&self.storage_root)
+                    .context("Failed to strip prefix")
+                    .and_then(RemotePath::new)
+                    .expect(
+                        "We list files for storage root, hence should be able to remote the prefix",
+                    );
+
+                if prefix.is_dir() {
+                    result.prefixes.push(stripped);
+                } else {
+                    result.keys.push(stripped);
+                }
+            }
+
+            Ok(result)
+        };
+
+        let timeout = async {
+            tokio::time::sleep(self.timeout).await;
+            Err(DownloadError::Timeout)
+        };
+
+        let cancelled = async {
+            cancel.cancelled().await;
+            Err(DownloadError::Cancelled)
+        };
+
+        tokio::select! {
+            res = op => res,
+            res = timeout => res,
+            res = cancelled => res,
+        }
+    }
+
+    async fn upload(
+        &self,
+        data: impl Stream<Item = std::io::Result<Bytes>> + Send + Sync,
+        data_size_bytes: usize,
+        to: &RemotePath,
+        metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let cancel = cancel.child_token();
+
+        let op = self.upload0(data, data_size_bytes, to, metadata, &cancel);
+        let mut op = std::pin::pin!(op);
+
+        // race the upload0 to the timeout; if it goes over, do a graceful shutdown
+        let (res, timeout) = tokio::select! {
+            res = &mut op => (res, false),
+            _ = tokio::time::sleep(self.timeout) => {
+                cancel.cancel();
+                (op.await, true)
+            }
+        };
+
+        match res {
+            Err(e) if timeout && TimeoutOrCancel::caused_by_cancel(&e) => {
+                // we caused this cancel (or they happened simultaneously) -- swap it out to
+                // Timeout
+                Err(TimeoutOrCancel::Timeout.into())
+            }
+            res => res,
+        }
+    }
+
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         let target_path = from.with_base(&self.storage_root);
         if file_exists(&target_path).map_err(DownloadError::BadInput)? {
             let source = ReaderStream::new(
@@ -330,6 +422,10 @@ impl RemoteStorage for LocalFs {
                 .read_storage_metadata(&target_path)
                 .await
                 .map_err(DownloadError::Other)?;
+
+            let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
+            let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
+
             Ok(Download {
                 metadata,
                 last_modified: None,
@@ -346,6 +442,7 @@ impl RemoteStorage for LocalFs {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         if let Some(end_exclusive) = end_exclusive {
             if end_exclusive <= start_inclusive {
@@ -365,34 +462,43 @@ impl RemoteStorage for LocalFs {
                     format!("Failed to open source file {target_path:?} to use in the download")
                 })
                 .map_err(DownloadError::Other)?;
+
+            let len = source
+                .metadata()
+                .await
+                .context("query file length")
+                .map_err(DownloadError::Other)?
+                .len();
+
             source
                 .seek(io::SeekFrom::Start(start_inclusive))
                 .await
                 .context("Failed to seek to the range start in a local storage file")
                 .map_err(DownloadError::Other)?;
+
             let metadata = self
                 .read_storage_metadata(&target_path)
                 .await
                 .map_err(DownloadError::Other)?;
 
-            let download_stream: DownloadStream = match end_exclusive {
-                Some(end_exclusive) => Box::pin(ReaderStream::new(
-                    source.take(end_exclusive - start_inclusive),
-                )),
-                None => Box::pin(ReaderStream::new(source)),
-            };
+            let source = source.take(end_exclusive.unwrap_or(len) - start_inclusive);
+            let source = ReaderStream::new(source);
+
+            let cancel_or_timeout = crate::support::cancel_or_timeout(self.timeout, cancel.clone());
+            let source = crate::support::DownloadStream::new(cancel_or_timeout, source);
+
             Ok(Download {
                 metadata,
                 last_modified: None,
                 etag: None,
-                download_stream,
+                download_stream: Box::pin(source),
             })
         } else {
             Err(DownloadError::NotFound)
         }
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+    async fn delete(&self, path: &RemotePath, _cancel: &CancellationToken) -> anyhow::Result<()> {
         let file_path = path.with_base(&self.storage_root);
         match fs::remove_file(&file_path).await {
             Ok(()) => Ok(()),
@@ -404,14 +510,23 @@ impl RemoteStorage for LocalFs {
         }
     }
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         for path in paths {
-            self.delete(path).await?
+            self.delete(path, cancel).await?
         }
         Ok(())
     }
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        _cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         let from_path = from.with_base(&self.storage_root);
         let to_path = to.with_base(&self.storage_root);
         create_target_directory(&to_path).await?;
@@ -425,7 +540,6 @@ impl RemoteStorage for LocalFs {
         Ok(())
     }
 
-    #[allow(clippy::diverging_sub_expression)]
     async fn time_travel_recover(
         &self,
         _prefix: Option<&RemotePath>,
@@ -514,15 +628,14 @@ mod fs_tests {
     use futures_util::Stream;
     use std::{collections::HashMap, io::Write};
 
-    async fn read_and_assert_remote_file_contents(
+    async fn read_and_check_metadata(
         storage: &LocalFs,
-        #[allow(clippy::ptr_arg)]
-        // have to use &Utf8PathBuf due to `storage.local_path` parameter requirements
         remote_storage_path: &RemotePath,
         expected_metadata: Option<&StorageMetadata>,
     ) -> anyhow::Result<String> {
+        let cancel = CancellationToken::new();
         let download = storage
-            .download(remote_storage_path)
+            .download(remote_storage_path, &cancel)
             .await
             .map_err(|e| anyhow::anyhow!("Download failed: {e}"))?;
         ensure!(
@@ -537,16 +650,16 @@ mod fs_tests {
 
     #[tokio::test]
     async fn upload_file() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
 
-        let target_path_1 = upload_dummy_file(&storage, "upload_1", None).await?;
+        let target_path_1 = upload_dummy_file(&storage, "upload_1", None, &cancel).await?;
         assert_eq!(
             storage.list_all().await?,
             vec![target_path_1.clone()],
             "Should list a single file after first upload"
         );
 
-        let target_path_2 = upload_dummy_file(&storage, "upload_2", None).await?;
+        let target_path_2 = upload_dummy_file(&storage, "upload_2", None, &cancel).await?;
         assert_eq!(
             list_files_sorted(&storage).await?,
             vec![target_path_1.clone(), target_path_2.clone()],
@@ -558,7 +671,7 @@ mod fs_tests {
 
     #[tokio::test]
     async fn upload_file_negatives() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
 
         let id = RemotePath::new(Utf8Path::new("dummy"))?;
         let content = Bytes::from_static(b"12345");
@@ -567,36 +680,36 @@ mod fs_tests {
         // Check that you get an error if the size parameter doesn't match the actual
         // size of the stream.
         storage
-            .upload(content(), 0, &id, None)
+            .upload(content(), 0, &id, None, &cancel)
             .await
             .expect_err("upload with zero size succeeded");
         storage
-            .upload(content(), 4, &id, None)
+            .upload(content(), 4, &id, None, &cancel)
             .await
             .expect_err("upload with too short size succeeded");
         storage
-            .upload(content(), 6, &id, None)
+            .upload(content(), 6, &id, None, &cancel)
             .await
             .expect_err("upload with too large size succeeded");
 
         // Correct size is 5, this should succeed.
-        storage.upload(content(), 5, &id, None).await?;
+        storage.upload(content(), 5, &id, None, &cancel).await?;
 
         Ok(())
     }
 
-    fn create_storage() -> anyhow::Result<LocalFs> {
+    fn create_storage() -> anyhow::Result<(LocalFs, CancellationToken)> {
         let storage_root = tempdir()?.path().to_path_buf();
-        LocalFs::new(storage_root)
+        LocalFs::new(storage_root, Duration::from_secs(120)).map(|s| (s, CancellationToken::new()))
     }
 
     #[tokio::test]
     async fn download_file() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
-        let contents = read_and_assert_remote_file_contents(&storage, &upload_target, None).await?;
+        let contents = read_and_check_metadata(&storage, &upload_target, None).await?;
         assert_eq!(
             dummy_contents(upload_name),
             contents,
@@ -604,7 +717,7 @@ mod fs_tests {
         );
 
         let non_existing_path = "somewhere/else";
-        match storage.download(&RemotePath::new(Utf8Path::new(non_existing_path))?).await {
+        match storage.download(&RemotePath::new(Utf8Path::new(non_existing_path))?, &cancel).await {
             Err(DownloadError::NotFound) => {} // Should get NotFound for non existing keys
             other => panic!("Should get a NotFound error when downloading non-existing storage files, but got: {other:?}"),
         }
@@ -613,12 +726,12 @@ mod fs_tests {
 
     #[tokio::test]
     async fn download_file_range_positive() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
         let full_range_download_contents =
-            read_and_assert_remote_file_contents(&storage, &upload_target, None).await?;
+            read_and_check_metadata(&storage, &upload_target, None).await?;
         assert_eq!(
             dummy_contents(upload_name),
             full_range_download_contents,
@@ -629,7 +742,12 @@ mod fs_tests {
         let (first_part_local, second_part_local) = uploaded_bytes.split_at(3);
 
         let first_part_download = storage
-            .download_byte_range(&upload_target, 0, Some(first_part_local.len() as u64))
+            .download_byte_range(
+                &upload_target,
+                0,
+                Some(first_part_local.len() as u64),
+                &cancel,
+            )
             .await?;
         assert!(
             first_part_download.metadata.is_none(),
@@ -647,6 +765,7 @@ mod fs_tests {
                 &upload_target,
                 first_part_local.len() as u64,
                 Some((first_part_local.len() + second_part_local.len()) as u64),
+                &cancel,
             )
             .await?;
         assert!(
@@ -660,14 +779,30 @@ mod fs_tests {
             "Second part bytes should be returned when requested"
         );
 
+        let suffix_bytes = storage
+            .download_byte_range(&upload_target, 13, None, &cancel)
+            .await?
+            .download_stream;
+        let suffix_bytes = aggregate(suffix_bytes).await?;
+        let suffix = std::str::from_utf8(&suffix_bytes)?;
+        assert_eq!(upload_name, suffix);
+
+        let all_bytes = storage
+            .download_byte_range(&upload_target, 0, None, &cancel)
+            .await?
+            .download_stream;
+        let all_bytes = aggregate(all_bytes).await?;
+        let all_bytes = std::str::from_utf8(&all_bytes)?;
+        assert_eq!(dummy_contents("upload_1"), all_bytes);
+
         Ok(())
     }
 
     #[tokio::test]
     async fn download_file_range_negative() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
         let start = 1_000_000_000;
         let end = start + 1;
@@ -676,6 +811,7 @@ mod fs_tests {
                 &upload_target,
                 start,
                 Some(end), // exclusive end
+                &cancel,
             )
             .await
         {
@@ -692,7 +828,7 @@ mod fs_tests {
         let end = 234;
         assert!(start > end, "Should test an incorrect range");
         match storage
-            .download_byte_range(&upload_target, start, Some(end))
+            .download_byte_range(&upload_target, start, Some(end), &cancel)
             .await
         {
             Ok(_) => panic!("Should not allow downloading wrong ranges"),
@@ -709,15 +845,15 @@ mod fs_tests {
 
     #[tokio::test]
     async fn delete_file() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
-        let upload_target = upload_dummy_file(&storage, upload_name, None).await?;
+        let upload_target = upload_dummy_file(&storage, upload_name, None, &cancel).await?;
 
-        storage.delete(&upload_target).await?;
+        storage.delete(&upload_target, &cancel).await?;
         assert!(storage.list_all().await?.is_empty());
 
         storage
-            .delete(&upload_target)
+            .delete(&upload_target, &cancel)
             .await
             .expect("Should allow deleting non-existing storage files");
 
@@ -726,17 +862,17 @@ mod fs_tests {
 
     #[tokio::test]
     async fn file_with_metadata() -> anyhow::Result<()> {
-        let storage = create_storage()?;
+        let (storage, cancel) = create_storage()?;
         let upload_name = "upload_1";
         let metadata = StorageMetadata(HashMap::from([
             ("one".to_string(), "1".to_string()),
             ("two".to_string(), "2".to_string()),
         ]));
         let upload_target =
-            upload_dummy_file(&storage, upload_name, Some(metadata.clone())).await?;
+            upload_dummy_file(&storage, upload_name, Some(metadata.clone()), &cancel).await?;
 
         let full_range_download_contents =
-            read_and_assert_remote_file_contents(&storage, &upload_target, Some(&metadata)).await?;
+            read_and_check_metadata(&storage, &upload_target, Some(&metadata)).await?;
         assert_eq!(
             dummy_contents(upload_name),
             full_range_download_contents,
@@ -747,7 +883,12 @@ mod fs_tests {
         let (first_part_local, _) = uploaded_bytes.split_at(3);
 
         let partial_download_with_metadata = storage
-            .download_byte_range(&upload_target, 0, Some(first_part_local.len() as u64))
+            .download_byte_range(
+                &upload_target,
+                0,
+                Some(first_part_local.len() as u64),
+                &cancel,
+            )
             .await?;
         let first_part_remote = aggregate(partial_download_with_metadata.download_stream).await?;
         assert_eq!(
@@ -768,16 +909,20 @@ mod fs_tests {
     #[tokio::test]
     async fn list() -> anyhow::Result<()> {
         // No delimiter: should recursively list everything
-        let storage = create_storage()?;
-        let child = upload_dummy_file(&storage, "grandparent/parent/child", None).await?;
-        let uncle = upload_dummy_file(&storage, "grandparent/uncle", None).await?;
+        let (storage, cancel) = create_storage()?;
+        let child = upload_dummy_file(&storage, "grandparent/parent/child", None, &cancel).await?;
+        let uncle = upload_dummy_file(&storage, "grandparent/uncle", None, &cancel).await?;
 
-        let listing = storage.list(None, ListingMode::NoDelimiter).await?;
+        let listing = storage
+            .list(None, ListingMode::NoDelimiter, None, &cancel)
+            .await?;
         assert!(listing.prefixes.is_empty());
         assert_eq!(listing.keys, [uncle.clone(), child.clone()].to_vec());
 
         // Delimiter: should only go one deep
-        let listing = storage.list(None, ListingMode::WithDelimiter).await?;
+        let listing = storage
+            .list(None, ListingMode::WithDelimiter, None, &cancel)
+            .await?;
 
         assert_eq!(
             listing.prefixes,
@@ -790,6 +935,8 @@ mod fs_tests {
             .list(
                 Some(&RemotePath::from_string("timelines/some_timeline/grandparent").unwrap()),
                 ListingMode::WithDelimiter,
+                None,
+                &cancel,
             )
             .await?;
         assert_eq!(
@@ -802,10 +949,75 @@ mod fs_tests {
         Ok(())
     }
 
+    #[tokio::test]
+    async fn overwrite_shorter_file() -> anyhow::Result<()> {
+        let (storage, cancel) = create_storage()?;
+
+        let path = RemotePath::new("does/not/matter/file".into())?;
+
+        let body = Bytes::from_static(b"long file contents is long");
+        {
+            let len = body.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
+            storage.upload(body, len, &path, None, &cancel).await?;
+        }
+
+        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        assert_eq!(body, read);
+
+        let shorter = Bytes::from_static(b"shorter body");
+        {
+            let len = shorter.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(shorter.clone())));
+            storage.upload(body, len, &path, None, &cancel).await?;
+        }
+
+        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        assert_eq!(shorter, read);
+        Ok(())
+    }
+
+    #[tokio::test]
+    async fn cancelled_upload_can_later_be_retried() -> anyhow::Result<()> {
+        let (storage, cancel) = create_storage()?;
+
+        let path = RemotePath::new("does/not/matter/file".into())?;
+
+        let body = Bytes::from_static(b"long file contents is long");
+        {
+            let len = body.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
+            let cancel = cancel.child_token();
+            cancel.cancel();
+            let e = storage
+                .upload(body, len, &path, None, &cancel)
+                .await
+                .unwrap_err();
+
+            assert!(TimeoutOrCancel::caused_by_cancel(&e));
+        }
+
+        {
+            let len = body.len();
+            let body =
+                futures::stream::once(futures::future::ready(std::io::Result::Ok(body.clone())));
+            storage.upload(body, len, &path, None, &cancel).await?;
+        }
+
+        let read = aggregate(storage.download(&path, &cancel).await?.download_stream).await?;
+        assert_eq!(body, read);
+
+        Ok(())
+    }
+
     async fn upload_dummy_file(
         storage: &LocalFs,
         name: &str,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<RemotePath> {
         let from_path = storage
             .storage_root
@@ -827,7 +1039,9 @@ mod fs_tests {
 
         let file = tokio_util::io::ReaderStream::new(file);
 
-        storage.upload(file, size, &relative_path, metadata).await?;
+        storage
+            .upload(file, size, &relative_path, metadata, cancel)
+            .await?;
         Ok(relative_path)
     }
 

libs/remote_storage/src/s3_bucket.rs

@@ -7,10 +7,11 @@
 use std::{
     borrow::Cow,
     collections::HashMap,
+    num::NonZeroU32,
     pin::Pin,
     sync::Arc,
     task::{Context, Poll},
-    time::SystemTime,
+    time::{Duration, SystemTime},
 };
 
 use anyhow::{anyhow, Context as _};
@@ -45,8 +46,9 @@ use utils::backoff;
 
 use super::StorageMetadata;
 use crate::{
-    ConcurrencyLimiter, Download, DownloadError, Listing, ListingMode, RemotePath, RemoteStorage,
-    S3Config, TimeTravelError, MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
+    error::Cancelled, support::PermitCarrying, ConcurrencyLimiter, Download, DownloadError,
+    Listing, ListingMode, RemotePath, RemoteStorage, S3Config, TimeTravelError, TimeoutOrCancel,
+    MAX_KEYS_PER_DELETE, REMOTE_STORAGE_PREFIX_SEPARATOR,
 };
 
 pub(super) mod metrics;
@@ -61,9 +63,10 @@ pub struct S3Bucket {
     prefix_in_bucket: Option<String>,
     max_keys_per_list_response: Option<i32>,
     concurrency_limiter: ConcurrencyLimiter,
+    // Per-request timeout. Accessible for tests.
+    pub timeout: Duration,
 }
 
-#[derive(Default)]
 struct GetObjectRequest {
     bucket: String,
     key: String,
@@ -71,7 +74,7 @@ struct GetObjectRequest {
 }
 impl S3Bucket {
     /// Creates the S3 storage, errors if incorrect AWS S3 configuration provided.
-    pub fn new(aws_config: &S3Config) -> anyhow::Result<Self> {
+    pub fn new(aws_config: &S3Config, timeout: Duration) -> anyhow::Result<Self> {
         tracing::debug!(
             "Creating s3 remote storage for S3 bucket {}",
             aws_config.bucket_name
@@ -151,6 +154,7 @@ impl S3Bucket {
             max_keys_per_list_response: aws_config.max_keys_per_list_response,
             prefix_in_bucket,
             concurrency_limiter: ConcurrencyLimiter::new(aws_config.concurrency_limit.get()),
+            timeout,
         })
     }
 
@@ -184,40 +188,55 @@ impl S3Bucket {
         }
     }
 
-    async fn permit(&self, kind: RequestKind) -> tokio::sync::SemaphorePermit<'_> {
+    async fn permit(
+        &self,
+        kind: RequestKind,
+        cancel: &CancellationToken,
+    ) -> Result<tokio::sync::SemaphorePermit<'_>, Cancelled> {
         let started_at = start_counting_cancelled_wait(kind);
-        let permit = self
-            .concurrency_limiter
-            .acquire(kind)
-            .await
-            .expect("semaphore is never closed");
+        let acquire = self.concurrency_limiter.acquire(kind);
+
+        let permit = tokio::select! {
+            permit = acquire => permit.expect("semaphore is never closed"),
+            _ = cancel.cancelled() => return Err(Cancelled),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
         metrics::BUCKET_METRICS
             .wait_seconds
             .observe_elapsed(kind, started_at);
 
-        permit
+        Ok(permit)
     }
 
-    async fn owned_permit(&self, kind: RequestKind) -> tokio::sync::OwnedSemaphorePermit {
+    async fn owned_permit(
+        &self,
+        kind: RequestKind,
+        cancel: &CancellationToken,
+    ) -> Result<tokio::sync::OwnedSemaphorePermit, Cancelled> {
         let started_at = start_counting_cancelled_wait(kind);
-        let permit = self
-            .concurrency_limiter
-            .acquire_owned(kind)
-            .await
-            .expect("semaphore is never closed");
+        let acquire = self.concurrency_limiter.acquire_owned(kind);
+
+        let permit = tokio::select! {
+            permit = acquire => permit.expect("semaphore is never closed"),
+            _ = cancel.cancelled() => return Err(Cancelled),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
         metrics::BUCKET_METRICS
             .wait_seconds
             .observe_elapsed(kind, started_at);
-        permit
+        Ok(permit)
     }
 
-    async fn download_object(&self, request: GetObjectRequest) -> Result<Download, DownloadError> {
+    async fn download_object(
+        &self,
+        request: GetObjectRequest,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         let kind = RequestKind::Get;
-        let permit = self.owned_permit(kind).await;
+
+        let permit = self.owned_permit(kind, cancel).await?;
 
         let started_at = start_measuring_requests(kind);
 
@@ -227,29 +246,18 @@ impl S3Bucket {
             .bucket(request.bucket)
             .key(request.key)
             .set_range(request.range)
-            .send()
-            .await;
+            .send();
+
+        let get_object = tokio::select! {
+            res = get_object => res,
+            _ = tokio::time::sleep(self.timeout) => return Err(DownloadError::Timeout),
+            _ = cancel.cancelled() => return Err(DownloadError::Cancelled),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
 
-        match get_object {
-            Ok(object_output) => {
-                let metadata = object_output.metadata().cloned().map(StorageMetadata);
-                let etag = object_output.e_tag.clone();
-                let last_modified = object_output.last_modified.and_then(|t| t.try_into().ok());
-
-                let body = object_output.body;
-                let body = ByteStreamAsStream::from(body);
-                let body = PermitCarrying::new(permit, body);
-                let body = TimedDownload::new(started_at, body);
-
-                Ok(Download {
-                    metadata,
-                    etag,
-                    last_modified,
-                    download_stream: Box::pin(body),
-                })
-            }
+        let object_output = match get_object {
+            Ok(object_output) => object_output,
             Err(SdkError::ServiceError(e)) if matches!(e.err(), GetObjectError::NoSuchKey(_)) => {
                 // Count this in the AttemptOutcome::Ok bucket, because 404 is not
                 // an error: we expect to sometimes fetch an object and find it missing,
@@ -259,7 +267,7 @@ impl S3Bucket {
                     AttemptOutcome::Ok,
                     started_at,
                 );
-                Err(DownloadError::NotFound)
+                return Err(DownloadError::NotFound);
             }
             Err(e) => {
                 metrics::BUCKET_METRICS.req_seconds.observe_elapsed(
@@ -268,42 +276,76 @@ impl S3Bucket {
                     started_at,
                 );
 
-                Err(DownloadError::Other(
+                return Err(DownloadError::Other(
                     anyhow::Error::new(e).context("download s3 object"),
-                ))
+                ));
             }
-        }
+        };
+
+        // even if we would have no timeout left, continue anyways. the caller can decide to ignore
+        // the errors considering timeouts and cancellation.
+        let remaining = self.timeout.saturating_sub(started_at.elapsed());
+
+        let metadata = object_output.metadata().cloned().map(StorageMetadata);
+        let etag = object_output.e_tag;
+        let last_modified = object_output.last_modified.and_then(|t| t.try_into().ok());
+
+        let body = object_output.body;
+        let body = ByteStreamAsStream::from(body);
+        let body = PermitCarrying::new(permit, body);
+        let body = TimedDownload::new(started_at, body);
+
+        let cancel_or_timeout = crate::support::cancel_or_timeout(remaining, cancel.clone());
+        let body = crate::support::DownloadStream::new(cancel_or_timeout, body);
+
+        Ok(Download {
+            metadata,
+            etag,
+            last_modified,
+            download_stream: Box::pin(body),
+        })
     }
 
     async fn delete_oids(
         &self,
-        kind: RequestKind,
+        _permit: &tokio::sync::SemaphorePermit<'_>,
         delete_objects: &[ObjectIdentifier],
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
+        let kind = RequestKind::Delete;
+        let mut cancel = std::pin::pin!(cancel.cancelled());
+
         for chunk in delete_objects.chunks(MAX_KEYS_PER_DELETE) {
             let started_at = start_measuring_requests(kind);
 
-            let resp = self
+            let req = self
                 .client
                 .delete_objects()
                 .bucket(self.bucket_name.clone())
                 .delete(
                     Delete::builder()
                         .set_objects(Some(chunk.to_vec()))
-                        .build()?,
+                        .build()
+                        .context("build request")?,
                 )
-                .send()
-                .await;
+                .send();
+
+            let resp = tokio::select! {
+                resp = req => resp,
+                _ = tokio::time::sleep(self.timeout) => return Err(TimeoutOrCancel::Timeout.into()),
+                _ = &mut cancel => return Err(TimeoutOrCancel::Cancel.into()),
+            };
 
             let started_at = ScopeGuard::into_inner(started_at);
             metrics::BUCKET_METRICS
                 .req_seconds
                 .observe_elapsed(kind, &resp, started_at);
 
-            let resp = resp?;
+            let resp = resp.context("request deletion")?;
             metrics::BUCKET_METRICS
                 .deleted_objects_total
                 .inc_by(chunk.len() as u64);
+
             if let Some(errors) = resp.errors {
                 // Log a bounded number of the errors within the response:
                 // these requests can carry 1000 keys so logging each one
@@ -319,9 +361,10 @@ impl S3Bucket {
                     );
                 }
 
-                return Err(anyhow::format_err!(
-                    "Failed to delete {} objects",
-                    errors.len()
+                return Err(anyhow::anyhow!(
+                    "Failed to delete {}/{} objects",
+                    errors.len(),
+                    chunk.len(),
                 ));
             }
         }
@@ -354,33 +397,6 @@ impl Stream for ByteStreamAsStream {
     // sense and Stream::size_hint does not really
 }
 
-pin_project_lite::pin_project! {
-    /// An `AsyncRead` adapter which carries a permit for the lifetime of the value.
-    struct PermitCarrying<S> {
-        permit: tokio::sync::OwnedSemaphorePermit,
-        #[pin]
-        inner: S,
-    }
-}
-
-impl<S> PermitCarrying<S> {
-    fn new(permit: tokio::sync::OwnedSemaphorePermit, inner: S) -> Self {
-        Self { permit, inner }
-    }
-}
-
-impl<S: Stream<Item = std::io::Result<Bytes>>> Stream for PermitCarrying<S> {
-    type Item = <S as Stream>::Item;
-
-    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
-        self.project().inner.poll_next(cx)
-    }
-
-    fn size_hint(&self) -> (usize, Option<usize>) {
-        self.inner.size_hint()
-    }
-}
-
 pin_project_lite::pin_project! {
     /// Times and tracks the outcome of the request.
     struct TimedDownload<S> {
@@ -435,8 +451,12 @@ impl RemoteStorage for S3Bucket {
         &self,
         prefix: Option<&RemotePath>,
         mode: ListingMode,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Listing, DownloadError> {
         let kind = RequestKind::List;
+        // s3 sdk wants i32
+        let mut max_keys = max_keys.map(|mk| mk.get() as i32);
         let mut result = Listing::default();
 
         // get the passed prefix or if it is not set use prefix_in_bucket value
@@ -454,27 +474,41 @@ impl RemoteStorage for S3Bucket {
                 p
             });
 
+        let _permit = self.permit(kind, cancel).await?;
+
         let mut continuation_token = None;
 
         loop {
-            let _guard = self.permit(kind).await;
             let started_at = start_measuring_requests(kind);
 
+            // min of two Options, returning Some if one is value and another is
+            // None (None is smaller than anything, so plain min doesn't work).
+            let request_max_keys = self
+                .max_keys_per_list_response
+                .into_iter()
+                .chain(max_keys.into_iter())
+                .min();
             let mut request = self
                 .client
                 .list_objects_v2()
                 .bucket(self.bucket_name.clone())
                 .set_prefix(list_prefix.clone())
                 .set_continuation_token(continuation_token)
-                .set_max_keys(self.max_keys_per_list_response);
+                .set_max_keys(request_max_keys);
 
             if let ListingMode::WithDelimiter = mode {
                 request = request.delimiter(REMOTE_STORAGE_PREFIX_SEPARATOR.to_string());
             }
 
-            let response = request
-                .send()
-                .await
+            let request = request.send();
+
+            let response = tokio::select! {
+                res = request => res,
+                _ = tokio::time::sleep(self.timeout) => return Err(DownloadError::Timeout),
+                _ = cancel.cancelled() => return Err(DownloadError::Cancelled),
+            };
+
+            let response = response
                 .context("Failed to list S3 prefixes")
                 .map_err(DownloadError::Other);
 
@@ -496,6 +530,14 @@ impl RemoteStorage for S3Bucket {
                 let object_path = object.key().expect("response does not contain a key");
                 let remote_path = self.s3_object_to_relative_path(object_path);
                 result.keys.push(remote_path);
+                if let Some(mut mk) = max_keys {
+                    assert!(mk > 0);
+                    mk -= 1;
+                    if mk == 0 {
+                        return Ok(result); // limit reached
+                    }
+                    max_keys = Some(mk);
+                }
             }
 
             result.prefixes.extend(
@@ -519,16 +561,17 @@ impl RemoteStorage for S3Bucket {
         from_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         let kind = RequestKind::Put;
-        let _guard = self.permit(kind).await;
+        let _permit = self.permit(kind, cancel).await?;
 
         let started_at = start_measuring_requests(kind);
 
         let body = Body::wrap_stream(from);
         let bytes_stream = ByteStream::new(SdkBody::from_body_0_4(body));
 
-        let res = self
+        let upload = self
             .client
             .put_object()
             .bucket(self.bucket_name.clone())
@@ -536,22 +579,40 @@ impl RemoteStorage for S3Bucket {
             .set_metadata(metadata.map(|m| m.0))
             .content_length(from_size_bytes.try_into()?)
             .body(bytes_stream)
-            .send()
-            .await;
+            .send();
 
-        let started_at = ScopeGuard::into_inner(started_at);
-        metrics::BUCKET_METRICS
-            .req_seconds
-            .observe_elapsed(kind, &res, started_at);
+        let upload = tokio::time::timeout(self.timeout, upload);
 
-        res?;
+        let res = tokio::select! {
+            res = upload => res,
+            _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
+        };
 
-        Ok(())
+        if let Ok(inner) = &res {
+            // do not incl. timeouts as errors in metrics but cancellations
+            let started_at = ScopeGuard::into_inner(started_at);
+            metrics::BUCKET_METRICS
+                .req_seconds
+                .observe_elapsed(kind, inner, started_at);
+        }
+
+        match res {
+            Ok(Ok(_put)) => Ok(()),
+            Ok(Err(sdk)) => Err(sdk.into()),
+            Err(_timeout) => Err(TimeoutOrCancel::Timeout.into()),
+        }
     }
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         let kind = RequestKind::Copy;
-        let _guard = self.permit(kind).await;
+        let _permit = self.permit(kind, cancel).await?;
+
+        let timeout = tokio::time::sleep(self.timeout);
 
         let started_at = start_measuring_requests(kind);
 
@@ -562,14 +623,19 @@ impl RemoteStorage for S3Bucket {
             self.relative_path_to_s3_object(from)
         );
 
-        let res = self
+        let op = self
             .client
             .copy_object()
             .bucket(self.bucket_name.clone())
             .key(self.relative_path_to_s3_object(to))
             .copy_source(copy_source)
-            .send()
-            .await;
+            .send();
+
+        let res = tokio::select! {
+            res = op => res,
+            _ = timeout => return Err(TimeoutOrCancel::Timeout.into()),
+            _ = cancel.cancelled() => return Err(TimeoutOrCancel::Cancel.into()),
+        };
 
         let started_at = ScopeGuard::into_inner(started_at);
         metrics::BUCKET_METRICS
@@ -581,14 +647,21 @@ impl RemoteStorage for S3Bucket {
         Ok(())
     }
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
         // if prefix is not none then download file `prefix/from`
         // if prefix is none then download file `from`
-        self.download_object(GetObjectRequest {
-            bucket: self.bucket_name.clone(),
-            key: self.relative_path_to_s3_object(from),
-            range: None,
-        })
+        self.download_object(
+            GetObjectRequest {
+                bucket: self.bucket_name.clone(),
+                key: self.relative_path_to_s3_object(from),
+                range: None,
+            },
+            cancel,
+        )
         .await
     }
 
@@ -597,6 +670,7 @@ impl RemoteStorage for S3Bucket {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         // S3 accepts ranges as https://www.w3.org/Protocols/rfc2616/rfc2616-sec14.html#sec14.35
         // and needs both ends to be exclusive
@@ -606,31 +680,39 @@ impl RemoteStorage for S3Bucket {
             None => format!("bytes={start_inclusive}-"),
         });
 
-        self.download_object(GetObjectRequest {
-            bucket: self.bucket_name.clone(),
-            key: self.relative_path_to_s3_object(from),
-            range,
-        })
+        self.download_object(
+            GetObjectRequest {
+                bucket: self.bucket_name.clone(),
+                key: self.relative_path_to_s3_object(from),
+                range,
+            },
+            cancel,
+        )
         .await
     }
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
-        let kind = RequestKind::Delete;
-        let _guard = self.permit(kind).await;
 
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
+        let kind = RequestKind::Delete;
+        let permit = self.permit(kind, cancel).await?;
         let mut delete_objects = Vec::with_capacity(paths.len());
         for path in paths {
             let obj_id = ObjectIdentifier::builder()
                 .set_key(Some(self.relative_path_to_s3_object(path)))
-                .build()?;
+                .build()
+                .context("convert path to oid")?;
             delete_objects.push(obj_id);
         }
 
-        self.delete_oids(kind, &delete_objects).await
+        self.delete_oids(&permit, &delete_objects, cancel).await
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
         let paths = std::array::from_ref(path);
-        self.delete_objects(paths).await
+        self.delete_objects(paths, cancel).await
     }
 
     async fn time_travel_recover(
@@ -641,7 +723,7 @@ impl RemoteStorage for S3Bucket {
         cancel: &CancellationToken,
     ) -> Result<(), TimeTravelError> {
         let kind = RequestKind::TimeTravel;
-        let _guard = self.permit(kind).await;
+        let permit = self.permit(kind, cancel).await?;
 
         let timestamp = DateTime::from(timestamp);
         let done_if_after = DateTime::from(done_if_after);
@@ -655,7 +737,7 @@ impl RemoteStorage for S3Bucket {
 
         let warn_threshold = 3;
         let max_retries = 10;
-        let is_permanent = |_e: &_| false;
+        let is_permanent = |e: &_| matches!(e, TimeTravelError::Cancelled);
 
         let mut key_marker = None;
         let mut version_id_marker = None;
@@ -664,15 +746,19 @@ impl RemoteStorage for S3Bucket {
         loop {
             let response = backoff::retry(
                 || async {
-                    self.client
+                    let op = self
+                        .client
                         .list_object_versions()
                         .bucket(self.bucket_name.clone())
                         .set_prefix(prefix.clone())
                         .set_key_marker(key_marker.clone())
                         .set_version_id_marker(version_id_marker.clone())
-                        .send()
-                        .await
-                        .map_err(|e| TimeTravelError::Other(e.into()))
+                        .send();
+
+                    tokio::select! {
+                        res = op => res.map_err(|e| TimeTravelError::Other(e.into())),
+                        _ = cancel.cancelled() => Err(TimeTravelError::Cancelled),
+                    }
                 },
                 is_permanent,
                 warn_threshold,
@@ -794,14 +880,18 @@ impl RemoteStorage for S3Bucket {
 
                         backoff::retry(
                             || async {
-                                self.client
+                                let op = self
+                                    .client
                                     .copy_object()
                                     .bucket(self.bucket_name.clone())
                                     .key(key)
                                     .copy_source(&source_id)
-                                    .send()
-                                    .await
-                                    .map_err(|e| TimeTravelError::Other(e.into()))
+                                    .send();
+
+                                tokio::select! {
+                                    res = op => res.map_err(|e| TimeTravelError::Other(e.into())),
+                                    _ = cancel.cancelled() => Err(TimeTravelError::Cancelled),
+                                }
                             },
                             is_permanent,
                             warn_threshold,
@@ -832,10 +922,18 @@ impl RemoteStorage for S3Bucket {
                     let oid = ObjectIdentifier::builder()
                         .key(key.to_owned())
                         .build()
-                        .map_err(|e| TimeTravelError::Other(anyhow::Error::new(e)))?;
-                    self.delete_oids(kind, &[oid])
+                        .map_err(|e| TimeTravelError::Other(e.into()))?;
+
+                    self.delete_oids(&permit, &[oid], cancel)
                         .await
-                        .map_err(TimeTravelError::Other)?;
+                        .map_err(|e| {
+                            // delete_oid0 will use TimeoutOrCancel
+                            if TimeoutOrCancel::caused_by_cancel(&e) {
+                                TimeTravelError::Cancelled
+                            } else {
+                                TimeTravelError::Other(e)
+                            }
+                        })?;
                 }
             }
         }
@@ -971,7 +1069,8 @@ mod tests {
                 concurrency_limit: NonZeroUsize::new(100).unwrap(),
                 max_keys_per_list_response: Some(5),
             };
-            let storage = S3Bucket::new(&config).expect("remote storage init");
+            let storage =
+                S3Bucket::new(&config, std::time::Duration::ZERO).expect("remote storage init");
             for (test_path_idx, test_path) in all_paths.iter().enumerate() {
                 let result = storage.relative_path_to_s3_object(test_path);
                 let expected = expected_outputs[prefix_idx][test_path_idx];

libs/remote_storage/src/simulate_failures.rs

@@ -4,6 +4,7 @@
 use bytes::Bytes;
 use futures::stream::Stream;
 use std::collections::HashMap;
+use std::num::NonZeroU32;
 use std::sync::Mutex;
 use std::time::SystemTime;
 use std::{collections::hash_map::Entry, sync::Arc};
@@ -60,7 +61,7 @@ impl UnreliableWrapper {
     /// On the first attempts of this operation, return an error. After 'attempts_to_fail'
     /// attempts, let the operation go ahead, and clear the counter.
     ///
-    fn attempt(&self, op: RemoteOp) -> Result<u64, DownloadError> {
+    fn attempt(&self, op: RemoteOp) -> anyhow::Result<u64> {
         let mut attempts = self.attempts.lock().unwrap();
 
         match attempts.entry(op) {
@@ -78,22 +79,27 @@ impl UnreliableWrapper {
                 } else {
                     let error =
                         anyhow::anyhow!("simulated failure of remote operation {:?}", e.key());
-                    Err(DownloadError::Other(error))
+                    Err(error)
                 }
             }
             Entry::Vacant(e) => {
                 let error = anyhow::anyhow!("simulated failure of remote operation {:?}", e.key());
                 e.insert(1);
-                Err(DownloadError::Other(error))
+                Err(error)
             }
         }
     }
 
-    async fn delete_inner(&self, path: &RemotePath, attempt: bool) -> anyhow::Result<()> {
+    async fn delete_inner(
+        &self,
+        path: &RemotePath,
+        attempt: bool,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         if attempt {
             self.attempt(RemoteOp::Delete(path.clone()))?;
         }
-        self.inner.delete(path).await
+        self.inner.delete(path, cancel).await
     }
 }
 
@@ -104,23 +110,34 @@ impl RemoteStorage for UnreliableWrapper {
     async fn list_prefixes(
         &self,
         prefix: Option<&RemotePath>,
+        cancel: &CancellationToken,
     ) -> Result<Vec<RemotePath>, DownloadError> {
-        self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))?;
-        self.inner.list_prefixes(prefix).await
+        self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))
+            .map_err(DownloadError::Other)?;
+        self.inner.list_prefixes(prefix, cancel).await
     }
 
-    async fn list_files(&self, folder: Option<&RemotePath>) -> anyhow::Result<Vec<RemotePath>> {
-        self.attempt(RemoteOp::ListPrefixes(folder.cloned()))?;
-        self.inner.list_files(folder).await
+    async fn list_files(
+        &self,
+        folder: Option<&RemotePath>,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
+    ) -> Result<Vec<RemotePath>, DownloadError> {
+        self.attempt(RemoteOp::ListPrefixes(folder.cloned()))
+            .map_err(DownloadError::Other)?;
+        self.inner.list_files(folder, max_keys, cancel).await
     }
 
     async fn list(
         &self,
         prefix: Option<&RemotePath>,
         mode: ListingMode,
+        max_keys: Option<NonZeroU32>,
+        cancel: &CancellationToken,
     ) -> Result<Listing, DownloadError> {
-        self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))?;
-        self.inner.list(prefix, mode).await
+        self.attempt(RemoteOp::ListPrefixes(prefix.cloned()))
+            .map_err(DownloadError::Other)?;
+        self.inner.list(prefix, mode, max_keys, cancel).await
     }
 
     async fn upload(
@@ -131,14 +148,22 @@ impl RemoteStorage for UnreliableWrapper {
         data_size_bytes: usize,
         to: &RemotePath,
         metadata: Option<StorageMetadata>,
+        cancel: &CancellationToken,
     ) -> anyhow::Result<()> {
         self.attempt(RemoteOp::Upload(to.clone()))?;
-        self.inner.upload(data, data_size_bytes, to, metadata).await
+        self.inner
+            .upload(data, data_size_bytes, to, metadata, cancel)
+            .await
     }
 
-    async fn download(&self, from: &RemotePath) -> Result<Download, DownloadError> {
-        self.attempt(RemoteOp::Download(from.clone()))?;
-        self.inner.download(from).await
+    async fn download(
+        &self,
+        from: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> Result<Download, DownloadError> {
+        self.attempt(RemoteOp::Download(from.clone()))
+            .map_err(DownloadError::Other)?;
+        self.inner.download(from, cancel).await
     }
 
     async fn download_byte_range(
@@ -146,26 +171,32 @@ impl RemoteStorage for UnreliableWrapper {
         from: &RemotePath,
         start_inclusive: u64,
         end_exclusive: Option<u64>,
+        cancel: &CancellationToken,
     ) -> Result<Download, DownloadError> {
         // Note: We treat any download_byte_range as an "attempt" of the same
         // operation. We don't pay attention to the ranges. That's good enough
         // for now.
-        self.attempt(RemoteOp::Download(from.clone()))?;
+        self.attempt(RemoteOp::Download(from.clone()))
+            .map_err(DownloadError::Other)?;
         self.inner
-            .download_byte_range(from, start_inclusive, end_exclusive)
+            .download_byte_range(from, start_inclusive, end_exclusive, cancel)
             .await
     }
 
-    async fn delete(&self, path: &RemotePath) -> anyhow::Result<()> {
-        self.delete_inner(path, true).await
+    async fn delete(&self, path: &RemotePath, cancel: &CancellationToken) -> anyhow::Result<()> {
+        self.delete_inner(path, true, cancel).await
     }
 
-    async fn delete_objects<'a>(&self, paths: &'a [RemotePath]) -> anyhow::Result<()> {
+    async fn delete_objects<'a>(
+        &self,
+        paths: &'a [RemotePath],
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         self.attempt(RemoteOp::DeleteObjects(paths.to_vec()))?;
         let mut error_counter = 0;
         for path in paths {
             // Dont record attempt because it was already recorded above
-            if (self.delete_inner(path, false).await).is_err() {
+            if (self.delete_inner(path, false, cancel).await).is_err() {
                 error_counter += 1;
             }
         }
@@ -178,11 +209,16 @@ impl RemoteStorage for UnreliableWrapper {
         Ok(())
     }
 
-    async fn copy(&self, from: &RemotePath, to: &RemotePath) -> anyhow::Result<()> {
+    async fn copy(
+        &self,
+        from: &RemotePath,
+        to: &RemotePath,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<()> {
         // copy is equivalent to download + upload
         self.attempt(RemoteOp::Download(from.clone()))?;
         self.attempt(RemoteOp::Upload(to.clone()))?;
-        self.inner.copy_object(from, to).await
+        self.inner.copy_object(from, to, cancel).await
     }
 
     async fn time_travel_recover(
@@ -193,7 +229,7 @@ impl RemoteStorage for UnreliableWrapper {
         cancel: &CancellationToken,
     ) -> Result<(), TimeTravelError> {
         self.attempt(RemoteOp::TimeTravelRecover(prefix.map(|p| p.to_owned())))
-            .map_err(|e| TimeTravelError::Other(anyhow::Error::new(e)))?;
+            .map_err(TimeTravelError::Other)?;
         self.inner
             .time_travel_recover(prefix, timestamp, done_if_after, cancel)
             .await

libs/remote_storage/src/support.rs

@@ -0,0 +1,169 @@
+use std::{
+    future::Future,
+    pin::Pin,
+    task::{Context, Poll},
+    time::Duration,
+};
+
+use bytes::Bytes;
+use futures_util::Stream;
+use tokio_util::sync::CancellationToken;
+
+use crate::TimeoutOrCancel;
+
+pin_project_lite::pin_project! {
+    /// An `AsyncRead` adapter which carries a permit for the lifetime of the value.
+    pub(crate) struct PermitCarrying<S> {
+        permit: tokio::sync::OwnedSemaphorePermit,
+        #[pin]
+        inner: S,
+    }
+}
+
+impl<S> PermitCarrying<S> {
+    pub(crate) fn new(permit: tokio::sync::OwnedSemaphorePermit, inner: S) -> Self {
+        Self { permit, inner }
+    }
+}
+
+impl<S: Stream> Stream for PermitCarrying<S> {
+    type Item = <S as Stream>::Item;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        self.project().inner.poll_next(cx)
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+pin_project_lite::pin_project! {
+    pub(crate) struct DownloadStream<F, S> {
+        hit: bool,
+        #[pin]
+        cancellation: F,
+        #[pin]
+        inner: S,
+    }
+}
+
+impl<F, S> DownloadStream<F, S> {
+    pub(crate) fn new(cancellation: F, inner: S) -> Self {
+        Self {
+            cancellation,
+            hit: false,
+            inner,
+        }
+    }
+}
+
+/// See documentation on [`crate::DownloadStream`] on rationale why `std::io::Error` is used.
+impl<E, F, S> Stream for DownloadStream<F, S>
+where
+    std::io::Error: From<E>,
+    F: Future<Output = E>,
+    S: Stream<Item = std::io::Result<Bytes>>,
+{
+    type Item = <S as Stream>::Item;
+
+    fn poll_next(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
+        let this = self.project();
+
+        if !*this.hit {
+            if let Poll::Ready(e) = this.cancellation.poll(cx) {
+                *this.hit = true;
+                let e = Err(std::io::Error::from(e));
+                return Poll::Ready(Some(e));
+            }
+        }
+
+        this.inner.poll_next(cx)
+    }
+
+    fn size_hint(&self) -> (usize, Option<usize>) {
+        self.inner.size_hint()
+    }
+}
+
+/// Fires only on the first cancel or timeout, not on both.
+pub(crate) async fn cancel_or_timeout(
+    timeout: Duration,
+    cancel: CancellationToken,
+) -> TimeoutOrCancel {
+    tokio::select! {
+        _ = tokio::time::sleep(timeout) => TimeoutOrCancel::Timeout,
+        _ = cancel.cancelled() => TimeoutOrCancel::Cancel,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::DownloadError;
+    use futures::stream::StreamExt;
+
+    #[tokio::test(start_paused = true)]
+    async fn cancelled_download_stream() {
+        let inner = futures::stream::pending();
+        let timeout = Duration::from_secs(120);
+        let cancel = CancellationToken::new();
+
+        let stream = DownloadStream::new(cancel_or_timeout(timeout, cancel.clone()), inner);
+        let mut stream = std::pin::pin!(stream);
+
+        let mut first = stream.next();
+
+        tokio::select! {
+            _ = &mut first => unreachable!("we haven't yet cancelled nor is timeout passed"),
+            _ = tokio::time::sleep(Duration::from_secs(1)) => {},
+        }
+
+        cancel.cancel();
+
+        let e = first.await.expect("there must be some").unwrap_err();
+        assert!(matches!(e.kind(), std::io::ErrorKind::Other), "{e:?}");
+        let inner = e.get_ref().expect("inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Cancelled)),
+            "{inner:?}"
+        );
+
+        tokio::select! {
+            _ = stream.next() => unreachable!("no timeout ever happens as we were already cancelled"),
+            _ = tokio::time::sleep(Duration::from_secs(121)) => {},
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn timeouted_download_stream() {
+        let inner = futures::stream::pending();
+        let timeout = Duration::from_secs(120);
+        let cancel = CancellationToken::new();
+
+        let stream = DownloadStream::new(cancel_or_timeout(timeout, cancel.clone()), inner);
+        let mut stream = std::pin::pin!(stream);
+
+        // because the stream uses 120s timeout we are paused, we advance to 120s right away.
+        let first = stream.next();
+
+        let e = first.await.expect("there must be some").unwrap_err();
+        assert!(matches!(e.kind(), std::io::ErrorKind::Other), "{e:?}");
+        let inner = e.get_ref().expect("inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Timeout)),
+            "{inner:?}"
+        );
+
+        cancel.cancel();
+
+        tokio::select! {
+            _ = stream.next() => unreachable!("no cancellation ever happens because we already timed out"),
+            _ = tokio::time::sleep(Duration::from_secs(121)) => {},
+        }
+    }
+}

libs/remote_storage/tests/common/mod.rs

@@ -10,6 +10,7 @@ use futures::stream::Stream;
 use once_cell::sync::OnceCell;
 use remote_storage::{Download, GenericRemoteStorage, RemotePath};
 use tokio::task::JoinSet;
+use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info};
 
 static LOGGING_DONE: OnceCell<()> = OnceCell::new();
@@ -58,8 +59,12 @@ pub(crate) async fn upload_simple_remote_data(
 ) -> ControlFlow<HashSet<RemotePath>, HashSet<RemotePath>> {
     info!("Creating {upload_tasks_count} remote files");
     let mut upload_tasks = JoinSet::new();
+    let cancel = CancellationToken::new();
+
     for i in 1..upload_tasks_count + 1 {
         let task_client = Arc::clone(client);
+        let cancel = cancel.clone();
+
         upload_tasks.spawn(async move {
             let blob_path = PathBuf::from(format!("folder{}/blob_{}.txt", i / 7, i));
             let blob_path = RemotePath::new(
@@ -69,7 +74,9 @@ pub(crate) async fn upload_simple_remote_data(
             debug!("Creating remote item {i} at path {blob_path:?}");
 
             let (data, len) = upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, len, &blob_path, None).await?;
+            task_client
+                .upload(data, len, &blob_path, None, &cancel)
+                .await?;
 
             Ok::<_, anyhow::Error>(blob_path)
         });
@@ -107,13 +114,15 @@ pub(crate) async fn cleanup(
         "Removing {} objects from the remote storage during cleanup",
         objects_to_delete.len()
     );
+    let cancel = CancellationToken::new();
     let mut delete_tasks = JoinSet::new();
     for object_to_delete in objects_to_delete {
         let task_client = Arc::clone(client);
+        let cancel = cancel.clone();
         delete_tasks.spawn(async move {
             debug!("Deleting remote item at path {object_to_delete:?}");
             task_client
-                .delete(&object_to_delete)
+                .delete(&object_to_delete, &cancel)
                 .await
                 .with_context(|| format!("{object_to_delete:?} removal"))
         });
@@ -141,8 +150,12 @@ pub(crate) async fn upload_remote_data(
 ) -> ControlFlow<Uploads, Uploads> {
     info!("Creating {upload_tasks_count} remote files");
     let mut upload_tasks = JoinSet::new();
+    let cancel = CancellationToken::new();
+
     for i in 1..upload_tasks_count + 1 {
         let task_client = Arc::clone(client);
+        let cancel = cancel.clone();
+
         upload_tasks.spawn(async move {
             let prefix = format!("{base_prefix_str}/sub_prefix_{i}/");
             let blob_prefix = RemotePath::new(Utf8Path::new(&prefix))
@@ -152,7 +165,9 @@ pub(crate) async fn upload_remote_data(
 
             let (data, data_len) =
                 upload_stream(format!("remote blob data {i}").into_bytes().into());
-            task_client.upload(data, data_len, &blob_path, None).await?;
+            task_client
+                .upload(data, data_len, &blob_path, None, &cancel)
+                .await?;
 
             Ok::<_, anyhow::Error>((blob_prefix, blob_path))
         });

libs/remote_storage/tests/common/tests.rs

@@ -1,9 +1,10 @@
 use anyhow::Context;
 use camino::Utf8Path;
 use remote_storage::RemotePath;
-use std::collections::HashSet;
 use std::sync::Arc;
+use std::{collections::HashSet, num::NonZeroU32};
 use test_context::test_context;
+use tokio_util::sync::CancellationToken;
 use tracing::debug;
 
 use crate::common::{download_to_vec, upload_stream, wrap_stream};
@@ -45,13 +46,15 @@ async fn pagination_should_work(ctx: &mut MaybeEnabledStorageWithTestBlobs) -> a
         }
     };
 
+    let cancel = CancellationToken::new();
+
     let test_client = Arc::clone(&ctx.enabled.client);
     let expected_remote_prefixes = ctx.remote_prefixes.clone();
 
     let base_prefix = RemotePath::new(Utf8Path::new(ctx.enabled.base_prefix))
         .context("common_prefix construction")?;
     let root_remote_prefixes = test_client
-        .list_prefixes(None)
+        .list_prefixes(None, &cancel)
         .await
         .context("client list root prefixes failure")?
         .into_iter()
@@ -62,7 +65,7 @@ async fn pagination_should_work(ctx: &mut MaybeEnabledStorageWithTestBlobs) -> a
     );
 
     let nested_remote_prefixes = test_client
-        .list_prefixes(Some(&base_prefix))
+        .list_prefixes(Some(&base_prefix), &cancel)
         .await
         .context("client list nested prefixes failure")?
         .into_iter()
@@ -99,11 +102,12 @@ async fn list_files_works(ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs) -> a
             anyhow::bail!("S3 init failed: {e:?}")
         }
     };
+    let cancel = CancellationToken::new();
     let test_client = Arc::clone(&ctx.enabled.client);
     let base_prefix =
         RemotePath::new(Utf8Path::new("folder1")).context("common_prefix construction")?;
     let root_files = test_client
-        .list_files(None)
+        .list_files(None, None, &cancel)
         .await
         .context("client list root files failure")?
         .into_iter()
@@ -113,8 +117,17 @@ async fn list_files_works(ctx: &mut MaybeEnabledStorageWithSimpleTestBlobs) -> a
         ctx.remote_blobs.clone(),
         "remote storage list_files on root mismatches with the uploads."
     );
+
+    // Test that max_keys limit works. In total there are about 21 files (see
+    // upload_simple_remote_data call in test_real_s3.rs).
+    let limited_root_files = test_client
+        .list_files(None, Some(NonZeroU32::new(2).unwrap()), &cancel)
+        .await
+        .context("client list root files failure")?;
+    assert_eq!(limited_root_files.len(), 2);
+
     let nested_remote_files = test_client
-        .list_files(Some(&base_prefix))
+        .list_files(Some(&base_prefix), None, &cancel)
         .await
         .context("client list nested files failure")?
         .into_iter()
@@ -141,12 +154,17 @@ async fn delete_non_exising_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Resu
         MaybeEnabledStorage::Disabled => return Ok(()),
     };
 
+    let cancel = CancellationToken::new();
+
     let path = RemotePath::new(Utf8Path::new(
         format!("{}/for_sure_there_is_nothing_there_really", ctx.base_prefix).as_str(),
     ))
     .with_context(|| "RemotePath conversion")?;
 
-    ctx.client.delete(&path).await.expect("should succeed");
+    ctx.client
+        .delete(&path, &cancel)
+        .await
+        .expect("should succeed");
 
     Ok(())
 }
@@ -159,6 +177,8 @@ async fn delete_objects_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<(
         MaybeEnabledStorage::Disabled => return Ok(()),
     };
 
+    let cancel = CancellationToken::new();
+
     let path1 = RemotePath::new(Utf8Path::new(format!("{}/path1", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
@@ -169,21 +189,21 @@ async fn delete_objects_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<(
         .with_context(|| "RemotePath conversion")?;
 
     let (data, len) = upload_stream("remote blob data1".as_bytes().into());
-    ctx.client.upload(data, len, &path1, None).await?;
+    ctx.client.upload(data, len, &path1, None, &cancel).await?;
 
     let (data, len) = upload_stream("remote blob data2".as_bytes().into());
-    ctx.client.upload(data, len, &path2, None).await?;
+    ctx.client.upload(data, len, &path2, None, &cancel).await?;
 
     let (data, len) = upload_stream("remote blob data3".as_bytes().into());
-    ctx.client.upload(data, len, &path3, None).await?;
+    ctx.client.upload(data, len, &path3, None, &cancel).await?;
 
-    ctx.client.delete_objects(&[path1, path2]).await?;
+    ctx.client.delete_objects(&[path1, path2], &cancel).await?;
 
-    let prefixes = ctx.client.list_prefixes(None).await?;
+    let prefixes = ctx.client.list_prefixes(None, &cancel).await?;
 
     assert_eq!(prefixes.len(), 1);
 
-    ctx.client.delete_objects(&[path3]).await?;
+    ctx.client.delete_objects(&[path3], &cancel).await?;
 
     Ok(())
 }
@@ -195,6 +215,8 @@ async fn upload_download_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<
         return Ok(());
     };
 
+    let cancel = CancellationToken::new();
+
     let path = RemotePath::new(Utf8Path::new(format!("{}/file", ctx.base_prefix).as_str()))
         .with_context(|| "RemotePath conversion")?;
 
@@ -202,47 +224,56 @@ async fn upload_download_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<
 
     let (data, len) = wrap_stream(orig.clone());
 
-    ctx.client.upload(data, len, &path, None).await?;
+    ctx.client.upload(data, len, &path, None, &cancel).await?;
 
     // Normal download request
-    let dl = ctx.client.download(&path).await?;
+    let dl = ctx.client.download(&path, &cancel).await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // Full range (end specified)
     let dl = ctx
         .client
-        .download_byte_range(&path, 0, Some(len as u64))
+        .download_byte_range(&path, 0, Some(len as u64), &cancel)
         .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     // partial range (end specified)
-    let dl = ctx.client.download_byte_range(&path, 4, Some(10)).await?;
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 4, Some(10), &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..10]);
 
     // partial range (end beyond real end)
     let dl = ctx
         .client
-        .download_byte_range(&path, 8, Some(len as u64 * 100))
+        .download_byte_range(&path, 8, Some(len as u64 * 100), &cancel)
         .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[8..]);
 
     // Partial range (end unspecified)
-    let dl = ctx.client.download_byte_range(&path, 4, None).await?;
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 4, None, &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig[4..]);
 
     // Full range (end unspecified)
-    let dl = ctx.client.download_byte_range(&path, 0, None).await?;
+    let dl = ctx
+        .client
+        .download_byte_range(&path, 0, None, &cancel)
+        .await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     debug!("Cleanup: deleting file at path {path:?}");
     ctx.client
-        .delete(&path)
+        .delete(&path, &cancel)
         .await
         .with_context(|| format!("{path:?} removal"))?;
 
@@ -256,6 +287,8 @@ async fn copy_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
         return Ok(());
     };
 
+    let cancel = CancellationToken::new();
+
     let path = RemotePath::new(Utf8Path::new(
         format!("{}/file_to_copy", ctx.base_prefix).as_str(),
     ))
@@ -269,18 +302,18 @@ async fn copy_works(ctx: &mut MaybeEnabledStorage) -> anyhow::Result<()> {
 
     let (data, len) = wrap_stream(orig.clone());
 
-    ctx.client.upload(data, len, &path, None).await?;
+    ctx.client.upload(data, len, &path, None, &cancel).await?;
 
     // Normal download request
-    ctx.client.copy_object(&path, &path_dest).await?;
+    ctx.client.copy_object(&path, &path_dest, &cancel).await?;
 
-    let dl = ctx.client.download(&path_dest).await?;
+    let dl = ctx.client.download(&path_dest, &cancel).await?;
     let buf = download_to_vec(dl).await?;
     assert_eq!(&buf, &orig);
 
     debug!("Cleanup: deleting file at path {path:?}");
     ctx.client
-        .delete_objects(&[path.clone(), path_dest.clone()])
+        .delete_objects(&[path.clone(), path_dest.clone()], &cancel)
         .await
         .with_context(|| format!("{path:?} removal"))?;
 

libs/remote_storage/tests/test_real_azure.rs

@@ -1,9 +1,9 @@
-use std::collections::HashSet;
 use std::env;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
 use std::sync::Arc;
 use std::time::UNIX_EPOCH;
+use std::{collections::HashSet, time::Duration};
 
 use anyhow::Context;
 use remote_storage::{
@@ -39,6 +39,17 @@ impl EnabledAzure {
             base_prefix: BASE_PREFIX,
         }
     }
+
+    #[allow(unused)] // this will be needed when moving the timeout integration tests back
+    fn configure_request_timeout(&mut self, timeout: Duration) {
+        match Arc::get_mut(&mut self.client).expect("outer Arc::get_mut") {
+            GenericRemoteStorage::AzureBlob(azure) => {
+                let azure = Arc::get_mut(azure).expect("inner Arc::get_mut");
+                azure.timeout = timeout;
+            }
+            _ => unreachable!(),
+        }
+    }
 }
 
 enum MaybeEnabledStorage {
@@ -213,6 +224,7 @@ fn create_azure_client(
             concurrency_limit: NonZeroUsize::new(100).unwrap(),
             max_keys_per_list_response,
         }),
+        timeout: Duration::from_secs(120),
     };
     Ok(Arc::new(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,

libs/remote_storage/tests/test_real_s3.rs

@@ -1,5 +1,6 @@
 use std::env;
 use std::fmt::{Debug, Display};
+use std::future::Future;
 use std::num::NonZeroUsize;
 use std::ops::ControlFlow;
 use std::sync::Arc;
@@ -9,9 +10,10 @@ use std::{collections::HashSet, time::SystemTime};
 use crate::common::{download_to_vec, upload_stream};
 use anyhow::Context;
 use camino::Utf8Path;
-use futures_util::Future;
+use futures_util::StreamExt;
 use remote_storage::{
-    GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind, S3Config,
+    DownloadError, GenericRemoteStorage, RemotePath, RemoteStorageConfig, RemoteStorageKind,
+    S3Config,
 };
 use test_context::test_context;
 use test_context::AsyncTestContext;
@@ -27,7 +29,6 @@ use common::{cleanup, ensure_logging_ready, upload_remote_data, upload_simple_re
 use utils::backoff;
 
 const ENABLE_REAL_S3_REMOTE_STORAGE_ENV_VAR_NAME: &str = "ENABLE_REAL_S3_REMOTE_STORAGE";
-
 const BASE_PREFIX: &str = "test";
 
 #[test_context(MaybeEnabledStorage)]
@@ -69,8 +70,11 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
         ret
     }
 
-    async fn list_files(client: &Arc<GenericRemoteStorage>) -> anyhow::Result<HashSet<RemotePath>> {
-        Ok(retry(|| client.list_files(None))
+    async fn list_files(
+        client: &Arc<GenericRemoteStorage>,
+        cancel: &CancellationToken,
+    ) -> anyhow::Result<HashSet<RemotePath>> {
+        Ok(retry(|| client.list_files(None, None, cancel))
             .await
             .context("list root files failure")?
             .into_iter()
@@ -90,11 +94,11 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream("remote blob data1".as_bytes().into());
-        ctx.client.upload(data, len, &path1, None)
+        ctx.client.upload(data, len, &path1, None, &cancel)
     })
     .await?;
 
-    let t0_files = list_files(&ctx.client).await?;
+    let t0_files = list_files(&ctx.client, &cancel).await?;
     let t0 = time_point().await;
     println!("at t0: {t0_files:?}");
 
@@ -102,17 +106,17 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream(old_data.as_bytes().into());
-        ctx.client.upload(data, len, &path2, None)
+        ctx.client.upload(data, len, &path2, None, &cancel)
     })
     .await?;
 
-    let t1_files = list_files(&ctx.client).await?;
+    let t1_files = list_files(&ctx.client, &cancel).await?;
     let t1 = time_point().await;
     println!("at t1: {t1_files:?}");
 
     // A little check to ensure that our clock is not too far off from the S3 clock
     {
-        let dl = retry(|| ctx.client.download(&path2)).await?;
+        let dl = retry(|| ctx.client.download(&path2, &cancel)).await?;
         let last_modified = dl.last_modified.unwrap();
         let half_wt = WAIT_TIME.mul_f32(0.5);
         let t0_hwt = t0 + half_wt;
@@ -125,7 +129,7 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream("remote blob data3".as_bytes().into());
-        ctx.client.upload(data, len, &path3, None)
+        ctx.client.upload(data, len, &path3, None, &cancel)
     })
     .await?;
 
@@ -133,12 +137,12 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
 
     retry(|| {
         let (data, len) = upload_stream(new_data.as_bytes().into());
-        ctx.client.upload(data, len, &path2, None)
+        ctx.client.upload(data, len, &path2, None, &cancel)
     })
     .await?;
 
-    retry(|| ctx.client.delete(&path1)).await?;
-    let t2_files = list_files(&ctx.client).await?;
+    retry(|| ctx.client.delete(&path1, &cancel)).await?;
+    let t2_files = list_files(&ctx.client, &cancel).await?;
     let t2 = time_point().await;
     println!("at t2: {t2_files:?}");
 
@@ -147,10 +151,10 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     ctx.client
         .time_travel_recover(None, t2, t_final, &cancel)
         .await?;
-    let t2_files_recovered = list_files(&ctx.client).await?;
+    let t2_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t2: {t2_files_recovered:?}");
     assert_eq!(t2_files, t2_files_recovered);
-    let path2_recovered_t2 = download_to_vec(ctx.client.download(&path2).await?).await?;
+    let path2_recovered_t2 = download_to_vec(ctx.client.download(&path2, &cancel).await?).await?;
     assert_eq!(path2_recovered_t2, new_data.as_bytes());
 
     // after recovery to t1: path1 is back, path2 has the old content
@@ -158,10 +162,10 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     ctx.client
         .time_travel_recover(None, t1, t_final, &cancel)
         .await?;
-    let t1_files_recovered = list_files(&ctx.client).await?;
+    let t1_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t1: {t1_files_recovered:?}");
     assert_eq!(t1_files, t1_files_recovered);
-    let path2_recovered_t1 = download_to_vec(ctx.client.download(&path2).await?).await?;
+    let path2_recovered_t1 = download_to_vec(ctx.client.download(&path2, &cancel).await?).await?;
     assert_eq!(path2_recovered_t1, old_data.as_bytes());
 
     // after recovery to t0: everything is gone except for path1
@@ -169,14 +173,14 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     ctx.client
         .time_travel_recover(None, t0, t_final, &cancel)
         .await?;
-    let t0_files_recovered = list_files(&ctx.client).await?;
+    let t0_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t0: {t0_files_recovered:?}");
     assert_eq!(t0_files, t0_files_recovered);
 
     // cleanup
 
     let paths = &[path1, path2, path3];
-    retry(|| ctx.client.delete_objects(paths)).await?;
+    retry(|| ctx.client.delete_objects(paths, &cancel)).await?;
 
     Ok(())
 }
@@ -197,6 +201,16 @@ impl EnabledS3 {
             base_prefix: BASE_PREFIX,
         }
     }
+
+    fn configure_request_timeout(&mut self, timeout: Duration) {
+        match Arc::get_mut(&mut self.client).expect("outer Arc::get_mut") {
+            GenericRemoteStorage::AwsS3(s3) => {
+                let s3 = Arc::get_mut(s3).expect("inner Arc::get_mut");
+                s3.timeout = timeout;
+            }
+            _ => unreachable!(),
+        }
+    }
 }
 
 enum MaybeEnabledStorage {
@@ -370,8 +384,169 @@ fn create_s3_client(
             concurrency_limit: NonZeroUsize::new(100).unwrap(),
             max_keys_per_list_response,
         }),
+        timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
     };
     Ok(Arc::new(
         GenericRemoteStorage::from_config(&remote_storage_config).context("remote storage init")?,
     ))
 }
+
+#[test_context(MaybeEnabledStorage)]
+#[tokio::test]
+async fn download_is_timeouted(ctx: &mut MaybeEnabledStorage) {
+    let MaybeEnabledStorage::Enabled(ctx) = ctx else {
+        return;
+    };
+
+    let cancel = CancellationToken::new();
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/file_to_copy", ctx.base_prefix).as_str(),
+    ))
+    .unwrap();
+
+    let len = upload_large_enough_file(&ctx.client, &path, &cancel).await;
+
+    let timeout = std::time::Duration::from_secs(5);
+
+    ctx.configure_request_timeout(timeout);
+
+    let started_at = std::time::Instant::now();
+    let mut stream = ctx
+        .client
+        .download(&path, &cancel)
+        .await
+        .expect("download succeeds")
+        .download_stream;
+
+    if started_at.elapsed().mul_f32(0.9) >= timeout {
+        tracing::warn!(
+            elapsed_ms = started_at.elapsed().as_millis(),
+            "timeout might be too low, consumed most of it during headers"
+        );
+    }
+
+    let first = stream
+        .next()
+        .await
+        .expect("should have the first blob")
+        .expect("should have succeeded");
+
+    tracing::info!(len = first.len(), "downloaded first chunk");
+
+    assert!(
+        first.len() < len,
+        "uploaded file is too small, we downloaded all on first chunk"
+    );
+
+    tokio::time::sleep(timeout).await;
+
+    {
+        let started_at = std::time::Instant::now();
+        let next = stream
+            .next()
+            .await
+            .expect("stream should not have ended yet");
+
+        tracing::info!(
+            next.is_err = next.is_err(),
+            elapsed_ms = started_at.elapsed().as_millis(),
+            "received item after timeout"
+        );
+
+        let e = next.expect_err("expected an error, but got a chunk?");
+
+        let inner = e.get_ref().expect("std::io::Error::inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Timeout)),
+            "{inner:?}"
+        );
+    }
+
+    ctx.configure_request_timeout(RemoteStorageConfig::DEFAULT_TIMEOUT);
+
+    ctx.client.delete_objects(&[path], &cancel).await.unwrap()
+}
+
+#[test_context(MaybeEnabledStorage)]
+#[tokio::test]
+async fn download_is_cancelled(ctx: &mut MaybeEnabledStorage) {
+    let MaybeEnabledStorage::Enabled(ctx) = ctx else {
+        return;
+    };
+
+    let cancel = CancellationToken::new();
+
+    let path = RemotePath::new(Utf8Path::new(
+        format!("{}/file_to_copy", ctx.base_prefix).as_str(),
+    ))
+    .unwrap();
+
+    let len = upload_large_enough_file(&ctx.client, &path, &cancel).await;
+
+    {
+        let mut stream = ctx
+            .client
+            .download(&path, &cancel)
+            .await
+            .expect("download succeeds")
+            .download_stream;
+
+        let first = stream
+            .next()
+            .await
+            .expect("should have the first blob")
+            .expect("should have succeeded");
+
+        tracing::info!(len = first.len(), "downloaded first chunk");
+
+        assert!(
+            first.len() < len,
+            "uploaded file is too small, we downloaded all on first chunk"
+        );
+
+        cancel.cancel();
+
+        let next = stream.next().await.expect("stream should have more");
+
+        let e = next.expect_err("expected an error, but got a chunk?");
+
+        let inner = e.get_ref().expect("std::io::Error::inner should be set");
+        assert!(
+            inner
+                .downcast_ref::<DownloadError>()
+                .is_some_and(|e| matches!(e, DownloadError::Cancelled)),
+            "{inner:?}"
+        );
+    }
+
+    let cancel = CancellationToken::new();
+
+    ctx.client.delete_objects(&[path], &cancel).await.unwrap();
+}
+
+/// Upload a long enough file so that we cannot download it in single chunk
+///
+/// For s3 the first chunk seems to be less than 10kB, so this has a bit of a safety margin
+async fn upload_large_enough_file(
+    client: &GenericRemoteStorage,
+    path: &RemotePath,
+    cancel: &CancellationToken,
+) -> usize {
+    let header = bytes::Bytes::from_static("remote blob data content".as_bytes());
+    let body = bytes::Bytes::from(vec![0u8; 1024]);
+    let contents = std::iter::once(header).chain(std::iter::repeat(body).take(128));
+
+    let len = contents.clone().fold(0, |acc, next| acc + next.len());
+
+    let contents = futures::stream::iter(contents.map(std::io::Result::Ok));
+
+    client
+        .upload(contents, len, path, None, cancel)
+        .await
+        .expect("upload succeeds");
+
+    len
+}

libs/safekeeper_api/src/models.rs

@@ -50,9 +50,6 @@ pub struct SkTimelineInfo {
     pub safekeeper_connstr: Option<String>,
     #[serde(default)]
     pub http_connstr: Option<String>,
-    // Minimum of all active RO replicas flush LSN
-    #[serde(default = "lsn_invalid")]
-    pub standby_horizon: Lsn,
 }
 
 #[derive(Debug, Clone, Deserialize, Serialize)]

libs/utils/benches/benchmarks.rs

@@ -1,5 +1,3 @@
-#![allow(unused)]
-
 use criterion::{criterion_group, criterion_main, Criterion};
 use utils::id;
 

libs/utils/src/completion.rs

@@ -27,6 +27,11 @@ impl Barrier {
             b.wait().await
         }
     }
+
+    /// Return true if a call to wait() would complete immediately
+    pub fn is_ready(&self) -> bool {
+        futures::future::FutureExt::now_or_never(self.0.wait()).is_some()
+    }
 }
 
 impl PartialEq for Barrier {

libs/utils/src/generation.rs

@@ -54,12 +54,10 @@ impl Generation {
     }
 
     #[track_caller]
-    pub fn get_suffix(&self) -> String {
+    pub fn get_suffix(&self) -> impl std::fmt::Display {
         match self {
-            Self::Valid(v) => {
-                format!("-{:08x}", v)
-            }
-            Self::None => "".into(),
+            Self::Valid(v) => GenerationFileSuffix(Some(*v)),
+            Self::None => GenerationFileSuffix(None),
             Self::Broken => {
                 panic!("Tried to use a broken generation");
             }
@@ -90,6 +88,7 @@ impl Generation {
         }
     }
 
+    #[track_caller]
     pub fn next(&self) -> Generation {
         match self {
             Self::Valid(n) => Self::Valid(*n + 1),
@@ -107,6 +106,18 @@ impl Generation {
     }
 }
 
+struct GenerationFileSuffix(Option<u32>);
+
+impl std::fmt::Display for GenerationFileSuffix {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        if let Some(g) = self.0 {
+            write!(f, "-{g:08x}")
+        } else {
+            Ok(())
+        }
+    }
+}
+
 impl Serialize for Generation {
     fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
     where
@@ -164,4 +175,24 @@ mod test {
         assert!(Generation::none() < Generation::new(0));
         assert!(Generation::none() < Generation::new(1));
     }
+
+    #[test]
+    fn suffix_is_stable() {
+        use std::fmt::Write as _;
+
+        // the suffix must remain stable through-out the pageserver remote storage evolution and
+        // not be changed accidentially without thinking about migration
+        let examples = [
+            (line!(), Generation::None, ""),
+            (line!(), Generation::Valid(0), "-00000000"),
+            (line!(), Generation::Valid(u32::MAX), "-ffffffff"),
+        ];
+
+        let mut s = String::new();
+        for (line, gen, expected) in examples {
+            s.clear();
+            write!(s, "{}", &gen.get_suffix()).expect("string grows");
+            assert_eq!(s, expected, "example on {line}");
+        }
+    }
 }

libs/utils/src/sync/heavier_once_cell.rs

@@ -1,6 +1,6 @@
 use std::sync::{
     atomic::{AtomicUsize, Ordering},
-    Arc,
+    Arc, Mutex, MutexGuard,
 };
 use tokio::sync::Semaphore;
 
@@ -12,7 +12,7 @@ use tokio::sync::Semaphore;
 ///
 /// [`OwnedSemaphorePermit`]: tokio::sync::OwnedSemaphorePermit
 pub struct OnceCell<T> {
-    inner: tokio::sync::RwLock<Inner<T>>,
+    inner: Mutex<Inner<T>>,
     initializers: AtomicUsize,
 }
 
@@ -50,7 +50,7 @@ impl<T> OnceCell<T> {
         let sem = Semaphore::new(1);
         sem.close();
         Self {
-            inner: tokio::sync::RwLock::new(Inner {
+            inner: Mutex::new(Inner {
                 init_semaphore: Arc::new(sem),
                 value: Some(value),
             }),
@@ -61,99 +61,63 @@ impl<T> OnceCell<T> {
     /// Returns a guard to an existing initialized value, or uniquely initializes the value before
     /// returning the guard.
     ///
-    /// Initializing might wait on any existing [`GuardMut::take_and_deinit`] deinitialization.
+    /// Initializing might wait on any existing [`Guard::take_and_deinit`] deinitialization.
     ///
     /// Initialization is panic-safe and cancellation-safe.
-    pub async fn get_mut_or_init<F, Fut, E>(&self, factory: F) -> Result<GuardMut<'_, T>, E>
+    pub async fn get_or_init<F, Fut, E>(&self, factory: F) -> Result<Guard<'_, T>, E>
     where
         F: FnOnce(InitPermit) -> Fut,
         Fut: std::future::Future<Output = Result<(T, InitPermit), E>>,
     {
-        let sem = {
-            let guard = self.inner.write().await;
-            if guard.value.is_some() {
-                return Ok(GuardMut(guard));
+        loop {
+            let sem = {
+                let guard = self.inner.lock().unwrap();
+                if guard.value.is_some() {
+                    return Ok(Guard(guard));
+                }
+                guard.init_semaphore.clone()
+            };
+
+            {
+                let permit = {
+                    // increment the count for the duration of queued
+                    let _guard = CountWaitingInitializers::start(self);
+                    sem.acquire().await
+                };
+
+                let Ok(permit) = permit else {
+                    let guard = self.inner.lock().unwrap();
+                    if !Arc::ptr_eq(&sem, &guard.init_semaphore) {
+                        // there was a take_and_deinit in between
+                        continue;
+                    }
+                    assert!(
+                        guard.value.is_some(),
+                        "semaphore got closed, must be initialized"
+                    );
+                    return Ok(Guard(guard));
+                };
+
+                permit.forget();
             }
-            guard.init_semaphore.clone()
-        };
 
-        let permit = {
-            // increment the count for the duration of queued
-            let _guard = CountWaitingInitializers::start(self);
-            sem.acquire_owned().await
-        };
+            let permit = InitPermit(sem);
+            let (value, _permit) = factory(permit).await?;
 
-        match permit {
-            Ok(permit) => {
-                let permit = InitPermit(permit);
-                let (value, _permit) = factory(permit).await?;
+            let guard = self.inner.lock().unwrap();
 
-                let guard = self.inner.write().await;
-
-                Ok(Self::set0(value, guard))
-            }
-            Err(_closed) => {
-                let guard = self.inner.write().await;
-                assert!(
-                    guard.value.is_some(),
-                    "semaphore got closed, must be initialized"
-                );
-                return Ok(GuardMut(guard));
-            }
+            return Ok(Self::set0(value, guard));
         }
     }
 
-    /// Returns a guard to an existing initialized value, or uniquely initializes the value before
-    /// returning the guard.
-    ///
-    /// Initialization is panic-safe and cancellation-safe.
-    pub async fn get_or_init<F, Fut, E>(&self, factory: F) -> Result<GuardRef<'_, T>, E>
-    where
-        F: FnOnce(InitPermit) -> Fut,
-        Fut: std::future::Future<Output = Result<(T, InitPermit), E>>,
-    {
-        let sem = {
-            let guard = self.inner.read().await;
-            if guard.value.is_some() {
-                return Ok(GuardRef(guard));
-            }
-            guard.init_semaphore.clone()
-        };
-
-        let permit = {
-            // increment the count for the duration of queued
-            let _guard = CountWaitingInitializers::start(self);
-            sem.acquire_owned().await
-        };
-
-        match permit {
-            Ok(permit) => {
-                let permit = InitPermit(permit);
-                let (value, _permit) = factory(permit).await?;
-
-                let guard = self.inner.write().await;
-
-                Ok(Self::set0(value, guard).downgrade())
-            }
-            Err(_closed) => {
-                let guard = self.inner.read().await;
-                assert!(
-                    guard.value.is_some(),
-                    "semaphore got closed, must be initialized"
-                );
-                return Ok(GuardRef(guard));
-            }
-        }
-    }
-
-    /// Assuming a permit is held after previous call to [`GuardMut::take_and_deinit`], it can be used
+    /// Assuming a permit is held after previous call to [`Guard::take_and_deinit`], it can be used
     /// to complete initializing the inner value.
     ///
     /// # Panics
     ///
     /// If the inner has already been initialized.
-    pub async fn set(&self, value: T, _permit: InitPermit) -> GuardMut<'_, T> {
-        let guard = self.inner.write().await;
+    pub fn set(&self, value: T, _permit: InitPermit) -> Guard<'_, T> {
+        let guard = self.inner.lock().unwrap();
 
         // cannot assert that this permit is for self.inner.semaphore, but we can assert it cannot
         // give more permits right now.
@@ -165,31 +129,21 @@ impl<T> OnceCell<T> {
         Self::set0(value, guard)
     }
 
-    fn set0(value: T, mut guard: tokio::sync::RwLockWriteGuard<'_, Inner<T>>) -> GuardMut<'_, T> {
+    fn set0(value: T, mut guard: std::sync::MutexGuard<'_, Inner<T>>) -> Guard<'_, T> {
         if guard.value.is_some() {
             drop(guard);
             unreachable!("we won permit, must not be initialized");
         }
         guard.value = Some(value);
         guard.init_semaphore.close();
-        GuardMut(guard)
+        Guard(guard)
     }
 
     /// Returns a guard to an existing initialized value, if any.
-    pub async fn get_mut(&self) -> Option<GuardMut<'_, T>> {
-        let guard = self.inner.write().await;
+    pub fn get(&self) -> Option<Guard<'_, T>> {
+        let guard = self.inner.lock().unwrap();
         if guard.value.is_some() {
-            Some(GuardMut(guard))
-        } else {
-            None
-        }
-    }
-
-    /// Returns a guard to an existing initialized value, if any.
-    pub async fn get(&self) -> Option<GuardRef<'_, T>> {
-        let guard = self.inner.read().await;
-        if guard.value.is_some() {
-            Some(GuardRef(guard))
+            Some(Guard(guard))
         } else {
             None
         }
@@ -221,9 +175,9 @@ impl<'a, T> Drop for CountWaitingInitializers<'a, T> {
 /// Uninteresting guard object to allow short-lived access to inspect or clone the held,
 /// initialized value.
 #[derive(Debug)]
-pub struct GuardMut<'a, T>(tokio::sync::RwLockWriteGuard<'a, Inner<T>>);
+pub struct Guard<'a, T>(MutexGuard<'a, Inner<T>>);
 
-impl<T> std::ops::Deref for GuardMut<'_, T> {
+impl<T> std::ops::Deref for Guard<'_, T> {
     type Target = T;
 
     fn deref(&self) -> &Self::Target {
@@ -234,7 +188,7 @@ impl<T> std::ops::Deref for GuardMut<'_, T> {
     }
 }
 
-impl<T> std::ops::DerefMut for GuardMut<'_, T> {
+impl<T> std::ops::DerefMut for Guard<'_, T> {
     fn deref_mut(&mut self) -> &mut Self::Target {
         self.0
             .value
@@ -243,52 +197,48 @@ impl<T> std::ops::DerefMut for GuardMut<'_, T> {
     }
 }
 
-impl<'a, T> GuardMut<'a, T> {
+impl<'a, T> Guard<'a, T> {
     /// Take the current value, and a new permit for it's deinitialization.
     ///
     /// The permit will be on a semaphore part of the new internal value, and any following
     /// [`OnceCell::get_or_init`] will wait on it to complete.
     pub fn take_and_deinit(&mut self) -> (T, InitPermit) {
         let mut swapped = Inner::default();
-        let permit = swapped
-            .init_semaphore
-            .clone()
-            .try_acquire_owned()
-            .expect("we just created this");
+        let sem = swapped.init_semaphore.clone();
+        // acquire and forget right away, moving the control over to InitPermit
+        sem.try_acquire().expect("we just created this").forget();
         std::mem::swap(&mut *self.0, &mut swapped);
         swapped
             .value
-            .map(|v| (v, InitPermit(permit)))
-            .expect("guard is not created unless value has been initialized")
-    }
-
-    pub fn downgrade(self) -> GuardRef<'a, T> {
-        GuardRef(self.0.downgrade())
-    }
-}
-
-#[derive(Debug)]
-pub struct GuardRef<'a, T>(tokio::sync::RwLockReadGuard<'a, Inner<T>>);
-
-impl<T> std::ops::Deref for GuardRef<'_, T> {
-    type Target = T;
-
-    fn deref(&self) -> &Self::Target {
-        self.0
-            .value
-            .as_ref()
+            .map(|v| (v, InitPermit(sem)))
             .expect("guard is not created unless value has been initialized")
     }
 }
 
 /// Type held by OnceCell (de)initializing task.
-pub struct InitPermit(tokio::sync::OwnedSemaphorePermit);
+///
+/// On drop, this type will return the permit.
+pub struct InitPermit(Arc<tokio::sync::Semaphore>);
+
+impl Drop for InitPermit {
+    fn drop(&mut self) {
+        assert_eq!(
+            self.0.available_permits(),
+            0,
+            "InitPermit should only exist as the unique permit"
+        );
+        self.0.add_permits(1);
+    }
+}
 
 #[cfg(test)]
 mod tests {
+    use futures::Future;
+
     use super::*;
     use std::{
         convert::Infallible,
+        pin::{pin, Pin},
         sync::atomic::{AtomicUsize, Ordering},
         time::Duration,
     };
@@ -319,7 +269,7 @@ mod tests {
                     barrier.wait().await;
                     let won = {
                         let g = cell
-                            .get_mut_or_init(|permit| {
+                            .get_or_init(|permit| {
                                 counters.factory_got_to_run.fetch_add(1, Ordering::Relaxed);
                                 async {
                                     counters.future_polled.fetch_add(1, Ordering::Relaxed);
@@ -366,11 +316,7 @@ mod tests {
             let cell = cell.clone();
             let deinitialization_started = deinitialization_started.clone();
             async move {
-                let (answer, _permit) = cell
-                    .get_mut()
-                    .await
-                    .expect("initialized to value")
-                    .take_and_deinit();
+                let (answer, _permit) = cell.get().expect("initialized to value").take_and_deinit();
                 assert_eq!(answer, initial);
 
                 deinitialization_started.wait().await;
@@ -381,7 +327,7 @@ mod tests {
         deinitialization_started.wait().await;
 
         let started_at = tokio::time::Instant::now();
-        cell.get_mut_or_init(|permit| async { Ok::<_, Infallible>((reinit, permit)) })
+        cell.get_or_init(|permit| async { Ok::<_, Infallible>((reinit, permit)) })
             .await
             .unwrap();
 
@@ -393,21 +339,21 @@ mod tests {
 
         jh.await.unwrap();
 
-        assert_eq!(*cell.get_mut().await.unwrap(), reinit);
+        assert_eq!(*cell.get().unwrap(), reinit);
     }
 
-    #[tokio::test]
-    async fn reinit_with_deinit_permit() {
+    #[test]
+    fn reinit_with_deinit_permit() {
         let cell = Arc::new(OnceCell::new(42));
 
-        let (mol, permit) = cell.get_mut().await.unwrap().take_and_deinit();
-        cell.set(5, permit).await;
-        assert_eq!(*cell.get_mut().await.unwrap(), 5);
+        let (mol, permit) = cell.get().unwrap().take_and_deinit();
+        cell.set(5, permit);
+        assert_eq!(*cell.get().unwrap(), 5);
 
-        let (five, permit) = cell.get_mut().await.unwrap().take_and_deinit();
+        let (five, permit) = cell.get().unwrap().take_and_deinit();
         assert_eq!(5, five);
-        cell.set(mol, permit).await;
-        assert_eq!(*cell.get_mut().await.unwrap(), 42);
+        cell.set(mol, permit);
+        assert_eq!(*cell.get().unwrap(), 42);
     }
 
     #[tokio::test]
@@ -415,13 +361,13 @@ mod tests {
         let cell = OnceCell::default();
 
         for _ in 0..10 {
-            cell.get_mut_or_init(|_permit| async { Err("whatever error") })
+            cell.get_or_init(|_permit| async { Err("whatever error") })
                 .await
                 .unwrap_err();
         }
 
         let g = cell
-            .get_mut_or_init(|permit| async { Ok::<_, Infallible>(("finally success", permit)) })
+            .get_or_init(|permit| async { Ok::<_, Infallible>(("finally success", permit)) })
             .await
             .unwrap();
         assert_eq!(*g, "finally success");
@@ -433,7 +379,7 @@ mod tests {
 
         let barrier = tokio::sync::Barrier::new(2);
 
-        let initializer = cell.get_mut_or_init(|permit| async {
+        let initializer = cell.get_or_init(|permit| async {
             barrier.wait().await;
             futures::future::pending::<()>().await;
 
@@ -447,12 +393,93 @@ mod tests {
 
         // now initializer is dropped
 
-        assert!(cell.get_mut().await.is_none());
+        assert!(cell.get().is_none());
 
         let g = cell
-            .get_mut_or_init(|permit| async { Ok::<_, Infallible>(("now initialized", permit)) })
+            .get_or_init(|permit| async { Ok::<_, Infallible>(("now initialized", permit)) })
             .await
             .unwrap();
         assert_eq!(*g, "now initialized");
     }
+
+    #[tokio::test(start_paused = true)]
+    async fn reproduce_init_take_deinit_race() {
+        init_take_deinit_scenario(|cell, factory| {
+            Box::pin(async {
+                cell.get_or_init(factory).await.unwrap();
+            })
+        })
+        .await;
+    }
+
+    type BoxedInitFuture<T, E> = Pin<Box<dyn Future<Output = Result<(T, InitPermit), E>>>>;
+    type BoxedInitFunction<T, E> = Box<dyn Fn(InitPermit) -> BoxedInitFuture<T, E>>;
+
+    /// Reproduce an assertion failure.
+    ///
+    /// This has interesting generics to be generic between `get_or_init` and `get_mut_or_init`.
+    /// We currently only have one, but the structure is kept.
+    async fn init_take_deinit_scenario<F>(init_way: F)
+    where
+        F: for<'a> Fn(
+            &'a OnceCell<&'static str>,
+            BoxedInitFunction<&'static str, Infallible>,
+        ) -> Pin<Box<dyn Future<Output = ()> + 'a>>,
+    {
+        let cell = OnceCell::default();
+
+        // acquire the init_semaphore only permit to drive initializing tasks in order to waiting
+        // on the same semaphore.
+        let permit = cell
+            .inner
+            .lock()
+            .unwrap()
+            .init_semaphore
+            .clone()
+            .try_acquire_owned()
+            .unwrap();
+
+        let mut t1 = pin!(init_way(
+            &cell,
+            Box::new(|permit| Box::pin(async move { Ok(("t1", permit)) })),
+        ));
+
+        let mut t2 = pin!(init_way(
+            &cell,
+            Box::new(|permit| Box::pin(async move { Ok(("t2", permit)) })),
+        ));
+
+        // drive t2 first to the init_semaphore -- the timeout will be hit once t2 future can
+        // no longer make progress
+        tokio::select! {
+            _ = &mut t2 => unreachable!("it cannot get permit"),
+            _ = tokio::time::sleep(Duration::from_secs(3600 * 24 * 7 * 365)) => {}
+        }
+
+        // followed by t1 in the init_semaphore
+        tokio::select! {
+            _ = &mut t1 => unreachable!("it cannot get permit"),
+            _ = tokio::time::sleep(Duration::from_secs(3600 * 24 * 7 * 365)) => {}
+        }
+
+        // now let t2 proceed and initialize
+        drop(permit);
+        t2.await;
+
+        let (s, permit) = { cell.get().unwrap().take_and_deinit() };
+        assert_eq!("t2", s);
+
+        // now originally t1 would see the semaphore it has as closed. it cannot yet get a permit from
+        // the new one.
+        tokio::select! {
+            _ = &mut t1 => unreachable!("it cannot get permit"),
+            _ = tokio::time::sleep(Duration::from_secs(3600 * 24 * 7 * 365)) => {}
+        }
+
+        // only now we get to initialize it
+        drop(permit);
+        t1.await;
+
+        assert_eq!("t1", *cell.get().unwrap());
+    }
 }

libs/walproposer/build.rs

@@ -34,6 +34,9 @@ fn main() -> anyhow::Result<()> {
     println!("cargo:rustc-link-lib=static=walproposer");
     println!("cargo:rustc-link-search={walproposer_lib_search_str}");
 
+    // Rebuild crate when libwalproposer.a changes
+    println!("cargo:rerun-if-changed={walproposer_lib_search_str}/libwalproposer.a");
+
     let pg_config_bin = pg_install_abs.join("v16").join("bin").join("pg_config");
     let inc_server_path: String = if pg_config_bin.exists() {
         let output = Command::new(pg_config_bin)
@@ -79,6 +82,7 @@ fn main() -> anyhow::Result<()> {
         .allowlist_function("WalProposerBroadcast")
         .allowlist_function("WalProposerPoll")
         .allowlist_function("WalProposerFree")
+        .allowlist_function("SafekeeperStateDesiredEvents")
         .allowlist_var("DEBUG5")
         .allowlist_var("DEBUG4")
         .allowlist_var("DEBUG3")

libs/walproposer/src/api_bindings.rs

@@ -22,6 +22,7 @@ use crate::bindings::WalProposerExecStatusType;
 use crate::bindings::WalproposerShmemState;
 use crate::bindings::XLogRecPtr;
 use crate::walproposer::ApiImpl;
+use crate::walproposer::StreamingCallback;
 use crate::walproposer::WaitResult;
 
 extern "C" fn get_shmem_state(wp: *mut WalProposer) -> *mut WalproposerShmemState {
@@ -36,7 +37,8 @@ extern "C" fn start_streaming(wp: *mut WalProposer, startpos: XLogRecPtr) {
     unsafe {
         let callback_data = (*(*wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
-        (*api).start_streaming(startpos)
+        let callback = StreamingCallback::new(wp);
+        (*api).start_streaming(startpos, &callback);
     }
 }
 
@@ -134,19 +136,18 @@ extern "C" fn conn_async_read(
     unsafe {
         let callback_data = (*(*(*sk).wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
-        let (res, result) = (*api).conn_async_read(&mut (*sk));
 
         // This function has guarantee that returned buf will be valid until
         // the next call. So we can store a Vec in each Safekeeper and reuse
         // it on the next call.
         let mut inbuf = take_vec_u8(&mut (*sk).inbuf).unwrap_or_default();
-
         inbuf.clear();
-        inbuf.extend_from_slice(res);
+
+        let result = (*api).conn_async_read(&mut (*sk), &mut inbuf);
 
         // Put a Vec back to sk->inbuf and return data ptr.
+        *amount = inbuf.len() as i32;
         *buf = store_vec_u8(&mut (*sk).inbuf, inbuf);
-        *amount = res.len() as i32;
 
         result
     }
@@ -182,6 +183,10 @@ extern "C" fn recovery_download(wp: *mut WalProposer, sk: *mut Safekeeper) -> bo
     unsafe {
         let callback_data = (*(*(*sk).wp).config).callback_data;
         let api = callback_data as *mut Box<dyn ApiImpl>;
+
+        // currently `recovery_download` is always called right after election
+        (*api).after_election(&mut (*wp));
+
         (*api).recovery_download(&mut (*wp), &mut (*sk))
     }
 }
@@ -277,7 +282,8 @@ extern "C" fn wait_event_set(
             }
             WaitResult::Timeout => {
                 *event_sk = std::ptr::null_mut();
-                *events = crate::bindings::WL_TIMEOUT;
+                // WaitEventSetWait returns 0 for timeout.
+                *events = 0;
                 0
             }
             WaitResult::Network(sk, event_mask) => {
@@ -340,7 +346,7 @@ extern "C" fn log_internal(
     }
 }
 
-#[derive(Debug)]
+#[derive(Debug, PartialEq)]
 pub enum Level {
     Debug5,
     Debug4,

libs/walproposer/src/walproposer.rs

@@ -1,13 +1,13 @@
 use std::ffi::CString;
 
 use postgres_ffi::WAL_SEGMENT_SIZE;
-use utils::id::TenantTimelineId;
+use utils::{id::TenantTimelineId, lsn::Lsn};
 
 use crate::{
     api_bindings::{create_api, take_vec_u8, Level},
     bindings::{
-        NeonWALReadResult, Safekeeper, WalProposer, WalProposerConfig, WalProposerCreate,
-        WalProposerFree, WalProposerStart,
+        NeonWALReadResult, Safekeeper, WalProposer, WalProposerBroadcast, WalProposerConfig,
+        WalProposerCreate, WalProposerFree, WalProposerPoll, WalProposerStart,
     },
 };
 
@@ -16,11 +16,11 @@ use crate::{
 ///
 /// Refer to `pgxn/neon/walproposer.h` for documentation.
 pub trait ApiImpl {
-    fn get_shmem_state(&self) -> &mut crate::bindings::WalproposerShmemState {
+    fn get_shmem_state(&self) -> *mut crate::bindings::WalproposerShmemState {
         todo!()
     }
 
-    fn start_streaming(&self, _startpos: u64) {
+    fn start_streaming(&self, _startpos: u64, _callback: &StreamingCallback) {
         todo!()
     }
 
@@ -70,7 +70,11 @@ pub trait ApiImpl {
         todo!()
     }
 
-    fn conn_async_read(&self, _sk: &mut Safekeeper) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+    fn conn_async_read(
+        &self,
+        _sk: &mut Safekeeper,
+        _vec: &mut Vec<u8>,
+    ) -> crate::bindings::PGAsyncReadResult {
         todo!()
     }
 
@@ -151,12 +155,14 @@ pub trait ApiImpl {
     }
 }
 
+#[derive(Debug)]
 pub enum WaitResult {
     Latch,
     Timeout,
     Network(*mut Safekeeper, u32),
 }
 
+#[derive(Clone)]
 pub struct Config {
     /// Tenant and timeline id
     pub ttid: TenantTimelineId,
@@ -242,6 +248,24 @@ impl Drop for Wrapper {
     }
 }
 
+pub struct StreamingCallback {
+    wp: *mut WalProposer,
+}
+
+impl StreamingCallback {
+    pub fn new(wp: *mut WalProposer) -> StreamingCallback {
+        StreamingCallback { wp }
+    }
+
+    pub fn broadcast(&self, startpos: Lsn, endpos: Lsn) {
+        unsafe { WalProposerBroadcast(self.wp, startpos.0, endpos.0) }
+    }
+
+    pub fn poll(&self) {
+        unsafe { WalProposerPoll(self.wp) }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use core::panic;
@@ -344,14 +368,13 @@ mod tests {
         fn conn_async_read(
             &self,
             _: &mut crate::bindings::Safekeeper,
-        ) -> (&[u8], crate::bindings::PGAsyncReadResult) {
+            vec: &mut Vec<u8>,
+        ) -> crate::bindings::PGAsyncReadResult {
             println!("conn_async_read");
             let reply = self.next_safekeeper_reply();
             println!("conn_async_read result: {:?}", reply);
-            (
-                reply,
-                crate::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS,
-            )
+            vec.extend_from_slice(reply);
+            crate::bindings::PGAsyncReadResult_PG_ASYNC_READ_SUCCESS
         }
 
         fn conn_blocking_write(&self, _: &mut crate::bindings::Safekeeper, buf: &[u8]) -> bool {
@@ -453,9 +476,12 @@ mod tests {
                 event_mask: 0,
             }),
             expected_messages: vec![
-                // Greeting(ProposerGreeting { protocol_version: 2, pg_version: 160001, proposer_id: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], system_id: 0, timeline_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tenant_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tli: 1, wal_seg_size: 16777216 })
+                // TODO: When updating Postgres versions, this test will cause
+                // problems. Postgres version in message needs updating.
+                //
+                // Greeting(ProposerGreeting { protocol_version: 2, pg_version: 160002, proposer_id: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], system_id: 0, timeline_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tenant_id: 9e4c8f36063c6c6e93bc20d65a820f3d, tli: 1, wal_seg_size: 16777216 })
                 vec![
-                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 1, 113, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                    103, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 2, 113, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
                     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 158, 76, 143, 54, 6, 60, 108, 110,
                     147, 188, 32, 214, 90, 130, 15, 61, 158, 76, 143, 54, 6, 60, 108, 110, 147,
                     188, 32, 214, 90, 130, 15, 61, 1, 0, 0, 0, 0, 0, 0, 1,

pageserver/Cargo.toml

@@ -21,6 +21,7 @@ camino.workspace = true
 camino-tempfile.workspace = true
 chrono = { workspace = true, features = ["serde"] }
 clap = { workspace = true, features = ["string"] }
+close_fds.workspace = true
 const_format.workspace = true
 consumption_metrics.workspace = true
 crc32c.workspace = true

pageserver/client/src/mgmt_api.rs

@@ -56,10 +56,18 @@ pub enum ForceAwaitLogicalSize {
 
 impl Client {
     pub fn new(mgmt_api_endpoint: String, jwt: Option<&str>) -> Self {
+        Self::from_client(reqwest::Client::new(), mgmt_api_endpoint, jwt)
+    }
+
+    pub fn from_client(
+        client: reqwest::Client,
+        mgmt_api_endpoint: String,
+        jwt: Option<&str>,
+    ) -> Self {
         Self {
             mgmt_api_endpoint,
             authorization_header: jwt.map(|jwt| format!("Bearer {jwt}")),
-            client: reqwest::Client::new(),
+            client,
         }
     }
 
@@ -310,6 +318,22 @@ impl Client {
             .map_err(Error::ReceiveBody)
     }
 
+    pub async fn tenant_shard_split(
+        &self,
+        tenant_shard_id: TenantShardId,
+        req: TenantShardSplitRequest,
+    ) -> Result<TenantShardSplitResponse> {
+        let uri = format!(
+            "{}/v1/tenant/{}/shard_split",
+            self.mgmt_api_endpoint, tenant_shard_id
+        );
+        self.request(Method::PUT, &uri, req)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
+
     pub async fn timeline_list(
         &self,
         tenant_shard_id: &TenantShardId,
@@ -339,4 +363,16 @@ impl Client {
             .await
             .map_err(Error::ReceiveBody)
     }
+
+    pub async fn put_io_engine(
+        &self,
+        engine: &pageserver_api::models::virtual_file::IoEngineKind,
+    ) -> Result<()> {
+        let uri = format!("{}/v1/io_engine", self.mgmt_api_endpoint);
+        self.request(Method::PUT, uri, engine)
+            .await?
+            .json()
+            .await
+            .map_err(Error::ReceiveBody)
+    }
 }

pageserver/ctl/src/layer_map_analyzer.rs

@@ -142,7 +142,7 @@ pub(crate) async fn main(cmd: &AnalyzeLayerMapCmd) -> Result<()> {
     let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
 
     // Initialize virtual_file (file desriptor cache) and page cache which are needed to access layer persistent B-Tree.
-    pageserver::virtual_file::init(10, virtual_file::IoEngineKind::StdFs);
+    pageserver::virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
     pageserver::page_cache::init(100);
 
     let mut total_delta_layers = 0usize;

pageserver/ctl/src/layers.rs

@@ -59,7 +59,7 @@ pub(crate) enum LayerCmd {
 
 async fn read_delta_file(path: impl AsRef<Path>, ctx: &RequestContext) -> Result<()> {
     let path = Utf8Path::from_path(path.as_ref()).expect("non-Unicode path");
-    virtual_file::init(10, virtual_file::IoEngineKind::StdFs);
+    virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
     page_cache::init(100);
     let file = FileBlockReader::new(VirtualFile::open(path).await?);
     let summary_blk = file.read_blk(0, ctx).await?;
@@ -187,7 +187,7 @@ pub(crate) async fn main(cmd: &LayerCmd) -> Result<()> {
             new_tenant_id,
             new_timeline_id,
         } => {
-            pageserver::virtual_file::init(10, virtual_file::IoEngineKind::StdFs);
+            pageserver::virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
             pageserver::page_cache::init(100);
 
             let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);

pageserver/ctl/src/main.rs

@@ -123,7 +123,7 @@ fn read_pg_control_file(control_file_path: &Utf8Path) -> anyhow::Result<()> {
 
 async fn print_layerfile(path: &Utf8Path) -> anyhow::Result<()> {
     // Basic initialization of things that don't change after startup
-    virtual_file::init(10, virtual_file::IoEngineKind::StdFs);
+    virtual_file::init(10, virtual_file::api::IoEngineKind::StdFs);
     page_cache::init(100);
     let ctx = RequestContext::new(TaskKind::DebugTool, DownloadBehavior::Error);
     dump_layerfile_from_path(path, true, &ctx).await

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -51,6 +51,10 @@ pub(crate) struct Args {
     /// It doesn't get invalidated if the keyspace changes under the hood, e.g., due to new ingested data or compaction.
     #[clap(long)]
     keyspace_cache: Option<Utf8PathBuf>,
+    /// Before starting the benchmark, live-reconfigure the pageserver to use the given
+    /// [`pageserver_api::models::virtual_file::IoEngineKind`].
+    #[clap(long)]
+    set_io_engine: Option<pageserver_api::models::virtual_file::IoEngineKind>,
     targets: Option<Vec<TenantTimelineId>>,
 }
 
@@ -109,6 +113,10 @@ async fn main_impl(
         args.pageserver_jwt.as_deref(),
     ));
 
+    if let Some(engine_str) = &args.set_io_engine {
+        mgmt_api_client.put_io_engine(engine_str).await?;
+    }
+
     // discover targets
     let timelines: Vec<TenantTimelineId> = crate::util::cli::targets::discover(
         &mgmt_api_client,
@@ -288,11 +296,7 @@ async fn main_impl(
                                 num_client: rng.gen_range(0..args.num_clients.get()),
                             },
                             PagestreamGetPageRequest {
-                                horizon: if rng.gen_bool(args.req_latest_probability) {
-                                    Lsn::MAX
-                                } else {
-                                    r.timeline_lsn
-                                },
+                                latest: rng.gen_bool(args.req_latest_probability),
                                 lsn: r.timeline_lsn,
                                 rel: rel_tag,
                                 blkno: block_no,
@@ -339,11 +343,7 @@ async fn main_impl(
                                     let (rel_tag, block_no) = key_to_rel_block(key)
                                         .expect("we filter non-rel-block keys out above");
                                     PagestreamGetPageRequest {
-                                        horizon: if rng.gen_bool(args.req_latest_probability) {
-                                            Lsn::MAX
-                                        } else {
-                                            r.timeline_lsn
-                                        },
+                                        latest: rng.gen_bool(args.req_latest_probability),
                                         lsn: r.timeline_lsn,
                                         rel: rel_tag,
                                         blkno: block_no,

pageserver/src/basebackup.rs

@@ -357,10 +357,9 @@ where
 
     /// Add contents of relfilenode `src`, naming it as `dst`.
     async fn add_rel(&mut self, src: RelTag, dst: RelTag) -> anyhow::Result<()> {
-        let horizon = self.lsn; // we do not need latest version
         let nblocks = self
             .timeline
-            .get_rel_size(src, Version::Lsn(self.lsn), horizon, self.ctx)
+            .get_rel_size(src, Version::Lsn(self.lsn), false, self.ctx)
             .await?;
 
         // If the relation is empty, create an empty file
@@ -381,7 +380,7 @@ where
             for blknum in startblk..endblk {
                 let img = self
                     .timeline
-                    .get_rel_page_at_lsn(src, blknum, Version::Lsn(self.lsn), horizon, self.ctx)
+                    .get_rel_page_at_lsn(src, blknum, Version::Lsn(self.lsn), false, self.ctx)
                     .await?;
                 segment_data.extend_from_slice(&img[..]);
             }

pageserver/src/bin/pageserver.rs

@@ -272,6 +272,12 @@ fn start_pageserver(
     );
     set_build_info_metric(GIT_VERSION, BUILD_TAG);
     set_launch_timestamp_metric(launch_ts);
+    #[cfg(target_os = "linux")]
+    metrics::register_internal(Box::new(metrics::more_process_metrics::Collector::new())).unwrap();
+    metrics::register_internal(Box::new(
+        pageserver::metrics::tokio_epoll_uring::Collector::new(),
+    ))
+    .unwrap();
     pageserver::preinitialize_metrics();
 
     // If any failpoints were set from FAILPOINTS environment variable,

pageserver/src/config.rs

@@ -1359,6 +1359,7 @@ broker_endpoint = '{broker_endpoint}'
                 parsed_remote_storage_config,
                 RemoteStorageConfig {
                     storage: RemoteStorageKind::LocalFs(local_storage_path.clone()),
+                    timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
                 },
                 "Remote storage config should correctly parse the local FS config and fill other storage defaults"
             );
@@ -1426,6 +1427,7 @@ broker_endpoint = '{broker_endpoint}'
                         concurrency_limit: s3_concurrency_limit,
                         max_keys_per_list_response: None,
                     }),
+                    timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
                 },
                 "Remote storage config should correctly parse the S3 config"
             );

pageserver/src/deletion_queue.rs

@@ -234,7 +234,7 @@ impl DeletionHeader {
         let header_bytes = serde_json::to_vec(self).context("serialize deletion header")?;
         let header_path = conf.deletion_header_path();
         let temp_path = path_with_suffix_extension(&header_path, TEMP_SUFFIX);
-        VirtualFile::crashsafe_overwrite(&header_path, &temp_path, &header_bytes)
+        VirtualFile::crashsafe_overwrite(&header_path, &temp_path, header_bytes)
             .await
             .maybe_fatal_err("save deletion header")?;
 
@@ -325,7 +325,7 @@ impl DeletionList {
         let temp_path = path_with_suffix_extension(&path, TEMP_SUFFIX);
 
         let bytes = serde_json::to_vec(self).expect("Failed to serialize deletion list");
-        VirtualFile::crashsafe_overwrite(&path, &temp_path, &bytes)
+        VirtualFile::crashsafe_overwrite(&path, &temp_path, bytes)
             .await
             .maybe_fatal_err("save deletion list")
             .map_err(Into::into)
@@ -834,7 +834,6 @@ mod test {
     }
 
     impl ControlPlaneGenerationsApi for MockControlPlane {
-        #[allow(clippy::diverging_sub_expression)] // False positive via async_trait
         async fn re_attach(&self) -> Result<HashMap<TenantShardId, Generation>, RetryForeverError> {
             unimplemented!()
         }
@@ -868,6 +867,7 @@ mod test {
         let remote_fs_dir = harness.conf.workdir.join("remote_fs").canonicalize_utf8()?;
         let storage_config = RemoteStorageConfig {
             storage: RemoteStorageKind::LocalFs(remote_fs_dir.clone()),
+            timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
         };
         let storage = GenericRemoteStorage::from_config(&storage_config).unwrap();
 
@@ -1171,6 +1171,7 @@ pub(crate) mod mock {
     pub struct ConsumerState {
         rx: tokio::sync::mpsc::UnboundedReceiver<ListWriterQueueMessage>,
         executor_rx: tokio::sync::mpsc::Receiver<DeleterMessage>,
+        cancel: CancellationToken,
     }
 
     impl ConsumerState {
@@ -1184,7 +1185,7 @@ pub(crate) mod mock {
                 match msg {
                     DeleterMessage::Delete(objects) => {
                         for path in objects {
-                            match remote_storage.delete(&path).await {
+                            match remote_storage.delete(&path, &self.cancel).await {
                                 Ok(_) => {
                                     debug!("Deleted {path}");
                                 }
@@ -1217,7 +1218,7 @@ pub(crate) mod mock {
 
                         for path in objects {
                             info!("Executing deletion {path}");
-                            match remote_storage.delete(&path).await {
+                            match remote_storage.delete(&path, &self.cancel).await {
                                 Ok(_) => {
                                     debug!("Deleted {path}");
                                 }
@@ -1267,7 +1268,11 @@ pub(crate) mod mock {
                 executor_tx,
                 executed,
                 remote_storage,
-                consumer: std::sync::Mutex::new(ConsumerState { rx, executor_rx }),
+                consumer: std::sync::Mutex::new(ConsumerState {
+                    rx,
+                    executor_rx,
+                    cancel: CancellationToken::new(),
+                }),
                 lsn_table: Arc::new(std::sync::RwLock::new(VisibleLsnUpdates::new())),
             }
         }

pageserver/src/deletion_queue/deleter.rs

@@ -8,6 +8,7 @@
 
 use remote_storage::GenericRemoteStorage;
 use remote_storage::RemotePath;
+use remote_storage::TimeoutOrCancel;
 use remote_storage::MAX_KEYS_PER_DELETE;
 use std::time::Duration;
 use tokio_util::sync::CancellationToken;
@@ -71,9 +72,11 @@ impl Deleter {
                     Err(anyhow::anyhow!("failpoint: deletion-queue-before-execute"))
                 });
 
-                self.remote_storage.delete_objects(&self.accumulator).await
+                self.remote_storage
+                    .delete_objects(&self.accumulator, &self.cancel)
+                    .await
             },
-            |_| false,
+            TimeoutOrCancel::caused_by_cancel,
             3,
             10,
             "executing deletion batch",

pageserver/src/disk_usage_eviction_task.rs

@@ -351,7 +351,6 @@ pub enum IterationOutcome<U> {
     Finished(IterationOutcomeFinished<U>),
 }
 
-#[allow(dead_code)]
 #[derive(Debug, Serialize)]
 pub struct IterationOutcomeFinished<U> {
     /// The actual usage observed before we started the iteration.
@@ -366,7 +365,6 @@ pub struct IterationOutcomeFinished<U> {
 }
 
 #[derive(Debug, Serialize)]
-#[allow(dead_code)]
 struct AssumedUsage<U> {
     /// The expected value for `after`, after phase 2.
     projected_after: U,
@@ -374,14 +372,12 @@ struct AssumedUsage<U> {
     failed: LayerCount,
 }
 
-#[allow(dead_code)]
 #[derive(Debug, Serialize)]
 struct PlannedUsage<U> {
     respecting_tenant_min_resident_size: U,
     fallback_to_global_lru: Option<U>,
 }
 
-#[allow(dead_code)]
 #[derive(Debug, Default, Serialize)]
 struct LayerCount {
     file_sizes: u64,
@@ -565,7 +561,6 @@ pub(crate) struct EvictionSecondaryLayer {
 #[derive(Clone)]
 pub(crate) enum EvictionLayer {
     Attached(Layer),
-    #[allow(dead_code)]
     Secondary(EvictionSecondaryLayer),
 }
 
@@ -623,6 +618,7 @@ impl std::fmt::Display for EvictionLayer {
     }
 }
 
+#[derive(Default)]
 pub(crate) struct DiskUsageEvictionInfo {
     /// Timeline's largest layer (remote or resident)
     pub max_layer_size: Option<u64>,
@@ -854,19 +850,27 @@ async fn collect_eviction_candidates(
 
         let total = tenant_candidates.len();
 
-        for (i, mut candidate) in tenant_candidates.into_iter().enumerate() {
-            // as we iterate this reverse sorted list, the most recently accessed layer will always
-            // be 1.0; this is for us to evict it last.
-            candidate.relative_last_activity = eviction_order.relative_last_activity(total, i);
+        let tenant_candidates =
+            tenant_candidates
+                .into_iter()
+                .enumerate()
+                .map(|(i, mut candidate)| {
+                    // as we iterate this reverse sorted list, the most recently accessed layer will always
+                    // be 1.0; this is for us to evict it last.
+                    candidate.relative_last_activity =
+                        eviction_order.relative_last_activity(total, i);
 
-            let partition = if cumsum > min_resident_size as i128 {
-                MinResidentSizePartition::Above
-            } else {
-                MinResidentSizePartition::Below
-            };
-            cumsum += i128::from(candidate.layer.get_file_size());
-            candidates.push((partition, candidate));
-        }
+                    let partition = if cumsum > min_resident_size as i128 {
+                        MinResidentSizePartition::Above
+                    } else {
+                        MinResidentSizePartition::Below
+                    };
+                    cumsum += i128::from(candidate.layer.get_file_size());
+
+                    (partition, candidate)
+                });
+
+        candidates.extend(tenant_candidates);
     }
 
     // Note: the same tenant ID might be hit twice, if it transitions from attached to
@@ -882,21 +886,41 @@ async fn collect_eviction_candidates(
     );
 
     for secondary_tenant in secondary_tenants {
-        let mut layer_info = secondary_tenant.get_layers_for_eviction();
+        // for secondary tenants we use a sum of on_disk layers and already evicted layers. this is
+        // to prevent repeated disk usage based evictions from completely draining less often
+        // updating secondaries.
+        let (mut layer_info, total_layers) = secondary_tenant.get_layers_for_eviction();
+
+        debug_assert!(
+            total_layers >= layer_info.resident_layers.len(),
+            "total_layers ({total_layers}) must be at least the resident_layers.len() ({})",
+            layer_info.resident_layers.len()
+        );
 
         layer_info
             .resident_layers
             .sort_unstable_by_key(|layer_info| std::cmp::Reverse(layer_info.last_activity_ts));
 
-        candidates.extend(layer_info.resident_layers.into_iter().map(|candidate| {
-            (
-                // Secondary locations' layers are always considered above the min resident size,
-                // i.e. secondary locations are permitted to be trimmed to zero layers if all
-                // the layers have sufficiently old access times.
-                MinResidentSizePartition::Above,
-                candidate,
-            )
-        }));
+        let tenant_candidates =
+            layer_info
+                .resident_layers
+                .into_iter()
+                .enumerate()
+                .map(|(i, mut candidate)| {
+                    candidate.relative_last_activity =
+                        eviction_order.relative_last_activity(total_layers, i);
+                    (
+                        // Secondary locations' layers are always considered above the min resident size,
+                        // i.e. secondary locations are permitted to be trimmed to zero layers if all
+                        // the layers have sufficiently old access times.
+                        MinResidentSizePartition::Above,
+                        candidate,
+                    )
+                });
+
+        candidates.extend(tenant_candidates);
+
+        tokio::task::yield_now().await;
     }
 
     debug_assert!(MinResidentSizePartition::Above < MinResidentSizePartition::Below,
@@ -1076,7 +1100,6 @@ mod filesystem_level_usage {
     use super::DiskUsageEvictionTaskConfig;
 
     #[derive(Debug, Clone, Copy)]
-    #[allow(dead_code)]
     pub struct Usage<'a> {
         config: &'a DiskUsageEvictionTaskConfig,
 

pageserver/src/http/routes.rs

@@ -19,11 +19,14 @@ use pageserver_api::models::ShardParameters;
 use pageserver_api::models::TenantDetails;
 use pageserver_api::models::TenantLocationConfigResponse;
 use pageserver_api::models::TenantShardLocation;
+use pageserver_api::models::TenantShardSplitRequest;
+use pageserver_api::models::TenantShardSplitResponse;
 use pageserver_api::models::TenantState;
 use pageserver_api::models::{
     DownloadRemoteLayersTaskSpawnRequest, LocationConfigMode, TenantAttachRequest,
     TenantLoadRequest, TenantLocationConfigRequest,
 };
+use pageserver_api::shard::ShardCount;
 use pageserver_api::shard::TenantShardId;
 use remote_storage::GenericRemoteStorage;
 use remote_storage::TimeTravelError;
@@ -419,6 +422,7 @@ async fn build_timeline_info_common(
             tenant::timeline::logical_size::Accuracy::Approximate => false,
             tenant::timeline::logical_size::Accuracy::Exact => true,
         },
+        directory_entries_counts: timeline.get_directory_metrics().to_vec(),
         current_physical_size,
         current_logical_size_non_incremental: None,
         timeline_dir_layer_file_size_sum: None,
@@ -485,7 +489,9 @@ async fn timeline_create_handler(
     let state = get_state(&request);
 
     async {
-        let tenant = state.tenant_manager.get_attached_tenant_shard(tenant_shard_id, false)?;
+        let tenant = state
+            .tenant_manager
+            .get_attached_tenant_shard(tenant_shard_id, false)?;
 
         tenant.wait_to_become_active(ACTIVE_TENANT_TIMEOUT).await?;
 
@@ -495,48 +501,62 @@ async fn timeline_create_handler(
             tracing::info!("bootstrapping");
         }
 
-        match tenant.create_timeline(
-            new_timeline_id,
-            request_data.ancestor_timeline_id.map(TimelineId::from),
-            request_data.ancestor_start_lsn,
-            request_data.pg_version.unwrap_or(crate::DEFAULT_PG_VERSION),
-            request_data.existing_initdb_timeline_id,
-            state.broker_client.clone(),
-            &ctx,
-        )
-        .await {
+        match tenant
+            .create_timeline(
+                new_timeline_id,
+                request_data.ancestor_timeline_id,
+                request_data.ancestor_start_lsn,
+                request_data.pg_version.unwrap_or(crate::DEFAULT_PG_VERSION),
+                request_data.existing_initdb_timeline_id,
+                state.broker_client.clone(),
+                &ctx,
+            )
+            .await
+        {
             Ok(new_timeline) => {
                 // Created. Construct a TimelineInfo for it.
-                let timeline_info = build_timeline_info_common(&new_timeline, &ctx, tenant::timeline::GetLogicalSizePriority::User)
-                    .await
-                    .map_err(ApiError::InternalServerError)?;
+                let timeline_info = build_timeline_info_common(
+                    &new_timeline,
+                    &ctx,
+                    tenant::timeline::GetLogicalSizePriority::User,
+                )
+                .await
+                .map_err(ApiError::InternalServerError)?;
                 json_response(StatusCode::CREATED, timeline_info)
             }
             Err(_) if tenant.cancel.is_cancelled() => {
                 // In case we get some ugly error type during shutdown, cast it into a clean 503.
-                json_response(StatusCode::SERVICE_UNAVAILABLE, HttpErrorBody::from_msg("Tenant shutting down".to_string()))
-            }
-            Err(tenant::CreateTimelineError::Conflict | tenant::CreateTimelineError::AlreadyCreating) => {
-                json_response(StatusCode::CONFLICT, ())
-            }
-            Err(tenant::CreateTimelineError::AncestorLsn(err)) => {
-                json_response(StatusCode::NOT_ACCEPTABLE, HttpErrorBody::from_msg(
-                    format!("{err:#}")
-                ))
-            }
-            Err(e @ tenant::CreateTimelineError::AncestorNotActive) => {
-                json_response(StatusCode::SERVICE_UNAVAILABLE, HttpErrorBody::from_msg(e.to_string()))
-            }
-            Err(tenant::CreateTimelineError::ShuttingDown) => {
-                json_response(StatusCode::SERVICE_UNAVAILABLE,HttpErrorBody::from_msg("tenant shutting down".to_string()))
+                json_response(
+                    StatusCode::SERVICE_UNAVAILABLE,
+                    HttpErrorBody::from_msg("Tenant shutting down".to_string()),
+                )
             }
+            Err(
+                tenant::CreateTimelineError::Conflict
+                | tenant::CreateTimelineError::AlreadyCreating,
+            ) => json_response(StatusCode::CONFLICT, ()),
+            Err(tenant::CreateTimelineError::AncestorLsn(err)) => json_response(
+                StatusCode::NOT_ACCEPTABLE,
+                HttpErrorBody::from_msg(format!("{err:#}")),
+            ),
+            Err(e @ tenant::CreateTimelineError::AncestorNotActive) => json_response(
+                StatusCode::SERVICE_UNAVAILABLE,
+                HttpErrorBody::from_msg(e.to_string()),
+            ),
+            Err(tenant::CreateTimelineError::ShuttingDown) => json_response(
+                StatusCode::SERVICE_UNAVAILABLE,
+                HttpErrorBody::from_msg("tenant shutting down".to_string()),
+            ),
             Err(tenant::CreateTimelineError::Other(err)) => Err(ApiError::InternalServerError(err)),
         }
     }
     .instrument(info_span!("timeline_create",
         tenant_id = %tenant_shard_id.tenant_id,
         shard_id = %tenant_shard_id.shard_slug(),
-        timeline_id = %new_timeline_id, lsn=?request_data.ancestor_start_lsn, pg_version=?request_data.pg_version))
+        timeline_id = %new_timeline_id,
+        lsn=?request_data.ancestor_start_lsn,
+        pg_version=?request_data.pg_version
+    ))
     .await
 }
 
@@ -875,7 +895,7 @@ async fn tenant_reset_handler(
     let state = get_state(&request);
     state
         .tenant_manager
-        .reset_tenant(tenant_shard_id, drop_cache.unwrap_or(false), ctx)
+        .reset_tenant(tenant_shard_id, drop_cache.unwrap_or(false), &ctx)
         .await
         .map_err(ApiError::InternalServerError)?;
 
@@ -1104,6 +1124,25 @@ async fn tenant_size_handler(
     )
 }
 
+async fn tenant_shard_split_handler(
+    mut request: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    let req: TenantShardSplitRequest = json_request(&mut request).await?;
+
+    let tenant_shard_id: TenantShardId = parse_request_param(&request, "tenant_shard_id")?;
+    let state = get_state(&request);
+    let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Warn);
+
+    let new_shards = state
+        .tenant_manager
+        .shard_split(tenant_shard_id, ShardCount(req.new_shard_count), &ctx)
+        .await
+        .map_err(ApiError::InternalServerError)?;
+
+    json_response(StatusCode::OK, TenantShardSplitResponse { new_shards })
+}
+
 async fn layer_map_info_handler(
     request: Request<Body>,
     _cancel: CancellationToken,
@@ -1908,6 +1947,16 @@ async fn post_tracing_event_handler(
     json_response(StatusCode::OK, ())
 }
 
+async fn put_io_engine_handler(
+    mut r: Request<Body>,
+    _cancel: CancellationToken,
+) -> Result<Response<Body>, ApiError> {
+    check_permission(&r, None)?;
+    let kind: crate::virtual_file::IoEngineKind = json_request(&mut r).await?;
+    crate::virtual_file::io_engine::set(kind);
+    json_response(StatusCode::OK, ())
+}
+
 /// Common functionality of all the HTTP API handlers.
 ///
 /// - Adds a tracing span to each request (by `request_span`)
@@ -2054,6 +2103,9 @@ pub fn make_router(
         .put("/v1/tenant/config", |r| {
             api_handler(r, update_tenant_config_handler)
         })
+        .put("/v1/tenant/:tenant_shard_id/shard_split", |r| {
+            api_handler(r, tenant_shard_split_handler)
+        })
         .get("/v1/tenant/:tenant_shard_id/config", |r| {
             api_handler(r, get_tenant_config_handler)
         })
@@ -2163,7 +2215,8 @@ pub fn make_router(
         )
         .get(
             "/v1/tenant/:tenant_shard_id/timeline/:timeline_id/keyspace",
-            |r| testing_api_handler("read out the keyspace", r, timeline_collect_keyspace),
+            |r| api_handler(r, timeline_collect_keyspace),
         )
+        .put("/v1/io_engine", |r| api_handler(r, put_io_engine_handler))
         .any(handler_404))
 }

pageserver/src/metrics.rs

@@ -602,6 +602,15 @@ pub(crate) mod initial_logical_size {
         });
 }
 
+static DIRECTORY_ENTRIES_COUNT: Lazy<UIntGaugeVec> = Lazy::new(|| {
+    register_uint_gauge_vec!(
+        "pageserver_directory_entries_count",
+        "Sum of the entries in pageserver-stored directory listings",
+        &["tenant_id", "shard_id", "timeline_id"]
+    )
+    .expect("failed to define a metric")
+});
+
 pub(crate) static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
     register_uint_gauge_vec!(
         "pageserver_tenant_states_count",
@@ -1809,6 +1818,7 @@ pub(crate) struct TimelineMetrics {
     resident_physical_size_gauge: UIntGauge,
     /// copy of LayeredTimeline.current_logical_size
     pub current_logical_size_gauge: UIntGauge,
+    pub directory_entries_count_gauge: Lazy<UIntGauge, Box<dyn Send + Fn() -> UIntGauge>>,
     pub num_persistent_files_created: IntCounter,
     pub persistent_bytes_written: IntCounter,
     pub evictions: IntCounter,
@@ -1818,12 +1828,12 @@ pub(crate) struct TimelineMetrics {
 impl TimelineMetrics {
     pub fn new(
         tenant_shard_id: &TenantShardId,
-        timeline_id: &TimelineId,
+        timeline_id_raw: &TimelineId,
         evictions_with_low_residence_duration_builder: EvictionsWithLowResidenceDurationBuilder,
     ) -> Self {
         let tenant_id = tenant_shard_id.tenant_id.to_string();
         let shard_id = format!("{}", tenant_shard_id.shard_slug());
-        let timeline_id = timeline_id.to_string();
+        let timeline_id = timeline_id_raw.to_string();
         let flush_time_histo = StorageTimeMetrics::new(
             StorageTimeOperation::LayerFlush,
             &tenant_id,
@@ -1876,6 +1886,22 @@ impl TimelineMetrics {
         let current_logical_size_gauge = CURRENT_LOGICAL_SIZE
             .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
             .unwrap();
+        // TODO use impl Trait syntax here once we have ability to use it: https://github.com/rust-lang/rust/issues/63065
+        let directory_entries_count_gauge_closure = {
+            let tenant_shard_id = *tenant_shard_id;
+            let timeline_id_raw = *timeline_id_raw;
+            move || {
+                let tenant_id = tenant_shard_id.tenant_id.to_string();
+                let shard_id = format!("{}", tenant_shard_id.shard_slug());
+                let timeline_id = timeline_id_raw.to_string();
+                let gauge: UIntGauge = DIRECTORY_ENTRIES_COUNT
+                    .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
+                    .unwrap();
+                gauge
+            }
+        };
+        let directory_entries_count_gauge: Lazy<UIntGauge, Box<dyn Send + Fn() -> UIntGauge>> =
+            Lazy::new(Box::new(directory_entries_count_gauge_closure));
         let num_persistent_files_created = NUM_PERSISTENT_FILES_CREATED
             .get_metric_with_label_values(&[&tenant_id, &shard_id, &timeline_id])
             .unwrap();
@@ -1902,6 +1928,7 @@ impl TimelineMetrics {
             last_record_gauge,
             resident_physical_size_gauge,
             current_logical_size_gauge,
+            directory_entries_count_gauge,
             num_persistent_files_created,
             persistent_bytes_written,
             evictions,
@@ -1944,6 +1971,9 @@ impl Drop for TimelineMetrics {
                 RESIDENT_PHYSICAL_SIZE.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
         }
         let _ = CURRENT_LOGICAL_SIZE.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
+        if let Some(metric) = Lazy::get(&DIRECTORY_ENTRIES_COUNT) {
+            let _ = metric.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
+        }
         let _ =
             NUM_PERSISTENT_FILES_CREATED.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
         let _ = PERSISTENT_BYTES_WRITTEN.remove_label_values(&[tenant_id, &shard_id, timeline_id]);
@@ -2400,6 +2430,72 @@ impl<F: Future<Output = Result<O, E>>, O, E> Future for MeasuredRemoteOp<F> {
     }
 }
 
+pub mod tokio_epoll_uring {
+    use metrics::UIntGauge;
+
+    pub struct Collector {
+        descs: Vec<metrics::core::Desc>,
+        systems_created: UIntGauge,
+        systems_destroyed: UIntGauge,
+    }
+
+    const NMETRICS: usize = 2;
+
+    impl metrics::core::Collector for Collector {
+        fn desc(&self) -> Vec<&metrics::core::Desc> {
+            self.descs.iter().collect()
+        }
+
+        fn collect(&self) -> Vec<metrics::proto::MetricFamily> {
+            let mut mfs = Vec::with_capacity(NMETRICS);
+            let tokio_epoll_uring::metrics::Metrics {
+                systems_created,
+                systems_destroyed,
+            } = tokio_epoll_uring::metrics::global();
+            self.systems_created.set(systems_created);
+            mfs.extend(self.systems_created.collect());
+            self.systems_destroyed.set(systems_destroyed);
+            mfs.extend(self.systems_destroyed.collect());
+            mfs
+        }
+    }
+
+    impl Collector {
+        #[allow(clippy::new_without_default)]
+        pub fn new() -> Self {
+            let mut descs = Vec::new();
+
+            let systems_created = UIntGauge::new(
+                "pageserver_tokio_epoll_uring_systems_created",
+                "counter of tokio-epoll-uring systems that were created",
+            )
+            .unwrap();
+            descs.extend(
+                metrics::core::Collector::desc(&systems_created)
+                    .into_iter()
+                    .cloned(),
+            );
+
+            let systems_destroyed = UIntGauge::new(
+                "pageserver_tokio_epoll_uring_systems_destroyed",
+                "counter of tokio-epoll-uring systems that were destroyed",
+            )
+            .unwrap();
+            descs.extend(
+                metrics::core::Collector::desc(&systems_destroyed)
+                    .into_iter()
+                    .cloned(),
+            );
+
+            Self {
+                descs,
+                systems_created,
+                systems_destroyed,
+            }
+        }
+    }
+}
+
 pub fn preinitialize_metrics() {
     // Python tests need these and on some we do alerting.
     //

pageserver/src/page_service.rs

@@ -91,8 +91,8 @@ const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(30000);
 /// `tokio_tar` already read the first such block. Read the second all-zeros block,
 /// and check that there is no more data after the EOF marker.
 ///
-/// XXX: Currently, any trailing data after the EOF marker prints a warning.
-/// Perhaps it should be a hard error?
+/// 'tar' command can also write extra blocks of zeros, up to a record
+/// size, controlled by the --record-size argument. Ignore them too.
 async fn read_tar_eof(mut reader: (impl AsyncRead + Unpin)) -> anyhow::Result<()> {
     use tokio::io::AsyncReadExt;
     let mut buf = [0u8; 512];
@@ -113,17 +113,24 @@ async fn read_tar_eof(mut reader: (impl AsyncRead + Unpin)) -> anyhow::Result<()
         anyhow::bail!("invalid tar EOF marker");
     }
 
-    // Drain any data after the EOF marker
+    // Drain any extra zero-blocks after the EOF marker
     let mut trailing_bytes = 0;
+    let mut seen_nonzero_bytes = false;
     loop {
         let nbytes = reader.read(&mut buf).await?;
         trailing_bytes += nbytes;
+        if !buf.iter().all(|&x| x == 0) {
+            seen_nonzero_bytes = true;
+        }
         if nbytes == 0 {
             break;
         }
     }
-    if trailing_bytes > 0 {
-        warn!("ignored {trailing_bytes} unexpected bytes after the tar archive");
+    if seen_nonzero_bytes {
+        anyhow::bail!("unexpected non-zero bytes after the tar archive");
+    }
+    if trailing_bytes % 512 != 0 {
+        anyhow::bail!("unexpected number of zeros ({trailing_bytes}), not divisible by tar block size (512 bytes), after the tar archive");
     }
     Ok(())
 }
@@ -631,25 +638,6 @@ impl PageServerHandler {
                         span,
                     )
                 }
-                PagestreamFeMessage::GetLatestPage(old_req) => {
-                    let req = PagestreamGetPageRequest {
-                        horizon: if old_req.latest {
-                            Lsn::MAX
-                        } else {
-                            old_req.lsn
-                        },
-                        lsn: old_req.lsn,
-                        rel: old_req.rel,
-                        blkno: old_req.blkno,
-                    };
-                    let span = tracing::info_span!("handle_get_page_at_lsn_request", rel = %req.rel, blkno = %req.blkno, req_lsn = %req.lsn);
-                    (
-                        self.handle_get_page_at_lsn_request(tenant_id, timeline_id, &req, &ctx)
-                            .instrument(span.clone())
-                            .await,
-                        span,
-                    )
-                }
                 PagestreamFeMessage::GetPage(req) => {
                     // shard_id is filled in by the handler
                     let span = tracing::info_span!("handle_get_page_at_lsn_request", rel = %req.rel, blkno = %req.blkno, req_lsn = %req.lsn);
@@ -861,41 +849,55 @@ impl PageServerHandler {
     /// the LSN that should be used to look up the page versions.
     async fn wait_or_get_last_lsn(
         timeline: &Timeline,
-        lsn: Lsn,
-        horizon: Lsn,
+        mut lsn: Lsn,
+        latest: bool,
         latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
         ctx: &RequestContext,
     ) -> Result<Lsn, PageStreamError> {
-        let last_record_lsn = timeline.get_last_record_lsn();
-        // Horizon = 0 (INVALID) is treated as LSN interval degenerated to point [lsn,lsn].
-        // It as done mostly for convenience (because such get_page commands are widely used in tests) and
-        // also seems to be logical: Lsn::MAX moves upper boundary of LSN interval till last_record_lsn and
-        // Lsn(0) moves upper boundary to lower boundary.
-        let request_horizon = if horizon == Lsn::INVALID {
-            lsn
+        if latest {
+            // Latest page version was requested. If LSN is given, it is a hint
+            // to the page server that there have been no modifications to the
+            // page after that LSN. If we haven't received WAL up to that point,
+            // wait until it arrives.
+            let last_record_lsn = timeline.get_last_record_lsn();
+
+            // Note: this covers the special case that lsn == Lsn(0). That
+            // special case means "return the latest version whatever it is",
+            // and it's used for bootstrapping purposes, when the page server is
+            // connected directly to the compute node. That is needed because
+            // when you connect to the compute node, to receive the WAL, the
+            // walsender process will do a look up in the pg_authid catalog
+            // table for authentication. That poses a deadlock problem: the
+            // catalog table lookup will send a GetPage request, but the GetPage
+            // request will block in the page server because the recent WAL
+            // hasn't been received yet, and it cannot be received until the
+            // walsender completes the authentication and starts streaming the
+            // WAL.
+            if lsn <= last_record_lsn {
+                lsn = last_record_lsn;
+            } else {
+                timeline.wait_lsn(lsn, ctx).await?;
+                // Since we waited for 'lsn' to arrive, that is now the last
+                // record LSN. (Or close enough for our purposes; the
+                // last-record LSN can advance immediately after we return
+                // anyway)
+            }
         } else {
-            horizon
-        };
-        let effective_lsn = Lsn::max(lsn, Lsn::min(request_horizon, last_record_lsn));
-        if effective_lsn > last_record_lsn {
-            timeline.wait_lsn(effective_lsn, ctx).await?;
-            // Since we waited for 'lsn' to arrive, that is now the last
-            // record LSN. (Or close enough for our purposes; the
-            // last-record LSN can advance immediately after we return
-            // anyway)
-        } else if effective_lsn == Lsn(0) {
-            return Err(PageStreamError::BadRequest(
-                "invalid LSN(0) in request".into(),
-            ));
+            if lsn == Lsn(0) {
+                return Err(PageStreamError::BadRequest(
+                    "invalid LSN(0) in request".into(),
+                ));
+            }
+            timeline.wait_lsn(lsn, ctx).await?;
         }
 
-        if effective_lsn < **latest_gc_cutoff_lsn {
+        if lsn < **latest_gc_cutoff_lsn {
             return Err(PageStreamError::BadRequest(format!(
                 "tried to request a page version that was garbage collected. requested at {} gc cutoff {}",
-                effective_lsn, **latest_gc_cutoff_lsn
+                lsn, **latest_gc_cutoff_lsn
             ).into()));
         }
-        Ok(effective_lsn)
+        Ok(lsn)
     }
 
     #[instrument(skip_all, fields(shard_id))]
@@ -913,11 +915,11 @@ impl PageServerHandler {
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
-            Self::wait_or_get_last_lsn(timeline, req.lsn, req.horizon, &latest_gc_cutoff_lsn, ctx)
+            Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest, &latest_gc_cutoff_lsn, ctx)
                 .await?;
 
         let exists = timeline
-            .get_rel_exists(req.rel, Version::Lsn(lsn), req.horizon, ctx)
+            .get_rel_exists(req.rel, Version::Lsn(lsn), req.latest, ctx)
             .await?;
 
         Ok(PagestreamBeMessage::Exists(PagestreamExistsResponse {
@@ -941,11 +943,11 @@ impl PageServerHandler {
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
-            Self::wait_or_get_last_lsn(timeline, req.lsn, req.horizon, &latest_gc_cutoff_lsn, ctx)
+            Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest, &latest_gc_cutoff_lsn, ctx)
                 .await?;
 
         let n_blocks = timeline
-            .get_rel_size(req.rel, Version::Lsn(lsn), req.horizon, ctx)
+            .get_rel_size(req.rel, Version::Lsn(lsn), req.latest, ctx)
             .await?;
 
         Ok(PagestreamBeMessage::Nblocks(PagestreamNblocksResponse {
@@ -969,7 +971,7 @@ impl PageServerHandler {
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
-            Self::wait_or_get_last_lsn(timeline, req.lsn, req.horizon, &latest_gc_cutoff_lsn, ctx)
+            Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest, &latest_gc_cutoff_lsn, ctx)
                 .await?;
 
         let total_blocks = timeline
@@ -977,7 +979,7 @@ impl PageServerHandler {
                 DEFAULTTABLESPACE_OID,
                 req.dbnode,
                 Version::Lsn(lsn),
-                req.horizon,
+                req.latest,
                 ctx,
             )
             .await?;
@@ -1147,11 +1149,11 @@ impl PageServerHandler {
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
-            Self::wait_or_get_last_lsn(timeline, req.lsn, req.horizon, &latest_gc_cutoff_lsn, ctx)
+            Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest, &latest_gc_cutoff_lsn, ctx)
                 .await?;
 
         let page = timeline
-            .get_rel_page_at_lsn(req.rel, req.blkno, Version::Lsn(lsn), req.horizon, ctx)
+            .get_rel_page_at_lsn(req.rel, req.blkno, Version::Lsn(lsn), req.latest, ctx)
             .await?;
 
         Ok(PagestreamBeMessage::GetPage(PagestreamGetPageResponse {
@@ -1175,7 +1177,7 @@ impl PageServerHandler {
 
         let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
         let lsn =
-            Self::wait_or_get_last_lsn(timeline, req.lsn, req.horizon, &latest_gc_cutoff_lsn, ctx)
+            Self::wait_or_get_last_lsn(timeline, req.lsn, req.latest, &latest_gc_cutoff_lsn, ctx)
                 .await?;
 
         let kind = SlruKind::from_repr(req.kind)

pageserver/src/pgdatadir_mapping.rs

@@ -14,6 +14,7 @@ use crate::span::debug_assert_current_span_has_tenant_and_timeline_id_no_shard_i
 use crate::walrecord::NeonWalRecord;
 use anyhow::{ensure, Context};
 use bytes::{Buf, Bytes, BytesMut};
+use enum_map::Enum;
 use pageserver_api::key::{
     dbdir_key_range, is_rel_block_key, is_slru_block_key, rel_block_to_key, rel_dir_to_key,
     rel_key_range, rel_size_to_key, relmap_file_key, slru_block_to_key, slru_dir_to_key,
@@ -155,6 +156,8 @@ impl Timeline {
             pending_updates: HashMap::new(),
             pending_deletions: Vec::new(),
             pending_nblocks: 0,
+            pending_aux_files: None,
+            pending_directory_entries: Vec::new(),
             lsn,
         }
     }
@@ -169,7 +172,7 @@ impl Timeline {
         tag: RelTag,
         blknum: BlockNumber,
         version: Version<'_>,
-        horizon: Lsn,
+        latest: bool,
         ctx: &RequestContext,
     ) -> Result<Bytes, PageReconstructError> {
         if tag.relnode == 0 {
@@ -178,7 +181,7 @@ impl Timeline {
             ));
         }
 
-        let nblocks = self.get_rel_size(tag, version, horizon, ctx).await?;
+        let nblocks = self.get_rel_size(tag, version, latest, ctx).await?;
         if blknum >= nblocks {
             debug!(
                 "read beyond EOF at {} blk {} at {}, size is {}: returning all-zeros page",
@@ -200,7 +203,7 @@ impl Timeline {
         spcnode: Oid,
         dbnode: Oid,
         version: Version<'_>,
-        horizon: Lsn,
+        latest: bool,
         ctx: &RequestContext,
     ) -> Result<usize, PageReconstructError> {
         let mut total_blocks = 0;
@@ -208,7 +211,7 @@ impl Timeline {
         let rels = self.list_rels(spcnode, dbnode, version, ctx).await?;
 
         for rel in rels {
-            let n_blocks = self.get_rel_size(rel, version, horizon, ctx).await?;
+            let n_blocks = self.get_rel_size(rel, version, latest, ctx).await?;
             total_blocks += n_blocks as usize;
         }
         Ok(total_blocks)
@@ -219,7 +222,7 @@ impl Timeline {
         &self,
         tag: RelTag,
         version: Version<'_>,
-        horizon: Lsn,
+        latest: bool,
         ctx: &RequestContext,
     ) -> Result<BlockNumber, PageReconstructError> {
         if tag.relnode == 0 {
@@ -233,7 +236,7 @@ impl Timeline {
         }
 
         if (tag.forknum == FSM_FORKNUM || tag.forknum == VISIBILITYMAP_FORKNUM)
-            && !self.get_rel_exists(tag, version, horizon, ctx).await?
+            && !self.get_rel_exists(tag, version, latest, ctx).await?
         {
             // FIXME: Postgres sometimes calls smgrcreate() to create
             // FSM, and smgrnblocks() on it immediately afterwards,
@@ -246,8 +249,14 @@ impl Timeline {
         let mut buf = version.get(self, key, ctx).await?;
         let nblocks = buf.get_u32_le();
 
-        if horizon == Lsn::MAX {
-            // Update relation size cache only if latest version is requested.
+        if latest {
+            // Update relation size cache only if "latest" flag is set.
+            // This flag is set by compute when it is working with most recent version of relation.
+            // Typically master compute node always set latest=true.
+            // Please notice, that even if compute node "by mistake" specifies old LSN but set
+            // latest=true, then it can not cause cache corruption, because with latest=true
+            // pageserver choose max(request_lsn, last_written_lsn) and so cached value will be
+            // associated with most recent value of LSN.
             self.update_cached_rel_size(tag, version.get_lsn(), nblocks);
         }
         Ok(nblocks)
@@ -258,7 +267,7 @@ impl Timeline {
         &self,
         tag: RelTag,
         version: Version<'_>,
-        _horizon: Lsn,
+        _latest: bool,
         ctx: &RequestContext,
     ) -> Result<bool, PageReconstructError> {
         if tag.relnode == 0 {
@@ -862,6 +871,15 @@ pub struct DatadirModification<'a> {
     pending_updates: HashMap<Key, Vec<(Lsn, Value)>>,
     pending_deletions: Vec<(Range<Key>, Lsn)>,
     pending_nblocks: i64,
+
+    // If we already wrote any aux file changes in this modification, stash the latest dir.  If set,
+    // [`Self::put_file`] may assume that it is safe to emit a delta rather than checking
+    // if AUX_FILES_KEY is already set.
+    pending_aux_files: Option<AuxFilesDirectory>,
+
+    /// For special "directory" keys that store key-value maps, track the size of the map
+    /// if it was updated in this modification.
+    pending_directory_entries: Vec<(DirectoryKind, usize)>,
 }
 
 impl<'a> DatadirModification<'a> {
@@ -893,6 +911,7 @@ impl<'a> DatadirModification<'a> {
         let buf = DbDirectory::ser(&DbDirectory {
             dbdirs: HashMap::new(),
         })?;
+        self.pending_directory_entries.push((DirectoryKind::Db, 0));
         self.put(DBDIR_KEY, Value::Image(buf.into()));
 
         // Create AuxFilesDirectory
@@ -901,16 +920,24 @@ impl<'a> DatadirModification<'a> {
         let buf = TwoPhaseDirectory::ser(&TwoPhaseDirectory {
             xids: HashSet::new(),
         })?;
+        self.pending_directory_entries
+            .push((DirectoryKind::TwoPhase, 0));
         self.put(TWOPHASEDIR_KEY, Value::Image(buf.into()));
 
         let buf: Bytes = SlruSegmentDirectory::ser(&SlruSegmentDirectory::default())?.into();
         let empty_dir = Value::Image(buf);
         self.put(slru_dir_to_key(SlruKind::Clog), empty_dir.clone());
+        self.pending_directory_entries
+            .push((DirectoryKind::SlruSegment(SlruKind::Clog), 0));
         self.put(
             slru_dir_to_key(SlruKind::MultiXactMembers),
             empty_dir.clone(),
         );
+        self.pending_directory_entries
+            .push((DirectoryKind::SlruSegment(SlruKind::Clog), 0));
         self.put(slru_dir_to_key(SlruKind::MultiXactOffsets), empty_dir);
+        self.pending_directory_entries
+            .push((DirectoryKind::SlruSegment(SlruKind::MultiXactOffsets), 0));
 
         Ok(())
     }
@@ -1011,6 +1038,7 @@ impl<'a> DatadirModification<'a> {
             let buf = RelDirectory::ser(&RelDirectory {
                 rels: HashSet::new(),
             })?;
+            self.pending_directory_entries.push((DirectoryKind::Rel, 0));
             self.put(
                 rel_dir_to_key(spcnode, dbnode),
                 Value::Image(Bytes::from(buf)),
@@ -1033,6 +1061,8 @@ impl<'a> DatadirModification<'a> {
         if !dir.xids.insert(xid) {
             anyhow::bail!("twophase file for xid {} already exists", xid);
         }
+        self.pending_directory_entries
+            .push((DirectoryKind::TwoPhase, dir.xids.len()));
         self.put(
             TWOPHASEDIR_KEY,
             Value::Image(Bytes::from(TwoPhaseDirectory::ser(&dir)?)),
@@ -1060,7 +1090,7 @@ impl<'a> DatadirModification<'a> {
     ) -> anyhow::Result<()> {
         let total_blocks = self
             .tline
-            .get_db_size(spcnode, dbnode, Version::Modified(self), Lsn::MAX, ctx)
+            .get_db_size(spcnode, dbnode, Version::Modified(self), true, ctx)
             .await?;
 
         // Remove entry from dbdir
@@ -1068,6 +1098,8 @@ impl<'a> DatadirModification<'a> {
         let mut dir = DbDirectory::des(&buf)?;
         if dir.dbdirs.remove(&(spcnode, dbnode)).is_some() {
             let buf = DbDirectory::ser(&dir)?;
+            self.pending_directory_entries
+                .push((DirectoryKind::Db, dir.dbdirs.len()));
             self.put(DBDIR_KEY, Value::Image(buf.into()));
         } else {
             warn!(
@@ -1105,6 +1137,8 @@ impl<'a> DatadirModification<'a> {
             // Didn't exist. Update dbdir
             dbdir.dbdirs.insert((rel.spcnode, rel.dbnode), false);
             let buf = DbDirectory::ser(&dbdir).context("serialize db")?;
+            self.pending_directory_entries
+                .push((DirectoryKind::Db, dbdir.dbdirs.len()));
             self.put(DBDIR_KEY, Value::Image(buf.into()));
 
             // and create the RelDirectory
@@ -1119,6 +1153,10 @@ impl<'a> DatadirModification<'a> {
         if !rel_dir.rels.insert((rel.relnode, rel.forknum)) {
             return Err(RelationError::AlreadyExists);
         }
+
+        self.pending_directory_entries
+            .push((DirectoryKind::Rel, rel_dir.rels.len()));
+
         self.put(
             rel_dir_key,
             Value::Image(Bytes::from(
@@ -1151,7 +1189,7 @@ impl<'a> DatadirModification<'a> {
         anyhow::ensure!(rel.relnode != 0, RelationError::InvalidRelnode);
         if self
             .tline
-            .get_rel_exists(rel, Version::Modified(self), Lsn::MAX, ctx)
+            .get_rel_exists(rel, Version::Modified(self), true, ctx)
             .await?
         {
             let size_key = rel_size_to_key(rel);
@@ -1210,6 +1248,9 @@ impl<'a> DatadirModification<'a> {
         let buf = self.get(dir_key, ctx).await?;
         let mut dir = RelDirectory::des(&buf)?;
 
+        self.pending_directory_entries
+            .push((DirectoryKind::Rel, dir.rels.len()));
+
         if dir.rels.remove(&(rel.relnode, rel.forknum)) {
             self.put(dir_key, Value::Image(Bytes::from(RelDirectory::ser(&dir)?)));
         } else {
@@ -1245,6 +1286,8 @@ impl<'a> DatadirModification<'a> {
         if !dir.segments.insert(segno) {
             anyhow::bail!("slru segment {kind:?}/{segno} already exists");
         }
+        self.pending_directory_entries
+            .push((DirectoryKind::SlruSegment(kind), dir.segments.len()));
         self.put(
             dir_key,
             Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
@@ -1289,6 +1332,8 @@ impl<'a> DatadirModification<'a> {
         if !dir.segments.remove(&segno) {
             warn!("slru segment {:?}/{} does not exist", kind, segno);
         }
+        self.pending_directory_entries
+            .push((DirectoryKind::SlruSegment(kind), dir.segments.len()));
         self.put(
             dir_key,
             Value::Image(Bytes::from(SlruSegmentDirectory::ser(&dir)?)),
@@ -1319,6 +1364,8 @@ impl<'a> DatadirModification<'a> {
         if !dir.xids.remove(&xid) {
             warn!("twophase file for xid {} does not exist", xid);
         }
+        self.pending_directory_entries
+            .push((DirectoryKind::TwoPhase, dir.xids.len()));
         self.put(
             TWOPHASEDIR_KEY,
             Value::Image(Bytes::from(TwoPhaseDirectory::ser(&dir)?)),
@@ -1334,6 +1381,8 @@ impl<'a> DatadirModification<'a> {
         let buf = AuxFilesDirectory::ser(&AuxFilesDirectory {
             files: HashMap::new(),
         })?;
+        self.pending_directory_entries
+            .push((DirectoryKind::AuxFiles, 0));
         self.put(AUX_FILES_KEY, Value::Image(Bytes::from(buf)));
         Ok(())
     }
@@ -1344,28 +1393,76 @@ impl<'a> DatadirModification<'a> {
         content: &[u8],
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
-        let mut dir = match self.get(AUX_FILES_KEY, ctx).await {
-            Ok(buf) => AuxFilesDirectory::des(&buf)?,
-            Err(e) => {
-                // This is expected: historical databases do not have the key.
-                debug!("Failed to get info about AUX files: {}", e);
-                AuxFilesDirectory {
-                    files: HashMap::new(),
+        let file_path = path.to_string();
+        let content = if content.is_empty() {
+            None
+        } else {
+            Some(Bytes::copy_from_slice(content))
+        };
+
+        let dir = if let Some(mut dir) = self.pending_aux_files.take() {
+            // We already updated aux files in `self`: emit a delta and update our latest value
+
+            self.put(
+                AUX_FILES_KEY,
+                Value::WalRecord(NeonWalRecord::AuxFile {
+                    file_path: file_path.clone(),
+                    content: content.clone(),
+                }),
+            );
+
+            dir.upsert(file_path, content);
+            dir
+        } else {
+            // Check if the AUX_FILES_KEY is initialized
+            match self.get(AUX_FILES_KEY, ctx).await {
+                Ok(dir_bytes) => {
+                    let mut dir = AuxFilesDirectory::des(&dir_bytes)?;
+                    // Key is already set, we may append a delta
+                    self.put(
+                        AUX_FILES_KEY,
+                        Value::WalRecord(NeonWalRecord::AuxFile {
+                            file_path: file_path.clone(),
+                            content: content.clone(),
+                        }),
+                    );
+                    dir.upsert(file_path, content);
+                    dir
+                }
+                Err(
+                    e @ (PageReconstructError::AncestorStopping(_)
+                    | PageReconstructError::Cancelled
+                    | PageReconstructError::AncestorLsnTimeout(_)),
+                ) => {
+                    // Important that we do not interpret a shutdown error as "not found" and thereby
+                    // reset the map.
+                    return Err(e.into());
+                }
+                // FIXME: PageReconstructError doesn't have an explicit variant for key-not-found, so
+                // we are assuming that all _other_ possible errors represents a missing key.  If some
+                // other error occurs, we may incorrectly reset the map of aux files.
+                Err(PageReconstructError::Other(_) | PageReconstructError::WalRedo(_)) => {
+                    // Key is missing, we must insert an image as the basis for subsequent deltas.
+
+                    let mut dir = AuxFilesDirectory {
+                        files: HashMap::new(),
+                    };
+                    dir.upsert(file_path, content);
+                    self.put(
+                        AUX_FILES_KEY,
+                        Value::Image(Bytes::from(
+                            AuxFilesDirectory::ser(&dir).context("serialize")?,
+                        )),
+                    );
+                    dir
                 }
             }
         };
-        let path = path.to_string();
-        if content.is_empty() {
-            dir.files.remove(&path);
-        } else {
-            dir.files.insert(path, Bytes::copy_from_slice(content));
-        }
-        self.put(
-            AUX_FILES_KEY,
-            Value::Image(Bytes::from(
-                AuxFilesDirectory::ser(&dir).context("serialize")?,
-            )),
-        );
+
+        self.pending_directory_entries
+            .push((DirectoryKind::AuxFiles, dir.files.len()));
+        self.pending_aux_files = Some(dir);
+
         Ok(())
     }
 
@@ -1421,6 +1518,10 @@ impl<'a> DatadirModification<'a> {
             self.pending_nblocks = 0;
         }
 
+        for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
+            writer.update_directory_entries_count(kind, count as u64);
+        }
+
         Ok(())
     }
 
@@ -1458,6 +1559,10 @@ impl<'a> DatadirModification<'a> {
             writer.update_current_logical_size(pending_nblocks * i64::from(BLCKSZ));
         }
 
+        for (kind, count) in std::mem::take(&mut self.pending_directory_entries) {
+            writer.update_directory_entries_count(kind, count as u64);
+        }
+
         Ok(())
     }
 
@@ -1567,8 +1672,18 @@ struct RelDirectory {
 }
 
 #[derive(Debug, Serialize, Deserialize, Default)]
-struct AuxFilesDirectory {
-    files: HashMap<String, Bytes>,
+pub(crate) struct AuxFilesDirectory {
+    pub(crate) files: HashMap<String, Bytes>,
+}
+
+impl AuxFilesDirectory {
+    pub(crate) fn upsert(&mut self, key: String, value: Option<Bytes>) {
+        if let Some(value) = value {
+            self.files.insert(key, value);
+        } else {
+            self.files.remove(&key);
+        }
+    }
 }
 
 #[derive(Debug, Serialize, Deserialize)]
@@ -1582,13 +1697,82 @@ struct SlruSegmentDirectory {
     segments: HashSet<u32>,
 }
 
+#[derive(Copy, Clone, PartialEq, Eq, Debug, enum_map::Enum)]
+#[repr(u8)]
+pub(crate) enum DirectoryKind {
+    Db,
+    TwoPhase,
+    Rel,
+    AuxFiles,
+    SlruSegment(SlruKind),
+}
+
+impl DirectoryKind {
+    pub(crate) const KINDS_NUM: usize = <DirectoryKind as Enum>::LENGTH;
+    pub(crate) fn offset(&self) -> usize {
+        self.into_usize()
+    }
+}
+
 static ZERO_PAGE: Bytes = Bytes::from_static(&[0u8; BLCKSZ as usize]);
 
 #[allow(clippy::bool_assert_comparison)]
 #[cfg(test)]
 mod tests {
-    //use super::repo_harness::*;
-    //use super::*;
+    use hex_literal::hex;
+    use utils::id::TimelineId;
+
+    use super::*;
+
+    use crate::{tenant::harness::TenantHarness, DEFAULT_PG_VERSION};
+
+    /// Test a round trip of aux file updates, from DatadirModification to reading back from the Timeline
+    #[tokio::test]
+    async fn aux_files_round_trip() -> anyhow::Result<()> {
+        let name = "aux_files_round_trip";
+        let harness = TenantHarness::create(name)?;
+
+        pub const TIMELINE_ID: TimelineId =
+            TimelineId::from_array(hex!("11223344556677881122334455667788"));
+
+        let (tenant, ctx) = harness.load().await;
+        let tline = tenant
+            .create_empty_timeline(TIMELINE_ID, Lsn(0), DEFAULT_PG_VERSION, &ctx)
+            .await?;
+        let tline = tline.raw_timeline().unwrap();
+
+        // First modification: insert two keys
+        let mut modification = tline.begin_modification(Lsn(0x1000));
+        modification.put_file("foo/bar1", b"content1", &ctx).await?;
+        modification.set_lsn(Lsn(0x1008))?;
+        modification.put_file("foo/bar2", b"content2", &ctx).await?;
+        modification.commit(&ctx).await?;
+        let expect_1008 = HashMap::from([
+            ("foo/bar1".to_string(), Bytes::from_static(b"content1")),
+            ("foo/bar2".to_string(), Bytes::from_static(b"content2")),
+        ]);
+
+        let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
+        assert_eq!(readback, expect_1008);
+
+        // Second modification: update one key, remove the other
+        let mut modification = tline.begin_modification(Lsn(0x2000));
+        modification.put_file("foo/bar1", b"content3", &ctx).await?;
+        modification.set_lsn(Lsn(0x2008))?;
+        modification.put_file("foo/bar2", b"", &ctx).await?;
+        modification.commit(&ctx).await?;
+        let expect_2008 =
+            HashMap::from([("foo/bar1".to_string(), Bytes::from_static(b"content3"))]);
+
+        let readback = tline.list_aux_files(Lsn(0x2008), &ctx).await?;
+        assert_eq!(readback, expect_2008);
+
+        // Reading back in time works
+        let readback = tline.list_aux_files(Lsn(0x1008), &ctx).await?;
+        assert_eq!(readback, expect_1008);
+
+        Ok(())
+    }
 
     /*
         fn assert_current_logical_size<R: Repository>(timeline: &DatadirTimeline<R>, lsn: Lsn) {

pageserver/src/task_mgr.rs

@@ -30,10 +30,6 @@
 //! only a single tenant or timeline.
 //!
 
-// Clippy 1.60 incorrectly complains about the tokio::task_local!() macro.
-// Silence it. See https://github.com/rust-lang/rust-clippy/issues/9224.
-#![allow(clippy::declare_interior_mutable_const)]
-
 use std::collections::HashMap;
 use std::fmt;
 use std::future::Future;
@@ -312,7 +308,6 @@ struct MutableTaskState {
 }
 
 struct PageServerTask {
-    #[allow(dead_code)] // unused currently
     task_id: PageserverTaskId,
 
     kind: TaskKind,
@@ -576,8 +571,8 @@ pub fn shutdown_token() -> CancellationToken {
 
 /// Has the current task been requested to shut down?
 pub fn is_shutdown_requested() -> bool {
-    if let Ok(cancel) = SHUTDOWN_TOKEN.try_with(|t| t.clone()) {
-        cancel.is_cancelled()
+    if let Ok(true_or_false) = SHUTDOWN_TOKEN.try_with(|t| t.is_cancelled()) {
+        true_or_false
     } else {
         if !cfg!(test) {
             warn!("is_shutdown_requested() called in an unexpected task or thread");

pageserver/src/tenant.rs

@@ -25,6 +25,7 @@ use pageserver_api::shard::ShardIdentity;
 use pageserver_api::shard::TenantShardId;
 use remote_storage::DownloadError;
 use remote_storage::GenericRemoteStorage;
+use remote_storage::TimeoutOrCancel;
 use std::fmt;
 use storage_broker::BrokerClientChannel;
 use tokio::io::BufReader;
@@ -53,6 +54,7 @@ use self::metadata::TimelineMetadata;
 use self::mgr::GetActiveTenantError;
 use self::mgr::GetTenantError;
 use self::mgr::TenantsMap;
+use self::remote_timeline_client::upload::upload_index_part;
 use self::remote_timeline_client::RemoteTimelineClient;
 use self::timeline::uninit::TimelineExclusionError;
 use self::timeline::uninit::TimelineUninitMark;
@@ -643,10 +645,10 @@ impl Tenant {
 
         // The attach task will carry a GateGuard, so that shutdown() reliably waits for it to drop out if
         // we shut down while attaching.
-        let Ok(attach_gate_guard) = tenant.gate.enter() else {
-            // We just created the Tenant: nothing else can have shut it down yet
-            unreachable!();
-        };
+        let attach_gate_guard = tenant
+            .gate
+            .enter()
+            .expect("We just created the Tenant: nothing else can have shut it down yet");
 
         // Do all the hard work in the background
         let tenant_clone = Arc::clone(&tenant);
@@ -754,36 +756,27 @@ impl Tenant {
                     AttachType::Normal
                 };
 
-                let preload_timer = TENANT.preload.start_timer();
-                let preload = match mode {
-                    SpawnMode::Create => {
-                        // Don't count the skipped preload into the histogram of preload durations
-                        preload_timer.stop_and_discard();
+                let preload = match (&mode, &remote_storage) {
+                    (SpawnMode::Create, _) => {
                         None
                     },
-                    SpawnMode::Normal => {
-                        match &remote_storage {
-                            Some(remote_storage) => Some(
-                                match tenant_clone
-                                    .preload(remote_storage, task_mgr::shutdown_token())
-                                    .instrument(
-                                        tracing::info_span!(parent: None, "attach_preload", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug()),
-                                    )
-                                    .await {
-                                        Ok(p) => {
-                                            preload_timer.observe_duration();
-                                            p
-                                        }
-                                            ,
-                                        Err(e) => {
-                                            make_broken(&tenant_clone, anyhow::anyhow!(e));
-                                                return Ok(());
-                                        }
-                                    },
-                            ),
-                            None => None,
+                    (SpawnMode::Normal, Some(remote_storage)) => {
+                        let _preload_timer = TENANT.preload.start_timer();
+                        let res = tenant_clone
+                            .preload(remote_storage, task_mgr::shutdown_token())
+                            .await;
+                        match res {
+                            Ok(p) => Some(p),
+                            Err(e) => {
+                                make_broken(&tenant_clone, anyhow::anyhow!(e));
+                                return Ok(());
+                            }
                         }
                     }
+                    (SpawnMode::Normal, None) => {
+                        let _preload_timer = TENANT.preload.start_timer();
+                        None
+                    }
                 };
 
                 // Remote preload is complete.
@@ -819,36 +812,37 @@ impl Tenant {
                         info!("ready for backgound jobs barrier");
                     }
 
-                    match DeleteTenantFlow::resume_from_attach(
+                    let deleted = DeleteTenantFlow::resume_from_attach(
                         deletion,
                         &tenant_clone,
                         preload,
                         tenants,
                         &ctx,
                     )
-                    .await
-                    {
-                        Err(err) => {
-                            make_broken(&tenant_clone, anyhow::anyhow!(err));
-                            return Ok(());
-                        }
-                        Ok(()) => return Ok(()),
+                    .await;
+
+                    if let Err(e) = deleted {
+                        make_broken(&tenant_clone, anyhow::anyhow!(e));
                     }
+
+                    return Ok(());
                 }
 
                 // We will time the duration of the attach phase unless this is a creation (attach will do no work)
-                let attach_timer = match mode {
-                    SpawnMode::Create => None,
-                    SpawnMode::Normal => {Some(TENANT.attach.start_timer())}
+                let attached = {
+                    let _attach_timer = match mode {
+                        SpawnMode::Create => None,
+                        SpawnMode::Normal => {Some(TENANT.attach.start_timer())}
+                    };
+                    tenant_clone.attach(preload, mode, &ctx).await
                 };
-                match tenant_clone.attach(preload, mode, &ctx).await {
+
+                match attached {
                     Ok(()) => {
                         info!("attach finished, activating");
-                        if let Some(t)=  attach_timer {t.observe_duration();}
                         tenant_clone.activate(broker_client, None, &ctx);
                     }
                     Err(e) => {
-                        if let Some(t)=  attach_timer {t.observe_duration();}
                         make_broken(&tenant_clone, anyhow::anyhow!(e));
                     }
                 }
@@ -861,34 +855,26 @@ impl Tenant {
                 // logical size calculations: if logical size calculation semaphore is saturated,
                 // then warmup will wait for that before proceeding to the next tenant.
                 if let AttachType::Warmup(_permit) = attach_type {
-                    let mut futs = FuturesUnordered::new();
-                    let timelines: Vec<_> = tenant_clone.timelines.lock().unwrap().values().cloned().collect();
-                    for t in timelines {
-                        futs.push(t.await_initial_logical_size())
-                    }
+                    let mut futs: FuturesUnordered<_> = tenant_clone.timelines.lock().unwrap().values().cloned().map(|t| t.await_initial_logical_size()).collect();
                     tracing::info!("Waiting for initial logical sizes while warming up...");
-                    while futs.next().await.is_some() {
-
-                    }
+                    while futs.next().await.is_some() {}
                     tracing::info!("Warm-up complete");
                 }
 
                 Ok(())
             }
-            .instrument({
-                let span = tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation);
-                span.follows_from(Span::current());
-                span
-            }),
+            .instrument(tracing::info_span!(parent: None, "attach", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), gen=?generation)),
         );
         Ok(tenant)
     }
 
+    #[instrument(skip_all)]
     pub(crate) async fn preload(
         self: &Arc<Tenant>,
         remote_storage: &GenericRemoteStorage,
         cancel: CancellationToken,
     ) -> anyhow::Result<TenantPreload> {
+        span::debug_assert_current_span_has_tenant_id();
         // Get list of remote timelines
         // download index files for every tenant timeline
         info!("listing remote timelines");
@@ -1376,7 +1362,7 @@ impl Tenant {
                 async move {
                     debug!("starting index part download");
 
-                    let index_part = client.download_index_file(cancel_clone).await;
+                    let index_part = client.download_index_file(&cancel_clone).await;
 
                     debug!("finished index part download");
 
@@ -2397,6 +2383,67 @@ impl Tenant {
     pub(crate) fn get_generation(&self) -> Generation {
         self.generation
     }
+
+    /// This function partially shuts down the tenant (it shuts down the Timelines) and is fallible,
+    /// and can leave the tenant in a bad state if it fails.  The caller is responsible for
+    /// resetting this tenant to a valid state if we fail.
+    pub(crate) async fn split_prepare(
+        &self,
+        child_shards: &Vec<TenantShardId>,
+    ) -> anyhow::Result<()> {
+        let timelines = self.timelines.lock().unwrap().clone();
+        for timeline in timelines.values() {
+            let Some(tl_client) = &timeline.remote_client else {
+                anyhow::bail!("Remote storage is mandatory");
+            };
+
+            let Some(remote_storage) = &self.remote_storage else {
+                anyhow::bail!("Remote storage is mandatory");
+            };
+
+            // We do not block timeline creation/deletion during splits inside the pageserver: it is up to higher levels
+            // to ensure that they do not start a split if currently in the process of doing these.
+
+            // Upload an index from the parent: this is partly to provide freshness for the
+            // child tenants that will copy it, and partly for general ease-of-debugging: there will
+            // always be a parent shard index in the same generation as we wrote the child shard index.
+            tl_client.schedule_index_upload_for_file_changes()?;
+            tl_client.wait_completion().await?;
+
+            // Shut down the timeline's remote client: this means that the indices we write
+            // for child shards will not be invalidated by the parent shard deleting layers.
+            tl_client.shutdown().await?;
+
+            // Download methods can still be used after shutdown, as they don't flow through the remote client's
+            // queue.  In principal the RemoteTimelineClient could provide this without downloading it, but this
+            // operation is rare, so it's simpler to just download it (and robustly guarantees that the index
+            // we use here really is the remotely persistent one).
+            let result = tl_client
+                .download_index_file(&self.cancel)
+                .instrument(info_span!("download_index_file", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), timeline_id=%timeline.timeline_id))
+                .await?;
+            let index_part = match result {
+                MaybeDeletedIndexPart::Deleted(_) => {
+                    anyhow::bail!("Timeline deletion happened concurrently with split")
+                }
+                MaybeDeletedIndexPart::IndexPart(p) => p,
+            };
+
+            for child_shard in child_shards {
+                upload_index_part(
+                    remote_storage,
+                    child_shard,
+                    &timeline.timeline_id,
+                    self.generation,
+                    &index_part,
+                    &self.cancel,
+                )
+                .await?;
+            }
+        }
+
+        Ok(())
+    }
 }
 
 /// Given a Vec of timelines and their ancestors (timeline_id, ancestor_id),
@@ -2834,7 +2881,7 @@ impl Tenant {
         let config_path = config_path.to_owned();
         tokio::task::spawn_blocking(move || {
             Handle::current().block_on(async move {
-                let conf_content = conf_content.as_bytes();
+                let conf_content = conf_content.into_bytes();
                 VirtualFile::crashsafe_overwrite(&config_path, &temp_path, conf_content)
                     .await
                     .with_context(|| {
@@ -2871,7 +2918,7 @@ impl Tenant {
         let target_config_path = target_config_path.to_owned();
         tokio::task::spawn_blocking(move || {
             Handle::current().block_on(async move {
-                let conf_content = conf_content.as_bytes();
+                let conf_content = conf_content.into_bytes();
                 VirtualFile::crashsafe_overwrite(&target_config_path, &temp_path, conf_content)
                     .await
                     .with_context(|| {
@@ -3229,7 +3276,7 @@ impl Tenant {
 
     /// For unit tests, make this visible so that other modules can directly create timelines
     #[cfg(test)]
-    #[tracing::instrument(fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
+    #[tracing::instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), %timeline_id))]
     pub(crate) async fn bootstrap_timeline_test(
         &self,
         timeline_id: TimelineId,
@@ -3293,7 +3340,7 @@ impl Tenant {
             &self.cancel,
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancelled"))
+        .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
         .and_then(|x| x)
     }
 
@@ -3343,8 +3390,10 @@ impl Tenant {
                 );
                 let dest_path =
                     &remote_initdb_archive_path(&self.tenant_shard_id.tenant_id, &timeline_id);
+
+                // if this fails, it will get retried by retried control plane requests
                 storage
-                    .copy_object(source_path, dest_path)
+                    .copy_object(source_path, dest_path, &self.cancel)
                     .await
                     .context("copy initdb tar")?;
             }
@@ -3732,6 +3781,10 @@ impl Tenant {
 
         Ok(())
     }
+
+    pub(crate) fn get_tenant_conf(&self) -> TenantConfOpt {
+        self.tenant_conf.read().unwrap().tenant_conf
+    }
 }
 
 fn remove_timeline_and_uninit_mark(
@@ -3864,6 +3917,7 @@ pub(crate) mod harness {
     use utils::lsn::Lsn;
 
     use crate::deletion_queue::mock::MockDeletionQueue;
+    use crate::walredo::apply_neon;
     use crate::{
         config::PageServerConf, repository::Key, tenant::Tenant, walrecord::NeonWalRecord,
     };
@@ -3879,8 +3933,7 @@ pub(crate) mod harness {
         TimelineId::from_array(hex!("AA223344556677881122334455667788"));
 
     /// Convenience function to create a page image with given string as the only content
-    #[allow(non_snake_case)]
-    pub fn TEST_IMG(s: &str) -> Bytes {
+    pub fn test_img(s: &str) -> Bytes {
         let mut buf = BytesMut::new();
         buf.extend_from_slice(s.as_bytes());
         buf.resize(64, 0);
@@ -3916,6 +3969,8 @@ pub(crate) mod harness {
         }
     }
 
+    #[cfg(test)]
+    #[derive(Debug)]
     enum LoadMode {
         Local,
         Remote,
@@ -3978,6 +4033,7 @@ pub(crate) mod harness {
             std::fs::create_dir_all(&remote_fs_dir).unwrap();
             let config = RemoteStorageConfig {
                 storage: RemoteStorageKind::LocalFs(remote_fs_dir.clone()),
+                timeout: RemoteStorageConfig::DEFAULT_TIMEOUT,
             };
             let remote_storage = GenericRemoteStorage::from_config(&config).unwrap();
             let deletion_queue = MockDeletionQueue::new(Some(remote_storage.clone()));
@@ -3998,7 +4054,7 @@ pub(crate) mod harness {
             info_span!("TenantHarness", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug())
         }
 
-        pub async fn load(&self) -> (Arc<Tenant>, RequestContext) {
+        pub(crate) async fn load(&self) -> (Arc<Tenant>, RequestContext) {
             let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
             (
                 self.try_load(&ctx)
@@ -4008,6 +4064,74 @@ pub(crate) mod harness {
             )
         }
 
+        /// For tests that specifically want to exercise the local load path, which does
+        /// not use remote storage.
+        pub(crate) async fn try_load_local(
+            &self,
+            ctx: &RequestContext,
+        ) -> anyhow::Result<Arc<Tenant>> {
+            self.do_try_load(ctx, LoadMode::Local).await
+        }
+
+        /// The 'load' in this function is either a local load or a normal attachment,
+        pub(crate) async fn try_load(&self, ctx: &RequestContext) -> anyhow::Result<Arc<Tenant>> {
+            // If we have nothing in remote storage, must use load_local instead of attach: attach
+            // will error out if there are no timelines.
+            //
+            // See https://github.com/neondatabase/neon/issues/5456 for how we will eliminate
+            // this weird state of a Tenant which exists but doesn't have any timelines.
+            let mode = match self.remote_empty() {
+                true => LoadMode::Local,
+                false => LoadMode::Remote,
+            };
+
+            self.do_try_load(ctx, mode).await
+        }
+
+        #[instrument(skip_all, fields(tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug(), ?mode))]
+        async fn do_try_load(
+            &self,
+            ctx: &RequestContext,
+            mode: LoadMode,
+        ) -> anyhow::Result<Arc<Tenant>> {
+            let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
+
+            let tenant = Arc::new(Tenant::new(
+                TenantState::Loading,
+                self.conf,
+                AttachedTenantConf::try_from(LocationConf::attached_single(
+                    TenantConfOpt::from(self.tenant_conf),
+                    self.generation,
+                    &ShardParameters::default(),
+                ))
+                .unwrap(),
+                // This is a legacy/test code path: sharding isn't supported here.
+                ShardIdentity::unsharded(),
+                Some(walredo_mgr),
+                self.tenant_shard_id,
+                Some(self.remote_storage.clone()),
+                self.deletion_queue.new_client(),
+            ));
+
+            match mode {
+                LoadMode::Local => {
+                    tenant.load_local(ctx).await?;
+                }
+                LoadMode::Remote => {
+                    let preload = tenant
+                        .preload(&self.remote_storage, CancellationToken::new())
+                        .await?;
+                    tenant.attach(Some(preload), SpawnMode::Normal, ctx).await?;
+                }
+            }
+
+            tenant.state.send_replace(TenantState::Active);
+            for timeline in tenant.timelines.lock().unwrap().values() {
+                timeline.set_state(TimelineState::Active);
+            }
+            Ok(tenant)
+        }
+
         fn remote_empty(&self) -> bool {
             let tenant_path = self.conf.tenant_path(&self.tenant_shard_id);
             let remote_tenant_dir = self
@@ -4033,77 +4157,6 @@ pub(crate) mod harness {
             }
         }
 
-        async fn do_try_load(
-            &self,
-            ctx: &RequestContext,
-            mode: LoadMode,
-        ) -> anyhow::Result<Arc<Tenant>> {
-            let walredo_mgr = Arc::new(WalRedoManager::from(TestRedoManager));
-
-            let tenant = Arc::new(Tenant::new(
-                TenantState::Loading,
-                self.conf,
-                AttachedTenantConf::try_from(LocationConf::attached_single(
-                    TenantConfOpt::from(self.tenant_conf),
-                    self.generation,
-                    &ShardParameters::default(),
-                ))
-                .unwrap(),
-                // This is a legacy/test code path: sharding isn't supported here.
-                ShardIdentity::unsharded(),
-                Some(walredo_mgr),
-                self.tenant_shard_id,
-                Some(self.remote_storage.clone()),
-                self.deletion_queue.new_client(),
-            ));
-
-            match mode {
-                LoadMode::Local => {
-                    tenant
-                        .load_local(ctx)
-                        .instrument(info_span!("try_load", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))
-                        .await?;
-                }
-                LoadMode::Remote => {
-                    let preload = tenant
-                        .preload(&self.remote_storage, CancellationToken::new())
-                        .instrument(info_span!("try_load_preload", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))
-                        .await?;
-                    tenant
-                        .attach(Some(preload), SpawnMode::Normal, ctx)
-                        .instrument(info_span!("try_load", tenant_id=%self.tenant_shard_id.tenant_id, shard_id=%self.tenant_shard_id.shard_slug()))
-                        .await?;
-                }
-            }
-
-            tenant.state.send_replace(TenantState::Active);
-            for timeline in tenant.timelines.lock().unwrap().values() {
-                timeline.set_state(TimelineState::Active);
-            }
-            Ok(tenant)
-        }
-
-        /// For tests that specifically want to exercise the local load path, which does
-        /// not use remote storage.
-        pub async fn try_load_local(&self, ctx: &RequestContext) -> anyhow::Result<Arc<Tenant>> {
-            self.do_try_load(ctx, LoadMode::Local).await
-        }
-
-        /// The 'load' in this function is either a local load or a normal attachment,
-        pub async fn try_load(&self, ctx: &RequestContext) -> anyhow::Result<Arc<Tenant>> {
-            // If we have nothing in remote storage, must use load_local instead of attach: attach
-            // will error out if there are no timelines.
-            //
-            // See https://github.com/neondatabase/neon/issues/5456 for how we will eliminate
-            // this weird state of a Tenant which exists but doesn't have any timelines.
-            let mode = match self.remote_empty() {
-                true => LoadMode::Local,
-                false => LoadMode::Remote,
-            };
-
-            self.do_try_load(ctx, mode).await
-        }
-
         pub fn timeline_path(&self, timeline_id: &TimelineId) -> Utf8PathBuf {
             self.conf.timeline_path(&self.tenant_shard_id, timeline_id)
         }
@@ -4124,20 +4177,33 @@ pub(crate) mod harness {
             records: Vec<(Lsn, NeonWalRecord)>,
             _pg_version: u32,
         ) -> anyhow::Result<Bytes> {
-            let s = format!(
-                "redo for {} to get to {}, with {} and {} records",
-                key,
-                lsn,
-                if base_img.is_some() {
-                    "base image"
-                } else {
-                    "no base image"
-                },
-                records.len()
-            );
-            println!("{s}");
+            let records_neon = records.iter().all(|r| apply_neon::can_apply_in_neon(&r.1));
+            if records_neon {
+                // For Neon wal records, we can decode without spawning postgres, so do so.
+                let base_img = base_img.expect("Neon WAL redo requires base image").1;
+                let mut page = BytesMut::new();
+                page.extend_from_slice(&base_img);
+                for (_record_lsn, record) in records {
+                    apply_neon::apply_in_neon(&record, key, &mut page)?;
+                }
+                Ok(page.freeze())
+            } else {
+                // We never spawn a postgres walredo process in unit tests: just log what we might have done.
+                let s = format!(
+                    "redo for {} to get to {}, with {} and {} records",
+                    key,
+                    lsn,
+                    if base_img.is_some() {
+                        "base image"
+                    } else {
+                        "no base image"
+                    },
+                    records.len()
+                );
+                println!("{s}");
 
-            Ok(TEST_IMG(&s))
+                Ok(test_img(&s))
+            }
         }
     }
 }
@@ -4171,7 +4237,7 @@ mod tests {
             .put(
                 *TEST_KEY,
                 Lsn(0x10),
-                &Value::Image(TEST_IMG("foo at 0x10")),
+                &Value::Image(test_img("foo at 0x10")),
                 &ctx,
             )
             .await?;
@@ -4183,7 +4249,7 @@ mod tests {
             .put(
                 *TEST_KEY,
                 Lsn(0x20),
-                &Value::Image(TEST_IMG("foo at 0x20")),
+                &Value::Image(test_img("foo at 0x20")),
                 &ctx,
             )
             .await?;
@@ -4192,15 +4258,15 @@ mod tests {
 
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
-            TEST_IMG("foo at 0x10")
+            test_img("foo at 0x10")
         );
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
-            TEST_IMG("foo at 0x10")
+            test_img("foo at 0x10")
         );
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
-            TEST_IMG("foo at 0x20")
+            test_img("foo at 0x20")
         );
 
         Ok(())
@@ -4309,7 +4375,6 @@ mod tests {
         ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         let mut lsn = start_lsn;
-        #[allow(non_snake_case)]
         {
             let writer = tline.writer().await;
             // Create a relation on the timeline
@@ -4317,7 +4382,7 @@ mod tests {
                 .put(
                     *TEST_KEY,
                     lsn,
-                    &Value::Image(TEST_IMG(&format!("foo at {}", lsn))),
+                    &Value::Image(test_img(&format!("foo at {}", lsn))),
                     ctx,
                 )
                 .await?;
@@ -4327,7 +4392,7 @@ mod tests {
                 .put(
                     *TEST_KEY,
                     lsn,
-                    &Value::Image(TEST_IMG(&format!("foo at {}", lsn))),
+                    &Value::Image(test_img(&format!("foo at {}", lsn))),
                     ctx,
                 )
                 .await?;
@@ -4341,7 +4406,7 @@ mod tests {
                 .put(
                     *TEST_KEY,
                     lsn,
-                    &Value::Image(TEST_IMG(&format!("foo at {}", lsn))),
+                    &Value::Image(test_img(&format!("foo at {}", lsn))),
                     ctx,
                 )
                 .await?;
@@ -4351,7 +4416,7 @@ mod tests {
                 .put(
                     *TEST_KEY,
                     lsn,
-                    &Value::Image(TEST_IMG(&format!("foo at {}", lsn))),
+                    &Value::Image(test_img(&format!("foo at {}", lsn))),
                     ctx,
                 )
                 .await?;
@@ -4506,7 +4571,7 @@ mod tests {
         // Broken, as long as you don't need to access data from the parent.
         assert_eq!(
             newtline.get(*TEST_KEY, Lsn(0x70), &ctx).await?,
-            TEST_IMG(&format!("foo at {}", Lsn(0x70)))
+            test_img(&format!("foo at {}", Lsn(0x70)))
         );
 
         // This needs to traverse to the parent, and fails.
@@ -4583,7 +4648,7 @@ mod tests {
         // Check that the data is still accessible on the branch.
         assert_eq!(
             newtline.get(*TEST_KEY, Lsn(0x50), &ctx).await?,
-            TEST_IMG(&format!("foo at {}", Lsn(0x40)))
+            test_img(&format!("foo at {}", Lsn(0x40)))
         );
 
         Ok(())
@@ -4758,7 +4823,7 @@ mod tests {
             .put(
                 *TEST_KEY,
                 Lsn(0x10),
-                &Value::Image(TEST_IMG("foo at 0x10")),
+                &Value::Image(test_img("foo at 0x10")),
                 &ctx,
             )
             .await?;
@@ -4775,7 +4840,7 @@ mod tests {
             .put(
                 *TEST_KEY,
                 Lsn(0x20),
-                &Value::Image(TEST_IMG("foo at 0x20")),
+                &Value::Image(test_img("foo at 0x20")),
                 &ctx,
             )
             .await?;
@@ -4792,7 +4857,7 @@ mod tests {
             .put(
                 *TEST_KEY,
                 Lsn(0x30),
-                &Value::Image(TEST_IMG("foo at 0x30")),
+                &Value::Image(test_img("foo at 0x30")),
                 &ctx,
             )
             .await?;
@@ -4809,7 +4874,7 @@ mod tests {
             .put(
                 *TEST_KEY,
                 Lsn(0x40),
-                &Value::Image(TEST_IMG("foo at 0x40")),
+                &Value::Image(test_img("foo at 0x40")),
                 &ctx,
             )
             .await?;
@@ -4823,23 +4888,23 @@ mod tests {
 
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x10), &ctx).await?,
-            TEST_IMG("foo at 0x10")
+            test_img("foo at 0x10")
         );
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x1f), &ctx).await?,
-            TEST_IMG("foo at 0x10")
+            test_img("foo at 0x10")
         );
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x20), &ctx).await?,
-            TEST_IMG("foo at 0x20")
+            test_img("foo at 0x20")
         );
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x30), &ctx).await?,
-            TEST_IMG("foo at 0x30")
+            test_img("foo at 0x30")
         );
         assert_eq!(
             tline.get(*TEST_KEY, Lsn(0x40), &ctx).await?,
-            TEST_IMG("foo at 0x40")
+            test_img("foo at 0x40")
         );
 
         Ok(())
@@ -4871,7 +4936,7 @@ mod tests {
                     .put(
                         test_key,
                         lsn,
-                        &Value::Image(TEST_IMG(&format!("{} at {}", blknum, lsn))),
+                        &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
                         &ctx,
                     )
                     .await?;
@@ -4933,7 +4998,7 @@ mod tests {
                 .put(
                     test_key,
                     lsn,
-                    &Value::Image(TEST_IMG(&format!("{} at {}", blknum, lsn))),
+                    &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
                     &ctx,
                 )
                 .await?;
@@ -4954,7 +5019,7 @@ mod tests {
                     .put(
                         test_key,
                         lsn,
-                        &Value::Image(TEST_IMG(&format!("{} at {}", blknum, lsn))),
+                        &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
                         &ctx,
                     )
                     .await?;
@@ -4968,7 +5033,7 @@ mod tests {
                 test_key.field6 = blknum as u32;
                 assert_eq!(
                     tline.get(test_key, lsn, &ctx).await?,
-                    TEST_IMG(&format!("{} at {}", blknum, last_lsn))
+                    test_img(&format!("{} at {}", blknum, last_lsn))
                 );
             }
 
@@ -5022,7 +5087,7 @@ mod tests {
                 .put(
                     test_key,
                     lsn,
-                    &Value::Image(TEST_IMG(&format!("{} at {}", blknum, lsn))),
+                    &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
                     &ctx,
                 )
                 .await?;
@@ -5051,7 +5116,7 @@ mod tests {
                     .put(
                         test_key,
                         lsn,
-                        &Value::Image(TEST_IMG(&format!("{} at {}", blknum, lsn))),
+                        &Value::Image(test_img(&format!("{} at {}", blknum, lsn))),
                         &ctx,
                     )
                     .await?;
@@ -5066,7 +5131,7 @@ mod tests {
                 test_key.field6 = blknum as u32;
                 assert_eq!(
                     tline.get(test_key, lsn, &ctx).await?,
-                    TEST_IMG(&format!("{} at {}", blknum, last_lsn))
+                    test_img(&format!("{} at {}", blknum, last_lsn))
                 );
             }
 
@@ -5128,7 +5193,7 @@ mod tests {
                     .put(
                         test_key,
                         lsn,
-                        &Value::Image(TEST_IMG(&format!("{} {} at {}", idx, blknum, lsn))),
+                        &Value::Image(test_img(&format!("{} {} at {}", idx, blknum, lsn))),
                         &ctx,
                     )
                     .await?;
@@ -5150,7 +5215,7 @@ mod tests {
                 test_key.field6 = blknum as u32;
                 assert_eq!(
                     tline.get(test_key, *lsn, &ctx).await?,
-                    TEST_IMG(&format!("{idx} {blknum} at {lsn}"))
+                    test_img(&format!("{idx} {blknum} at {lsn}"))
                 );
             }
         }

pageserver/src/tenant/blob_io.rs

@@ -11,6 +11,9 @@
 //! len <  128: 0XXXXXXX
 //! len >= 128: 1XXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
 //!
+use bytes::{BufMut, BytesMut};
+use tokio_epoll_uring::{BoundedBuf, Slice};
+
 use crate::context::RequestContext;
 use crate::page_cache::PAGE_SZ;
 use crate::tenant::block_io::BlockCursor;
@@ -100,6 +103,8 @@ pub struct BlobWriter<const BUFFERED: bool> {
     offset: u64,
     /// A buffer to save on write calls, only used if BUFFERED=true
     buf: Vec<u8>,
+    /// We do tiny writes for the length headers; they need to be in an owned buffer;
+    io_buf: Option<BytesMut>,
 }
 
 impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
@@ -108,6 +113,7 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
             inner,
             offset: start_offset,
             buf: Vec::with_capacity(Self::CAPACITY),
+            io_buf: Some(BytesMut::new()),
         }
     }
 
@@ -117,21 +123,31 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
 
     const CAPACITY: usize = if BUFFERED { PAGE_SZ } else { 0 };
 
-    #[inline(always)]
     /// Writes the given buffer directly to the underlying `VirtualFile`.
     /// You need to make sure that the internal buffer is empty, otherwise
     /// data will be written in wrong order.
-    async fn write_all_unbuffered(&mut self, src_buf: &[u8]) -> Result<(), Error> {
-        self.inner.write_all(src_buf).await?;
-        self.offset += src_buf.len() as u64;
-        Ok(())
+    #[inline(always)]
+    async fn write_all_unbuffered<B: BoundedBuf>(
+        &mut self,
+        src_buf: B,
+    ) -> (B::Buf, Result<(), Error>) {
+        let (src_buf, res) = self.inner.write_all(src_buf).await;
+        let nbytes = match res {
+            Ok(nbytes) => nbytes,
+            Err(e) => return (src_buf, Err(e)),
+        };
+        self.offset += nbytes as u64;
+        (src_buf, Ok(()))
     }
 
     #[inline(always)]
     /// Flushes the internal buffer to the underlying `VirtualFile`.
     pub async fn flush_buffer(&mut self) -> Result<(), Error> {
-        self.inner.write_all(&self.buf).await?;
-        self.buf.clear();
+        let buf = std::mem::take(&mut self.buf);
+        let (mut buf, res) = self.inner.write_all(buf).await;
+        res?;
+        buf.clear();
+        self.buf = buf;
         Ok(())
     }
 
@@ -146,62 +162,91 @@ impl<const BUFFERED: bool> BlobWriter<BUFFERED> {
     }
 
     /// Internal, possibly buffered, write function
-    async fn write_all(&mut self, mut src_buf: &[u8]) -> Result<(), Error> {
+    async fn write_all<B: BoundedBuf>(&mut self, src_buf: B) -> (B::Buf, Result<(), Error>) {
         if !BUFFERED {
             assert!(self.buf.is_empty());
-            self.write_all_unbuffered(src_buf).await?;
-            return Ok(());
+            return self.write_all_unbuffered(src_buf).await;
         }
         let remaining = Self::CAPACITY - self.buf.len();
+        let src_buf_len = src_buf.bytes_init();
+        if src_buf_len == 0 {
+            return (Slice::into_inner(src_buf.slice_full()), Ok(()));
+        }
+        let mut src_buf = src_buf.slice(0..src_buf_len);
         // First try to copy as much as we can into the buffer
         if remaining > 0 {
-            let copied = self.write_into_buffer(src_buf);
-            src_buf = &src_buf[copied..];
+            let copied = self.write_into_buffer(&src_buf);
+            src_buf = src_buf.slice(copied..);
         }
         // Then, if the buffer is full, flush it out
         if self.buf.len() == Self::CAPACITY {
-            self.flush_buffer().await?;
+            if let Err(e) = self.flush_buffer().await {
+                return (Slice::into_inner(src_buf), Err(e));
+            }
         }
         // Finally, write the tail of src_buf:
         // If it wholly fits into the buffer without
         // completely filling it, then put it there.
         // If not, write it out directly.
-        if !src_buf.is_empty() {
+        let src_buf = if !src_buf.is_empty() {
             assert_eq!(self.buf.len(), 0);
             if src_buf.len() < Self::CAPACITY {
-                let copied = self.write_into_buffer(src_buf);
+                let copied = self.write_into_buffer(&src_buf);
                 // We just verified above that src_buf fits into our internal buffer.
                 assert_eq!(copied, src_buf.len());
+                Slice::into_inner(src_buf)
             } else {
-                self.write_all_unbuffered(src_buf).await?;
+                let (src_buf, res) = self.write_all_unbuffered(src_buf).await;
+                if let Err(e) = res {
+                    return (src_buf, Err(e));
+                }
+                src_buf
             }
-        }
-        Ok(())
+        } else {
+            Slice::into_inner(src_buf)
+        };
+        (src_buf, Ok(()))
     }
 
     /// Write a blob of data. Returns the offset that it was written to,
     /// which can be used to retrieve the data later.
-    pub async fn write_blob(&mut self, srcbuf: &[u8]) -> Result<u64, Error> {
+    pub async fn write_blob<B: BoundedBuf>(&mut self, srcbuf: B) -> (B::Buf, Result<u64, Error>) {
         let offset = self.offset;
 
-        if srcbuf.len() < 128 {
-            // Short blob. Write a 1-byte length header
-            let len_buf = srcbuf.len() as u8;
-            self.write_all(&[len_buf]).await?;
-        } else {
-            // Write a 4-byte length header
-            if srcbuf.len() > 0x7fff_ffff {
-                return Err(Error::new(
-                    ErrorKind::Other,
-                    format!("blob too large ({} bytes)", srcbuf.len()),
-                ));
+        let len = srcbuf.bytes_init();
+
+        let mut io_buf = self.io_buf.take().expect("we always put it back below");
+        io_buf.clear();
+        let (io_buf, hdr_res) = async {
+            if len < 128 {
+                // Short blob. Write a 1-byte length header
+                io_buf.put_u8(len as u8);
+                self.write_all(io_buf).await
+            } else {
+                // Write a 4-byte length header
+                if len > 0x7fff_ffff {
+                    return (
+                        io_buf,
+                        Err(Error::new(
+                            ErrorKind::Other,
+                            format!("blob too large ({} bytes)", len),
+                        )),
+                    );
+                }
+                let mut len_buf = (len as u32).to_be_bytes();
+                len_buf[0] |= 0x80;
+                io_buf.extend_from_slice(&len_buf[..]);
+                self.write_all(io_buf).await
             }
-            let mut len_buf = ((srcbuf.len()) as u32).to_be_bytes();
-            len_buf[0] |= 0x80;
-            self.write_all(&len_buf).await?;
         }
-        self.write_all(srcbuf).await?;
-        Ok(offset)
+        .await;
+        self.io_buf = Some(io_buf);
+        match hdr_res {
+            Ok(_) => (),
+            Err(e) => return (Slice::into_inner(srcbuf.slice(..)), Err(e)),
+        }
+        let (srcbuf, res) = self.write_all(srcbuf).await;
+        (srcbuf, res.map(|_| offset))
     }
 }
 
@@ -248,12 +293,14 @@ mod tests {
             let file = VirtualFile::create(pathbuf.as_path()).await?;
             let mut wtr = BlobWriter::<BUFFERED>::new(file, 0);
             for blob in blobs.iter() {
-                let offs = wtr.write_blob(blob).await?;
+                let (_, res) = wtr.write_blob(blob.clone()).await;
+                let offs = res?;
                 offsets.push(offs);
             }
             // Write out one page worth of zeros so that we can
             // read again with read_blk
-            let offs = wtr.write_blob(&vec![0; PAGE_SZ]).await?;
+            let (_, res) = wtr.write_blob(vec![0; PAGE_SZ]).await;
+            let offs = res?;
             println!("Writing final blob at offs={offs}");
             wtr.flush_buffer().await?;
         }

pageserver/src/tenant/delete.rs

@@ -3,10 +3,10 @@ use std::sync::Arc;
 use anyhow::Context;
 use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::{models::TenantState, shard::TenantShardId};
-use remote_storage::{GenericRemoteStorage, RemotePath};
+use remote_storage::{GenericRemoteStorage, RemotePath, TimeoutOrCancel};
 use tokio::sync::OwnedMutexGuard;
 use tokio_util::sync::CancellationToken;
-use tracing::{error, instrument, Instrument, Span};
+use tracing::{error, instrument, Instrument};
 
 use utils::{backoff, completion, crashsafe, fs_ext, id::TimelineId};
 
@@ -84,17 +84,17 @@ async fn create_remote_delete_mark(
             let data = bytes::Bytes::from_static(data);
             let stream = futures::stream::once(futures::future::ready(Ok(data)));
             remote_storage
-                .upload(stream, 0, &remote_mark_path, None)
+                .upload(stream, 0, &remote_mark_path, None, cancel)
                 .await
         },
-        |_e| false,
+        TimeoutOrCancel::caused_by_cancel,
         FAILED_UPLOAD_WARN_THRESHOLD,
         FAILED_REMOTE_OP_RETRIES,
         "mark_upload",
         cancel,
     )
     .await
-    .ok_or_else(|| anyhow::anyhow!("Cancelled"))
+    .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
     .and_then(|x| x)
     .context("mark_upload")?;
 
@@ -184,15 +184,15 @@ async fn remove_tenant_remote_delete_mark(
     if let Some(remote_storage) = remote_storage {
         let path = remote_tenant_delete_mark_path(conf, tenant_shard_id)?;
         backoff::retry(
-            || async { remote_storage.delete(&path).await },
-            |_e| false,
+            || async { remote_storage.delete(&path, cancel).await },
+            TimeoutOrCancel::caused_by_cancel,
             FAILED_UPLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "remove_tenant_remote_delete_mark",
             cancel,
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancelled"))
+        .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
         .and_then(|x| x)
         .context("remove_tenant_remote_delete_mark")?;
     }
@@ -496,11 +496,7 @@ impl DeleteTenantFlow {
                 };
                 Ok(())
             }
-            .instrument({
-                let span = tracing::info_span!(parent: None, "delete_tenant", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug());
-                span.follows_from(Span::current());
-                span
-            }),
+            .instrument(tracing::info_span!(parent: None, "delete_tenant", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug())),
         );
     }
 

pageserver/src/tenant/disk_btree.rs

@@ -36,7 +36,6 @@ use crate::{
 pub const VALUE_SZ: usize = 5;
 pub const MAX_VALUE: u64 = 0x007f_ffff_ffff;
 
-#[allow(dead_code)]
 pub const PAGE_SZ: usize = 8192;
 
 #[derive(Clone, Copy, Debug)]

pageserver/src/tenant/ephemeral_file.rs

@@ -6,6 +6,7 @@ use crate::context::RequestContext;
 use crate::page_cache::{self, PAGE_SZ};
 use crate::tenant::block_io::{BlockCursor, BlockLease, BlockReader};
 use crate::virtual_file::{self, VirtualFile};
+use bytes::BytesMut;
 use camino::Utf8PathBuf;
 use pageserver_api::shard::TenantShardId;
 use std::cmp::min;
@@ -26,7 +27,10 @@ pub struct EphemeralFile {
     /// An ephemeral file is append-only.
     /// We keep the last page, which can still be modified, in [`Self::mutable_tail`].
     /// The other pages, which can no longer be modified, are accessed through the page cache.
-    mutable_tail: [u8; PAGE_SZ],
+    ///
+    /// None <=> IO is ongoing.
+    /// Size is fixed to PAGE_SZ at creation time and must not be changed.
+    mutable_tail: Option<BytesMut>,
 }
 
 impl EphemeralFile {
@@ -60,7 +64,7 @@ impl EphemeralFile {
             _timeline_id: timeline_id,
             file,
             len: 0,
-            mutable_tail: [0u8; PAGE_SZ],
+            mutable_tail: Some(BytesMut::zeroed(PAGE_SZ)),
         })
     }
 
@@ -103,7 +107,13 @@ impl EphemeralFile {
             };
         } else {
             debug_assert_eq!(blknum as u64, self.len / PAGE_SZ as u64);
-            Ok(BlockLease::EphemeralFileMutableTail(&self.mutable_tail))
+            Ok(BlockLease::EphemeralFileMutableTail(
+                self.mutable_tail
+                    .as_deref()
+                    .expect("we're not doing IO, it must be Some()")
+                    .try_into()
+                    .expect("we ensure that it's always PAGE_SZ"),
+            ))
         }
     }
 
@@ -135,21 +145,27 @@ impl EphemeralFile {
             ) -> Result<(), io::Error> {
                 let mut src_remaining = src;
                 while !src_remaining.is_empty() {
-                    let dst_remaining = &mut self.ephemeral_file.mutable_tail[self.off..];
+                    let dst_remaining = &mut self
+                        .ephemeral_file
+                        .mutable_tail
+                        .as_deref_mut()
+                        .expect("IO is not yet ongoing")[self.off..];
                     let n = min(dst_remaining.len(), src_remaining.len());
                     dst_remaining[..n].copy_from_slice(&src_remaining[..n]);
                     self.off += n;
                     src_remaining = &src_remaining[n..];
                     if self.off == PAGE_SZ {
-                        match self
+                        let mutable_tail = std::mem::take(&mut self.ephemeral_file.mutable_tail)
+                            .expect("IO is not yet ongoing");
+                        let (mutable_tail, res) = self
                             .ephemeral_file
                             .file
-                            .write_all_at(
-                                &self.ephemeral_file.mutable_tail,
-                                self.blknum as u64 * PAGE_SZ as u64,
-                            )
-                            .await
-                        {
+                            .write_all_at(mutable_tail, self.blknum as u64 * PAGE_SZ as u64)
+                            .await;
+                        // TODO: If we panic before we can put the mutable_tail back, subsequent calls will fail.
+                        // I.e., the IO isn't retryable if we panic.
+                        self.ephemeral_file.mutable_tail = Some(mutable_tail);
+                        match res {
                             Ok(_) => {
                                 // Pre-warm the page cache with what we just wrote.
                                 // This isn't necessary for coherency/correctness, but it's how we've always done it.
@@ -169,7 +185,12 @@ impl EphemeralFile {
                                     Ok(page_cache::ReadBufResult::NotFound(mut write_guard)) => {
                                         let buf: &mut [u8] = write_guard.deref_mut();
                                         debug_assert_eq!(buf.len(), PAGE_SZ);
-                                        buf.copy_from_slice(&self.ephemeral_file.mutable_tail);
+                                        buf.copy_from_slice(
+                                            self.ephemeral_file
+                                                .mutable_tail
+                                                .as_deref()
+                                                .expect("IO is not ongoing"),
+                                        );
                                         let _ = write_guard.mark_valid();
                                         // pre-warm successful
                                     }
@@ -181,7 +202,11 @@ impl EphemeralFile {
                                 // Zero the buffer for re-use.
                                 // Zeroing is critical for correcntess because the write_blob code below
                                 // and similarly read_blk expect zeroed pages.
-                                self.ephemeral_file.mutable_tail.fill(0);
+                                self.ephemeral_file
+                                    .mutable_tail
+                                    .as_deref_mut()
+                                    .expect("IO is not ongoing")
+                                    .fill(0);
                                 // This block is done, move to next one.
                                 self.blknum += 1;
                                 self.off = 0;

pageserver/src/tenant/metadata.rs

@@ -279,7 +279,7 @@ pub async fn save_metadata(
     let path = conf.metadata_path(tenant_shard_id, timeline_id);
     let temp_path = path_with_suffix_extension(&path, TEMP_FILE_SUFFIX);
     let metadata_bytes = data.to_bytes().context("serialize metadata")?;
-    VirtualFile::crashsafe_overwrite(&path, &temp_path, &metadata_bytes)
+    VirtualFile::crashsafe_overwrite(&path, &temp_path, metadata_bytes)
         .await
         .context("write metadata")?;
     Ok(())

pageserver/src/tenant/mgr.rs

@@ -2,6 +2,8 @@
 //! page server.
 
 use camino::{Utf8DirEntry, Utf8Path, Utf8PathBuf};
+use futures::stream::StreamExt;
+use itertools::Itertools;
 use pageserver_api::key::Key;
 use pageserver_api::models::ShardParameters;
 use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, TenantShardId};
@@ -22,7 +24,7 @@ use tokio_util::sync::CancellationToken;
 use tracing::*;
 
 use remote_storage::GenericRemoteStorage;
-use utils::crashsafe;
+use utils::{completion, crashsafe};
 
 use crate::config::PageServerConf;
 use crate::context::{DownloadBehavior, RequestContext};
@@ -644,8 +646,6 @@ pub(crate) async fn shutdown_all_tenants() {
 }
 
 async fn shutdown_all_tenants0(tenants: &std::sync::RwLock<TenantsMap>) {
-    use utils::completion;
-
     let mut join_set = JoinSet::new();
 
     // Atomically, 1. create the shutdown tasks and 2. prevent creation of new tenants.
@@ -1200,7 +1200,7 @@ impl TenantManager {
         &self,
         tenant_shard_id: TenantShardId,
         drop_cache: bool,
-        ctx: RequestContext,
+        ctx: &RequestContext,
     ) -> anyhow::Result<()> {
         let mut slot_guard = tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
         let Some(old_slot) = slot_guard.get_old_value() else {
@@ -1253,7 +1253,7 @@ impl TenantManager {
             None,
             self.tenants,
             SpawnMode::Normal,
-            &ctx,
+            ctx,
         )?;
 
         slot_guard.upsert(TenantSlot::Attached(tenant))?;
@@ -1375,6 +1375,309 @@ impl TenantManager {
         slot_guard.revert();
         result
     }
+
+    #[instrument(skip_all, fields(tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), new_shard_count=%new_shard_count.0))]
+    pub(crate) async fn shard_split(
+        &self,
+        tenant_shard_id: TenantShardId,
+        new_shard_count: ShardCount,
+        ctx: &RequestContext,
+    ) -> anyhow::Result<Vec<TenantShardId>> {
+        let tenant = get_tenant(tenant_shard_id, true)?;
+
+        // Plan: identify what the new child shards will be
+        let effective_old_shard_count = std::cmp::max(tenant_shard_id.shard_count.0, 1);
+        if new_shard_count <= ShardCount(effective_old_shard_count) {
+            anyhow::bail!("Requested shard count is not an increase");
+        }
+        let expansion_factor = new_shard_count.0 / effective_old_shard_count;
+        if !expansion_factor.is_power_of_two() {
+            anyhow::bail!("Requested split is not a power of two");
+        }
+
+        let parent_shard_identity = tenant.shard_identity;
+        let parent_tenant_conf = tenant.get_tenant_conf();
+        let parent_generation = tenant.generation;
+
+        let child_shards = tenant_shard_id.split(new_shard_count);
+        tracing::info!(
+            "Shard {} splits into: {}",
+            tenant_shard_id.to_index(),
+            child_shards
+                .iter()
+                .map(|id| format!("{}", id.to_index()))
+                .join(",")
+        );
+
+        // Phase 1: Write out child shards' remote index files, in the parent tenant's current generation
+        if let Err(e) = tenant.split_prepare(&child_shards).await {
+            // If [`Tenant::split_prepare`] fails, we must reload the tenant, because it might
+            // have been left in a partially-shut-down state.
+            tracing::warn!("Failed to prepare for split: {e}, reloading Tenant before returning");
+            self.reset_tenant(tenant_shard_id, false, ctx).await?;
+            return Err(e);
+        }
+
+        self.resources.deletion_queue_client.flush_advisory();
+
+        // Phase 2: Put the parent shard to InProgress and grab a reference to the parent Tenant
+        drop(tenant);
+        let mut parent_slot_guard =
+            tenant_map_acquire_slot(&tenant_shard_id, TenantSlotAcquireMode::Any)?;
+        let parent = match parent_slot_guard.get_old_value() {
+            Some(TenantSlot::Attached(t)) => t,
+            Some(TenantSlot::Secondary(_)) => anyhow::bail!("Tenant location in secondary mode"),
+            Some(TenantSlot::InProgress(_)) => {
+                // tenant_map_acquire_slot never returns InProgress, if a slot was InProgress
+                // it would return an error.
+                unreachable!()
+            }
+            None => {
+                // We don't actually need the parent shard to still be attached to do our work, but it's
+                // a weird enough situation that the caller probably didn't want us to continue working
+                // if they had detached the tenant they requested the split on.
+                anyhow::bail!("Detached parent shard in the middle of split!")
+            }
+        };
+
+        // Optimization: hardlink layers from the parent into the children, so that they don't have to
+        // re-download & duplicate the data referenced in their initial IndexPart
+        self.shard_split_hardlink(parent, child_shards.clone())
+            .await?;
+
+        // Take a snapshot of where the parent's WAL ingest had got to: we will wait for
+        // child shards to reach this point.
+        let mut target_lsns = HashMap::new();
+        for timeline in parent.timelines.lock().unwrap().clone().values() {
+            target_lsns.insert(timeline.timeline_id, timeline.get_last_record_lsn());
+        }
+
+        // TODO: we should have the parent shard stop its WAL ingest here, it's a waste of resources
+        // and could slow down the children trying to catch up.
+
+        // Phase 3: Spawn the child shards
+        for child_shard in &child_shards {
+            let mut child_shard_identity = parent_shard_identity;
+            child_shard_identity.count = child_shard.shard_count;
+            child_shard_identity.number = child_shard.shard_number;
+
+            let child_location_conf = LocationConf {
+                mode: LocationMode::Attached(AttachedLocationConfig {
+                    generation: parent_generation,
+                    attach_mode: AttachmentMode::Single,
+                }),
+                shard: child_shard_identity,
+                tenant_conf: parent_tenant_conf,
+            };
+
+            self.upsert_location(
+                *child_shard,
+                child_location_conf,
+                None,
+                SpawnMode::Normal,
+                ctx,
+            )
+            .await?;
+        }
+
+        // Phase 4: wait for child chards WAL ingest to catch up to target LSN
+        for child_shard_id in &child_shards {
+            let child_shard_id = *child_shard_id;
+            let child_shard = {
+                let locked = TENANTS.read().unwrap();
+                let peek_slot =
+                    tenant_map_peek_slot(&locked, &child_shard_id, TenantSlotPeekMode::Read)?;
+                peek_slot.and_then(|s| s.get_attached()).cloned()
+            };
+            if let Some(t) = child_shard {
+                let timelines = t.timelines.lock().unwrap().clone();
+                for timeline in timelines.values() {
+                    let Some(target_lsn) = target_lsns.get(&timeline.timeline_id) else {
+                        continue;
+                    };
+
+                    tracing::info!(
+                        "Waiting for child shard {}/{} to reach target lsn {}...",
+                        child_shard_id,
+                        timeline.timeline_id,
+                        target_lsn
+                    );
+                    if let Err(e) = timeline.wait_lsn(*target_lsn, ctx).await {
+                        // Failure here might mean shutdown, in any case this part is an optimization
+                        // and we shouldn't hold up the split operation.
+                        tracing::warn!(
+                            "Failed to wait for timeline {} to reach lsn {target_lsn}: {e}",
+                            timeline.timeline_id
+                        );
+                    } else {
+                        tracing::info!(
+                            "Child shard {}/{} reached target lsn {}",
+                            child_shard_id,
+                            timeline.timeline_id,
+                            target_lsn
+                        );
+                    }
+                }
+            }
+        }
+
+        // Phase 5: Shut down the parent shard, and erase it from disk
+        let (_guard, progress) = completion::channel();
+        match parent.shutdown(progress, false).await {
+            Ok(()) => {}
+            Err(other) => {
+                other.wait().await;
+            }
+        }
+        let local_tenant_directory = self.conf.tenant_path(&tenant_shard_id);
+        let tmp_path = safe_rename_tenant_dir(&local_tenant_directory)
+            .await
+            .with_context(|| format!("local tenant directory {local_tenant_directory:?} rename"))?;
+        task_mgr::spawn(
+            task_mgr::BACKGROUND_RUNTIME.handle(),
+            TaskKind::MgmtRequest,
+            None,
+            None,
+            "tenant_files_delete",
+            false,
+            async move {
+                fs::remove_dir_all(tmp_path.as_path())
+                    .await
+                    .with_context(|| format!("tenant directory {:?} deletion", tmp_path))
+            },
+        );
+
+        parent_slot_guard.drop_old_value()?;
+
+        // Phase 6: Release the InProgress on the parent shard
+        drop(parent_slot_guard);
+
+        Ok(child_shards)
+    }
+
+    /// Part of [`Self::shard_split`]: hard link parent shard layers into child shards, as an optimization
+    /// to avoid the children downloading them again.
+    ///
+    /// For each resident layer in the parent shard, we will hard link it into all of the child shards.
+    async fn shard_split_hardlink(
+        &self,
+        parent_shard: &Tenant,
+        child_shards: Vec<TenantShardId>,
+    ) -> anyhow::Result<()> {
+        debug_assert_current_span_has_tenant_id();
+
+        let parent_path = self.conf.tenant_path(parent_shard.get_tenant_shard_id());
+        let (parent_timelines, parent_layers) = {
+            let mut parent_layers = Vec::new();
+            let timelines = parent_shard.timelines.lock().unwrap().clone();
+            let parent_timelines = timelines.keys().cloned().collect::<Vec<_>>();
+            for timeline in timelines.values() {
+                let timeline_layers = timeline
+                    .layers
+                    .read()
+                    .await
+                    .resident_layers()
+                    .collect::<Vec<_>>()
+                    .await;
+                for layer in timeline_layers {
+                    let relative_path = layer
+                        .local_path()
+                        .strip_prefix(&parent_path)
+                        .context("Removing prefix from parent layer path")?;
+                    parent_layers.push(relative_path.to_owned());
+                }
+            }
+            debug_assert!(
+                !parent_layers.is_empty(),
+                "shutdown cannot empty the layermap"
+            );
+            (parent_timelines, parent_layers)
+        };
+
+        let mut child_prefixes = Vec::new();
+        let mut create_dirs = Vec::new();
+
+        for child in child_shards {
+            let child_prefix = self.conf.tenant_path(&child);
+            create_dirs.push(child_prefix.clone());
+            create_dirs.extend(
+                parent_timelines
+                    .iter()
+                    .map(|t| self.conf.timeline_path(&child, t)),
+            );
+
+            child_prefixes.push(child_prefix);
+        }
+
+        // Since we will do a large number of small filesystem metadata operations, batch them into
+        // spawn_blocking calls rather than doing each one as a tokio::fs round-trip.
+        let jh = tokio::task::spawn_blocking(move || -> anyhow::Result<usize> {
+            for dir in &create_dirs {
+                if let Err(e) = std::fs::create_dir_all(dir) {
+                    // Ignore AlreadyExists errors, drop out on all other errors
+                    match e.kind() {
+                        std::io::ErrorKind::AlreadyExists => {}
+                        _ => {
+                            return Err(anyhow::anyhow!(e).context(format!("Creating {dir}")));
+                        }
+                    }
+                }
+            }
+
+            for child_prefix in child_prefixes {
+                for relative_layer in &parent_layers {
+                    let parent_path = parent_path.join(relative_layer);
+                    let child_path = child_prefix.join(relative_layer);
+                    if let Err(e) = std::fs::hard_link(&parent_path, &child_path) {
+                        match e.kind() {
+                            std::io::ErrorKind::AlreadyExists => {}
+                            std::io::ErrorKind::NotFound => {
+                                tracing::info!(
+                                    "Layer {} not found during hard-linking, evicted during split?",
+                                    relative_layer
+                                );
+                            }
+                            _ => {
+                                return Err(anyhow::anyhow!(e).context(format!(
+                                    "Hard linking {relative_layer} into {child_prefix}"
+                                )))
+                            }
+                        }
+                    }
+                }
+            }
+
+            // Durability is not required for correctness, but if we crashed during split and
+            // then came restarted with empty timeline dirs, it would be very inefficient to
+            // re-populate from remote storage.
+            for dir in create_dirs {
+                if let Err(e) = crashsafe::fsync(&dir) {
+                    // Something removed a newly created timeline dir out from underneath us?  Extremely
+                    // unexpected, but not worth panic'ing over as this whole function is just an
+                    // optimization.
+                    tracing::warn!("Failed to fsync directory {dir}: {e}")
+                }
+            }
+
+            Ok(parent_layers.len())
+        });
+
+        match jh.await {
+            Ok(Ok(layer_count)) => {
+                tracing::info!(count = layer_count, "Hard linked layers into child shards");
+            }
+            Ok(Err(e)) => {
+                // This is an optimization, so we tolerate failure.
+                tracing::warn!("Error hard-linking layers, proceeding anyway: {e}")
+            }
+            Err(e) => {
+                // This is something totally unexpected like a panic, so bail out.
+                anyhow::bail!("Error joining hard linking task: {e}");
+            }
+        }
+
+        Ok(())
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -2209,8 +2512,6 @@ async fn remove_tenant_from_memory<V, F>(
 where
     F: std::future::Future<Output = anyhow::Result<V>>,
 {
-    use utils::completion;
-
     let mut slot_guard =
         tenant_map_acquire_slot_impl(&tenant_shard_id, tenants, TenantSlotAcquireMode::MustExist)?;
 

pageserver/src/tenant/remote_timeline_client.rs

@@ -196,14 +196,12 @@ pub(crate) use upload::upload_initdb_dir;
 use utils::backoff::{
     self, exponential_backoff, DEFAULT_BASE_BACKOFF_SECONDS, DEFAULT_MAX_BACKOFF_SECONDS,
 };
-use utils::timeout::{timeout_cancellable, TimeoutCancellableError};
 
 use std::collections::{HashMap, VecDeque};
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Mutex};
-use std::time::Duration;
 
-use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath};
+use remote_storage::{DownloadError, GenericRemoteStorage, RemotePath, TimeoutOrCancel};
 use std::ops::DerefMut;
 use tracing::{debug, error, info, instrument, warn};
 use tracing::{info_span, Instrument};
@@ -217,6 +215,7 @@ use crate::metrics::{
 };
 use crate::task_mgr::shutdown_token;
 use crate::tenant::debug_assert_current_span_has_tenant_and_timeline_id;
+use crate::tenant::remote_timeline_client::download::download_retry;
 use crate::tenant::storage_layer::AsLayerDesc;
 use crate::tenant::upload_queue::Delete;
 use crate::tenant::TIMELINES_SEGMENT_NAME;
@@ -325,45 +324,6 @@ pub struct RemoteTimelineClient {
     cancel: CancellationToken,
 }
 
-/// This timeout is intended to deal with hangs in lower layers, e.g. stuck TCP flows.  It is not
-/// intended to be snappy enough for prompt shutdown, as we have a CancellationToken for that.
-const UPLOAD_TIMEOUT: Duration = Duration::from_secs(120);
-const DOWNLOAD_TIMEOUT: Duration = Duration::from_secs(120);
-
-/// Wrapper for timeout_cancellable that flattens result and converts TimeoutCancellableError to anyhow.
-///
-/// This is a convenience for the various upload functions.  In future
-/// the anyhow::Error result should be replaced with a more structured type that
-/// enables callers to avoid handling shutdown as an error.
-async fn upload_cancellable<F>(cancel: &CancellationToken, future: F) -> anyhow::Result<()>
-where
-    F: std::future::Future<Output = anyhow::Result<()>>,
-{
-    match timeout_cancellable(UPLOAD_TIMEOUT, cancel, future).await {
-        Ok(Ok(())) => Ok(()),
-        Ok(Err(e)) => Err(e),
-        Err(TimeoutCancellableError::Timeout) => Err(anyhow::anyhow!("Timeout")),
-        Err(TimeoutCancellableError::Cancelled) => Err(anyhow::anyhow!("Shutting down")),
-    }
-}
-/// Wrapper for timeout_cancellable that flattens result and converts TimeoutCancellableError to DownloaDError.
-async fn download_cancellable<F, R>(
-    cancel: &CancellationToken,
-    future: F,
-) -> Result<R, DownloadError>
-where
-    F: std::future::Future<Output = Result<R, DownloadError>>,
-{
-    match timeout_cancellable(DOWNLOAD_TIMEOUT, cancel, future).await {
-        Ok(Ok(r)) => Ok(r),
-        Ok(Err(e)) => Err(e),
-        Err(TimeoutCancellableError::Timeout) => {
-            Err(DownloadError::Other(anyhow::anyhow!("Timed out")))
-        }
-        Err(TimeoutCancellableError::Cancelled) => Err(DownloadError::Cancelled),
-    }
-}
-
 impl RemoteTimelineClient {
     ///
     /// Create a remote storage client for given timeline
@@ -506,7 +466,7 @@ impl RemoteTimelineClient {
     /// Download index file
     pub async fn download_index_file(
         &self,
-        cancel: CancellationToken,
+        cancel: &CancellationToken,
     ) -> Result<MaybeDeletedIndexPart, DownloadError> {
         let _unfinished_gauge_guard = self.metrics.call_begin(
             &RemoteOpFileKind::Index,
@@ -1049,7 +1009,7 @@ impl RemoteTimelineClient {
             &self.cancel,
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancelled"))
+        .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
         .and_then(|x| x)?;
 
         // all good, disarm the guard and mark as success
@@ -1081,14 +1041,14 @@ impl RemoteTimelineClient {
                 upload::preserve_initdb_archive(&self.storage_impl, tenant_id, timeline_id, cancel)
                     .await
             },
-            |_e| false,
+            TimeoutOrCancel::caused_by_cancel,
             FAILED_DOWNLOAD_WARN_THRESHOLD,
             FAILED_REMOTE_OP_RETRIES,
             "preserve_initdb_tar_zst",
             &cancel.clone(),
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancellled"))
+        .ok_or_else(|| anyhow::Error::new(TimeoutOrCancel::Cancel))
         .and_then(|x| x)
         .context("backing up initdb archive")?;
         Ok(())
@@ -1147,22 +1107,17 @@ impl RemoteTimelineClient {
 
         let cancel = shutdown_token();
 
-        let remaining = backoff::retry(
+        let remaining = download_retry(
             || async {
                 self.storage_impl
-                    .list_files(Some(&timeline_storage_path))
+                    .list_files(Some(&timeline_storage_path), None, &cancel)
                     .await
             },
-            |_e| false,
-            FAILED_DOWNLOAD_WARN_THRESHOLD,
-            FAILED_REMOTE_OP_RETRIES,
-            "list_prefixes",
+            "list remaining files",
             &cancel,
         )
         .await
-        .ok_or_else(|| anyhow::anyhow!("Cancelled!"))
-        .and_then(|x| x)
-        .context("list prefixes")?;
+        .context("list files remaining files")?;
 
         // We will delete the current index_part object last, since it acts as a deletion
         // marker via its deleted_at attribute
@@ -1351,6 +1306,7 @@ impl RemoteTimelineClient {
     /// queue.
     ///
     async fn perform_upload_task(self: &Arc<Self>, task: Arc<UploadTask>) {
+        let cancel = shutdown_token();
         // Loop to retry until it completes.
         loop {
             // If we're requested to shut down, close up shop and exit.
@@ -1362,7 +1318,7 @@ impl RemoteTimelineClient {
             // the Future, but we're not 100% sure if the remote storage library
             // is cancellation safe, so we don't dare to do that. Hopefully, the
             // upload finishes or times out soon enough.
-            if task_mgr::is_shutdown_requested() {
+            if cancel.is_cancelled() {
                 info!("upload task cancelled by shutdown request");
                 match self.stop() {
                     Ok(()) => {}
@@ -1448,6 +1404,10 @@ impl RemoteTimelineClient {
                 Ok(()) => {
                     break;
                 }
+                Err(e) if TimeoutOrCancel::caused_by_cancel(&e) => {
+                    // loop around to do the proper stopping
+                    continue;
+                }
                 Err(e) => {
                     let retries = task.retries.fetch_add(1, Ordering::SeqCst);
 
@@ -1473,7 +1433,7 @@ impl RemoteTimelineClient {
                         retries,
                         DEFAULT_BASE_BACKOFF_SECONDS,
                         DEFAULT_MAX_BACKOFF_SECONDS,
-                        &shutdown_token(),
+                        &cancel,
                     )
                     .await;
                 }
@@ -1703,23 +1663,6 @@ impl RemoteTimelineClient {
             }
         }
     }
-
-    pub(crate) fn get_layers_metadata(
-        &self,
-        layers: Vec<LayerFileName>,
-    ) -> anyhow::Result<Vec<Option<LayerFileMetadata>>> {
-        let q = self.upload_queue.lock().unwrap();
-        let q = match &*q {
-            UploadQueue::Stopped(_) | UploadQueue::Uninitialized => {
-                anyhow::bail!("queue is in state {}", q.as_str())
-            }
-            UploadQueue::Initialized(inner) => inner,
-        };
-
-        let decorated = layers.into_iter().map(|l| q.latest_files.get(&l).cloned());
-
-        Ok(decorated.collect())
-    }
 }
 
 pub fn remote_timelines_path(tenant_shard_id: &TenantShardId) -> RemotePath {
@@ -1990,7 +1933,7 @@ mod tests {
 
         // Download back the index.json, and check that the list of files is correct
         let initial_index_part = match client
-            .download_index_file(CancellationToken::new())
+            .download_index_file(&CancellationToken::new())
             .await
             .unwrap()
         {
@@ -2084,7 +2027,7 @@ mod tests {
 
         // Download back the index.json, and check that the list of files is correct
         let index_part = match client
-            .download_index_file(CancellationToken::new())
+            .download_index_file(&CancellationToken::new())
             .await
             .unwrap()
         {
@@ -2286,7 +2229,7 @@ mod tests {
         let client = test_state.build_client(get_generation);
 
         let download_r = client
-            .download_index_file(CancellationToken::new())
+            .download_index_file(&CancellationToken::new())
             .await
             .expect("download should always succeed");
         assert!(matches!(download_r, MaybeDeletedIndexPart::IndexPart(_)));

pageserver/src/tenant/remote_timeline_client/download.rs

@@ -11,16 +11,14 @@ use camino::{Utf8Path, Utf8PathBuf};
 use pageserver_api::shard::TenantShardId;
 use tokio::fs::{self, File, OpenOptions};
 use tokio::io::{AsyncSeekExt, AsyncWriteExt};
+use tokio_util::io::StreamReader;
 use tokio_util::sync::CancellationToken;
 use tracing::warn;
-use utils::timeout::timeout_cancellable;
 use utils::{backoff, crashsafe};
 
 use crate::config::PageServerConf;
 use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
-use crate::tenant::remote_timeline_client::{
-    download_cancellable, remote_layer_path, remote_timelines_path, DOWNLOAD_TIMEOUT,
-};
+use crate::tenant::remote_timeline_client::{remote_layer_path, remote_timelines_path};
 use crate::tenant::storage_layer::LayerFileName;
 use crate::tenant::Generation;
 use crate::virtual_file::on_fatal_io_error;
@@ -83,15 +81,13 @@ pub async fn download_layer_file<'a>(
                 .with_context(|| format!("create a destination file for layer '{temp_file_path}'"))
                 .map_err(DownloadError::Other)?;
 
-            // Cancellation safety: it is safe to cancel this future, because it isn't writing to a local
-            // file: the write to local file doesn't start until after the request header is returned
-            // and we start draining the body stream below
-            let download = download_cancellable(cancel, storage.download(&remote_path))
+            let download = storage
+                .download(&remote_path, cancel)
                 .await
                 .with_context(|| {
                     format!(
-                    "open a download stream for layer with remote storage path '{remote_path:?}'"
-                )
+                        "open a download stream for layer with remote storage path '{remote_path:?}'"
+                    )
                 })
                 .map_err(DownloadError::Other)?;
 
@@ -100,43 +96,26 @@ pub async fn download_layer_file<'a>(
 
             let mut reader = tokio_util::io::StreamReader::new(download.download_stream);
 
-            // Cancellation safety: it is safe to cancel this future because it is writing into a temporary file,
-            // and we will unlink the temporary file if there is an error.  This unlink is important because we
-            // are in a retry loop, and we wouldn't want to leave behind a rogue write I/O to a file that
-            // we will imminiently try and write to again.
-            let bytes_amount: u64 = match timeout_cancellable(
-                DOWNLOAD_TIMEOUT,
-                cancel,
-                tokio::io::copy_buf(&mut reader, &mut destination_file),
-            )
-            .await
-            .with_context(|| {
-                format!(
+            let bytes_amount = tokio::io::copy_buf(&mut reader, &mut destination_file)
+                .await
+                .with_context(|| format!(
                     "download layer at remote path '{remote_path:?}' into file {temp_file_path:?}"
-                )
-            })
-            .map_err(DownloadError::Other)?
-            {
-                Ok(b) => Ok(b),
+                ))
+                .map_err(DownloadError::Other);
+
+            match bytes_amount {
+                Ok(bytes_amount) => {
+                    let destination_file = destination_file.into_inner();
+                    Ok((destination_file, bytes_amount))
+                }
                 Err(e) => {
-                    // Remove incomplete files: on restart Timeline would do this anyway, but we must
-                    // do it here for the retry case.
                     if let Err(e) = tokio::fs::remove_file(&temp_file_path).await {
                         on_fatal_io_error(&e, &format!("Removing temporary file {temp_file_path}"));
                     }
+
                     Err(e)
                 }
             }
-            .with_context(|| {
-                format!(
-                    "download layer at remote path '{remote_path:?}' into file {temp_file_path:?}"
-                )
-            })
-            .map_err(DownloadError::Other)?;
-
-            let destination_file = destination_file.into_inner();
-
-            Ok((destination_file, bytes_amount))
         },
         &format!("download {remote_path:?}"),
         cancel,
@@ -216,16 +195,17 @@ pub async fn list_remote_timelines(
         anyhow::bail!("storage-sync-list-remote-timelines");
     });
 
-    let cancel_inner = cancel.clone();
     let listing = download_retry_forever(
         || {
-            download_cancellable(
-                &cancel_inner,
-                storage.list(Some(&remote_path), ListingMode::WithDelimiter),
+            storage.list(
+                Some(&remote_path),
+                ListingMode::WithDelimiter,
+                None,
+                &cancel,
             )
         },
         &format!("list timelines for {tenant_shard_id}"),
-        cancel,
+        &cancel,
     )
     .await?;
 
@@ -258,29 +238,25 @@ async fn do_download_index_part(
     tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     index_generation: Generation,
-    cancel: CancellationToken,
+    cancel: &CancellationToken,
 ) -> Result<IndexPart, DownloadError> {
-    use futures::stream::StreamExt;
-
     let remote_path = remote_index_path(tenant_shard_id, timeline_id, index_generation);
 
-    let cancel_inner = cancel.clone();
     let index_part_bytes = download_retry_forever(
         || async {
-            // Cancellation: if is safe to cancel this future because we're just downloading into
-            // a memory buffer, not touching local disk.
-            let index_part_download =
-                download_cancellable(&cancel_inner, storage.download(&remote_path)).await?;
+            let download = storage.download(&remote_path, cancel).await?;
 
-            let mut index_part_bytes = Vec::new();
-            let mut stream = std::pin::pin!(index_part_download.download_stream);
-            while let Some(chunk) = stream.next().await {
-                let chunk = chunk
-                    .with_context(|| format!("download index part at {remote_path:?}"))
-                    .map_err(DownloadError::Other)?;
-                index_part_bytes.extend_from_slice(&chunk[..]);
-            }
-            Ok(index_part_bytes)
+            let mut bytes = Vec::new();
+
+            let stream = download.download_stream;
+            let mut stream = StreamReader::new(stream);
+
+            tokio::io::copy_buf(&mut stream, &mut bytes)
+                .await
+                .with_context(|| format!("download index part at {remote_path:?}"))
+                .map_err(DownloadError::Other)?;
+
+            Ok(bytes)
         },
         &format!("download {remote_path:?}"),
         cancel,
@@ -288,7 +264,7 @@ async fn do_download_index_part(
     .await?;
 
     let index_part: IndexPart = serde_json::from_slice(&index_part_bytes)
-        .with_context(|| format!("download index part file at {remote_path:?}"))
+        .with_context(|| format!("deserialize index part file at {remote_path:?}"))
         .map_err(DownloadError::Other)?;
 
     Ok(index_part)
@@ -305,7 +281,7 @@ pub(super) async fn download_index_part(
     tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
     my_generation: Generation,
-    cancel: CancellationToken,
+    cancel: &CancellationToken,
 ) -> Result<IndexPart, DownloadError> {
     debug_assert_current_span_has_tenant_and_timeline_id();
 
@@ -325,14 +301,8 @@ pub(super) async fn download_index_part(
     // index in our generation.
     //
     // This is an optimization to avoid doing the listing for the general case below.
-    let res = do_download_index_part(
-        storage,
-        tenant_shard_id,
-        timeline_id,
-        my_generation,
-        cancel.clone(),
-    )
-    .await;
+    let res =
+        do_download_index_part(storage, tenant_shard_id, timeline_id, my_generation, cancel).await;
     match res {
         Ok(index_part) => {
             tracing::debug!(
@@ -357,7 +327,7 @@ pub(super) async fn download_index_part(
         tenant_shard_id,
         timeline_id,
         my_generation.previous(),
-        cancel.clone(),
+        cancel,
     )
     .await;
     match res {
@@ -379,18 +349,13 @@ pub(super) async fn download_index_part(
     // objects, and select the highest one with a generation <= my_generation.  Constructing the prefix is equivalent
     // to constructing a full index path with no generation, because the generation is a suffix.
     let index_prefix = remote_index_path(tenant_shard_id, timeline_id, Generation::none());
-    let indices = backoff::retry(
-        || async { storage.list_files(Some(&index_prefix)).await },
-        |_| false,
-        FAILED_DOWNLOAD_WARN_THRESHOLD,
-        FAILED_REMOTE_OP_RETRIES,
-        "listing index_part files",
-        &cancel,
+
+    let indices = download_retry(
+        || async { storage.list_files(Some(&index_prefix), None, cancel).await },
+        "list index_part files",
+        cancel,
     )
-    .await
-    .ok_or_else(|| anyhow::anyhow!("Cancelled"))
-    .and_then(|x| x)
-    .map_err(DownloadError::Other)?;
+    .await?;
 
     // General case logic for which index to use: the latest index whose generation
     // is <= our own.  See "Finding the remote indices for timelines" in docs/rfcs/025-generation-numbers.md
@@ -447,8 +412,6 @@ pub(crate) async fn download_initdb_tar_zst(
         "{INITDB_PATH}.download-{timeline_id}.{TEMP_FILE_SUFFIX}"
     ));
 
-    let cancel_inner = cancel.clone();
-
     let file = download_retry(
         || async {
             let file = OpenOptions::new()
@@ -461,13 +424,10 @@ pub(crate) async fn download_initdb_tar_zst(
                 .with_context(|| format!("tempfile creation {temp_path}"))
                 .map_err(DownloadError::Other)?;
 
-            let download = match download_cancellable(&cancel_inner, storage.download(&remote_path))
-                .await
-            {
+            let download = match storage.download(&remote_path, cancel).await {
                 Ok(dl) => dl,
                 Err(DownloadError::NotFound) => {
-                    download_cancellable(&cancel_inner, storage.download(&remote_preserved_path))
-                        .await?
+                    storage.download(&remote_preserved_path, cancel).await?
                 }
                 Err(other) => Err(other)?,
             };
@@ -477,6 +437,7 @@ pub(crate) async fn download_initdb_tar_zst(
             // TODO: this consumption of the response body should be subject to timeout + cancellation, but
             // not without thinking carefully about how to recover safely from cancelling a write to
             // local storage (e.g. by writing into a temp file as we do in download_layer)
+            // FIXME: flip the weird error wrapping
             tokio::io::copy_buf(&mut download, &mut writer)
                 .await
                 .with_context(|| format!("download initdb.tar.zst at {remote_path:?}"))
@@ -516,7 +477,7 @@ pub(crate) async fn download_initdb_tar_zst(
 /// with backoff.
 ///
 /// (See similar logic for uploads in `perform_upload_task`)
-async fn download_retry<T, O, F>(
+pub(super) async fn download_retry<T, O, F>(
     op: O,
     description: &str,
     cancel: &CancellationToken,
@@ -527,7 +488,7 @@ where
 {
     backoff::retry(
         op,
-        |e| matches!(e, DownloadError::BadInput(_) | DownloadError::NotFound),
+        DownloadError::is_permanent,
         FAILED_DOWNLOAD_WARN_THRESHOLD,
         FAILED_REMOTE_OP_RETRIES,
         description,
@@ -541,7 +502,7 @@ where
 async fn download_retry_forever<T, O, F>(
     op: O,
     description: &str,
-    cancel: CancellationToken,
+    cancel: &CancellationToken,
 ) -> Result<T, DownloadError>
 where
     O: FnMut() -> F,
@@ -549,11 +510,11 @@ where
 {
     backoff::retry(
         op,
-        |e| matches!(e, DownloadError::BadInput(_) | DownloadError::NotFound),
+        DownloadError::is_permanent,
         FAILED_DOWNLOAD_WARN_THRESHOLD,
         u32::MAX,
         description,
-        &cancel,
+        cancel,
     )
     .await
     .ok_or_else(|| DownloadError::Cancelled)

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -16,7 +16,7 @@ use crate::{
     config::PageServerConf,
     tenant::remote_timeline_client::{
         index::IndexPart, remote_index_path, remote_initdb_archive_path,
-        remote_initdb_preserved_archive_path, remote_path, upload_cancellable,
+        remote_initdb_preserved_archive_path, remote_path,
     },
 };
 use remote_storage::{GenericRemoteStorage, TimeTravelError};
@@ -27,7 +27,7 @@ use super::index::LayerFileMetadata;
 use tracing::info;
 
 /// Serializes and uploads the given index part data to the remote storage.
-pub(super) async fn upload_index_part<'a>(
+pub(crate) async fn upload_index_part<'a>(
     storage: &'a GenericRemoteStorage,
     tenant_shard_id: &TenantShardId,
     timeline_id: &TimelineId,
@@ -49,16 +49,15 @@ pub(super) async fn upload_index_part<'a>(
     let index_part_bytes = bytes::Bytes::from(index_part_bytes);
 
     let remote_path = remote_index_path(tenant_shard_id, timeline_id, generation);
-    upload_cancellable(
-        cancel,
-        storage.upload_storage_object(
+    storage
+        .upload_storage_object(
             futures::stream::once(futures::future::ready(Ok(index_part_bytes))),
             index_part_size,
             &remote_path,
-        ),
-    )
-    .await
-    .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
+            cancel,
+        )
+        .await
+        .with_context(|| format!("upload index part for '{tenant_shard_id} / {timeline_id}'"))
 }
 
 /// Attempts to upload given layer files.
@@ -115,11 +114,10 @@ pub(super) async fn upload_timeline_layer<'a>(
 
     let reader = tokio_util::io::ReaderStream::with_capacity(source_file, super::BUFFER_SIZE);
 
-    upload_cancellable(cancel, storage.upload(reader, fs_size, &storage_path, None))
+    storage
+        .upload(reader, fs_size, &storage_path, None, cancel)
         .await
-        .with_context(|| format!("upload layer from local path '{source_path}'"))?;
-
-    Ok(())
+        .with_context(|| format!("upload layer from local path '{source_path}'"))
 }
 
 /// Uploads the given `initdb` data to the remote storage.
@@ -139,12 +137,10 @@ pub(crate) async fn upload_initdb_dir(
     let file = tokio_util::io::ReaderStream::with_capacity(initdb_tar_zst, super::BUFFER_SIZE);
 
     let remote_path = remote_initdb_archive_path(tenant_id, timeline_id);
-    upload_cancellable(
-        cancel,
-        storage.upload_storage_object(file, size as usize, &remote_path),
-    )
-    .await
-    .with_context(|| format!("upload initdb dir for '{tenant_id} / {timeline_id}'"))
+    storage
+        .upload_storage_object(file, size as usize, &remote_path, cancel)
+        .await
+        .with_context(|| format!("upload initdb dir for '{tenant_id} / {timeline_id}'"))
 }
 
 pub(crate) async fn preserve_initdb_archive(
@@ -155,7 +151,8 @@ pub(crate) async fn preserve_initdb_archive(
 ) -> anyhow::Result<()> {
     let source_path = remote_initdb_archive_path(tenant_id, timeline_id);
     let dest_path = remote_initdb_preserved_archive_path(tenant_id, timeline_id);
-    upload_cancellable(cancel, storage.copy_object(&source_path, &dest_path))
+    storage
+        .copy_object(&source_path, &dest_path, cancel)
         .await
         .with_context(|| format!("backing up initdb archive for '{tenant_id} / {timeline_id}'"))
 }