refactor thread local accesses

## Problem

As reported in #10441 the `control_plane/README/md` incorrectly
specified that `--pg-version` should be specified in the `cargo neon
init` command. This is not the case and causes an invalid argument
error.

## Summary of changes

Fix the README

## Test Plan

I verified that the steps in the README now work locally. I connected to
the started postgres endpoint and executed some basic metadata queries.

.config/hakari.toml

@@ -30,6 +30,7 @@ workspace-members = [
     "vm_monitor",
     # All of these exist in libs and are not usually built independently.
     # Putting workspace hack there adds a bottleneck for cargo builds.
+    "alloc-metrics",
     "compute_api",
     "consumption_metrics",
     "desim",

.dockerignore

@@ -27,4 +27,4 @@
 !storage_controller/
 !vendor/postgres-*/
 !workspace_hack/
-!build_tools/patches
+!build-tools/patches

.github/actionlint.yml

@@ -31,6 +31,7 @@ config-variables:
   - NEON_PROD_AWS_ACCOUNT_ID
   - PGREGRESS_PG16_PROJECT_ID
   - PGREGRESS_PG17_PROJECT_ID
+  - PREWARM_PGBENCH_SIZE
   - REMOTE_STORAGE_AZURE_CONTAINER
   - REMOTE_STORAGE_AZURE_REGION
   - SLACK_CICD_CHANNEL_ID

.github/actions/run-python-test-set/action.yml

@@ -176,7 +176,13 @@ runs:
         fi
 
         if [[ $BUILD_TYPE == "debug" && $RUNNER_ARCH == 'X64' ]]; then
-          cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          # We don't use code coverage for regression tests (the step is disabled),
+          # so there's no need to collect it.
+          # Ref https://github.com/neondatabase/neon/issues/4540
+          # cov_prefix=(scripts/coverage "--profraw-prefix=$GITHUB_JOB" --dir=/tmp/coverage run)
+          cov_prefix=()
+          # Explicitly set LLVM_PROFILE_FILE to /dev/null to avoid writing *.profraw files
+          export LLVM_PROFILE_FILE=/dev/null
         else
           cov_prefix=()
         fi

.github/workflows/_build-and-test-locally.yml

@@ -150,7 +150,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v14
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v14_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v15 build
         id: cache_pg_15
@@ -162,7 +162,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v15
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v15_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v16 build
         id: cache_pg_16
@@ -174,7 +174,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v16
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v16_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Cache postgres v17 build
         id: cache_pg_17
@@ -186,7 +186,7 @@ jobs:
           secretKey: ${{ secrets.HETZNER_CACHE_SECRET_KEY }}
           use-fallback: false
           path: pg_install/v17
-          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools.Dockerfile') }}
+          key: v1-${{ runner.os }}-${{ runner.arch }}-${{ inputs.build-type }}-pg-${{ steps.pg_v17_rev.outputs.pg_rev }}-bookworm-${{ hashFiles('Makefile', 'build-tools/Dockerfile') }}
 
       - name: Build all
         # Note: the Makefile picks up BUILD_TYPE and CARGO_PROFILE from the env variables

.github/workflows/benchmarking.yml

@@ -219,6 +219,7 @@ jobs:
           --ignore test_runner/performance/test_cumulative_statistics_persistence.py
           --ignore test_runner/performance/test_perf_many_relations.py
           --ignore test_runner/performance/test_perf_oltp_large_tenant.py
+          --ignore test_runner/performance/test_lfc_prewarm.py
       env:
         BENCHMARK_CONNSTR: ${{ steps.create-neon-project.outputs.dsn }}
         VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
@@ -410,6 +411,77 @@ jobs:
       env:
         SLACK_BOT_TOKEN: ${{ secrets.SLACK_BOT_TOKEN }}
 
+  prewarm-test:
+    if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
+    permissions:
+      contents: write
+      statuses: write
+      id-token: write # aws-actions/configure-aws-credentials
+    env:
+      PGBENCH_SIZE: ${{ vars.PREWARM_PGBENCH_SIZE }}
+      POSTGRES_DISTRIB_DIR: /tmp/neon/pg_install
+      DEFAULT_PG_VERSION: 17
+      TEST_OUTPUT: /tmp/test_output
+      BUILD_TYPE: remote
+      SAVE_PERF_REPORT: ${{ github.event.inputs.save_perf_report || ( github.ref_name == 'main' ) }}
+      PLATFORM: "neon-staging"
+
+    runs-on: [ self-hosted, us-east-2, x64 ]
+    container:
+      image: ghcr.io/neondatabase/build-tools:pinned-bookworm
+      credentials:
+        username: ${{ github.actor }}
+        password: ${{ secrets.GITHUB_TOKEN }}
+      options: --init
+
+    steps:
+    - name: Harden the runner (Audit all outbound calls)
+      uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
+      with:
+        egress-policy: audit
+
+    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+
+    - name: Configure AWS credentials
+      uses: aws-actions/configure-aws-credentials@e3dd6a429d7300a6a4c196c26e071d42e0343502 # v4.0.2
+      with:
+        aws-region: eu-central-1
+        role-to-assume: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+        role-duration-seconds: 18000 # 5 hours
+
+    - name: Download Neon artifact
+      uses: ./.github/actions/download
+      with:
+        name: neon-${{ runner.os }}-${{ runner.arch }}-release-artifact
+        path: /tmp/neon/
+        prefix: latest
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+
+    - name: Run prewarm benchmark
+      uses: ./.github/actions/run-python-test-set
+      with:
+        build_type: ${{ env.BUILD_TYPE }}
+        test_selection: performance/test_lfc_prewarm.py
+        run_in_parallel: false
+        save_perf_report: ${{ env.SAVE_PERF_REPORT }}
+        extra_params: -m remote_cluster --timeout 5400
+        pg_version: ${{ env.DEFAULT_PG_VERSION }}
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        VIP_VAP_ACCESS_TOKEN: "${{ secrets.VIP_VAP_ACCESS_TOKEN }}"
+        PERF_TEST_RESULT_CONNSTR: "${{ secrets.PERF_TEST_RESULT_CONNSTR }}"
+        NEON_API_KEY: ${{ secrets.NEON_STAGING_API_KEY }}
+
+    - name: Create Allure report
+      id: create-allure-report
+      if: ${{ !cancelled() }}
+      uses: ./.github/actions/allure-report-generate
+      with:
+        store-test-results-into-db: true
+        aws-oidc-role-arn: ${{ vars.DEV_AWS_OIDC_ROLE_ARN }}
+      env:
+        REGRESS_TEST_RESULT_CONNSTR_NEW: ${{ secrets.REGRESS_TEST_RESULT_CONNSTR_NEW }}
+
   generate-matrices:
     if: ${{ github.event.inputs.run_only_pgvector_tests == 'false' || github.event.inputs.run_only_pgvector_tests == null }}
     # Create matrices for the benchmarking jobs, so we run benchmarks on rds only once a week (on Saturday)

.github/workflows/build-build-tools-image.yml

@@ -72,7 +72,7 @@ jobs:
           ARCHS: ${{ inputs.archs || '["x64","arm64"]' }}
           DEBIANS: ${{ inputs.debians || '["bullseye","bookworm"]' }}
           IMAGE_TAG: |
-            ${{ hashFiles('build-tools.Dockerfile',
+            ${{ hashFiles('build-tools/Dockerfile',
                           '.github/workflows/build-build-tools-image.yml') }}
         run: |
           echo "archs=${ARCHS}"           | tee -a ${GITHUB_OUTPUT}
@@ -144,7 +144,7 @@ jobs:
 
       - uses: docker/build-push-action@471d1dc4e07e5cdedd4c2171150001c434f0b7a4 # v6.15.0
         with:
-          file: build-tools.Dockerfile
+          file: build-tools/Dockerfile
           context: .
           provenance: false
           push: true

.github/workflows/build_and_test.yml

@@ -87,22 +87,27 @@ jobs:
     uses: ./.github/workflows/build-build-tools-image.yml
     secrets: inherit
 
-  lint-openapi-spec:
-    runs-on: ubuntu-22.04
-    needs: [ meta, check-permissions ]
+  lint-yamls:
+    needs: [ meta, check-permissions, build-build-tools-image ]
     # We do need to run this in `.*-rc-pr` because of hotfixes.
     if: ${{ contains(fromJSON('["pr", "push-main", "storage-rc-pr", "proxy-rc-pr", "compute-rc-pr"]'), needs.meta.outputs.run-kind) }}
+    runs-on: [ self-hosted, small ]
+    container:
+      image: ${{ needs.build-build-tools-image.outputs.image }}
+      credentials:
+        username: ${{ github.actor }}
+        password: ${{ secrets.GITHUB_TOKEN }}
+      options: --init
+
     steps:
       - name: Harden the runner (Audit all outbound calls)
         uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
         with:
           egress-policy: audit
+
       - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-      - uses: docker/login-action@74a5d142397b4f367a81961eba4e8cd7edddf772 # v3.4.0
-        with:
-          registry: ghcr.io
-          username: ${{ github.actor }}
-          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - run: make -C compute manifest-schema-validation
       - run: make lint-openapi-spec
 
   check-codestyle-python:
@@ -217,28 +222,6 @@ jobs:
       build-tools-image: ${{ needs.build-build-tools-image.outputs.image }}-bookworm
     secrets: inherit
 
-  validate-compute-manifest:
-    runs-on: ubuntu-22.04
-    needs: [ meta, check-permissions ]
-    # We do need to run this in `.*-rc-pr` because of hotfixes.
-    if: ${{ contains(fromJSON('["pr", "push-main", "storage-rc-pr", "proxy-rc-pr", "compute-rc-pr"]'), needs.meta.outputs.run-kind) }}
-    steps:
-      - name: Harden the runner (Audit all outbound calls)
-        uses: step-security/harden-runner@4d991eb9b905ef189e4c376166672c3f2f230481 # v2.11.0
-        with:
-          egress-policy: audit
-
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Set up Node.js
-        uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020 # v4.4.0
-        with:
-          node-version: '24'
-
-      - name: Validate manifest against schema
-        run: |
-          make -C compute manifest-schema-validation
-
   build-and-test-locally:
     needs: [ meta, build-build-tools-image ]
     # We do need to run this in `.*-rc-pr` because of hotfixes.

.gitignore

@@ -29,3 +29,6 @@ docker-compose/docker-compose-parallel.yml
 
 # pgindent typedef lists
 *.list
+
+# Node
+**/node_modules/

.gitmodules

@@ -1,16 +1,16 @@
 [submodule "vendor/postgres-v14"]
 	path = vendor/postgres-v14
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_14_STABLE_neon
 [submodule "vendor/postgres-v15"]
 	path = vendor/postgres-v15
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_15_STABLE_neon
 [submodule "vendor/postgres-v16"]
 	path = vendor/postgres-v16
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_16_STABLE_neon
 [submodule "vendor/postgres-v17"]
 	path = vendor/postgres-v17
-	url = https://github.com/neondatabase/postgres.git
+	url = ../postgres.git
 	branch = REL_17_STABLE_neon

Cargo.lock

@@ -61,6 +61,17 @@ dependencies = [
  "equator",
 ]
 
+[[package]]
+name = "alloc-metrics"
+version = "0.1.0"
+dependencies = [
+ "criterion",
+ "measured",
+ "metrics",
+ "thread_local",
+ "tikv-jemallocator",
+]
+
 [[package]]
 name = "allocator-api2"
 version = "0.2.16"
@@ -687,40 +698,13 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "axum"
-version = "0.7.9"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "edca88bc138befd0323b20752846e6587272d3b03b0343c8ea28a6f819e6e71f"
-dependencies = [
- "async-trait",
- "axum-core 0.4.5",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "itoa",
- "matchit 0.7.3",
- "memchr",
- "mime",
- "percent-encoding",
- "pin-project-lite",
- "rustversion",
- "serde",
- "sync_wrapper 1.0.1",
- "tower 0.5.2",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum"
 version = "0.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6d6fd624c75e18b3b4c6b9caf42b1afe24437daaee904069137d8bab077be8b8"
 dependencies = [
- "axum-core 0.5.0",
+ "axum-core",
  "base64 0.22.1",
  "bytes",
  "form_urlencoded",
@@ -731,7 +715,7 @@ dependencies = [
  "hyper 1.4.1",
  "hyper-util",
  "itoa",
- "matchit 0.8.4",
+ "matchit",
  "memchr",
  "mime",
  "percent-encoding",
@@ -751,26 +735,6 @@ dependencies = [
  "tracing",
 ]
 
-[[package]]
-name = "axum-core"
-version = "0.4.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09f2bd6146b97ae3359fa0cc6d6b376d9539582c7b4220f041a33ec24c226199"
-dependencies = [
- "async-trait",
- "bytes",
- "futures-util",
- "http 1.1.0",
- "http-body 1.0.0",
- "http-body-util",
- "mime",
- "pin-project-lite",
- "rustversion",
- "sync_wrapper 1.0.1",
- "tower-layer",
- "tower-service",
-]
-
 [[package]]
 name = "axum-core"
 version = "0.5.0"
@@ -797,8 +761,8 @@ version = "0.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "460fc6f625a1f7705c6cf62d0d070794e94668988b1c38111baeec177c715f7b"
 dependencies = [
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
  "bytes",
  "form_urlencoded",
  "futures-util",
@@ -1009,24 +973,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "bindgen"
-version = "0.70.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f49d8fed880d473ea71efb9bf597651e77201bdd4893efe54c9e5d65ae04ce6f"
-dependencies = [
- "bitflags 2.8.0",
- "cexpr",
- "clang-sys",
- "itertools 0.12.1",
- "proc-macro2",
- "quote",
- "regex",
- "rustc-hash 1.1.0",
- "shlex",
- "syn 2.0.100",
-]
-
 [[package]]
 name = "bindgen"
 version = "0.71.1"
@@ -1386,7 +1332,7 @@ dependencies = [
  "aws-sdk-kms",
  "aws-sdk-s3",
  "aws-smithy-types",
- "axum 0.8.1",
+ "axum",
  "axum-extra",
  "base64 0.22.1",
  "bytes",
@@ -1401,10 +1347,8 @@ dependencies = [
  "hostname-validator",
  "http 1.1.0",
  "indexmap 2.9.0",
- "inferno 0.12.0",
  "itertools 0.10.5",
  "jsonwebtoken",
- "libproc",
  "metrics",
  "nix 0.30.1",
  "notify",
@@ -1418,8 +1362,6 @@ dependencies = [
  "postgres-types",
  "postgres_initdb",
  "postgres_versioninfo",
- "pprof 0.15.0",
- "prost 0.12.6",
  "regex",
  "remote_storage",
  "reqwest",
@@ -1431,7 +1373,6 @@ dependencies = [
  "serde_with",
  "signal-hook",
  "tar",
- "tempfile",
  "thiserror 1.0.69",
  "tokio",
  "tokio-postgres",
@@ -1942,6 +1883,7 @@ dependencies = [
  "diesel_derives",
  "itoa",
  "serde_json",
+ "uuid",
 ]
 
 [[package]]
@@ -2152,7 +2094,7 @@ name = "endpoint_storage"
 version = "0.0.1"
 dependencies = [
  "anyhow",
- "axum 0.8.1",
+ "axum",
  "axum-extra",
  "camino",
  "camino-tempfile",
@@ -2268,12 +2210,12 @@ checksum = "5443807d6dff69373d433ab9ef5378ad8df50ca6298caf15de6e52e24aaf54d5"
 
 [[package]]
 name = "errno"
-version = "0.3.13"
+version = "0.3.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "778e2ac28f6c47af28e4907f13ffd1e1ddbd400980a9abd7c8df189bf578a5ad"
+checksum = "a258e46cdc063eb8519c00b9fc845fc47bcfca4130e2f08e88665ceda8474245"
 dependencies = [
  "libc",
- "windows-sys 0.59.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -2872,15 +2814,6 @@ dependencies = [
  "digest",
 ]
 
-[[package]]
-name = "home"
-version = "0.5.11"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "589533453244b0995c858700322199b2becb13b627df2851f64a2775d024abcf"
-dependencies = [
- "windows-sys 0.59.0",
-]
-
 [[package]]
 name = "hostname"
 version = "0.4.0"
@@ -2990,7 +2923,7 @@ dependencies = [
  "jsonwebtoken",
  "metrics",
  "once_cell",
- "pprof 0.14.0",
+ "pprof",
  "regex",
  "routerify",
  "rustls 0.23.27",
@@ -3652,7 +3585,7 @@ version = "1.5.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bbd2bcb4c963f2ddae06a2efc7e9f3591312473c50c6685e1f298068316e66fe"
 dependencies = [
- "spin 0.9.8",
+ "spin",
 ]
 
 [[package]]
@@ -3677,17 +3610,6 @@ version = "0.2.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4ec2a862134d2a7d32d7983ddcdd1c4923530833c9f2ea1a44fc5fa473989058"
 
-[[package]]
-name = "libproc"
-version = "0.14.10"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e78a09b56be5adbcad5aa1197371688dc6bb249a26da3bca2011ee2fb987ebfb"
-dependencies = [
- "bindgen 0.70.1",
- "errno",
- "libc",
-]
-
 [[package]]
 name = "linux-raw-sys"
 version = "0.4.14"
@@ -3700,12 +3622,6 @@ version = "0.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f0b5399f6804fbab912acbd8878ed3532d506b7c951b8f9f164ef90fef39e3f4"
 
-[[package]]
-name = "linux-raw-sys"
-version = "0.9.4"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd945864f07fe9f5371a27ad7b52a172b4b499999f1d97574c9fa68373937e12"
-
 [[package]]
 name = "litemap"
 version = "0.7.4"
@@ -3714,9 +3630,9 @@ checksum = "4ee93343901ab17bd981295f2cf0026d4ad018c7c31ba84549a4ddbb47a45104"
 
 [[package]]
 name = "lock_api"
-version = "0.4.10"
+version = "0.4.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c1cc9717a20b1bb222f333e6a92fd32f7d8a18ddc5a3191a11af45dcbf4dcd16"
+checksum = "96936507f153605bddfcda068dd804796c84324ed2510809e5b2a624c81da765"
 dependencies = [
  "autocfg",
  "scopeguard",
@@ -3759,12 +3675,6 @@ dependencies = [
  "regex-automata 0.1.10",
 ]
 
-[[package]]
-name = "matchit"
-version = "0.7.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0e7465ac9959cc2b1404e8e2367b43684a6d13790fe23056cc8c6c5a6b7bcb94"
-
 [[package]]
 name = "matchit"
 version = "0.8.4"
@@ -3872,7 +3782,7 @@ dependencies = [
  "procfs",
  "prometheus",
  "rand 0.8.5",
- "rand_distr",
+ "rand_distr 0.4.3",
  "twox-hash",
 ]
 
@@ -3960,7 +3870,12 @@ checksum = "e5ce46fe64a9d73be07dcbe690a38ce1b293be448fd8ce1e6c1b8062c9f72c6a"
 name = "neon-shmem"
 version = "0.1.0"
 dependencies = [
+ "libc",
+ "lock_api",
  "nix 0.30.1",
+ "rand 0.9.1",
+ "rand_distr 0.5.1",
+ "rustc-hash 2.1.1",
  "tempfile",
  "thiserror 1.0.69",
  "workspace_hack",
@@ -4410,6 +4325,7 @@ dependencies = [
  "pageserver_client",
  "pageserver_client_grpc",
  "pageserver_page_api",
+ "pprof",
  "rand 0.8.5",
  "reqwest",
  "serde",
@@ -4504,7 +4420,7 @@ dependencies = [
  "postgres_ffi_types",
  "postgres_initdb",
  "posthog_client_lite",
- "pprof 0.14.0",
+ "pprof",
  "pq_proto",
  "procfs",
  "rand 0.8.5",
@@ -5079,7 +4995,7 @@ name = "postgres_ffi"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "bindgen 0.71.1",
+ "bindgen",
  "bytes",
  "crc32c",
  "criterion",
@@ -5090,7 +5006,7 @@ dependencies = [
  "postgres",
  "postgres_ffi_types",
  "postgres_versioninfo",
- "pprof 0.14.0",
+ "pprof",
  "regex",
  "serde",
  "thiserror 1.0.69",
@@ -5180,30 +5096,6 @@ dependencies = [
  "thiserror 1.0.69",
 ]
 
-[[package]]
-name = "pprof"
-version = "0.15.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "38a01da47675efa7673b032bf8efd8214f1917d89685e07e395ab125ea42b187"
-dependencies = [
- "aligned-vec",
- "backtrace",
- "cfg-if",
- "findshlibs",
- "inferno 0.11.21",
- "libc",
- "log",
- "nix 0.26.4",
- "once_cell",
- "protobuf",
- "protobuf-codegen",
- "smallvec",
- "spin 0.10.0",
- "symbolic-demangle",
- "tempfile",
- "thiserror 2.0.11",
-]
-
 [[package]]
 name = "pprof_util"
 version = "0.7.0"
@@ -5279,7 +5171,7 @@ dependencies = [
  "hex",
  "lazy_static",
  "procfs-core",
- "rustix 0.38.41",
+ "rustix",
 ]
 
 [[package]]
@@ -5415,62 +5307,12 @@ dependencies = [
  "prost 0.13.5",
 ]
 
-[[package]]
-name = "protobuf"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d65a1d4ddae7d8b5de68153b48f6aa3bba8cb002b243dbdbc55a5afbc98f99f4"
-dependencies = [
- "once_cell",
- "protobuf-support",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "protobuf-codegen"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5d3976825c0014bbd2f3b34f0001876604fe87e0c86cd8fa54251530f1544ace"
-dependencies = [
- "anyhow",
- "once_cell",
- "protobuf",
- "protobuf-parse",
- "regex",
- "tempfile",
- "thiserror 1.0.69",
-]
-
-[[package]]
-name = "protobuf-parse"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b4aeaa1f2460f1d348eeaeed86aea999ce98c1bded6f089ff8514c9d9dbdc973"
-dependencies = [
- "anyhow",
- "indexmap 2.9.0",
- "log",
- "protobuf",
- "protobuf-support",
- "tempfile",
- "thiserror 1.0.69",
- "which",
-]
-
-[[package]]
-name = "protobuf-support"
-version = "3.7.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e36c2f31e0a47f9280fb347ef5e461ffcd2c52dd520d8e216b52f93b0b0d7d6"
-dependencies = [
- "thiserror 1.0.69",
-]
-
 [[package]]
 name = "proxy"
 version = "0.1.0"
 dependencies = [
  "ahash",
+ "alloc-metrics",
  "anyhow",
  "arc-swap",
  "assert-json-diff",
@@ -5478,6 +5320,7 @@ dependencies = [
  "async-trait",
  "atomic-take",
  "aws-config",
+ "aws-credential-types",
  "aws-sdk-iam",
  "aws-sigv4",
  "base64 0.22.1",
@@ -5517,6 +5360,7 @@ dependencies = [
  "itoa",
  "jose-jwa",
  "jose-jwk",
+ "json",
  "lasso",
  "measured",
  "metrics",
@@ -5533,7 +5377,7 @@ dependencies = [
  "postgres_backend",
  "pq_proto",
  "rand 0.8.5",
- "rand_distr",
+ "rand_distr 0.4.3",
  "rcgen",
  "redis",
  "regex",
@@ -5544,7 +5388,7 @@ dependencies = [
  "reqwest-tracing",
  "rsa",
  "rstest",
- "rustc-hash 1.1.0",
+ "rustc-hash 2.1.1",
  "rustls 0.23.27",
  "rustls-native-certs 0.8.0",
  "rustls-pemfile 2.1.1",
@@ -5667,6 +5511,16 @@ dependencies = [
  "rand_core 0.6.4",
 ]
 
+[[package]]
+name = "rand"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9fbfd9d094a40bf3ae768db9361049ace4c0e04a4fd6b359518bd7b73a73dd97"
+dependencies = [
+ "rand_chacha 0.9.0",
+ "rand_core 0.9.3",
+]
+
 [[package]]
 name = "rand_chacha"
 version = "0.2.2"
@@ -5687,6 +5541,16 @@ dependencies = [
  "rand_core 0.6.4",
 ]
 
+[[package]]
+name = "rand_chacha"
+version = "0.9.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
+dependencies = [
+ "ppv-lite86",
+ "rand_core 0.9.3",
+]
+
 [[package]]
 name = "rand_core"
 version = "0.5.1"
@@ -5705,6 +5569,15 @@ dependencies = [
  "getrandom 0.2.11",
 ]
 
+[[package]]
+name = "rand_core"
+version = "0.9.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "99d9a13982dcf210057a8a78572b2217b667c3beacbf3a0d8b454f6f82837d38"
+dependencies = [
+ "getrandom 0.3.3",
+]
+
 [[package]]
 name = "rand_distr"
 version = "0.4.3"
@@ -5715,6 +5588,16 @@ dependencies = [
  "rand 0.8.5",
 ]
 
+[[package]]
+name = "rand_distr"
+version = "0.5.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6a8615d50dcf34fa31f7ab52692afec947c4dd0ab803cc87cb3b0b4570ff7463"
+dependencies = [
+ "num-traits",
+ "rand 0.9.1",
+]
+
 [[package]]
 name = "rand_hc"
 version = "0.2.0"
@@ -6031,7 +5914,7 @@ dependencies = [
  "async-trait",
  "getrandom 0.2.11",
  "http 1.1.0",
- "matchit 0.8.4",
+ "matchit",
  "opentelemetry",
  "reqwest",
  "reqwest-middleware",
@@ -6222,19 +6105,6 @@ dependencies = [
  "windows-sys 0.52.0",
 ]
 
-[[package]]
-name = "rustix"
-version = "1.0.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c71e83d6afe7ff64890ec6b71d6a69bb8a610ab78ce364b3352876bb4c801266"
-dependencies = [
- "bitflags 2.8.0",
- "errno",
- "libc",
- "linux-raw-sys 0.9.4",
- "windows-sys 0.59.0",
-]
-
 [[package]]
 name = "rustls"
 version = "0.21.12"
@@ -6409,6 +6279,7 @@ dependencies = [
  "itertools 0.10.5",
  "jsonwebtoken",
  "metrics",
+ "nix 0.30.1",
  "once_cell",
  "pageserver_api",
  "parking_lot 0.12.1",
@@ -6416,8 +6287,9 @@ dependencies = [
  "postgres-protocol",
  "postgres_backend",
  "postgres_ffi",
+ "postgres_ffi_types",
  "postgres_versioninfo",
- "pprof 0.14.0",
+ "pprof",
  "pq_proto",
  "rand 0.8.5",
  "regex",
@@ -6460,7 +6332,7 @@ dependencies = [
  "anyhow",
  "const_format",
  "pageserver_api",
- "postgres_ffi",
+ "postgres_ffi_types",
  "postgres_versioninfo",
  "pq_proto",
  "serde",
@@ -7005,18 +6877,6 @@ name = "spin"
 version = "0.9.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6980e8d7511241f8acf4aebddbb1ff938df5eebe98691418c4468d0b72a96a67"
-dependencies = [
- "lock_api",
-]
-
-[[package]]
-name = "spin"
-version = "0.10.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d5fe4ccb98d9c292d56fec89a5e07da7fc4cf0dc11e156b41793132775d3e591"
-dependencies = [
- "lock_api",
-]
 
 [[package]]
 name = "spinning_top"
@@ -7148,6 +7008,7 @@ dependencies = [
  "tokio-util",
  "tracing",
  "utils",
+ "uuid",
  "workspace_hack",
 ]
 
@@ -7211,6 +7072,7 @@ dependencies = [
  "pageserver_api",
  "pageserver_client",
  "reqwest",
+ "safekeeper_api",
  "serde_json",
  "storage_controller_client",
  "tokio",
@@ -7377,14 +7239,14 @@ dependencies = [
 
 [[package]]
 name = "tempfile"
-version = "3.20.0"
+version = "3.14.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e8a64e3985349f2441a1a9ef0b853f869006c3855f2cda6862a94d26ebb9d6a1"
+checksum = "28cce251fcbc87fac86a866eeb0d6c2d536fc16d06f184bb61aeae11aa4cee0c"
 dependencies = [
+ "cfg-if",
  "fastrand 2.2.0",
- "getrandom 0.3.3",
  "once_cell",
- "rustix 1.0.7",
+ "rustix",
  "windows-sys 0.59.0",
 ]
 
@@ -7482,12 +7344,10 @@ dependencies = [
 
 [[package]]
 name = "thread_local"
-version = "1.1.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3fdd6f064ccff2d6567adcb3873ca630700f00b5ad3f060c25b5dcfd9a4ce152"
+version = "1.1.9"
+source = "git+https://github.com/conradludgate/thread_local-rs?branch=no-tls-destructor-get#f9ca3d375745c14a632ae3ffe6a7a646dc8421a0"
 dependencies = [
  "cfg-if",
- "once_cell",
 ]
 
 [[package]]
@@ -7780,6 +7640,7 @@ dependencies = [
  "futures-core",
  "pin-project-lite",
  "tokio",
+ "tokio-util",
 ]
 
 [[package]]
@@ -7878,25 +7739,16 @@ version = "0.12.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "877c5b330756d856ffcc4553ab34a5684481ade925ecc54bcd1bf02b1d0d4d52"
 dependencies = [
- "async-stream",
  "async-trait",
- "axum 0.7.9",
  "base64 0.22.1",
  "bytes",
- "h2 0.4.4",
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.4.1",
- "hyper-timeout",
- "hyper-util",
  "percent-encoding",
  "pin-project",
  "prost 0.13.5",
- "socket2",
- "tokio",
  "tokio-stream",
- "tower 0.4.13",
  "tower-layer",
  "tower-service",
  "tracing",
@@ -7909,7 +7761,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7e581ba15a835f4d9ea06c55ab1bd4dce26fc53752c69a04aac00703bfb49ba9"
 dependencies = [
  "async-trait",
- "axum 0.8.1",
+ "axum",
  "base64 0.22.1",
  "bytes",
  "flate2",
@@ -7970,16 +7822,11 @@ checksum = "b8fa9be0de6cf49e536ce1851f987bd21a43b771b09473c3549a6c853db37c1c"
 dependencies = [
  "futures-core",
  "futures-util",
- "indexmap 1.9.3",
  "pin-project",
  "pin-project-lite",
- "rand 0.8.5",
- "slab",
  "tokio",
- "tokio-util",
  "tower-layer",
  "tower-service",
- "tracing",
 ]
 
 [[package]]
@@ -8411,7 +8258,7 @@ dependencies = [
  "pem",
  "pin-project-lite",
  "postgres_connection",
- "pprof 0.14.0",
+ "pprof",
  "pq_proto",
  "rand 0.8.5",
  "regex",
@@ -8433,6 +8280,7 @@ dependencies = [
  "tracing-error",
  "tracing-subscriber",
  "tracing-utils",
+ "uuid",
  "walkdir",
 ]
 
@@ -8463,7 +8311,7 @@ name = "vm_monitor"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "axum 0.8.1",
+ "axum",
  "cgroups-rs",
  "clap",
  "futures",
@@ -8518,7 +8366,7 @@ dependencies = [
  "pageserver_api",
  "postgres_ffi",
  "postgres_ffi_types",
- "pprof 0.14.0",
+ "pprof",
  "prost 0.13.5",
  "remote_storage",
  "serde",
@@ -8548,7 +8396,7 @@ name = "walproposer"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "bindgen 0.71.1",
+ "bindgen",
  "postgres_ffi",
  "utils",
 ]
@@ -8713,18 +8561,6 @@ dependencies = [
  "rustls-pki-types",
 ]
 
-[[package]]
-name = "which"
-version = "4.4.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "87ba24419a2078cd2b0f2ede2691b6c66d8e47836da3b6db8265ebad47afbfc7"
-dependencies = [
- "either",
- "home",
- "once_cell",
- "rustix 0.38.41",
-]
-
 [[package]]
 name = "whoami"
 version = "1.5.1"
@@ -8969,19 +8805,17 @@ dependencies = [
  "ahash",
  "anstream",
  "anyhow",
- "axum 0.8.1",
- "axum-core 0.5.0",
+ "axum",
+ "axum-core",
  "base64 0.21.7",
  "base64ct",
  "bytes",
  "camino",
  "cc",
  "chrono",
- "clang-sys",
  "clap",
  "clap_builder",
  "const-oid",
- "criterion",
  "crypto-bigint 0.5.5",
  "der 0.7.8",
  "deranged",
@@ -9008,7 +8842,6 @@ dependencies = [
  "hyper 0.14.30",
  "hyper 1.4.1",
  "hyper-util",
- "indexmap 1.9.3",
  "indexmap 2.9.0",
  "itertools 0.12.1",
  "lazy_static",
@@ -9025,6 +8858,7 @@ dependencies = [
  "num-iter",
  "num-rational",
  "num-traits",
+ "once_cell",
  "p256 0.13.2",
  "parquet",
  "prettyplease",
@@ -9046,7 +8880,6 @@ dependencies = [
  "sha2",
  "signature 2.2.0",
  "smallvec",
- "spin 0.9.8",
  "spki 0.7.3",
  "stable_deref_trait",
  "subtle",
@@ -9061,14 +8894,12 @@ dependencies = [
  "tokio-stream",
  "tokio-util",
  "toml_edit",
- "tonic 0.12.3",
  "tower 0.5.2",
  "tracing",
  "tracing-core",
  "tracing-log",
  "tracing-subscriber",
  "url",
- "uuid",
  "zeroize",
  "zstd",
  "zstd-safe",

Cargo.toml

@@ -130,7 +130,7 @@ jemalloc_pprof = { version = "0.7", features = ["symbolize", "flamegraph"] }
 jsonwebtoken = "9"
 lasso = "0.7"
 libc = "0.2"
-libproc = "0.14"
+lock_api = "0.4.13"
 md5 = "0.7.0"
 measured = { version = "0.0.22", features=["lasso"] }
 measured-process = { version = "0.0.22" }
@@ -166,7 +166,7 @@ reqwest-middleware = "0.4"
 reqwest-retry = "0.7"
 routerify = "3"
 rpds = "0.13"
-rustc-hash = "1.1.0"
+rustc-hash = "2.1.1"
 rustls = { version = "0.23.16", default-features = false }
 rustls-pemfile = "2"
 rustls-pki-types = "1.11"
@@ -195,6 +195,7 @@ sync_wrapper = "0.1.2"
 tar = "0.4"
 test-context = "0.3"
 thiserror = "1.0"
+thread_local = "1.1.9"
 tikv-jemallocator = { version = "0.6", features = ["profiling", "stats", "unprefixed_malloc_on_supported_platforms"] }
 tikv-jemalloc-ctl = { version = "0.6", features = ["stats"] }
 tokio = { version = "1.43.1", features = ["macros"] }
@@ -202,7 +203,7 @@ tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.g
 tokio-io-timeout = "1.2.0"
 tokio-postgres-rustls = "0.12.0"
 tokio-rustls = { version = "0.26.0", default-features = false, features = ["tls12", "ring"]}
-tokio-stream = "0.1"
+tokio-stream = { version = "0.1", features = ["sync"] }
 tokio-tar = "0.3"
 tokio-util = { version = "0.7.10", features = ["io", "io-util", "rt"] }
 toml = "0.8"
@@ -253,6 +254,7 @@ azure_storage = { git = "https://github.com/neondatabase/azure-sdk-for-rust.git"
 azure_storage_blobs = { git = "https://github.com/neondatabase/azure-sdk-for-rust.git", branch = "neon", default-features = false, features = ["enable_reqwest_rustls"] }
 
 ## Local libraries
+alloc-metrics = { version = "0.1", path = "./libs/alloc-metrics/" }
 compute_api = { version = "0.1", path = "./libs/compute_api/" }
 consumption_metrics = { version = "0.1", path = "./libs/consumption_metrics/" }
 desim = { version = "0.1", path = "./libs/desim" }
@@ -279,7 +281,6 @@ safekeeper_api = { version = "0.1", path = "./libs/safekeeper_api" }
 safekeeper_client = { path = "./safekeeper/client" }
 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.
 storage_controller_client = { path = "./storage_controller/client" }
-tempfile = "3"
 tenant_size_model = { version = "0.1", path = "./libs/tenant_size_model/" }
 tracing-utils = { version = "0.1", path = "./libs/tracing-utils/" }
 utils = { version = "0.1", path = "./libs/utils/" }
@@ -303,6 +304,9 @@ tonic-build = "0.13.1"
 # Needed to get `tokio-postgres-rustls` to depend on our fork.
 tokio-postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 
+# Needed to fix a bug in alloc-metrics
+thread_local = { git = "https://github.com/conradludgate/thread_local-rs", branch = "no-tls-destructor-get" }
+
 ################# Binary contents sections
 
 [profile.release]

Dockerfile

@@ -109,8 +109,6 @@ RUN set -e \
         libreadline-dev \
         libseccomp-dev \
         ca-certificates \
-        bpfcc-tools \
-        sudo \
         openssl \
         unzip \
         curl \

Makefile

@@ -2,7 +2,7 @@ ROOT_PROJECT_DIR := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
 
 # Where to install Postgres, default is ./pg_install, maybe useful for package
 # managers.
-POSTGRES_INSTALL_DIR ?= $(ROOT_PROJECT_DIR)/pg_install/
+POSTGRES_INSTALL_DIR ?= $(ROOT_PROJECT_DIR)/pg_install
 
 # Supported PostgreSQL versions
 POSTGRES_VERSIONS = v17 v16 v15 v14
@@ -14,7 +14,7 @@ POSTGRES_VERSIONS = v17 v16 v15 v14
 # it is derived from BUILD_TYPE.
 
 # All intermediate build artifacts are stored here.
-BUILD_DIR := build
+BUILD_DIR := $(ROOT_PROJECT_DIR)/build
 
 ICU_PREFIX_DIR := /usr/local/icu
 
@@ -212,7 +212,7 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 		FIND_TYPEDEF=$(ROOT_PROJECT_DIR)/vendor/postgres-v17/src/tools/find_typedef \
 		INDENT=$(BUILD_DIR)/v17/src/tools/pg_bsd_indent/pg_bsd_indent \
 		PGINDENT_SCRIPT=$(ROOT_PROJECT_DIR)/vendor/postgres-v17/src/tools/pgindent/pgindent \
-		-C $(BUILD_DIR)/neon-v17 \
+		-C $(BUILD_DIR)/pgxn-v17/neon \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile pgindent
 
 
@@ -220,11 +220,15 @@ neon-pgindent: postgres-v17-pg-bsd-indent neon-pg-ext-v17
 setup-pre-commit-hook:
 	ln -s -f $(ROOT_PROJECT_DIR)/pre-commit.py .git/hooks/pre-commit
 
+build-tools/node_modules: build-tools/package.json
+	cd build-tools && $(if $(CI),npm ci,npm install)
+	touch build-tools/node_modules
+
 .PHONY: lint-openapi-spec
-lint-openapi-spec:
+lint-openapi-spec: build-tools/node_modules
 	# operation-2xx-response: pageserver timeline delete returns 404 on success
 	find . -iname "openapi_spec.y*ml" -exec\
-		docker run --rm -v ${PWD}:/spec ghcr.io/redocly/cli:1.34.4\
+		npx --prefix=build-tools/ redocly\
 			--skip-rule=operation-operationId --skip-rule=operation-summary --extends=minimal\
 			--skip-rule=no-server-example.com --skip-rule=operation-2xx-response\
 			lint {} \+

build-tools/Dockerfile

@@ -35,7 +35,7 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
     echo -e "retry_connrefused=on\ntimeout=15\ntries=5\nretry-on-host-error=on\n" > /root/.wgetrc && \
     echo -e "--retry-connrefused\n--connect-timeout 15\n--retry 5\n--max-time 300\n" > /root/.curlrc
 
-COPY build_tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
+COPY build-tools/patches/pgcopydbv017.patch /pgcopydbv017.patch
 
 RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
         set -e && \
@@ -61,9 +61,6 @@ RUN if [ "${DEBIAN_VERSION}" = "bookworm" ]; then \
         libpq5 \
         libpq-dev \
         libzstd-dev \
-        linux-perf \
-        bpfcc-tools \
-        linux-headers-$(case "$(uname -m)" in x86_64) echo amd64;; aarch64) echo arm64;; esac) \
         postgresql-16 \
         postgresql-server-dev-16 \
         postgresql-common  \
@@ -108,21 +105,15 @@ RUN echo 'Acquire::Retries "5";' > /etc/apt/apt.conf.d/80-retries && \
 #
 # 'gdb' is included so that we get backtraces of core dumps produced in
 # regression tests
-RUN set -ex \
-    && KERNEL_VERSION="$(uname -r | cut -d'-' -f1 | sed 's/\.0$//')" \
-    && echo KERNEL_VERSION=${KERNEL_VERSION} >> /etc/environment \
-    && KERNEL_ARCH=$(uname -m | awk '{ if ($1 ~ /^(x86_64|i[3-6]86)$/) print "x86"; else if ($1 ~ /^(aarch64|arm.*)$/) print "aarch"; else print $1 }') \
-    && echo KERNEL_ARCH=${KERNEL_ARCH} >> /etc/environment \
+RUN set -e \
     && apt update \
     && apt install -y \
         autoconf \
         automake \
-        bc \
         bison \
         build-essential \
         ca-certificates \
         cmake \
-        cpio \
         curl \
         flex \
         gdb \
@@ -131,10 +122,8 @@ RUN set -ex \
         gzip \
         jq \
         jsonnet \
-        kmod \
         libcurl4-openssl-dev \
         libbz2-dev \
-        libelf-dev \
         libffi-dev \
         liblzma-dev \
         libncurses5-dev \
@@ -148,11 +137,6 @@ RUN set -ex \
         libxml2-dev \
         libxmlsec1-dev \
         libxxhash-dev \
-        linux-perf \
-        bpfcc-tools \
-        libbpfcc \
-        libbpfcc-dev \
-        linux-headers-$(case "$(uname -m)" in x86_64) echo amd64;; aarch64) echo arm64;; esac) \
         lsof \
         make \
         netcat-openbsd \
@@ -160,8 +144,6 @@ RUN set -ex \
         openssh-client \
         parallel \
         pkg-config \
-        rsync \
-        sudo \
         unzip \
         wget \
         xz-utils \
@@ -206,6 +188,12 @@ RUN curl -fsSL 'https://apt.llvm.org/llvm-snapshot.gpg.key' | apt-key add - \
     && bash -c 'for f in /usr/bin/clang*-${LLVM_VERSION} /usr/bin/llvm*-${LLVM_VERSION}; do ln -s "${f}" "${f%-${LLVM_VERSION}}"; done' \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
 
+# Install node
+ENV NODE_VERSION=24
+RUN curl -fsSL https://deb.nodesource.com/setup_${NODE_VERSION}.x | bash - \
+    && apt install -y nodejs \
+    && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
+
 # Install docker
 RUN curl -fsSL https://download.docker.com/linux/ubuntu/gpg | gpg --dearmor -o /usr/share/keyrings/docker-archive-keyring.gpg \
     && echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/docker-archive-keyring.gpg] https://download.docker.com/linux/debian ${DEBIAN_VERSION} stable" > /etc/apt/sources.list.d/docker.list \
@@ -216,8 +204,6 @@ RUN curl -fsSL https://download.docker.com/linux/ubuntu/gpg | gpg --dearmor -o /
 # Configure sudo & docker
 RUN usermod -aG sudo nonroot && \
     echo '%sudo ALL=(ALL) NOPASSWD:ALL' >> /etc/sudoers && \
-    mkdir -p /etc/sudoers.d && \
-    echo 'nonroot ALL=(ALL) NOPASSWD:ALL' > /etc/sudoers.d/nonroot && \
     usermod -aG docker nonroot
 
 # AWS CLI
@@ -331,14 +317,14 @@ RUN curl -sSO https://static.rust-lang.org/rustup/dist/$(uname -m)-unknown-linux
     . "$HOME/.cargo/env" && \
     cargo --version && rustup --version && \
     rustup component add llvm-tools rustfmt clippy && \
-    cargo install rustfilt            --version ${RUSTFILT_VERSION} --locked && \
-    cargo install cargo-hakari        --version ${CARGO_HAKARI_VERSION} --locked && \
-    cargo install cargo-deny          --version ${CARGO_DENY_VERSION} --locked && \
-    cargo install cargo-hack          --version ${CARGO_HACK_VERSION} --locked && \
-    cargo install cargo-nextest       --version ${CARGO_NEXTEST_VERSION} --locked && \
-    cargo install cargo-chef          --version ${CARGO_CHEF_VERSION} --locked && \
-    cargo install diesel_cli          --version ${CARGO_DIESEL_CLI_VERSION} --locked \
-                                      --features postgres-bundled --no-default-features && \
+    cargo install rustfilt      --locked --version ${RUSTFILT_VERSION} && \
+    cargo install cargo-hakari  --locked --version ${CARGO_HAKARI_VERSION} && \
+    cargo install cargo-deny    --locked --version ${CARGO_DENY_VERSION} && \
+    cargo install cargo-hack    --locked --version ${CARGO_HACK_VERSION} && \
+    cargo install cargo-nextest --locked --version ${CARGO_NEXTEST_VERSION} && \
+    cargo install cargo-chef    --locked --version ${CARGO_CHEF_VERSION} && \
+    cargo install diesel_cli    --locked --version ${CARGO_DIESEL_CLI_VERSION} \
+                                --features postgres-bundled --no-default-features && \
     rm -rf /home/nonroot/.cargo/registry && \
     rm -rf /home/nonroot/.cargo/git
 

build-tools/package.json

@@ -0,0 +1,8 @@
+{
+  "name": "build-tools",
+  "private": true,
+  "devDependencies": {
+    "@redocly/cli": "1.34.4",
+    "@sourcemeta/jsonschema": "10.0.0"
+  }
+}

compute/Makefile

@@ -50,9 +50,9 @@ jsonnetfmt-format:
 	jsonnetfmt --in-place $(jsonnet_files)
 
 .PHONY: manifest-schema-validation
-manifest-schema-validation: node_modules
-	node_modules/.bin/jsonschema validate -d https://json-schema.org/draft/2020-12/schema manifest.schema.json manifest.yaml
+manifest-schema-validation: ../build-tools/node_modules
+	npx --prefix=../build-tools/ jsonschema validate -d https://json-schema.org/draft/2020-12/schema manifest.schema.json manifest.yaml
 
-node_modules: package.json
-	npm install
-	touch node_modules
+../build-tools/node_modules: ../build-tools/package.json
+	cd ../build-tools && $(if $(CI),npm ci,npm install)
+	touch ../build-tools/node_modules

compute/compute-node.Dockerfile

@@ -9,7 +9,7 @@
 #
 # build-tools:   This contains Rust compiler toolchain and other tools needed at compile
 #                time. This is also used for the storage builds. This image is defined in
-#                build-tools.Dockerfile.
+#                build-tools/Dockerfile.
 #
 # build-deps:    Contains C compiler, other build tools, and compile-time dependencies
 #                needed to compile PostgreSQL and most extensions. (Some extensions need
@@ -115,7 +115,7 @@ ARG EXTENSIONS=all
 FROM $BASE_IMAGE_SHA AS build-deps
 ARG DEBIAN_VERSION
 
-# Keep in sync with build-tools.Dockerfile
+# Keep in sync with build-tools/Dockerfile
 ENV PROTOC_VERSION=25.1
 
 # Use strict mode for bash to catch errors early
@@ -149,9 +149,6 @@ RUN case $DEBIAN_VERSION in \
     ninja-build git autoconf automake libtool build-essential bison flex libreadline-dev \
     zlib1g-dev libxml2-dev libcurl4-openssl-dev libossp-uuid-dev wget ca-certificates pkg-config libssl-dev \
     libicu-dev libxslt1-dev liblz4-dev libzstd-dev zstd curl unzip g++ \
-    bpfcc-tools \
-    libbpfcc \
-    libbpfcc-dev \
     libclang-dev \
     jsonnet \
     $VERSION_INSTALLS \
@@ -173,7 +170,29 @@ RUN case $DEBIAN_VERSION in \
 FROM build-deps AS pg-build
 ARG PG_VERSION
 COPY vendor/postgres-${PG_VERSION:?} postgres
+COPY compute/patches/postgres_fdw.patch .
+COPY compute/patches/pg_stat_statements_pg14-16.patch .
+COPY compute/patches/pg_stat_statements_pg17.patch .
 RUN cd postgres && \
+    # Apply patches to some contrib extensions
+    # For example, we need to grant EXECUTE on pg_stat_statements_reset() to {privileged_role_name}.
+    # In vanilla Postgres this function is limited to Postgres role superuser.
+    # In Neon we have {privileged_role_name} role that is not a superuser but replaces superuser in some cases.
+    # We could add the additional grant statements to the Postgres repository but it would be hard to maintain,
+    # whenever we need to pick up a new Postgres version and we want to limit the changes in our Postgres fork,
+    # so we do it here.
+    case "${PG_VERSION}" in \
+    "v14" | "v15" | "v16") \
+    patch -p1 < /pg_stat_statements_pg14-16.patch; \
+    ;; \
+    "v17") \
+    patch -p1 < /pg_stat_statements_pg17.patch; \
+    ;; \
+    *) \
+    # To do not forget to migrate patches to the next major version
+    echo "No contrib patches for this PostgreSQL version" && exit 1;; \
+    esac && \
+    patch -p1 < /postgres_fdw.patch && \
     export CONFIGURE_CMD="./configure CFLAGS='-O2 -g3 -fsigned-char' --enable-debug --with-openssl --with-uuid=ossp \
     --with-icu --with-libxml --with-libxslt --with-lz4" && \
     if [ "${PG_VERSION:?}" != "v14" ]; then \
@@ -187,8 +206,6 @@ RUN cd postgres && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/autoinc.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/dblink.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/postgres_fdw.control && \
-    file=/usr/local/pgsql/share/extension/postgres_fdw--1.0.sql && [ -e $file ] && \
-    echo 'GRANT USAGE ON FOREIGN DATA WRAPPER postgres_fdw TO neon_superuser;' >> $file && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/bloom.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/earthdistance.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/insert_username.control && \
@@ -198,34 +215,7 @@ RUN cd postgres && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgrowlocks.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/pgstattuple.control && \
     echo 'trusted = true' >> /usr/local/pgsql/share/extension/refint.control && \
-    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control && \
-    # We need to grant EXECUTE on pg_stat_statements_reset() to neon_superuser.
-    # In vanilla postgres this function is limited to Postgres role superuser.
-    # In neon we have neon_superuser role that is not a superuser but replaces superuser in some cases.
-    # We could add the additional grant statements to the postgres repository but it would be hard to maintain,
-    # whenever we need to pick up a new postgres version and we want to limit the changes in our postgres fork,
-    # so we do it here.
-    for file in /usr/local/pgsql/share/extension/pg_stat_statements--*.sql; do \
-        filename=$(basename "$file"); \
-        # Note that there are no downgrade scripts for pg_stat_statements, so we \
-        # don't have to modify any downgrade paths or (much) older versions: we only \
-        # have to make sure every creation of the pg_stat_statements_reset function \
-        # also adds execute permissions to the neon_superuser.
-        case $filename in \
-          pg_stat_statements--1.4.sql) \
-            # pg_stat_statements_reset is first created with 1.4
-            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO neon_superuser;' >> $file; \
-            ;; \
-          pg_stat_statements--1.6--1.7.sql) \
-            # Then with the 1.6-1.7 migration it is re-created with a new signature, thus add the permissions back
-            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO neon_superuser;' >> $file; \
-            ;; \
-          pg_stat_statements--1.10--1.11.sql) \
-            # Then with the 1.10-1.11 migration it is re-created with a new signature again, thus add the permissions back
-            echo 'GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint, boolean) TO neon_superuser;' >> $file; \
-            ;; \
-        esac; \
-    done;
+    echo 'trusted = true' >> /usr/local/pgsql/share/extension/xml2.control
 
 # Set PATH for all the subsequent build steps
 ENV PATH="/usr/local/pgsql/bin:$PATH"
@@ -1527,7 +1517,7 @@ WORKDIR /ext-src
 COPY compute/patches/pg_duckdb_v031.patch .
 COPY compute/patches/duckdb_v120.patch .
 # pg_duckdb build requires source dir to be a git repo to get submodules
-# allow neon_superuser to execute some functions that in pg_duckdb are available to superuser only:
+# allow {privileged_role_name} to execute some functions that in pg_duckdb are available to superuser only:
 # - extension management function duckdb.install_extension()
 # - access to duckdb.extensions table and its sequence
 RUN git clone --depth 1 --branch v0.3.1 https://github.com/duckdb/pg_duckdb.git pg_duckdb-src && \
@@ -1793,7 +1783,7 @@ RUN set -e \
 #########################################################################################
 FROM build-deps AS exporters
 ARG TARGETARCH
-# Keep sql_exporter version same as in build-tools.Dockerfile and
+# Keep sql_exporter version same as in build-tools/Dockerfile and
 # test_runner/regress/test_compute_metrics.py
 # See comment on the top of the file regading `echo`, `-e` and `\n`
 RUN if [ "$TARGETARCH" = "amd64" ]; then\
@@ -1991,10 +1981,6 @@ RUN apt update && \
         locales \
         lsof \
         procps \
-        bpfcc-tools \
-        libbpfcc \
-        libbpfcc-dev \
-        libclang-dev \
         rsyslog-gnutls \
         screen \
         tcpdump \

compute/package.json

@@ -1,7 +0,0 @@
-{
-  "name": "neon-compute",
-  "private": true,
-  "dependencies": {
-    "@sourcemeta/jsonschema": "9.3.4"
-  }
-} 

compute/patches/anon_v2.patch

@@ -1,22 +1,26 @@
 diff --git a/sql/anon.sql b/sql/anon.sql
-index 0cdc769..b450327 100644
+index 0cdc769..5eab1d6 100644
 --- a/sql/anon.sql
 +++ b/sql/anon.sql
-@@ -1141,3 +1141,15 @@ $$
+@@ -1141,3 +1141,19 @@ $$
  -- TODO : https://en.wikipedia.org/wiki/L-diversity
  
  -- TODO : https://en.wikipedia.org/wiki/T-closeness
 +
 +-- NEON Patches
 +
-+GRANT ALL ON SCHEMA anon to neon_superuser;
-+GRANT ALL ON ALL TABLES IN SCHEMA anon TO neon_superuser;
-+
 +DO $$
++DECLARE
++  privileged_role_name text;
 +BEGIN
-+    IF current_setting('server_version_num')::int >= 150000 THEN
-+        GRANT SET ON PARAMETER anon.transparent_dynamic_masking TO neon_superuser;
-+    END IF;
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT ALL ON SCHEMA anon to %I', privileged_role_name);
++  EXECUTE format('GRANT ALL ON ALL TABLES IN SCHEMA anon TO %I', privileged_role_name);
++
++  IF current_setting('server_version_num')::int >= 150000 THEN
++    EXECUTE format('GRANT SET ON PARAMETER anon.transparent_dynamic_masking TO %I', privileged_role_name);
++  END IF;
 +END $$;
 diff --git a/sql/init.sql b/sql/init.sql
 index 7da6553..9b6164b 100644

compute/patches/pg_duckdb_v031.patch

@@ -21,13 +21,21 @@ index 3235cc8..6b892bc 100644
  include Makefile.global
  
 diff --git a/sql/pg_duckdb--0.2.0--0.3.0.sql b/sql/pg_duckdb--0.2.0--0.3.0.sql
-index d777d76..af60106 100644
+index d777d76..3b54396 100644
 --- a/sql/pg_duckdb--0.2.0--0.3.0.sql
 +++ b/sql/pg_duckdb--0.2.0--0.3.0.sql
-@@ -1056,3 +1056,6 @@ GRANT ALL ON FUNCTION duckdb.cache(TEXT, TEXT) TO PUBLIC;
+@@ -1056,3 +1056,14 @@ GRANT ALL ON FUNCTION duckdb.cache(TEXT, TEXT) TO PUBLIC;
  GRANT ALL ON FUNCTION duckdb.cache_info() TO PUBLIC;
  GRANT ALL ON FUNCTION duckdb.cache_delete(TEXT) TO PUBLIC;
  GRANT ALL ON PROCEDURE duckdb.recycle_ddb() TO PUBLIC;
-+GRANT ALL ON FUNCTION duckdb.install_extension(TEXT) TO neon_superuser;
-+GRANT ALL ON TABLE duckdb.extensions TO neon_superuser;
-+GRANT ALL ON SEQUENCE duckdb.extensions_table_seq TO neon_superuser;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT ALL ON FUNCTION duckdb.install_extension(TEXT) TO %I', privileged_role_name);
++  EXECUTE format('GRANT ALL ON TABLE duckdb.extensions TO %I', privileged_role_name);
++  EXECUTE format('GRANT ALL ON SEQUENCE duckdb.extensions_table_seq TO %I', privileged_role_name);
++END $$;

compute/patches/pg_stat_statements_pg14-16.patch

@@ -0,0 +1,34 @@
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+index 58cdf600fce..8be57a996f6 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+@@ -46,3 +46,12 @@ GRANT SELECT ON pg_stat_statements TO PUBLIC;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset() FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO %I', privileged_role_name);
++END $$;
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+index 6fc3fed4c93..256345a8f79 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+@@ -20,3 +20,12 @@ LANGUAGE C STRICT PARALLEL SAFE;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO %I', privileged_role_name);
++END $$;

compute/patches/pg_stat_statements_pg17.patch

@@ -0,0 +1,52 @@
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql b/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql
+index 0bb2c397711..32764db1d8b 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.10--1.11.sql
+@@ -80,3 +80,12 @@ LANGUAGE C STRICT PARALLEL SAFE;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint, boolean) FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint, boolean) TO %I', privileged_role_name);
++END $$;
+\ No newline at end of file
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+index 58cdf600fce..8be57a996f6 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.4.sql
+@@ -46,3 +46,12 @@ GRANT SELECT ON pg_stat_statements TO PUBLIC;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset() FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset() TO %I', privileged_role_name);
++END $$;
+diff --git a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+index 6fc3fed4c93..256345a8f79 100644
+--- a/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
++++ b/contrib/pg_stat_statements/pg_stat_statements--1.6--1.7.sql
+@@ -20,3 +20,12 @@ LANGUAGE C STRICT PARALLEL SAFE;
+ 
+ -- Don't want this to be available to non-superusers.
+ REVOKE ALL ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) FROM PUBLIC;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT EXECUTE ON FUNCTION pg_stat_statements_reset(Oid, Oid, bigint) TO %I', privileged_role_name);
++END $$;

compute/patches/postgres_fdw.patch

@@ -0,0 +1,17 @@
+diff --git a/contrib/postgres_fdw/postgres_fdw--1.0.sql b/contrib/postgres_fdw/postgres_fdw--1.0.sql
+index a0f0fc1bf45..ee077f2eea6 100644
+--- a/contrib/postgres_fdw/postgres_fdw--1.0.sql
++++ b/contrib/postgres_fdw/postgres_fdw--1.0.sql
+@@ -16,3 +16,12 @@ LANGUAGE C STRICT;
+ CREATE FOREIGN DATA WRAPPER postgres_fdw
+   HANDLER postgres_fdw_handler
+   VALIDATOR postgres_fdw_validator;
++
++DO $$
++DECLARE
++  privileged_role_name text;
++BEGIN
++  privileged_role_name := current_setting('neon.privileged_role_name');
++
++  EXECUTE format('GRANT USAGE ON FOREIGN DATA WRAPPER postgres_fdw TO %I', privileged_role_name);
++END $$;

compute/vm-image-spec-bookworm.yaml

@@ -39,14 +39,6 @@ commands:
     user: nobody
     sysvInitAction: respawn
     shell: '/bin/sql_exporter -config.file=/etc/sql_exporter_autoscaling.yml -web.listen-address=:9499'
-  - name: enable-kernel-modules
-    user: root
-    sysvInitAction: sysinit
-    shell: mkdir -p /lib/ && ln -s /neonvm/tools/lib/modules /lib/
-  - name: enable-bpfs
-    user: root
-    sysvInitAction: sysinit
-    shell: mkdir -p /sys/kernel/debug && mount -t debugfs debugfs /sys/kernel/debug && mount -t bpf bpf /sys/fs/bpf && chmod 755 /sys/fs/bpf
   # Rsyslog by default creates a unix socket under /dev/log . That's where Postgres sends logs also.
   # We run syslog with postgres user so it can't create /dev/log. Instead we configure rsyslog to
   # use a different path for the socket. The symlink actually points to our custom path.
@@ -73,7 +65,7 @@ files:
       # regardless of hostname (ALL)
       #
       # Also allow it to shut down the VM. The fast_import job does that when it's finished.
-      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd, /neonvm/tools/bin/perf, /usr/sbin/profile-bpfcc
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd
   - filename: cgconfig.conf
     content: |
       # Configuration for cgroups in VM compute nodes
@@ -160,8 +152,6 @@ merge: |
   RUN set -e \
       && chmod 0644 /etc/cgconfig.conf
 
-  ENV PERF_BINARY_PATH=/neonvm/tools/bin/perf
-
 
   COPY compute_rsyslog.conf /etc/compute_rsyslog.conf
   RUN chmod 0666 /etc/compute_rsyslog.conf

compute/vm-image-spec-bullseye.yaml

@@ -39,14 +39,6 @@ commands:
     user: nobody
     sysvInitAction: respawn
     shell: '/bin/sql_exporter -config.file=/etc/sql_exporter_autoscaling.yml -web.listen-address=:9499'
-  - name: enable-kernel-modules
-    user: root
-    sysvInitAction: sysinit
-    shell: mkdir -p /lib/ && ln -s /neonvm/tools/lib/modules /lib/
-  - name: enable-bpfs
-    user: root
-    sysvInitAction: sysinit
-    shell: mkdir -p /sys/kernel/debug && mount -t debugfs debugfs /sys/kernel/debug && mount -t bpf bpf /sys/fs/bpf && chmod 755 /sys/fs/bpf
   # Rsyslog by default creates a unix socket under /dev/log . That's where Postgres sends logs also.
   # We run syslog with postgres user so it can't create /dev/log. Instead we configure rsyslog to
   # use a different path for the socket. The symlink actually points to our custom path.
@@ -73,7 +65,7 @@ files:
       # regardless of hostname (ALL)
       #
       # Also allow it to shut down the VM. The fast_import job does that when it's finished.
-      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd, /neonvm/tools/bin/perf, /usr/sbin/profile-bpfcc
+      postgres ALL=(root) NOPASSWD: /neonvm/bin/resize-swap, /neonvm/bin/set-disk-quota, /neonvm/bin/poweroff, /usr/sbin/rsyslogd
   - filename: cgconfig.conf
     content: |
       # Configuration for cgroups in VM compute nodes
@@ -156,8 +148,6 @@ merge: |
   RUN set -e \
       && chmod 0644 /etc/cgconfig.conf
 
-  ENV PERF_BINARY_PATH=/neonvm/tools/bin/perf
-
   COPY compute_rsyslog.conf /etc/compute_rsyslog.conf
   RUN chmod 0666 /etc/compute_rsyslog.conf
   RUN mkdir /var/log/rsyslog && chown -R postgres /var/log/rsyslog

compute_tools/Cargo.toml

@@ -31,7 +31,6 @@ hostname-validator = "1.1"
 indexmap.workspace = true
 itertools.workspace = true
 jsonwebtoken.workspace = true
-libproc.workspace = true
 metrics.workspace = true
 nix.workspace = true
 notify.workspace = true
@@ -50,7 +49,6 @@ serde_with.workspace = true
 serde_json.workspace = true
 signal-hook.workspace = true
 tar.workspace = true
-tempfile.workspace = true
 tower.workspace = true
 tower-http.workspace = true
 tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
@@ -80,10 +78,3 @@ zstd = "0.13"
 bytes = "1.0"
 rust-ini = "0.20.0"
 rlimit = "0.10.1"
-
-inferno = { version = "0.12", default-features = false, features = [
-    "multithreaded",
-    "nameattr",
-] }
-pprof = { version = "0.15", features = ["protobuf-codec", "flamegraph"] }
-prost = "0.12"

compute_tools/src/bin/compute_ctl.rs

@@ -87,6 +87,14 @@ struct Cli {
     #[arg(short = 'C', long, value_name = "DATABASE_URL")]
     pub connstr: String,
 
+    #[arg(
+        long,
+        default_value = "neon_superuser",
+        value_name = "PRIVILEGED_ROLE_NAME",
+        value_parser = Self::parse_privileged_role_name
+    )]
+    pub privileged_role_name: String,
+
     #[cfg(target_os = "linux")]
     #[arg(long, default_value = "neon-postgres")]
     pub cgroup: String,
@@ -149,6 +157,21 @@ impl Cli {
 
         Ok(url)
     }
+
+    /// For simplicity, we do not escape `privileged_role_name` anywhere in the code.
+    /// Since it's a system role, which we fully control, that's fine. Still, let's
+    /// validate it to avoid any surprises.
+    fn parse_privileged_role_name(value: &str) -> Result<String> {
+        use regex::Regex;
+
+        let pattern = Regex::new(r"^[a-z_]+$").unwrap();
+
+        if !pattern.is_match(value) {
+            bail!("--privileged-role-name can only contain lowercase letters and underscores")
+        }
+
+        Ok(value.to_string())
+    }
 }
 
 fn main() -> Result<()> {
@@ -178,6 +201,7 @@ fn main() -> Result<()> {
         ComputeNodeParams {
             compute_id: cli.compute_id,
             connstr,
+            privileged_role_name: cli.privileged_role_name.clone(),
             pgdata: cli.pgdata.clone(),
             pgbin: cli.pgbin.clone(),
             pgversion: get_pg_version_string(&cli.pgbin),
@@ -327,4 +351,49 @@ mod test {
         ])
         .expect_err("URL parameters are not allowed");
     }
+
+    #[test]
+    fn verify_privileged_role_name() {
+        // Valid name
+        let cli = Cli::parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "my_superuser",
+        ]);
+        assert_eq!(cli.privileged_role_name, "my_superuser");
+
+        // Invalid names
+        Cli::try_parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "NeonSuperuser",
+        ])
+        .expect_err("uppercase letters are not allowed");
+
+        Cli::try_parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "$'neon_superuser",
+        ])
+        .expect_err("special characters are not allowed");
+
+        Cli::try_parse_from([
+            "compute_ctl",
+            "--pgdata=test",
+            "--connstr=test",
+            "--compute-id=test",
+            "--privileged-role-name",
+            "",
+        ])
+        .expect_err("empty name is not allowed");
+    }
 }

compute_tools/src/compute.rs

@@ -74,12 +74,20 @@ const DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL: u64 = 3600;
 
 /// Static configuration params that don't change after startup. These mostly
 /// come from the CLI args, or are derived from them.
+#[derive(Clone, Debug)]
 pub struct ComputeNodeParams {
     /// The ID of the compute
     pub compute_id: String,
-    // Url type maintains proper escaping
+
+    /// Url type maintains proper escaping
     pub connstr: url::Url,
 
+    /// The name of the 'weak' superuser role, which we give to the users.
+    /// It follows the allow list approach, i.e., we take a standard role
+    /// and grant it extra permissions with explicit GRANTs here and there,
+    /// and core patches.
+    pub privileged_role_name: String,
+
     pub resize_swap_on_bind: bool,
     pub set_disk_quota_for_fs: Option<String>,
 
@@ -371,9 +379,7 @@ fn maybe_cgexec(cmd: &str) -> Command {
     }
 }
 
-/// A handle to the Postgres process that is running in the compute
-/// node.
-pub struct PostgresHandle {
+struct PostgresHandle {
     postgres: std::process::Child,
     log_collector: JoinHandle<Result<()>>,
 }
@@ -1042,6 +1048,8 @@ impl ComputeNode {
             PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
         };
 
+        self.fix_zenith_signal_neon_signal()?;
+
         let mut state = self.state.lock().unwrap();
         state.metrics.pageserver_connect_micros =
             connected.duration_since(started).as_micros() as u64;
@@ -1051,6 +1059,27 @@ impl ComputeNode {
         Ok(())
     }
 
+    /// Move the Zenith signal file to Neon signal file location.
+    /// This makes Compute compatible with older PageServers that don't yet
+    /// know about the Zenith->Neon rename.
+    fn fix_zenith_signal_neon_signal(&self) -> Result<()> {
+        let datadir = Path::new(&self.params.pgdata);
+
+        let neonsig = datadir.join("neon.signal");
+
+        if neonsig.is_file() {
+            return Ok(());
+        }
+
+        let zenithsig = datadir.join("zenith.signal");
+
+        if zenithsig.is_file() {
+            fs::copy(zenithsig, neonsig)?;
+        }
+
+        Ok(())
+    }
+
     /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
     /// the connection was established, and the (compressed) size of the basebackup.
     fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
@@ -1265,9 +1294,7 @@ impl ComputeNode {
 
         // In case of error, log and fail the check, but don't crash.
         // We're playing it safe because these errors could be transient
-        // and we don't yet retry. Also being careful here allows us to
-        // be backwards compatible with safekeepers that don't have the
-        // TIMELINE_STATUS API yet.
+        // and we don't yet retry.
         if responses.len() < quorum {
             error!(
                 "failed sync safekeepers check {:?} {:?} {:?}",
@@ -1370,6 +1397,7 @@ impl ComputeNode {
         self.create_pgdata()?;
         config::write_postgres_conf(
             pgdata_path,
+            &self.params,
             &pspec.spec,
             self.params.internal_http_port,
             tls_config,
@@ -1718,6 +1746,7 @@ impl ComputeNode {
         }
 
         // Run migrations separately to not hold up cold starts
+        let params = self.params.clone();
         tokio::spawn(async move {
             let mut conf = conf.as_ref().clone();
             conf.application_name("compute_ctl:migrations");
@@ -1729,7 +1758,7 @@ impl ComputeNode {
                             eprintln!("connection error: {e}");
                         }
                     });
-                    if let Err(e) = handle_migrations(&mut client).await {
+                    if let Err(e) = handle_migrations(params, &mut client).await {
                         error!("Failed to run migrations: {}", e);
                     }
                 }
@@ -1808,6 +1837,7 @@ impl ComputeNode {
         let pgdata_path = Path::new(&self.params.pgdata);
         config::write_postgres_conf(
             pgdata_path,
+            &self.params,
             &spec,
             self.params.internal_http_port,
             tls_config,
@@ -2420,14 +2450,31 @@ LIMIT 100",
     pub fn spawn_lfc_offload_task(self: &Arc<Self>, interval: Duration) {
         self.terminate_lfc_offload_task();
         let secs = interval.as_secs();
-        info!("spawning lfc offload worker with {secs}s interval");
         let this = self.clone();
+
+        info!("spawning LFC offload worker with {secs}s interval");
         let handle = spawn(async move {
             let mut interval = time::interval(interval);
             interval.tick().await; // returns immediately
             loop {
                 interval.tick().await;
-                this.offload_lfc_async().await;
+
+                let prewarm_state = this.state.lock().unwrap().lfc_prewarm_state.clone();
+                // Do not offload LFC state if we are currently prewarming or any issue occurred.
+                // If we'd do that, we might override the LFC state in endpoint storage with some
+                // incomplete state. Imagine a situation:
+                // 1. Endpoint started with `autoprewarm: true`
+                // 2. While prewarming is not completed, we upload the new incomplete state
+                // 3. Compute gets interrupted and restarts
+                // 4. We start again and try to prewarm with the state from 2. instead of the previous complete state
+                if matches!(
+                    prewarm_state,
+                    LfcPrewarmState::Completed
+                        | LfcPrewarmState::NotPrewarmed
+                        | LfcPrewarmState::Skipped
+                ) {
+                    this.offload_lfc_async().await;
+                }
             }
         });
         *self.lfc_offload_task.lock().unwrap() = Some(handle);
@@ -2466,7 +2513,7 @@ pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
                 serde_json::to_string(&extensions).expect("failed to serialize extensions list")
             );
         }
-        Err(err) => error!("could not get installed extensions: {err:?}"),
+        Err(err) => error!("could not get installed extensions: {err}"),
     }
     Ok(())
 }

compute_tools/src/compute_prewarm.rs

@@ -89,7 +89,7 @@ impl ComputeNode {
         self.state.lock().unwrap().lfc_offload_state.clone()
     }
 
-    /// If there is a prewarm request ongoing, return false, true otherwise
+    /// If there is a prewarm request ongoing, return `false`, `true` otherwise.
     pub fn prewarm_lfc(self: &Arc<Self>, from_endpoint: Option<String>) -> bool {
         {
             let state = &mut self.state.lock().unwrap().lfc_prewarm_state;
@@ -101,15 +101,25 @@ impl ComputeNode {
 
         let cloned = self.clone();
         spawn(async move {
-            let Err(err) = cloned.prewarm_impl(from_endpoint).await else {
-                cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Completed;
-                return;
-            };
-            crate::metrics::LFC_PREWARM_ERRORS.inc();
-            error!(%err, "prewarming lfc");
-            cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Failed {
-                error: err.to_string(),
+            let state = match cloned.prewarm_impl(from_endpoint).await {
+                Ok(true) => LfcPrewarmState::Completed,
+                Ok(false) => {
+                    info!(
+                        "skipping LFC prewarm because LFC state is not found in endpoint storage"
+                    );
+                    LfcPrewarmState::Skipped
+                }
+                Err(err) => {
+                    crate::metrics::LFC_PREWARM_ERRORS.inc();
+                    error!(%err, "could not prewarm LFC");
+
+                    LfcPrewarmState::Failed {
+                        error: err.to_string(),
+                    }
+                }
             };
+
+            cloned.state.lock().unwrap().lfc_prewarm_state = state;
         });
         true
     }
@@ -120,15 +130,21 @@ impl ComputeNode {
         EndpointStoragePair::from_spec_and_endpoint(state.pspec.as_ref().unwrap(), from_endpoint)
     }
 
-    async fn prewarm_impl(&self, from_endpoint: Option<String>) -> Result<()> {
+    /// Request LFC state from endpoint storage and load corresponding pages into Postgres.
+    /// Returns a result with `false` if the LFC state is not found in endpoint storage.
+    async fn prewarm_impl(&self, from_endpoint: Option<String>) -> Result<bool> {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(from_endpoint)?;
-        info!(%url, "requesting LFC state from endpoint storage");
 
+        info!(%url, "requesting LFC state from endpoint storage");
         let request = Client::new().get(&url).bearer_auth(token);
         let res = request.send().await.context("querying endpoint storage")?;
         let status = res.status();
-        if status != StatusCode::OK {
-            bail!("{status} querying endpoint storage")
+        match status {
+            StatusCode::OK => (),
+            StatusCode::NOT_FOUND => {
+                return Ok(false);
+            }
+            _ => bail!("{status} querying endpoint storage"),
         }
 
         let mut uncompressed = Vec::new();
@@ -141,7 +157,8 @@ impl ComputeNode {
             .await
             .context("decoding LFC state")?;
         let uncompressed_len = uncompressed.len();
-        info!(%url, "downloaded LFC state, uncompressed size {uncompressed_len}, loading into postgres");
+
+        info!(%url, "downloaded LFC state, uncompressed size {uncompressed_len}, loading into Postgres");
 
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
             .await
@@ -149,7 +166,9 @@ impl ComputeNode {
             .query_one("select neon.prewarm_local_cache($1)", &[&uncompressed])
             .await
             .context("loading LFC state into postgres")
-            .map(|_| ())
+            .map(|_| ())?;
+
+        Ok(true)
     }
 
     /// If offload request is ongoing, return false, true otherwise
@@ -177,12 +196,14 @@ impl ComputeNode {
 
     async fn offload_lfc_with_state_update(&self) {
         crate::metrics::LFC_OFFLOADS.inc();
+
         let Err(err) = self.offload_lfc_impl().await else {
             self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Completed;
             return;
         };
+
         crate::metrics::LFC_OFFLOAD_ERRORS.inc();
-        error!(%err, "offloading lfc");
+        error!(%err, "could not offload LFC state to endpoint storage");
         self.state.lock().unwrap().lfc_offload_state = LfcOffloadState::Failed {
             error: err.to_string(),
         };
@@ -190,7 +211,7 @@ impl ComputeNode {
 
     async fn offload_lfc_impl(&self) -> Result<()> {
         let EndpointStoragePair { url, token } = self.endpoint_storage_pair(None)?;
-        info!(%url, "requesting LFC state from postgres");
+        info!(%url, "requesting LFC state from Postgres");
 
         let mut compressed = Vec::new();
         ComputeNode::get_maintenance_client(&self.tokio_conn_conf)
@@ -205,13 +226,17 @@ impl ComputeNode {
             .read_to_end(&mut compressed)
             .await
             .context("compressing LFC state")?;
+
         let compressed_len = compressed.len();
         info!(%url, "downloaded LFC state, compressed size {compressed_len}, writing to endpoint storage");
 
         let request = Client::new().put(url).bearer_auth(token).body(compressed);
         match request.send().await {
             Ok(res) if res.status() == StatusCode::OK => Ok(()),
-            Ok(res) => bail!("Error writing to endpoint storage: {}", res.status()),
+            Ok(res) => bail!(
+                "Request to endpoint storage failed with status: {}",
+                res.status()
+            ),
             Err(err) => Err(err).context("writing to endpoint storage"),
         }
     }

compute_tools/src/config.rs

@@ -9,6 +9,7 @@ use std::path::Path;
 use compute_api::responses::TlsConfig;
 use compute_api::spec::{ComputeAudit, ComputeMode, ComputeSpec, GenericOption};
 
+use crate::compute::ComputeNodeParams;
 use crate::pg_helpers::{
     GenericOptionExt, GenericOptionsSearch, PgOptionsSerialize, escape_conf_value,
 };
@@ -41,6 +42,7 @@ pub fn line_in_file(path: &Path, line: &str) -> Result<bool> {
 /// Create or completely rewrite configuration file specified by `path`
 pub fn write_postgres_conf(
     pgdata_path: &Path,
+    params: &ComputeNodeParams,
     spec: &ComputeSpec,
     extension_server_port: u16,
     tls_config: &Option<TlsConfig>,
@@ -54,14 +56,15 @@ pub fn write_postgres_conf(
         writeln!(file, "{conf}")?;
     }
 
+    // Stripe size GUC should be defined prior to connection string
+    if let Some(stripe_size) = spec.shard_stripe_size {
+        writeln!(file, "neon.stripe_size={stripe_size}")?;
+    }
     // Add options for connecting to storage
     writeln!(file, "# Neon storage settings")?;
     if let Some(s) = &spec.pageserver_connstring {
         writeln!(file, "neon.pageserver_connstring={}", escape_conf_value(s))?;
     }
-    if let Some(stripe_size) = spec.shard_stripe_size {
-        writeln!(file, "neon.stripe_size={stripe_size}")?;
-    }
     if !spec.safekeeper_connstrings.is_empty() {
         let mut neon_safekeepers_value = String::new();
         tracing::info!(
@@ -161,6 +164,12 @@ pub fn write_postgres_conf(
         }
     }
 
+    writeln!(
+        file,
+        "neon.privileged_role_name={}",
+        escape_conf_value(params.privileged_role_name.as_str())
+    )?;
+
     // If there are any extra options in the 'settings' field, append those
     if spec.cluster.settings.is_some() {
         writeln!(file, "# Managed by compute_ctl: begin")?;

compute_tools/src/http/openapi_spec.yaml

@@ -613,11 +613,11 @@ components:
         - skipped
       properties:
         status:
-          description: Lfc prewarm status
-          enum: [not_prewarmed, prewarming, completed, failed]
+          description: LFC prewarm status
+          enum: [not_prewarmed, prewarming, completed, failed, skipped]
           type: string
         error:
-          description: Lfc prewarm error, if any
+          description: LFC prewarm error, if any
           type: string
         total:
           description: Total pages processed
@@ -635,11 +635,11 @@ components:
         - status
       properties:
         status:
-          description: Lfc offload status
+          description: LFC offload status
           enum: [not_offloaded, offloading, completed, failed]
           type: string
         error:
-          description: Lfc offload error, if any
+          description: LFC offload error, if any
           type: string
 
     PromoteState:

compute_tools/src/http/routes/mod.rs

@@ -15,7 +15,6 @@ pub(in crate::http) mod lfc;
 pub(in crate::http) mod metrics;
 pub(in crate::http) mod metrics_json;
 pub(in crate::http) mod promote;
-pub(in crate::http) mod profile;
 pub(in crate::http) mod status;
 pub(in crate::http) mod terminate;
 

compute_tools/src/http/routes/profile.rs

@@ -1,217 +0,0 @@
-//! Contains the route for profiling the compute.
-//!
-//! Profiling the compute means generating a pprof profile of the
-//! postgres processes.
-//!
-//! The profiling is done using the `perf` tool, which is expected to be
-//! available somewhere in `$PATH`.
-use std::sync::atomic::Ordering;
-
-use axum::Json;
-use axum::response::IntoResponse;
-use http::StatusCode;
-use nix::unistd::Pid;
-use once_cell::sync::Lazy;
-use tokio::sync::Mutex;
-
-use crate::http::JsonResponse;
-
-static CANCEL_CHANNEL: Lazy<Mutex<Option<tokio::sync::broadcast::Sender<()>>>> =
-    Lazy::new(|| Mutex::new(None));
-
-fn default_sampling_frequency() -> u16 {
-    100
-}
-
-fn default_timeout_seconds() -> u8 {
-    5
-}
-
-fn deserialize_sampling_frequency<'de, D>(deserializer: D) -> Result<u16, D::Error>
-where
-    D: serde::Deserializer<'de>,
-{
-    use serde::Deserialize;
-
-    const MIN_SAMPLING_FREQUENCY: u16 = 1;
-    const MAX_SAMPLING_FREQUENCY: u16 = 1000;
-
-    let value = u16::deserialize(deserializer)?;
-
-    if !(MIN_SAMPLING_FREQUENCY..=MAX_SAMPLING_FREQUENCY).contains(&value) {
-        return Err(serde::de::Error::custom(format!(
-            "sampling_frequency must be between {MIN_SAMPLING_FREQUENCY} and {MAX_SAMPLING_FREQUENCY}, got {value}"
-        )));
-    }
-    Ok(value)
-}
-
-fn deserialize_profiling_timeout<'de, D>(deserializer: D) -> Result<u8, D::Error>
-where
-    D: serde::Deserializer<'de>,
-{
-    use serde::Deserialize;
-
-    const MIN_TIMEOUT_SECONDS: u8 = 1;
-    const MAX_TIMEOUT_SECONDS: u8 = 60;
-
-    let value = u8::deserialize(deserializer)?;
-
-    if !(MIN_TIMEOUT_SECONDS..=MAX_TIMEOUT_SECONDS).contains(&value) {
-        return Err(serde::de::Error::custom(format!(
-            "timeout_seconds must be between {MIN_TIMEOUT_SECONDS} and {MAX_TIMEOUT_SECONDS}, got {value}"
-        )));
-    }
-    Ok(value)
-}
-
-/// Request parameters for profiling the compute.
-#[derive(Debug, Clone, serde::Deserialize)]
-pub(in crate::http) struct ProfileRequest {
-    /// The profiling tool to use, currently only `perf` is supported.
-    profiler: crate::profiling::ProfileGenerator,
-    #[serde(default = "default_sampling_frequency")]
-    #[serde(deserialize_with = "deserialize_sampling_frequency")]
-    sampling_frequency: u16,
-    #[serde(default = "default_timeout_seconds")]
-    #[serde(deserialize_with = "deserialize_profiling_timeout")]
-    timeout_seconds: u8,
-    #[serde(default)]
-    archive: bool,
-}
-
-/// The HTTP request handler for reporting the profiling status of
-/// the compute.
-pub(in crate::http) async fn profile_status() -> impl IntoResponse {
-    tracing::info!("Profile status request received.");
-
-    let cancel_channel = CANCEL_CHANNEL.lock().await;
-
-    if let Some(tx) = cancel_channel.as_ref() {
-        if tx.receiver_count() > 0 {
-            return JsonResponse::create_response(
-                StatusCode::OK,
-                "Profiling is currently in progress.",
-            );
-        }
-    }
-
-    JsonResponse::create_response(StatusCode::NO_CONTENT, "Profiling is not in progress.")
-}
-
-/// The HTTP request handler for stopping profiling the compute.
-pub(in crate::http) async fn profile_stop() -> impl IntoResponse {
-    tracing::info!("Profile stop request received.");
-
-    match CANCEL_CHANNEL.lock().await.take() {
-        Some(tx) => {
-            if tx.send(()).is_err() {
-                tracing::error!("Failed to send cancellation signal.");
-                return JsonResponse::create_response(
-                    StatusCode::INTERNAL_SERVER_ERROR,
-                    "Failed to send cancellation signal",
-                );
-            }
-            JsonResponse::create_response(StatusCode::OK, "Profiling stopped successfully.")
-        }
-        None => JsonResponse::create_response(
-            StatusCode::PRECONDITION_FAILED,
-            "Profiling is not in progress, there is nothing to stop.",
-        ),
-    }
-}
-
-/// The HTTP request handler for starting profiling the compute.
-pub(in crate::http) async fn profile_start(
-    Json(request): Json<ProfileRequest>,
-) -> impl IntoResponse {
-    tracing::info!("Profile start request received: {request:?}");
-
-    let tx = tokio::sync::broadcast::Sender::<()>::new(1);
-
-    {
-        let mut cancel_channel = CANCEL_CHANNEL.lock().await;
-
-        if cancel_channel.is_some() {
-            return JsonResponse::create_response(
-                StatusCode::CONFLICT,
-                "Profiling is already in progress.",
-            );
-        }
-        *cancel_channel = Some(tx.clone());
-    }
-
-    tracing::info!("Profiling will start with parameters: {request:?}");
-    let pg_pid = Pid::from_raw(crate::compute::PG_PID.load(Ordering::SeqCst) as _);
-
-    let run_with_sudo = !cfg!(feature = "testing");
-
-    let options = crate::profiling::ProfileGenerationOptions {
-        profiler: request.profiler,
-        run_with_sudo,
-        pids: [pg_pid].into_iter().collect(),
-        follow_forks: true,
-        sampling_frequency: request.sampling_frequency as u32,
-        blocklist_symbols: vec![
-            "libc".to_owned(),
-            "libgcc".to_owned(),
-            "pthread".to_owned(),
-            "vdso".to_owned(),
-        ],
-        archive: request.archive,
-    };
-
-    let options = crate::profiling::ProfileGenerationTaskOptions {
-        options,
-        timeout: std::time::Duration::from_secs(request.timeout_seconds as u64),
-        should_stop: Some(tx),
-    };
-
-    let pprof_data = crate::profiling::generate_pprof_profile(options).await;
-
-    if CANCEL_CHANNEL.lock().await.take().is_none() {
-        tracing::error!("Profiling was cancelled from another request.");
-
-        return JsonResponse::create_response(
-            StatusCode::NO_CONTENT,
-            "Profiling was cancelled from another request.",
-        );
-    }
-
-    let pprof_data = match pprof_data {
-        Ok(data) => data,
-        Err(e) => {
-            tracing::error!(error = ?e, "failed to generate pprof data");
-            return JsonResponse::create_response(
-                StatusCode::INTERNAL_SERVER_ERROR,
-                format!("Failed to generate pprof data: {e:?}"),
-            );
-        }
-    };
-
-    tracing::info!("Profiling has completed successfully.");
-
-    let mut headers = http::HeaderMap::new();
-
-    if request.archive {
-        headers.insert(
-            http::header::CONTENT_TYPE,
-            http::HeaderValue::from_static("application/gzip"),
-        );
-        headers.insert(
-            http::header::CONTENT_DISPOSITION,
-            http::HeaderValue::from_static("attachment; filename=\"profile.pb.gz\""),
-        );
-    } else {
-        headers.insert(
-            http::header::CONTENT_TYPE,
-            http::HeaderValue::from_static("application/octet-stream"),
-        );
-        headers.insert(
-            http::header::CONTENT_DISPOSITION,
-            http::HeaderValue::from_static("attachment; filename=\"profile.pb\""),
-        );
-    }
-
-    (headers, pprof_data.0).into_response()
-}

compute_tools/src/http/server.rs

@@ -27,7 +27,6 @@ use super::{
     },
 };
 use crate::compute::ComputeNode;
-use crate::http::routes::profile;
 
 /// `compute_ctl` has two servers: internal and external. The internal server
 /// binds to the loopback interface and handles communication from clients on
@@ -82,14 +81,8 @@ impl From<&Server> for Router<Arc<ComputeNode>> {
             Server::External {
                 config, compute_id, ..
             } => {
-                let unauthenticated_router = Router::<Arc<ComputeNode>>::new()
-                    .route("/metrics", get(metrics::get_metrics))
-                    .route(
-                        "/profile/cpu",
-                        get(profile::profile_status)
-                            .post(profile::profile_start)
-                            .delete(profile::profile_stop),
-                    );
+                let unauthenticated_router =
+                    Router::<Arc<ComputeNode>>::new().route("/metrics", get(metrics::get_metrics));
 
                 let authenticated_router = Router::<Arc<ComputeNode>>::new()
                     .route("/lfc/prewarm", get(lfc::prewarm_state).post(lfc::prewarm))

compute_tools/src/installed_extensions.rs

@@ -2,6 +2,7 @@ use std::collections::HashMap;
 
 use anyhow::Result;
 use compute_api::responses::{InstalledExtension, InstalledExtensions};
+use tokio_postgres::error::Error as PostgresError;
 use tokio_postgres::{Client, Config, NoTls};
 
 use crate::metrics::INSTALLED_EXTENSIONS;
@@ -10,7 +11,7 @@ use crate::metrics::INSTALLED_EXTENSIONS;
 /// and to make database listing query here more explicit.
 ///
 /// Limit the number of databases to 500 to avoid excessive load.
-async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
+async fn list_dbs(client: &mut Client) -> Result<Vec<String>, PostgresError> {
     // `pg_database.datconnlimit = -2` means that the database is in the
     // invalid state
     let databases = client
@@ -37,7 +38,9 @@ async fn list_dbs(client: &mut Client) -> Result<Vec<String>> {
 /// Same extension can be installed in multiple databases with different versions,
 /// so we report a separate metric (number of databases where it is installed)
 /// for each extension version.
-pub async fn get_installed_extensions(mut conf: Config) -> Result<InstalledExtensions> {
+pub async fn get_installed_extensions(
+    mut conf: Config,
+) -> Result<InstalledExtensions, PostgresError> {
     conf.application_name("compute_ctl:get_installed_extensions");
     let databases: Vec<String> = {
         let (mut client, connection) = conf.connect(NoTls).await?;

compute_tools/src/lib.rs

@@ -24,7 +24,6 @@ pub mod monitor;
 pub mod params;
 pub mod pg_helpers;
 pub mod pgbouncer;
-pub mod profiling;
 pub mod rsyslog;
 pub mod spec;
 mod spec_apply;

compute_tools/src/migrations/0001-add_bypass_rls_to_privileged_role.sql

@@ -0,0 +1 @@
+ALTER ROLE {privileged_role_name} BYPASSRLS;

compute_tools/src/migrations/0001-neon_superuser_bypass_rls.sql

@@ -1 +0,0 @@
-ALTER ROLE neon_superuser BYPASSRLS;

compute_tools/src/migrations/0002-alter_roles.sql

@@ -15,7 +15,7 @@ DO $$
 DECLARE
     role_name text;
 BEGIN
-    FOR role_name IN SELECT rolname FROM pg_roles WHERE pg_has_role(rolname, 'neon_superuser', 'member')
+    FOR role_name IN SELECT rolname FROM pg_roles WHERE pg_has_role(rolname, '{privileged_role_name}', 'member')
     LOOP
         RAISE NOTICE 'EXECUTING ALTER ROLE % INHERIT', quote_ident(role_name);
         EXECUTE 'ALTER ROLE ' || quote_ident(role_name) || ' INHERIT';
@@ -23,7 +23,7 @@ BEGIN
 
     FOR role_name IN SELECT rolname FROM pg_roles
         WHERE
-            NOT pg_has_role(rolname, 'neon_superuser', 'member') AND NOT starts_with(rolname, 'pg_')
+            NOT pg_has_role(rolname, '{privileged_role_name}', 'member') AND NOT starts_with(rolname, 'pg_')
     LOOP
         RAISE NOTICE 'EXECUTING ALTER ROLE % NOBYPASSRLS', quote_ident(role_name);
         EXECUTE 'ALTER ROLE ' || quote_ident(role_name) || ' NOBYPASSRLS';

compute_tools/src/migrations/0003-grant_pg_create_subscription_to_privileged_role.sql

@@ -1,6 +1,6 @@
 DO $$
 BEGIN
     IF (SELECT setting::numeric >= 160000 FROM pg_settings WHERE name = 'server_version_num') THEN
-        EXECUTE 'GRANT pg_create_subscription TO neon_superuser';
+        EXECUTE 'GRANT pg_create_subscription TO {privileged_role_name}';
     END IF;
 END $$;

compute_tools/src/migrations/0004-grant_pg_monitor_to_neon_superuser.sql

@@ -1 +0,0 @@
-GRANT pg_monitor TO neon_superuser WITH ADMIN OPTION;

compute_tools/src/migrations/0004-grant_pg_monitor_to_privileged_role.sql

@@ -0,0 +1 @@
+GRANT pg_monitor TO {privileged_role_name} WITH ADMIN OPTION;

compute_tools/src/migrations/0005-grant_all_on_tables_to_privileged_role.sql

@@ -1,4 +1,4 @@
 -- SKIP: Deemed insufficient for allowing relations created by extensions to be
---       interacted with by neon_superuser without permission issues.
+--       interacted with by {privileged_role_name} without permission issues.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO neon_superuser;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO {privileged_role_name};

compute_tools/src/migrations/0006-grant_all_on_sequences_to_privileged_role.sql

@@ -1,4 +1,4 @@
 -- SKIP: Deemed insufficient for allowing relations created by extensions to be
---       interacted with by neon_superuser without permission issues.
+--       interacted with by {privileged_role_name} without permission issues.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO neon_superuser;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO {privileged_role_name};

compute_tools/src/migrations/0007-grant_all_on_tables_with_grant_option_to_privileged_role.sql

@@ -1,3 +1,3 @@
 -- SKIP: Moved inline to the handle_grants() functions.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO neon_superuser WITH GRANT OPTION;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON TABLES TO {privileged_role_name} WITH GRANT OPTION;

compute_tools/src/migrations/0008-grant_all_on_sequences_with_grant_option_to_privileged_role.sql

@@ -1,3 +1,3 @@
 -- SKIP: Moved inline to the handle_grants() functions.
 
-ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO neon_superuser WITH GRANT OPTION;
+ALTER DEFAULT PRIVILEGES IN SCHEMA public GRANT ALL ON SEQUENCES TO {privileged_role_name} WITH GRANT OPTION;

compute_tools/src/migrations/0010-grant_snapshot_synchronization_funcs_to_privileged_role.sql

@@ -1,7 +1,7 @@
 DO $$
 BEGIN
     IF (SELECT setting::numeric >= 160000 FROM pg_settings WHERE name = 'server_version_num') THEN
-       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_export_snapshot TO neon_superuser';
-       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_log_standby_snapshot TO neon_superuser';
+       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_export_snapshot TO {privileged_role_name}';
+       EXECUTE 'GRANT EXECUTE ON FUNCTION pg_log_standby_snapshot TO {privileged_role_name}';
     END IF;
 END $$;

compute_tools/src/migrations/0011-grant_pg_show_replication_origin_status_to_neon_superuser.sql

@@ -1 +0,0 @@
-GRANT EXECUTE ON FUNCTION pg_show_replication_origin_status TO neon_superuser;

compute_tools/src/migrations/0011-grant_pg_show_replication_origin_status_to_privileged_role.sql

@@ -0,0 +1 @@
+GRANT EXECUTE ON FUNCTION pg_show_replication_origin_status TO {privileged_role_name};

compute_tools/src/migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql

@@ -1 +0,0 @@
-GRANT pg_signal_backend TO neon_superuser WITH ADMIN OPTION;

compute_tools/src/migrations/0012-grant_pg_signal_backend_to_privileged_role.sql

@@ -0,0 +1 @@
+GRANT pg_signal_backend TO {privileged_role_name} WITH ADMIN OPTION;

compute_tools/src/spec.rs

@@ -9,6 +9,7 @@ use reqwest::StatusCode;
 use tokio_postgres::Client;
 use tracing::{error, info, instrument};
 
+use crate::compute::ComputeNodeParams;
 use crate::config;
 use crate::metrics::{CPLANE_REQUESTS_TOTAL, CPlaneRequestRPC, UNKNOWN_HTTP_STATUS};
 use crate::migration::MigrationRunner;
@@ -169,7 +170,7 @@ pub async fn handle_neon_extension_upgrade(client: &mut Client) -> Result<()> {
 }
 
 #[instrument(skip_all)]
-pub async fn handle_migrations(client: &mut Client) -> Result<()> {
+pub async fn handle_migrations(params: ComputeNodeParams, client: &mut Client) -> Result<()> {
     info!("handle migrations");
 
     // !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
@@ -178,26 +179,59 @@ pub async fn handle_migrations(client: &mut Client) -> Result<()> {
 
     // Add new migrations in numerical order.
     let migrations = [
-        include_str!("./migrations/0001-neon_superuser_bypass_rls.sql"),
-        include_str!("./migrations/0002-alter_roles.sql"),
-        include_str!("./migrations/0003-grant_pg_create_subscription_to_neon_superuser.sql"),
-        include_str!("./migrations/0004-grant_pg_monitor_to_neon_superuser.sql"),
-        include_str!("./migrations/0005-grant_all_on_tables_to_neon_superuser.sql"),
-        include_str!("./migrations/0006-grant_all_on_sequences_to_neon_superuser.sql"),
-        include_str!(
-            "./migrations/0007-grant_all_on_tables_to_neon_superuser_with_grant_option.sql"
+        &format!(
+            include_str!("./migrations/0001-add_bypass_rls_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
         ),
-        include_str!(
-            "./migrations/0008-grant_all_on_sequences_to_neon_superuser_with_grant_option.sql"
+        &format!(
+            include_str!("./migrations/0002-alter_roles.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0003-grant_pg_create_subscription_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0004-grant_pg_monitor_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0005-grant_all_on_tables_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0006-grant_all_on_sequences_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!(
+                "./migrations/0007-grant_all_on_tables_with_grant_option_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!(
+                "./migrations/0008-grant_all_on_sequences_with_grant_option_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
         ),
         include_str!("./migrations/0009-revoke_replication_for_previously_allowed_roles.sql"),
-        include_str!(
-            "./migrations/0010-grant_snapshot_synchronization_funcs_to_neon_superuser.sql"
+        &format!(
+            include_str!(
+                "./migrations/0010-grant_snapshot_synchronization_funcs_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
         ),
-        include_str!(
-            "./migrations/0011-grant_pg_show_replication_origin_status_to_neon_superuser.sql"
+        &format!(
+            include_str!(
+                "./migrations/0011-grant_pg_show_replication_origin_status_to_privileged_role.sql"
+            ),
+            privileged_role_name = params.privileged_role_name
+        ),
+        &format!(
+            include_str!("./migrations/0012-grant_pg_signal_backend_to_privileged_role.sql"),
+            privileged_role_name = params.privileged_role_name
         ),
-        include_str!("./migrations/0012-grant_pg_signal_backend_to_neon_superuser.sql"),
     ];
 
     MigrationRunner::new(client, &migrations)

compute_tools/src/spec_apply.rs

@@ -13,14 +13,14 @@ use tokio_postgres::Client;
 use tokio_postgres::error::SqlState;
 use tracing::{Instrument, debug, error, info, info_span, instrument, warn};
 
-use crate::compute::{ComputeNode, ComputeState};
+use crate::compute::{ComputeNode, ComputeNodeParams, ComputeState};
 use crate::pg_helpers::{
     DatabaseExt, Escaping, GenericOptionsSearch, RoleExt, get_existing_dbs_async,
     get_existing_roles_async,
 };
 use crate::spec_apply::ApplySpecPhase::{
-    CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreateNeonSuperuser,
-    CreatePgauditExtension, CreatePgauditlogtofileExtension, CreateSchemaNeon,
+    CreateAndAlterDatabases, CreateAndAlterRoles, CreateAvailabilityCheck, CreatePgauditExtension,
+    CreatePgauditlogtofileExtension, CreatePrivilegedRole, CreateSchemaNeon,
     DisablePostgresDBPgAudit, DropInvalidDatabases, DropRoles, FinalizeDropLogicalSubscriptions,
     HandleNeonExtension, HandleOtherExtensions, RenameAndDeleteDatabases, RenameRoles,
     RunInEachDatabase,
@@ -49,6 +49,7 @@ impl ComputeNode {
             // Proceed with post-startup configuration. Note, that order of operations is important.
             let client = Self::get_maintenance_client(&conf).await?;
             let spec = spec.clone();
+            let params = Arc::new(self.params.clone());
 
             let databases = get_existing_dbs_async(&client).await?;
             let roles = get_existing_roles_async(&client)
@@ -157,6 +158,7 @@ impl ComputeNode {
 
                     let conf = Arc::new(conf);
                     let fut = Self::apply_spec_sql_db(
+                        params.clone(),
                         spec.clone(),
                         conf,
                         ctx.clone(),
@@ -185,7 +187,7 @@ impl ComputeNode {
             }
 
             for phase in [
-                CreateNeonSuperuser,
+                CreatePrivilegedRole,
                 DropInvalidDatabases,
                 RenameRoles,
                 CreateAndAlterRoles,
@@ -195,6 +197,7 @@ impl ComputeNode {
             ] {
                 info!("Applying phase {:?}", &phase);
                 apply_operations(
+                    params.clone(),
                     spec.clone(),
                     ctx.clone(),
                     jwks_roles.clone(),
@@ -243,6 +246,7 @@ impl ComputeNode {
                     }
 
                     let fut = Self::apply_spec_sql_db(
+                        params.clone(),
                         spec.clone(),
                         conf,
                         ctx.clone(),
@@ -293,6 +297,7 @@ impl ComputeNode {
             for phase in phases {
                 debug!("Applying phase {:?}", &phase);
                 apply_operations(
+                    params.clone(),
                     spec.clone(),
                     ctx.clone(),
                     jwks_roles.clone(),
@@ -313,7 +318,9 @@ impl ComputeNode {
     /// May opt to not connect to databases that don't have any scheduled
     /// operations.  The function is concurrency-controlled with the provided
     /// semaphore.  The caller has to make sure the semaphore isn't exhausted.
+    #[allow(clippy::too_many_arguments)] // TODO: needs bigger refactoring
     async fn apply_spec_sql_db(
+        params: Arc<ComputeNodeParams>,
         spec: Arc<ComputeSpec>,
         conf: Arc<tokio_postgres::Config>,
         ctx: Arc<tokio::sync::RwLock<MutableApplyContext>>,
@@ -328,6 +335,7 @@ impl ComputeNode {
 
         for subphase in subphases {
             apply_operations(
+                params.clone(),
                 spec.clone(),
                 ctx.clone(),
                 jwks_roles.clone(),
@@ -467,7 +475,7 @@ pub enum PerDatabasePhase {
 
 #[derive(Clone, Debug)]
 pub enum ApplySpecPhase {
-    CreateNeonSuperuser,
+    CreatePrivilegedRole,
     DropInvalidDatabases,
     RenameRoles,
     CreateAndAlterRoles,
@@ -510,6 +518,7 @@ pub struct MutableApplyContext {
 /// - No timeouts have (yet) been implemented.
 /// - The caller is responsible for limiting and/or applying concurrency.
 pub async fn apply_operations<'a, Fut, F>(
+    params: Arc<ComputeNodeParams>,
     spec: Arc<ComputeSpec>,
     ctx: Arc<RwLock<MutableApplyContext>>,
     jwks_roles: Arc<HashSet<String>>,
@@ -527,7 +536,7 @@ where
         debug!("Processing phase {:?}", &apply_spec_phase);
         let ctx = ctx;
 
-        let mut ops = get_operations(&spec, &ctx, &jwks_roles, &apply_spec_phase)
+        let mut ops = get_operations(&params, &spec, &ctx, &jwks_roles, &apply_spec_phase)
             .await?
             .peekable();
 
@@ -588,14 +597,18 @@ where
 /// sort/merge/batch execution, but for now this is a nice way to improve
 /// batching behavior of the commands.
 async fn get_operations<'a>(
+    params: &'a ComputeNodeParams,
     spec: &'a ComputeSpec,
     ctx: &'a RwLock<MutableApplyContext>,
     jwks_roles: &'a HashSet<String>,
     apply_spec_phase: &'a ApplySpecPhase,
 ) -> Result<Box<dyn Iterator<Item = Operation> + 'a + Send>> {
     match apply_spec_phase {
-        ApplySpecPhase::CreateNeonSuperuser => Ok(Box::new(once(Operation {
-            query: include_str!("sql/create_neon_superuser.sql").to_string(),
+        ApplySpecPhase::CreatePrivilegedRole => Ok(Box::new(once(Operation {
+            query: format!(
+                include_str!("sql/create_privileged_role.sql"),
+                privileged_role_name = params.privileged_role_name
+            ),
             comment: None,
         }))),
         ApplySpecPhase::DropInvalidDatabases => {
@@ -697,8 +710,9 @@ async fn get_operations<'a>(
                         None => {
                             let query = if !jwks_roles.contains(role.name.as_str()) {
                                 format!(
-                                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE neon_superuser {}",
+                                    "CREATE ROLE {} INHERIT CREATEROLE CREATEDB BYPASSRLS REPLICATION IN ROLE {} {}",
                                     role.name.pg_quote(),
+                                    params.privileged_role_name,
                                     role.to_pg_options(),
                                 )
                             } else {
@@ -849,8 +863,9 @@ async fn get_operations<'a>(
                                 // ALL PRIVILEGES grants CREATE, CONNECT, and TEMPORARY on the database
                                 // (see https://www.postgresql.org/docs/current/ddl-priv.html)
                                 query: format!(
-                                    "GRANT ALL PRIVILEGES ON DATABASE {} TO neon_superuser",
-                                    db.name.pg_quote()
+                                    "GRANT ALL PRIVILEGES ON DATABASE {} TO {}",
+                                    db.name.pg_quote(),
+                                    params.privileged_role_name
                                 ),
                                 comment: None,
                             },

compute_tools/src/sql/create_neon_superuser.sql

@@ -1,8 +0,0 @@
-DO $$
-    BEGIN
-        IF NOT EXISTS (SELECT FROM pg_catalog.pg_roles WHERE rolname = 'neon_superuser')
-        THEN
-            CREATE ROLE neon_superuser CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
-        END IF;
-    END
-$$;

compute_tools/src/sql/create_privileged_role.sql

@@ -0,0 +1,8 @@
+DO $$
+    BEGIN
+        IF NOT EXISTS (SELECT FROM pg_catalog.pg_roles WHERE rolname = '{privileged_role_name}')
+        THEN
+            CREATE ROLE {privileged_role_name} CREATEDB CREATEROLE NOLOGIN REPLICATION BYPASSRLS IN ROLE pg_read_all_data, pg_write_all_data;
+        END IF;
+    END
+$$;

control_plane/README.md

@@ -8,10 +8,10 @@ code changes locally, but not suitable for running production systems.
 
 ## Example: Start with Postgres 16
 
-To create and start a local development environment with Postgres 16, you will need to provide `--pg-version` flag to 3 of the start-up commands.
+To create and start a local development environment with Postgres 16, you will need to provide `--pg-version` flag to 2 of the start-up commands.
 
 ```shell
-cargo neon init --pg-version 16
+cargo neon init
 cargo neon start
 cargo neon tenant create --set-default --pg-version 16
 cargo neon endpoint create main --pg-version 16

control_plane/src/bin/neon_local.rs

@@ -631,6 +631,10 @@ struct EndpointCreateCmdArgs {
         help = "Allow multiple primary endpoints running on the same branch. Shouldn't be used normally, but useful for tests."
     )]
     allow_multiple: bool,
+
+    /// Only allow changing it on creation
+    #[clap(long, help = "Name of the privileged role for the endpoint")]
+    privileged_role_name: Option<String>,
 }
 
 #[derive(clap::Args)]
@@ -1480,6 +1484,7 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 args.grpc,
                 !args.update_catalog,
                 false,
+                args.privileged_role_name.clone(),
             )?;
         }
         EndpointCmd::Start(args) => {

control_plane/src/broker.rs

@@ -36,7 +36,7 @@ impl StorageBroker {
     pub async fn start(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         let broker = &self.env.broker;
 
-        print!("Starting neon broker at {}", broker.client_url());
+        println!("Starting neon broker at {}", broker.client_url());
 
         let mut args = Vec::new();
 

control_plane/src/endpoint.rs

@@ -32,7 +32,8 @@
 //!     config.json                 - passed to `compute_ctl`
 //!     pgdata/
 //!         postgresql.conf       - copy of postgresql.conf created by `compute_ctl`
-//!         zenith.signal
+//!         neon.signal
+//!         zenith.signal         - copy of neon.signal, for backward compatibility
 //!         <other PostgreSQL files>
 //! ```
 //!
@@ -98,6 +99,7 @@ pub struct EndpointConf {
     features: Vec<ComputeFeature>,
     cluster: Option<Cluster>,
     compute_ctl_config: ComputeCtlConfig,
+    privileged_role_name: Option<String>,
 }
 
 //
@@ -198,6 +200,7 @@ impl ComputeControlPlane {
         grpc: bool,
         skip_pg_catalog_updates: bool,
         drop_subscriptions_before_start: bool,
+        privileged_role_name: Option<String>,
     ) -> Result<Arc<Endpoint>> {
         let pg_port = pg_port.unwrap_or_else(|| self.get_port());
         let external_http_port = external_http_port.unwrap_or_else(|| self.get_port() + 1);
@@ -235,6 +238,7 @@ impl ComputeControlPlane {
             features: vec![],
             cluster: None,
             compute_ctl_config: compute_ctl_config.clone(),
+            privileged_role_name: privileged_role_name.clone(),
         });
 
         ep.create_endpoint_dir()?;
@@ -256,6 +260,7 @@ impl ComputeControlPlane {
                 features: vec![],
                 cluster: None,
                 compute_ctl_config,
+                privileged_role_name,
             })?,
         )?;
         std::fs::write(
@@ -331,6 +336,9 @@ pub struct Endpoint {
 
     /// The compute_ctl config for the endpoint's compute.
     compute_ctl_config: ComputeCtlConfig,
+
+    /// The name of the privileged role for the endpoint.
+    privileged_role_name: Option<String>,
 }
 
 #[derive(PartialEq, Eq)]
@@ -431,6 +439,7 @@ impl Endpoint {
             features: conf.features,
             cluster: conf.cluster,
             compute_ctl_config: conf.compute_ctl_config,
+            privileged_role_name: conf.privileged_role_name,
         })
     }
 
@@ -463,7 +472,7 @@ impl Endpoint {
         conf.append("max_connections", "100");
         conf.append("wal_level", "logical");
         // wal_sender_timeout is the maximum time to wait for WAL replication.
-        // It also defines how often the walreciever will send a feedback message to the wal sender.
+        // It also defines how often the walreceiver will send a feedback message to the wal sender.
         conf.append("wal_sender_timeout", "5s");
         conf.append("listen_addresses", &self.pg_address.ip().to_string());
         conf.append("port", &self.pg_address.port().to_string());
@@ -869,6 +878,10 @@ impl Endpoint {
             cmd.arg("--dev");
         }
 
+        if let Some(privileged_role_name) = self.privileged_role_name.clone() {
+            cmd.args(["--privileged-role-name", &privileged_role_name]);
+        }
+
         let child = cmd.spawn()?;
         // set up a scopeguard to kill & wait for the child in case we panic or bail below
         let child = scopeguard::guard(child, |mut child| {

control_plane/src/local_env.rs

@@ -217,6 +217,9 @@ pub struct NeonStorageControllerConf {
     pub posthog_config: Option<PostHogConfig>,
 
     pub kick_secondary_downloads: Option<bool>,
+
+    #[serde(with = "humantime_serde")]
+    pub shard_split_request_timeout: Option<Duration>,
 }
 
 impl NeonStorageControllerConf {
@@ -250,6 +253,7 @@ impl Default for NeonStorageControllerConf {
             timeline_safekeeper_count: None,
             posthog_config: None,
             kick_secondary_downloads: None,
+            shard_split_request_timeout: None,
         }
     }
 }

control_plane/src/pageserver.rs

@@ -303,7 +303,7 @@ impl PageServerNode {
     async fn start_node(&self, retry_timeout: &Duration) -> anyhow::Result<()> {
         // TODO: using a thread here because start_process() is not async but we need to call check_status()
         let datadir = self.repo_path();
-        print!(
+        println!(
             "Starting pageserver node {} at '{}' in {:?}, retrying for {:?}",
             self.conf.id,
             self.pg_connection_config.raw_address(),

control_plane/src/safekeeper.rs

@@ -127,7 +127,7 @@ impl SafekeeperNode {
         extra_opts: &[String],
         retry_timeout: &Duration,
     ) -> anyhow::Result<()> {
-        print!(
+        println!(
             "Starting safekeeper at '{}' in '{}', retrying for {:?}",
             self.pg_connection_config.raw_address(),
             self.datadir_path().display(),

control_plane/src/storage_controller.rs

@@ -648,6 +648,13 @@ impl StorageController {
             args.push(format!("--timeline-safekeeper-count={sk_cnt}"));
         }
 
+        if let Some(duration) = self.config.shard_split_request_timeout {
+            args.push(format!(
+                "--shard-split-request-timeout={}",
+                humantime::Duration::from(duration)
+            ));
+        }
+
         let mut envs = vec![
             ("LD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
             ("DYLD_LIBRARY_PATH".to_owned(), pg_lib_dir.to_string()),
@@ -660,7 +667,7 @@ impl StorageController {
             ));
         }
 
-        println!("Starting storage controller");
+        println!("Starting storage controller at {scheme}://{host}:{listen_port}");
 
         background_process::start_process(
             COMMAND,

control_plane/storcon_cli/Cargo.toml

@@ -14,6 +14,7 @@ humantime.workspace = true
 pageserver_api.workspace = true
 pageserver_client.workspace = true
 reqwest.workspace = true
+safekeeper_api.workspace=true
 serde_json = { workspace = true, features = ["raw_value"] }
 storage_controller_client.workspace = true
 tokio.workspace = true

control_plane/storcon_cli/src/main.rs

@@ -11,7 +11,7 @@ use pageserver_api::controller_api::{
     PlacementPolicy, SafekeeperDescribeResponse, SafekeeperSchedulingPolicyRequest,
     ShardSchedulingPolicy, ShardsPreferredAzsRequest, ShardsPreferredAzsResponse,
     SkSchedulingPolicy, TenantCreateRequest, TenantDescribeResponse, TenantPolicyRequest,
-    TenantShardMigrateRequest, TenantShardMigrateResponse,
+    TenantShardMigrateRequest, TenantShardMigrateResponse, TimelineSafekeeperMigrateRequest,
 };
 use pageserver_api::models::{
     EvictionPolicy, EvictionPolicyLayerAccessThreshold, ShardParameters, TenantConfig,
@@ -21,6 +21,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{ShardStripeSize, TenantShardId};
 use pageserver_client::mgmt_api::{self};
 use reqwest::{Certificate, Method, StatusCode, Url};
+use safekeeper_api::models::TimelineLocateResponse;
 use storage_controller_client::control_api::Client;
 use utils::id::{NodeId, TenantId, TimelineId};
 
@@ -75,6 +76,12 @@ enum Command {
     NodeStartDelete {
         #[arg(long)]
         node_id: NodeId,
+        /// When `force` is true, skip waiting for shards to prewarm during migration.
+        /// This can significantly speed up node deletion since prewarming all shards
+        /// can take considerable time, but may result in slower initial access to
+        /// migrated shards until they warm up naturally.
+        #[arg(long)]
+        force: bool,
     },
     /// Cancel deletion of the specified pageserver and wait for `timeout`
     /// for the operation to be canceled. May be retried.
@@ -279,6 +286,23 @@ enum Command {
         #[arg(long)]
         concurrency: Option<usize>,
     },
+    /// Locate safekeepers for a timeline from the storcon DB.
+    TimelineLocate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+    },
+    /// Migrate a timeline to a new set of safekeepers
+    TimelineSafekeeperMigrate {
+        #[arg(long)]
+        tenant_id: TenantId,
+        #[arg(long)]
+        timeline_id: TimelineId,
+        /// Example: --new-sk-set 1,2,3
+        #[arg(long, required = true, value_delimiter = ',')]
+        new_sk_set: Vec<NodeId>,
+    },
 }
 
 #[derive(Parser)]
@@ -458,6 +482,7 @@ async fn main() -> anyhow::Result<()> {
                         listen_http_port,
                         listen_https_port,
                         availability_zone_id: AvailabilityZone(availability_zone_id),
+                        node_ip_addr: None,
                     }),
                 )
                 .await?;
@@ -933,13 +958,14 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::DELETE, format!("control/v1/node/{node_id}"), None)
                 .await?;
         }
-        Command::NodeStartDelete { node_id } => {
+        Command::NodeStartDelete { node_id, force } => {
+            let query = if force {
+                format!("control/v1/node/{node_id}/delete?force=true")
+            } else {
+                format!("control/v1/node/{node_id}/delete")
+            };
             storcon_client
-                .dispatch::<(), ()>(
-                    Method::PUT,
-                    format!("control/v1/node/{node_id}/delete"),
-                    None,
-                )
+                .dispatch::<(), ()>(Method::PUT, query, None)
                 .await?;
             println!("Delete started for {node_id}");
         }
@@ -1324,7 +1350,7 @@ async fn main() -> anyhow::Result<()> {
             concurrency,
         } => {
             let mut path = format!(
-                "/v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
+                "v1/tenant/{tenant_shard_id}/timeline/{timeline_id}/download_heatmap_layers",
             );
 
             if let Some(c) = concurrency {
@@ -1335,6 +1361,41 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::POST, path, None)
                 .await?;
         }
+        Command::TimelineLocate {
+            tenant_id,
+            timeline_id,
+        } => {
+            let path = format!("debug/v1/tenant/{tenant_id}/timeline/{timeline_id}/locate");
+
+            let resp = storcon_client
+                .dispatch::<(), TimelineLocateResponse>(Method::GET, path, None)
+                .await?;
+
+            let sk_set = resp.sk_set.iter().map(|id| id.0 as i64).collect::<Vec<_>>();
+            let new_sk_set = resp
+                .new_sk_set
+                .as_ref()
+                .map(|ids| ids.iter().map(|id| id.0 as i64).collect::<Vec<_>>());
+
+            println!("generation = {}", resp.generation);
+            println!("sk_set = {sk_set:?}");
+            println!("new_sk_set = {new_sk_set:?}");
+        }
+        Command::TimelineSafekeeperMigrate {
+            tenant_id,
+            timeline_id,
+            new_sk_set,
+        } => {
+            let path = format!("v1/tenant/{tenant_id}/timeline/{timeline_id}/safekeeper_migrate");
+
+            storcon_client
+                .dispatch::<_, ()>(
+                    Method::POST,
+                    path,
+                    Some(TimelineSafekeeperMigrateRequest { new_sk_set }),
+                )
+                .await?;
+        }
     }
 
     Ok(())

docs/continuous-profiling.md

@@ -1,58 +0,0 @@
-# Continuous Crofiling (Compute)
-
-The continuous profiling of the compute node is performed by `perf` or `bcc-tools`, the latter is preferred.
-
-The executables profiled are all the postgres-related ones only, excluding the actual compute code (Rust). This can be done as well but
-was not the main goal.
-
-## Tools
-
-The aforementioned tools are available within the same Docker image as
-the compute node itself, but the corresponding dependencies linux the
-linux kernel headers and the linux kernel itself are not and can't be
-for obvious reasons. To solve that, as we run the compute nodes as a
-virtual machine (qemu), we need to deliver these dependencies to it.
-This is done by the `autoscaling` part, which builds and deploys the
-kernel headers, needed modules, and the `perf` binary into an ext4-fs
-disk image, which is later attached to the VM and is symlinked to be
-made available for the compute node.
-
-## Output
-
-The output of the profiling is always a binary file in the same format
-of `pprof`. It can, however, be archived by `gzip` additionally, if the
-corresponding argument is provided in the JSON request.
-
-## REST API
-
-### Test profiling
-
-One can test the profiling after connecting to the VM and running:
-
-```sh
-curl -X POST -H "Content-Type: application/json" http://localhost:3080/profile/cpu -d '{"profiler": {"BccProfile": null}, "sampling_frequency": 99, "timeout_seconds": 5, "archive": false}' -v --output profile.pb
-```
-
-This uses the `Bcc` profiler and does not archive the output. The
-profiling data will be saved into the `profile.pb` file locally.
-
-**Only one profiling session can be run at a time.**
-
-To check the profiling status (to see whether it is already running or
-not), one can perform the `GET` request:
-
-```sh
-curl http://localhost:3080/profile/cpu -v
-```
-
-The profiling can be stopped by performing the `DELETE` request:
-
-```sh
-curl -X DELETE http://localhost:3080/profile/cpu -v
-```
-
-## Supported profiling data
-
-For now, only the CPU profiling is done and ther is no heap profiling.
-Also, only the postgres-related executables are tracked, the compute
-(Rust) part itself **is not tracked**.

docs/core_changes.md

@@ -129,9 +129,10 @@ segment to bootstrap the WAL writing, but it doesn't contain the checkpoint reco
 changes in xlog.c, to allow starting the compute node without reading the last checkpoint record
 from WAL.
 
-This includes code to read the `zenith.signal` file, which tells the startup code the LSN to start
-at. When the `zenith.signal` file is present, the startup uses that LSN instead of the last
-checkpoint's LSN. The system is known to be consistent at that LSN, without any WAL redo.
+This includes code to read the `neon.signal` (also `zenith.signal`) file, which tells the startup 
+code the LSN to start at. When the `neon.signal` file is present, the startup uses that LSN
+instead of the last checkpoint's LSN. The system is known to be consistent at that LSN, without 
+any WAL redo.
 
 
 ### How to get rid of the patch

docs/pageserver-services.md

@@ -75,7 +75,7 @@ CLI examples:
 * AWS S3  : `env AWS_ACCESS_KEY_ID='SOMEKEYAAAAASADSAH*#' AWS_SECRET_ACCESS_KEY='SOMEsEcReTsd292v' ${PAGESERVER_BIN} -c "remote_storage={bucket_name='some-sample-bucket',bucket_region='eu-north-1', prefix_in_bucket='/test_prefix/'}"`
 
 For Amazon AWS S3, a key id and secret access key could be located in `~/.aws/credentials` if awscli was ever configured to work with the desired bucket, on the AWS Settings page for a certain user. Also note, that the bucket names does not contain any protocols when used on AWS.
-For local S3 installations, refer to the their documentation for name format and credentials.
+For local S3 installations, refer to their documentation for name format and credentials.
 
 Similar to other pageserver settings, toml config file can be used to configure either of the storages as backup targets.
 Required sections are:

libs/alloc-metrics/Cargo.toml

@@ -0,0 +1,18 @@
+[package]
+name = "alloc-metrics"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+metrics.workspace = true
+measured.workspace = true
+thread_local.workspace = true
+
+[dev-dependencies]
+criterion.workspace = true
+tikv-jemallocator.workspace = true
+
+[[bench]]
+harness = false
+name = "alloc"

libs/alloc-metrics/benches/alloc.rs

@@ -0,0 +1,110 @@
+use std::alloc::{GlobalAlloc, Layout, System, handle_alloc_error};
+
+use alloc_metrics::TrackedAllocator;
+use criterion::{
+    AxisScale, BenchmarkGroup, BenchmarkId, Criterion, PlotConfiguration, measurement::Measurement,
+};
+use measured::FixedCardinalityLabel;
+use tikv_jemallocator::Jemalloc;
+
+fn main() {
+    let mut c = Criterion::default().configure_from_args();
+    bench(&mut c);
+    c.final_summary();
+}
+
+#[rustfmt::skip]
+fn bench(c: &mut Criterion) {
+    bench_alloc(c.benchmark_group("alloc/system"),  &System, &ALLOC_SYSTEM);
+    bench_alloc(c.benchmark_group("alloc/jemalloc"), &Jemalloc, &ALLOC_JEMALLOC);
+
+    bench_dealloc(c.benchmark_group("dealloc/system"), &System, &ALLOC_SYSTEM);
+    bench_dealloc(c.benchmark_group("dealloc/jemalloc"), &Jemalloc, &ALLOC_JEMALLOC);
+}
+
+#[derive(FixedCardinalityLabel, Clone, Copy, Debug)]
+#[label(singleton = "memory_context")]
+pub enum MemoryContext {
+    Root,
+    Test,
+}
+
+static ALLOC_SYSTEM: TrackedAllocator<System, MemoryContext> =
+    unsafe { TrackedAllocator::new(System, MemoryContext::Root) };
+static ALLOC_JEMALLOC: TrackedAllocator<Jemalloc, MemoryContext> =
+    unsafe { TrackedAllocator::new(Jemalloc, MemoryContext::Root) };
+
+const KB: u64 = 1024;
+const SIZES: [u64; 6] = [64, 256, KB, 4 * KB, 16 * KB, KB * KB];
+
+fn bench_alloc<A: GlobalAlloc>(
+    mut g: BenchmarkGroup<'_, impl Measurement>,
+    alloc1: &'static A,
+    alloc2: &'static TrackedAllocator<A, MemoryContext>,
+) {
+    g.plot_config(PlotConfiguration::default().summary_scale(AxisScale::Logarithmic));
+    for size in SIZES {
+        let layout = Layout::from_size_align(size as usize, 8).unwrap();
+
+        g.throughput(criterion::Throughput::Bytes(size));
+        g.bench_with_input(BenchmarkId::new("default", size), &layout, |b, &layout| {
+            let bs = criterion::BatchSize::NumBatches(10 + size.ilog2() as u64);
+            b.iter_batched(|| {}, |()| Alloc::new(alloc1, layout), bs);
+        });
+        g.bench_with_input(BenchmarkId::new("tracked", size), &layout, |b, &layout| {
+            let _scope = alloc2.scope(MemoryContext::Test);
+
+            let bs = criterion::BatchSize::NumBatches(10 + size.ilog2() as u64);
+            b.iter_batched(|| {}, |()| Alloc::new(alloc2, layout), bs);
+        });
+    }
+}
+
+fn bench_dealloc<A: GlobalAlloc>(
+    mut g: BenchmarkGroup<'_, impl Measurement>,
+    alloc1: &'static A,
+    alloc2: &'static TrackedAllocator<A, MemoryContext>,
+) {
+    g.plot_config(PlotConfiguration::default().summary_scale(AxisScale::Logarithmic));
+    for size in SIZES {
+        let layout = Layout::from_size_align(size as usize, 8).unwrap();
+
+        g.throughput(criterion::Throughput::Bytes(size));
+        g.bench_with_input(BenchmarkId::new("default", size), &layout, |b, &layout| {
+            let bs = criterion::BatchSize::NumBatches(10 + size.ilog2() as u64);
+            b.iter_batched(|| Alloc::new(alloc1, layout), drop, bs);
+        });
+        g.bench_with_input(BenchmarkId::new("tracked", size), &layout, |b, &layout| {
+            let _scope = alloc2.scope(MemoryContext::Test);
+
+            let bs = criterion::BatchSize::NumBatches(10 + size.ilog2() as u64);
+            b.iter_batched(|| Alloc::new(alloc2, layout), drop, bs);
+        });
+    }
+}
+
+struct Alloc<'a, A: GlobalAlloc> {
+    alloc: &'a A,
+    ptr: *mut u8,
+    layout: Layout,
+}
+
+impl<'a, A: GlobalAlloc> Alloc<'a, A> {
+    fn new(alloc: &'a A, layout: Layout) -> Self {
+        let ptr = unsafe { alloc.alloc(layout) };
+        if ptr.is_null() {
+            handle_alloc_error(layout);
+        }
+
+        // actually make the page resident.
+        unsafe { ptr.cast::<u8>().write(1) };
+
+        Self { alloc, ptr, layout }
+    }
+}
+
+impl<'a, A: GlobalAlloc> Drop for Alloc<'a, A> {
+    fn drop(&mut self) {
+        unsafe { self.alloc.dealloc(self.ptr, self.layout) };
+    }
+}

libs/alloc-metrics/src/counters.rs

@@ -0,0 +1,48 @@
+use std::marker::PhantomData;
+
+use measured::{
+    FixedCardinalityLabel, LabelGroup, label::StaticLabelSet, metric::MetricFamilyEncoding,
+};
+use metrics::{CounterPairAssoc, Dec, Inc, MeasuredCounterPairState};
+
+use crate::metric_vec::DenseMetricVec;
+
+pub struct DenseCounterPairVec<
+    A: CounterPairAssoc<LabelGroupSet = StaticLabelSet<L>>,
+    L: FixedCardinalityLabel + LabelGroup,
+> {
+    pub vec: DenseMetricVec<MeasuredCounterPairState, L>,
+    pub _marker: PhantomData<A>,
+}
+
+impl<A: CounterPairAssoc<LabelGroupSet = StaticLabelSet<L>>, L: FixedCardinalityLabel + LabelGroup>
+    DenseCounterPairVec<A, L>
+{
+    pub fn new() -> Self {
+        Self {
+            vec: DenseMetricVec::new(),
+            _marker: PhantomData,
+        }
+    }
+}
+
+impl<T, A, L> ::measured::metric::group::MetricGroup<T> for DenseCounterPairVec<A, L>
+where
+    T: ::measured::metric::group::Encoding,
+    ::measured::metric::counter::CounterState: ::measured::metric::MetricEncoding<T>,
+    A: CounterPairAssoc<LabelGroupSet = StaticLabelSet<L>>,
+    L: FixedCardinalityLabel + LabelGroup,
+{
+    fn collect_group_into(&self, enc: &mut T) -> Result<(), T::Err> {
+        // write decrement first to avoid a race condition where inc - dec < 0
+        T::write_help(enc, A::DEC_NAME, A::DEC_HELP)?;
+        self.vec
+            .collect_family_into(A::DEC_NAME, &mut Dec(&mut *enc))?;
+
+        T::write_help(enc, A::INC_NAME, A::INC_HELP)?;
+        self.vec
+            .collect_family_into(A::INC_NAME, &mut Inc(&mut *enc))?;
+
+        Ok(())
+    }
+}

libs/alloc-metrics/src/lib.rs

@@ -0,0 +1,441 @@
+//! Tagged allocator measurements.
+
+mod counters;
+mod metric_vec;
+
+use std::{
+    alloc::{GlobalAlloc, Layout},
+    cell::Cell,
+    marker::PhantomData,
+    sync::{
+        OnceLock,
+        atomic::{AtomicU64, Ordering::Relaxed},
+    },
+};
+
+use measured::{
+    FixedCardinalityLabel, LabelGroup, MetricGroup,
+    label::StaticLabelSet,
+    metric::{MetricEncoding, counter::CounterState, group::Encoding, name::MetricName},
+};
+use metrics::{CounterPairAssoc, MeasuredCounterPairState};
+use thread_local::ThreadLocal;
+
+type AllocCounter<T> = counters::DenseCounterPairVec<AllocPair<T>, T>;
+
+pub struct TrackedAllocator<A, T: 'static + Send + Sync + FixedCardinalityLabel + LabelGroup> {
+    inner: A,
+
+    /// potentially high-content fallback if the thread was not registered.
+    default_counters: MeasuredCounterPairState,
+    /// Default tag to use if this thread is not registered.
+    default_tag: T,
+
+    thread: OnceLock<RegisteredThread<T>>,
+
+    /// where thread alloc data is eventually saved to, even if threads are shutdown.
+    global: OnceLock<AllocCounter<T>>,
+}
+
+impl<A, T> TrackedAllocator<A, T>
+where
+    T: 'static + Send + Sync + FixedCardinalityLabel + LabelGroup,
+{
+    /// # Safety
+    ///
+    /// [`FixedCardinalityLabel`] must be implemented correctly, fully dense, and must not panic.
+    pub const unsafe fn new(alloc: A, default: T) -> Self {
+        TrackedAllocator {
+            inner: alloc,
+            default_tag: default,
+            default_counters: MeasuredCounterPairState {
+                inc: CounterState {
+                    count: AtomicU64::new(0),
+                },
+                dec: CounterState {
+                    count: AtomicU64::new(0),
+                },
+            },
+            thread: OnceLock::new(),
+            global: OnceLock::new(),
+        }
+    }
+
+    /// Allocations
+    pub fn register_thread(&'static self) {
+        self.register_thread_inner();
+    }
+
+    pub fn scope(&'static self, tag: T) -> AllocScope<'static, T> {
+        let cell = self.register_thread_inner();
+        let last = cell.replace(tag);
+        AllocScope { cell, last }
+    }
+
+    fn register_thread_inner(&'static self) -> &'static Cell<T> {
+        let thread = self.thread.get_or_init(|| RegisteredThread {
+            scope: ThreadLocal::new(),
+            state: ThreadLocal::new(),
+        });
+
+        thread.state.get_or(|| ThreadState {
+            counters: AllocCounter::new(),
+            global: self.global.get_or_init(AllocCounter::new),
+        });
+
+        thread.scope.get_or(|| Cell::new(self.default_tag))
+    }
+}
+
+macro_rules! alloc {
+    ($alloc_fn:ident) => {
+        unsafe fn $alloc_fn(&self, layout: Layout) -> *mut u8 {
+            let Ok((tagged_layout, tag_offset)) = layout.extend(Layout::new::<T>()) else {
+                return std::ptr::null_mut();
+            };
+            let tagged_layout = tagged_layout.pad_to_align();
+
+            // Safety: The layout is not zero-sized.
+            let ptr = unsafe { self.inner.$alloc_fn(tagged_layout) };
+
+            // allocation failed.
+            if ptr.is_null() {
+                return ptr;
+            }
+
+            // We are being very careful here to not allocate or panic.
+            let thread = self.thread.get().map(|s| (s.scope.get(), s.state.get()));
+            let tag = thread.and_then(|t| t.0).map_or(self.default_tag, Cell::get);
+
+            // Allocation successful. Write our tag
+            // Safety: tag_offset is inbounds of the ptr
+            unsafe { ptr.add(tag_offset).cast::<T>().write(tag) }
+
+            let counters = thread.and_then(|t| t.1).map(|s| &s.counters);
+            let metric = if let Some(counters) = counters {
+                counters.vec.get_metric(tag)
+            } else {
+                // if tag is not default, then the thread state would have been registered, therefore tag must be default.
+                &self.default_counters
+            };
+
+            metric.inc.count.fetch_add(layout.size() as u64, Relaxed);
+
+            ptr
+        }
+    };
+}
+
+// We will tag our allocation by adding `T` to the end of the layout.
+// This is ok only as long as it does not overflow. If it does, we will
+// just fail the allocation by returning null.
+//
+// Safety: we will not unwind during alloc, and we will ensure layouts are handled correctly.
+unsafe impl<A, T> GlobalAlloc for TrackedAllocator<A, T>
+where
+    A: GlobalAlloc,
+    T: 'static + Send + Sync + FixedCardinalityLabel + LabelGroup,
+{
+    alloc!(alloc);
+    alloc!(alloc_zeroed);
+
+    unsafe fn realloc(&self, ptr: *mut u8, layout: Layout, new_size: usize) -> *mut u8 {
+        // SAFETY: the caller must ensure that the `new_size` does not overflow.
+        // `layout.align()` comes from a `Layout` and is thus guaranteed to be valid.
+        let new_layout = unsafe { Layout::from_size_align_unchecked(new_size, layout.align()) };
+
+        let Ok((new_tagged_layout, new_tag_offset)) = new_layout.extend(Layout::new::<T>()) else {
+            return std::ptr::null_mut();
+        };
+        let new_tagged_layout = new_tagged_layout.pad_to_align();
+
+        let Ok((tagged_layout, tag_offset)) = layout.extend(Layout::new::<T>()) else {
+            // Safety: This layout clearly did not match what was originally allocated,
+            // otherwise alloc() would have caught this error and returned null.
+            unsafe { std::hint::unreachable_unchecked() }
+        };
+        let tagged_layout = tagged_layout.pad_to_align();
+
+        // get the tag set during alloc
+        // Safety: tag_offset is inbounds of the ptr
+        let tag = unsafe { ptr.add(tag_offset).cast::<T>().read() };
+
+        // Safety: layout sizes are correct
+        let new_ptr = unsafe {
+            self.inner
+                .realloc(ptr, tagged_layout, new_tagged_layout.size())
+        };
+
+        // allocation failed.
+        if new_ptr.is_null() {
+            return new_ptr;
+        }
+
+        // We are being very careful here to not allocate or panic.
+        let thread = self.thread.get().map(|s| (s.scope.get(), s.state.get()));
+        let new_tag = thread.and_then(|t| t.0).map_or(self.default_tag, Cell::get);
+
+        // Allocation successful. Write our tag
+        // Safety: new_tag_offset is inbounds of the ptr
+        unsafe { new_ptr.add(new_tag_offset).cast::<T>().write(new_tag) }
+
+        let counters = thread.and_then(|t| t.1).map(|s| &s.counters);
+        let counters = counters.or_else(|| self.global.get());
+        let (new_metric, old_metric) = if let Some(counters) = counters {
+            let new_metric = counters.vec.get_metric(new_tag);
+            let old_metric = counters.vec.get_metric(tag);
+
+            (new_metric, old_metric)
+        } else {
+            // no tag was registered at all, therefore both tags must be default.
+            (&self.default_counters, &self.default_counters)
+        };
+
+        let (inc, dec) = if tag.encode() != new_tag.encode() {
+            (new_layout.size() as u64, layout.size() as u64)
+        } else if new_layout.size() > layout.size() {
+            ((new_layout.size() - layout.size()) as u64, 0)
+        } else {
+            (0, (layout.size() - new_layout.size()) as u64)
+        };
+
+        new_metric.inc.count.fetch_add(inc, Relaxed);
+        old_metric.dec.count.fetch_add(dec, Relaxed);
+
+        new_ptr
+    }
+
+    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
+        let Ok((tagged_layout, tag_offset)) = layout.extend(Layout::new::<T>()) else {
+            // Safety: This layout clearly did not match what was originally allocated,
+            // otherwise alloc() would have caught this error and returned null.
+            unsafe { std::hint::unreachable_unchecked() }
+        };
+        let tagged_layout = tagged_layout.pad_to_align();
+
+        // get the tag set during alloc
+        // Safety: tag_offset is inbounds of the ptr
+        let tag = unsafe { ptr.add(tag_offset).cast::<T>().read() };
+
+        // Safety: caller upholds contract for us
+        unsafe { self.inner.dealloc(ptr, tagged_layout) }
+
+        // We are being very careful here to not allocate or panic.
+        let thread = self.thread.get().map(|s| (s.scope.get(), s.state.get()));
+        let counters = thread.and_then(|t| t.1).map(|s| &s.counters);
+        let counters = counters.or_else(|| self.global.get());
+
+        let metric = if let Some(counters) = counters {
+            counters.vec.get_metric(tag)
+        } else {
+            // if tag is not default, then global would have been registered, therefore tag must be default.
+            &self.default_counters
+        };
+
+        metric.dec.count.fetch_add(layout.size() as u64, Relaxed);
+    }
+}
+
+pub struct AllocScope<'a, T: FixedCardinalityLabel> {
+    cell: &'a Cell<T>,
+    last: T,
+}
+
+impl<'a, T: FixedCardinalityLabel> Drop for AllocScope<'a, T> {
+    fn drop(&mut self) {
+        self.cell.set(self.last);
+    }
+}
+
+struct AllocPair<T>(PhantomData<T>);
+
+impl<T: FixedCardinalityLabel + LabelGroup> CounterPairAssoc for AllocPair<T> {
+    const INC_NAME: &'static MetricName = MetricName::from_str("allocated_bytes");
+    const DEC_NAME: &'static MetricName = MetricName::from_str("deallocated_bytes");
+
+    const INC_HELP: &'static str = "total number of bytes allocated";
+    const DEC_HELP: &'static str = "total number of bytes deallocated";
+
+    type LabelGroupSet = StaticLabelSet<T>;
+}
+
+struct RegisteredThread<T: 'static + Send + Sync + FixedCardinalityLabel + LabelGroup> {
+    /// Current memory context for this thread.
+    scope: ThreadLocal<Cell<T>>,
+    /// per thread state containing low contention counters for faster allocations.
+    state: ThreadLocal<ThreadState<T>>,
+}
+
+struct ThreadState<T: 'static + FixedCardinalityLabel + LabelGroup> {
+    counters: AllocCounter<T>,
+    global: &'static AllocCounter<T>,
+}
+
+// Ensure the counters are measured on thread destruction.
+impl<T: 'static + FixedCardinalityLabel + LabelGroup> Drop for ThreadState<T> {
+    fn drop(&mut self) {
+        // iterate over all labels
+        for tag in (0..T::cardinality()).map(T::decode) {
+            // load and reset the counts in the thread-local counters.
+            let m = self.counters.vec.get_metric_mut(tag);
+            let inc = *m.inc.count.get_mut();
+            let dec = *m.dec.count.get_mut();
+
+            // add the counts into the global counters.
+            let m = self.global.vec.get_metric(tag);
+            m.inc.count.fetch_add(inc, Relaxed);
+            m.dec.count.fetch_add(dec, Relaxed);
+        }
+    }
+}
+
+impl<A, T, Enc> MetricGroup<Enc> for TrackedAllocator<A, T>
+where
+    T: 'static + Send + Sync + FixedCardinalityLabel + LabelGroup,
+    Enc: Encoding,
+    CounterState: MetricEncoding<Enc>,
+{
+    fn collect_group_into(&self, enc: &mut Enc) -> Result<(), Enc::Err> {
+        let global = self.global.get_or_init(AllocCounter::new);
+
+        // iterate over all counter threads
+        for s in self.thread.get().into_iter().flat_map(|s| s.state.iter()) {
+            // iterate over all labels
+            for tag in (0..T::cardinality()).map(T::decode) {
+                sample(global, s.counters.vec.get_metric(tag), tag);
+            }
+        }
+
+        sample(global, &self.default_counters, self.default_tag);
+
+        global.collect_group_into(enc)
+    }
+}
+
+fn sample<T: FixedCardinalityLabel + LabelGroup>(
+    global: &AllocCounter<T>,
+    local: &MeasuredCounterPairState,
+    tag: T,
+) {
+    // load and reset the counts in the thread-local counters.
+    let inc = local.inc.count.swap(0, Relaxed);
+    let dec = local.dec.count.swap(0, Relaxed);
+
+    // add the counts into the global counters.
+    let m = global.vec.get_metric(tag);
+    m.inc.count.fetch_add(inc, Relaxed);
+    m.dec.count.fetch_add(dec, Relaxed);
+}
+
+#[cfg(test)]
+mod tests {
+    use std::alloc::{GlobalAlloc, Layout, System};
+
+    use measured::{FixedCardinalityLabel, MetricGroup, text::BufferedTextEncoder};
+
+    use crate::TrackedAllocator;
+
+    #[derive(FixedCardinalityLabel, Clone, Copy, Debug)]
+    #[label(singleton = "memory_context")]
+    pub enum MemoryContext {
+        Root,
+        Test,
+    }
+
+    #[test]
+    fn alloc() {
+        // Safety: `MemoryContext` upholds the safety requirements.
+        static GLOBAL: TrackedAllocator<System, MemoryContext> =
+            unsafe { TrackedAllocator::new(System, MemoryContext::Root) };
+
+        GLOBAL.register_thread();
+
+        let _test = GLOBAL.scope(MemoryContext::Test);
+
+        let ptr = unsafe { GLOBAL.alloc(Layout::for_value(&[0_i32])) };
+        let ptr = unsafe { GLOBAL.realloc(ptr, Layout::for_value(&[0_i32]), 8) };
+
+        drop(_test);
+
+        let ptr = unsafe { GLOBAL.realloc(ptr, Layout::for_value(&[0_i32, 1_i32]), 4) };
+        unsafe { GLOBAL.dealloc(ptr, Layout::for_value(&[0_i32])) };
+
+        let mut text = BufferedTextEncoder::new();
+        GLOBAL.collect_group_into(&mut text).unwrap();
+        let text = String::from_utf8(text.finish().into()).unwrap();
+        assert_eq!(
+            text,
+            r#"# HELP deallocated_bytes total number of bytes deallocated
+# TYPE deallocated_bytes counter
+deallocated_bytes{memory_context="root"} 4
+deallocated_bytes{memory_context="test"} 8
+
+# HELP allocated_bytes total number of bytes allocated
+# TYPE allocated_bytes counter
+allocated_bytes{memory_context="root"} 4
+allocated_bytes{memory_context="test"} 8
+"#
+        );
+    }
+
+    #[test]
+    fn unregistered_thread() {
+        // Safety: `MemoryContext` upholds the safety requirements.
+        static GLOBAL: TrackedAllocator<System, MemoryContext> =
+            unsafe { TrackedAllocator::new(System, MemoryContext::Root) };
+
+        GLOBAL.register_thread();
+
+        // unregistered thread
+        std::thread::spawn(|| {
+            let ptr = unsafe { GLOBAL.alloc(Layout::for_value(&[0_i32])) };
+            unsafe { GLOBAL.dealloc(ptr, Layout::for_value(&[0_i32])) };
+        })
+        .join()
+        .unwrap();
+
+        let mut text = BufferedTextEncoder::new();
+        GLOBAL.collect_group_into(&mut text).unwrap();
+        let text = String::from_utf8(text.finish().into()).unwrap();
+        assert_eq!(
+            text,
+            r#"# HELP deallocated_bytes total number of bytes deallocated
+# TYPE deallocated_bytes counter
+deallocated_bytes{memory_context="root"} 4
+deallocated_bytes{memory_context="test"} 0
+
+# HELP allocated_bytes total number of bytes allocated
+# TYPE allocated_bytes counter
+allocated_bytes{memory_context="root"} 4
+allocated_bytes{memory_context="test"} 0
+"#
+        );
+    }
+
+    #[test]
+    fn fully_unregistered() {
+        // Safety: `MemoryContext` upholds the safety requirements.
+        static GLOBAL: TrackedAllocator<System, MemoryContext> =
+            unsafe { TrackedAllocator::new(System, MemoryContext::Root) };
+
+        let ptr = unsafe { GLOBAL.alloc(Layout::for_value(&[0_i32])) };
+        unsafe { GLOBAL.dealloc(ptr, Layout::for_value(&[0_i32])) };
+
+        let mut text = BufferedTextEncoder::new();
+        GLOBAL.collect_group_into(&mut text).unwrap();
+        let text = String::from_utf8(text.finish().into()).unwrap();
+        assert_eq!(
+            text,
+            r#"# HELP deallocated_bytes total number of bytes deallocated
+# TYPE deallocated_bytes counter
+deallocated_bytes{memory_context="root"} 4
+deallocated_bytes{memory_context="test"} 0
+
+# HELP allocated_bytes total number of bytes allocated
+# TYPE allocated_bytes counter
+allocated_bytes{memory_context="root"} 4
+allocated_bytes{memory_context="test"} 0
+"#
+        );
+    }
+}

libs/alloc-metrics/src/metric_vec.rs

@@ -0,0 +1,72 @@
+//! Dense metric vec
+
+use measured::{
+    FixedCardinalityLabel, LabelGroup,
+    label::StaticLabelSet,
+    metric::{
+        MetricEncoding, MetricFamilyEncoding, MetricType, group::Encoding, name::MetricNameEncoder,
+    },
+};
+
+pub struct DenseMetricVec<M: MetricType, L: FixedCardinalityLabel + LabelGroup> {
+    metrics: Box<[M]>,
+    metadata: M::Metadata,
+    _label_set: StaticLabelSet<L>,
+}
+
+fn new_dense<M: MetricType>(c: usize) -> Box<[M]> {
+    let mut vec = Vec::with_capacity(c);
+    vec.resize_with(c, M::default);
+    vec.into_boxed_slice()
+}
+
+impl<M: MetricType, L: FixedCardinalityLabel + LabelGroup> DenseMetricVec<M, L>
+where
+    M::Metadata: Default,
+{
+    /// Create a new metric vec with the given label set and metric metadata
+    pub fn new() -> Self {
+        Self::with_metadata(<M::Metadata>::default())
+    }
+}
+
+impl<M: MetricType, L: FixedCardinalityLabel + LabelGroup> DenseMetricVec<M, L> {
+    /// Create a new metric vec with the given label set and metric metadata
+    pub fn with_metadata(metadata: M::Metadata) -> Self {
+        Self {
+            metrics: new_dense(L::cardinality()),
+            metadata,
+            _label_set: StaticLabelSet::new(),
+        }
+    }
+
+    /// Get the individual metric at the given identifier.
+    ///
+    /// # Panics
+    /// Can panic or cause strange behaviour if the label ID comes from a different metric family.
+    pub fn get_metric(&self, label: L) -> &M {
+        // safety: The caller has guarantees that the label encoding is valid.
+        unsafe { self.metrics.get_unchecked(label.encode()) }
+    }
+
+    /// Get the individual metric at the given identifier.
+    ///
+    /// # Panics
+    /// Can panic or cause strange behaviour if the label ID comes from a different metric family.
+    pub fn get_metric_mut(&mut self, label: L) -> &mut M {
+        // safety: The caller has guarantees that the label encoding is valid.
+        unsafe { self.metrics.get_unchecked_mut(label.encode()) }
+    }
+}
+
+impl<M: MetricEncoding<T>, L: FixedCardinalityLabel + LabelGroup, T: Encoding>
+    MetricFamilyEncoding<T> for DenseMetricVec<M, L>
+{
+    fn collect_family_into(&self, name: impl MetricNameEncoder, enc: &mut T) -> Result<(), T::Err> {
+        M::write_type(&name, enc)?;
+        for (index, value) in self.metrics.iter().enumerate() {
+            value.collect_into(&self.metadata, L::decode(index), &name, enc)?;
+        }
+        Ok(())
+    }
+}

libs/compute_api/src/responses.rs

@@ -46,16 +46,33 @@ pub struct ExtensionInstallResponse {
     pub version: ExtVersion,
 }
 
+/// Status of the LFC prewarm process. The same state machine is reused for
+/// both autoprewarm (prewarm after compute/Postgres start using the previously
+/// stored LFC state) and explicit prewarming via API.
 #[derive(Serialize, Default, Debug, Clone, PartialEq)]
 #[serde(tag = "status", rename_all = "snake_case")]
 pub enum LfcPrewarmState {
+    /// Default value when compute boots up.
     #[default]
     NotPrewarmed,
+    /// Prewarming thread is active and loading pages into LFC.
     Prewarming,
+    /// We found requested LFC state in the endpoint storage and
+    /// completed prewarming successfully.
     Completed,
-    Failed {
-        error: String,
-    },
+    /// Unexpected error happened during prewarming. Note, `Not Found 404`
+    /// response from the endpoint storage is explicitly excluded here
+    /// because it can normally happen on the first compute start,
+    /// since LFC state is not available yet.
+    Failed { error: String },
+    /// We tried to fetch the corresponding LFC state from the endpoint storage,
+    /// but received `Not Found 404`. This should normally happen only during the
+    /// first endpoint start after creation with `autoprewarm: true`.
+    ///
+    /// During the orchestrated prewarm via API, when a caller explicitly
+    /// provides the LFC state key to prewarm from, it's the caller responsibility
+    /// to handle this status as an error state in this case.
+    Skipped,
 }
 
 impl Display for LfcPrewarmState {
@@ -64,6 +81,7 @@ impl Display for LfcPrewarmState {
             LfcPrewarmState::NotPrewarmed => f.write_str("NotPrewarmed"),
             LfcPrewarmState::Prewarming => f.write_str("Prewarming"),
             LfcPrewarmState::Completed => f.write_str("Completed"),
+            LfcPrewarmState::Skipped => f.write_str("Skipped"),
             LfcPrewarmState::Failed { error } => write!(f, "Error({error})"),
         }
     }

libs/metrics/src/lib.rs

@@ -4,12 +4,14 @@
 //! a default registry.
 #![deny(clippy::undocumented_unsafe_blocks)]
 
+use std::sync::RwLock;
+
 use measured::label::{LabelGroupSet, LabelGroupVisitor, LabelName, NoLabels};
 use measured::metric::counter::CounterState;
 use measured::metric::gauge::GaugeState;
 use measured::metric::group::Encoding;
 use measured::metric::name::{MetricName, MetricNameEncoder};
-use measured::metric::{MetricEncoding, MetricFamilyEncoding};
+use measured::metric::{MetricEncoding, MetricFamilyEncoding, MetricType};
 use measured::{FixedCardinalityLabel, LabelGroup, MetricGroup};
 use once_cell::sync::Lazy;
 use prometheus::Registry;
@@ -116,12 +118,52 @@ pub fn pow2_buckets(start: usize, end: usize) -> Vec<f64> {
         .collect()
 }
 
+pub struct InfoMetric<L: LabelGroup, M: MetricType = GaugeState> {
+    label: RwLock<L>,
+    metric: M,
+}
+
+impl<L: LabelGroup> InfoMetric<L> {
+    pub fn new(label: L) -> Self {
+        Self::with_metric(label, GaugeState::new(1))
+    }
+}
+
+impl<L: LabelGroup, M: MetricType<Metadata = ()>> InfoMetric<L, M> {
+    pub fn with_metric(label: L, metric: M) -> Self {
+        Self {
+            label: RwLock::new(label),
+            metric,
+        }
+    }
+
+    pub fn set_label(&self, label: L) {
+        *self.label.write().unwrap() = label;
+    }
+}
+
+impl<L, M, E> MetricFamilyEncoding<E> for InfoMetric<L, M>
+where
+    L: LabelGroup,
+    M: MetricEncoding<E, Metadata = ()>,
+    E: Encoding,
+{
+    fn collect_family_into(
+        &self,
+        name: impl measured::metric::name::MetricNameEncoder,
+        enc: &mut E,
+    ) -> Result<(), E::Err> {
+        M::write_type(&name, enc)?;
+        self.metric
+            .collect_into(&(), &*self.label.read().unwrap(), name, enc)
+    }
+}
+
 pub struct BuildInfo {
     pub revision: &'static str,
     pub build_tag: &'static str,
 }
 
-// todo: allow label group without the set
 impl LabelGroup for BuildInfo {
     fn visit_values(&self, v: &mut impl LabelGroupVisitor) {
         const REVISION: &LabelName = LabelName::from_str("revision");
@@ -131,24 +173,6 @@ impl LabelGroup for BuildInfo {
     }
 }
 
-impl<T: Encoding> MetricFamilyEncoding<T> for BuildInfo
-where
-    GaugeState: MetricEncoding<T>,
-{
-    fn collect_family_into(
-        &self,
-        name: impl measured::metric::name::MetricNameEncoder,
-        enc: &mut T,
-    ) -> Result<(), T::Err> {
-        enc.write_help(&name, "Build/version information")?;
-        GaugeState::write_type(&name, enc)?;
-        GaugeState {
-            count: std::sync::atomic::AtomicI64::new(1),
-        }
-        .collect_into(&(), self, name, enc)
-    }
-}
-
 #[derive(MetricGroup)]
 #[metric(new(build_info: BuildInfo))]
 pub struct NeonMetrics {
@@ -165,8 +189,8 @@ pub struct NeonMetrics {
 #[derive(MetricGroup)]
 #[metric(new(build_info: BuildInfo))]
 pub struct LibMetrics {
-    #[metric(init = build_info)]
-    build_info: BuildInfo,
+    #[metric(init = InfoMetric::new(build_info))]
+    build_info: InfoMetric<BuildInfo>,
 
     #[metric(flatten)]
     rusage: Rusage,
@@ -454,7 +478,7 @@ pub trait CounterPairAssoc {
 }
 
 pub struct CounterPairVec<A: CounterPairAssoc> {
-    vec: measured::metric::MetricVec<MeasuredCounterPairState, A::LabelGroupSet>,
+    pub vec: measured::metric::MetricVec<MeasuredCounterPairState, A::LabelGroupSet>,
 }
 
 impl<A: CounterPairAssoc> Default for CounterPairVec<A>
@@ -468,6 +492,17 @@ where
     }
 }
 
+impl<A: CounterPairAssoc> CounterPairVec<A>
+where
+    A::LabelGroupSet: Default,
+{
+    pub fn dense() -> Self {
+        Self {
+            vec: measured::metric::MetricVec::dense(),
+        }
+    }
+}
+
 impl<A: CounterPairAssoc> CounterPairVec<A> {
     pub fn guard(
         &self,
@@ -477,14 +512,31 @@ impl<A: CounterPairAssoc> CounterPairVec<A> {
         self.vec.get_metric(id).inc.inc();
         MeasuredCounterPairGuard { vec: &self.vec, id }
     }
+
+    #[inline]
     pub fn inc(&self, labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>) {
         let id = self.vec.with_labels(labels);
         self.vec.get_metric(id).inc.inc();
     }
+
+    #[inline]
     pub fn dec(&self, labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>) {
         let id = self.vec.with_labels(labels);
         self.vec.get_metric(id).dec.inc();
     }
+
+    #[inline]
+    pub fn inc_by(&self, labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>, x: u64) {
+        let id = self.vec.with_labels(labels);
+        self.vec.get_metric(id).inc.inc_by(x);
+    }
+
+    #[inline]
+    pub fn dec_by(&self, labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>, x: u64) {
+        let id = self.vec.with_labels(labels);
+        self.vec.get_metric(id).dec.inc_by(x);
+    }
+
     pub fn remove_metric(
         &self,
         labels: <A::LabelGroupSet as LabelGroupSet>::Group<'_>,
@@ -529,6 +581,28 @@ pub struct MeasuredCounterPairState {
     pub dec: CounterState,
 }
 
+impl MeasuredCounterPairState {
+    #[inline]
+    pub fn inc(&self) {
+        self.inc.inc();
+    }
+
+    #[inline]
+    pub fn dec(&self) {
+        self.dec.inc();
+    }
+
+    #[inline]
+    pub fn inc_by(&self, x: u64) {
+        self.inc.inc_by(x);
+    }
+
+    #[inline]
+    pub fn dec_by(&self, x: u64) {
+        self.dec.inc_by(x);
+    }
+}
+
 impl measured::metric::MetricType for MeasuredCounterPairState {
     type Metadata = ();
 }
@@ -545,9 +619,9 @@ impl<A: CounterPairAssoc> Drop for MeasuredCounterPairGuard<'_, A> {
 }
 
 /// [`MetricEncoding`] for [`MeasuredCounterPairState`] that only writes the inc counter to the inner encoder.
-struct Inc<T>(T);
+pub struct Inc<T>(pub T);
 /// [`MetricEncoding`] for [`MeasuredCounterPairState`] that only writes the dec counter to the inner encoder.
-struct Dec<T>(T);
+pub struct Dec<T>(pub T);
 
 impl<T: Encoding> Encoding for Inc<T> {
     type Err = T::Err;

libs/neon-shmem/Cargo.toml

@@ -6,8 +6,15 @@ license.workspace = true
 
 [dependencies]
 thiserror.workspace = true
-nix.workspace = true
+nix.workspace=true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
+libc.workspace = true
+lock_api.workspace = true
+rustc-hash.workspace = true
 
 [target.'cfg(target_os = "macos")'.dependencies]
-tempfile = "3.20.0"
+tempfile = "3.14.0"
+
+[dev-dependencies]
+rand = "0.9"
+rand_distr = "0.5.1"

libs/neon-shmem/src/hash.rs

@@ -0,0 +1,583 @@
+//! Resizable hash table implementation on top of byte-level storage (either a [`ShmemHandle`] or a fixed byte array).
+//!
+//! This hash table has two major components: the bucket array and the dictionary. Each bucket within the
+//! bucket array contains a `Option<(K, V)>` and an index of another bucket. In this way there is both an
+//! implicit freelist within the bucket array (`None` buckets point to other `None` entries) and various hash
+//! chains within the bucket array (a Some bucket will point to other Some buckets that had the same hash).
+//!
+//! Buckets are never moved unless they are within a region that is being shrunk, and so the actual hash-
+//! dependent component is done with the dictionary. When a new key is inserted into the map, a position
+//! within the dictionary is decided based on its hash, the data is inserted into an empty bucket based
+//! off of the freelist, and then the index of said bucket is placed in the dictionary.
+//!
+//! This map is resizable (if initialized on top of a [`ShmemHandle`]). Both growing and shrinking happen
+//! in-place and are at a high level achieved by expanding/reducing the bucket array and rebuilding the
+//! dictionary by rehashing all keys.
+//!
+//! Concurrency is managed very simply: the entire map is guarded by one shared-memory RwLock.
+
+use std::hash::{BuildHasher, Hash};
+use std::mem::MaybeUninit;
+
+use crate::shmem::ShmemHandle;
+use crate::{shmem, sync::*};
+
+mod core;
+pub mod entry;
+
+#[cfg(test)]
+mod tests;
+
+use core::{Bucket, CoreHashMap, INVALID_POS};
+use entry::{Entry, OccupiedEntry, PrevPos, VacantEntry};
+
+use thiserror::Error;
+
+/// Error type for a hashmap shrink operation.
+#[derive(Error, Debug)]
+pub enum HashMapShrinkError {
+    /// There was an error encountered while resizing the memory area.
+    #[error("shmem resize failed: {0}")]
+    ResizeError(shmem::Error),
+    /// Occupied entries in to-be-shrunk space were encountered beginning at the given index.
+    #[error("occupied entry in deallocated space found at {0}")]
+    RemainingEntries(usize),
+}
+
+/// This represents a hash table that (possibly) lives in shared memory.
+/// If a new process is launched with fork(), the child process inherits
+/// this struct.
+#[must_use]
+pub struct HashMapInit<'a, K, V, S = rustc_hash::FxBuildHasher> {
+    shmem_handle: Option<ShmemHandle>,
+    shared_ptr: *mut HashMapShared<'a, K, V>,
+    shared_size: usize,
+    hasher: S,
+    num_buckets: u32,
+}
+
+/// This is a per-process handle to a hash table that (possibly) lives in shared memory.
+/// If a child process is launched with fork(), the child process should
+/// get its own HashMapAccess by calling HashMapInit::attach_writer/reader().
+///
+/// XXX: We're not making use of it at the moment, but this struct could
+/// hold process-local information in the future.
+pub struct HashMapAccess<'a, K, V, S = rustc_hash::FxBuildHasher> {
+    shmem_handle: Option<ShmemHandle>,
+    shared_ptr: *mut HashMapShared<'a, K, V>,
+    hasher: S,
+}
+
+unsafe impl<K: Sync, V: Sync, S> Sync for HashMapAccess<'_, K, V, S> {}
+unsafe impl<K: Send, V: Send, S> Send for HashMapAccess<'_, K, V, S> {}
+
+impl<'a, K: Clone + Hash + Eq, V, S> HashMapInit<'a, K, V, S> {
+    /// Change the 'hasher' used by the hash table.
+    ///
+    /// NOTE: This must be called right after creating the hash table,
+    /// before inserting any entries and before calling attach_writer/reader.
+    /// Otherwise different accessors could be using different hash function,
+    /// with confusing results.
+    pub fn with_hasher<T: BuildHasher>(self, hasher: T) -> HashMapInit<'a, K, V, T> {
+        HashMapInit {
+            hasher,
+            shmem_handle: self.shmem_handle,
+            shared_ptr: self.shared_ptr,
+            shared_size: self.shared_size,
+            num_buckets: self.num_buckets,
+        }
+    }
+
+    /// Loosely (over)estimate the size needed to store a hash table with `num_buckets` buckets.
+    pub fn estimate_size(num_buckets: u32) -> usize {
+        // add some margin to cover alignment etc.
+        CoreHashMap::<K, V>::estimate_size(num_buckets) + size_of::<HashMapShared<K, V>>() + 1000
+    }
+
+    fn new(
+        num_buckets: u32,
+        shmem_handle: Option<ShmemHandle>,
+        area_ptr: *mut u8,
+        area_size: usize,
+        hasher: S,
+    ) -> Self {
+        let mut ptr: *mut u8 = area_ptr;
+        let end_ptr: *mut u8 = unsafe { ptr.add(area_size) };
+
+        // carve out area for the One Big Lock (TM) and the HashMapShared.
+        ptr = unsafe { ptr.add(ptr.align_offset(align_of::<libc::pthread_rwlock_t>())) };
+        let raw_lock_ptr = ptr;
+        ptr = unsafe { ptr.add(size_of::<libc::pthread_rwlock_t>()) };
+        ptr = unsafe { ptr.add(ptr.align_offset(align_of::<HashMapShared<K, V>>())) };
+        let shared_ptr: *mut HashMapShared<K, V> = ptr.cast();
+        ptr = unsafe { ptr.add(size_of::<HashMapShared<K, V>>()) };
+
+        // carve out the buckets
+        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<core::Bucket<K, V>>())) };
+        let buckets_ptr = ptr;
+        ptr = unsafe { ptr.add(size_of::<core::Bucket<K, V>>() * num_buckets as usize) };
+
+        // use remaining space for the dictionary
+        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<u32>())) };
+        assert!(ptr.addr() < end_ptr.addr());
+        let dictionary_ptr = ptr;
+        let dictionary_size = unsafe { end_ptr.byte_offset_from(ptr) / size_of::<u32>() as isize };
+        assert!(dictionary_size > 0);
+
+        let buckets =
+            unsafe { std::slice::from_raw_parts_mut(buckets_ptr.cast(), num_buckets as usize) };
+        let dictionary = unsafe {
+            std::slice::from_raw_parts_mut(dictionary_ptr.cast(), dictionary_size as usize)
+        };
+
+        let hashmap = CoreHashMap::new(buckets, dictionary);
+        unsafe {
+            let lock = RwLock::from_raw(PthreadRwLock::new(raw_lock_ptr.cast()), hashmap);
+            std::ptr::write(shared_ptr, lock);
+        }
+
+        Self {
+            num_buckets,
+            shmem_handle,
+            shared_ptr,
+            shared_size: area_size,
+            hasher,
+        }
+    }
+
+    /// Attach to a hash table for writing.
+    pub fn attach_writer(self) -> HashMapAccess<'a, K, V, S> {
+        HashMapAccess {
+            shmem_handle: self.shmem_handle,
+            shared_ptr: self.shared_ptr,
+            hasher: self.hasher,
+        }
+    }
+
+    /// Initialize a table for reading. Currently identical to [`HashMapInit::attach_writer`].
+    ///
+    /// This is a holdover from a previous implementation and is being kept around for
+    /// backwards compatibility reasons.
+    pub fn attach_reader(self) -> HashMapAccess<'a, K, V, S> {
+        self.attach_writer()
+    }
+}
+
+/// Hash table data that is actually stored in the shared memory area.
+///
+/// NOTE: We carve out the parts from a contiguous chunk. Growing and shrinking the hash table
+/// relies on the memory layout! The data structures are laid out in the contiguous shared memory
+/// area as follows:
+///
+/// [`libc::pthread_rwlock_t`]
+/// [`HashMapShared`]
+/// buckets
+/// dictionary
+///
+/// In between the above parts, there can be padding bytes to align the parts correctly.
+type HashMapShared<'a, K, V> = RwLock<CoreHashMap<'a, K, V>>;
+
+impl<'a, K, V> HashMapInit<'a, K, V, rustc_hash::FxBuildHasher>
+where
+    K: Clone + Hash + Eq,
+{
+    /// Place the hash table within a user-supplied fixed memory area.
+    pub fn with_fixed(num_buckets: u32, area: &'a mut [MaybeUninit<u8>]) -> Self {
+        Self::new(
+            num_buckets,
+            None,
+            area.as_mut_ptr().cast(),
+            area.len(),
+            rustc_hash::FxBuildHasher,
+        )
+    }
+
+    /// Place a new hash map in the given shared memory area
+    ///
+    /// # Panics
+    /// Will panic on failure to resize area to expected map size.
+    pub fn with_shmem(num_buckets: u32, shmem: ShmemHandle) -> Self {
+        let size = Self::estimate_size(num_buckets);
+        shmem
+            .set_size(size)
+            .expect("could not resize shared memory area");
+        let ptr = shmem.data_ptr.as_ptr().cast();
+        Self::new(
+            num_buckets,
+            Some(shmem),
+            ptr,
+            size,
+            rustc_hash::FxBuildHasher,
+        )
+    }
+
+    /// Make a resizable hash map within a new shared memory area with the given name.
+    pub fn new_resizeable_named(num_buckets: u32, max_buckets: u32, name: &str) -> Self {
+        let size = Self::estimate_size(num_buckets);
+        let max_size = Self::estimate_size(max_buckets);
+        let shmem =
+            ShmemHandle::new(name, size, max_size).expect("failed to make shared memory area");
+        let ptr = shmem.data_ptr.as_ptr().cast();
+
+        Self::new(
+            num_buckets,
+            Some(shmem),
+            ptr,
+            size,
+            rustc_hash::FxBuildHasher,
+        )
+    }
+
+    /// Make a resizable hash map within a new anonymous shared memory area.
+    pub fn new_resizeable(num_buckets: u32, max_buckets: u32) -> Self {
+        use std::sync::atomic::{AtomicUsize, Ordering};
+        static COUNTER: AtomicUsize = AtomicUsize::new(0);
+        let val = COUNTER.fetch_add(1, Ordering::Relaxed);
+        let name = format!("neon_shmem_hmap{val}");
+        Self::new_resizeable_named(num_buckets, max_buckets, &name)
+    }
+}
+
+impl<'a, K, V, S: BuildHasher> HashMapAccess<'a, K, V, S>
+where
+    K: Clone + Hash + Eq,
+{
+    /// Hash a key using the map's hasher.
+    #[inline]
+    fn get_hash_value(&self, key: &K) -> u64 {
+        self.hasher.hash_one(key)
+    }
+
+    fn entry_with_hash(&self, key: K, hash: u64) -> Entry<'a, '_, K, V> {
+        let mut map = unsafe { self.shared_ptr.as_ref() }.unwrap().write();
+        let dict_pos = hash as usize % map.dictionary.len();
+        let first = map.dictionary[dict_pos];
+        if first == INVALID_POS {
+            // no existing entry
+            return Entry::Vacant(VacantEntry {
+                map,
+                key,
+                dict_pos: dict_pos as u32,
+            });
+        }
+
+        let mut prev_pos = PrevPos::First(dict_pos as u32);
+        let mut next = first;
+        loop {
+            let bucket = &mut map.buckets[next as usize];
+            let (bucket_key, _bucket_value) = bucket.inner.as_mut().expect("entry is in use");
+            if *bucket_key == key {
+                // found existing entry
+                return Entry::Occupied(OccupiedEntry {
+                    map,
+                    _key: key,
+                    prev_pos,
+                    bucket_pos: next,
+                });
+            }
+
+            if bucket.next == INVALID_POS {
+                // No existing entry
+                return Entry::Vacant(VacantEntry {
+                    map,
+                    key,
+                    dict_pos: dict_pos as u32,
+                });
+            }
+            prev_pos = PrevPos::Chained(next);
+            next = bucket.next;
+        }
+    }
+
+    /// Get a reference to the corresponding value for a key.
+    pub fn get<'e>(&'e self, key: &K) -> Option<ValueReadGuard<'e, V>> {
+        let hash = self.get_hash_value(key);
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        RwLockReadGuard::try_map(map, |m| m.get_with_hash(key, hash)).ok()
+    }
+
+    /// Get a reference to the entry containing a key.
+    ///
+    /// NB: THis takes a write lock as there's no way to distinguish whether the intention
+    /// is to use the entry for reading or for writing in advance.
+    pub fn entry(&self, key: K) -> Entry<'a, '_, K, V> {
+        let hash = self.get_hash_value(&key);
+        self.entry_with_hash(key, hash)
+    }
+
+    /// Remove a key given its hash. Returns the associated value if it existed.
+    pub fn remove(&self, key: &K) -> Option<V> {
+        let hash = self.get_hash_value(key);
+        match self.entry_with_hash(key.clone(), hash) {
+            Entry::Occupied(e) => Some(e.remove()),
+            Entry::Vacant(_) => None,
+        }
+    }
+
+    /// Insert/update a key. Returns the previous associated value if it existed.
+    ///
+    /// # Errors
+    /// Will return [`core::FullError`] if there is no more space left in the map.
+    pub fn insert(&self, key: K, value: V) -> Result<Option<V>, core::FullError> {
+        let hash = self.get_hash_value(&key);
+        match self.entry_with_hash(key.clone(), hash) {
+            Entry::Occupied(mut e) => Ok(Some(e.insert(value))),
+            Entry::Vacant(e) => {
+                _ = e.insert(value)?;
+                Ok(None)
+            }
+        }
+    }
+
+    /// Optionally return the entry for a bucket at a given index if it exists.
+    ///
+    /// Has more overhead than one would intuitively expect: performs both a clone of the key
+    /// due to the [`OccupiedEntry`] type owning the key and also a hash of the key in order
+    /// to enable repairing the hash chain if the entry is removed.
+    pub fn entry_at_bucket(&self, pos: usize) -> Option<OccupiedEntry<'a, '_, K, V>> {
+        let map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        if pos >= map.buckets.len() {
+            return None;
+        }
+
+        let entry = map.buckets[pos].inner.as_ref();
+        match entry {
+            Some((key, _)) => Some(OccupiedEntry {
+                _key: key.clone(),
+                bucket_pos: pos as u32,
+                prev_pos: entry::PrevPos::Unknown(self.get_hash_value(key)),
+                map,
+            }),
+            _ => None,
+        }
+    }
+
+    /// Returns the number of buckets in the table.
+    pub fn get_num_buckets(&self) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        map.get_num_buckets()
+    }
+
+    /// Return the key and value stored in bucket with given index. This can be used to
+    /// iterate through the hash map.
+    // TODO: An Iterator might be nicer. The communicator's clock algorithm needs to
+    // _slowly_ iterate through all buckets with its clock hand,  without holding a lock.
+    // If we switch to an Iterator, it must not hold the lock.
+    pub fn get_at_bucket(&self, pos: usize) -> Option<ValueReadGuard<(K, V)>> {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        if pos >= map.buckets.len() {
+            return None;
+        }
+        RwLockReadGuard::try_map(map, |m| m.buckets[pos].inner.as_ref()).ok()
+    }
+
+    /// Returns the index of the bucket a given value corresponds to.
+    pub fn get_bucket_for_value(&self, val_ptr: *const V) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+
+        let origin = map.buckets.as_ptr();
+        let idx = (val_ptr as usize - origin as usize) / size_of::<Bucket<K, V>>();
+        assert!(idx < map.buckets.len());
+
+        idx
+    }
+
+    /// Returns the number of occupied buckets in the table.
+    pub fn get_num_buckets_in_use(&self) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap().read();
+        map.buckets_in_use as usize
+    }
+
+    /// Clears all entries in a table. Does not reset any shrinking operations.
+    pub fn clear(&self) {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        map.clear();
+    }
+
+    /// Perform an in-place rehash of some region (0..`rehash_buckets`) of the table and reset
+    /// the `buckets` and `dictionary` slices to be as long as `num_buckets`. Resets the freelist
+    /// in the process.
+    fn rehash_dict(
+        &self,
+        inner: &mut CoreHashMap<'a, K, V>,
+        buckets_ptr: *mut core::Bucket<K, V>,
+        end_ptr: *mut u8,
+        num_buckets: u32,
+        rehash_buckets: u32,
+    ) {
+        inner.free_head = INVALID_POS;
+
+        let buckets;
+        let dictionary;
+        unsafe {
+            let buckets_end_ptr = buckets_ptr.add(num_buckets as usize);
+            let dictionary_ptr: *mut u32 = buckets_end_ptr
+                .byte_add(buckets_end_ptr.align_offset(align_of::<u32>()))
+                .cast();
+            let dictionary_size: usize =
+                end_ptr.byte_offset_from(buckets_end_ptr) as usize / size_of::<u32>();
+
+            buckets = std::slice::from_raw_parts_mut(buckets_ptr, num_buckets as usize);
+            dictionary = std::slice::from_raw_parts_mut(dictionary_ptr, dictionary_size);
+        }
+        for e in dictionary.iter_mut() {
+            *e = INVALID_POS;
+        }
+
+        for (i, bucket) in buckets.iter_mut().enumerate().take(rehash_buckets as usize) {
+            if bucket.inner.is_none() {
+                bucket.next = inner.free_head;
+                inner.free_head = i as u32;
+                continue;
+            }
+
+            let hash = self.hasher.hash_one(&bucket.inner.as_ref().unwrap().0);
+            let pos: usize = (hash % dictionary.len() as u64) as usize;
+            bucket.next = dictionary[pos];
+            dictionary[pos] = i as u32;
+        }
+
+        inner.dictionary = dictionary;
+        inner.buckets = buckets;
+    }
+
+    /// Rehash the map without growing or shrinking.
+    pub fn shuffle(&self) {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        let num_buckets = map.get_num_buckets() as u32;
+        let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
+        let end_ptr: *mut u8 = unsafe { self.shared_ptr.byte_add(size_bytes).cast() };
+        let buckets_ptr = map.buckets.as_mut_ptr();
+        self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, num_buckets);
+    }
+
+    /// Grow the number of buckets within the table.
+    ///
+    /// 1. Grows the underlying shared memory area
+    /// 2. Initializes new buckets and overwrites the current dictionary
+    /// 3. Rehashes the dictionary
+    ///
+    /// # Panics
+    /// Panics if called on a map initialized with [`HashMapInit::with_fixed`].
+    ///
+    /// # Errors
+    /// Returns an [`shmem::Error`] if any errors occur resizing the memory region.
+    pub fn grow(&self, num_buckets: u32) -> Result<(), shmem::Error> {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        let old_num_buckets = map.buckets.len() as u32;
+
+        assert!(
+            num_buckets >= old_num_buckets,
+            "grow called with a smaller number of buckets"
+        );
+        if num_buckets == old_num_buckets {
+            return Ok(());
+        }
+        let shmem_handle = self
+            .shmem_handle
+            .as_ref()
+            .expect("grow called on a fixed-size hash table");
+
+        let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
+        shmem_handle.set_size(size_bytes)?;
+        let end_ptr: *mut u8 = unsafe { shmem_handle.data_ptr.as_ptr().add(size_bytes) };
+
+        // Initialize new buckets. The new buckets are linked to the free list.
+        // NB: This overwrites the dictionary!
+        let buckets_ptr = map.buckets.as_mut_ptr();
+        unsafe {
+            for i in old_num_buckets..num_buckets {
+                let bucket = buckets_ptr.add(i as usize);
+                bucket.write(core::Bucket {
+                    next: if i < num_buckets - 1 {
+                        i + 1
+                    } else {
+                        map.free_head
+                    },
+                    inner: None,
+                });
+            }
+        }
+
+        self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, old_num_buckets);
+        map.free_head = old_num_buckets;
+
+        Ok(())
+    }
+
+    /// Begin a shrink, limiting all new allocations to be in buckets with index below `num_buckets`.
+    ///
+    /// # Panics
+    /// Panics if called on a map initialized with [`HashMapInit::with_fixed`] or if `num_buckets` is
+    /// greater than the number of buckets in the map.
+    pub fn begin_shrink(&mut self, num_buckets: u32) {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        assert!(
+            num_buckets <= map.get_num_buckets() as u32,
+            "shrink called with a larger number of buckets"
+        );
+        _ = self
+            .shmem_handle
+            .as_ref()
+            .expect("shrink called on a fixed-size hash table");
+        map.alloc_limit = num_buckets;
+    }
+
+    /// If a shrink operation is underway, returns the target size of the map. Otherwise, returns None.
+    pub fn shrink_goal(&self) -> Option<usize> {
+        let map = unsafe { self.shared_ptr.as_mut() }.unwrap().read();
+        let goal = map.alloc_limit;
+        if goal == INVALID_POS {
+            None
+        } else {
+            Some(goal as usize)
+        }
+    }
+
+    /// Complete a shrink after caller has evicted entries, removing the unused buckets and rehashing.
+    ///
+    /// # Panics
+    /// The following cases result in a panic:
+    /// - Calling this function on a map initialized with [`HashMapInit::with_fixed`].
+    /// - Calling this function on a map when no shrink operation is in progress.
+    pub fn finish_shrink(&self) -> Result<(), HashMapShrinkError> {
+        let mut map = unsafe { self.shared_ptr.as_mut() }.unwrap().write();
+        assert!(
+            map.alloc_limit != INVALID_POS,
+            "called finish_shrink when no shrink is in progress"
+        );
+
+        let num_buckets = map.alloc_limit;
+
+        if map.get_num_buckets() == num_buckets as usize {
+            return Ok(());
+        }
+
+        assert!(
+            map.buckets_in_use <= num_buckets,
+            "called finish_shrink before enough entries were removed"
+        );
+
+        for i in (num_buckets as usize)..map.buckets.len() {
+            if map.buckets[i].inner.is_some() {
+                return Err(HashMapShrinkError::RemainingEntries(i));
+            }
+        }
+
+        let shmem_handle = self
+            .shmem_handle
+            .as_ref()
+            .expect("shrink called on a fixed-size hash table");
+
+        let size_bytes = HashMapInit::<K, V, S>::estimate_size(num_buckets);
+        if let Err(e) = shmem_handle.set_size(size_bytes) {
+            return Err(HashMapShrinkError::ResizeError(e));
+        }
+        let end_ptr: *mut u8 = unsafe { shmem_handle.data_ptr.as_ptr().add(size_bytes) };
+        let buckets_ptr = map.buckets.as_mut_ptr();
+        self.rehash_dict(&mut map, buckets_ptr, end_ptr, num_buckets, num_buckets);
+        map.alloc_limit = INVALID_POS;
+
+        Ok(())
+    }
+}

libs/neon-shmem/src/hash/core.rs

@@ -0,0 +1,174 @@
+//! Simple hash table with chaining.
+
+use std::hash::Hash;
+use std::mem::MaybeUninit;
+
+use crate::hash::entry::*;
+
+/// Invalid position within the map (either within the dictionary or bucket array).
+pub(crate) const INVALID_POS: u32 = u32::MAX;
+
+/// Fundamental storage unit within the hash table. Either empty or contains a key-value pair.
+/// Always part of a chain of some kind (either a freelist if empty or a hash chain if full).
+pub(crate) struct Bucket<K, V> {
+    /// Index of next bucket in the chain.
+    pub(crate) next: u32,
+    /// Key-value pair contained within bucket.
+    pub(crate) inner: Option<(K, V)>,
+}
+
+/// Core hash table implementation.
+pub(crate) struct CoreHashMap<'a, K, V> {
+    /// Dictionary used to map hashes to bucket indices.
+    pub(crate) dictionary: &'a mut [u32],
+    /// Buckets containing key-value pairs.
+    pub(crate) buckets: &'a mut [Bucket<K, V>],
+    /// Head of the freelist.
+    pub(crate) free_head: u32,
+    /// Maximum index of a bucket allowed to be allocated. [`INVALID_POS`] if no limit.
+    pub(crate) alloc_limit: u32,
+    /// The number of currently occupied buckets.
+    pub(crate) buckets_in_use: u32,
+}
+
+/// Error for when there are no empty buckets left but one is needed.
+#[derive(Debug, PartialEq)]
+pub struct FullError;
+
+impl<'a, K: Clone + Hash + Eq, V> CoreHashMap<'a, K, V> {
+    const FILL_FACTOR: f32 = 0.60;
+
+    /// Estimate the size of data contained within the the hash map.
+    pub fn estimate_size(num_buckets: u32) -> usize {
+        let mut size = 0;
+
+        // buckets
+        size += size_of::<Bucket<K, V>>() * num_buckets as usize;
+
+        // dictionary
+        size += (f32::ceil((size_of::<u32>() * num_buckets as usize) as f32 / Self::FILL_FACTOR))
+            as usize;
+
+        size
+    }
+
+    pub fn new(
+        buckets: &'a mut [MaybeUninit<Bucket<K, V>>],
+        dictionary: &'a mut [MaybeUninit<u32>],
+    ) -> Self {
+        // Initialize the buckets
+        for i in 0..buckets.len() {
+            buckets[i].write(Bucket {
+                next: if i < buckets.len() - 1 {
+                    i as u32 + 1
+                } else {
+                    INVALID_POS
+                },
+                inner: None,
+            });
+        }
+
+        // Initialize the dictionary
+        for e in dictionary.iter_mut() {
+            e.write(INVALID_POS);
+        }
+
+        // TODO: use std::slice::assume_init_mut() once it stabilizes
+        let buckets =
+            unsafe { std::slice::from_raw_parts_mut(buckets.as_mut_ptr().cast(), buckets.len()) };
+        let dictionary = unsafe {
+            std::slice::from_raw_parts_mut(dictionary.as_mut_ptr().cast(), dictionary.len())
+        };
+
+        Self {
+            dictionary,
+            buckets,
+            free_head: 0,
+            buckets_in_use: 0,
+            alloc_limit: INVALID_POS,
+        }
+    }
+
+    /// Get the value associated with a key (if it exists) given its hash.
+    pub fn get_with_hash(&self, key: &K, hash: u64) -> Option<&V> {
+        let mut next = self.dictionary[hash as usize % self.dictionary.len()];
+        loop {
+            if next == INVALID_POS {
+                return None;
+            }
+
+            let bucket = &self.buckets[next as usize];
+            let (bucket_key, bucket_value) = bucket.inner.as_ref().expect("entry is in use");
+            if bucket_key == key {
+                return Some(bucket_value);
+            }
+            next = bucket.next;
+        }
+    }
+
+    /// Get number of buckets in map.
+    pub fn get_num_buckets(&self) -> usize {
+        self.buckets.len()
+    }
+
+    /// Clears all entries from the hashmap.
+    ///
+    /// Does not reset any allocation limits, but does clear any entries beyond them.
+    pub fn clear(&mut self) {
+        for i in 0..self.buckets.len() {
+            self.buckets[i] = Bucket {
+                next: if i < self.buckets.len() - 1 {
+                    i as u32 + 1
+                } else {
+                    INVALID_POS
+                },
+                inner: None,
+            }
+        }
+        for i in 0..self.dictionary.len() {
+            self.dictionary[i] = INVALID_POS;
+        }
+
+        self.free_head = 0;
+        self.buckets_in_use = 0;
+    }
+
+    /// Find the position of an unused bucket via the freelist and initialize it.
+    pub(crate) fn alloc_bucket(&mut self, key: K, value: V) -> Result<u32, FullError> {
+        let mut pos = self.free_head;
+
+        // Find the first bucket we're *allowed* to use.
+        let mut prev = PrevPos::First(self.free_head);
+        while pos != INVALID_POS && pos >= self.alloc_limit {
+            let bucket = &mut self.buckets[pos as usize];
+            prev = PrevPos::Chained(pos);
+            pos = bucket.next;
+        }
+        if pos == INVALID_POS {
+            return Err(FullError);
+        }
+
+        // Repair the freelist.
+        match prev {
+            PrevPos::First(_) => {
+                let next_pos = self.buckets[pos as usize].next;
+                self.free_head = next_pos;
+            }
+            PrevPos::Chained(p) => {
+                if p != INVALID_POS {
+                    let next_pos = self.buckets[pos as usize].next;
+                    self.buckets[p as usize].next = next_pos;
+                }
+            }
+            _ => unreachable!(),
+        }
+
+        // Initialize the bucket.
+        let bucket = &mut self.buckets[pos as usize];
+        self.buckets_in_use += 1;
+        bucket.next = INVALID_POS;
+        bucket.inner = Some((key, value));
+
+        Ok(pos)
+    }
+}

libs/neon-shmem/src/hash/entry.rs

@@ -0,0 +1,130 @@
+//! Equivalent of [`std::collections::hash_map::Entry`] for this hashmap.
+
+use crate::hash::core::{CoreHashMap, FullError, INVALID_POS};
+use crate::sync::{RwLockWriteGuard, ValueWriteGuard};
+
+use std::hash::Hash;
+use std::mem;
+
+pub enum Entry<'a, 'b, K, V> {
+    Occupied(OccupiedEntry<'a, 'b, K, V>),
+    Vacant(VacantEntry<'a, 'b, K, V>),
+}
+
+/// Enum representing the previous position within a chain.
+#[derive(Clone, Copy)]
+pub(crate) enum PrevPos {
+    /// Starting index within the dictionary.  
+    First(u32),
+    /// Regular index within the buckets.
+    Chained(u32),
+    /// Unknown - e.g. the associated entry was retrieved by index instead of chain.
+    Unknown(u64),
+}
+
+pub struct OccupiedEntry<'a, 'b, K, V> {
+    /// Mutable reference to the map containing this entry.
+    pub(crate) map: RwLockWriteGuard<'b, CoreHashMap<'a, K, V>>,
+    /// The key of the occupied entry
+    pub(crate) _key: K,
+    /// The index of the previous entry in the chain.
+    pub(crate) prev_pos: PrevPos,
+    /// The position of the bucket in the [`CoreHashMap`] bucket array.
+    pub(crate) bucket_pos: u32,
+}
+
+impl<K, V> OccupiedEntry<'_, '_, K, V> {
+    pub fn get(&self) -> &V {
+        &self.map.buckets[self.bucket_pos as usize]
+            .inner
+            .as_ref()
+            .unwrap()
+            .1
+    }
+
+    pub fn get_mut(&mut self) -> &mut V {
+        &mut self.map.buckets[self.bucket_pos as usize]
+            .inner
+            .as_mut()
+            .unwrap()
+            .1
+    }
+
+    /// Inserts a value into the entry, replacing (and returning) the existing value.
+    pub fn insert(&mut self, value: V) -> V {
+        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
+        // This assumes inner is Some, which it must be for an OccupiedEntry
+        mem::replace(&mut bucket.inner.as_mut().unwrap().1, value)
+    }
+
+    /// Removes the entry from the hash map, returning the value originally stored within it.
+    ///
+    /// This may result in multiple bucket accesses if the entry was obtained by index as the
+    /// previous chain entry needs to be discovered in this case.
+    pub fn remove(mut self) -> V {
+        // If this bucket was queried by index, go ahead and follow its chain from the start.
+        let prev = if let PrevPos::Unknown(hash) = self.prev_pos {
+            let dict_idx = hash as usize % self.map.dictionary.len();
+            let mut prev = PrevPos::First(dict_idx as u32);
+            let mut curr = self.map.dictionary[dict_idx];
+            while curr != self.bucket_pos {
+                assert!(curr != INVALID_POS);
+                prev = PrevPos::Chained(curr);
+                curr = self.map.buckets[curr as usize].next;
+            }
+            prev
+        } else {
+            self.prev_pos
+        };
+
+        // CoreHashMap::remove returns Option<(K, V)>. We know it's Some for an OccupiedEntry.
+        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
+
+        // unlink it from the chain
+        match prev {
+            PrevPos::First(dict_pos) => {
+                self.map.dictionary[dict_pos as usize] = bucket.next;
+            }
+            PrevPos::Chained(bucket_pos) => {
+                self.map.buckets[bucket_pos as usize].next = bucket.next;
+            }
+            _ => unreachable!(),
+        }
+
+        // and add it to the freelist
+        let free = self.map.free_head;
+        let bucket = &mut self.map.buckets[self.bucket_pos as usize];
+        let old_value = bucket.inner.take();
+        bucket.next = free;
+        self.map.free_head = self.bucket_pos;
+        self.map.buckets_in_use -= 1;
+
+        old_value.unwrap().1
+    }
+}
+
+/// An abstract view into a vacant entry within the map.
+pub struct VacantEntry<'a, 'b, K, V> {
+    /// Mutable reference to the map containing this entry.
+    pub(crate) map: RwLockWriteGuard<'b, CoreHashMap<'a, K, V>>,
+    /// The key to be inserted into this entry.
+    pub(crate) key: K,
+    /// The position within the dictionary corresponding to the key's hash.
+    pub(crate) dict_pos: u32,
+}
+
+impl<'b, K: Clone + Hash + Eq, V> VacantEntry<'_, 'b, K, V> {
+    /// Insert a value into the vacant entry, finding and populating an empty bucket in the process.
+    ///
+    /// # Errors
+    /// Will return [`FullError`] if there are no unoccupied buckets in the map.
+    pub fn insert(mut self, value: V) -> Result<ValueWriteGuard<'b, V>, FullError> {
+        let pos = self.map.alloc_bucket(self.key, value)?;
+        self.map.buckets[pos as usize].next = self.map.dictionary[self.dict_pos as usize];
+        self.map.dictionary[self.dict_pos as usize] = pos;
+
+        Ok(RwLockWriteGuard::map(self.map, |m| {
+            &mut m.buckets[pos as usize].inner.as_mut().unwrap().1
+        }))
+    }
+}

libs/neon-shmem/src/hash/tests.rs

@@ -0,0 +1,428 @@
+use std::collections::BTreeMap;
+use std::collections::HashSet;
+use std::fmt::Debug;
+use std::mem::MaybeUninit;
+
+use crate::hash::Entry;
+use crate::hash::HashMapAccess;
+use crate::hash::HashMapInit;
+use crate::hash::core::FullError;
+
+use rand::seq::SliceRandom;
+use rand::{Rng, RngCore};
+use rand_distr::Zipf;
+
+const TEST_KEY_LEN: usize = 16;
+
+#[derive(Clone, Copy, Debug, Hash, PartialEq, Eq, PartialOrd, Ord)]
+struct TestKey([u8; TEST_KEY_LEN]);
+
+impl From<&TestKey> for u128 {
+    fn from(val: &TestKey) -> u128 {
+        u128::from_be_bytes(val.0)
+    }
+}
+
+impl From<u128> for TestKey {
+    fn from(val: u128) -> TestKey {
+        TestKey(val.to_be_bytes())
+    }
+}
+
+impl<'a> From<&'a [u8]> for TestKey {
+    fn from(bytes: &'a [u8]) -> TestKey {
+        TestKey(bytes.try_into().unwrap())
+    }
+}
+
+fn test_inserts<K: Into<TestKey> + Copy>(keys: &[K]) {
+    let w = HashMapInit::<TestKey, usize>::new_resizeable_named(100000, 120000, "test_inserts")
+        .attach_writer();
+
+    for (idx, k) in keys.iter().enumerate() {
+        let res = w.entry((*k).into());
+        match res {
+            Entry::Occupied(mut e) => {
+                e.insert(idx);
+            }
+            Entry::Vacant(e) => {
+                let res = e.insert(idx);
+                assert!(res.is_ok());
+            }
+        };
+    }
+
+    for (idx, k) in keys.iter().enumerate() {
+        let x = w.get(&(*k).into());
+        let value = x.as_deref().copied();
+        assert_eq!(value, Some(idx));
+    }
+}
+
+#[test]
+fn dense() {
+    // This exercises splitting a node with prefix
+    let keys: &[u128] = &[0, 1, 2, 3, 256];
+    test_inserts(keys);
+
+    // Dense keys
+    let mut keys: Vec<u128> = (0..10000).collect();
+    test_inserts(&keys);
+
+    // Do the same in random orders
+    for _ in 1..10 {
+        keys.shuffle(&mut rand::rng());
+        test_inserts(&keys);
+    }
+}
+
+#[test]
+fn sparse() {
+    // sparse keys
+    let mut keys: Vec<TestKey> = Vec::new();
+    let mut used_keys = HashSet::new();
+    for _ in 0..10000 {
+        loop {
+            let key = rand::random::<u128>();
+            if used_keys.contains(&key) {
+                continue;
+            }
+            used_keys.insert(key);
+            keys.push(key.into());
+            break;
+        }
+    }
+    test_inserts(&keys);
+}
+
+#[derive(Clone, Debug)]
+struct TestOp(TestKey, Option<usize>);
+
+fn apply_op(
+    op: &TestOp,
+    map: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+) {
+    // apply the change to the shadow tree first
+    let shadow_existing = if let Some(v) = op.1 {
+        shadow.insert(op.0, v)
+    } else {
+        shadow.remove(&op.0)
+    };
+
+    let entry = map.entry(op.0);
+    let hash_existing = match op.1 {
+        Some(new) => match entry {
+            Entry::Occupied(mut e) => Some(e.insert(new)),
+            Entry::Vacant(e) => {
+                _ = e.insert(new).unwrap();
+                None
+            }
+        },
+        None => match entry {
+            Entry::Occupied(e) => Some(e.remove()),
+            Entry::Vacant(_) => None,
+        },
+    };
+
+    assert_eq!(shadow_existing, hash_existing);
+}
+
+fn do_random_ops(
+    num_ops: usize,
+    size: u32,
+    del_prob: f64,
+    writer: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+    rng: &mut rand::rngs::ThreadRng,
+) {
+    for i in 0..num_ops {
+        let key: TestKey = ((rng.next_u32() % size) as u128).into();
+        let op = TestOp(
+            key,
+            if rng.random_bool(del_prob) {
+                Some(i)
+            } else {
+                None
+            },
+        );
+        apply_op(&op, writer, shadow);
+    }
+}
+
+fn do_deletes(
+    num_ops: usize,
+    writer: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+) {
+    for _ in 0..num_ops {
+        let (k, _) = shadow.pop_first().unwrap();
+        writer.remove(&k);
+    }
+}
+
+fn do_shrink(
+    writer: &mut HashMapAccess<TestKey, usize>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+    from: u32,
+    to: u32,
+) {
+    assert!(writer.shrink_goal().is_none());
+    writer.begin_shrink(to);
+    assert_eq!(writer.shrink_goal(), Some(to as usize));
+    for i in to..from {
+        if let Some(entry) = writer.entry_at_bucket(i as usize) {
+            shadow.remove(&entry._key);
+            entry.remove();
+        }
+    }
+    let old_usage = writer.get_num_buckets_in_use();
+    writer.finish_shrink().unwrap();
+    assert!(writer.shrink_goal().is_none());
+    assert_eq!(writer.get_num_buckets_in_use(), old_usage);
+}
+
+#[test]
+fn random_ops() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(100000, 120000, "test_random")
+            .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+
+    let distribution = Zipf::new(u128::MAX as f64, 1.1).unwrap();
+    let mut rng = rand::rng();
+    for i in 0..100000 {
+        let key: TestKey = (rng.sample(distribution) as u128).into();
+
+        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
+
+        apply_op(&op, &mut writer, &mut shadow);
+    }
+}
+
+#[test]
+fn test_shuffle() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 1200, "test_shuf")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
+    writer.shuffle();
+    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
+}
+
+#[test]
+fn test_grow() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 2000, "test_grow")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(10000, 1000, 0.75, &mut writer, &mut shadow, &mut rng);
+    let old_usage = writer.get_num_buckets_in_use();
+    writer.grow(1500).unwrap();
+    assert_eq!(writer.get_num_buckets_in_use(), old_usage);
+    assert_eq!(writer.get_num_buckets(), 1500);
+    do_random_ops(10000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+}
+
+#[test]
+fn test_clear() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_clear")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+    do_random_ops(2000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+    writer.clear();
+    assert_eq!(writer.get_num_buckets_in_use(), 0);
+    assert_eq!(writer.get_num_buckets(), 1500);
+    while let Some((key, _)) = shadow.pop_first() {
+        assert!(writer.get(&key).is_none());
+    }
+    do_random_ops(2000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+    for i in 0..(1500 - writer.get_num_buckets_in_use()) {
+        writer.insert((1500 + i as u128).into(), 0).unwrap();
+    }
+    assert_eq!(writer.insert(5000.into(), 0), Err(FullError {}));
+    writer.clear();
+    assert!(writer.insert(5000.into(), 0).is_ok());
+}
+
+#[test]
+fn test_idx_remove() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_clear")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+    do_random_ops(2000, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
+    for _ in 0..100 {
+        let idx = (rng.next_u32() % 1500) as usize;
+        if let Some(e) = writer.entry_at_bucket(idx) {
+            shadow.remove(&e._key);
+            e.remove();
+        }
+    }
+    while let Some((key, val)) = shadow.pop_first() {
+        assert_eq!(*writer.get(&key).unwrap(), val);
+    }
+}
+
+#[test]
+fn test_idx_get() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_clear")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+    do_random_ops(2000, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
+    for _ in 0..100 {
+        let idx = (rng.next_u32() % 1500) as usize;
+        if let Some(pair) = writer.get_at_bucket(idx) {
+            {
+                let v: *const usize = &pair.1;
+                assert_eq!(writer.get_bucket_for_value(v), idx);
+            }
+            {
+                let v: *const usize = &pair.1;
+                assert_eq!(writer.get_bucket_for_value(v), idx);
+            }
+        }
+    }
+}
+
+#[test]
+fn test_shrink() {
+    let mut writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_shrink")
+        .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(10000, 1500, 0.75, &mut writer, &mut shadow, &mut rng);
+    do_shrink(&mut writer, &mut shadow, 1500, 1000);
+    assert_eq!(writer.get_num_buckets(), 1000);
+    do_deletes(500, &mut writer, &mut shadow);
+    do_random_ops(10000, 500, 0.75, &mut writer, &mut shadow, &mut rng);
+    assert!(writer.get_num_buckets_in_use() <= 1000);
+}
+
+#[test]
+fn test_shrink_grow_seq() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 20000, "test_grow_seq")
+            .attach_writer();
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+    let mut rng = rand::rng();
+
+    do_random_ops(500, 1000, 0.1, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Shrinking to 750");
+    do_shrink(&mut writer, &mut shadow, 1000, 750);
+    do_random_ops(200, 1000, 0.5, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Growing to 1500");
+    writer.grow(1500).unwrap();
+    do_random_ops(600, 1500, 0.1, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Shrinking to 200");
+    while shadow.len() > 100 {
+        do_deletes(1, &mut writer, &mut shadow);
+    }
+    do_shrink(&mut writer, &mut shadow, 1500, 200);
+    do_random_ops(50, 1500, 0.25, &mut writer, &mut shadow, &mut rng);
+    eprintln!("Growing to 10k");
+    writer.grow(10000).unwrap();
+    do_random_ops(10000, 5000, 0.25, &mut writer, &mut shadow, &mut rng);
+}
+
+#[test]
+fn test_bucket_ops() {
+    let writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1000, 1200, "test_bucket_ops")
+        .attach_writer();
+    match writer.entry(1.into()) {
+        Entry::Occupied(mut e) => {
+            e.insert(2);
+        }
+        Entry::Vacant(e) => {
+            _ = e.insert(2).unwrap();
+        }
+    }
+    assert_eq!(writer.get_num_buckets_in_use(), 1);
+    assert_eq!(writer.get_num_buckets(), 1000);
+    assert_eq!(*writer.get(&1.into()).unwrap(), 2);
+    let pos = match writer.entry(1.into()) {
+        Entry::Occupied(e) => {
+            assert_eq!(e._key, 1.into());
+            e.bucket_pos as usize
+        }
+        Entry::Vacant(_) => {
+            panic!("Insert didn't affect entry");
+        }
+    };
+    assert_eq!(writer.entry_at_bucket(pos).unwrap()._key, 1.into());
+    assert_eq!(*writer.get_at_bucket(pos).unwrap(), (1.into(), 2));
+    {
+        let ptr: *const usize = &*writer.get(&1.into()).unwrap();
+        assert_eq!(writer.get_bucket_for_value(ptr), pos);
+    }
+    writer.remove(&1.into());
+    assert!(writer.get(&1.into()).is_none());
+}
+
+#[test]
+fn test_shrink_zero() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_shrink_zero")
+            .attach_writer();
+    writer.begin_shrink(0);
+    for i in 0..1500 {
+        writer.entry_at_bucket(i).map(|x| x.remove());
+    }
+    writer.finish_shrink().unwrap();
+    assert_eq!(writer.get_num_buckets_in_use(), 0);
+    let entry = writer.entry(1.into());
+    if let Entry::Vacant(v) = entry {
+        assert!(v.insert(2).is_err());
+    } else {
+        panic!("Somehow got non-vacant entry in empty map.")
+    }
+    writer.grow(50).unwrap();
+    let entry = writer.entry(1.into());
+    if let Entry::Vacant(v) = entry {
+        assert!(v.insert(2).is_ok());
+    } else {
+        panic!("Somehow got non-vacant entry in empty map.")
+    }
+    assert_eq!(writer.get_num_buckets_in_use(), 1);
+}
+
+#[test]
+#[should_panic]
+fn test_grow_oom() {
+    let writer = HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2000, "test_grow_oom")
+        .attach_writer();
+    writer.grow(20000).unwrap();
+}
+
+#[test]
+#[should_panic]
+fn test_shrink_bigger() {
+    let mut writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2500, "test_shrink_bigger")
+            .attach_writer();
+    writer.begin_shrink(2000);
+}
+
+#[test]
+#[should_panic]
+fn test_shrink_early_finish() {
+    let writer =
+        HashMapInit::<TestKey, usize>::new_resizeable_named(1500, 2500, "test_shrink_early_finish")
+            .attach_writer();
+    writer.finish_shrink().unwrap();
+}
+
+#[test]
+#[should_panic]
+fn test_shrink_fixed_size() {
+    let mut area = [MaybeUninit::uninit(); 10000];
+    let init_struct = HashMapInit::<TestKey, usize>::with_fixed(3, &mut area);
+    let mut writer = init_struct.attach_writer();
+    writer.begin_shrink(1);
+}

libs/neon-shmem/src/lib.rs

@@ -1,418 +1,3 @@
-//! Shared memory utilities for neon communicator
-
-use std::num::NonZeroUsize;
-use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
-use std::ptr::NonNull;
-use std::sync::atomic::{AtomicUsize, Ordering};
-
-use nix::errno::Errno;
-use nix::sys::mman::MapFlags;
-use nix::sys::mman::ProtFlags;
-use nix::sys::mman::mmap as nix_mmap;
-use nix::sys::mman::munmap as nix_munmap;
-use nix::unistd::ftruncate as nix_ftruncate;
-
-/// ShmemHandle represents a shared memory area that can be shared by processes over fork().
-/// Unlike shared memory allocated by Postgres, this area is resizable, up to 'max_size' that's
-/// specified at creation.
-///
-/// The area is backed by an anonymous file created with memfd_create(). The full address space for
-/// 'max_size' is reserved up-front with mmap(), but whenever you call [`ShmemHandle::set_size`],
-/// the underlying file is resized. Do not access the area beyond the current size. Currently, that
-/// will cause the file to be expanded, but we might use mprotect() etc. to enforce that in the
-/// future.
-pub struct ShmemHandle {
-    /// memfd file descriptor
-    fd: OwnedFd,
-
-    max_size: usize,
-
-    // Pointer to the beginning of the shared memory area. The header is stored there.
-    shared_ptr: NonNull<SharedStruct>,
-
-    // Pointer to the beginning of the user data
-    pub data_ptr: NonNull<u8>,
-}
-
-/// This is stored at the beginning in the shared memory area.
-struct SharedStruct {
-    max_size: usize,
-
-    /// Current size of the backing file. The high-order bit is used for the RESIZE_IN_PROGRESS flag
-    current_size: AtomicUsize,
-}
-
-const RESIZE_IN_PROGRESS: usize = 1 << 63;
-
-const HEADER_SIZE: usize = std::mem::size_of::<SharedStruct>();
-
-/// Error type returned by the ShmemHandle functions.
-#[derive(thiserror::Error, Debug)]
-#[error("{msg}: {errno}")]
-pub struct Error {
-    pub msg: String,
-    pub errno: Errno,
-}
-
-impl Error {
-    fn new(msg: &str, errno: Errno) -> Error {
-        Error {
-            msg: msg.to_string(),
-            errno,
-        }
-    }
-}
-
-impl ShmemHandle {
-    /// Create a new shared memory area. To communicate between processes, the processes need to be
-    /// fork()'d after calling this, so that the ShmemHandle is inherited by all processes.
-    ///
-    /// If the ShmemHandle is dropped, the memory is unmapped from the current process. Other
-    /// processes can continue using it, however.
-    pub fn new(name: &str, initial_size: usize, max_size: usize) -> Result<ShmemHandle, Error> {
-        // create the backing anonymous file.
-        let fd = create_backing_file(name)?;
-
-        Self::new_with_fd(fd, initial_size, max_size)
-    }
-
-    fn new_with_fd(
-        fd: OwnedFd,
-        initial_size: usize,
-        max_size: usize,
-    ) -> Result<ShmemHandle, Error> {
-        // We reserve the high-order bit for the RESIZE_IN_PROGRESS flag, and the actual size
-        // is a little larger than this because of the SharedStruct header. Make the upper limit
-        // somewhat smaller than that, because with anything close to that, you'll run out of
-        // memory anyway.
-        if max_size >= 1 << 48 {
-            panic!("max size {max_size} too large");
-        }
-        if initial_size > max_size {
-            panic!("initial size {initial_size} larger than max size {max_size}");
-        }
-
-        // The actual initial / max size is the one given by the caller, plus the size of
-        // 'SharedStruct'.
-        let initial_size = HEADER_SIZE + initial_size;
-        let max_size = NonZeroUsize::new(HEADER_SIZE + max_size).unwrap();
-
-        // Reserve address space for it with mmap
-        //
-        // TODO: Use MAP_HUGETLB if possible
-        let start_ptr = unsafe {
-            nix_mmap(
-                None,
-                max_size,
-                ProtFlags::PROT_READ | ProtFlags::PROT_WRITE,
-                MapFlags::MAP_SHARED,
-                &fd,
-                0,
-            )
-        }
-        .map_err(|e| Error::new("mmap failed: {e}", e))?;
-
-        // Reserve space for the initial size
-        enlarge_file(fd.as_fd(), initial_size as u64)?;
-
-        // Initialize the header
-        let shared: NonNull<SharedStruct> = start_ptr.cast();
-        unsafe {
-            shared.write(SharedStruct {
-                max_size: max_size.into(),
-                current_size: AtomicUsize::new(initial_size),
-            })
-        };
-
-        // The user data begins after the header
-        let data_ptr = unsafe { start_ptr.cast().add(HEADER_SIZE) };
-
-        Ok(ShmemHandle {
-            fd,
-            max_size: max_size.into(),
-            shared_ptr: shared,
-            data_ptr,
-        })
-    }
-
-    // return reference to the header
-    fn shared(&self) -> &SharedStruct {
-        unsafe { self.shared_ptr.as_ref() }
-    }
-
-    /// Resize the shared memory area. 'new_size' must not be larger than the 'max_size' specified
-    /// when creating the area.
-    ///
-    /// This may only be called from one process/thread concurrently. We detect that case
-    /// and return an Error.
-    pub fn set_size(&self, new_size: usize) -> Result<(), Error> {
-        let new_size = new_size + HEADER_SIZE;
-        let shared = self.shared();
-
-        if new_size > self.max_size {
-            panic!(
-                "new size ({} is greater than max size ({})",
-                new_size, self.max_size
-            );
-        }
-        assert_eq!(self.max_size, shared.max_size);
-
-        // Lock the area by setting the bit in 'current_size'
-        //
-        // Ordering::Relaxed would probably be sufficient here, as we don't access any other memory
-        // and the posix_fallocate/ftruncate call is surely a synchronization point anyway. But
-        // since this is not performance-critical, better safe than sorry .
-        let mut old_size = shared.current_size.load(Ordering::Acquire);
-        loop {
-            if (old_size & RESIZE_IN_PROGRESS) != 0 {
-                return Err(Error::new(
-                    "concurrent resize detected",
-                    Errno::UnknownErrno,
-                ));
-            }
-            match shared.current_size.compare_exchange(
-                old_size,
-                new_size,
-                Ordering::Acquire,
-                Ordering::Relaxed,
-            ) {
-                Ok(_) => break,
-                Err(x) => old_size = x,
-            }
-        }
-
-        // Ok, we got the lock.
-        //
-        // NB: If anything goes wrong, we *must* clear the bit!
-        let result = {
-            use std::cmp::Ordering::{Equal, Greater, Less};
-            match new_size.cmp(&old_size) {
-                Less => nix_ftruncate(&self.fd, new_size as i64).map_err(|e| {
-                    Error::new("could not shrink shmem segment, ftruncate failed: {e}", e)
-                }),
-                Equal => Ok(()),
-                Greater => enlarge_file(self.fd.as_fd(), new_size as u64),
-            }
-        };
-
-        // Unlock
-        shared.current_size.store(
-            if result.is_ok() { new_size } else { old_size },
-            Ordering::Release,
-        );
-
-        result
-    }
-
-    /// Returns the current user-visible size of the shared memory segment.
-    ///
-    /// NOTE: a concurrent set_size() call can change the size at any time. It is the caller's
-    /// responsibility not to access the area beyond the current size.
-    pub fn current_size(&self) -> usize {
-        let total_current_size =
-            self.shared().current_size.load(Ordering::Relaxed) & !RESIZE_IN_PROGRESS;
-        total_current_size - HEADER_SIZE
-    }
-}
-
-impl Drop for ShmemHandle {
-    fn drop(&mut self) {
-        // SAFETY: The pointer was obtained from mmap() with the given size.
-        // We unmap the entire region.
-        let _ = unsafe { nix_munmap(self.shared_ptr.cast(), self.max_size) };
-        // The fd is dropped automatically by OwnedFd.
-    }
-}
-
-/// Create a "backing file" for the shared memory area. On Linux, use memfd_create(), to create an
-/// anonymous in-memory file. One macos, fall back to a regular file. That's good enough for
-/// development and testing, but in production we want the file to stay in memory.
-///
-/// disable 'unused_variables' warnings, because in the macos path, 'name' is unused.
-#[allow(unused_variables)]
-fn create_backing_file(name: &str) -> Result<OwnedFd, Error> {
-    #[cfg(not(target_os = "macos"))]
-    {
-        nix::sys::memfd::memfd_create(name, nix::sys::memfd::MFdFlags::empty())
-            .map_err(|e| Error::new("memfd_create failed: {e}", e))
-    }
-    #[cfg(target_os = "macos")]
-    {
-        let file = tempfile::tempfile().map_err(|e| {
-            Error::new(
-                "could not create temporary file to back shmem area: {e}",
-                nix::errno::Errno::from_raw(e.raw_os_error().unwrap_or(0)),
-            )
-        })?;
-        Ok(OwnedFd::from(file))
-    }
-}
-
-fn enlarge_file(fd: BorrowedFd, size: u64) -> Result<(), Error> {
-    // Use posix_fallocate() to enlarge the file. It reserves the space correctly, so that
-    // we don't get a segfault later when trying to actually use it.
-    #[cfg(not(target_os = "macos"))]
-    {
-        nix::fcntl::posix_fallocate(fd, 0, size as i64).map_err(|e| {
-            Error::new(
-                "could not grow shmem segment, posix_fallocate failed: {e}",
-                e,
-            )
-        })
-    }
-    // As a fallback on macos, which doesn't have posix_fallocate, use plain 'fallocate'
-    #[cfg(target_os = "macos")]
-    {
-        nix::unistd::ftruncate(fd, size as i64)
-            .map_err(|e| Error::new("could not grow shmem segment, ftruncate failed: {e}", e))
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    use nix::unistd::ForkResult;
-    use std::ops::Range;
-
-    /// check that all bytes in given range have the expected value.
-    fn assert_range(ptr: *const u8, expected: u8, range: Range<usize>) {
-        for i in range {
-            let b = unsafe { *(ptr.add(i)) };
-            assert_eq!(expected, b, "unexpected byte at offset {i}");
-        }
-    }
-
-    /// Write 'b' to all bytes in the given range
-    fn write_range(ptr: *mut u8, b: u8, range: Range<usize>) {
-        unsafe { std::ptr::write_bytes(ptr.add(range.start), b, range.end - range.start) };
-    }
-
-    // simple single-process test of growing and shrinking
-    #[test]
-    fn test_shmem_resize() -> Result<(), Error> {
-        let max_size = 1024 * 1024;
-        let init_struct = ShmemHandle::new("test_shmem_resize", 0, max_size)?;
-
-        assert_eq!(init_struct.current_size(), 0);
-
-        // Initial grow
-        let size1 = 10000;
-        init_struct.set_size(size1).unwrap();
-        assert_eq!(init_struct.current_size(), size1);
-
-        // Write some data
-        let data_ptr = init_struct.data_ptr.as_ptr();
-        write_range(data_ptr, 0xAA, 0..size1);
-        assert_range(data_ptr, 0xAA, 0..size1);
-
-        // Shrink
-        let size2 = 5000;
-        init_struct.set_size(size2).unwrap();
-        assert_eq!(init_struct.current_size(), size2);
-
-        // Grow again
-        let size3 = 20000;
-        init_struct.set_size(size3).unwrap();
-        assert_eq!(init_struct.current_size(), size3);
-
-        // Try to read it. The area that was shrunk and grown again should read as all zeros now
-        assert_range(data_ptr, 0xAA, 0..5000);
-        assert_range(data_ptr, 0, 5000..size1);
-
-        // Try to grow beyond max_size
-        //let size4 = max_size + 1;
-        //assert!(init_struct.set_size(size4).is_err());
-
-        // Dropping init_struct should unmap the memory
-        drop(init_struct);
-
-        Ok(())
-    }
-
-    /// This is used in tests to coordinate between test processes. It's like std::sync::Barrier,
-    /// but is stored in the shared memory area and works across processes. It's implemented by
-    /// polling, because e.g. standard rust mutexes are not guaranteed to work across processes.
-    struct SimpleBarrier {
-        num_procs: usize,
-        count: AtomicUsize,
-    }
-
-    impl SimpleBarrier {
-        unsafe fn init(ptr: *mut SimpleBarrier, num_procs: usize) {
-            unsafe {
-                *ptr = SimpleBarrier {
-                    num_procs,
-                    count: AtomicUsize::new(0),
-                }
-            }
-        }
-
-        pub fn wait(&self) {
-            let old = self.count.fetch_add(1, Ordering::Relaxed);
-
-            let generation = old / self.num_procs;
-
-            let mut current = old + 1;
-            while current < (generation + 1) * self.num_procs {
-                std::thread::sleep(std::time::Duration::from_millis(10));
-                current = self.count.load(Ordering::Relaxed);
-            }
-        }
-    }
-
-    #[test]
-    fn test_multi_process() {
-        // Initialize
-        let max_size = 1_000_000_000_000;
-        let init_struct = ShmemHandle::new("test_multi_process", 0, max_size).unwrap();
-        let ptr = init_struct.data_ptr.as_ptr();
-
-        // Store the SimpleBarrier in the first 1k of the area.
-        init_struct.set_size(10000).unwrap();
-        let barrier_ptr: *mut SimpleBarrier = unsafe {
-            ptr.add(ptr.align_offset(std::mem::align_of::<SimpleBarrier>()))
-                .cast()
-        };
-        unsafe { SimpleBarrier::init(barrier_ptr, 2) };
-        let barrier = unsafe { barrier_ptr.as_ref().unwrap() };
-
-        // Fork another test process. The code after this runs in both processes concurrently.
-        let fork_result = unsafe { nix::unistd::fork().unwrap() };
-
-        // In the parent, fill bytes between 1000..2000. In the child, between 2000..3000
-        if fork_result.is_parent() {
-            write_range(ptr, 0xAA, 1000..2000);
-        } else {
-            write_range(ptr, 0xBB, 2000..3000);
-        }
-        barrier.wait();
-        // Verify the contents. (in both processes)
-        assert_range(ptr, 0xAA, 1000..2000);
-        assert_range(ptr, 0xBB, 2000..3000);
-
-        // Grow, from the child this time
-        let size = 10_000_000;
-        if !fork_result.is_parent() {
-            init_struct.set_size(size).unwrap();
-        }
-        barrier.wait();
-
-        // make some writes at the end
-        if fork_result.is_parent() {
-            write_range(ptr, 0xAA, (size - 10)..size);
-        } else {
-            write_range(ptr, 0xBB, (size - 20)..(size - 10));
-        }
-        barrier.wait();
-
-        // Verify the contents. (This runs in both processes)
-        assert_range(ptr, 0, (size - 1000)..(size - 20));
-        assert_range(ptr, 0xBB, (size - 20)..(size - 10));
-        assert_range(ptr, 0xAA, (size - 10)..size);
-
-        if let ForkResult::Parent { child } = fork_result {
-            nix::sys::wait::waitpid(child, None).unwrap();
-        }
-    }
-}
+pub mod hash;
+pub mod shmem;
+pub mod sync;

libs/neon-shmem/src/shmem.rs

@@ -0,0 +1,409 @@
+//! Dynamically resizable contiguous chunk of shared memory
+
+use std::num::NonZeroUsize;
+use std::os::fd::{AsFd, BorrowedFd, OwnedFd};
+use std::ptr::NonNull;
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+use nix::errno::Errno;
+use nix::sys::mman::MapFlags;
+use nix::sys::mman::ProtFlags;
+use nix::sys::mman::mmap as nix_mmap;
+use nix::sys::mman::munmap as nix_munmap;
+use nix::unistd::ftruncate as nix_ftruncate;
+
+/// `ShmemHandle` represents a shared memory area that can be shared by processes over `fork()`.
+/// Unlike shared memory allocated by Postgres, this area is resizable, up to `max_size` that's
+/// specified at creation.
+///
+/// The area is backed by an anonymous file created with `memfd_create()`. The full address space for
+/// `max_size` is reserved up-front with `mmap()`, but whenever you call [`ShmemHandle::set_size`],
+/// the underlying file is resized. Do not access the area beyond the current size. Currently, that
+/// will cause the file to be expanded, but we might use `mprotect()` etc. to enforce that in the
+/// future.
+pub struct ShmemHandle {
+    /// memfd file descriptor
+    fd: OwnedFd,
+
+    max_size: usize,
+
+    // Pointer to the beginning of the shared memory area. The header is stored there.
+    shared_ptr: NonNull<SharedStruct>,
+
+    // Pointer to the beginning of the user data
+    pub data_ptr: NonNull<u8>,
+}
+
+/// This is stored at the beginning in the shared memory area.
+struct SharedStruct {
+    max_size: usize,
+
+    /// Current size of the backing file. The high-order bit is used for the [`RESIZE_IN_PROGRESS`] flag.
+    current_size: AtomicUsize,
+}
+
+const RESIZE_IN_PROGRESS: usize = 1 << 63;
+
+const HEADER_SIZE: usize = std::mem::size_of::<SharedStruct>();
+
+/// Error type returned by the [`ShmemHandle`] functions.
+#[derive(thiserror::Error, Debug)]
+#[error("{msg}: {errno}")]
+pub struct Error {
+    pub msg: String,
+    pub errno: Errno,
+}
+
+impl Error {
+    fn new(msg: &str, errno: Errno) -> Self {
+        Self {
+            msg: msg.to_string(),
+            errno,
+        }
+    }
+}
+
+impl ShmemHandle {
+    /// Create a new shared memory area. To communicate between processes, the processes need to be
+    /// `fork()`'d after calling this, so that the `ShmemHandle` is inherited by all processes.
+    ///
+    /// If the `ShmemHandle` is dropped, the memory is unmapped from the current process. Other
+    /// processes can continue using it, however.
+    pub fn new(name: &str, initial_size: usize, max_size: usize) -> Result<Self, Error> {
+        // create the backing anonymous file.
+        let fd = create_backing_file(name)?;
+
+        Self::new_with_fd(fd, initial_size, max_size)
+    }
+
+    fn new_with_fd(fd: OwnedFd, initial_size: usize, max_size: usize) -> Result<Self, Error> {
+        // We reserve the high-order bit for the `RESIZE_IN_PROGRESS` flag, and the actual size
+        // is a little larger than this because of the SharedStruct header. Make the upper limit
+        // somewhat smaller than that, because with anything close to that, you'll run out of
+        // memory anyway.
+        assert!(max_size < 1 << 48, "max size {max_size} too large");
+
+        assert!(
+            initial_size <= max_size,
+            "initial size {initial_size} larger than max size {max_size}"
+        );
+
+        // The actual initial / max size is the one given by the caller, plus the size of
+        // 'SharedStruct'.
+        let initial_size = HEADER_SIZE + initial_size;
+        let max_size = NonZeroUsize::new(HEADER_SIZE + max_size).unwrap();
+
+        // Reserve address space for it with mmap
+        //
+        // TODO: Use MAP_HUGETLB if possible
+        let start_ptr = unsafe {
+            nix_mmap(
+                None,
+                max_size,
+                ProtFlags::PROT_READ | ProtFlags::PROT_WRITE,
+                MapFlags::MAP_SHARED,
+                &fd,
+                0,
+            )
+        }
+        .map_err(|e| Error::new("mmap failed", e))?;
+
+        // Reserve space for the initial size
+        enlarge_file(fd.as_fd(), initial_size as u64)?;
+
+        // Initialize the header
+        let shared: NonNull<SharedStruct> = start_ptr.cast();
+        unsafe {
+            shared.write(SharedStruct {
+                max_size: max_size.into(),
+                current_size: AtomicUsize::new(initial_size),
+            });
+        }
+
+        // The user data begins after the header
+        let data_ptr = unsafe { start_ptr.cast().add(HEADER_SIZE) };
+
+        Ok(Self {
+            fd,
+            max_size: max_size.into(),
+            shared_ptr: shared,
+            data_ptr,
+        })
+    }
+
+    // return reference to the header
+    fn shared(&self) -> &SharedStruct {
+        unsafe { self.shared_ptr.as_ref() }
+    }
+
+    /// Resize the shared memory area. `new_size` must not be larger than the `max_size` specified
+    /// when creating the area.
+    ///
+    /// This may only be called from one process/thread concurrently. We detect that case
+    /// and return an [`shmem::Error`](Error).
+    pub fn set_size(&self, new_size: usize) -> Result<(), Error> {
+        let new_size = new_size + HEADER_SIZE;
+        let shared = self.shared();
+
+        assert!(
+            new_size <= self.max_size,
+            "new size ({new_size}) is greater than max size ({})",
+            self.max_size
+        );
+
+        assert_eq!(self.max_size, shared.max_size);
+
+        // Lock the area by setting the bit in `current_size`
+        //
+        // Ordering::Relaxed would probably be sufficient here, as we don't access any other memory
+        // and the `posix_fallocate`/`ftruncate` call is surely a synchronization point anyway. But
+        // since this is not performance-critical, better safe than sorry.
+        let mut old_size = shared.current_size.load(Ordering::Acquire);
+        loop {
+            if (old_size & RESIZE_IN_PROGRESS) != 0 {
+                return Err(Error::new(
+                    "concurrent resize detected",
+                    Errno::UnknownErrno,
+                ));
+            }
+            match shared.current_size.compare_exchange(
+                old_size,
+                new_size,
+                Ordering::Acquire,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => break,
+                Err(x) => old_size = x,
+            }
+        }
+
+        // Ok, we got the lock.
+        //
+        // NB: If anything goes wrong, we *must* clear the bit!
+        let result = {
+            use std::cmp::Ordering::{Equal, Greater, Less};
+            match new_size.cmp(&old_size) {
+                Less => nix_ftruncate(&self.fd, new_size as i64)
+                    .map_err(|e| Error::new("could not shrink shmem segment, ftruncate failed", e)),
+                Equal => Ok(()),
+                Greater => enlarge_file(self.fd.as_fd(), new_size as u64),
+            }
+        };
+
+        // Unlock
+        shared.current_size.store(
+            if result.is_ok() { new_size } else { old_size },
+            Ordering::Release,
+        );
+
+        result
+    }
+
+    /// Returns the current user-visible size of the shared memory segment.
+    ///
+    /// NOTE: a concurrent [`ShmemHandle::set_size()`] call can change the size at any time.
+    /// It is the caller's responsibility not to access the area beyond the current size.
+    pub fn current_size(&self) -> usize {
+        let total_current_size =
+            self.shared().current_size.load(Ordering::Relaxed) & !RESIZE_IN_PROGRESS;
+        total_current_size - HEADER_SIZE
+    }
+}
+
+impl Drop for ShmemHandle {
+    fn drop(&mut self) {
+        // SAFETY: The pointer was obtained from mmap() with the given size.
+        // We unmap the entire region.
+        let _ = unsafe { nix_munmap(self.shared_ptr.cast(), self.max_size) };
+        // The fd is dropped automatically by OwnedFd.
+    }
+}
+
+/// Create a "backing file" for the shared memory area. On Linux, use `memfd_create()`, to create an
+/// anonymous in-memory file. One macos, fall back to a regular file. That's good enough for
+/// development and testing, but in production we want the file to stay in memory.
+///
+/// Disable unused variables warnings because `name` is unused in the macos path.
+#[allow(unused_variables)]
+fn create_backing_file(name: &str) -> Result<OwnedFd, Error> {
+    #[cfg(not(target_os = "macos"))]
+    {
+        nix::sys::memfd::memfd_create(name, nix::sys::memfd::MFdFlags::empty())
+            .map_err(|e| Error::new("memfd_create failed", e))
+    }
+    #[cfg(target_os = "macos")]
+    {
+        let file = tempfile::tempfile().map_err(|e| {
+            Error::new(
+                "could not create temporary file to back shmem area",
+                nix::errno::Errno::from_raw(e.raw_os_error().unwrap_or(0)),
+            )
+        })?;
+        Ok(OwnedFd::from(file))
+    }
+}
+
+fn enlarge_file(fd: BorrowedFd, size: u64) -> Result<(), Error> {
+    // Use posix_fallocate() to enlarge the file. It reserves the space correctly, so that
+    // we don't get a segfault later when trying to actually use it.
+    #[cfg(not(target_os = "macos"))]
+    {
+        nix::fcntl::posix_fallocate(fd, 0, size as i64)
+            .map_err(|e| Error::new("could not grow shmem segment, posix_fallocate failed", e))
+    }
+    // As a fallback on macos, which doesn't have posix_fallocate, use plain 'fallocate'
+    #[cfg(target_os = "macos")]
+    {
+        nix::unistd::ftruncate(fd, size as i64)
+            .map_err(|e| Error::new("could not grow shmem segment, ftruncate failed", e))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    use nix::unistd::ForkResult;
+    use std::ops::Range;
+
+    /// check that all bytes in given range have the expected value.
+    fn assert_range(ptr: *const u8, expected: u8, range: Range<usize>) {
+        for i in range {
+            let b = unsafe { *(ptr.add(i)) };
+            assert_eq!(expected, b, "unexpected byte at offset {i}");
+        }
+    }
+
+    /// Write 'b' to all bytes in the given range
+    fn write_range(ptr: *mut u8, b: u8, range: Range<usize>) {
+        unsafe { std::ptr::write_bytes(ptr.add(range.start), b, range.end - range.start) };
+    }
+
+    // simple single-process test of growing and shrinking
+    #[test]
+    fn test_shmem_resize() -> Result<(), Error> {
+        let max_size = 1024 * 1024;
+        let init_struct = ShmemHandle::new("test_shmem_resize", 0, max_size)?;
+
+        assert_eq!(init_struct.current_size(), 0);
+
+        // Initial grow
+        let size1 = 10000;
+        init_struct.set_size(size1).unwrap();
+        assert_eq!(init_struct.current_size(), size1);
+
+        // Write some data
+        let data_ptr = init_struct.data_ptr.as_ptr();
+        write_range(data_ptr, 0xAA, 0..size1);
+        assert_range(data_ptr, 0xAA, 0..size1);
+
+        // Shrink
+        let size2 = 5000;
+        init_struct.set_size(size2).unwrap();
+        assert_eq!(init_struct.current_size(), size2);
+
+        // Grow again
+        let size3 = 20000;
+        init_struct.set_size(size3).unwrap();
+        assert_eq!(init_struct.current_size(), size3);
+
+        // Try to read it. The area that was shrunk and grown again should read as all zeros now
+        assert_range(data_ptr, 0xAA, 0..5000);
+        assert_range(data_ptr, 0, 5000..size1);
+
+        // Try to grow beyond max_size
+        //let size4 = max_size + 1;
+        //assert!(init_struct.set_size(size4).is_err());
+
+        // Dropping init_struct should unmap the memory
+        drop(init_struct);
+
+        Ok(())
+    }
+
+    /// This is used in tests to coordinate between test processes. It's like `std::sync::Barrier`,
+    /// but is stored in the shared memory area and works across processes. It's implemented by
+    /// polling, because e.g. standard rust mutexes are not guaranteed to work across processes.
+    struct SimpleBarrier {
+        num_procs: usize,
+        count: AtomicUsize,
+    }
+
+    impl SimpleBarrier {
+        unsafe fn init(ptr: *mut SimpleBarrier, num_procs: usize) {
+            unsafe {
+                *ptr = SimpleBarrier {
+                    num_procs,
+                    count: AtomicUsize::new(0),
+                }
+            }
+        }
+
+        pub fn wait(&self) {
+            let old = self.count.fetch_add(1, Ordering::Relaxed);
+
+            let generation = old / self.num_procs;
+
+            let mut current = old + 1;
+            while current < (generation + 1) * self.num_procs {
+                std::thread::sleep(std::time::Duration::from_millis(10));
+                current = self.count.load(Ordering::Relaxed);
+            }
+        }
+    }
+
+    #[test]
+    fn test_multi_process() {
+        // Initialize
+        let max_size = 1_000_000_000_000;
+        let init_struct = ShmemHandle::new("test_multi_process", 0, max_size).unwrap();
+        let ptr = init_struct.data_ptr.as_ptr();
+
+        // Store the SimpleBarrier in the first 1k of the area.
+        init_struct.set_size(10000).unwrap();
+        let barrier_ptr: *mut SimpleBarrier = unsafe {
+            ptr.add(ptr.align_offset(std::mem::align_of::<SimpleBarrier>()))
+                .cast()
+        };
+        unsafe { SimpleBarrier::init(barrier_ptr, 2) };
+        let barrier = unsafe { barrier_ptr.as_ref().unwrap() };
+
+        // Fork another test process. The code after this runs in both processes concurrently.
+        let fork_result = unsafe { nix::unistd::fork().unwrap() };
+
+        // In the parent, fill bytes between 1000..2000. In the child, between 2000..3000
+        if fork_result.is_parent() {
+            write_range(ptr, 0xAA, 1000..2000);
+        } else {
+            write_range(ptr, 0xBB, 2000..3000);
+        }
+        barrier.wait();
+        // Verify the contents. (in both processes)
+        assert_range(ptr, 0xAA, 1000..2000);
+        assert_range(ptr, 0xBB, 2000..3000);
+
+        // Grow, from the child this time
+        let size = 10_000_000;
+        if !fork_result.is_parent() {
+            init_struct.set_size(size).unwrap();
+        }
+        barrier.wait();
+
+        // make some writes at the end
+        if fork_result.is_parent() {
+            write_range(ptr, 0xAA, (size - 10)..size);
+        } else {
+            write_range(ptr, 0xBB, (size - 20)..(size - 10));
+        }
+        barrier.wait();
+
+        // Verify the contents. (This runs in both processes)
+        assert_range(ptr, 0, (size - 1000)..(size - 20));
+        assert_range(ptr, 0xBB, (size - 20)..(size - 10));
+        assert_range(ptr, 0xAA, (size - 10)..size);
+
+        if let ForkResult::Parent { child } = fork_result {
+            nix::sys::wait::waitpid(child, None).unwrap();
+        }
+    }
+}

libs/neon-shmem/src/sync.rs

@@ -0,0 +1,111 @@
+//! Simple utilities akin to what's in [`std::sync`] but designed to work with shared memory.
+
+use std::mem::MaybeUninit;
+use std::ptr::NonNull;
+
+use nix::errno::Errno;
+
+pub type RwLock<T> = lock_api::RwLock<PthreadRwLock, T>;
+pub type RwLockReadGuard<'a, T> = lock_api::RwLockReadGuard<'a, PthreadRwLock, T>;
+pub type RwLockWriteGuard<'a, T> = lock_api::RwLockWriteGuard<'a, PthreadRwLock, T>;
+pub type ValueReadGuard<'a, T> = lock_api::MappedRwLockReadGuard<'a, PthreadRwLock, T>;
+pub type ValueWriteGuard<'a, T> = lock_api::MappedRwLockWriteGuard<'a, PthreadRwLock, T>;
+
+/// Shared memory read-write lock.
+pub struct PthreadRwLock(Option<NonNull<libc::pthread_rwlock_t>>);
+
+/// Simple macro that calls a function in the libc namespace and panics if return value is nonzero.
+macro_rules! libc_checked {
+    ($fn_name:ident ( $($arg:expr),* )) => {{
+        let res = libc::$fn_name($($arg),*);
+        if res != 0 {
+            panic!("{} failed with {}", stringify!($fn_name), Errno::from_raw(res));
+        }
+    }};
+}
+
+impl PthreadRwLock {
+    /// Creates a new `PthreadRwLock` on top of a pointer to a pthread rwlock.
+    ///
+    /// # Safety
+    /// `lock` must be non-null. Every unsafe operation will panic in the event of an error.
+    pub unsafe fn new(lock: *mut libc::pthread_rwlock_t) -> Self {
+        unsafe {
+            let mut attrs = MaybeUninit::uninit();
+            libc_checked!(pthread_rwlockattr_init(attrs.as_mut_ptr()));
+            libc_checked!(pthread_rwlockattr_setpshared(
+                attrs.as_mut_ptr(),
+                libc::PTHREAD_PROCESS_SHARED
+            ));
+            libc_checked!(pthread_rwlock_init(lock, attrs.as_mut_ptr()));
+            // Safety: POSIX specifies that "any function affecting the attributes
+            // object (including destruction) shall not affect any previously
+            // initialized read-write locks".
+            libc_checked!(pthread_rwlockattr_destroy(attrs.as_mut_ptr()));
+            Self(Some(NonNull::new_unchecked(lock)))
+        }
+    }
+
+    fn inner(&self) -> NonNull<libc::pthread_rwlock_t> {
+        match self.0 {
+            None => {
+                panic!("PthreadRwLock constructed badly - something likely used RawRwLock::INIT")
+            }
+            Some(x) => x,
+        }
+    }
+}
+
+unsafe impl lock_api::RawRwLock for PthreadRwLock {
+    type GuardMarker = lock_api::GuardSend;
+    const INIT: Self = Self(None);
+
+    fn try_lock_shared(&self) -> bool {
+        unsafe {
+            let res = libc::pthread_rwlock_tryrdlock(self.inner().as_ptr());
+            match res {
+                0 => true,
+                libc::EAGAIN => false,
+                _ => panic!(
+                    "pthread_rwlock_tryrdlock failed with {}",
+                    Errno::from_raw(res)
+                ),
+            }
+        }
+    }
+
+    fn try_lock_exclusive(&self) -> bool {
+        unsafe {
+            let res = libc::pthread_rwlock_trywrlock(self.inner().as_ptr());
+            match res {
+                0 => true,
+                libc::EAGAIN => false,
+                _ => panic!("try_wrlock failed with {}", Errno::from_raw(res)),
+            }
+        }
+    }
+
+    fn lock_shared(&self) {
+        unsafe {
+            libc_checked!(pthread_rwlock_rdlock(self.inner().as_ptr()));
+        }
+    }
+
+    fn lock_exclusive(&self) {
+        unsafe {
+            libc_checked!(pthread_rwlock_wrlock(self.inner().as_ptr()));
+        }
+    }
+
+    unsafe fn unlock_exclusive(&self) {
+        unsafe {
+            libc_checked!(pthread_rwlock_unlock(self.inner().as_ptr()));
+        }
+    }
+
+    unsafe fn unlock_shared(&self) {
+        unsafe {
+            libc_checked!(pthread_rwlock_unlock(self.inner().as_ptr()));
+        }
+    }
+}

libs/pageserver_api/src/controller_api.rs

@@ -1,5 +1,6 @@
 use std::collections::{HashMap, HashSet};
 use std::fmt::Display;
+use std::net::IpAddr;
 use std::str::FromStr;
 use std::time::{Duration, Instant};
 
@@ -10,7 +11,7 @@ use serde::{Deserialize, Serialize};
 use utils::id::{NodeId, TenantId, TimelineId};
 use utils::lsn::Lsn;
 
-use crate::models::{PageserverUtilization, ShardParameters, TenantConfig};
+use crate::models::{PageserverUtilization, ShardParameters, TenantConfig, TimelineInfo};
 use crate::shard::{ShardStripeSize, TenantShardId};
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -60,6 +61,11 @@ pub struct NodeRegisterRequest {
     pub listen_https_port: Option<u16>,
 
     pub availability_zone_id: AvailabilityZone,
+
+    // Reachable IP address of the PS/SK registering, if known.
+    // Hadron Cluster Coordiantor will update the DNS record of the registering node
+    // with this IP address.
+    pub node_ip_addr: Option<IpAddr>,
 }
 
 #[derive(Serialize, Deserialize)]
@@ -126,6 +132,13 @@ pub struct TenantDescribeResponse {
     pub config: TenantConfig,
 }
 
+#[derive(Serialize, Deserialize, Debug)]
+pub struct TenantTimelineDescribeResponse {
+    pub shards: Vec<TimelineInfo>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
+}
+
 #[derive(Serialize, Deserialize, Debug)]
 pub struct NodeShardResponse {
     pub node_id: NodeId,
@@ -538,6 +551,39 @@ pub struct SafekeeperDescribeResponse {
     pub scheduling_policy: SkSchedulingPolicy,
 }
 
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct TimelineSafekeeperPeer {
+    pub node_id: NodeId,
+    pub listen_http_addr: String,
+    pub http_port: i32,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SCSafekeeperTimeline {
+    // SC does not know the tenant id.
+    pub timeline_id: TimelineId,
+    pub peers: Vec<NodeId>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SCSafekeeperTimelinesResponse {
+    pub timelines: Vec<SCSafekeeperTimeline>,
+    pub safekeeper_peers: Vec<TimelineSafekeeperPeer>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SafekeeperTimeline {
+    pub tenant_id: TenantId,
+    pub timeline_id: TimelineId,
+    pub peers: Vec<NodeId>,
+}
+
+#[derive(Serialize, Deserialize, Clone, Debug)]
+pub struct SafekeeperTimelinesResponse {
+    pub timelines: Vec<SafekeeperTimeline>,
+    pub safekeeper_peers: Vec<TimelineSafekeeperPeer>,
+}
+
 #[derive(Serialize, Deserialize, Clone)]
 pub struct SafekeeperSchedulingPolicyRequest {
     pub scheduling_policy: SkSchedulingPolicy,

libs/pageserver_api/src/models.rs

@@ -1622,6 +1622,9 @@ pub struct TimelineInfo {
 
     /// Whether the timeline is invisible in synthetic size calculations.
     pub is_invisible: Option<bool>,
+    // HADRON: the largest LSN below which all page updates have been included in the image layers.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub image_consistent_lsn: Option<Lsn>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]

libs/postgres_backend/src/lib.rs

@@ -749,7 +749,18 @@ impl<IO: AsyncRead + AsyncWrite + Unpin> PostgresBackend<IO> {
                 trace!("got query {query_string:?}");
                 if let Err(e) = handler.process_query(self, query_string).await {
                     match e {
-                        QueryError::Shutdown => return Ok(ProcessMsgResult::Break),
+                        err @ QueryError::Shutdown => {
+                            // Notify postgres of the connection shutdown at the libpq
+                            // protocol level. This avoids postgres having to tell apart
+                            // from an idle connection and a stale one, which is bug prone.
+                            let shutdown_error = short_error(&err);
+                            self.write_message_noflush(&BeMessage::ErrorResponse(
+                                &shutdown_error,
+                                Some(err.pg_error_code()),
+                            ))?;
+
+                            return Ok(ProcessMsgResult::Break);
+                        }
                         QueryError::SimulatedConnectionError => {
                             return Err(QueryError::SimulatedConnectionError);
                         }