Build changes to make communicator build on macos.

Does not include splitting requests across shards.

.gitignore

@@ -13,6 +13,7 @@ neon.iml
 /.neon
 /integration_tests/.neon
 compaction-suite-results.*
+pgxn/neon/communicator/communicator_bindings.h
 
 # Coverage
 *.profraw

Cargo.lock

@@ -253,6 +253,17 @@ version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a8ab6b55fe97976e46f91ddbed8d147d966475dc29b2032757ba47e02376fbc3"
 
+[[package]]
+name = "atomic_enum"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "99e1aca718ea7b89985790c94aad72d77533063fe00bc497bb79a7c2dae6a661"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.100",
+]
+
 [[package]]
 name = "autocfg"
 version = "1.1.0"
@@ -687,13 +698,40 @@ 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",
+ "axum-core 0.5.0",
  "base64 0.22.1",
  "bytes",
  "form_urlencoded",
@@ -701,10 +739,10 @@ dependencies = [
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-util",
  "itoa",
- "matchit",
+ "matchit 0.8.4",
  "memchr",
  "mime",
  "percent-encoding",
@@ -724,6 +762,26 @@ 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"
@@ -750,10 +808,9 @@ version = "0.10.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "460fc6f625a1f7705c6cf62d0d070794e94668988b1c38111baeec177c715f7b"
 dependencies = [
- "axum",
- "axum-core",
+ "axum 0.8.1",
+ "axum-core 0.5.0",
  "bytes",
- "form_urlencoded",
  "futures-util",
  "headers",
  "http 1.1.0",
@@ -762,8 +819,6 @@ dependencies = [
  "mime",
  "pin-project-lite",
  "serde",
- "serde_html_form",
- "serde_path_to_error",
  "tower 0.5.2",
  "tower-layer",
  "tower-service",
@@ -903,6 +958,12 @@ version = "0.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9e1b586273c5702936fe7b7d6896644d8be71e6314cfe09d3167c95f712589e8"
 
+[[package]]
+name = "base64"
+version = "0.20.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ea22880d78093b0cbe17c89f64a7d457941e65759157ec6cb31a31d652b05e5"
+
 [[package]]
 name = "base64"
 version = "0.21.7"
@@ -1083,6 +1144,25 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"
 
+[[package]]
+name = "cbindgen"
+version = "0.28.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "eadd868a2ce9ca38de7eeafdcec9c7065ef89b42b32f0839278d55f35c54d1ff"
+dependencies = [
+ "clap",
+ "heck 0.4.1",
+ "indexmap 2.9.0",
+ "log",
+ "proc-macro2",
+ "quote",
+ "serde",
+ "serde_json",
+ "syn 2.0.100",
+ "tempfile",
+ "toml",
+]
+
 [[package]]
 name = "cc"
 version = "1.2.16"
@@ -1209,7 +1289,7 @@ version = "4.5.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4ac6a0c7b1a9e9a5186361f67dfa1b88213572f427fb9ab038efb2bd8c582dab"
 dependencies = [
- "heck",
+ "heck 0.5.0",
  "proc-macro2",
  "quote",
  "syn 2.0.100",
@@ -1267,6 +1347,34 @@ dependencies = [
  "unicode-width",
 ]
 
+[[package]]
+name = "communicator"
+version = "0.1.0"
+dependencies = [
+ "atomic_enum",
+ "axum 0.8.1",
+ "bytes",
+ "cbindgen",
+ "clashmap",
+ "http 1.1.0",
+ "libc",
+ "metrics",
+ "neon-shmem",
+ "nix 0.30.1",
+ "pageserver_client_grpc",
+ "pageserver_page_api",
+ "prometheus",
+ "prost 0.13.5",
+ "thiserror 1.0.69",
+ "tokio",
+ "tokio-pipe",
+ "tonic 0.12.3",
+ "tracing",
+ "tracing-subscriber",
+ "uring-common",
+ "utils",
+]
+
 [[package]]
 name = "compute_api"
 version = "0.1.0"
@@ -1292,9 +1400,9 @@ dependencies = [
  "aws-sdk-kms",
  "aws-sdk-s3",
  "aws-smithy-types",
- "axum",
+ "axum 0.8.1",
  "axum-extra",
- "base64 0.22.1",
+ "base64 0.13.1",
  "bytes",
  "camino",
  "cfg-if",
@@ -1316,6 +1424,7 @@ dependencies = [
  "opentelemetry",
  "opentelemetry_sdk",
  "p256 0.13.2",
+ "pageserver_page_api",
  "postgres",
  "postgres_initdb",
  "regex",
@@ -1334,6 +1443,7 @@ dependencies = [
  "tokio-postgres",
  "tokio-stream",
  "tokio-util",
+ "tonic 0.13.1",
  "tower 0.5.2",
  "tower-http",
  "tower-otel",
@@ -1420,7 +1530,7 @@ name = "control_plane"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "base64 0.22.1",
+ "base64 0.13.1",
  "camino",
  "clap",
  "comfy-table",
@@ -1442,7 +1552,6 @@ dependencies = [
  "regex",
  "reqwest",
  "safekeeper_api",
- "safekeeper_client",
  "scopeguard",
  "serde",
  "serde_json",
@@ -1592,9 +1701,9 @@ dependencies = [
 
 [[package]]
 name = "crossbeam-utils"
-version = "0.8.19"
+version = "0.8.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "248e3bacc7dc6baa3b21e405ee045c3047101a49145e7e9eca583ab4c2ca5345"
+checksum = "d0a5c400df2834b80a4c3327b3aad3a4c4cd4de0629063962b03235697506a28"
 
 [[package]]
 name = "crossterm"
@@ -1934,7 +2043,7 @@ checksum = "0892a17df262a24294c382f0d5997571006e7a4348b4327557c4ff1cd4a8bccc"
 dependencies = [
  "darling",
  "either",
- "heck",
+ "heck 0.5.0",
  "proc-macro2",
  "quote",
  "syn 2.0.100",
@@ -2048,11 +2157,10 @@ name = "endpoint_storage"
 version = "0.0.1"
 dependencies = [
  "anyhow",
- "axum",
+ "axum 0.8.1",
  "axum-extra",
  "camino",
  "camino-tempfile",
- "clap",
  "futures",
  "http-body-util",
  "itertools 0.10.5",
@@ -2329,7 +2437,7 @@ dependencies = [
  "futures-core",
  "futures-sink",
  "http-body-util",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-util",
  "pin-project",
  "rand 0.8.5",
@@ -2499,6 +2607,18 @@ dependencies = [
  "wasm-bindgen",
 ]
 
+[[package]]
+name = "getrandom"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "73fea8450eea4bac3940448fb7ae50d91f034f941199fcd9d909a5a07aa455f0"
+dependencies = [
+ "cfg-if",
+ "libc",
+ "r-efi",
+ "wasi 0.14.2+wasi-0.2.4",
+]
+
 [[package]]
 name = "gettid"
 version = "0.1.3"
@@ -2711,6 +2831,12 @@ dependencies = [
  "http 1.1.0",
 ]
 
+[[package]]
+name = "heck"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"
+
 [[package]]
 name = "heck"
 version = "0.5.0"
@@ -2882,9 +3008,9 @@ dependencies = [
 
 [[package]]
 name = "httparse"
-version = "1.8.0"
+version = "1.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d897f394bad6a705d5f4104762e116a75639e470d80901eed05a860a95cb1904"
+checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
 
 [[package]]
 name = "httpdate"
@@ -2934,9 +3060,9 @@ dependencies = [
 
 [[package]]
 name = "hyper"
-version = "1.4.1"
+version = "1.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "50dfd22e0e76d0f662d429a5f80fcaf3855009297eab6a0a9f8543834744ba05"
+checksum = "cc2b571658e38e0c01b1fdca3bbbe93c00d3d71693ff2770043f8c29bc7d6f80"
 dependencies = [
  "bytes",
  "futures-channel",
@@ -2976,7 +3102,7 @@ checksum = "a0bea761b46ae2b24eb4aef630d8d1c398157b6fc29e6350ecf090a0b70c952c"
 dependencies = [
  "futures-util",
  "http 1.1.0",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-util",
  "rustls 0.22.4",
  "rustls-pki-types",
@@ -2991,7 +3117,7 @@ version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3203a961e5c83b6f5498933e78b6b263e208c197b63e9c6c53cc82ffd3f63793"
 dependencies = [
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-util",
  "pin-project-lite",
  "tokio",
@@ -3000,20 +3126,20 @@ dependencies = [
 
 [[package]]
 name = "hyper-util"
-version = "0.1.7"
+version = "0.1.12"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cde7055719c54e36e95e8719f95883f22072a48ede39db7fc17a4e1d5281e9b9"
+checksum = "cf9f1e950e0d9d1d3c47184416723cf29c0d1f93bd8cccf37e4beb6b44f31710"
 dependencies = [
  "bytes",
  "futures-channel",
  "futures-util",
  "http 1.1.0",
  "http-body 1.0.0",
- "hyper 1.4.1",
+ "hyper 1.6.0",
+ "libc",
  "pin-project-lite",
  "socket2",
  "tokio",
- "tower 0.4.13",
  "tower-service",
  "tracing",
 ]
@@ -3602,6 +3728,12 @@ 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"
@@ -3647,7 +3779,7 @@ version = "0.0.22"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b9e6777fc80a575f9503d908c8b498782a6c3ee88a06cb416dc3941401e43b94"
 dependencies = [
- "heck",
+ "heck 0.5.0",
  "proc-macro2",
  "quote",
  "syn 2.0.100",
@@ -3709,7 +3841,7 @@ dependencies = [
  "procfs",
  "prometheus",
  "rand 0.8.5",
- "rand_distr",
+ "rand_distr 0.4.3",
  "twox-hash",
 ]
 
@@ -3798,11 +3930,25 @@ name = "neon-shmem"
 version = "0.1.0"
 dependencies = [
  "nix 0.30.1",
+ "rand 0.9.1",
+ "rand_distr 0.5.1",
+ "spin",
  "tempfile",
  "thiserror 1.0.69",
  "workspace_hack",
 ]
 
+[[package]]
+name = "neonart"
+version = "0.1.0"
+dependencies = [
+ "crossbeam-utils",
+ "rand 0.9.1",
+ "rand_distr 0.5.1",
+ "spin",
+ "tracing",
+]
+
 [[package]]
 name = "never-say-never"
 version = "6.6.666"
@@ -4236,15 +4382,19 @@ version = "0.1.0"
 dependencies = [
  "anyhow",
  "async-trait",
+ "axum 0.8.1",
  "bytes",
  "camino",
  "clap",
  "futures",
  "hdrhistogram",
+ "http 1.1.0",
  "humantime",
  "humantime-serde",
+ "metrics",
  "pageserver_api",
  "pageserver_client",
+ "pageserver_client_grpc",
  "pageserver_page_api",
  "rand 0.8.5",
  "reqwest",
@@ -4327,6 +4477,7 @@ dependencies = [
  "pageserver_client",
  "pageserver_compaction",
  "pageserver_page_api",
+ "peekable",
  "pem",
  "pin-project-lite",
  "postgres-protocol",
@@ -4339,6 +4490,7 @@ dependencies = [
  "pprof",
  "pq_proto",
  "procfs",
+ "prost 0.13.5",
  "rand 0.8.5",
  "range-set-blaze",
  "regex",
@@ -4438,6 +4590,34 @@ dependencies = [
  "workspace_hack",
 ]
 
+[[package]]
+name = "pageserver_client_grpc"
+version = "0.1.0"
+dependencies = [
+ "async-trait",
+ "bytes",
+ "chrono",
+ "dashmap 5.5.0",
+ "futures",
+ "http 1.1.0",
+ "hyper 1.6.0",
+ "hyper-util",
+ "metrics",
+ "pageserver_api",
+ "pageserver_page_api",
+ "priority-queue",
+ "rand 0.8.5",
+ "thiserror 1.0.69",
+ "tokio",
+ "tokio-stream",
+ "tokio-util",
+ "tonic 0.13.1",
+ "tower 0.4.13",
+ "tracing",
+ "utils",
+ "uuid",
+]
+
 [[package]]
 name = "pageserver_compaction"
 version = "0.1.0"
@@ -4602,6 +4782,15 @@ dependencies = [
  "sha2",
 ]
 
+[[package]]
+name = "peekable"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "225f9651e475709164f871dc2f5724956be59cb9edb055372ffeeab01ec2d20b"
+dependencies = [
+ "smallvec",
+]
+
 [[package]]
 name = "pem"
 version = "3.0.3"
@@ -4812,7 +5001,7 @@ dependencies = [
 name = "postgres-protocol2"
 version = "0.1.0"
 dependencies = [
- "base64 0.22.1",
+ "base64 0.20.0",
  "byteorder",
  "bytes",
  "fallible-iterator",
@@ -5013,6 +5202,17 @@ dependencies = [
  "elliptic-curve 0.13.8",
 ]
 
+[[package]]
+name = "priority-queue"
+version = "2.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ef08705fa1589a1a59aa924ad77d14722cb0cd97b67dd5004ed5f4a4873fce8d"
+dependencies = [
+ "autocfg",
+ "equivalent",
+ "indexmap 2.9.0",
+]
+
 [[package]]
 name = "proc-macro2"
 version = "1.0.94"
@@ -5091,7 +5291,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "22505a5c94da8e3b7c2996394d1c933236c4d743e81a410bcca4e6989fc066a4"
 dependencies = [
  "bytes",
- "heck",
+ "heck 0.5.0",
  "itertools 0.12.1",
  "log",
  "multimap",
@@ -5112,7 +5312,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0c1318b19085f08681016926435853bbf7858f9c082d0999b80550ff5d9abe15"
 dependencies = [
  "bytes",
- "heck",
+ "heck 0.5.0",
  "itertools 0.12.1",
  "log",
  "multimap",
@@ -5184,7 +5384,7 @@ dependencies = [
  "aws-config",
  "aws-sdk-iam",
  "aws-sigv4",
- "base64 0.22.1",
+ "base64 0.13.1",
  "bstr",
  "bytes",
  "camino",
@@ -5213,7 +5413,7 @@ dependencies = [
  "humantime",
  "humantime-serde",
  "hyper 0.14.30",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-util",
  "indexmap 2.9.0",
  "ipnet",
@@ -5237,7 +5437,7 @@ dependencies = [
  "postgres_backend",
  "pq_proto",
  "rand 0.8.5",
- "rand_distr",
+ "rand_distr 0.4.3",
  "rcgen",
  "redis",
  "regex",
@@ -5341,6 +5541,12 @@ dependencies = [
  "proc-macro2",
 ]
 
+[[package]]
+name = "r-efi"
+version = "5.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "74765f6d916ee2faa39bc8e68e4f3ed8949b48cccdac59983d287a7cb71ce9c5"
+
 [[package]]
 name = "rand"
 version = "0.7.3"
@@ -5365,6 +5571,16 @@ 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"
@@ -5385,6 +5601,16 @@ 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"
@@ -5403,6 +5629,15 @@ 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 0.3.2",
+]
+
 [[package]]
 name = "rand_distr"
 version = "0.4.3"
@@ -5413,6 +5648,16 @@ 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"
@@ -5609,7 +5854,7 @@ dependencies = [
  "http-body-util",
  "http-types",
  "humantime-serde",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "itertools 0.10.5",
  "metrics",
  "once_cell",
@@ -5649,7 +5894,7 @@ dependencies = [
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-rustls 0.26.0",
  "hyper-util",
  "ipnet",
@@ -5706,7 +5951,7 @@ dependencies = [
  "futures",
  "getrandom 0.2.11",
  "http 1.1.0",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "parking_lot 0.11.2",
  "reqwest",
  "reqwest-middleware",
@@ -5727,7 +5972,7 @@ dependencies = [
  "async-trait",
  "getrandom 0.2.11",
  "http 1.1.0",
- "matchit",
+ "matchit 0.8.4",
  "opentelemetry",
  "reqwest",
  "reqwest-middleware",
@@ -6419,19 +6664,6 @@ dependencies = [
  "syn 2.0.100",
 ]
 
-[[package]]
-name = "serde_html_form"
-version = "0.2.7"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d2de91cf02bbc07cde38891769ccd5d4f073d22a40683aa4bc7a95781aaa2c4"
-dependencies = [
- "form_urlencoded",
- "indexmap 2.9.0",
- "itoa",
- "ryu",
- "serde",
-]
-
 [[package]]
 name = "serde_json"
 version = "1.0.125"
@@ -6488,17 +6720,15 @@ dependencies = [
 
 [[package]]
 name = "serde_with"
-version = "3.12.0"
+version = "2.3.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6b6f7f2fcb69f747921f79f3926bd1e203fce4fef62c268dd3abfb6d86029aa"
+checksum = "07ff71d2c147a7b57362cead5e22f772cd52f6ab31cfcd9edcd7f6aeb2a0afbe"
 dependencies = [
- "base64 0.22.1",
+ "base64 0.13.1",
  "chrono",
  "hex",
  "indexmap 1.9.3",
- "indexmap 2.9.0",
  "serde",
- "serde_derive",
  "serde_json",
  "serde_with_macros",
  "time",
@@ -6506,9 +6736,9 @@ dependencies = [
 
 [[package]]
 name = "serde_with_macros"
-version = "3.12.0"
+version = "2.3.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8d00caa5193a3c8362ac2b73be6b9e768aa5a4b2f721d8f4b339600c3cb51f8e"
+checksum = "881b6f881b17d13214e5d494c939ebab463d01264ce1811e9d4ac3a882e7695f"
 dependencies = [
  "darling",
  "proc-macro2",
@@ -6662,12 +6892,12 @@ dependencies = [
 
 [[package]]
 name = "socket2"
-version = "0.5.5"
+version = "0.5.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7b5fac59a5cb5dd637972e5fca70daf0523c9067fcdc4842f053dae04a18f8e9"
+checksum = "4f5fd57c80058a56cf5c777ab8a126398ece8e442983605d280a44ce79d0edef"
 dependencies = [
  "libc",
- "windows-sys 0.48.0",
+ "windows-sys 0.52.0",
 ]
 
 [[package]]
@@ -6675,6 +6905,9 @@ name = "spin"
 version = "0.9.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6980e8d7511241f8acf4aebddbb1ff938df5eebe98691418c4468d0b72a96a67"
+dependencies = [
+ "lock_api",
+]
 
 [[package]]
 name = "spinning_top"
@@ -6733,7 +6966,7 @@ dependencies = [
  "http-body-util",
  "http-utils",
  "humantime",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-util",
  "metrics",
  "once_cell",
@@ -6914,7 +7147,7 @@ version = "0.26.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4c6bee85a5a24955dc440386795aa378cd9cf82acd5f764469152d2270e581be"
 dependencies = [
- "heck",
+ "heck 0.5.0",
  "proc-macro2",
  "quote",
  "rustversion",
@@ -7339,6 +7572,16 @@ dependencies = [
  "syn 2.0.100",
 ]
 
+[[package]]
+name = "tokio-pipe"
+version = "0.2.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f213a84bffbd61b8fa0ba8a044b4bbe35d471d0b518867181e82bd5c15542784"
+dependencies = [
+ "libc",
+ "tokio",
+]
+
 [[package]]
 name = "tokio-postgres"
 version = "0.7.10"
@@ -7533,16 +7776,25 @@ 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 1.6.0",
+ "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",
@@ -7555,14 +7807,15 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7e581ba15a835f4d9ea06c55ab1bd4dce26fc53752c69a04aac00703bfb49ba9"
 dependencies = [
  "async-trait",
- "axum",
+ "axum 0.8.1",
  "base64 0.22.1",
  "bytes",
+ "flate2",
  "h2 0.4.4",
  "http 1.1.0",
  "http-body 1.0.0",
  "http-body-util",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-timeout",
  "hyper-util",
  "percent-encoding",
@@ -7614,11 +7867,16 @@ 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]]
@@ -8102,7 +8360,7 @@ name = "vm_monitor"
 version = "0.1.0"
 dependencies = [
  "anyhow",
- "axum",
+ "axum 0.8.1",
  "cgroups-rs",
  "clap",
  "futures",
@@ -8213,6 +8471,15 @@ version = "0.11.0+wasi-snapshot-preview1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9c8d87e72b64a3b4db28d11ce29237c246188f4f51057d65a7eab63b7987e423"
 
+[[package]]
+name = "wasi"
+version = "0.14.2+wasi-0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9683f9a5a998d873c0d21fcbe3c083009670149a8fab228644b8bd36b2c48cb3"
+dependencies = [
+ "wit-bindgen-rt",
+]
+
 [[package]]
 name = "wasite"
 version = "0.1.0"
@@ -8570,6 +8837,15 @@ dependencies = [
  "windows-sys 0.48.0",
 ]
 
+[[package]]
+name = "wit-bindgen-rt"
+version = "0.39.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6f42320e61fe2cfd34354ecb597f86f413484a798ba44a8ca1165c58d42da6c1"
+dependencies = [
+ "bitflags 2.8.0",
+]
+
 [[package]]
 name = "workspace_hack"
 version = "0.1.0"
@@ -8577,8 +8853,9 @@ dependencies = [
  "ahash",
  "anstream",
  "anyhow",
- "axum",
- "axum-core",
+ "axum 0.8.1",
+ "axum-core 0.5.0",
+ "base64 0.13.1",
  "base64 0.21.7",
  "base64ct",
  "bytes",
@@ -8610,7 +8887,7 @@ dependencies = [
  "hex",
  "hmac",
  "hyper 0.14.30",
- "hyper 1.4.1",
+ "hyper 1.6.0",
  "hyper-util",
  "indexmap 2.9.0",
  "itertools 0.12.1",

Cargo.toml

@@ -8,6 +8,7 @@ members = [
     "pageserver/compaction",
     "pageserver/ctl",
     "pageserver/client",
+    "pageserver/client_grpc",
     "pageserver/pagebench",
     "pageserver/page_api",
     "proxy",
@@ -32,6 +33,7 @@ members = [
     "libs/pq_proto",
     "libs/tenant_size_model",
     "libs/metrics",
+    "libs/neonart",
     "libs/postgres_connection",
     "libs/remote_storage",
     "libs/tracing-utils",
@@ -44,6 +46,7 @@ members = [
     "libs/proxy/postgres-types2",
     "libs/proxy/tokio-postgres2",
     "endpoint_storage",
+    "pgxn/neon/communicator",
 ]
 
 [workspace.package]
@@ -71,8 +74,8 @@ aws-credential-types = "1.2.0"
 aws-sigv4 = { version = "1.2", features = ["sign-http"] }
 aws-types = "1.3"
 axum = { version = "0.8.1", features = ["ws"] }
-axum-extra = { version = "0.10.0", features = ["typed-header", "query"] }
-base64 = "0.22"
+axum-extra = { version = "0.10.0", features = ["typed-header"] }
+base64 = "0.13.0"
 bincode = "1.3"
 bindgen = "0.71"
 bit_field = "0.10.2"
@@ -87,6 +90,7 @@ clap = { version = "4.0", features = ["derive", "env"] }
 clashmap = { version = "1.0", features = ["raw-api"] }
 comfy-table = "7.1"
 const_format = "0.2"
+crossbeam-utils = "0.8.21"
 crc32c = "0.6"
 diatomic-waker = { version = "0.2.3" }
 either = "1.8"
@@ -145,6 +149,7 @@ parquet = { version = "53", default-features = false, features = ["zstd"] }
 parquet_derive = "53"
 pbkdf2 = { version = "0.12.1", features = ["simple", "std"] }
 pem = "3.0.3"
+peekable = "0.3.0"
 pin-project-lite = "0.2"
 pprof = { version = "0.14", features = ["criterion", "flamegraph", "frame-pointer", "prost-codec"] }
 procfs = "0.16"
@@ -171,7 +176,7 @@ sentry = { version = "0.37", default-features = false, features = ["backtrace",
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1"
 serde_path_to_error = "0.1"
-serde_with = { version = "3", features = [ "base64" ] }
+serde_with = { version = "2.0", features = [ "base64" ] }
 serde_assert = "0.5.0"
 sha2 = "0.10.2"
 signal-hook = "0.3"
@@ -179,6 +184,7 @@ smallvec = "1.11"
 smol_str = { version = "0.2.0", features = ["serde"] }
 socket2 = "0.5"
 spki = "0.7.3"
+spin = "0.9.8"
 strum = "0.26"
 strum_macros = "0.26"
 "subtle"  = "2.5.0"
@@ -190,16 +196,15 @@ thiserror = "1.0"
 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"] }
-tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
 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-tar = "0.3"
-tokio-util = { version = "0.7.10", features = ["io", "rt"] }
+tokio-util = { version = "0.7.10", features = ["io", "io-util", "rt"] }
 toml = "0.8"
 toml_edit = "0.22"
-tonic = { version = "0.13.1", default-features = false, features = ["channel", "codegen", "prost", "router", "server", "tls-ring", "tls-native-roots"] }
+tonic = { version = "0.13.1", default-features = false, features = ["channel", "codegen", "gzip", "prost", "router", "server", "tls-ring", "tls-native-roots"] }
 tonic-reflection = { version = "0.13.1", features = ["server"] }
 tower = { version = "0.5.2", default-features = false }
 tower-http = { version = "0.6.2", features = ["auth", "request-id", "trace"] }
@@ -232,6 +237,9 @@ x509-cert = { version = "0.2.5" }
 env_logger = "0.11"
 log = "0.4"
 
+tokio-epoll-uring = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
+uring-common = { git = "https://github.com/neondatabase/tokio-epoll-uring.git" , branch = "main" }
+
 ## Libraries from neondatabase/ git forks, ideally with changes to be upstreamed
 postgres = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
 postgres-protocol = { git = "https://github.com/neondatabase/rust-postgres.git", branch = "neon" }
@@ -251,9 +259,12 @@ desim = { version = "0.1", path = "./libs/desim" }
 endpoint_storage = { version = "0.0.1", path = "./endpoint_storage/" }
 http-utils = { version = "0.1", path = "./libs/http-utils/" }
 metrics = { version = "0.1", path = "./libs/metrics/" }
+neonart = { version = "0.1", path = "./libs/neonart/" }
+neon-shmem = { version = "0.1", path = "./libs/neon-shmem/" }
 pageserver = { path = "./pageserver" }
 pageserver_api = { version = "0.1", path = "./libs/pageserver_api/" }
 pageserver_client = { path = "./pageserver/client" }
+pageserver_client_grpc = { path = "./pageserver/client_grpc" }
 pageserver_compaction = { version = "0.1", path = "./pageserver/compaction/" }
 pageserver_page_api = { path = "./pageserver/page_api" }
 postgres_backend = { version = "0.1", path = "./libs/postgres_backend/" }
@@ -278,6 +289,7 @@ walproposer = { version = "0.1", path = "./libs/walproposer/" }
 workspace_hack = { version = "0.1", path = "./workspace_hack/" }
 
 ## Build dependencies
+cbindgen = "0.28.0"
 criterion = "0.5.1"
 rcgen = "0.13"
 rstest = "0.18"

Dockerfile

@@ -110,19 +110,6 @@ RUN set -e \
 	# System postgres for use with client libraries (e.g. in storage controller)
         postgresql-15 \
         openssl \
-        unzip \
-        curl \
-    && ARCH=$(uname -m) \
-    && if [ "$ARCH" = "x86_64" ]; then \
-        curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"; \
-    elif [ "$ARCH" = "aarch64" ]; then \
-        curl "https://awscli.amazonaws.com/awscli-exe-linux-aarch64.zip" -o "awscliv2.zip"; \
-    else \
-        echo "Unsupported architecture: $ARCH" && exit 1; \
-    fi \
-    && unzip awscliv2.zip \
-    && ./aws/install \
-    && rm -rf aws awscliv2.zip \
     && rm -f /etc/apt/apt.conf.d/80-retries \
     && rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/* \
     && useradd -d /data neon \

Makefile

@@ -18,10 +18,12 @@ ifeq ($(BUILD_TYPE),release)
 	PG_LDFLAGS = $(LDFLAGS)
 	# Unfortunately, `--profile=...` is a nightly feature
 	CARGO_BUILD_FLAGS += --release
+	NEON_CARGO_ARTIFACT_TARGET_DIR = $(ROOT_PROJECT_DIR)/target/release
 else ifeq ($(BUILD_TYPE),debug)
 	PG_CONFIGURE_OPTS = --enable-debug --with-openssl --enable-cassert --enable-depend
 	PG_CFLAGS += -O0 -g3 $(CFLAGS)
 	PG_LDFLAGS = $(LDFLAGS)
+	NEON_CARGO_ARTIFACT_TARGET_DIR = $(ROOT_PROJECT_DIR)/target/debug
 else
 	$(error Bad build type '$(BUILD_TYPE)', see Makefile for options)
 endif
@@ -180,11 +182,16 @@ postgres-check-%: postgres-%
 
 .PHONY: neon-pg-ext-%
 neon-pg-ext-%: postgres-%
+	+@echo "Compiling communicator $*"
+	$(CARGO_CMD_PREFIX) cargo build -p communicator $(CARGO_BUILD_FLAGS)
+
 	+@echo "Compiling neon $*"
 	mkdir -p $(POSTGRES_INSTALL_DIR)/build/neon-$*
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config COPT='$(COPT)' \
+		LIBCOMMUNICATOR_PATH=$(NEON_CARGO_ARTIFACT_TARGET_DIR) \
 		-C $(POSTGRES_INSTALL_DIR)/build/neon-$* \
 		-f $(ROOT_PROJECT_DIR)/pgxn/neon/Makefile install
+
 	+@echo "Compiling neon_walredo $*"
 	mkdir -p $(POSTGRES_INSTALL_DIR)/build/neon-walredo-$*
 	$(MAKE) PG_CONFIG=$(POSTGRES_INSTALL_DIR)/$*/bin/pg_config COPT='$(COPT)' \

compute_tools/Cargo.toml

@@ -38,6 +38,7 @@ once_cell.workspace = true
 opentelemetry.workspace = true
 opentelemetry_sdk.workspace = true
 p256 = { version = "0.13", features = ["pem"] }
+pageserver_page_api.workspace = true
 postgres.workspace = true
 regex.workspace = true
 reqwest = { workspace = true, features = ["json"] }
@@ -53,6 +54,7 @@ tokio = { workspace = true, features = ["rt", "rt-multi-thread"] }
 tokio-postgres.workspace = true
 tokio-util.workspace = true
 tokio-stream.workspace = true
+tonic.workspace = true
 tower-otel.workspace = true
 tracing.workspace = true
 tracing-opentelemetry.workspace = true

compute_tools/src/compute.rs

@@ -1,4 +1,4 @@
-use anyhow::{Context, Result};
+use anyhow::{Context, Result, anyhow};
 use chrono::{DateTime, Utc};
 use compute_api::privilege::Privilege;
 use compute_api::responses::{
@@ -15,6 +15,7 @@ use itertools::Itertools;
 use nix::sys::signal::{Signal, kill};
 use nix::unistd::Pid;
 use once_cell::sync::Lazy;
+use pageserver_page_api as page_api;
 use postgres;
 use postgres::NoTls;
 use postgres::error::SqlState;
@@ -29,7 +30,9 @@ use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::{Arc, Condvar, Mutex, RwLock};
 use std::time::{Duration, Instant};
 use std::{env, fs};
+use tokio::io::AsyncReadExt;
 use tokio::spawn;
+use tokio_util::io::StreamReader;
 use tracing::{Instrument, debug, error, info, instrument, warn};
 use url::Url;
 use utils::id::{TenantId, TimelineId};
@@ -369,7 +372,7 @@ impl ComputeNode {
 
         let mut new_state = ComputeState::new();
         if let Some(spec) = config.spec {
-            let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
+            let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow!(msg))?;
             new_state.pspec = Some(pspec);
         }
 
@@ -785,7 +788,7 @@ impl ComputeNode {
         self.spawn_extension_stats_task();
 
         if pspec.spec.autoprewarm {
-            self.prewarm_lfc(None);
+            self.prewarm_lfc();
         }
         Ok(())
     }
@@ -941,6 +944,74 @@ impl ComputeNode {
     #[instrument(skip_all, fields(%lsn))]
     fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
         let spec = compute_state.pspec.as_ref().expect("spec must be set");
+        let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
+
+        match Url::parse(shard0_connstr)?.scheme() {
+            "postgres" | "postgresql" => self.try_get_basebackup_libpq(spec, lsn),
+            "grpc" => self.try_get_basebackup_grpc(spec, lsn),
+            scheme => return Err(anyhow!("unknown URL scheme {scheme}")),
+        }
+    }
+
+    fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<()> {
+        let start_time = Instant::now();
+
+        let shard0_connstr = spec
+            .pageserver_connstr
+            .split(',')
+            .next()
+            .unwrap()
+            .to_string();
+
+        let chunks = tokio::runtime::Handle::current().block_on(async move {
+            let mut client = page_api::proto::PageServiceClient::connect(shard0_connstr).await?;
+
+            let req = page_api::proto::GetBaseBackupRequest {
+                lsn: lsn.0,
+                replica: false, // TODO: handle replicas, with LSN 0
+            };
+            let mut req = tonic::Request::new(req);
+            let metadata = req.metadata_mut();
+            metadata.insert("neon-tenant-id", spec.tenant_id.to_string().parse()?);
+            metadata.insert("neon-timeline-id", spec.timeline_id.to_string().parse()?);
+            metadata.insert("neon-shard-id", "0000".to_string().parse()?); // TODO: shard count
+            if let Some(auth) = spec.storage_auth_token.as_ref() {
+                metadata.insert("authorization", format!("Bearer {auth}").parse()?);
+            }
+
+            let chunks = client.get_base_backup(req).await?.into_inner();
+            anyhow::Ok(chunks)
+        })?;
+        let pageserver_connect_micros = start_time.elapsed().as_micros() as u64;
+
+        // Convert the chunks stream into an AsyncRead
+        let stream_reader = StreamReader::new(
+            chunks.map(|chunk| chunk.map(|c| c.chunk).map_err(std::io::Error::other)),
+        );
+
+        // Wrap the AsyncRead into a blocking reader for compatibility with tar::Archive
+        let reader = tokio_util::io::SyncIoBridge::new(stream_reader);
+        let mut measured_reader = MeasuredReader::new(reader);
+        let mut bufreader = std::io::BufReader::new(&mut measured_reader);
+
+        // Read the archive directly from the `CopyOutReader`
+        //
+        // Set `ignore_zeros` so that unpack() reads all the Copy data and
+        // doesn't stop at the end-of-archive marker. Otherwise, if the server
+        // sends an Error after finishing the tarball, we will not notice it.
+        let mut ar = tar::Archive::new(&mut bufreader);
+        ar.set_ignore_zeros(true);
+        ar.unpack(&self.params.pgdata)?;
+
+        // Report metrics
+        let mut state = self.state.lock().unwrap();
+        state.metrics.pageserver_connect_micros = pageserver_connect_micros;
+        state.metrics.basebackup_bytes = measured_reader.get_byte_count() as u64;
+        state.metrics.basebackup_ms = start_time.elapsed().as_millis() as u64;
+        Ok(())
+    }
+
+    fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<()> {
         let start_time = Instant::now();
 
         let shard0_connstr = spec.pageserver_connstr.split(',').next().unwrap();
@@ -956,12 +1027,10 @@ impl ComputeNode {
         }
 
         config.application_name("compute_ctl");
-        if let Some(spec) = &compute_state.pspec {
-            config.options(&format!(
-                "-c neon.compute_mode={}",
-                spec.spec.mode.to_type_str()
-            ));
-        }
+        config.options(&format!(
+            "-c neon.compute_mode={}",
+            spec.spec.mode.to_type_str()
+        ));
 
         // Connect to pageserver
         let mut client = config.connect(NoTls)?;
@@ -1035,10 +1104,7 @@ impl ComputeNode {
                     return result;
                 }
                 Err(ref e) if attempts < max_attempts => {
-                    warn!(
-                        "Failed to get basebackup: {} (attempt {}/{})",
-                        e, attempts, max_attempts
-                    );
+                    warn!("Failed to get basebackup: {e:?} (attempt {attempts}/{max_attempts})");
                     std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
                     retry_period_ms *= 1.5;
                 }
@@ -1916,7 +1982,7 @@ LIMIT 100",
             self.params
                 .remote_ext_base_url
                 .as_ref()
-                .ok_or(DownloadError::BadInput(anyhow::anyhow!(
+                .ok_or(DownloadError::BadInput(anyhow!(
                     "Remote extensions storage is not configured",
                 )))?;
 
@@ -2112,7 +2178,7 @@ LIMIT 100",
         let remote_extensions = spec
             .remote_extensions
             .as_ref()
-            .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
+            .ok_or(anyhow!("Remote extensions are not configured"))?;
 
         info!("parse shared_preload_libraries from spec.cluster.settings");
         let mut libs_vec = Vec::new();

compute_tools/src/compute_prewarm.rs

@@ -25,16 +25,11 @@ struct EndpointStoragePair {
 }
 
 const KEY: &str = "lfc_state";
-impl EndpointStoragePair {
-    /// endpoint_id is set to None while prewarming from other endpoint, see replica promotion
-    /// If not None, takes precedence over pspec.spec.endpoint_id
-    fn from_spec_and_endpoint(
-        pspec: &crate::compute::ParsedSpec,
-        endpoint_id: Option<String>,
-    ) -> Result<Self> {
-        let endpoint_id = endpoint_id.as_ref().or(pspec.spec.endpoint_id.as_ref());
-        let Some(ref endpoint_id) = endpoint_id else {
-            bail!("pspec.endpoint_id missing, other endpoint_id not provided")
+impl TryFrom<&crate::compute::ParsedSpec> for EndpointStoragePair {
+    type Error = anyhow::Error;
+    fn try_from(pspec: &crate::compute::ParsedSpec) -> Result<Self, Self::Error> {
+        let Some(ref endpoint_id) = pspec.spec.endpoint_id else {
+            bail!("pspec.endpoint_id missing")
         };
         let Some(ref base_uri) = pspec.endpoint_storage_addr else {
             bail!("pspec.endpoint_storage_addr missing")
@@ -89,7 +84,7 @@ impl ComputeNode {
     }
 
     /// Returns false if there is a prewarm request ongoing, true otherwise
-    pub fn prewarm_lfc(self: &Arc<Self>, from_endpoint: Option<String>) -> bool {
+    pub fn prewarm_lfc(self: &Arc<Self>) -> bool {
         crate::metrics::LFC_PREWARM_REQUESTS.inc();
         {
             let state = &mut self.state.lock().unwrap().lfc_prewarm_state;
@@ -102,7 +97,7 @@ impl ComputeNode {
 
         let cloned = self.clone();
         spawn(async move {
-            let Err(err) = cloned.prewarm_impl(from_endpoint).await else {
+            let Err(err) = cloned.prewarm_impl().await else {
                 cloned.state.lock().unwrap().lfc_prewarm_state = LfcPrewarmState::Completed;
                 return;
             };
@@ -114,14 +109,13 @@ impl ComputeNode {
         true
     }
 
-    /// from_endpoint: None for endpoint managed by this compute_ctl
-    fn endpoint_storage_pair(&self, from_endpoint: Option<String>) -> Result<EndpointStoragePair> {
+    fn endpoint_storage_pair(&self) -> Result<EndpointStoragePair> {
         let state = self.state.lock().unwrap();
-        EndpointStoragePair::from_spec_and_endpoint(state.pspec.as_ref().unwrap(), from_endpoint)
+        state.pspec.as_ref().unwrap().try_into()
     }
 
-    async fn prewarm_impl(&self, from_endpoint: Option<String>) -> Result<()> {
-        let EndpointStoragePair { url, token } = self.endpoint_storage_pair(from_endpoint)?;
+    async fn prewarm_impl(&self) -> Result<()> {
+        let EndpointStoragePair { url, token } = self.endpoint_storage_pair()?;
         info!(%url, "requesting LFC state from endpoint storage");
 
         let request = Client::new().get(&url).bearer_auth(token);
@@ -179,7 +173,7 @@ impl ComputeNode {
     }
 
     async fn offload_lfc_impl(&self) -> Result<()> {
-        let EndpointStoragePair { url, token } = self.endpoint_storage_pair(None)?;
+        let EndpointStoragePair { url, token } = self.endpoint_storage_pair()?;
         info!(%url, "requesting LFC state from postgres");
 
         let mut compressed = Vec::new();

compute_tools/src/http/routes/lfc.rs

@@ -2,7 +2,6 @@ use crate::compute_prewarm::LfcPrewarmStateWithProgress;
 use crate::http::JsonResponse;
 use axum::response::{IntoResponse, Response};
 use axum::{Json, http::StatusCode};
-use axum_extra::extract::OptionalQuery;
 use compute_api::responses::LfcOffloadState;
 type Compute = axum::extract::State<std::sync::Arc<crate::compute::ComputeNode>>;
 
@@ -17,16 +16,8 @@ pub(in crate::http) async fn offload_state(compute: Compute) -> Json<LfcOffloadS
     Json(compute.lfc_offload_state())
 }
 
-#[derive(serde::Deserialize)]
-pub struct PrewarmQuery {
-    pub from_endpoint: String,
-}
-
-pub(in crate::http) async fn prewarm(
-    compute: Compute,
-    OptionalQuery(query): OptionalQuery<PrewarmQuery>,
-) -> Response {
-    if compute.prewarm_lfc(query.map(|q| q.from_endpoint)) {
+pub(in crate::http) async fn prewarm(compute: Compute) -> Response {
+    if compute.prewarm_lfc() {
         StatusCode::ACCEPTED.into_response()
     } else {
         JsonResponse::error(

control_plane/Cargo.toml

@@ -36,7 +36,6 @@ pageserver_api.workspace = true
 pageserver_client.workspace = true
 postgres_backend.workspace = true
 safekeeper_api.workspace = true
-safekeeper_client.workspace = true
 postgres_connection.workspace = true
 storage_broker.workspace = true
 http-utils.workspace = true

control_plane/src/bin/neon_local.rs

@@ -18,7 +18,7 @@ use clap::Parser;
 use compute_api::requests::ComputeClaimsScope;
 use compute_api::spec::ComputeMode;
 use control_plane::broker::StorageBroker;
-use control_plane::endpoint::ComputeControlPlane;
+use control_plane::endpoint::{ComputeControlPlane, PageserverProtocol};
 use control_plane::endpoint_storage::{ENDPOINT_STORAGE_DEFAULT_ADDR, EndpointStorage};
 use control_plane::local_env;
 use control_plane::local_env::{
@@ -45,7 +45,7 @@ use pageserver_api::models::{
 use pageserver_api::shard::{DEFAULT_STRIPE_SIZE, ShardCount, ShardStripeSize, TenantShardId};
 use postgres_backend::AuthType;
 use postgres_connection::parse_host_port;
-use safekeeper_api::membership::{SafekeeperGeneration, SafekeeperId};
+use safekeeper_api::membership::SafekeeperGeneration;
 use safekeeper_api::{
     DEFAULT_HTTP_LISTEN_PORT as DEFAULT_SAFEKEEPER_HTTP_PORT,
     DEFAULT_PG_LISTEN_PORT as DEFAULT_SAFEKEEPER_PG_PORT,
@@ -664,6 +664,10 @@ struct EndpointStartCmdArgs {
     #[clap(short = 't', long, value_parser= humantime::parse_duration, help = "timeout until we fail the command")]
     #[arg(default_value = "90s")]
     start_timeout: Duration,
+
+    /// If enabled, use gRPC (and the communicator) to talk to Pageservers.
+    #[clap(long)]
+    grpc: bool,
 }
 
 #[derive(clap::Args)]
@@ -682,6 +686,10 @@ struct EndpointReconfigureCmdArgs {
 
     #[clap(long)]
     safekeepers: Option<String>,
+
+    /// If enabled, use gRPC (and communicator) to talk to Pageservers.
+    #[clap(long)]
+    grpc: bool,
 }
 
 #[derive(clap::Args)]
@@ -1255,45 +1263,6 @@ async fn handle_timeline(cmd: &TimelineCmd, env: &mut local_env::LocalEnv) -> Re
             pageserver
                 .timeline_import(tenant_id, timeline_id, base, pg_wal, args.pg_version)
                 .await?;
-            if env.storage_controller.timelines_onto_safekeepers {
-                println!("Creating timeline on safekeeper ...");
-                let timeline_info = pageserver
-                    .timeline_info(
-                        TenantShardId::unsharded(tenant_id),
-                        timeline_id,
-                        pageserver_client::mgmt_api::ForceAwaitLogicalSize::No,
-                    )
-                    .await?;
-                let default_sk = SafekeeperNode::from_env(env, env.safekeepers.first().unwrap());
-                let default_host = default_sk
-                    .conf
-                    .listen_addr
-                    .clone()
-                    .unwrap_or_else(|| "localhost".to_string());
-                let mconf = safekeeper_api::membership::Configuration {
-                    generation: SafekeeperGeneration::new(1),
-                    members: safekeeper_api::membership::MemberSet {
-                        m: vec![SafekeeperId {
-                            host: default_host,
-                            id: default_sk.conf.id,
-                            pg_port: default_sk.conf.pg_port,
-                        }],
-                    },
-                    new_members: None,
-                };
-                let pg_version = args.pg_version * 10000;
-                let req = safekeeper_api::models::TimelineCreateRequest {
-                    tenant_id,
-                    timeline_id,
-                    mconf,
-                    pg_version,
-                    system_id: None,
-                    wal_seg_size: None,
-                    start_lsn: timeline_info.last_record_lsn,
-                    commit_lsn: None,
-                };
-                default_sk.create_timeline(&req).await?;
-            }
             env.register_branch_mapping(branch_name.to_string(), tenant_id, timeline_id)?;
             println!("Done");
         }
@@ -1491,13 +1460,18 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
 
             let (pageservers, stripe_size) = if let Some(pageserver_id) = pageserver_id {
                 let conf = env.get_pageserver_conf(pageserver_id).unwrap();
-                let parsed = parse_host_port(&conf.listen_pg_addr).expect("Bad config");
-                (
-                    vec![(parsed.0, parsed.1.unwrap_or(5432))],
-                    // If caller is telling us what pageserver to use, this is not a tenant which is
-                    // full managed by storage controller, therefore not sharded.
-                    DEFAULT_STRIPE_SIZE,
-                )
+                // Use gRPC if requested.
+                let (protocol, host, port) = if args.grpc {
+                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
+                    let (host, port) = parse_host_port(grpc_addr).expect("bad config");
+                    (PageserverProtocol::Grpc, host, port.unwrap_or(51051))
+                } else {
+                    let (host, port) = parse_host_port(&conf.listen_pg_addr).expect("bad config");
+                    (PageserverProtocol::Libpq, host, port.unwrap_or(5432))
+                };
+                // If caller is telling us what pageserver to use, this is not a tenant which is
+                // fully managed by storage controller, therefore not sharded.
+                (vec![(protocol, host, port)], DEFAULT_STRIPE_SIZE)
             } else {
                 // Look up the currently attached location of the tenant, and its striping metadata,
                 // to pass these on to postgres.
@@ -1516,11 +1490,22 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                                 .await?;
                         }
 
-                        anyhow::Ok((
-                            Host::parse(&shard.listen_pg_addr)
-                                .expect("Storage controller reported bad hostname"),
-                            shard.listen_pg_port,
-                        ))
+                        let pageserver = if args.grpc {
+                            (
+                                PageserverProtocol::Grpc,
+                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC addr"))
+                                    .expect("bad hostname"),
+                                shard.listen_grpc_port.expect("no gRPC port"),
+                            )
+                        } else {
+                            (
+                                PageserverProtocol::Libpq,
+                                Host::parse(&shard.listen_pg_addr).expect("bad hostname"),
+                                shard.listen_pg_port,
+                            )
+                        };
+
+                        anyhow::Ok(pageserver)
                     }),
                 )
                 .await?;
@@ -1575,11 +1560,17 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                 .get(endpoint_id.as_str())
                 .with_context(|| format!("postgres endpoint {endpoint_id} is not found"))?;
             let pageservers = if let Some(ps_id) = args.endpoint_pageserver_id {
-                let pageserver = PageServerNode::from_env(env, env.get_pageserver_conf(ps_id)?);
-                vec![(
-                    pageserver.pg_connection_config.host().clone(),
-                    pageserver.pg_connection_config.port(),
-                )]
+                let conf = env.get_pageserver_conf(ps_id)?;
+                // Use gRPC if requested.
+                let (protocol, host, port) = if args.grpc {
+                    let grpc_addr = conf.listen_grpc_addr.as_ref().expect("bad config");
+                    let (host, port) = parse_host_port(grpc_addr).expect("bad config");
+                    (PageserverProtocol::Grpc, host, port.unwrap_or(51051))
+                } else {
+                    let (host, port) = parse_host_port(&conf.listen_pg_addr).expect("bad config");
+                    (PageserverProtocol::Libpq, host, port.unwrap_or(5432))
+                };
+                vec![(protocol, host, port)]
             } else {
                 let storage_controller = StorageController::from_env(env);
                 storage_controller
@@ -1588,11 +1579,20 @@ async fn handle_endpoint(subcmd: &EndpointCmd, env: &local_env::LocalEnv) -> Res
                     .shards
                     .into_iter()
                     .map(|shard| {
-                        (
-                            Host::parse(&shard.listen_pg_addr)
-                                .expect("Storage controller reported malformed host"),
-                            shard.listen_pg_port,
-                        )
+                        if args.grpc {
+                            (
+                                PageserverProtocol::Grpc,
+                                Host::parse(&shard.listen_grpc_addr.expect("no gRPC addr"))
+                                    .expect("bad hostname"),
+                                shard.listen_grpc_port.expect("no gRPC port"),
+                            )
+                        } else {
+                            (
+                                PageserverProtocol::Libpq,
+                                Host::parse(&shard.listen_pg_addr).expect("bad hostname"),
+                                shard.listen_pg_port,
+                            )
+                        }
                     })
                     .collect::<Vec<_>>()
             };

control_plane/src/endpoint.rs

@@ -37,6 +37,7 @@
 //! ```
 //!
 use std::collections::BTreeMap;
+use std::fmt::Display;
 use std::net::{IpAddr, Ipv4Addr, SocketAddr, TcpStream};
 use std::path::PathBuf;
 use std::process::Command;
@@ -45,8 +46,6 @@ use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use anyhow::{Context, Result, anyhow, bail};
-use base64::Engine;
-use base64::prelude::BASE64_URL_SAFE_NO_PAD;
 use compute_api::requests::{
     COMPUTE_AUDIENCE, ComputeClaims, ComputeClaimsScope, ConfigurationRequest,
 };
@@ -76,7 +75,6 @@ use utils::id::{NodeId, TenantId, TimelineId};
 
 use crate::local_env::LocalEnv;
 use crate::postgresql_conf::PostgresConf;
-use crate::storage_controller::StorageController;
 
 // contents of a endpoint.json file
 #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
@@ -166,7 +164,7 @@ impl ComputeControlPlane {
                     public_key_use: Some(PublicKeyUse::Signature),
                     key_operations: Some(vec![KeyOperations::Verify]),
                     key_algorithm: Some(KeyAlgorithm::EdDSA),
-                    key_id: Some(BASE64_URL_SAFE_NO_PAD.encode(key_hash)),
+                    key_id: Some(base64::encode_config(key_hash, base64::URL_SAFE_NO_PAD)),
                     x509_url: None::<String>,
                     x509_chain: None::<Vec<String>>,
                     x509_sha1_fingerprint: None::<String>,
@@ -175,7 +173,7 @@ impl ComputeControlPlane {
                 algorithm: AlgorithmParameters::OctetKeyPair(OctetKeyPairParameters {
                     key_type: OctetKeyPairType::OctetKeyPair,
                     curve: EllipticCurve::Ed25519,
-                    x: BASE64_URL_SAFE_NO_PAD.encode(public_key),
+                    x: base64::encode_config(public_key, base64::URL_SAFE_NO_PAD),
                 }),
             }],
         })
@@ -333,7 +331,7 @@ pub enum EndpointStatus {
     RunningNoPidfile,
 }
 
-impl std::fmt::Display for EndpointStatus {
+impl Display for EndpointStatus {
     fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
         let s = match self {
             Self::Running => "running",
@@ -345,6 +343,28 @@ impl std::fmt::Display for EndpointStatus {
     }
 }
 
+#[derive(Clone, Copy, Debug)]
+pub enum PageserverProtocol {
+    Libpq,
+    Grpc,
+}
+
+impl PageserverProtocol {
+    /// Returns the URL scheme for the protocol, used in connstrings.
+    pub fn scheme(&self) -> &'static str {
+        match self {
+            Self::Libpq => "postgresql",
+            Self::Grpc => "grpc",
+        }
+    }
+}
+
+impl Display for PageserverProtocol {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.scheme())
+    }
+}
+
 impl Endpoint {
     fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
         if !entry.file_type()?.is_dir() {
@@ -608,10 +628,10 @@ impl Endpoint {
         }
     }
 
-    fn build_pageserver_connstr(pageservers: &[(Host, u16)]) -> String {
+    fn build_pageserver_connstr(pageservers: &[(PageserverProtocol, Host, u16)]) -> String {
         pageservers
             .iter()
-            .map(|(host, port)| format!("postgresql://no_user@{host}:{port}"))
+            .map(|(scheme, host, port)| format!("{scheme}://no_user@{host}:{port}"))
             .collect::<Vec<_>>()
             .join(",")
     }
@@ -656,7 +676,7 @@ impl Endpoint {
         endpoint_storage_addr: String,
         safekeepers_generation: Option<SafekeeperGeneration>,
         safekeepers: Vec<NodeId>,
-        pageservers: Vec<(Host, u16)>,
+        pageservers: Vec<(PageserverProtocol, Host, u16)>,
         remote_ext_base_url: Option<&String>,
         shard_stripe_size: usize,
         create_test_user: bool,
@@ -941,10 +961,12 @@ impl Endpoint {
 
     pub async fn reconfigure(
         &self,
-        mut pageservers: Vec<(Host, u16)>,
+        pageservers: Vec<(PageserverProtocol, Host, u16)>,
         stripe_size: Option<ShardStripeSize>,
         safekeepers: Option<Vec<NodeId>>,
     ) -> Result<()> {
+        anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
+
         let (mut spec, compute_ctl_config) = {
             let config_path = self.endpoint_path().join("config.json");
             let file = std::fs::File::open(config_path)?;
@@ -956,25 +978,7 @@ impl Endpoint {
         let postgresql_conf = self.read_postgresql_conf()?;
         spec.cluster.postgresql_conf = Some(postgresql_conf);
 
-        // If we weren't given explicit pageservers, query the storage controller
-        if pageservers.is_empty() {
-            let storage_controller = StorageController::from_env(&self.env);
-            let locate_result = storage_controller.tenant_locate(self.tenant_id).await?;
-            pageservers = locate_result
-                .shards
-                .into_iter()
-                .map(|shard| {
-                    (
-                        Host::parse(&shard.listen_pg_addr)
-                            .expect("Storage controller reported bad hostname"),
-                        shard.listen_pg_port,
-                    )
-                })
-                .collect::<Vec<_>>();
-        }
-
         let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
-        assert!(!pageserver_connstr.is_empty());
         spec.pageserver_connstring = Some(pageserver_connstr);
         if stripe_size.is_some() {
             spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);

control_plane/src/pageserver.rs

@@ -265,6 +265,14 @@ impl PageServerNode {
             None => None,
         };
 
+        let mut grpc_host = None;
+        let mut grpc_port = None;
+        if let Some(grpc_addr) = &self.conf.listen_grpc_addr {
+            let (_, port) = parse_host_port(grpc_addr).expect("Unable to parse listen_grpc_addr");
+            grpc_host = Some("localhost".to_string());
+            grpc_port = Some(port.unwrap_or(51051));
+        }
+
         // Intentionally hand-craft JSON: this acts as an implicit format compat test
         // in case the pageserver-side structure is edited, and reflects the real life
         // situation: the metadata is written by some other script.
@@ -273,6 +281,8 @@ impl PageServerNode {
             serde_json::to_vec(&pageserver_api::config::NodeMetadata {
                 postgres_host: "localhost".to_string(),
                 postgres_port: self.pg_connection_config.port(),
+                grpc_host,
+                grpc_port,
                 http_host: "localhost".to_string(),
                 http_port,
                 https_port,
@@ -635,16 +645,4 @@ impl PageServerNode {
 
         Ok(())
     }
-    pub async fn timeline_info(
-        &self,
-        tenant_shard_id: TenantShardId,
-        timeline_id: TimelineId,
-        force_await_logical_size: mgmt_api::ForceAwaitLogicalSize,
-    ) -> anyhow::Result<TimelineInfo> {
-        let timeline_info = self
-            .http_client
-            .timeline_info(tenant_shard_id, timeline_id, force_await_logical_size)
-            .await?;
-        Ok(timeline_info)
-    }
 }

control_plane/src/safekeeper.rs

@@ -6,6 +6,7 @@
 //!   .neon/safekeepers/<safekeeper id>
 //! ```
 use std::error::Error as _;
+use std::future::Future;
 use std::io::Write;
 use std::path::PathBuf;
 use std::time::Duration;
@@ -13,9 +14,9 @@ use std::{io, result};
 
 use anyhow::Context;
 use camino::Utf8PathBuf;
+use http_utils::error::HttpErrorBody;
 use postgres_connection::PgConnectionConfig;
-use safekeeper_api::models::TimelineCreateRequest;
-use safekeeper_client::mgmt_api;
+use reqwest::{IntoUrl, Method};
 use thiserror::Error;
 use utils::auth::{Claims, Scope};
 use utils::id::NodeId;
@@ -34,14 +35,25 @@ pub enum SafekeeperHttpError {
 
 type Result<T> = result::Result<T, SafekeeperHttpError>;
 
-fn err_from_client_err(err: mgmt_api::Error) -> SafekeeperHttpError {
-    use mgmt_api::Error::*;
-    match err {
-        ApiError(_, str) => SafekeeperHttpError::Response(str),
-        Cancelled => SafekeeperHttpError::Response("Cancelled".to_owned()),
-        ReceiveBody(err) => SafekeeperHttpError::Transport(err),
-        ReceiveErrorBody(err) => SafekeeperHttpError::Response(err),
-        Timeout(str) => SafekeeperHttpError::Response(format!("timeout: {str}")),
+pub(crate) trait ResponseErrorMessageExt: Sized {
+    fn error_from_body(self) -> impl Future<Output = Result<Self>> + Send;
+}
+
+impl ResponseErrorMessageExt for reqwest::Response {
+    async fn error_from_body(self) -> Result<Self> {
+        let status = self.status();
+        if !(status.is_client_error() || status.is_server_error()) {
+            return Ok(self);
+        }
+
+        // reqwest does not export its error construction utility functions, so let's craft the message ourselves
+        let url = self.url().to_owned();
+        Err(SafekeeperHttpError::Response(
+            match self.json::<HttpErrorBody>().await {
+                Ok(err_body) => format!("Error: {}", err_body.msg),
+                Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
+            },
+        ))
     }
 }
 
@@ -58,8 +70,9 @@ pub struct SafekeeperNode {
 
     pub pg_connection_config: PgConnectionConfig,
     pub env: LocalEnv,
-    pub http_client: mgmt_api::Client,
+    pub http_client: reqwest::Client,
     pub listen_addr: String,
+    pub http_base_url: String,
 }
 
 impl SafekeeperNode {
@@ -69,14 +82,13 @@ impl SafekeeperNode {
         } else {
             "127.0.0.1".to_string()
         };
-        let jwt = None;
-        let http_base_url = format!("http://{}:{}", listen_addr, conf.http_port);
         SafekeeperNode {
             id: conf.id,
             conf: conf.clone(),
             pg_connection_config: Self::safekeeper_connection_config(&listen_addr, conf.pg_port),
             env: env.clone(),
-            http_client: mgmt_api::Client::new(env.create_http_client(), http_base_url, jwt),
+            http_client: env.create_http_client(),
+            http_base_url: format!("http://{}:{}/v1", listen_addr, conf.http_port),
             listen_addr,
         }
     }
@@ -266,19 +278,20 @@ impl SafekeeperNode {
         )
     }
 
-    pub async fn check_status(&self) -> Result<()> {
-        self.http_client
-            .status()
-            .await
-            .map_err(err_from_client_err)?;
-        Ok(())
+    fn http_request<U: IntoUrl>(&self, method: Method, url: U) -> reqwest::RequestBuilder {
+        // TODO: authentication
+        //if self.env.auth_type == AuthType::NeonJWT {
+        //    builder = builder.bearer_auth(&self.env.safekeeper_auth_token)
+        //}
+        self.http_client.request(method, url)
     }
 
-    pub async fn create_timeline(&self, req: &TimelineCreateRequest) -> Result<()> {
-        self.http_client
-            .create_timeline(req)
-            .await
-            .map_err(err_from_client_err)?;
+    pub async fn check_status(&self) -> Result<()> {
+        self.http_request(Method::GET, format!("{}/{}", self.http_base_url, "status"))
+            .send()
+            .await?
+            .error_from_body()
+            .await?;
         Ok(())
     }
 }

control_plane/storcon_cli/src/main.rs

@@ -37,6 +37,11 @@ enum Command {
         #[arg(long)]
         listen_pg_port: u16,
 
+        #[arg(long)]
+        listen_grpc_addr: Option<String>,
+        #[arg(long)]
+        listen_grpc_port: Option<u16>,
+
         #[arg(long)]
         listen_http_addr: String,
         #[arg(long)]
@@ -61,16 +66,10 @@ enum Command {
         #[arg(long)]
         scheduling: Option<NodeSchedulingPolicy>,
     },
-    // Set a node status as deleted.
     NodeDelete {
         #[arg(long)]
         node_id: NodeId,
     },
-    /// Delete a tombstone of node from the storage controller.
-    NodeDeleteTombstone {
-        #[arg(long)]
-        node_id: NodeId,
-    },
     /// Modify a tenant's policies in the storage controller
     TenantPolicy {
         #[arg(long)]
@@ -88,8 +87,6 @@ enum Command {
     },
     /// List nodes known to the storage controller
     Nodes {},
-    /// List soft deleted nodes known to the storage controller
-    NodeTombstones {},
     /// List tenants known to the storage controller
     Tenants {
         /// If this field is set, it will list the tenants on a specific node
@@ -418,6 +415,8 @@ async fn main() -> anyhow::Result<()> {
             node_id,
             listen_pg_addr,
             listen_pg_port,
+            listen_grpc_addr,
+            listen_grpc_port,
             listen_http_addr,
             listen_http_port,
             listen_https_port,
@@ -431,6 +430,8 @@ async fn main() -> anyhow::Result<()> {
                         node_id,
                         listen_pg_addr,
                         listen_pg_port,
+                        listen_grpc_addr,
+                        listen_grpc_port,
                         listen_http_addr,
                         listen_http_port,
                         listen_https_port,
@@ -908,39 +909,6 @@ async fn main() -> anyhow::Result<()> {
                 .dispatch::<(), ()>(Method::DELETE, format!("control/v1/node/{node_id}"), None)
                 .await?;
         }
-        Command::NodeDeleteTombstone { node_id } => {
-            storcon_client
-                .dispatch::<(), ()>(
-                    Method::DELETE,
-                    format!("debug/v1/tombstone/{node_id}"),
-                    None,
-                )
-                .await?;
-        }
-        Command::NodeTombstones {} => {
-            let mut resp = storcon_client
-                .dispatch::<(), Vec<NodeDescribeResponse>>(
-                    Method::GET,
-                    "debug/v1/tombstone".to_string(),
-                    None,
-                )
-                .await?;
-
-            resp.sort_by(|a, b| a.listen_http_addr.cmp(&b.listen_http_addr));
-
-            let mut table = comfy_table::Table::new();
-            table.set_header(["Id", "Hostname", "AZ", "Scheduling", "Availability"]);
-            for node in resp {
-                table.add_row([
-                    format!("{}", node.id),
-                    node.listen_http_addr,
-                    node.availability_zone_id,
-                    format!("{:?}", node.scheduling),
-                    format!("{:?}", node.availability),
-                ]);
-            }
-            println!("{table}");
-        }
         Command::TenantSetTimeBasedEviction {
             tenant_id,
             period,

docker-compose/ext-src/postgis-src/neon-test.sh

@@ -1,6 +1,9 @@
-#!/bin/sh
+#!/bin/bash
 set -ex
 cd "$(dirname "$0")"
-patch -p1 <"postgis-common-${PG_VERSION}.patch"
-trap 'echo Cleaning up; patch -R -p1 <postgis-common-${PG_VERSION}.patch' EXIT
+if [[ ${PG_VERSION} = v17 ]]; then
+  sed -i '/computed_columns/d' regress/core/tests.mk
+fi
+patch -p1 <postgis-no-upgrade-test.patch
+trap 'echo Cleaning up; patch -R -p1 <postgis-no-upgrade-test.patch' EXIT
 make installcheck-base

docker-compose/ext-src/postgis-src/postgis-common-v17.patch

@@ -1,35 +0,0 @@
-diff --git a/regress/core/tests.mk b/regress/core/tests.mk
-index 9e05244..90987df 100644
---- a/regress/core/tests.mk
-+++ b/regress/core/tests.mk
-@@ -143,8 +143,7 @@ TESTS += \
- 	$(top_srcdir)/regress/core/oriented_envelope \
- 	$(top_srcdir)/regress/core/point_coordinates \
- 	$(top_srcdir)/regress/core/out_geojson \
--  $(top_srcdir)/regress/core/wrapx \
--	$(top_srcdir)/regress/core/computed_columns
-+  $(top_srcdir)/regress/core/wrapx
- 
- # Slow slow tests
- TESTS_SLOW = \
-diff --git a/regress/runtest.mk b/regress/runtest.mk
-index 4b95b7e..449d5a2 100644
---- a/regress/runtest.mk
-+++ b/regress/runtest.mk
-@@ -24,16 +24,6 @@ check-regress:
- 
- 	@POSTGIS_TOP_BUILD_DIR=$(abs_top_builddir) $(PERL) $(top_srcdir)/regress/run_test.pl $(RUNTESTFLAGS) $(RUNTESTFLAGS_INTERNAL) $(TESTS)
- 
--	@if echo "$(RUNTESTFLAGS)" | grep -vq -- --upgrade; then \
--		echo "Running upgrade test as RUNTESTFLAGS did not contain that"; \
--		POSTGIS_TOP_BUILD_DIR=$(abs_top_builddir) $(PERL) $(top_srcdir)/regress/run_test.pl \
--      --upgrade \
--      $(RUNTESTFLAGS) \
--      $(RUNTESTFLAGS_INTERNAL) \
--      $(TESTS); \
--	else \
--		echo "Skipping upgrade test as RUNTESTFLAGS already requested upgrades"; \
--	fi
- 
- check-long:
- 	$(PERL) $(top_srcdir)/regress/run_test.pl $(RUNTESTFLAGS) $(TESTS) $(TESTS_SLOW)

docker-compose/ext-src/postgis-src/postgis-no-upgrade-test.patch

@@ -1,19 +1,3 @@
-diff --git a/regress/core/tests.mk b/regress/core/tests.mk
-index 3abd7bc..64a9254 100644
---- a/regress/core/tests.mk
-+++ b/regress/core/tests.mk
-@@ -144,11 +144,6 @@ TESTS_SLOW = \
- 	$(top_srcdir)/regress/core/concave_hull_hard \
- 	$(top_srcdir)/regress/core/knn_recheck
- 
--ifeq ($(shell expr "$(POSTGIS_PGSQL_VERSION)" ">=" 120),1)
--	TESTS += \
--		$(top_srcdir)/regress/core/computed_columns
--endif
--
- ifeq ($(shell expr "$(POSTGIS_GEOS_VERSION)" ">=" 30700),1)
- 	# GEOS-3.7 adds:
- 	# ST_FrechetDistance
 diff --git a/regress/runtest.mk b/regress/runtest.mk
 index c051f03..010e493 100644
 --- a/regress/runtest.mk

docker-compose/ext-src/postgis-src/postgis-regular-v16.patch

@@ -125,7 +125,7 @@ index 7a36b65..ad78fc7 100644
  DROP SCHEMA tm CASCADE;
 +
 diff --git a/regress/core/tests.mk b/regress/core/tests.mk
-index 64a9254..94903c3 100644
+index 3abd7bc..94903c3 100644
 --- a/regress/core/tests.mk
 +++ b/regress/core/tests.mk
 @@ -23,7 +23,6 @@ current_dir := $(dir $(abspath $(lastword $(MAKEFILE_LIST))))
@@ -160,6 +160,18 @@ index 64a9254..94903c3 100644
  	$(top_srcdir)/regress/core/wkb \
  	$(top_srcdir)/regress/core/wkt \
  	$(top_srcdir)/regress/core/wmsservers \
+@@ -144,11 +140,6 @@ TESTS_SLOW = \
+ 	$(top_srcdir)/regress/core/concave_hull_hard \
+ 	$(top_srcdir)/regress/core/knn_recheck
+ 
+-ifeq ($(shell expr "$(POSTGIS_PGSQL_VERSION)" ">=" 120),1)
+-	TESTS += \
+-		$(top_srcdir)/regress/core/computed_columns
+-endif
+-
+ ifeq ($(shell expr "$(POSTGIS_GEOS_VERSION)" ">=" 30700),1)
+ 	# GEOS-3.7 adds:
+ 	# ST_FrechetDistance
 diff --git a/regress/loader/tests.mk b/regress/loader/tests.mk
 index 1fc77ac..c3cb9de 100644
 --- a/regress/loader/tests.mk

docker-compose/ext-src/postgis-src/postgis-regular-v17.patch

@@ -125,7 +125,7 @@ index 7a36b65..ad78fc7 100644
  DROP SCHEMA tm CASCADE;
 +
 diff --git a/regress/core/tests.mk b/regress/core/tests.mk
-index 90987df..74fe3f1 100644
+index 9e05244..a63a3e1 100644
 --- a/regress/core/tests.mk
 +++ b/regress/core/tests.mk
 @@ -16,14 +16,13 @@ POSTGIS_PGSQL_VERSION=170
@@ -168,6 +168,16 @@ index 90987df..74fe3f1 100644
  	$(top_srcdir)/regress/core/wkb \
  	$(top_srcdir)/regress/core/wkt \
  	$(top_srcdir)/regress/core/wmsservers \
+@@ -143,8 +139,7 @@ TESTS += \
+ 	$(top_srcdir)/regress/core/oriented_envelope \
+ 	$(top_srcdir)/regress/core/point_coordinates \
+ 	$(top_srcdir)/regress/core/out_geojson \
+-  $(top_srcdir)/regress/core/wrapx \
+-	$(top_srcdir)/regress/core/computed_columns
++  $(top_srcdir)/regress/core/wrapx 
+ 
+ # Slow slow tests
+ TESTS_SLOW = \
 diff --git a/regress/loader/tests.mk b/regress/loader/tests.mk
 index ac4f8ad..4bad4fc 100644
 --- a/regress/loader/tests.mk

docker-compose/ext-src/postgis-src/regular-test.sh

@@ -10,8 +10,8 @@ psql -d contrib_regression -c "ALTER DATABASE contrib_regression SET TimeZone='U
      -c "CREATE EXTENSION postgis_tiger_geocoder CASCADE" \
      -c "CREATE EXTENSION postgis_raster SCHEMA public" \
      -c "CREATE EXTENSION postgis_sfcgal SCHEMA public"
-patch -p1 <"postgis-common-${PG_VERSION}.patch"
+patch -p1 <postgis-no-upgrade-test.patch
 patch -p1 <"postgis-regular-${PG_VERSION}.patch"
 psql -d contrib_regression -f raster_outdb_template.sql
-trap 'patch -R -p1 <postgis-regular-${PG_VERSION}.patch && patch -R -p1 <"postgis-common-${PG_VERSION}.patch"' EXIT
+trap 'patch -R -p1 <postgis-no-upgrade-test.patch && patch -R -p1 <"postgis-regular-${PG_VERSION}.patch"' EXIT
 POSTGIS_REGRESS_DB=contrib_regression RUNTESTFLAGS=--nocreate make installcheck-base

endpoint_storage/Cargo.toml

@@ -8,7 +8,6 @@ anyhow.workspace = true
 axum-extra.workspace = true
 axum.workspace = true
 camino.workspace = true
-clap.workspace = true
 futures.workspace = true
 jsonwebtoken.workspace = true
 prometheus.workspace = true

endpoint_storage/src/main.rs

@@ -4,8 +4,6 @@
 //! for large computes.
 mod app;
 use anyhow::Context;
-use clap::Parser;
-use std::net::{IpAddr, Ipv4Addr, SocketAddr};
 use tracing::info;
 use utils::logging;
 
@@ -14,26 +12,9 @@ const fn max_upload_file_limit() -> usize {
     100 * 1024 * 1024
 }
 
-const fn listen() -> SocketAddr {
-    SocketAddr::new(IpAddr::V4(Ipv4Addr::new(0, 0, 0, 0)), 51243)
-}
-
-#[derive(Parser)]
-struct Args {
-    #[arg(exclusive = true)]
-    config_file: Option<String>,
-    #[arg(long, default_value = "false", requires = "config")]
-    /// to allow testing k8s helm chart where we don't have s3 credentials
-    no_s3_check_on_startup: bool,
-    #[arg(long, value_name = "FILE")]
-    /// inline config mode for k8s helm chart
-    config: Option<String>,
-}
-
 #[derive(serde::Deserialize)]
 #[serde(tag = "type")]
 struct Config {
-    #[serde(default = "listen")]
     listen: std::net::SocketAddr,
     pemfile: camino::Utf8PathBuf,
     #[serde(flatten)]
@@ -50,18 +31,13 @@ async fn main() -> anyhow::Result<()> {
         logging::Output::Stdout,
     )?;
 
-    let args = Args::parse();
-    let config: Config = if let Some(config_path) = args.config_file {
-        info!("Reading config from {config_path}");
-        let config = std::fs::read_to_string(config_path)?;
-        serde_json::from_str(&config).context("parsing config")?
-    } else if let Some(config) = args.config {
-        info!("Reading inline config");
-        serde_json::from_str(&config).context("parsing config")?
-    } else {
-        anyhow::bail!("Supply either config file path or --config=inline-config");
-    };
-
+    let config: String = std::env::args().skip(1).take(1).collect();
+    if config.is_empty() {
+        anyhow::bail!("Usage: endpoint_storage config.json")
+    }
+    info!("Reading config from {config}");
+    let config = std::fs::read_to_string(config.clone())?;
+    let config: Config = serde_json::from_str(&config).context("parsing config")?;
     info!("Reading pemfile from {}", config.pemfile.clone());
     let pemfile = std::fs::read(config.pemfile.clone())?;
     info!("Loading public key from {}", config.pemfile.clone());
@@ -72,9 +48,7 @@ async fn main() -> anyhow::Result<()> {
 
     let storage = remote_storage::GenericRemoteStorage::from_config(&config.storage_config).await?;
     let cancel = tokio_util::sync::CancellationToken::new();
-    if !args.no_s3_check_on_startup {
-        app::check_storage_permissions(&storage, cancel.clone()).await?;
-    }
+    app::check_storage_permissions(&storage, cancel.clone()).await?;
 
     let proxy = std::sync::Arc::new(endpoint_storage::Storage {
         auth,

libs/compute_api/src/spec.rs

@@ -4,6 +4,7 @@
 //! provide it by calling the compute_ctl's `/compute_ctl` endpoint, or
 //! compute_ctl can fetch it by calling the control plane's API.
 use std::collections::HashMap;
+use std::fmt::Display;
 
 use indexmap::IndexMap;
 use regex::Regex;
@@ -319,6 +320,12 @@ impl ComputeMode {
     }
 }
 
+impl Display for ComputeMode {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.write_str(self.to_type_str())
+    }
+}
+
 /// Log level for audit logging
 #[derive(Clone, Debug, Default, Eq, PartialEq, Deserialize, Serialize)]
 pub enum ComputeAudit {

libs/neon-shmem/Cargo.toml

@@ -6,8 +6,13 @@ license.workspace = true
 
 [dependencies]
 thiserror.workspace = true
-nix.workspace=true
+nix.workspace = true
+spin.workspace = true
 workspace_hack = { version = "0.1", path = "../../workspace_hack" }
 
+[dev-dependencies]
+rand = "0.9.1"
+rand_distr = "0.5.1"
+
 [target.'cfg(target_os = "macos")'.dependencies]
 tempfile = "3.14.0"

libs/neon-shmem/src/hash.rs

@@ -0,0 +1,366 @@
+//! Hash table implementation on top of 'shmem'
+//!
+//! Features required in the long run by the communicator project:
+//!
+//! [X] Accessible from both Postgres processes and rust threads in the communicator process
+//! [X] Low latency
+//! [ ] Scalable to lots of concurrent accesses (currently uses a single spinlock)
+//! [ ] Resizable
+
+use std::fmt::Debug;
+use std::hash::Hash;
+use std::mem::MaybeUninit;
+use std::ops::Deref;
+
+use crate::shmem::ShmemHandle;
+
+use spin;
+
+mod core;
+
+#[cfg(test)]
+mod tests;
+
+use core::CoreHashMap;
+
+pub enum UpdateAction<V> {
+    Nothing,
+    Insert(V),
+    Remove,
+}
+
+#[derive(Debug)]
+pub struct OutOfMemoryError();
+
+pub struct HashMapInit<'a, K, V> {
+    // Hash table can be allocated in a fixed memory area, or in a resizeable ShmemHandle.
+    shmem_handle: Option<ShmemHandle>,
+    shared_ptr: *mut HashMapShared<'a, K, V>,
+}
+
+pub struct HashMapAccess<'a, K, V> {
+    shmem_handle: Option<ShmemHandle>,
+    shared_ptr: *mut HashMapShared<'a, K, V>,
+}
+
+unsafe impl<'a, K: Sync, V: Sync> Sync for HashMapAccess<'a, K, V> {}
+unsafe impl<'a, K: Send, V: Send> Send for HashMapAccess<'a, K, V> {}
+
+impl<'a, K, V> HashMapInit<'a, K, V> {
+    pub fn attach_writer(self) -> HashMapAccess<'a, K, V> {
+        HashMapAccess {
+            shmem_handle: self.shmem_handle,
+            shared_ptr: self.shared_ptr,
+        }
+    }
+
+    pub fn attach_reader(self) -> HashMapAccess<'a, K, V> {
+        // no difference to attach_writer currently
+        self.attach_writer()
+    }
+}
+
+// This is stored in the shared memory area
+struct HashMapShared<'a, K, V> {
+    inner: spin::RwLock<CoreHashMap<'a, K, V>>,
+}
+
+impl<'a, K, V> HashMapInit<'a, K, V>
+where
+    K: Clone + Hash + Eq,
+{
+    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
+    }
+
+    pub fn init_in_fixed_area(
+        num_buckets: u32,
+        area: &'a mut [MaybeUninit<u8>],
+    ) -> HashMapInit<'a, K, V> {
+        Self::init_common(num_buckets, None, area.as_mut_ptr().cast(), area.len())
+    }
+
+    /// Initialize a new hash map in the given shared memory area
+    pub fn init_in_shmem(num_buckets: u32, mut shmem: ShmemHandle) -> HashMapInit<'a, K, V> {
+        let size = Self::estimate_size(num_buckets);
+        shmem
+            .set_size(size)
+            .expect("could not resize shared memory area");
+
+        let ptr = unsafe { shmem.data_ptr.as_mut() };
+        Self::init_common(num_buckets, Some(shmem), ptr, size)
+    }
+
+    fn init_common(
+        num_buckets: u32,
+        shmem_handle: Option<ShmemHandle>,
+        area_ptr: *mut u8,
+        area_len: usize,
+    ) -> HashMapInit<'a, K, V> {
+        // carve out HashMapShared from the area. This does not include the hashmap's dictionary
+        // and buckets.
+        let mut ptr: *mut u8 = area_ptr;
+        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>>()) };
+
+        // the rest of the space is given to the hash map's dictionary and buckets
+        let remaining_area = unsafe {
+            std::slice::from_raw_parts_mut(ptr, area_len - ptr.offset_from(area_ptr) as usize)
+        };
+
+        let hashmap = CoreHashMap::new(num_buckets, remaining_area);
+        unsafe {
+            std::ptr::write(
+                shared_ptr,
+                HashMapShared {
+                    inner: spin::RwLock::new(hashmap),
+                },
+            );
+        }
+
+        HashMapInit {
+            shmem_handle: shmem_handle,
+            shared_ptr,
+        }
+    }
+}
+
+impl<'a, K, V> HashMapAccess<'a, K, V>
+where
+    K: Clone + Hash + Eq,
+{
+    pub fn get<'e>(&'e self, key: &K) -> Option<ValueReadGuard<'e, K, V>> {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
+        let lock_guard = map.inner.read();
+
+        match lock_guard.get(key) {
+            None => None,
+            Some(val_ref) => {
+                let val_ptr = std::ptr::from_ref(val_ref);
+                Some(ValueReadGuard {
+                    _lock_guard: lock_guard,
+                    value: val_ptr,
+                })
+            }
+        }
+    }
+
+    /// Insert a value
+    pub fn insert(&self, key: &K, value: V) -> Result<bool, OutOfMemoryError> {
+        let mut success = None;
+
+        self.update_with_fn(key, |existing| {
+            if let Some(_) = existing {
+                success = Some(false);
+                UpdateAction::Nothing
+            } else {
+                success = Some(true);
+                UpdateAction::Insert(value)
+            }
+        })?;
+        Ok(success.expect("value_fn not called"))
+    }
+
+    /// Remove value. Returns true if it existed
+    pub fn remove(&self, key: &K) -> bool {
+        let mut result = false;
+        self.update_with_fn(key, |existing| match existing {
+            Some(_) => {
+                result = true;
+                UpdateAction::Remove
+            }
+            None => UpdateAction::Nothing,
+        })
+        .expect("out of memory while removing");
+        result
+    }
+
+    /// Update key using the given function. All the other modifying operations are based on this.
+    pub fn update_with_fn<F>(&self, key: &K, value_fn: F) -> Result<(), OutOfMemoryError>
+    where
+        F: FnOnce(Option<&V>) -> UpdateAction<V>,
+    {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
+        let mut lock_guard = map.inner.write();
+
+        let old_val = lock_guard.get(key);
+        let action = value_fn(old_val);
+        match (old_val, action) {
+            (_, UpdateAction::Nothing) => {}
+            (_, UpdateAction::Insert(new_val)) => {
+                let _ = lock_guard.insert(key, new_val);
+            }
+            (None, UpdateAction::Remove) => panic!("Remove action with no old value"),
+            (Some(_), UpdateAction::Remove) => {
+                let _ = lock_guard.remove(key);
+            }
+        }
+
+        Ok(())
+    }
+
+    /// Update key using the given function. All the other modifying operations are based on this.
+    pub fn update_with_fn_at_bucket<F>(
+        &self,
+        pos: usize,
+        value_fn: F,
+    ) -> Result<(), OutOfMemoryError>
+    where
+        F: FnOnce(Option<&V>) -> UpdateAction<V>,
+    {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
+        let mut lock_guard = map.inner.write();
+
+        let old_val = lock_guard.get_bucket(pos);
+        let action = value_fn(old_val.map(|(_k, v)| v));
+        match (old_val, action) {
+            (_, UpdateAction::Nothing) => {}
+            (_, UpdateAction::Insert(_new_val)) => panic!("cannot insert without key"),
+            (None, UpdateAction::Remove) => panic!("Remove action with no old value"),
+            (Some((key, _value)), UpdateAction::Remove) => {
+                let key = key.clone();
+                let _ = lock_guard.remove(&key);
+            }
+        }
+
+        Ok(())
+    }
+
+    pub fn get_num_buckets(&self) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
+        map.inner.read().get_num_buckets()
+    }
+
+    /// Return the key and value stored in bucket with given index. This can be used to
+    /// iterate through the hash map. (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_bucket<'e>(&'e self, pos: usize) -> Option<ValueReadGuard<'e, K, V>> {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
+        let lock_guard = map.inner.read();
+
+        match lock_guard.get_bucket(pos) {
+            None => None,
+            Some((_key, val_ref)) => {
+                let val_ptr = std::ptr::from_ref(val_ref);
+                Some(ValueReadGuard {
+                    _lock_guard: lock_guard,
+                    value: val_ptr,
+                })
+            }
+        }
+    }
+
+    // for metrics
+    pub fn get_num_buckets_in_use(&self) -> usize {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
+        map.inner.read().buckets_in_use as usize
+    }
+
+    /// Grow
+    ///
+    /// 1. grow the underlying shared memory area
+    /// 2. Initialize new buckets. This overwrites the current dictionary
+    /// 3. Recalculate the dictionary
+    pub fn grow(&self, num_buckets: u32) -> Result<(), crate::shmem::Error> {
+        let map = unsafe { self.shared_ptr.as_ref() }.unwrap();
+        let mut lock_guard = map.inner.write();
+        let inner = &mut *lock_guard;
+        let old_num_buckets = inner.buckets.len() as u32;
+
+        if num_buckets < old_num_buckets {
+            panic!("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>::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 = inner.buckets.as_mut_ptr();
+        unsafe {
+            for i in old_num_buckets..num_buckets {
+                let bucket_ptr = buckets_ptr.add(i as usize);
+                bucket_ptr.write(core::Bucket {
+                    hash: 0,
+                    next: if i < num_buckets {
+                        i as u32 + 1
+                    } else {
+                        inner.free_head
+                    },
+                    inner: None,
+                });
+            }
+        }
+
+        // Recalculate the dictionary
+        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 i in 0..dictionary.len() {
+            dictionary[i] = core::INVALID_POS;
+        }
+
+        for i in 0..old_num_buckets as usize {
+            if buckets[i].inner.is_none() {
+                continue;
+            }
+            let pos: usize = (buckets[i].hash % dictionary.len() as u64) as usize;
+            buckets[i].next = dictionary[pos];
+            dictionary[pos] = i as u32;
+        }
+
+        // Finally, update the CoreHashMap struct
+        inner.dictionary = dictionary;
+        inner.buckets = buckets;
+        inner.free_head = old_num_buckets;
+
+        Ok(())
+    }
+
+    // TODO: Shrinking is a multi-step process that requires co-operation from the caller
+    //
+    // 1. The caller must first call begin_shrink(). That forbids allocation of higher-numbered
+    // buckets.
+    //
+    // 2. Next, the caller must evict all entries in higher-numbered buckets.
+    //
+    // 3. Finally, call finish_shrink(). This recomputes the dictionary and shrinks the underlying
+    //    shmem area
+}
+
+pub struct ValueReadGuard<'a, K, V> {
+    _lock_guard: spin::RwLockReadGuard<'a, CoreHashMap<'a, K, V>>,
+    value: *const V,
+}
+
+impl<'a, K, V> Deref for ValueReadGuard<'a, K, V> {
+    type Target = V;
+
+    fn deref(&self) -> &Self::Target {
+        // SAFETY: The `lock_guard` ensures that the underlying map (and thus the value pointed to
+        // by `value`) remains valid for the lifetime `'a`. The `value` has been obtained from a
+        // valid reference within the map.
+        unsafe { &*self.value }
+    }
+}

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

@@ -0,0 +1,233 @@
+//! Simple hash table with chaining
+//!
+//! # Resizing
+//!
+
+use std::hash::{DefaultHasher, Hash, Hasher};
+use std::mem::MaybeUninit;
+
+pub(crate) const INVALID_POS: u32 = u32::MAX;
+
+// Bucket
+pub(crate) struct Bucket<K, V> {
+    pub(crate) hash: u64,
+    pub(crate) next: u32,
+    pub(crate) inner: Option<(K, V)>,
+}
+
+pub(crate) struct CoreHashMap<'a, K, V> {
+    pub(crate) dictionary: &'a mut [u32],
+    pub(crate) buckets: &'a mut [Bucket<K, V>],
+    pub(crate) free_head: u32,
+
+    // metrics
+    pub(crate) buckets_in_use: u32,
+}
+
+pub struct FullError();
+
+impl<'a, K, V> CoreHashMap<'a, K, V>
+where
+    K: Clone + Hash + Eq,
+{
+    const FILL_FACTOR: f32 = 0.60;
+
+    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(num_buckets: u32, area: &'a mut [u8]) -> CoreHashMap<'a, K, V> {
+        let len = area.len();
+
+        let mut ptr: *mut u8 = area.as_mut_ptr();
+        let end_ptr: *mut u8 = unsafe { area.as_mut_ptr().add(len) };
+
+        // carve out the buckets
+        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<Bucket<K, V>>())) };
+        let buckets_ptr = ptr;
+        ptr = unsafe { ptr.add(size_of::<Bucket<K, V>>() * num_buckets as usize) };
+
+        // use remaining space for the dictionary
+        ptr = unsafe { ptr.byte_add(ptr.align_offset(align_of::<u32>())) };
+        let dictionary_ptr = ptr;
+
+        assert!(ptr.addr() < end_ptr.addr());
+        let dictionary_size = unsafe { end_ptr.byte_offset_from(ptr) / size_of::<u32>() as isize };
+        assert!(dictionary_size > 0);
+
+        // Initialize the buckets
+        let buckets = {
+            let buckets_ptr: *mut MaybeUninit<Bucket<K, V>> = buckets_ptr.cast();
+            let buckets =
+                unsafe { std::slice::from_raw_parts_mut(buckets_ptr, num_buckets as usize) };
+            for i in 0..buckets.len() {
+                buckets[i].write(Bucket {
+                    hash: 0,
+                    next: if i < buckets.len() - 1 {
+                        i as u32 + 1
+                    } else {
+                        INVALID_POS
+                    },
+                    inner: None,
+                });
+            }
+            // TODO: use std::slice::assume_init_mut() once it stabilizes
+            unsafe { std::slice::from_raw_parts_mut(buckets_ptr.cast(), num_buckets as usize) }
+        };
+
+        // Initialize the dictionary
+        let dictionary = {
+            let dictionary_ptr: *mut MaybeUninit<u32> = dictionary_ptr.cast();
+            let dictionary =
+                unsafe { std::slice::from_raw_parts_mut(dictionary_ptr, dictionary_size as usize) };
+
+            for i in 0..dictionary.len() {
+                dictionary[i].write(INVALID_POS);
+            }
+            // TODO: use std::slice::assume_init_mut() once it stabilizes
+            unsafe {
+                std::slice::from_raw_parts_mut(dictionary_ptr.cast(), dictionary_size as usize)
+            }
+        };
+
+        CoreHashMap {
+            dictionary,
+            buckets,
+            free_head: 0,
+            buckets_in_use: 0,
+        }
+    }
+
+    pub fn get(&self, key: &K) -> Option<&V> {
+        let mut hasher = DefaultHasher::new();
+        key.hash(&mut hasher);
+        let hash = hasher.finish();
+
+        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;
+        }
+    }
+
+    pub fn insert(&mut self, key: &K, value: V) -> Result<(), FullError> {
+        let mut hasher = DefaultHasher::new();
+        key.hash(&mut hasher);
+        let hash = hasher.finish();
+
+        let first = self.dictionary[hash as usize % self.dictionary.len()];
+        if first == INVALID_POS {
+            // no existing entry
+            let pos = self.alloc_bucket(key.clone(), value, hash)?;
+            if pos == INVALID_POS {
+                return Err(FullError());
+            }
+            self.dictionary[hash as usize % self.dictionary.len()] = pos;
+            return Ok(());
+        }
+
+        let mut next = first;
+        loop {
+            let bucket = &mut self.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, update its value
+                *bucket_value = value;
+                return Ok(());
+            }
+
+            if bucket.next == INVALID_POS {
+                // No existing entry found. Append to the chain
+                let pos = self.alloc_bucket(key.clone(), value, hash)?;
+                if pos == INVALID_POS {
+                    return Err(FullError());
+                }
+                self.buckets[next as usize].next = pos;
+                return Ok(());
+            }
+            next = bucket.next;
+        }
+    }
+
+    pub fn remove(&mut self, key: &K) -> Result<(), FullError> {
+        let mut hasher = DefaultHasher::new();
+        key.hash(&mut hasher);
+        let hash = hasher.finish();
+
+        let mut next = self.dictionary[hash as usize % self.dictionary.len()];
+        let mut prev_pos: u32 = INVALID_POS;
+        loop {
+            if next == INVALID_POS {
+                // no existing entry
+                return Ok(());
+            }
+            let bucket = &mut self.buckets[next as usize];
+            let (bucket_key, _) = bucket.inner.as_mut().expect("entry is in use");
+            if bucket_key == key {
+                // found existing entry, unlink it from the chain
+                if prev_pos == INVALID_POS {
+                    self.dictionary[hash as usize % self.dictionary.len()] = bucket.next;
+                } else {
+                    self.buckets[prev_pos as usize].next = bucket.next;
+                }
+
+                // and add it to the freelist
+                let bucket = &mut self.buckets[next as usize];
+                bucket.hash = 0;
+                bucket.inner = None;
+                bucket.next = self.free_head;
+                self.free_head = next;
+                self.buckets_in_use -= 1;
+                return Ok(());
+            }
+            prev_pos = next;
+            next = bucket.next;
+        }
+    }
+
+    pub fn get_num_buckets(&self) -> usize {
+        self.buckets.len()
+    }
+
+    pub fn get_bucket(&self, pos: usize) -> Option<&(K, V)> {
+        if pos >= self.buckets.len() {
+            return None;
+        }
+
+        self.buckets[pos].inner.as_ref()
+    }
+
+    fn alloc_bucket(&mut self, key: K, value: V, hash: u64) -> Result<u32, FullError> {
+        let pos = self.free_head;
+        if pos == INVALID_POS {
+            return Err(FullError());
+        }
+
+        let bucket = &mut self.buckets[pos as usize];
+        self.free_head = bucket.next;
+        self.buckets_in_use += 1;
+
+        bucket.hash = hash;
+        bucket.next = INVALID_POS;
+        bucket.inner = Some((key, value));
+
+        return Ok(pos);
+    }
+}

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

@@ -0,0 +1,220 @@
+use std::collections::BTreeMap;
+use std::collections::HashSet;
+use std::fmt::{Debug, Formatter};
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+use crate::hash::HashMapAccess;
+use crate::hash::HashMapInit;
+use crate::hash::UpdateAction;
+use crate::shmem::ShmemHandle;
+
+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]) {
+    const MAX_MEM_SIZE: usize = 10000000;
+    let shmem = ShmemHandle::new("test_inserts", 0, MAX_MEM_SIZE).unwrap();
+
+    let init_struct = HashMapInit::<TestKey, usize>::init_in_shmem(100000, shmem);
+    let w = init_struct.attach_writer();
+
+    for (idx, k) in keys.iter().enumerate() {
+        let res = w.insert(&(*k).into(), 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));
+    }
+
+    //eprintln!("stats: {:?}", tree_writer.get_statistics());
+}
+
+#[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.get(&key).is_some() {
+                continue;
+            }
+            used_keys.insert(key);
+            keys.push(key.into());
+            break;
+        }
+    }
+    test_inserts(&keys);
+}
+
+struct TestValue(AtomicUsize);
+
+impl TestValue {
+    fn new(val: usize) -> TestValue {
+        TestValue(AtomicUsize::new(val))
+    }
+
+    fn load(&self) -> usize {
+        self.0.load(Ordering::Relaxed)
+    }
+}
+
+impl Clone for TestValue {
+    fn clone(&self) -> TestValue {
+        TestValue::new(self.load())
+    }
+}
+
+impl Debug for TestValue {
+    fn fmt(&self, fmt: &mut Formatter<'_>) -> Result<(), std::fmt::Error> {
+        write!(fmt, "{:?}", self.load())
+    }
+}
+
+#[derive(Clone, Debug)]
+struct TestOp(TestKey, Option<usize>);
+
+fn apply_op(
+    op: &TestOp,
+    sut: &HashMapAccess<TestKey, TestValue>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+) {
+    eprintln!("applying op: {op:?}");
+
+    // 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)
+    };
+
+    // apply to Art tree
+    sut.update_with_fn(&op.0, |existing| {
+        assert_eq!(existing.map(TestValue::load), shadow_existing);
+
+        match (existing, op.1) {
+            (None, None) => UpdateAction::Nothing,
+            (None, Some(new_val)) => UpdateAction::Insert(TestValue::new(new_val)),
+            (Some(_old_val), None) => UpdateAction::Remove,
+            (Some(old_val), Some(new_val)) => {
+                old_val.0.store(new_val, Ordering::Relaxed);
+                UpdateAction::Nothing
+            }
+        }
+    })
+    .expect("out of memory");
+}
+
+#[test]
+fn random_ops() {
+    const MAX_MEM_SIZE: usize = 10000000;
+    let shmem = ShmemHandle::new("test_inserts", 0, MAX_MEM_SIZE).unwrap();
+
+    let init_struct = HashMapInit::<TestKey, TestValue>::init_in_shmem(100000, shmem);
+    let writer = init_struct.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, &writer, &mut shadow);
+
+        if i % 1000 == 0 {
+            eprintln!("{i} ops processed");
+            //eprintln!("stats: {:?}", tree_writer.get_statistics());
+            //test_iter(&tree_writer, &shadow);
+        }
+    }
+}
+
+#[test]
+fn test_grow() {
+    const MEM_SIZE: usize = 10000000;
+    let shmem = ShmemHandle::new("test_grow", 0, MEM_SIZE).unwrap();
+
+    let init_struct = HashMapInit::<TestKey, TestValue>::init_in_shmem(1000, shmem);
+    let writer = init_struct.attach_writer();
+
+    let mut shadow: std::collections::BTreeMap<TestKey, usize> = BTreeMap::new();
+
+    let mut rng = rand::rng();
+    for i in 0..10000 {
+        let key: TestKey = ((rng.next_u32() % 1000) as u128).into();
+
+        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
+
+        apply_op(&op, &writer, &mut shadow);
+
+        if i % 1000 == 0 {
+            eprintln!("{i} ops processed");
+            //eprintln!("stats: {:?}", tree_writer.get_statistics());
+            //test_iter(&tree_writer, &shadow);
+        }
+    }
+
+    writer.grow(1500).unwrap();
+
+    for i in 0..10000 {
+        let key: TestKey = ((rng.next_u32() % 1500) as u128).into();
+
+        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
+
+        apply_op(&op, &writer, &mut shadow);
+
+        if i % 1000 == 0 {
+            eprintln!("{i} ops processed");
+            //eprintln!("stats: {:?}", tree_writer.get_statistics());
+            //test_iter(&tree_writer, &shadow);
+        }
+    }
+}

libs/neon-shmem/src/lib.rs

@@ -1,418 +1,4 @@
 //! 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 {} too large", max_size);
-        }
-        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;

libs/neon-shmem/src/shmem.rs

@@ -0,0 +1,418 @@
+//! 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) -> 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 {} too large", max_size);
+        }
+        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();
+        }
+    }
+}

libs/neonart/Cargo.toml

@@ -0,0 +1,14 @@
+[package]
+name = "neonart"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+
+[dependencies]
+crossbeam-utils.workspace = true
+spin.workspace = true
+tracing.workspace = true
+
+[dev-dependencies]
+rand = "0.9.1"
+rand_distr = "0.5.1"

libs/neonart/src/algorithm.rs

@@ -0,0 +1,594 @@
+mod lock_and_version;
+pub(crate) mod node_ptr;
+mod node_ref;
+
+use std::vec::Vec;
+
+use crate::algorithm::lock_and_version::ConcurrentUpdateError;
+use crate::algorithm::node_ptr::MAX_PREFIX_LEN;
+use crate::algorithm::node_ref::{NewNodeRef, NodeRef, ReadLockedNodeRef, WriteLockedNodeRef};
+use crate::allocator::OutOfMemoryError;
+
+use crate::TreeWriteGuard;
+use crate::UpdateAction;
+use crate::allocator::ArtAllocator;
+use crate::epoch::EpochPin;
+use crate::{Key, Value};
+
+pub(crate) type RootPtr<V> = node_ptr::NodePtr<V>;
+
+#[derive(Debug)]
+pub enum ArtError {
+    ConcurrentUpdate, // need to retry
+    OutOfMemory,
+}
+
+impl From<ConcurrentUpdateError> for ArtError {
+    fn from(_: ConcurrentUpdateError) -> ArtError {
+        ArtError::ConcurrentUpdate
+    }
+}
+
+impl From<OutOfMemoryError> for ArtError {
+    fn from(_: OutOfMemoryError) -> ArtError {
+        ArtError::OutOfMemory
+    }
+}
+
+pub fn new_root<V: Value>(
+    allocator: &impl ArtAllocator<V>,
+) -> Result<RootPtr<V>, OutOfMemoryError> {
+    node_ptr::new_root(allocator)
+}
+
+pub(crate) fn search<'e, K: Key, V: Value>(
+    key: &K,
+    root: RootPtr<V>,
+    epoch_pin: &'e EpochPin,
+) -> Option<&'e V> {
+    loop {
+        let root_ref = NodeRef::from_root_ptr(root);
+        if let Ok(result) = lookup_recurse(key.as_bytes(), root_ref, None, epoch_pin) {
+            break result;
+        }
+        // retry
+    }
+}
+
+pub(crate) fn iter_next<'e, V: Value>(
+    key: &[u8],
+    root: RootPtr<V>,
+    epoch_pin: &'e EpochPin,
+) -> Option<(Vec<u8>, &'e V)> {
+    loop {
+        let mut path = Vec::new();
+        let root_ref = NodeRef::from_root_ptr(root);
+
+        match next_recurse(key, &mut path, root_ref, epoch_pin) {
+            Ok(Some(v)) => {
+                assert_eq!(path.len(), key.len());
+                break Some((path, v));
+            }
+            Ok(None) => break None,
+            Err(ConcurrentUpdateError()) => {
+                // retry
+                continue;
+            }
+        }
+    }
+}
+
+pub(crate) fn update_fn<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>, F>(
+    key: &K,
+    value_fn: F,
+    root: RootPtr<V>,
+    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
+) -> Result<(), OutOfMemoryError>
+where
+    F: FnOnce(Option<&V>) -> UpdateAction<V>,
+{
+    let value_fn_cell = std::cell::Cell::new(Some(value_fn));
+    loop {
+        let root_ref = NodeRef::from_root_ptr(root);
+        let this_value_fn = |arg: Option<&V>| value_fn_cell.take().unwrap()(arg);
+        let key_bytes = key.as_bytes();
+
+        match update_recurse(
+            key_bytes,
+            this_value_fn,
+            root_ref,
+            None,
+            None,
+            guard,
+            0,
+            key_bytes,
+        ) {
+            Ok(()) => break Ok(()),
+            Err(ArtError::ConcurrentUpdate) => {
+                continue; // retry
+            }
+            Err(ArtError::OutOfMemory) => break Err(OutOfMemoryError()),
+        }
+    }
+}
+
+// Error means you must retry.
+//
+// This corresponds to the 'lookupOpt' function in the paper
+fn lookup_recurse<'e, V: Value>(
+    key: &[u8],
+    node: NodeRef<'e, V>,
+    parent: Option<ReadLockedNodeRef<V>>,
+    epoch_pin: &'e EpochPin,
+) -> Result<Option<&'e V>, ConcurrentUpdateError> {
+    let rnode = node.read_lock_or_restart()?;
+    if let Some(parent) = parent {
+        parent.read_unlock_or_restart()?;
+    }
+
+    // check if the prefix matches, may increment level
+    let prefix_len = if let Some(prefix_len) = rnode.prefix_matches(key) {
+        prefix_len
+    } else {
+        rnode.read_unlock_or_restart()?;
+        return Ok(None);
+    };
+
+    if rnode.is_leaf() {
+        assert_eq!(key.len(), prefix_len);
+        let vptr = rnode.get_leaf_value_ptr()?;
+        // safety: It's OK to return a ref of the pointer because we checked the version
+        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
+        // as long as the epoch is pinned.
+        let v = unsafe { vptr.as_ref().unwrap() };
+        return Ok(Some(v));
+    }
+
+    let key = &key[prefix_len..];
+
+    // find child (or leaf value)
+    let next_node = rnode.find_child_or_restart(key[0])?;
+
+    match next_node {
+        None => Ok(None), // key not found
+        Some(child) => lookup_recurse(&key[1..], child, Some(rnode), epoch_pin),
+    }
+}
+
+fn next_recurse<'e, V: Value>(
+    min_key: &[u8],
+    path: &mut Vec<u8>,
+    node: NodeRef<'e, V>,
+    epoch_pin: &'e EpochPin,
+) -> Result<Option<&'e V>, ConcurrentUpdateError> {
+    let rnode = node.read_lock_or_restart()?;
+    let prefix = rnode.get_prefix();
+    if prefix.len() != 0 {
+        path.extend_from_slice(prefix);
+    }
+
+    use std::cmp::Ordering;
+    let comparison = path.as_slice().cmp(&min_key[0..path.len()]);
+    if comparison == Ordering::Less {
+        rnode.read_unlock_or_restart()?;
+        return Ok(None);
+    }
+
+    if rnode.is_leaf() {
+        assert_eq!(path.len(), min_key.len());
+        let vptr = rnode.get_leaf_value_ptr()?;
+        // safety: It's OK to return a ref of the pointer because we checked the version
+        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
+        // as long as the epoch is pinned.
+        let v = unsafe { vptr.as_ref().unwrap() };
+        return Ok(Some(v));
+    }
+
+    let mut min_key_byte = match comparison {
+        Ordering::Less => unreachable!(), // checked this above already
+        Ordering::Equal => min_key[path.len()],
+        Ordering::Greater => 0,
+    };
+
+    loop {
+        match rnode.find_next_child_or_restart(min_key_byte)? {
+            None => {
+                return Ok(None);
+            }
+            Some((key_byte, child_ref)) => {
+                let path_len = path.len();
+                path.push(key_byte);
+                let result = next_recurse(min_key, path, child_ref, epoch_pin)?;
+                if result.is_some() {
+                    return Ok(result);
+                }
+                if key_byte == u8::MAX {
+                    return Ok(None);
+                }
+                path.truncate(path_len);
+                min_key_byte = key_byte + 1;
+            }
+        }
+    }
+}
+
+// This corresponds to the 'insertOpt' function in the paper
+pub(crate) fn update_recurse<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>, F>(
+    key: &[u8],
+    value_fn: F,
+    node: NodeRef<'e, V>,
+    rparent: Option<(ReadLockedNodeRef<V>, u8)>,
+    rgrandparent: Option<(ReadLockedNodeRef<V>, u8)>,
+    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
+    level: usize,
+    orig_key: &[u8],
+) -> Result<(), ArtError>
+where
+    F: FnOnce(Option<&V>) -> UpdateAction<V>,
+{
+    let rnode = node.read_lock_or_restart()?;
+
+    let prefix_match_len = rnode.prefix_matches(key);
+    if prefix_match_len.is_none() {
+        let (rparent, parent_key) = rparent.expect("direct children of the root have no prefix");
+        let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
+        let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
+
+        match value_fn(None) {
+            UpdateAction::Nothing => {}
+            UpdateAction::Insert(new_value) => {
+                insert_split_prefix(key, new_value, &mut wnode, &mut wparent, parent_key, guard)?;
+            }
+            UpdateAction::Remove => {
+                panic!("unexpected Remove action on insertion");
+            }
+        }
+        wnode.write_unlock();
+        wparent.write_unlock();
+        return Ok(());
+    }
+    let prefix_match_len = prefix_match_len.unwrap();
+    let key = &key[prefix_match_len as usize..];
+    let level = level + prefix_match_len as usize;
+
+    if rnode.is_leaf() {
+        assert_eq!(key.len(), 0);
+        let (rparent, parent_key) = rparent.expect("root cannot be leaf");
+        let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
+        let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
+
+        // safety: Now that we have acquired the write lock, we have exclusive access to the
+        // value. XXX: There might be concurrent reads though?
+        let value_mut = wnode.get_leaf_value_mut();
+
+        match value_fn(Some(value_mut)) {
+            UpdateAction::Nothing => {
+                wparent.write_unlock();
+                wnode.write_unlock();
+            }
+            UpdateAction::Insert(_) => panic!("cannot insert over existing value"),
+            UpdateAction::Remove => {
+                guard.remember_obsolete_node(wnode.as_ptr());
+                wparent.delete_child(parent_key);
+                wnode.write_unlock_obsolete();
+
+                if let Some(rgrandparent) = rgrandparent {
+                    // FIXME: Ignore concurrency error. It doesn't lead to
+                    // corruption, but it means we might leak something. Until
+                    // another update cleans it up.
+                    let _ = cleanup_parent(wparent, rgrandparent, guard);
+                }
+            }
+        }
+
+        return Ok(());
+    }
+
+    let next_node = rnode.find_child_or_restart(key[0])?;
+
+    if next_node.is_none() {
+        if rnode.is_full() {
+            let (rparent, parent_key) = rparent.expect("root node cannot become full");
+            let mut wparent = rparent.upgrade_to_write_lock_or_restart()?;
+            let wnode = rnode.upgrade_to_write_lock_or_restart()?;
+
+            match value_fn(None) {
+                UpdateAction::Nothing => {
+                    wnode.write_unlock();
+                    wparent.write_unlock();
+                }
+                UpdateAction::Insert(new_value) => {
+                    insert_and_grow(key, new_value, wnode, &mut wparent, parent_key, guard)?;
+                    wparent.write_unlock();
+                }
+                UpdateAction::Remove => {
+                    panic!("unexpected Remove action on insertion");
+                }
+            };
+        } else {
+            let mut wnode = rnode.upgrade_to_write_lock_or_restart()?;
+            if let Some((rparent, _)) = rparent {
+                rparent.read_unlock_or_restart()?;
+            }
+            match value_fn(None) {
+                UpdateAction::Nothing => {}
+                UpdateAction::Insert(new_value) => {
+                    insert_to_node(&mut wnode, key, new_value, guard)?;
+                }
+                UpdateAction::Remove => {
+                    panic!("unexpected Remove action on insertion");
+                }
+            };
+            wnode.write_unlock();
+        }
+        return Ok(());
+    } else {
+        let next_child = next_node.unwrap(); // checked above it's not None
+        if let Some((ref rparent, _)) = rparent {
+            rparent.check_or_restart()?;
+        }
+
+        // recurse to next level
+        update_recurse(
+            &key[1..],
+            value_fn,
+            next_child,
+            Some((rnode, key[0])),
+            rparent,
+            guard,
+            level + 1,
+            orig_key,
+        )
+    }
+}
+
+#[derive(Clone)]
+enum PathElement {
+    Prefix(Vec<u8>),
+    KeyByte(u8),
+}
+
+impl std::fmt::Debug for PathElement {
+    fn fmt(&self, fmt: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> {
+        match self {
+            PathElement::Prefix(prefix) => write!(fmt, "{:?}", prefix),
+            PathElement::KeyByte(key_byte) => write!(fmt, "{}", key_byte),
+        }
+    }
+}
+
+pub(crate) fn dump_tree<'e, V: Value + std::fmt::Debug>(
+    root: RootPtr<V>,
+    epoch_pin: &'e EpochPin,
+    dst: &mut dyn std::io::Write,
+) {
+    let root_ref = NodeRef::from_root_ptr(root);
+
+    let _ = dump_recurse(&[], root_ref, &epoch_pin, 0, dst);
+}
+
+// TODO: return an Err if writeln!() returns error, instead of unwrapping
+fn dump_recurse<'e, V: Value + std::fmt::Debug>(
+    path: &[PathElement],
+    node: NodeRef<'e, V>,
+    epoch_pin: &'e EpochPin,
+    level: usize,
+    dst: &mut dyn std::io::Write,
+) -> Result<(), ConcurrentUpdateError> {
+    let indent = str::repeat(" ", level);
+
+    let rnode = node.read_lock_or_restart()?;
+    let mut path = Vec::from(path);
+    let prefix = rnode.get_prefix();
+    if prefix.len() != 0 {
+        path.push(PathElement::Prefix(Vec::from(prefix)));
+    }
+
+    if rnode.is_leaf() {
+        let vptr = rnode.get_leaf_value_ptr()?;
+        // safety: It's OK to return a ref of the pointer because we checked the version
+        // and the lifetime of 'epoch_pin' enforces that the reference is only accessible
+        // as long as the epoch is pinned.
+        let val = unsafe { vptr.as_ref().unwrap() };
+        writeln!(dst, "{} {:?}: {:?}", indent, path, val).unwrap();
+        return Ok(());
+    }
+
+    for key_byte in 0..=u8::MAX {
+        match rnode.find_child_or_restart(key_byte)? {
+            None => continue,
+            Some(child_ref) => {
+                let rchild = child_ref.read_lock_or_restart()?;
+                writeln!(
+                    dst,
+                    "{} {:?}, {}: prefix {:?}",
+                    indent,
+                    &path,
+                    key_byte,
+                    rchild.get_prefix()
+                )
+                .unwrap();
+
+                let mut child_path = path.clone();
+                child_path.push(PathElement::KeyByte(key_byte));
+
+                dump_recurse(&child_path, child_ref, epoch_pin, level + 1, dst)?;
+            }
+        }
+    }
+
+    Ok(())
+}
+
+///```text
+///        [fooba]r -> value
+///
+/// [foo]b -> [a]r  -> value
+///      e -> [ls]e -> value
+///```
+fn insert_split_prefix<'e, K: Key, V: Value, A: ArtAllocator<V>>(
+    key: &[u8],
+    value: V,
+    node: &mut WriteLockedNodeRef<V>,
+    parent: &mut WriteLockedNodeRef<V>,
+    parent_key: u8,
+    guard: &'e TreeWriteGuard<K, V, A>,
+) -> Result<(), OutOfMemoryError> {
+    let old_node = node;
+    let old_prefix = old_node.get_prefix();
+    let common_prefix_len = common_prefix(key, old_prefix);
+
+    // Allocate a node for the new value.
+    let new_value_node = allocate_node_for_value(
+        &key[common_prefix_len + 1..],
+        value,
+        guard.tree_writer.allocator,
+    )?;
+
+    // Allocate a new internal node with the common prefix
+    // FIXME: deallocate 'new_value_node' on OOM
+    let mut prefix_node =
+        node_ref::new_internal(&key[..common_prefix_len], guard.tree_writer.allocator)?;
+
+    // Add the old node and the new nodes to the new internal node
+    prefix_node.insert_old_child(old_prefix[common_prefix_len], old_node);
+    prefix_node.insert_new_child(key[common_prefix_len], new_value_node);
+
+    // Modify the prefix of the old child in place
+    old_node.truncate_prefix(old_prefix.len() - common_prefix_len - 1);
+
+    // replace the pointer in the parent
+    parent.replace_child(parent_key, prefix_node.into_ptr());
+
+    Ok(())
+}
+
+fn insert_to_node<'e, K: Key, V: Value, A: ArtAllocator<V>>(
+    wnode: &mut WriteLockedNodeRef<V>,
+    key: &[u8],
+    value: V,
+    guard: &'e TreeWriteGuard<K, V, A>,
+) -> Result<(), OutOfMemoryError> {
+    let value_child = allocate_node_for_value(&key[1..], value, guard.tree_writer.allocator)?;
+    wnode.insert_child(key[0], value_child.into_ptr());
+    Ok(())
+}
+
+// On entry: 'parent' and 'node' are locked
+fn insert_and_grow<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>>(
+    key: &[u8],
+    value: V,
+    wnode: WriteLockedNodeRef<V>,
+    parent: &mut WriteLockedNodeRef<V>,
+    parent_key_byte: u8,
+    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
+) -> Result<(), ArtError> {
+    let mut bigger_node = wnode.grow(guard.tree_writer.allocator)?;
+
+    // FIXME: deallocate 'bigger_node' on OOM
+    let value_child = allocate_node_for_value(&key[1..], value, guard.tree_writer.allocator)?;
+    bigger_node.insert_new_child(key[0], value_child);
+
+    // Replace the pointer in the parent
+    parent.replace_child(parent_key_byte, bigger_node.into_ptr());
+
+    guard.remember_obsolete_node(wnode.as_ptr());
+    wnode.write_unlock_obsolete();
+
+    Ok(())
+}
+
+fn cleanup_parent<'e, 'g, K: Key, V: Value, A: ArtAllocator<V>>(
+    wparent: WriteLockedNodeRef<V>,
+    rgrandparent: (ReadLockedNodeRef<V>, u8),
+    guard: &'g mut TreeWriteGuard<'e, K, V, A>,
+) -> Result<(), ArtError> {
+    let (rgrandparent, grandparent_key_byte) = rgrandparent;
+
+    // If the parent becomes completely empty after the deletion, remove the parent from the
+    // grandparent. (This case is possible because we reserve only 8 bytes for the prefix.)
+    // TODO: not implemented.
+
+    // If the parent has only one child, replace the parent with the remaining child. (This is not
+    // possible if the child's prefix field cannot absorb the parent's)
+    if wparent.num_children() == 1 {
+        // Try to lock the remaining child. This can fail if the child is updated
+        // concurrently.
+        let (key_byte, remaining_child) = wparent.find_remaining_child();
+
+        let mut wremaining_child = remaining_child.write_lock_or_restart()?;
+
+        if 1 + wremaining_child.get_prefix().len() + wparent.get_prefix().len() <= MAX_PREFIX_LEN {
+            let mut wgrandparent = rgrandparent.upgrade_to_write_lock_or_restart()?;
+
+            // Ok, we have locked the leaf, the parent, the grandparent, and the parent's only
+            // remaining leaf. Proceed with the updates.
+
+            // Update the prefix on the remaining leaf
+            wremaining_child.prepend_prefix(wparent.get_prefix(), key_byte);
+
+            // Replace the pointer in the grandparent to point directly to the remaining leaf
+            wgrandparent.replace_child(grandparent_key_byte, wremaining_child.as_ptr());
+
+            // Mark the parent as deleted.
+            guard.remember_obsolete_node(wparent.as_ptr());
+            wparent.write_unlock_obsolete();
+            return Ok(());
+        }
+    }
+
+    // If the parent's children would fit on a smaller node type after the deletion, replace it with
+    // a smaller node.
+    if wparent.can_shrink() {
+        let mut wgrandparent = rgrandparent.upgrade_to_write_lock_or_restart()?;
+        let smaller_node = wparent.shrink(guard.tree_writer.allocator)?;
+
+        // Replace the pointer in the grandparent
+        wgrandparent.replace_child(grandparent_key_byte, smaller_node.into_ptr());
+
+        guard.remember_obsolete_node(wparent.as_ptr());
+        wparent.write_unlock_obsolete();
+        return Ok(());
+    }
+
+    // nothing to do
+    wparent.write_unlock();
+    Ok(())
+}
+
+// Allocate a new leaf node to hold 'value'. If the key is long, we
+// may need to allocate new internal nodes to hold it too
+fn allocate_node_for_value<'a, V: Value, A: ArtAllocator<V>>(
+    key: &[u8],
+    value: V,
+    allocator: &'a A,
+) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError> {
+    let mut prefix_off = key.len().saturating_sub(MAX_PREFIX_LEN);
+
+    let leaf_node = node_ref::new_leaf(&key[prefix_off..key.len()], value, allocator)?;
+
+    let mut node = leaf_node;
+    while prefix_off > 0 {
+        // Need another internal node
+        let remain_prefix = &key[0..prefix_off];
+
+        prefix_off = remain_prefix.len().saturating_sub(MAX_PREFIX_LEN + 1);
+        let mut internal_node = node_ref::new_internal(
+            &remain_prefix[prefix_off..remain_prefix.len() - 1],
+            allocator,
+        )?;
+        internal_node.insert_new_child(*remain_prefix.last().unwrap(), node);
+        node = internal_node;
+    }
+
+    Ok(node)
+}
+
+fn common_prefix(a: &[u8], b: &[u8]) -> usize {
+    for i in 0..MAX_PREFIX_LEN {
+        if a[i] != b[i] {
+            return i;
+        }
+    }
+    panic!("prefixes are equal");
+}

libs/neonart/src/algorithm/lock_and_version.rs

@@ -0,0 +1,117 @@
+//! Each node in the tree has contains one atomic word that stores three things:
+//!
+//! Bit 0: set if the node is "obsolete". An obsolete node has been removed from the tree,
+//!        but might still be accessed by concurrent readers until the epoch expires.
+//! Bit 1: set if the node is currently write-locked. Used as a spinlock.
+//! Bits 2-63: Version number, incremented every time the node is modified.
+//!
+//! AtomicLockAndVersion represents that.
+
+use std::sync::atomic::{AtomicU64, Ordering};
+
+pub(crate) struct ConcurrentUpdateError();
+
+pub(crate) struct AtomicLockAndVersion {
+    inner: AtomicU64,
+}
+
+impl AtomicLockAndVersion {
+    pub(crate) fn new() -> AtomicLockAndVersion {
+        AtomicLockAndVersion {
+            inner: AtomicU64::new(0),
+        }
+    }
+}
+
+impl AtomicLockAndVersion {
+    pub(crate) fn read_lock_or_restart(&self) -> Result<u64, ConcurrentUpdateError> {
+        let version = self.await_node_unlocked();
+        if is_obsolete(version) {
+            return Err(ConcurrentUpdateError());
+        }
+        Ok(version)
+    }
+
+    pub(crate) fn check_or_restart(&self, version: u64) -> Result<(), ConcurrentUpdateError> {
+        self.read_unlock_or_restart(version)
+    }
+
+    pub(crate) fn read_unlock_or_restart(&self, version: u64) -> Result<(), ConcurrentUpdateError> {
+        if self.inner.load(Ordering::Acquire) != version {
+            return Err(ConcurrentUpdateError());
+        }
+        Ok(())
+    }
+
+    pub(crate) fn upgrade_to_write_lock_or_restart(
+        &self,
+        version: u64,
+    ) -> Result<(), ConcurrentUpdateError> {
+        if self
+            .inner
+            .compare_exchange(
+                version,
+                set_locked_bit(version),
+                Ordering::Acquire,
+                Ordering::Relaxed,
+            )
+            .is_err()
+        {
+            return Err(ConcurrentUpdateError());
+        }
+        Ok(())
+    }
+
+    pub(crate) fn write_lock_or_restart(&self) -> Result<(), ConcurrentUpdateError> {
+        let old = self.inner.load(Ordering::Relaxed);
+        if is_obsolete(old) || is_locked(old) {
+            return Err(ConcurrentUpdateError());
+        }
+        if self
+            .inner
+            .compare_exchange(
+                old,
+                set_locked_bit(old),
+                Ordering::Acquire,
+                Ordering::Relaxed,
+            )
+            .is_err()
+        {
+            return Err(ConcurrentUpdateError());
+        }
+        Ok(())
+    }
+
+    pub(crate) fn write_unlock(&self) {
+        // reset locked bit and overflow into version
+        self.inner.fetch_add(2, Ordering::Release);
+    }
+
+    pub(crate) fn write_unlock_obsolete(&self) {
+        // set obsolete, reset locked, overflow into version
+        self.inner.fetch_add(3, Ordering::Release);
+    }
+
+    // Helper functions
+    fn await_node_unlocked(&self) -> u64 {
+        let mut version = self.inner.load(Ordering::Acquire);
+        while is_locked(version) {
+            // spinlock
+            std::thread::yield_now();
+            version = self.inner.load(Ordering::Acquire)
+        }
+        version
+    }
+}
+
+fn set_locked_bit(version: u64) -> u64 {
+    return version + 2;
+}
+
+fn is_obsolete(version: u64) -> bool {
+    return (version & 1) == 1;
+}
+
+fn is_locked(version: u64) -> bool {
+    return (version & 2) == 2;
+}

libs/neonart/src/algorithm/node_ref.rs

@@ -0,0 +1,349 @@
+use std::fmt::Debug;
+use std::marker::PhantomData;
+
+use super::node_ptr;
+use super::node_ptr::NodePtr;
+use crate::EpochPin;
+use crate::Value;
+use crate::algorithm::lock_and_version::AtomicLockAndVersion;
+use crate::algorithm::lock_and_version::ConcurrentUpdateError;
+use crate::allocator::ArtAllocator;
+use crate::allocator::OutOfMemoryError;
+
+pub struct NodeRef<'e, V> {
+    ptr: NodePtr<V>,
+
+    phantom: PhantomData<&'e EpochPin<'e>>,
+}
+
+impl<'e, V> Debug for NodeRef<'e, V> {
+    fn fmt(&self, fmt: &mut std::fmt::Formatter<'_>) -> Result<(), std::fmt::Error> {
+        write!(fmt, "{:?}", self.ptr)
+    }
+}
+
+impl<'e, V: Value> NodeRef<'e, V> {
+    pub(crate) fn from_root_ptr(root_ptr: NodePtr<V>) -> NodeRef<'e, V> {
+        NodeRef {
+            ptr: root_ptr,
+            phantom: PhantomData,
+        }
+    }
+
+    pub(crate) fn read_lock_or_restart(
+        &self,
+    ) -> Result<ReadLockedNodeRef<'e, V>, ConcurrentUpdateError> {
+        let version = self.lockword().read_lock_or_restart()?;
+        Ok(ReadLockedNodeRef {
+            ptr: self.ptr,
+            version,
+            phantom: self.phantom,
+        })
+    }
+
+    pub(crate) fn write_lock_or_restart(
+        &self,
+    ) -> Result<WriteLockedNodeRef<'e, V>, ConcurrentUpdateError> {
+        self.lockword().write_lock_or_restart()?;
+        Ok(WriteLockedNodeRef {
+            ptr: self.ptr,
+            phantom: self.phantom,
+        })
+    }
+
+    fn lockword(&self) -> &AtomicLockAndVersion {
+        self.ptr.lockword()
+    }
+}
+
+/// A reference to a node that has been optimistically read-locked. The functions re-check
+/// the version after each read.
+pub struct ReadLockedNodeRef<'e, V> {
+    ptr: NodePtr<V>,
+    version: u64,
+
+    phantom: PhantomData<&'e EpochPin<'e>>,
+}
+
+impl<'e, V: Value> ReadLockedNodeRef<'e, V> {
+    pub(crate) fn is_leaf(&self) -> bool {
+        self.ptr.is_leaf()
+    }
+
+    pub(crate) fn is_full(&self) -> bool {
+        self.ptr.is_full()
+    }
+
+    pub(crate) fn get_prefix(&self) -> &[u8] {
+        self.ptr.get_prefix()
+    }
+
+    /// Note: because we're only holding a read lock, the prefix can change concurrently.
+    /// You must be prepared to restart, if read_unlock() returns error later.
+    ///
+    /// Returns the length of the prefix, or None if it's not a match
+    pub(crate) fn prefix_matches(&self, key: &[u8]) -> Option<usize> {
+        self.ptr.prefix_matches(key)
+    }
+
+    pub(crate) fn find_child_or_restart(
+        &self,
+        key_byte: u8,
+    ) -> Result<Option<NodeRef<'e, V>>, ConcurrentUpdateError> {
+        let child_or_value = self.ptr.find_child(key_byte);
+        self.ptr.lockword().check_or_restart(self.version)?;
+
+        match child_or_value {
+            None => Ok(None),
+            Some(child_ptr) => Ok(Some(NodeRef {
+                ptr: child_ptr,
+                phantom: self.phantom,
+            })),
+        }
+    }
+
+    pub(crate) fn find_next_child_or_restart(
+        &self,
+        min_key_byte: u8,
+    ) -> Result<Option<(u8, NodeRef<'e, V>)>, ConcurrentUpdateError> {
+        let child_or_value = self.ptr.find_next_child(min_key_byte);
+        self.ptr.lockword().check_or_restart(self.version)?;
+
+        match child_or_value {
+            None => Ok(None),
+            Some((k, child_ptr)) => Ok(Some((
+                k,
+                NodeRef {
+                    ptr: child_ptr,
+                    phantom: self.phantom,
+                },
+            ))),
+        }
+    }
+
+    pub(crate) fn get_leaf_value_ptr(&self) -> Result<*const V, ConcurrentUpdateError> {
+        let result = self.ptr.get_leaf_value();
+        self.ptr.lockword().check_or_restart(self.version)?;
+
+        // Extend the lifetime.
+        let result = std::ptr::from_ref(result);
+
+        Ok(result)
+    }
+
+    pub(crate) fn upgrade_to_write_lock_or_restart(
+        self,
+    ) -> Result<WriteLockedNodeRef<'e, V>, ConcurrentUpdateError> {
+        self.ptr
+            .lockword()
+            .upgrade_to_write_lock_or_restart(self.version)?;
+
+        Ok(WriteLockedNodeRef {
+            ptr: self.ptr,
+            phantom: self.phantom,
+        })
+    }
+
+    pub(crate) fn read_unlock_or_restart(self) -> Result<(), ConcurrentUpdateError> {
+        self.ptr.lockword().check_or_restart(self.version)?;
+        Ok(())
+    }
+
+    pub(crate) fn check_or_restart(&self) -> Result<(), ConcurrentUpdateError> {
+        self.ptr.lockword().check_or_restart(self.version)?;
+        Ok(())
+    }
+}
+
+/// A reference to a node that has been optimistically read-locked. The functions re-check
+/// the version after each read.
+pub struct WriteLockedNodeRef<'e, V> {
+    ptr: NodePtr<V>,
+    phantom: PhantomData<&'e EpochPin<'e>>,
+}
+
+impl<'e, V: Value> WriteLockedNodeRef<'e, V> {
+    pub(crate) fn can_shrink(&self) -> bool {
+        self.ptr.can_shrink()
+    }
+
+    pub(crate) fn num_children(&self) -> usize {
+        self.ptr.num_children()
+    }
+
+    pub(crate) fn write_unlock(mut self) {
+        self.ptr.lockword().write_unlock();
+        self.ptr = NodePtr::null();
+    }
+
+    pub(crate) fn write_unlock_obsolete(mut self) {
+        self.ptr.lockword().write_unlock_obsolete();
+        self.ptr = NodePtr::null();
+    }
+
+    pub(crate) fn get_prefix(&self) -> &[u8] {
+        self.ptr.get_prefix()
+    }
+
+    pub(crate) fn truncate_prefix(&mut self, new_prefix_len: usize) {
+        self.ptr.truncate_prefix(new_prefix_len)
+    }
+
+    pub(crate) fn prepend_prefix(&mut self, prefix: &[u8], prefix_byte: u8) {
+        self.ptr.prepend_prefix(prefix, prefix_byte)
+    }
+
+    pub(crate) fn insert_child(&mut self, key_byte: u8, child: NodePtr<V>) {
+        self.ptr.insert_child(key_byte, child)
+    }
+
+    pub(crate) fn get_leaf_value_mut(&mut self) -> &mut V {
+        self.ptr.get_leaf_value_mut()
+    }
+
+    pub(crate) fn grow<'a, A>(
+        &self,
+        allocator: &'a A,
+    ) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
+    where
+        A: ArtAllocator<V>,
+    {
+        let new_node = self.ptr.grow(allocator)?;
+        Ok(NewNodeRef {
+            ptr: new_node,
+            allocator,
+            extra_nodes: Vec::new(),
+        })
+    }
+
+    pub(crate) fn shrink<'a, A>(
+        &self,
+        allocator: &'a A,
+    ) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
+    where
+        A: ArtAllocator<V>,
+    {
+        let new_node = self.ptr.shrink(allocator)?;
+        Ok(NewNodeRef {
+            ptr: new_node,
+            allocator,
+            extra_nodes: Vec::new(),
+        })
+    }
+
+    pub(crate) fn as_ptr(&self) -> NodePtr<V> {
+        self.ptr
+    }
+
+    pub(crate) fn replace_child(&mut self, key_byte: u8, replacement: NodePtr<V>) {
+        self.ptr.replace_child(key_byte, replacement);
+    }
+
+    pub(crate) fn delete_child(&mut self, key_byte: u8) {
+        self.ptr.delete_child(key_byte);
+    }
+
+    pub(crate) fn find_remaining_child(&self) -> (u8, NodeRef<'e, V>) {
+        assert_eq!(self.num_children(), 1);
+        let child_or_value = self.ptr.find_next_child(0);
+
+        match child_or_value {
+            None => panic!("could not find only child in node"),
+            Some((k, child_ptr)) => (
+                k,
+                NodeRef {
+                    ptr: child_ptr,
+                    phantom: self.phantom,
+                },
+            ),
+        }
+    }
+}
+
+impl<'e, V> Drop for WriteLockedNodeRef<'e, V> {
+    fn drop(&mut self) {
+        if !self.ptr.is_null() {
+            self.ptr.lockword().write_unlock();
+        }
+    }
+}
+
+pub(crate) struct NewNodeRef<'a, V, A>
+where
+    V: Value,
+    A: ArtAllocator<V>,
+{
+    ptr: NodePtr<V>,
+    allocator: &'a A,
+
+    extra_nodes: Vec<NodePtr<V>>,
+}
+
+impl<'a, V, A> NewNodeRef<'a, V, A>
+where
+    V: Value,
+    A: ArtAllocator<V>,
+{
+    pub(crate) fn insert_old_child(&mut self, key_byte: u8, child: &WriteLockedNodeRef<V>) {
+        self.ptr.insert_child(key_byte, child.as_ptr())
+    }
+
+    pub(crate) fn into_ptr(mut self) -> NodePtr<V> {
+        let ptr = self.ptr;
+        self.ptr = NodePtr::null();
+        ptr
+    }
+
+    pub(crate) fn insert_new_child(&mut self, key_byte: u8, child: NewNodeRef<'a, V, A>) {
+        let child_ptr = child.into_ptr();
+        self.ptr.insert_child(key_byte, child_ptr);
+        self.extra_nodes.push(child_ptr);
+    }
+}
+
+impl<'a, V, A> Drop for NewNodeRef<'a, V, A>
+where
+    V: Value,
+    A: ArtAllocator<V>,
+{
+    /// This drop implementation deallocates the newly allocated node, if into_ptr() was not called.
+    fn drop(&mut self) {
+        if !self.ptr.is_null() {
+            self.ptr.deallocate(self.allocator);
+            for p in self.extra_nodes.iter() {
+                p.deallocate(self.allocator);
+            }
+        }
+    }
+}
+
+pub(crate) fn new_internal<'a, V, A>(
+    prefix: &[u8],
+    allocator: &'a A,
+) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
+where
+    V: Value,
+    A: ArtAllocator<V>,
+{
+    Ok(NewNodeRef {
+        ptr: node_ptr::new_internal(prefix, allocator)?,
+        allocator,
+        extra_nodes: Vec::new(),
+    })
+}
+
+pub(crate) fn new_leaf<'a, V, A>(
+    prefix: &[u8],
+    value: V,
+    allocator: &'a A,
+) -> Result<NewNodeRef<'a, V, A>, OutOfMemoryError>
+where
+    V: Value,
+    A: ArtAllocator<V>,
+{
+    Ok(NewNodeRef {
+        ptr: node_ptr::new_leaf(prefix, value, allocator)?,
+        allocator,
+        extra_nodes: Vec::new(),
+    })
+}

libs/neonart/src/allocator.rs

@@ -0,0 +1,158 @@
+pub mod block;
+mod multislab;
+mod slab;
+pub mod r#static;
+
+use std::alloc::Layout;
+use std::marker::PhantomData;
+use std::mem::MaybeUninit;
+use std::sync::atomic::Ordering;
+
+use crate::allocator::multislab::MultiSlabAllocator;
+use crate::allocator::r#static::alloc_from_slice;
+
+use spin;
+
+use crate::Tree;
+pub use crate::algorithm::node_ptr::{
+    NodeInternal4, NodeInternal16, NodeInternal48, NodeInternal256, NodeLeaf,
+};
+
+#[derive(Debug)]
+pub struct OutOfMemoryError();
+
+pub trait ArtAllocator<V: crate::Value> {
+    fn alloc_tree(&self) -> *mut Tree<V>;
+
+    fn alloc_node_internal4(&self) -> *mut NodeInternal4<V>;
+    fn alloc_node_internal16(&self) -> *mut NodeInternal16<V>;
+    fn alloc_node_internal48(&self) -> *mut NodeInternal48<V>;
+    fn alloc_node_internal256(&self) -> *mut NodeInternal256<V>;
+    fn alloc_node_leaf(&self) -> *mut NodeLeaf<V>;
+
+    fn dealloc_node_internal4(&self, ptr: *mut NodeInternal4<V>);
+    fn dealloc_node_internal16(&self, ptr: *mut NodeInternal16<V>);
+    fn dealloc_node_internal48(&self, ptr: *mut NodeInternal48<V>);
+    fn dealloc_node_internal256(&self, ptr: *mut NodeInternal256<V>);
+    fn dealloc_node_leaf(&self, ptr: *mut NodeLeaf<V>);
+}
+
+pub struct ArtMultiSlabAllocator<'t, V>
+where
+    V: crate::Value,
+{
+    tree_area: spin::Mutex<Option<&'t mut MaybeUninit<Tree<V>>>>,
+
+    pub(crate) inner: MultiSlabAllocator<'t, 5>,
+
+    phantom_val: PhantomData<V>,
+}
+
+impl<'t, V: crate::Value> ArtMultiSlabAllocator<'t, V> {
+    const LAYOUTS: [Layout; 5] = [
+        Layout::new::<NodeInternal4<V>>(),
+        Layout::new::<NodeInternal16<V>>(),
+        Layout::new::<NodeInternal48<V>>(),
+        Layout::new::<NodeInternal256<V>>(),
+        Layout::new::<NodeLeaf<V>>(),
+    ];
+
+    pub fn new(area: &'t mut [MaybeUninit<u8>]) -> &'t mut ArtMultiSlabAllocator<'t, V> {
+        let (allocator_area, remain) = alloc_from_slice::<ArtMultiSlabAllocator<V>>(area);
+        let (tree_area, remain) = alloc_from_slice::<Tree<V>>(remain);
+
+        let allocator = allocator_area.write(ArtMultiSlabAllocator {
+            tree_area: spin::Mutex::new(Some(tree_area)),
+            inner: MultiSlabAllocator::new(remain, &Self::LAYOUTS),
+            phantom_val: PhantomData,
+        });
+
+        allocator
+    }
+}
+
+impl<'t, V: crate::Value> ArtAllocator<V> for ArtMultiSlabAllocator<'t, V> {
+    fn alloc_tree(&self) -> *mut Tree<V> {
+        let mut t = self.tree_area.lock();
+        if let Some(tree_area) = t.take() {
+            return tree_area.as_mut_ptr().cast();
+        }
+        panic!("cannot allocate more than one tree");
+    }
+
+    fn alloc_node_internal4(&self) -> *mut NodeInternal4<V> {
+        self.inner.alloc_slab(0).cast()
+    }
+    fn alloc_node_internal16(&self) -> *mut NodeInternal16<V> {
+        self.inner.alloc_slab(1).cast()
+    }
+    fn alloc_node_internal48(&self) -> *mut NodeInternal48<V> {
+        self.inner.alloc_slab(2).cast()
+    }
+    fn alloc_node_internal256(&self) -> *mut NodeInternal256<V> {
+        self.inner.alloc_slab(3).cast()
+    }
+    fn alloc_node_leaf(&self) -> *mut NodeLeaf<V> {
+        self.inner.alloc_slab(4).cast()
+    }
+
+    fn dealloc_node_internal4(&self, ptr: *mut NodeInternal4<V>) {
+        self.inner.dealloc_slab(0, ptr.cast())
+    }
+
+    fn dealloc_node_internal16(&self, ptr: *mut NodeInternal16<V>) {
+        self.inner.dealloc_slab(1, ptr.cast())
+    }
+    fn dealloc_node_internal48(&self, ptr: *mut NodeInternal48<V>) {
+        self.inner.dealloc_slab(2, ptr.cast())
+    }
+    fn dealloc_node_internal256(&self, ptr: *mut NodeInternal256<V>) {
+        self.inner.dealloc_slab(3, ptr.cast())
+    }
+    fn dealloc_node_leaf(&self, ptr: *mut NodeLeaf<V>) {
+        self.inner.dealloc_slab(4, ptr.cast())
+    }
+}
+
+impl<'t, V: crate::Value> ArtMultiSlabAllocator<'t, V> {
+    pub(crate) fn get_statistics(&self) -> ArtMultiSlabStats {
+        ArtMultiSlabStats {
+            num_internal4: self.inner.slab_descs[0]
+                .num_allocated
+                .load(Ordering::Relaxed),
+            num_internal16: self.inner.slab_descs[1]
+                .num_allocated
+                .load(Ordering::Relaxed),
+            num_internal48: self.inner.slab_descs[2]
+                .num_allocated
+                .load(Ordering::Relaxed),
+            num_internal256: self.inner.slab_descs[3]
+                .num_allocated
+                .load(Ordering::Relaxed),
+            num_leaf: self.inner.slab_descs[4]
+                .num_allocated
+                .load(Ordering::Relaxed),
+
+            num_blocks_internal4: self.inner.slab_descs[0].num_blocks.load(Ordering::Relaxed),
+            num_blocks_internal16: self.inner.slab_descs[1].num_blocks.load(Ordering::Relaxed),
+            num_blocks_internal48: self.inner.slab_descs[2].num_blocks.load(Ordering::Relaxed),
+            num_blocks_internal256: self.inner.slab_descs[3].num_blocks.load(Ordering::Relaxed),
+            num_blocks_leaf: self.inner.slab_descs[4].num_blocks.load(Ordering::Relaxed),
+        }
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct ArtMultiSlabStats {
+    pub num_internal4: u64,
+    pub num_internal16: u64,
+    pub num_internal48: u64,
+    pub num_internal256: u64,
+    pub num_leaf: u64,
+
+    pub num_blocks_internal4: u64,
+    pub num_blocks_internal16: u64,
+    pub num_blocks_internal48: u64,
+    pub num_blocks_internal256: u64,
+    pub num_blocks_leaf: u64,
+}

libs/neonart/src/allocator/block.rs

@@ -0,0 +1,191 @@
+//! Simple allocator of fixed-size blocks
+
+use std::mem::MaybeUninit;
+use std::sync::atomic::{AtomicU64, Ordering};
+
+use spin;
+
+pub const BLOCK_SIZE: usize = 16 * 1024;
+
+const INVALID_BLOCK: u64 = u64::MAX;
+
+pub(crate) struct BlockAllocator<'t> {
+    blocks_ptr: &'t [MaybeUninit<u8>],
+    num_blocks: u64,
+    num_initialized: AtomicU64,
+
+    freelist_head: spin::Mutex<u64>,
+}
+
+struct FreeListBlock {
+    inner: spin::Mutex<FreeListBlockInner>,
+}
+
+struct FreeListBlockInner {
+    next: u64,
+
+    num_free_blocks: u64,
+    free_blocks: [u64; 100], // FIXME: fill the rest of the block
+}
+
+impl<'t> BlockAllocator<'t> {
+    pub(crate) fn new(area: &'t mut [MaybeUninit<u8>]) -> Self {
+        // Use all the space for the blocks
+        let padding = area.as_ptr().align_offset(BLOCK_SIZE);
+        let remain = &mut area[padding..];
+
+        let num_blocks = (remain.len() / BLOCK_SIZE) as u64;
+
+        BlockAllocator {
+            blocks_ptr: remain,
+            num_blocks,
+            num_initialized: AtomicU64::new(0),
+            freelist_head: spin::Mutex::new(INVALID_BLOCK),
+        }
+    }
+
+    /// safety: you must hold a lock on the pointer to this block, otherwise it might get
+    /// reused for another kind of block
+    fn read_freelist_block(&self, blkno: u64) -> &FreeListBlock {
+        let ptr: *const FreeListBlock = self.get_block_ptr(blkno).cast();
+        unsafe { ptr.as_ref().unwrap() }
+    }
+
+    fn get_block_ptr(&self, blkno: u64) -> *mut u8 {
+        assert!(blkno < self.num_blocks);
+        unsafe {
+            self.blocks_ptr
+                .as_ptr()
+                .byte_offset(blkno as isize * BLOCK_SIZE as isize)
+        }
+        .cast_mut()
+        .cast()
+    }
+
+    #[allow(clippy::mut_from_ref)]
+    pub(crate) fn alloc_block(&self) -> &mut [MaybeUninit<u8>] {
+        // FIXME: handle OOM
+        let blkno = self.alloc_block_internal();
+        if blkno == INVALID_BLOCK {
+            panic!("out of memory");
+        }
+
+        let ptr: *mut MaybeUninit<u8> = self.get_block_ptr(blkno).cast();
+        unsafe { std::slice::from_raw_parts_mut(ptr, BLOCK_SIZE) }
+    }
+
+    fn alloc_block_internal(&self) -> u64 {
+        //  check the free list.
+        {
+            let mut freelist_head = self.freelist_head.lock();
+            if *freelist_head != INVALID_BLOCK {
+                let freelist_block = self.read_freelist_block(*freelist_head);
+
+                // acquire lock on the freelist block before releasing the lock on the parent (i.e. lock coupling)
+                let mut g = freelist_block.inner.lock();
+
+                if g.num_free_blocks > 0 {
+                    g.num_free_blocks -= 1;
+                    let result = g.free_blocks[g.num_free_blocks as usize];
+                    return result;
+                } else {
+                    // consume the freelist block itself
+                    let result = *freelist_head;
+                    *freelist_head = g.next;
+                    // This freelist block is now unlinked and can be repurposed
+                    drop(g);
+                    return result;
+                }
+            }
+        }
+
+        // If there are some blocks left that we've never used, pick next such block
+        let mut next_uninitialized = self.num_initialized.load(Ordering::Relaxed);
+        while next_uninitialized < self.num_blocks {
+            match self.num_initialized.compare_exchange(
+                next_uninitialized,
+                next_uninitialized + 1,
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            ) {
+                Ok(_) => {
+                    return next_uninitialized;
+                }
+                Err(old) => {
+                    next_uninitialized = old;
+                    continue;
+                }
+            }
+        }
+
+        // out of blocks
+        return INVALID_BLOCK;
+    }
+
+    // TODO: this is currently unused. The slab allocator never releases blocks
+    #[allow(dead_code)]
+    pub(crate) fn release_block(&self, block_ptr: *mut u8) {
+        let blockno = unsafe { block_ptr.byte_offset_from(self.blocks_ptr) / BLOCK_SIZE as isize };
+        self.release_block_internal(blockno as u64);
+    }
+
+    fn release_block_internal(&self, blockno: u64) {
+        let mut freelist_head = self.freelist_head.lock();
+        if *freelist_head != INVALID_BLOCK {
+            let freelist_block = self.read_freelist_block(*freelist_head);
+
+            // acquire lock on the freelist block before releasing the lock on the parent (i.e. lock coupling)
+            let mut g = freelist_block.inner.lock();
+
+            let num_free_blocks = g.num_free_blocks;
+            if num_free_blocks < g.free_blocks.len() as u64 {
+                g.free_blocks[num_free_blocks as usize] = blockno;
+                g.num_free_blocks += 1;
+                return;
+            }
+        }
+
+        // Convert the block into a new freelist block
+        let block_ptr: *mut FreeListBlock = self.get_block_ptr(blockno).cast();
+        let init = FreeListBlock {
+            inner: spin::Mutex::new(FreeListBlockInner {
+                next: *freelist_head,
+                num_free_blocks: 0,
+                free_blocks: [INVALID_BLOCK; 100],
+            }),
+        };
+        unsafe { (*block_ptr) = init };
+        *freelist_head = blockno;
+    }
+
+    // for debugging
+    pub(crate) fn get_statistics(&self) -> BlockAllocatorStats {
+        let mut num_free_blocks = 0;
+
+        let mut _prev_lock = None;
+        let head_lock = self.freelist_head.lock();
+        let mut next_blk = *head_lock;
+        let mut _head_lock = Some(head_lock);
+        while next_blk != INVALID_BLOCK {
+            let freelist_block = self.read_freelist_block(next_blk);
+            let lock = freelist_block.inner.lock();
+            num_free_blocks += lock.num_free_blocks;
+            next_blk = lock.next;
+            _prev_lock = Some(lock); // hold the lock until we've read the next block
+            _head_lock = None;
+        }
+
+        BlockAllocatorStats {
+            num_blocks: self.num_blocks,
+            num_initialized: self.num_initialized.load(Ordering::Relaxed),
+            num_free_blocks,
+        }
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct BlockAllocatorStats {
+    pub num_blocks: u64,
+    pub num_initialized: u64,
+    pub num_free_blocks: u64,
+}

libs/neonart/src/allocator/multislab.rs

@@ -0,0 +1,33 @@
+use std::alloc::Layout;
+use std::mem::MaybeUninit;
+
+use crate::allocator::block::BlockAllocator;
+use crate::allocator::slab::SlabDesc;
+
+pub struct MultiSlabAllocator<'t, const N: usize> {
+    pub(crate) block_allocator: BlockAllocator<'t>,
+
+    pub(crate) slab_descs: [SlabDesc; N],
+}
+
+impl<'t, const N: usize> MultiSlabAllocator<'t, N> {
+    pub(crate) fn new(
+        area: &'t mut [MaybeUninit<u8>],
+        layouts: &[Layout; N],
+    ) -> MultiSlabAllocator<'t, N> {
+        let block_allocator = BlockAllocator::new(area);
+        MultiSlabAllocator {
+            block_allocator,
+
+            slab_descs: std::array::from_fn(|i| SlabDesc::new(&layouts[i])),
+        }
+    }
+
+    pub(crate) fn alloc_slab(&self, slab_idx: usize) -> *mut u8 {
+        self.slab_descs[slab_idx].alloc_chunk(&self.block_allocator)
+    }
+
+    pub(crate) fn dealloc_slab(&self, slab_idx: usize, ptr: *mut u8) {
+        self.slab_descs[slab_idx].dealloc_chunk(ptr, &self.block_allocator)
+    }
+}

libs/neonart/src/allocator/slab.rs

@@ -0,0 +1,432 @@
+//! A slab allocator that carves out fixed-size chunks from larger blocks.
+//!
+//!
+
+use std::alloc::Layout;
+use std::mem::MaybeUninit;
+use std::ops::Deref;
+use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
+
+use spin;
+
+use super::alloc_from_slice;
+use super::block::BlockAllocator;
+
+use crate::allocator::block::BLOCK_SIZE;
+
+pub(crate) struct SlabDesc {
+    pub(crate) layout: Layout,
+
+    block_lists: spin::RwLock<BlockLists>,
+
+    pub(crate) num_blocks: AtomicU64,
+    pub(crate) num_allocated: AtomicU64,
+}
+
+// FIXME: Not sure if SlabDesc is really Sync or Send. It probably is when it's empty, but
+// 'block_lists' contains pointers when it's not empty. In the current use as part of the
+// the art tree, SlabDescs are only moved during initialization.
+unsafe impl Sync for SlabDesc {}
+unsafe impl Send for SlabDesc {}
+
+#[derive(Default, Debug)]
+struct BlockLists {
+    full_blocks: BlockList,
+    nonfull_blocks: BlockList,
+}
+
+impl BlockLists {
+    // Unlink a node. It must be in either one of the two lists.
+    unsafe fn unlink(&mut self, elem: *mut SlabBlockHeader) {
+        let list = unsafe {
+            if (*elem).next.is_null() {
+                if self.full_blocks.tail == elem {
+                    Some(&mut self.full_blocks)
+                } else {
+                    Some(&mut self.nonfull_blocks)
+                }
+            } else if (*elem).prev.is_null() {
+                if self.full_blocks.head == elem {
+                    Some(&mut self.full_blocks)
+                } else {
+                    Some(&mut self.nonfull_blocks)
+                }
+            } else {
+                None
+            }
+        };
+        unsafe { unlink_slab_block(list, elem) };
+    }
+}
+
+unsafe fn unlink_slab_block(mut list: Option<&mut BlockList>, elem: *mut SlabBlockHeader) {
+    unsafe {
+        if (*elem).next.is_null() {
+            assert_eq!(list.as_ref().unwrap().tail, elem);
+            list.as_mut().unwrap().tail = (*elem).prev;
+        } else {
+            assert_eq!((*(*elem).next).prev, elem);
+            (*(*elem).next).prev = (*elem).prev;
+        }
+        if (*elem).prev.is_null() {
+            assert_eq!(list.as_ref().unwrap().head, elem);
+            list.as_mut().unwrap().head = (*elem).next;
+        } else {
+            assert_eq!((*(*elem).prev).next, elem);
+            (*(*elem).prev).next = (*elem).next;
+        }
+    }
+}
+
+#[derive(Debug)]
+struct BlockList {
+    head: *mut SlabBlockHeader,
+    tail: *mut SlabBlockHeader,
+}
+
+impl Default for BlockList {
+    fn default() -> Self {
+        BlockList {
+            head: std::ptr::null_mut(),
+            tail: std::ptr::null_mut(),
+        }
+    }
+}
+
+impl BlockList {
+    unsafe fn push_head(&mut self, elem: *mut SlabBlockHeader) {
+        unsafe {
+            if self.is_empty() {
+                self.tail = elem;
+                (*elem).next = std::ptr::null_mut();
+            } else {
+                (*elem).next = self.head;
+                (*self.head).prev = elem;
+            }
+            (*elem).prev = std::ptr::null_mut();
+            self.head = elem;
+        }
+    }
+
+    fn is_empty(&self) -> bool {
+        self.head.is_null()
+    }
+
+    unsafe fn unlink(&mut self, elem: *mut SlabBlockHeader) {
+        unsafe { unlink_slab_block(Some(self), elem) }
+    }
+
+    #[cfg(test)]
+    fn dump(&self) {
+        let mut next = self.head;
+
+        while !next.is_null() {
+            let n = unsafe { next.as_ref() }.unwrap();
+            eprintln!(
+                "  blk {:?} (free {}/{})",
+                next,
+                n.num_free_chunks.load(Ordering::Relaxed),
+                n.num_chunks
+            );
+            next = n.next;
+        }
+    }
+}
+
+impl SlabDesc {
+    pub(crate) fn new(layout: &Layout) -> SlabDesc {
+        SlabDesc {
+            layout: *layout,
+            block_lists: spin::RwLock::new(BlockLists::default()),
+            num_allocated: AtomicU64::new(0),
+            num_blocks: AtomicU64::new(0),
+        }
+    }
+}
+
+#[derive(Debug)]
+struct SlabBlockHeader {
+    free_chunks_head: spin::Mutex<*mut FreeChunk>,
+    num_free_chunks: AtomicU32,
+    num_chunks: u32, // this is really a constant for a given Layout
+
+    // these fields are protected by the lock on the BlockLists
+    prev: *mut SlabBlockHeader,
+    next: *mut SlabBlockHeader,
+}
+
+struct FreeChunk {
+    next: *mut FreeChunk,
+}
+
+enum ReadOrWriteGuard<'a, T> {
+    Read(spin::RwLockReadGuard<'a, T>),
+    Write(spin::RwLockWriteGuard<'a, T>),
+}
+
+impl<'a, T> Deref for ReadOrWriteGuard<'a, T> {
+    type Target = T;
+
+    fn deref(&self) -> &<Self as Deref>::Target {
+        match self {
+            ReadOrWriteGuard::Read(g) => g.deref(),
+            ReadOrWriteGuard::Write(g) => g.deref(),
+        }
+    }
+}
+
+impl SlabDesc {
+    pub fn alloc_chunk(&self, block_allocator: &BlockAllocator) -> *mut u8 {
+        // Are there any free chunks?
+        let mut acquire_write = false;
+        'outer: loop {
+            let mut block_lists_guard = if acquire_write {
+                ReadOrWriteGuard::Write(self.block_lists.write())
+            } else {
+                ReadOrWriteGuard::Read(self.block_lists.read())
+            };
+            'inner: loop {
+                let block_ptr = block_lists_guard.nonfull_blocks.head;
+                if block_ptr.is_null() {
+                    break 'outer;
+                }
+                unsafe {
+                    let mut free_chunks_head = (*block_ptr).free_chunks_head.lock();
+                    if !(*free_chunks_head).is_null() {
+                        let result = *free_chunks_head;
+                        (*free_chunks_head) = (*result).next;
+                        let _old = (*block_ptr).num_free_chunks.fetch_sub(1, Ordering::Relaxed);
+
+                        self.num_allocated.fetch_add(1, Ordering::Relaxed);
+                        return result.cast();
+                    }
+                }
+
+                // The block at the head of the list was full. Grab write lock and retry
+                match block_lists_guard {
+                    ReadOrWriteGuard::Read(_) => {
+                        acquire_write = true;
+                        continue 'outer;
+                    }
+                    ReadOrWriteGuard::Write(ref mut g) => {
+                        // move the node to the list of full blocks
+                        unsafe {
+                            g.nonfull_blocks.unlink(block_ptr);
+                            g.full_blocks.push_head(block_ptr);
+                        };
+                        continue 'inner;
+                    }
+                }
+            }
+        }
+
+        // no free chunks. Allocate a new block (and the chunk from that)
+        let (new_block, new_chunk) = self.alloc_block_and_chunk(block_allocator);
+        self.num_blocks.fetch_add(1, Ordering::Relaxed);
+
+        // Add the block to the list in the SlabDesc
+        unsafe {
+            let mut block_lists_guard = self.block_lists.write();
+            block_lists_guard.nonfull_blocks.push_head(new_block);
+        }
+        self.num_allocated.fetch_add(1, Ordering::Relaxed);
+        new_chunk
+    }
+
+    pub fn dealloc_chunk(&self, chunk_ptr: *mut u8, _block_allocator: &BlockAllocator) {
+        // Find the block it belongs to. You can find the block from the address. (And knowing the
+        // layout, you could calculate the chunk number too.)
+        let block_ptr: *mut SlabBlockHeader = {
+            let block_addr = (chunk_ptr.addr() / BLOCK_SIZE) * BLOCK_SIZE;
+            chunk_ptr.with_addr(block_addr).cast()
+        };
+        let chunk_ptr: *mut FreeChunk = chunk_ptr.cast();
+
+        // Mark the chunk as free in 'freechunks' list
+        let num_chunks;
+        let num_free_chunks;
+        unsafe {
+            let mut free_chunks_head = (*block_ptr).free_chunks_head.lock();
+            (*chunk_ptr).next = *free_chunks_head;
+            *free_chunks_head = chunk_ptr;
+
+            num_free_chunks = (*block_ptr).num_free_chunks.fetch_add(1, Ordering::Relaxed) + 1;
+            num_chunks = (*block_ptr).num_chunks;
+        }
+
+        if num_free_chunks == 1 {
+            // If the block was full previously, add it to the nonfull blocks list. Note that
+            // we're not holding the lock anymore, so it can immediately become full again.
+            // That's harmless, it will be moved back to the full list again when a call
+            // to alloc_chunk() sees it.
+            let mut block_lists = self.block_lists.write();
+            unsafe {
+                block_lists.unlink(block_ptr);
+                block_lists.nonfull_blocks.push_head(block_ptr);
+            };
+        } else if num_free_chunks == num_chunks {
+            // If the block became completely empty, move it to the free list
+            // TODO
+            // FIXME: we're still holding the spinlock. It's not exactly safe to return it to
+            // the free blocks list, is it? Defer it as garbage to wait out concurrent updates?
+            //block_allocator.release_block()
+        }
+
+        // update stats
+        self.num_allocated.fetch_sub(1, Ordering::Relaxed);
+    }
+
+    fn alloc_block_and_chunk(
+        &self,
+        block_allocator: &BlockAllocator,
+    ) -> (*mut SlabBlockHeader, *mut u8) {
+        // fixme: handle OOM
+        let block_slice: &mut [MaybeUninit<u8>] = block_allocator.alloc_block();
+        let (block_header, remain) = alloc_from_slice::<SlabBlockHeader>(block_slice);
+
+        let padding = remain.as_ptr().align_offset(self.layout.align());
+
+        let num_chunks = (remain.len() - padding) / self.layout.size();
+
+        let first_chunk_ptr: *mut FreeChunk = remain[padding..].as_mut_ptr().cast();
+
+        unsafe {
+            let mut chunk_ptr = first_chunk_ptr;
+            for _ in 0..num_chunks - 1 {
+                let next_chunk_ptr = chunk_ptr.byte_add(self.layout.size());
+                (*chunk_ptr).next = next_chunk_ptr;
+                chunk_ptr = next_chunk_ptr;
+            }
+            (*chunk_ptr).next = std::ptr::null_mut();
+
+            let result_chunk = first_chunk_ptr;
+
+            let block_header = block_header.write(SlabBlockHeader {
+                free_chunks_head: spin::Mutex::new((*first_chunk_ptr).next),
+                prev: std::ptr::null_mut(),
+                next: std::ptr::null_mut(),
+                num_chunks: num_chunks as u32,
+                num_free_chunks: AtomicU32::new(num_chunks as u32 - 1),
+            });
+
+            (block_header, result_chunk.cast())
+        }
+    }
+
+    #[cfg(test)]
+    fn dump(&self) {
+        eprintln!(
+            "slab dump ({} blocks, {} allocated chunks)",
+            self.num_blocks.load(Ordering::Relaxed),
+            self.num_allocated.load(Ordering::Relaxed)
+        );
+        let lists = self.block_lists.read();
+
+        eprintln!("nonfull blocks:");
+        lists.nonfull_blocks.dump();
+        eprintln!("full blocks:");
+        lists.full_blocks.dump();
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    use rand::Rng;
+    use rand_distr::Zipf;
+
+    struct TestObject {
+        val: usize,
+        _dummy: [u8; BLOCK_SIZE / 4],
+    }
+
+    struct TestObjectSlab<'a>(SlabDesc, BlockAllocator<'a>);
+    impl<'a> TestObjectSlab<'a> {
+        fn new(block_allocator: BlockAllocator) -> TestObjectSlab {
+            TestObjectSlab(SlabDesc::new(&Layout::new::<TestObject>()), block_allocator)
+        }
+
+        fn alloc(&self, val: usize) -> *mut TestObject {
+            let obj: *mut TestObject = self.0.alloc_chunk(&self.1).cast();
+            unsafe { (*obj).val = val };
+            obj
+        }
+
+        fn dealloc(&self, obj: *mut TestObject) {
+            self.0.dealloc_chunk(obj.cast(), &self.1)
+        }
+    }
+
+    #[test]
+    fn test_slab_alloc() {
+        const MEM_SIZE: usize = 100000000;
+        let mut area = Box::new_uninit_slice(MEM_SIZE);
+        let block_allocator = BlockAllocator::new(&mut area);
+
+        let slab = TestObjectSlab::new(block_allocator);
+
+        let mut all: Vec<*mut TestObject> = Vec::new();
+        for i in 0..11 {
+            all.push(slab.alloc(i));
+        }
+        for i in 0..11 {
+            assert!(unsafe { (*all[i]).val == i });
+        }
+
+        let distribution = Zipf::new(10 as f64, 1.1).unwrap();
+        let mut rng = rand::rng();
+        for _ in 0..100000 {
+            slab.0.dump();
+            let idx = (rng.sample(distribution) as usize).into();
+            let ptr: *mut TestObject = all[idx];
+            if !ptr.is_null() {
+                assert_eq!(unsafe { (*ptr).val }, idx);
+                slab.dealloc(ptr);
+                all[idx] = std::ptr::null_mut();
+            } else {
+                all[idx] = slab.alloc(idx);
+            }
+        }
+    }
+
+    fn new_test_blk(i: u32) -> *mut SlabBlockHeader {
+        Box::into_raw(Box::new(SlabBlockHeader {
+            free_chunks_head: spin::Mutex::new(std::ptr::null_mut()),
+            num_free_chunks: AtomicU32::new(0),
+            num_chunks: i,
+            prev: std::ptr::null_mut(),
+            next: std::ptr::null_mut(),
+        }))
+    }
+
+    #[test]
+    fn test_block_linked_list() {
+        // note: these are leaked, but that's OK for tests
+        let a = new_test_blk(0);
+        let b = new_test_blk(1);
+
+        let mut list = BlockList::default();
+        assert!(list.is_empty());
+
+        unsafe {
+            list.push_head(a);
+            assert!(!list.is_empty());
+            list.unlink(a);
+        }
+        assert!(list.is_empty());
+
+        unsafe {
+            list.push_head(b);
+            list.push_head(a);
+            assert_eq!(list.head, a);
+            assert_eq!((*a).next, b);
+            assert_eq!((*b).prev, a);
+            assert_eq!(list.tail, b);
+
+            list.unlink(a);
+            list.unlink(b);
+            assert!(list.is_empty());
+        }
+    }
+}

libs/neonart/src/allocator/static.rs

@@ -0,0 +1,44 @@
+use std::mem::MaybeUninit;
+
+pub fn alloc_from_slice<T>(
+    area: &mut [MaybeUninit<u8>],
+) -> (&mut MaybeUninit<T>, &mut [MaybeUninit<u8>]) {
+    let layout = std::alloc::Layout::new::<T>();
+
+    let area_start = area.as_mut_ptr();
+
+    // pad to satisfy alignment requirements
+    let padding = area_start.align_offset(layout.align());
+    if padding + layout.size() > area.len() {
+        panic!("out of memory");
+    }
+    let area = &mut area[padding..];
+    let (result_area, remain) = area.split_at_mut(layout.size());
+
+    let result_ptr: *mut MaybeUninit<T> = result_area.as_mut_ptr().cast();
+    let result = unsafe { result_ptr.as_mut().unwrap() };
+
+    (result, remain)
+}
+
+pub fn alloc_array_from_slice<T>(
+    area: &mut [MaybeUninit<u8>],
+    len: usize,
+) -> (&mut [MaybeUninit<T>], &mut [MaybeUninit<u8>]) {
+    let layout = std::alloc::Layout::new::<T>();
+
+    let area_start = area.as_mut_ptr();
+
+    // pad to satisfy alignment requirements
+    let padding = area_start.align_offset(layout.align());
+    if padding + layout.size() * len > area.len() {
+        panic!("out of memory");
+    }
+    let area = &mut area[padding..];
+    let (result_area, remain) = area.split_at_mut(layout.size() * len);
+
+    let result_ptr: *mut MaybeUninit<T> = result_area.as_mut_ptr().cast();
+    let result = unsafe { std::slice::from_raw_parts_mut(result_ptr.as_mut().unwrap(), len) };
+
+    (result, remain)
+}

libs/neonart/src/epoch.rs

@@ -0,0 +1,147 @@
+//! This is similar to crossbeam_epoch crate, but works in shared memory
+
+use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
+
+use crossbeam_utils::CachePadded;
+use spin;
+
+const NUM_SLOTS: usize = 1000;
+
+/// This is the struct that is stored in shmem
+///
+/// bit 0: is it pinned or not?
+/// rest of the bits are the epoch counter.
+pub struct EpochShared {
+    global_epoch: AtomicU64,
+    participants: [CachePadded<AtomicU64>; NUM_SLOTS],
+
+    broadcast_lock: spin::Mutex<()>,
+}
+
+impl EpochShared {
+    pub fn new() -> EpochShared {
+        EpochShared {
+            global_epoch: AtomicU64::new(2),
+            participants: [const { CachePadded::new(AtomicU64::new(2)) }; NUM_SLOTS],
+            broadcast_lock: spin::Mutex::new(()),
+        }
+    }
+
+    pub fn register(&self) -> LocalHandle {
+        LocalHandle {
+            global: self,
+            last_slot: AtomicUsize::new(0), // todo: choose more intelligently
+        }
+    }
+
+    fn release_pin(&self, slot: usize, _epoch: u64) {
+        let global_epoch = self.global_epoch.load(Ordering::Relaxed);
+        self.participants[slot].store(global_epoch, Ordering::Relaxed);
+    }
+
+    fn pin_internal(&self, slot_hint: usize) -> (usize, u64) {
+        // pick a slot
+        let mut slot = slot_hint;
+        let epoch = loop {
+            let old = self.participants[slot].fetch_or(1, Ordering::Relaxed);
+            if old & 1 == 0 {
+                // Got this slot
+                break old;
+            }
+
+            // the slot was busy by another thread / process. try a different slot
+            slot += 1;
+            if slot == NUM_SLOTS {
+                slot = 0;
+            }
+            continue;
+        };
+        (slot, epoch)
+    }
+
+    pub(crate) fn advance(&self) -> u64 {
+        // Advance the global epoch
+        let old_epoch = self.global_epoch.fetch_add(2, Ordering::Relaxed);
+        let new_epoch = old_epoch + 2;
+
+        // Anyone that release their pin after this will update their slot.
+        new_epoch
+    }
+
+    pub(crate) fn broadcast(&self) {
+        let Some(_guard) = self.broadcast_lock.try_lock() else {
+            return;
+        };
+
+        let epoch = self.global_epoch.load(Ordering::Relaxed);
+        let old_epoch = epoch.wrapping_sub(2);
+
+        // Update all free slots.
+        for i in 0..NUM_SLOTS {
+            // TODO: check result, as a sanity check. It should either be the old epoch, or pinned
+            let _ = self.participants[i].compare_exchange(
+                old_epoch,
+                epoch,
+                Ordering::Relaxed,
+                Ordering::Relaxed,
+            );
+        }
+
+        // FIXME: memory fence here, since we used Relaxed?
+    }
+
+    pub(crate) fn get_oldest(&self) -> u64 {
+        // Read all slots.
+        let now = self.global_epoch.load(Ordering::Relaxed);
+        let mut oldest = now;
+        for i in 0..NUM_SLOTS {
+            let this_epoch = self.participants[i].load(Ordering::Relaxed);
+            let delta = now.wrapping_sub(this_epoch);
+            if delta > u64::MAX / 2 {
+                // this is very recent
+            } else {
+                if delta > now.wrapping_sub(oldest) {
+                    oldest = this_epoch;
+                }
+            }
+        }
+        oldest
+    }
+
+    pub(crate) fn get_current(&self) -> u64 {
+        self.global_epoch.load(Ordering::Relaxed)
+    }
+}
+
+pub(crate) struct EpochPin<'e> {
+    slot: usize,
+    pub(crate) epoch: u64,
+
+    handle: &'e LocalHandle<'e>,
+}
+
+impl<'e> Drop for EpochPin<'e> {
+    fn drop(&mut self) {
+        self.handle.global.release_pin(self.slot, self.epoch);
+    }
+}
+
+pub struct LocalHandle<'g> {
+    global: &'g EpochShared,
+
+    last_slot: AtomicUsize,
+}
+
+impl<'g> LocalHandle<'g> {
+    pub fn pin(&self) -> EpochPin {
+        let (slot, epoch) = self
+            .global
+            .pin_internal(self.last_slot.load(Ordering::Relaxed));
+        self.last_slot.store(slot, Ordering::Relaxed);
+        EpochPin {
+            handle: self,
+            epoch,
+            slot,
+        }
+    }
+}

libs/neonart/src/lib.rs

@@ -0,0 +1,587 @@
+//! Adaptive Radix Tree (ART) implementation, with Optimistic Lock Coupling.
+//!
+//! The data structure is described in these two papers:
+//!
+//! [1] Leis, V. & Kemper, Alfons & Neumann, Thomas. (2013).
+//!     The adaptive radix tree: ARTful indexing for main-memory databases.
+//!     Proceedings - International Conference on Data Engineering. 38-49. 10.1109/ICDE.2013.6544812.
+//!     https://db.in.tum.de/~leis/papers/ART.pdf
+//!
+//! [2] Leis, Viktor & Scheibner, Florian & Kemper, Alfons & Neumann, Thomas. (2016).
+//!     The ART of practical synchronization.
+//!     1-8. 10.1145/2933349.2933352.
+//!     https://db.in.tum.de/~leis/papers/artsync.pdf
+//!
+//! [1] describes the base data structure, and [2] describes the Optimistic Lock Coupling that we
+//! use.
+//!
+//! The papers mention a few different variants. We have made the following choices in this
+//! implementation:
+//!
+//! - All keys have the same length
+//!
+//! - Single-value leaves.
+//!
+//! - For collapsing inner nodes, we use the Pessimistic approach, where each inner node stores a
+//!   variable length "prefix", which stores the keys of all the one-way nodes which have been
+//!   removed. However, similar to the "hybrid" approach described in the paper, each node only has
+//!   space for a constant-size prefix of 8 bytes. If a node would have a longer prefix, then we
+//!   create create one-way nodes to store them. (There was no particular reason for this choice,
+//!   the "hybrid" approach described in the paper might be better.)
+//!
+//! - For concurrency, we use Optimistic Lock Coupling. The paper [2] also describes another method,
+//!   ROWEX, which generally performs better when there is contention, but that is not important
+//!   for use and Optimisic Lock Coupling is simpler to implement.
+//!
+//! ## Requirements
+//!
+//! This data structure is currently used for the integrated LFC, relsize and last-written LSN cache
+//! in the compute communicator, part of the 'neon' Postgres extension. We have some unique
+//! requirements, which is why we had to write our own. Namely:
+//!
+//! - The data structure has to live in fixed-sized shared memory segment. That rules out any
+//!   built-in Rust collections and most crates. (Except possibly with the 'allocator_api' rust
+//!   feature, which still nightly-only experimental as of this writing).
+//!
+//! - The data structure is accessed from multiple processes. Only one process updates the data
+//!   structure, but other processes perform reads. That rules out using built-in Rust locking
+//!   primitives like Mutex and RwLock, and most crates too.
+//!
+//! - Within the one process with write-access, multiple threads can perform updates concurrently.
+//!   That rules out using PostgreSQL LWLocks for the locking.
+//!
+//! The implementation is generic, and doesn't depend on any PostgreSQL specifics, but it has been
+//! written with that usage and the above constraints in mind. Some noteworthy assumptions:
+//!
+//! - Contention is assumed to be rare. In the integrated cache in PostgreSQL, there's higher level
+//!   locking in the PostgreSQL buffer manager, which ensures that two backends should not try to
+//!   read / write the same page at the same time. (Prefetching can conflict with actual reads,
+//!   however.)
+//!
+//!  - The keys in the integrated cache are 17 bytes long.
+//!
+//! ## Usage
+//!
+//! Because this is designed to be used as a Postgres shared memory data structure, initialization
+//! happens in three stages:
+//!
+//! 0. A fixed area of shared memory is allocated at postmaster startup.
+//!
+//! 1. TreeInitStruct::new() is called to initialize it, still in Postmaster process, before any
+//!    other process or thread is running. It returns a TreeInitStruct, which is inherited by all
+//!    the processes through fork().
+//!
+//! 2. One process may have write-access to the struct, by calling
+//!    [TreeInitStruct::attach_writer]. (That process is the communicator process.)
+//!
+//! 3. Other processes get read-access to the struct, by calling [TreeInitStruct::attach_reader]
+//!
+//! "Write access" means that you can insert / update / delete values in the tree.
+//!
+//! NOTE: The Values stored in the tree are sometimes moved, when a leaf node fills up and a new
+//! larger node needs to be allocated. The versioning and epoch-based allocator ensure that the data
+//! structure stays consistent, but if the Value has interior mutability, like atomic fields,
+//! updates to such fields might be lost if the leaf node is concurrently moved! If that becomes a
+//! problem, the version check could be passed up to the caller, so that the caller could detect the
+//! lost updates and retry the operation.
+//!
+//! ## Implementation
+//!
+//! node_ptr: Provides low-level implementations of the four different node types (eight actually,
+//! since there is an Internal and Leaf variant of each)
+//!
+//! lock_and_version.rs: Provides an abstraction for the combined lock and version counter on each
+//! node.
+//!
+//! node_ref.rs: The code in node_ptr.rs deals with raw pointers. node_ref.rs provides more type-safe
+//!   abstractions on top.
+//!
+//! algorithm.rs: Contains the functions to implement lookups and updates in the tree
+//!
+//! allocator.rs: Provides a facility to allocate memory for the tree nodes. (We must provide our
+//!   own abstraction for that because we need the data structure to live in a pre-allocated shared
+//!   memory segment).
+//!
+//! epoch.rs: The data structure requires that when a node is removed from the tree, it is not
+//!   immediately deallocated, but stays around for as long as concurrent readers might still have
+//!   pointers to them. This is enforced by an epoch system. This is similar to
+//!   e.g. crossbeam_epoch, but we couldn't use that either because it has to work across processes
+//!   communicating over the shared memory segment.
+//!
+//! ## See also
+//!
+//! There are some existing Rust ART implementations out there, but none of them filled all
+//! the requirements:
+//!
+//! - https://github.com/XiangpengHao/congee
+//! - https://github.com/declanvk/blart
+//!
+//! ## TODO
+//!
+//! - Removing values has not been implemented
+
+mod algorithm;
+pub mod allocator;
+mod epoch;
+
+use algorithm::RootPtr;
+use algorithm::node_ptr::NodePtr;
+
+use std::collections::VecDeque;
+use std::fmt::Debug;
+use std::marker::PhantomData;
+use std::ptr::NonNull;
+use std::sync::atomic::{AtomicBool, Ordering};
+
+use crate::epoch::EpochPin;
+
+#[cfg(test)]
+mod tests;
+
+use allocator::ArtAllocator;
+pub use allocator::ArtMultiSlabAllocator;
+pub use allocator::OutOfMemoryError;
+
+/// Fixed-length key type.
+///
+pub trait Key: Debug {
+    const KEY_LEN: usize;
+
+    fn as_bytes(&self) -> &[u8];
+}
+
+/// Values stored in the tree
+///
+/// Values need to be Cloneable, because when a node "grows", the value is copied to a new node and
+/// the old sticks around until all readers that might see the old value are gone.
+// fixme obsolete, no longer needs Clone
+pub trait Value {}
+
+const MAX_GARBAGE: usize = 1024;
+
+/// The root of the tree, plus other tree-wide data. This is stored in the shared memory.
+pub struct Tree<V: Value> {
+    /// For simplicity, so that we never need to grow or shrink the root, the root node is always an
+    /// Internal256 node. Also, it never has a prefix (that's actually a bit wasteful, incurring one
+    /// indirection to every lookup)
+    root: RootPtr<V>,
+
+    writer_attached: AtomicBool,
+
+    epoch: epoch::EpochShared,
+}
+
+unsafe impl<V: Value + Sync> Sync for Tree<V> {}
+unsafe impl<V: Value + Send> Send for Tree<V> {}
+
+struct GarbageQueue<V>(VecDeque<(NodePtr<V>, u64)>);
+
+unsafe impl<V: Value + Sync> Sync for GarbageQueue<V> {}
+unsafe impl<V: Value + Send> Send for GarbageQueue<V> {}
+
+impl<V> GarbageQueue<V> {
+    fn new() -> GarbageQueue<V> {
+        GarbageQueue(VecDeque::with_capacity(MAX_GARBAGE))
+    }
+
+    fn remember_obsolete_node(&mut self, ptr: NodePtr<V>, epoch: u64) {
+        self.0.push_front((ptr, epoch));
+    }
+
+    fn next_obsolete(&mut self, cutoff_epoch: u64) -> Option<NodePtr<V>> {
+        if let Some(back) = self.0.back() {
+            if back.1 < cutoff_epoch {
+                return Some(self.0.pop_back().unwrap().0);
+            }
+        }
+        None
+    }
+}
+
+/// Struct created at postmaster startup
+pub struct TreeInitStruct<'t, K: Key, V: Value, A: ArtAllocator<V>> {
+    tree: &'t Tree<V>,
+
+    allocator: &'t A,
+
+    phantom_key: PhantomData<K>,
+}
+
+/// The worker process has a reference to this. The write operations are only safe
+/// from the worker process
+pub struct TreeWriteAccess<'t, K: Key, V: Value, A: ArtAllocator<V>>
+where
+    K: Key,
+    V: Value,
+{
+    tree: &'t Tree<V>,
+
+    pub allocator: &'t A,
+
+    epoch_handle: epoch::LocalHandle<'t>,
+
+    phantom_key: PhantomData<K>,
+
+    /// Obsolete nodes that cannot be recycled until their epoch expires.
+    garbage: spin::Mutex<GarbageQueue<V>>,
+}
+
+/// The backends have a reference to this. It cannot be used to modify the tree
+pub struct TreeReadAccess<'t, K: Key, V: Value>
+where
+    K: Key,
+    V: Value,
+{
+    tree: &'t Tree<V>,
+
+    epoch_handle: epoch::LocalHandle<'t>,
+
+    phantom_key: PhantomData<K>,
+}
+
+impl<'a, 't: 'a, K: Key, V: Value, A: ArtAllocator<V>> TreeInitStruct<'t, K, V, A> {
+    pub fn new(allocator: &'t A) -> TreeInitStruct<'t, K, V, A> {
+        let tree_ptr = allocator.alloc_tree();
+        let tree_ptr = NonNull::new(tree_ptr).expect("out of memory");
+        let init = Tree {
+            root: algorithm::new_root(allocator).expect("out of memory"),
+            writer_attached: AtomicBool::new(false),
+            epoch: epoch::EpochShared::new(),
+        };
+        unsafe { tree_ptr.write(init) };
+
+        TreeInitStruct {
+            tree: unsafe { tree_ptr.as_ref() },
+            allocator,
+            phantom_key: PhantomData,
+        }
+    }
+
+    pub fn attach_writer(self) -> TreeWriteAccess<'t, K, V, A> {
+        let previously_attached = self.tree.writer_attached.swap(true, Ordering::Relaxed);
+        if previously_attached {
+            panic!("writer already attached");
+        }
+        TreeWriteAccess {
+            tree: self.tree,
+            allocator: self.allocator,
+            phantom_key: PhantomData,
+            epoch_handle: self.tree.epoch.register(),
+            garbage: spin::Mutex::new(GarbageQueue::new()),
+        }
+    }
+
+    pub fn attach_reader(self) -> TreeReadAccess<'t, K, V> {
+        TreeReadAccess {
+            tree: self.tree,
+            phantom_key: PhantomData,
+            epoch_handle: self.tree.epoch.register(),
+        }
+    }
+}
+
+impl<'t, K: Key, V: Value, A: ArtAllocator<V>> TreeWriteAccess<'t, K, V, A> {
+    pub fn start_write<'g>(&'t self) -> TreeWriteGuard<'g, K, V, A>
+    where
+        't: 'g,
+    {
+        TreeWriteGuard {
+            tree_writer: self,
+            epoch_pin: self.epoch_handle.pin(),
+            phantom_key: PhantomData,
+            created_garbage: false,
+        }
+    }
+
+    pub fn start_read(&'t self) -> TreeReadGuard<'t, K, V> {
+        TreeReadGuard {
+            tree: &self.tree,
+            epoch_pin: self.epoch_handle.pin(),
+            phantom_key: PhantomData,
+        }
+    }
+}
+
+impl<'t, K: Key, V: Value> TreeReadAccess<'t, K, V> {
+    pub fn start_read(&'t self) -> TreeReadGuard<'t, K, V> {
+        TreeReadGuard {
+            tree: &self.tree,
+            epoch_pin: self.epoch_handle.pin(),
+            phantom_key: PhantomData,
+        }
+    }
+}
+
+pub struct TreeReadGuard<'e, K, V>
+where
+    K: Key,
+    V: Value,
+{
+    tree: &'e Tree<V>,
+
+    epoch_pin: EpochPin<'e>,
+    phantom_key: PhantomData<K>,
+}
+
+impl<'e, K: Key, V: Value> TreeReadGuard<'e, K, V> {
+    pub fn get(&'e self, key: &K) -> Option<&'e V> {
+        algorithm::search(key, self.tree.root, &self.epoch_pin)
+    }
+}
+
+pub struct TreeWriteGuard<'e, K, V, A>
+where
+    K: Key,
+    V: Value,
+    A: ArtAllocator<V>,
+{
+    tree_writer: &'e TreeWriteAccess<'e, K, V, A>,
+
+    epoch_pin: EpochPin<'e>,
+    phantom_key: PhantomData<K>,
+
+    created_garbage: bool,
+}
+
+pub enum UpdateAction<V> {
+    Nothing,
+    Insert(V),
+    Remove,
+}
+
+impl<'e, K: Key, V: Value, A: ArtAllocator<V>> TreeWriteGuard<'e, K, V, A> {
+    /// Get a value
+    pub fn get(&'e mut self, key: &K) -> Option<&'e V> {
+        algorithm::search(key, self.tree_writer.tree.root, &self.epoch_pin)
+    }
+
+    /// Insert a value
+    pub fn insert(self, key: &K, value: V) -> Result<bool, OutOfMemoryError> {
+        let mut success = None;
+
+        self.update_with_fn(key, |existing| {
+            if let Some(_) = existing {
+                success = Some(false);
+                UpdateAction::Nothing
+            } else {
+                success = Some(true);
+                UpdateAction::Insert(value)
+            }
+        })?;
+        Ok(success.expect("value_fn not called"))
+    }
+
+    /// Remove value. Returns true if it existed
+    pub fn remove(self, key: &K) -> bool {
+        let mut result = false;
+        // FIXME: It's not clear if OOM is expected while removing. It seems
+        // not nice, but shrinking a node can OOM. Then again, we could opt
+        // to not shrink a node if we cannot allocate, to live a little longer.
+        self.update_with_fn(key, |existing| match existing {
+            Some(_) => {
+                result = true;
+                UpdateAction::Remove
+            }
+            None => UpdateAction::Nothing,
+        })
+        .expect("out of memory while removing");
+        result
+    }
+
+    /// Try to remove value and return the old value.
+    pub fn remove_and_return(self, key: &K) -> Option<V>
+    where
+        V: Clone,
+    {
+        let mut old = None;
+        self.update_with_fn(key, |existing| {
+            old = existing.cloned();
+            UpdateAction::Remove
+        })
+        .expect("out of memory while removing");
+        old
+    }
+
+    /// Update key using the given function. All the other modifying operations are based on this.
+    ///
+    /// The function is passed a reference to the existing value, if any. If the function
+    /// returns None, the value is removed from the tree (or if there was no existing value,
+    /// does nothing). If the function returns Some, the existing value is replaced, of if there
+    /// was no existing value, it is inserted. FIXME: update comment
+    pub fn update_with_fn<F>(mut self, key: &K, value_fn: F) -> Result<(), OutOfMemoryError>
+    where
+        F: FnOnce(Option<&V>) -> UpdateAction<V>,
+    {
+        algorithm::update_fn(key, value_fn, self.tree_writer.tree.root, &mut self)?;
+
+        if self.created_garbage {
+            let _ = self.collect_garbage();
+        }
+        Ok(())
+    }
+
+    fn remember_obsolete_node(&mut self, ptr: NodePtr<V>) {
+        self.tree_writer
+            .garbage
+            .lock()
+            .remember_obsolete_node(ptr, self.epoch_pin.epoch);
+        self.created_garbage = true;
+    }
+
+    // returns number of nodes recycled
+    fn collect_garbage(&self) -> usize {
+        self.tree_writer.tree.epoch.advance();
+        self.tree_writer.tree.epoch.broadcast();
+
+        let cutoff_epoch = self.tree_writer.tree.epoch.get_oldest();
+
+        let mut result = 0;
+        let mut garbage_queue = self.tree_writer.garbage.lock();
+        while let Some(ptr) = garbage_queue.next_obsolete(cutoff_epoch) {
+            ptr.deallocate(self.tree_writer.allocator);
+            result += 1;
+        }
+        result
+    }
+}
+
+pub struct TreeIterator<K>
+where
+    K: Key + for<'a> From<&'a [u8]>,
+{
+    done: bool,
+    pub next_key: Vec<u8>,
+    max_key: Option<Vec<u8>>,
+
+    phantom_key: PhantomData<K>,
+}
+
+impl<K> TreeIterator<K>
+where
+    K: Key + for<'a> From<&'a [u8]>,
+{
+    pub fn new_wrapping() -> TreeIterator<K> {
+        let mut next_key = Vec::new();
+        next_key.resize(K::KEY_LEN, 0);
+        TreeIterator {
+            done: false,
+            next_key,
+            max_key: None,
+            phantom_key: PhantomData,
+        }
+    }
+
+    pub fn new(range: &std::ops::Range<K>) -> TreeIterator<K> {
+        let result = TreeIterator {
+            done: false,
+            next_key: Vec::from(range.start.as_bytes()),
+            max_key: Some(Vec::from(range.end.as_bytes())),
+            phantom_key: PhantomData,
+        };
+        assert_eq!(result.next_key.len(), K::KEY_LEN);
+        assert_eq!(result.max_key.as_ref().unwrap().len(), K::KEY_LEN);
+
+        result
+    }
+
+    pub fn next<'g, V>(&mut self, read_guard: &'g TreeReadGuard<'g, K, V>) -> Option<(K, &'g V)>
+    where
+        V: Value,
+    {
+        if self.done {
+            return None;
+        }
+
+        let mut wrapped_around = false;
+        loop {
+            assert_eq!(self.next_key.len(), K::KEY_LEN);
+            if let Some((k, v)) = algorithm::iter_next(
+                &mut self.next_key,
+                read_guard.tree.root,
+                &read_guard.epoch_pin,
+            ) {
+                assert_eq!(k.len(), K::KEY_LEN);
+                assert_eq!(self.next_key.len(), K::KEY_LEN);
+
+                // Check if we reached the end of the range
+                if let Some(max_key) = &self.max_key {
+                    if k.as_slice() >= max_key.as_slice() {
+                        self.done = true;
+                        break None;
+                    }
+                }
+
+                // increment the key
+                self.next_key = k.clone();
+                increment_key(self.next_key.as_mut_slice());
+                let k = k.as_slice().into();
+
+                break Some((k, v));
+            } else {
+                if self.max_key.is_some() {
+                    self.done = true;
+                } else {
+                    // Start from beginning
+                    if !wrapped_around {
+                        for i in 0..K::KEY_LEN {
+                            self.next_key[i] = 0;
+                        }
+                        wrapped_around = true;
+                        continue;
+                    } else {
+                        // The tree is completely empty
+                        // FIXME: perhaps we should remember the starting point instead.
+                        // Currently this will scan some ranges twice.
+                        break None;
+                    }
+                }
+                break None;
+            }
+        }
+    }
+}
+
+fn increment_key(key: &mut [u8]) -> bool {
+    for i in (0..key.len()).rev() {
+        let (byte, overflow) = key[i].overflowing_add(1);
+        key[i] = byte;
+        if !overflow {
+            return false;
+        }
+    }
+    true
+}
+
+// Debugging functions
+impl<'e, K: Key, V: Value + Debug, A: ArtAllocator<V>> TreeWriteGuard<'e, K, V, A> {
+    pub fn dump(&mut self, dst: &mut dyn std::io::Write) {
+        algorithm::dump_tree(self.tree_writer.tree.root, &self.epoch_pin, dst)
+    }
+}
+impl<'e, K: Key, V: Value + Debug> TreeReadGuard<'e, K, V> {
+    pub fn dump(&mut self, dst: &mut dyn std::io::Write) {
+        algorithm::dump_tree(self.tree.root, &self.epoch_pin, dst)
+    }
+}
+impl<'e, K: Key, V: Value> TreeWriteAccess<'e, K, V, ArtMultiSlabAllocator<'e, V>> {
+    pub fn get_statistics(&self) -> ArtTreeStatistics {
+        self.allocator.get_statistics();
+        ArtTreeStatistics {
+            blocks: self.allocator.inner.block_allocator.get_statistics(),
+            slabs: self.allocator.get_statistics(),
+            epoch: self.tree.epoch.get_current(),
+            oldest_epoch: self.tree.epoch.get_oldest(),
+            num_garbage: self.garbage.lock().0.len() as u64,
+        }
+    }
+}
+
+#[derive(Clone, Debug)]
+pub struct ArtTreeStatistics {
+    pub blocks: allocator::block::BlockAllocatorStats,
+    pub slabs: allocator::ArtMultiSlabStats,
+
+    pub epoch: u64,
+    pub oldest_epoch: u64,
+    pub num_garbage: u64,
+}

libs/neonart/src/tests.rs

@@ -0,0 +1,243 @@
+use std::collections::BTreeMap;
+use std::collections::HashSet;
+use std::fmt::{Debug, Formatter};
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+use crate::ArtAllocator;
+use crate::ArtMultiSlabAllocator;
+use crate::TreeInitStruct;
+use crate::TreeIterator;
+use crate::TreeWriteAccess;
+use crate::UpdateAction;
+
+use crate::{Key, Value};
+
+use rand::Rng;
+use rand::seq::SliceRandom;
+use rand_distr::Zipf;
+
+const TEST_KEY_LEN: usize = 16;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
+struct TestKey([u8; TEST_KEY_LEN]);
+
+impl TestKey {
+    const MIN: TestKey = TestKey([0; TEST_KEY_LEN]);
+    const MAX: TestKey = TestKey([u8::MAX; TEST_KEY_LEN]);
+}
+
+impl Key for TestKey {
+    const KEY_LEN: usize = TEST_KEY_LEN;
+    fn as_bytes(&self) -> &[u8] {
+        &self.0
+    }
+}
+
+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())
+    }
+}
+
+impl Value for usize {}
+
+fn test_inserts<K: Into<TestKey> + Copy>(keys: &[K]) {
+    const MEM_SIZE: usize = 10000000;
+    let mut area = Box::new_uninit_slice(MEM_SIZE);
+
+    let allocator = ArtMultiSlabAllocator::new(&mut area);
+
+    let init_struct = TreeInitStruct::<TestKey, usize, _>::new(allocator);
+    let tree_writer = init_struct.attach_writer();
+
+    for (idx, k) in keys.iter().enumerate() {
+        let w = tree_writer.start_write();
+        let res = w.insert(&(*k).into(), idx);
+        assert!(res.is_ok());
+    }
+
+    for (idx, k) in keys.iter().enumerate() {
+        let r = tree_writer.start_read();
+        let value = r.get(&(*k).into());
+        assert_eq!(value, Some(idx).as_ref());
+    }
+
+    eprintln!("stats: {:?}", tree_writer.get_statistics());
+}
+
+#[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.get(&key).is_some() {
+                continue;
+            }
+            used_keys.insert(key);
+            keys.push(key.into());
+            break;
+        }
+    }
+    test_inserts(&keys);
+}
+
+struct TestValue(AtomicUsize);
+
+impl TestValue {
+    fn new(val: usize) -> TestValue {
+        TestValue(AtomicUsize::new(val))
+    }
+
+    fn load(&self) -> usize {
+        self.0.load(Ordering::Relaxed)
+    }
+}
+
+impl Value for TestValue {}
+
+impl Clone for TestValue {
+    fn clone(&self) -> TestValue {
+        TestValue::new(self.load())
+    }
+}
+
+impl Debug for TestValue {
+    fn fmt(&self, fmt: &mut Formatter<'_>) -> Result<(), std::fmt::Error> {
+        write!(fmt, "{:?}", self.load())
+    }
+}
+
+#[derive(Clone, Debug)]
+struct TestOp(TestKey, Option<usize>);
+
+fn apply_op<A: ArtAllocator<TestValue>>(
+    op: &TestOp,
+    tree: &TreeWriteAccess<TestKey, TestValue, A>,
+    shadow: &mut BTreeMap<TestKey, usize>,
+) {
+    eprintln!("applying op: {op:?}");
+
+    // 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)
+    };
+
+    // apply to Art tree
+    let w = tree.start_write();
+    w.update_with_fn(&op.0, |existing| {
+        assert_eq!(existing.map(TestValue::load), shadow_existing);
+
+        match (existing, op.1) {
+            (None, None) => UpdateAction::Nothing,
+            (None, Some(new_val)) => UpdateAction::Insert(TestValue::new(new_val)),
+            (Some(_old_val), None) => UpdateAction::Remove,
+            (Some(old_val), Some(new_val)) => {
+                old_val.0.store(new_val, Ordering::Relaxed);
+                UpdateAction::Nothing
+            }
+        }
+    })
+    .expect("out of memory");
+}
+
+fn test_iter<A: ArtAllocator<TestValue>>(
+    tree: &TreeWriteAccess<TestKey, TestValue, A>,
+    shadow: &BTreeMap<TestKey, usize>,
+) {
+    let mut shadow_iter = shadow.iter();
+    let mut iter = TreeIterator::new(&(TestKey::MIN..TestKey::MAX));
+
+    loop {
+        let shadow_item = shadow_iter.next().map(|(k, v)| (k.clone(), v.clone()));
+        let r = tree.start_read();
+        let item = iter.next(&r);
+
+        if shadow_item != item.map(|(k, v)| (k, v.load())) {
+            eprintln!(
+                "FAIL: iterator returned {:?}, expected {:?}",
+                item, shadow_item
+            );
+            tree.start_read().dump(&mut std::io::stderr());
+
+            eprintln!("SHADOW:");
+            let mut si = shadow.iter();
+            while let Some(si) = si.next() {
+                eprintln!("key: {:?}, val: {}", si.0, si.1);
+            }
+            panic!(
+                "FAIL: iterator returned {:?}, expected {:?}",
+                item, shadow_item
+            );
+        }
+        if item.is_none() {
+            break;
+        }
+    }
+}
+
+#[test]
+fn random_ops() {
+    const MEM_SIZE: usize = 10000000;
+    let mut area = Box::new_uninit_slice(MEM_SIZE);
+
+    let allocator = ArtMultiSlabAllocator::new(&mut area);
+
+    let init_struct = TreeInitStruct::<TestKey, TestValue, _>::new(allocator);
+    let tree_writer = init_struct.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 mut key: TestKey = (rng.sample(distribution) as u128).into();
+
+        if rng.random_bool(0.10) {
+            key = TestKey::from(u128::from(&key) | 0xffffffff);
+        }
+
+        let op = TestOp(key, if rng.random_bool(0.75) { Some(i) } else { None });
+
+        apply_op(&op, &tree_writer, &mut shadow);
+
+        if i % 1000 == 0 {
+            eprintln!("{i} ops processed");
+            eprintln!("stats: {:?}", tree_writer.get_statistics());
+            test_iter(&tree_writer, &shadow);
+        }
+    }
+}

libs/pageserver_api/src/config.rs

@@ -34,6 +34,8 @@ pub struct NodeMetadata {
     pub postgres_host: String,
     #[serde(rename = "port")]
     pub postgres_port: u16,
+    pub grpc_host: Option<String>,
+    pub grpc_port: Option<u16>,
     pub http_host: String,
     pub http_port: u16,
     pub https_port: Option<u16>,

libs/pageserver_api/src/config/tests.rs

@@ -14,6 +14,8 @@ fn test_node_metadata_v1_backward_compatibilty() {
         NodeMetadata {
             postgres_host: "localhost".to_string(),
             postgres_port: 23,
+            grpc_host: None,
+            grpc_port: None,
             http_host: "localhost".to_string(),
             http_port: 42,
             https_port: None,
@@ -37,6 +39,35 @@ fn test_node_metadata_v2_backward_compatibilty() {
         NodeMetadata {
             postgres_host: "localhost".to_string(),
             postgres_port: 23,
+            grpc_host: None,
+            grpc_port: None,
+            http_host: "localhost".to_string(),
+            http_port: 42,
+            https_port: Some(123),
+            other: HashMap::new(),
+        }
+    )
+}
+
+#[test]
+fn test_node_metadata_v3_backward_compatibilty() {
+    let v3 = serde_json::to_vec(&serde_json::json!({
+        "host": "localhost",
+        "port": 23,
+        "grpc_host": "localhost",
+        "grpc_port": 51,
+        "http_host": "localhost",
+        "http_port": 42,
+        "https_port": 123,
+    }));
+
+    assert_eq!(
+        serde_json::from_slice::<NodeMetadata>(&v3.unwrap()).unwrap(),
+        NodeMetadata {
+            postgres_host: "localhost".to_string(),
+            postgres_port: 23,
+            grpc_host: Some("localhost".to_string()),
+            grpc_port: Some(51),
             http_host: "localhost".to_string(),
             http_port: 42,
             https_port: Some(123),

libs/pageserver_api/src/controller_api.rs

@@ -53,6 +53,9 @@ pub struct NodeRegisterRequest {
     pub listen_pg_addr: String,
     pub listen_pg_port: u16,
 
+    pub listen_grpc_addr: Option<String>,
+    pub listen_grpc_port: Option<u16>,
+
     pub listen_http_addr: String,
     pub listen_http_port: u16,
     pub listen_https_port: Option<u16>,
@@ -102,6 +105,9 @@ pub struct TenantLocateResponseShard {
     pub listen_pg_addr: String,
     pub listen_pg_port: u16,
 
+    pub listen_grpc_addr: Option<String>,
+    pub listen_grpc_port: Option<u16>,
+
     pub listen_http_addr: String,
     pub listen_http_port: u16,
     pub listen_https_port: Option<u16>,
@@ -152,6 +158,8 @@ pub struct NodeDescribeResponse {
 
     pub listen_pg_addr: String,
     pub listen_pg_port: u16,
+    pub listen_grpc_addr: Option<String>,
+    pub listen_grpc_port: Option<u16>,
 }
 
 #[derive(Serialize, Deserialize, Debug)]
@@ -344,35 +352,6 @@ impl Default for ShardSchedulingPolicy {
     }
 }
 
-#[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
-pub enum NodeLifecycle {
-    Active,
-    Deleted,
-}
-
-impl FromStr for NodeLifecycle {
-    type Err = anyhow::Error;
-
-    fn from_str(s: &str) -> Result<Self, Self::Err> {
-        match s {
-            "active" => Ok(Self::Active),
-            "deleted" => Ok(Self::Deleted),
-            _ => Err(anyhow::anyhow!("Unknown node lifecycle '{s}'")),
-        }
-    }
-}
-
-impl From<NodeLifecycle> for String {
-    fn from(value: NodeLifecycle) -> String {
-        use NodeLifecycle::*;
-        match value {
-            Active => "active",
-            Deleted => "deleted",
-        }
-        .to_string()
-    }
-}
-
 #[derive(Serialize, Deserialize, Clone, Copy, Eq, PartialEq, Debug)]
 pub enum NodeSchedulingPolicy {
     Active,

libs/pageserver_api/src/upcall_api.rs

@@ -9,7 +9,7 @@ use utils::id::{NodeId, TimelineId};
 
 use crate::controller_api::NodeRegisterRequest;
 use crate::models::{LocationConfigMode, ShardImportStatus};
-use crate::shard::{ShardStripeSize, TenantShardId};
+use crate::shard::TenantShardId;
 
 /// Upcall message sent by the pageserver to the configured `control_plane_api` on
 /// startup.
@@ -36,10 +36,6 @@ pub struct ReAttachResponseTenant {
     /// Default value only for backward compat: this field should be set
     #[serde(default = "default_mode")]
     pub mode: LocationConfigMode,
-
-    // Default value only for backward compat: this field should be set
-    #[serde(default = "ShardStripeSize::default")]
-    pub stripe_size: ShardStripeSize,
 }
 #[derive(Serialize, Deserialize)]
 pub struct ReAttachResponse {

libs/posthog_client_lite/src/background_loop.rs

@@ -55,16 +55,9 @@ impl FeatureResolverBackgroundLoop {
                             continue;
                         }
                     };
-                    let project_id = this.posthog_client.config.project_id.parse::<u64>().ok();
-                    match FeatureStore::new_with_flags(resp.flags, project_id) {
-                        Ok(feature_store) => {
-                            this.feature_store.store(Arc::new(feature_store));
-                            tracing::info!("Feature flag updated");
-                        }
-                        Err(e) => {
-                            tracing::warn!("Cannot process feature flag spec: {}", e);
-                        }
-                    }
+                    let feature_store = FeatureStore::new_with_flags(resp.flags);
+                    this.feature_store.store(Arc::new(feature_store));
+                    tracing::info!("Feature flag updated");
                 }
                 tracing::info!("PostHog feature resolver stopped");
             }

libs/posthog_client_lite/src/lib.rs

@@ -39,9 +39,6 @@ pub struct LocalEvaluationResponse {
 
 #[derive(Deserialize)]
 pub struct LocalEvaluationFlag {
-    #[allow(dead_code)]
-    id: u64,
-    team_id: u64,
     key: String,
     filters: LocalEvaluationFlagFilters,
     active: bool,
@@ -110,32 +107,17 @@ impl FeatureStore {
         }
     }
 
-    pub fn new_with_flags(
-        flags: Vec<LocalEvaluationFlag>,
-        project_id: Option<u64>,
-    ) -> Result<Self, &'static str> {
+    pub fn new_with_flags(flags: Vec<LocalEvaluationFlag>) -> Self {
         let mut store = Self::new();
-        store.set_flags(flags, project_id)?;
-        Ok(store)
+        store.set_flags(flags);
+        store
     }
 
-    pub fn set_flags(
-        &mut self,
-        flags: Vec<LocalEvaluationFlag>,
-        project_id: Option<u64>,
-    ) -> Result<(), &'static str> {
+    pub fn set_flags(&mut self, flags: Vec<LocalEvaluationFlag>) {
         self.flags.clear();
         for flag in flags {
-            if let Some(project_id) = project_id {
-                if flag.team_id != project_id {
-                    return Err(
-                        "Retrieved a spec with different project id, wrong config? Discarding the feature flags.",
-                    );
-                }
-            }
             self.flags.insert(flag.key.clone(), flag);
         }
-        Ok(())
     }
 
     /// Generate a consistent hash for a user ID (e.g., tenant ID).
@@ -552,13 +534,6 @@ impl PostHogClient {
         })
     }
 
-    /// Check if the server API key is a feature flag secure API key. This key can only be
-    /// used to fetch the feature flag specs and can only be used on a undocumented API
-    /// endpoint.
-    fn is_feature_flag_secure_api_key(&self) -> bool {
-        self.config.server_api_key.starts_with("phs_")
-    }
-
     /// Fetch the feature flag specs from the server.
     ///
     /// This is unfortunately an undocumented API at:
@@ -572,22 +547,10 @@ impl PostHogClient {
     ) -> anyhow::Result<LocalEvaluationResponse> {
         // BASE_URL/api/projects/:project_id/feature_flags/local_evaluation
         // with bearer token of self.server_api_key
-        // OR
-        // BASE_URL/api/feature_flag/local_evaluation/
-        // with bearer token of feature flag specific self.server_api_key
-        let url = if self.is_feature_flag_secure_api_key() {
-            // The new feature local evaluation secure API token
-            format!(
-                "{}/api/feature_flag/local_evaluation",
-                self.config.private_api_url
-            )
-        } else {
-            // The old personal API token
-            format!(
-                "{}/api/projects/{}/feature_flags/local_evaluation",
-                self.config.private_api_url, self.config.project_id
-            )
-        };
+        let url = format!(
+            "{}/api/projects/{}/feature_flags/local_evaluation",
+            self.config.private_api_url, self.config.project_id
+        );
         let response = self
             .client
             .get(url)
@@ -840,7 +803,7 @@ mod tests {
     fn evaluate_multivariate() {
         let mut store = FeatureStore::new();
         let response: LocalEvaluationResponse = serde_json::from_str(data()).unwrap();
-        store.set_flags(response.flags, None).unwrap();
+        store.set_flags(response.flags);
 
         // This lacks the required properties and cannot be evaluated.
         let variant =
@@ -910,7 +873,7 @@ mod tests {
 
         let mut store = FeatureStore::new();
         let response: LocalEvaluationResponse = serde_json::from_str(data()).unwrap();
-        store.set_flags(response.flags, None).unwrap();
+        store.set_flags(response.flags);
 
         // This lacks the required properties and cannot be evaluated.
         let variant = store.evaluate_boolean_inner("boolean-flag", 1.00, &HashMap::new());
@@ -966,7 +929,7 @@ mod tests {
 
         let mut store = FeatureStore::new();
         let response: LocalEvaluationResponse = serde_json::from_str(data()).unwrap();
-        store.set_flags(response.flags, None).unwrap();
+        store.set_flags(response.flags);
 
         // This lacks the required properties and cannot be evaluated.
         let variant =

libs/proxy/postgres-protocol2/Cargo.toml

@@ -5,7 +5,7 @@ edition = "2024"
 license = "MIT/Apache-2.0"
 
 [dependencies]
-base64.workspace = true
+base64 = "0.20"
 byteorder.workspace = true
 bytes.workspace = true
 fallible-iterator.workspace = true

libs/proxy/postgres-protocol2/src/authentication/sasl.rs

@@ -3,8 +3,6 @@
 use std::fmt::Write;
 use std::{io, iter, mem, str};
 
-use base64::Engine as _;
-use base64::prelude::BASE64_STANDARD;
 use hmac::{Hmac, Mac};
 use rand::{self, Rng};
 use sha2::digest::FixedOutput;
@@ -228,7 +226,7 @@ impl ScramSha256 {
 
         let (client_key, server_key) = match password {
             Credentials::Password(password) => {
-                let salt = match BASE64_STANDARD.decode(parsed.salt) {
+                let salt = match base64::decode(parsed.salt) {
                     Ok(salt) => salt,
                     Err(e) => return Err(io::Error::new(io::ErrorKind::InvalidInput, e)),
                 };
@@ -257,7 +255,7 @@ impl ScramSha256 {
         let mut cbind_input = vec![];
         cbind_input.extend(channel_binding.gs2_header().as_bytes());
         cbind_input.extend(channel_binding.cbind_data());
-        let cbind_input = BASE64_STANDARD.encode(&cbind_input);
+        let cbind_input = base64::encode(&cbind_input);
 
         self.message.clear();
         write!(&mut self.message, "c={},r={}", cbind_input, parsed.nonce).unwrap();
@@ -274,12 +272,7 @@ impl ScramSha256 {
             *proof ^= signature;
         }
 
-        write!(
-            &mut self.message,
-            ",p={}",
-            BASE64_STANDARD.encode(client_proof)
-        )
-        .unwrap();
+        write!(&mut self.message, ",p={}", base64::encode(client_proof)).unwrap();
 
         self.state = State::Finish {
             server_key,
@@ -313,7 +306,7 @@ impl ScramSha256 {
             ServerFinalMessage::Verifier(verifier) => verifier,
         };
 
-        let verifier = match BASE64_STANDARD.decode(verifier) {
+        let verifier = match base64::decode(verifier) {
             Ok(verifier) => verifier,
             Err(e) => return Err(io::Error::new(io::ErrorKind::InvalidInput, e)),
         };

libs/proxy/postgres-protocol2/src/password/mod.rs

@@ -6,8 +6,6 @@
 //! side. This is good because it ensures the cleartext password won't
 //! end up in logs pg_stat displays, etc.
 
-use base64::Engine as _;
-use base64::prelude::BASE64_STANDARD;
 use hmac::{Hmac, Mac};
 use rand::RngCore;
 use sha2::digest::FixedOutput;
@@ -85,8 +83,8 @@ pub(crate) async fn scram_sha_256_salt(
     format!(
         "SCRAM-SHA-256${}:{}${}:{}",
         SCRAM_DEFAULT_ITERATIONS,
-        BASE64_STANDARD.encode(salt),
-        BASE64_STANDARD.encode(stored_key),
-        BASE64_STANDARD.encode(server_key)
+        base64::encode(salt),
+        base64::encode(stored_key),
+        base64::encode(server_key)
     )
 }

libs/proxy/tokio-postgres2/src/cancel_query.rs

@@ -10,7 +10,7 @@ use crate::{Error, cancel_query_raw, connect_socket};
 pub(crate) async fn cancel_query<T>(
     config: Option<SocketConfig>,
     ssl_mode: SslMode,
-    tls: T,
+    mut tls: T,
     process_id: i32,
     secret_key: i32,
 ) -> Result<(), Error>

libs/proxy/tokio-postgres2/src/config.rs

@@ -17,6 +17,7 @@ use crate::{Client, Connection, Error};
 
 /// TLS configuration.
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[non_exhaustive]
 pub enum SslMode {
     /// Do not use TLS.
     Disable,
@@ -230,7 +231,7 @@ impl Config {
     /// Requires the `runtime` Cargo feature (enabled by default).
     pub async fn connect<T>(
         &self,
-        tls: &T,
+        tls: T,
     ) -> Result<(Client, Connection<TcpStream, T::Stream>), Error>
     where
         T: MakeTlsConnect<TcpStream>,

libs/proxy/tokio-postgres2/src/connect.rs

@@ -13,7 +13,7 @@ use crate::tls::{MakeTlsConnect, TlsConnect};
 use crate::{Client, Config, Connection, Error, RawConnection};
 
 pub async fn connect<T>(
-    tls: &T,
+    mut tls: T,
     config: &Config,
 ) -> Result<(Client, Connection<TcpStream, T::Stream>), Error>
 where

libs/proxy/tokio-postgres2/src/tls.rs

@@ -47,7 +47,7 @@ pub trait MakeTlsConnect<S> {
     /// Creates a new `TlsConnect`or.
     ///
     /// The domain name is provided for certificate verification and SNI.
-    fn make_tls_connect(&self, domain: &str) -> Result<Self::TlsConnect, Self::Error>;
+    fn make_tls_connect(&mut self, domain: &str) -> Result<Self::TlsConnect, Self::Error>;
 }
 
 /// An asynchronous function wrapping a stream in a TLS session.
@@ -85,7 +85,7 @@ impl<S> MakeTlsConnect<S> for NoTls {
     type TlsConnect = NoTls;
     type Error = NoTlsError;
 
-    fn make_tls_connect(&self, _: &str) -> Result<NoTls, NoTlsError> {
+    fn make_tls_connect(&mut self, _: &str) -> Result<NoTls, NoTlsError> {
         Ok(NoTls)
     }
 }

libs/remote_storage/src/azure_blob.rs

@@ -10,7 +10,7 @@ use std::sync::Arc;
 use std::time::{Duration, SystemTime};
 use std::{env, io};
 
-use anyhow::{Context, Result, anyhow};
+use anyhow::{Context, Result};
 use azure_core::request_options::{IfMatchCondition, MaxResults, Metadata, Range};
 use azure_core::{Continuable, HttpClient, RetryOptions, TransportOptions};
 use azure_storage::StorageCredentials;
@@ -37,7 +37,6 @@ use crate::metrics::{AttemptOutcome, RequestKind, start_measuring_requests};
 use crate::{
     ConcurrencyLimiter, Download, DownloadError, DownloadKind, DownloadOpts, Listing, ListingMode,
     ListingObject, RemotePath, RemoteStorage, StorageMetadata, TimeTravelError, TimeoutOrCancel,
-    Version, VersionKind,
 };
 
 pub struct AzureBlobStorage {
@@ -406,39 +405,6 @@ impl AzureBlobStorage {
     pub fn container_name(&self) -> &str {
         &self.container_name
     }
-
-    async fn list_versions_with_permit(
-        &self,
-        _permit: &tokio::sync::SemaphorePermit<'_>,
-        prefix: Option<&RemotePath>,
-        mode: ListingMode,
-        max_keys: Option<NonZeroU32>,
-        cancel: &CancellationToken,
-    ) -> Result<crate::VersionListing, DownloadError> {
-        let customize_builder = |mut builder: ListBlobsBuilder| {
-            builder = builder.include_versions(true);
-            // We do not return this info back to `VersionListing` yet.
-            builder = builder.include_deleted(true);
-            builder
-        };
-        let kind = RequestKind::ListVersions;
-
-        let mut stream = std::pin::pin!(self.list_streaming_for_fn(
-            prefix,
-            mode,
-            max_keys,
-            cancel,
-            kind,
-            customize_builder
-        ));
-        let mut combined: crate::VersionListing =
-            stream.next().await.expect("At least one item required")?;
-        while let Some(list) = stream.next().await {
-            let list = list?;
-            combined.versions.extend(list.versions.into_iter());
-        }
-        Ok(combined)
-    }
 }
 
 trait ListingCollector {
@@ -522,10 +488,27 @@ impl RemoteStorage for AzureBlobStorage {
         max_keys: Option<NonZeroU32>,
         cancel: &CancellationToken,
     ) -> std::result::Result<crate::VersionListing, DownloadError> {
+        let customize_builder = |mut builder: ListBlobsBuilder| {
+            builder = builder.include_versions(true);
+            builder
+        };
         let kind = RequestKind::ListVersions;
-        let permit = self.permit(kind, cancel).await?;
-        self.list_versions_with_permit(&permit, prefix, mode, max_keys, cancel)
-            .await
+
+        let mut stream = std::pin::pin!(self.list_streaming_for_fn(
+            prefix,
+            mode,
+            max_keys,
+            cancel,
+            kind,
+            customize_builder
+        ));
+        let mut combined: crate::VersionListing =
+            stream.next().await.expect("At least one item required")?;
+        while let Some(list) = stream.next().await {
+            let list = list?;
+            combined.versions.extend(list.versions.into_iter());
+        }
+        Ok(combined)
     }
 
     async fn head_object(
@@ -820,159 +803,14 @@ impl RemoteStorage for AzureBlobStorage {
 
     async fn time_travel_recover(
         &self,
-        prefix: Option<&RemotePath>,
-        timestamp: SystemTime,
-        done_if_after: SystemTime,
-        cancel: &CancellationToken,
-        _complexity_limit: Option<NonZeroU32>,
+        _prefix: Option<&RemotePath>,
+        _timestamp: SystemTime,
+        _done_if_after: SystemTime,
+        _cancel: &CancellationToken,
     ) -> Result<(), TimeTravelError> {
-        let msg = "PLEASE NOTE: Azure Blob storage time-travel recovery may not work as expected "
-            .to_string()
-            + "for some specific files. If a file gets deleted but then overwritten and we want to recover "
-            + "to the time during the file was not present, this functionality will recover the file. Only "
-            + "use the functionality for services that can tolerate this. For example, recovering a state of the "
-            + "pageserver tenants.";
-        tracing::error!("{}", msg);
-
-        let kind = RequestKind::TimeTravel;
-        let permit = self.permit(kind, cancel).await?;
-
-        let mode = ListingMode::NoDelimiter;
-        let version_listing = self
-            .list_versions_with_permit(&permit, prefix, mode, None, cancel)
-            .await
-            .map_err(|err| match err {
-                DownloadError::Other(e) => TimeTravelError::Other(e),
-                DownloadError::Cancelled => TimeTravelError::Cancelled,
-                other => TimeTravelError::Other(other.into()),
-            })?;
-        let versions_and_deletes = version_listing.versions;
-
-        tracing::info!(
-            "Built list for time travel with {} versions and deletions",
-            versions_and_deletes.len()
-        );
-
-        // Work on the list of references instead of the objects directly,
-        // otherwise we get lifetime errors in the sort_by_key call below.
-        let mut versions_and_deletes = versions_and_deletes.iter().collect::<Vec<_>>();
-
-        versions_and_deletes.sort_by_key(|vd| (&vd.key, &vd.last_modified));
-
-        let mut vds_for_key = HashMap::<_, Vec<_>>::new();
-
-        for vd in &versions_and_deletes {
-            let Version { key, .. } = &vd;
-            let version_id = vd.version_id().map(|v| v.0.as_str());
-            if version_id == Some("null") {
-                return Err(TimeTravelError::Other(anyhow!(
-                    "Received ListVersions response for key={key} with version_id='null', \
-                        indicating either disabled versioning, or legacy objects with null version id values"
-                )));
-            }
-            tracing::trace!("Parsing version key={key} kind={:?}", vd.kind);
-
-            vds_for_key.entry(key).or_default().push(vd);
-        }
-
-        let warn_threshold = 3;
-        let max_retries = 10;
-        let is_permanent = |e: &_| matches!(e, TimeTravelError::Cancelled);
-
-        for (key, versions) in vds_for_key {
-            let last_vd = versions.last().unwrap();
-            let key = self.relative_path_to_name(key);
-            if last_vd.last_modified > done_if_after {
-                tracing::debug!("Key {key} has version later than done_if_after, skipping");
-                continue;
-            }
-            // the version we want to restore to.
-            let version_to_restore_to =
-                match versions.binary_search_by_key(&timestamp, |tpl| tpl.last_modified) {
-                    Ok(v) => v,
-                    Err(e) => e,
-                };
-            if version_to_restore_to == versions.len() {
-                tracing::debug!("Key {key} has no changes since timestamp, skipping");
-                continue;
-            }
-            let mut do_delete = false;
-            if version_to_restore_to == 0 {
-                // All versions more recent, so the key didn't exist at the specified time point.
-                tracing::debug!(
-                    "All {} versions more recent for {key}, deleting",
-                    versions.len()
-                );
-                do_delete = true;
-            } else {
-                match &versions[version_to_restore_to - 1] {
-                    Version {
-                        kind: VersionKind::Version(version_id),
-                        ..
-                    } => {
-                        let source_url = format!(
-                            "{}/{}?versionid={}",
-                            self.client
-                                .url()
-                                .map_err(|e| TimeTravelError::Other(anyhow!("{e}")))?,
-                            key,
-                            version_id.0
-                        );
-                        tracing::debug!(
-                            "Promoting old version {} for {key} at {}...",
-                            version_id.0,
-                            source_url
-                        );
-                        backoff::retry(
-                            || async {
-                                let blob_client = self.client.blob_client(key.clone());
-                                let op = blob_client.copy(Url::from_str(&source_url).unwrap());
-                                tokio::select! {
-                                    res = op => res.map_err(|e| TimeTravelError::Other(e.into())),
-                                    _ = cancel.cancelled() => Err(TimeTravelError::Cancelled),
-                                }
-                            },
-                            is_permanent,
-                            warn_threshold,
-                            max_retries,
-                            "copying object version for time_travel_recover",
-                            cancel,
-                        )
-                        .await
-                        .ok_or_else(|| TimeTravelError::Cancelled)
-                        .and_then(|x| x)?;
-                        tracing::info!(?version_id, %key, "Copied old version in Azure blob storage");
-                    }
-                    Version {
-                        kind: VersionKind::DeletionMarker,
-                        ..
-                    } => {
-                        do_delete = true;
-                    }
-                }
-            };
-            if do_delete {
-                if matches!(last_vd.kind, VersionKind::DeletionMarker) {
-                    // Key has since been deleted (but there was some history), no need to do anything
-                    tracing::debug!("Key {key} already deleted, skipping.");
-                } else {
-                    tracing::debug!("Deleting {key}...");
-
-                    self.delete(&RemotePath::from_string(&key).unwrap(), cancel)
-                        .await
-                        .map_err(|e| {
-                            // delete_oid0 will use TimeoutOrCancel
-                            if TimeoutOrCancel::caused_by_cancel(&e) {
-                                TimeTravelError::Cancelled
-                            } else {
-                                TimeTravelError::Other(e)
-                            }
-                        })?;
-                }
-            }
-        }
-
-        Ok(())
+        // TODO use Azure point in time recovery feature for this
+        // https://learn.microsoft.com/en-us/azure/storage/blobs/point-in-time-restore-overview
+        Err(TimeTravelError::Unimplemented)
     }
 }
 

libs/remote_storage/src/lib.rs

@@ -440,7 +440,6 @@ pub trait RemoteStorage: Send + Sync + 'static {
         timestamp: SystemTime,
         done_if_after: SystemTime,
         cancel: &CancellationToken,
-        complexity_limit: Option<NonZeroU32>,
     ) -> Result<(), TimeTravelError>;
 }
 
@@ -652,23 +651,22 @@ impl<Other: RemoteStorage> GenericRemoteStorage<Arc<Other>> {
         timestamp: SystemTime,
         done_if_after: SystemTime,
         cancel: &CancellationToken,
-        complexity_limit: Option<NonZeroU32>,
     ) -> Result<(), TimeTravelError> {
         match self {
             Self::LocalFs(s) => {
-                s.time_travel_recover(prefix, timestamp, done_if_after, cancel, complexity_limit)
+                s.time_travel_recover(prefix, timestamp, done_if_after, cancel)
                     .await
             }
             Self::AwsS3(s) => {
-                s.time_travel_recover(prefix, timestamp, done_if_after, cancel, complexity_limit)
+                s.time_travel_recover(prefix, timestamp, done_if_after, cancel)
                     .await
             }
             Self::AzureBlob(s) => {
-                s.time_travel_recover(prefix, timestamp, done_if_after, cancel, complexity_limit)
+                s.time_travel_recover(prefix, timestamp, done_if_after, cancel)
                     .await
             }
             Self::Unreliable(s) => {
-                s.time_travel_recover(prefix, timestamp, done_if_after, cancel, complexity_limit)
+                s.time_travel_recover(prefix, timestamp, done_if_after, cancel)
                     .await
             }
         }

libs/remote_storage/src/local_fs.rs

@@ -610,7 +610,6 @@ impl RemoteStorage for LocalFs {
         _timestamp: SystemTime,
         _done_if_after: SystemTime,
         _cancel: &CancellationToken,
-        _complexity_limit: Option<NonZeroU32>,
     ) -> Result<(), TimeTravelError> {
         Err(TimeTravelError::Unimplemented)
     }

libs/remote_storage/src/s3_bucket.rs

@@ -981,16 +981,22 @@ impl RemoteStorage for S3Bucket {
         timestamp: SystemTime,
         done_if_after: SystemTime,
         cancel: &CancellationToken,
-        complexity_limit: Option<NonZeroU32>,
     ) -> Result<(), TimeTravelError> {
         let kind = RequestKind::TimeTravel;
         let permit = self.permit(kind, cancel).await?;
 
         tracing::trace!("Target time: {timestamp:?}, done_if_after {done_if_after:?}");
 
+        // Limit the number of versions deletions, mostly so that we don't
+        // keep requesting forever if the list is too long, as we'd put the
+        // list in RAM.
+        // Building a list of 100k entries that reaches the limit roughly takes
+        // 40 seconds, and roughly corresponds to tenants of 2 TiB physical size.
+        const COMPLEXITY_LIMIT: Option<NonZeroU32> = NonZeroU32::new(100_000);
+
         let mode = ListingMode::NoDelimiter;
         let version_listing = self
-            .list_versions_with_permit(&permit, prefix, mode, complexity_limit, cancel)
+            .list_versions_with_permit(&permit, prefix, mode, COMPLEXITY_LIMIT, cancel)
             .await
             .map_err(|err| match err {
                 DownloadError::Other(e) => TimeTravelError::Other(e),
@@ -1016,7 +1022,6 @@ impl RemoteStorage for S3Bucket {
             let Version { key, .. } = &vd;
             let version_id = vd.version_id().map(|v| v.0.as_str());
             if version_id == Some("null") {
-                // TODO: check the behavior of using the SDK on a non-versioned container
                 return Err(TimeTravelError::Other(anyhow!(
                     "Received ListVersions response for key={key} with version_id='null', \
                     indicating either disabled versioning, or legacy objects with null version id values"

libs/remote_storage/src/simulate_failures.rs

@@ -240,12 +240,11 @@ impl RemoteStorage for UnreliableWrapper {
         timestamp: SystemTime,
         done_if_after: SystemTime,
         cancel: &CancellationToken,
-        complexity_limit: Option<NonZeroU32>,
     ) -> Result<(), TimeTravelError> {
         self.attempt(RemoteOp::TimeTravelRecover(prefix.map(|p| p.to_owned())))
             .map_err(TimeTravelError::Other)?;
         self.inner
-            .time_travel_recover(prefix, timestamp, done_if_after, cancel, complexity_limit)
+            .time_travel_recover(prefix, timestamp, done_if_after, cancel)
             .await
     }
 }

libs/remote_storage/tests/test_real_s3.rs

@@ -157,7 +157,7 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     // No changes after recovery to t2 (no-op)
     let t_final = time_point().await;
     ctx.client
-        .time_travel_recover(None, t2, t_final, &cancel, None)
+        .time_travel_recover(None, t2, t_final, &cancel)
         .await?;
     let t2_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t2: {t2_files_recovered:?}");
@@ -173,7 +173,7 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     // after recovery to t1: path1 is back, path2 has the old content
     let t_final = time_point().await;
     ctx.client
-        .time_travel_recover(None, t1, t_final, &cancel, None)
+        .time_travel_recover(None, t1, t_final, &cancel)
         .await?;
     let t1_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t1: {t1_files_recovered:?}");
@@ -189,7 +189,7 @@ async fn s3_time_travel_recovery_works(ctx: &mut MaybeEnabledStorage) -> anyhow:
     // after recovery to t0: everything is gone except for path1
     let t_final = time_point().await;
     ctx.client
-        .time_travel_recover(None, t0, t_final, &cancel, None)
+        .time_travel_recover(None, t0, t_final, &cancel)
         .await?;
     let t0_files_recovered = list_files(&ctx.client, &cancel).await?;
     println!("after recovery to t0: {t0_files_recovered:?}");

libs/safekeeper_api/src/models.rs

@@ -13,7 +13,7 @@ use utils::pageserver_feedback::PageserverFeedback;
 use crate::membership::Configuration;
 use crate::{ServerInfo, Term};
 
-#[derive(Debug, Serialize, Deserialize)]
+#[derive(Debug, Serialize)]
 pub struct SafekeeperStatus {
     pub id: NodeId,
 }

pageserver/Cargo.toml

@@ -51,6 +51,7 @@ pageserver_api.workspace = true
 pageserver_client.workspace = true # for ResponseErrorMessageExt TOOD refactor that
 pageserver_compaction.workspace = true
 pageserver_page_api.workspace = true
+peekable.workspace = true
 pem.workspace = true
 pin-project-lite.workspace = true
 postgres_backend.workspace = true
@@ -62,6 +63,7 @@ postgres-types.workspace = true
 posthog_client_lite.workspace = true
 pprof.workspace = true
 pq_proto.workspace = true
+prost.workspace = true
 rand.workspace = true
 range-set-blaze = { version = "0.1.16", features = ["alloc"] }
 regex.workspace = true

pageserver/client_grpc/Cargo.toml

@@ -0,0 +1,30 @@
+[package]
+name = "pageserver_client_grpc"
+version = "0.1.0"
+edition = "2024"
+
+[dependencies]
+bytes.workspace = true
+futures.workspace = true
+http.workspace = true
+thiserror.workspace = true
+tonic.workspace = true
+tracing.workspace = true
+tokio = { version = "1.43.1", features = ["full", "macros", "net", "io-util", "rt", "rt-multi-thread"] }
+uuid = { version = "1", features = ["v4"] }
+tower = {  version = "0.4", features = ["timeout", "util"] }
+rand = "0.8"
+tokio-util = { version = "0.7", features = ["compat"] }
+hyper-util = "0.1.9"
+hyper = "1.6.0"
+metrics.workspace = true
+priority-queue = "2.3.1"
+async-trait = { version = "0.1" }
+tokio-stream = "0.1"
+dashmap = "5"
+chrono = { version = "0.4", features = ["serde"] }
+
+
+pageserver_page_api.workspace = true
+pageserver_api.workspace = true
+utils.workspace = true

pageserver/client_grpc/examples/load_test.rs

@@ -0,0 +1,296 @@
+// examples/load_test.rs, generated by AI
+
+use std::collections::{HashMap, HashSet};
+use std::sync::{
+    Arc,
+    Mutex,
+    atomic::{AtomicU64, AtomicUsize, Ordering},
+};
+use std::time::{Duration, Instant};
+
+use tokio::task;
+use tokio::time::sleep;
+use rand::Rng;
+use tonic::Status;
+use uuid::Uuid;
+
+// Pull in your ConnectionPool and PooledItemFactory from the pageserver_client_grpc crate.
+// Adjust these paths if necessary.
+use pageserver_client_grpc::client_cache::ConnectionPool;
+use pageserver_client_grpc::client_cache::PooledItemFactory;
+
+// --------------------------------------
+// GLOBAL COUNTERS FOR “CREATED” / “DROPPED” MockConnections
+// --------------------------------------
+static CREATED: AtomicU64 = AtomicU64::new(0);
+static DROPPED: AtomicU64 = AtomicU64::new(0);
+
+// --------------------------------------
+// MockConnection + Factory
+// --------------------------------------
+
+#[derive(Debug)]
+pub struct MockConnection {
+    pub id: u64,
+}
+
+impl Clone for MockConnection {
+    fn clone(&self) -> Self {
+        // Cloning a MockConnection does NOT count as “creating” a brand‐new connection,
+        // so we do NOT bump CREATED here. We only bump CREATED in the factory’s `create()`.
+        CREATED.fetch_add(1, Ordering::Relaxed);
+        MockConnection { id: self.id }
+    }
+}
+
+impl Drop for MockConnection {
+    fn drop(&mut self) {
+        // When a MockConnection actually gets dropped, bump the counter.
+        DROPPED.fetch_add(1, Ordering::SeqCst);
+    }
+}
+
+pub struct MockConnectionFactory {
+    counter: AtomicU64,
+}
+
+impl MockConnectionFactory {
+    pub fn new() -> Self {
+        MockConnectionFactory {
+            counter: AtomicU64::new(1),
+        }
+    }
+}
+
+#[async_trait::async_trait]
+impl PooledItemFactory<MockConnection> for MockConnectionFactory {
+    /// The trait on ConnectionPool expects:
+    ///   async fn create(&self, timeout: Duration)
+    ///       -> Result<Result<MockConnection, Status>, tokio::time::error::Elapsed>;
+    ///
+    /// On success: Ok(Ok(MockConnection))
+    /// On a simulated “gRPC” failure: Ok(Err(Status::…))
+    /// On a transport/factory error: Err(Box<…>)
+    async fn create(
+        &self,
+        _timeout: Duration,
+    ) -> Result<Result<MockConnection, Status>, tokio::time::error::Elapsed> {
+        // Simulate connection creation immediately succeeding.
+        CREATED.fetch_add(1, Ordering::SeqCst);
+        let next_id = self.counter.fetch_add(1, Ordering::Relaxed);
+        Ok(Ok(MockConnection { id: next_id }))
+    }
+}
+
+// --------------------------------------
+// CLIENT WORKER
+// --------------------------------------
+//
+// Each worker repeatedly calls `pool.get_client().await`. When it succeeds, we:
+//  1. Lock the shared Mutex<HashMap<u64, Arc<AtomicUsize>>> to fetch/insert an Arc<AtomicUsize> for this conn_id.
+//  2. Lock the shared Mutex<HashSet<u64>> to record this conn_id as “seen.”
+//  3. Drop both locks, then atomically increment that counter and assert it ≤ max_consumers.
+//  4. Sleep 10–100 ms to simulate “work.”
+//  5. Atomically decrement the counter.
+//  6. Call `pooled.finish(Ok(()))` to return to the pool.
+
+async fn client_worker(
+    pool: Arc<ConnectionPool<MockConnection>>,
+    usage_map: Arc<Mutex<HashMap<u64, Arc<AtomicUsize>>>>,
+    seen_set: Arc<Mutex<HashSet<u64>>>,
+    max_consumers: usize,
+    worker_id: usize,
+) {
+    for iteration in 0..10 {
+        match pool.clone().get_client().await {
+            Ok(pooled) => {
+                let conn: MockConnection = pooled.channel();
+                let conn_id = conn.id;
+
+                // 1. Fetch or insert the Arc<AtomicUsize> for this conn_id:
+                let counter_arc: Arc<AtomicUsize> = {
+                    let mut guard = usage_map.lock().unwrap();
+                    guard
+                        .entry(conn_id)
+                        .or_insert_with(|| Arc::new(AtomicUsize::new(0)))
+                        .clone()
+                    // MutexGuard is dropped here
+                };
+
+                // 2. Record this conn_id in the shared HashSet of “seen” IDs:
+                {
+                    let mut seen_guard = seen_set.lock().unwrap();
+                    seen_guard.insert(conn_id);
+                    // MutexGuard is dropped immediately
+                }
+
+                // 3. Atomically bump the count for this connection ID
+                let prev = counter_arc.fetch_add(1, Ordering::SeqCst);
+                let current = prev + 1;
+                assert!(
+                    current <= max_consumers,
+                    "Connection {} exceeded max_consumers (got {})",
+                    conn_id,
+                    current
+                );
+
+                println!(
+                    "[worker {}][iter {}] got MockConnection id={} ({} concurrent)",
+                    worker_id, iteration, conn_id, current
+                );
+
+                // 4. Simulate some work (10–100 ms)
+                let delay_ms = rand::thread_rng().gen_range(10..100);
+                sleep(Duration::from_millis(delay_ms)).await;
+
+                // 5. Decrement the usage counter
+                let prev2 = counter_arc.fetch_sub(1, Ordering::SeqCst);
+                let after = prev2 - 1;
+                println!(
+                    "[worker {}][iter {}] returning MockConnection id={} (now {} remain)",
+                    worker_id, iteration, conn_id, after
+                );
+
+                // 6. Return to the pool (mark success)
+                pooled.finish(Ok(())).await;
+            }
+            Err(status) => {
+                eprintln!(
+                    "[worker {}][iter {}] failed to get client: {:?}",
+                    worker_id, iteration, status
+                );
+            }
+        }
+
+        // Small random pause before next iteration to spread out load
+        let pause = rand::thread_rng().gen_range(0..20);
+        sleep(Duration::from_millis(pause)).await;
+    }
+}
+
+#[tokio::main(flavor = "multi_thread", worker_threads = 8)]
+async fn main() {
+    // --------------------------------------
+    // 1. Create factory and shared instrumentation
+    // --------------------------------------
+    let factory = Arc::new(MockConnectionFactory::new());
+
+    // Shared map: connection ID → Arc<AtomicUsize>
+    let usage_map: Arc<Mutex<HashMap<u64, Arc<AtomicUsize>>>> =
+        Arc::new(Mutex::new(HashMap::new()));
+
+    // Shared set: record each unique connection ID we actually saw
+    let seen_set: Arc<Mutex<HashSet<u64>>> = Arc::new(Mutex::new(HashSet::new()));
+
+    // --------------------------------------
+    // 2. Pool parameters
+    // --------------------------------------
+    let connect_timeout    = Duration::from_millis(500);
+    let connect_backoff    = Duration::from_millis(100);
+    let max_consumers      = 100;                 // test limit
+    let error_threshold    = 2;                 // mock never fails
+    let max_idle_duration  = Duration::from_secs(2);
+    let max_total_connections  = 3;
+    let aggregate_metrics  = None;
+
+    let pool: Arc<ConnectionPool<MockConnection>> = ConnectionPool::new(
+        factory,
+        connect_timeout,
+        connect_backoff,
+        max_consumers,
+        error_threshold,
+        max_idle_duration,
+        max_total_connections,
+        aggregate_metrics,
+    );
+
+    // --------------------------------------
+    // 3. Spawn worker tasks
+    // --------------------------------------
+    let num_workers = 10000;
+    let mut handles = Vec::with_capacity(num_workers);
+    let start_time = Instant::now();
+
+    for worker_id in 0..num_workers {
+        let pool_clone   = Arc::clone(&pool);
+        let usage_clone  = Arc::clone(&usage_map);
+        let seen_clone   = Arc::clone(&seen_set);
+        let mc           = max_consumers;
+
+        let handle = task::spawn(async move {
+            client_worker(pool_clone, usage_clone, seen_clone, mc, worker_id).await;
+        });
+        handles.push(handle);
+    }
+
+    // --------------------------------------
+    // 4. Wait for workers to finish
+    // --------------------------------------
+    for handle in handles {
+        let _ = handle.await;
+    }
+    let elapsed = Instant::now().duration_since(start_time);
+    println!(
+        "All {} workers completed in {:?}",
+        num_workers, elapsed
+    );
+
+    // --------------------------------------
+    // 5. Print the total number of unique connections seen so far
+    // --------------------------------------
+    let unique_count = {
+        let seen_guard = seen_set.lock().unwrap();
+        seen_guard.len()
+    };
+    println!("Total unique connections used by workers: {}", unique_count);
+
+    // --------------------------------------
+    // 6. Sleep so the background sweeper can run (max_idle_duration = 2 s)
+    // --------------------------------------
+    sleep(Duration::from_secs(3)).await;
+
+    // --------------------------------------
+    // 7. Shutdown the pool
+    // --------------------------------------
+    let shutdown_pool = Arc::clone(&pool);
+    shutdown_pool.shutdown().await;
+    println!("Pool.shutdown() returned.");
+
+    // --------------------------------------
+    // 8. Verify that no background task still holds an Arc clone of `pool`.
+    //    If any task is still alive (sweeper/create_connection), strong_count > 1.
+    // --------------------------------------
+    sleep(Duration::from_secs(1)).await; // give tasks time to exit
+    let sc = Arc::strong_count(&pool);
+    assert!(
+        sc == 1,
+        "Pool tasks did not all terminate: Arc::strong_count = {} (expected 1)",
+        sc
+    );
+    println!("Verified: all pool tasks have terminated (strong_count == 1).");
+
+    // --------------------------------------
+    // 9. Verify no MockConnection was leaked:
+    //    CREATED must equal DROPPED.
+    // --------------------------------------
+    let created = CREATED.load(Ordering::SeqCst);
+    let dropped = DROPPED.load(Ordering::SeqCst);
+    assert!(
+        created == dropped,
+        "Leaked connections: created={} but dropped={}",
+        created,
+        dropped
+    );
+    println!(
+        "Verified: no connections leaked (created = {}, dropped = {}).",
+        created, dropped
+    );
+
+    // --------------------------------------
+    // 10. Because `client_worker` asserted inside that no connection
+    //     ever exceeded `max_consumers`, reaching this point means that check passed.
+    // --------------------------------------
+    println!("All per-connection usage stayed within max_consumers = {}.", max_consumers);
+
+    println!("Load test complete; exiting cleanly.");
+}

pageserver/client_grpc/examples/request_tracker_load_test.rs

@@ -0,0 +1,160 @@
+// examples/request_tracker_load_test.rs
+
+use std::{sync::Arc, time::Duration};
+use tokio;
+use pageserver_client_grpc::request_tracker::RequestTracker;
+use pageserver_client_grpc::request_tracker::MockStreamFactory;
+use pageserver_client_grpc::request_tracker::StreamReturner;
+use pageserver_client_grpc::client_cache::ConnectionPool;
+use pageserver_client_grpc::client_cache::PooledItemFactory;
+use pageserver_client_grpc::ClientCacheOptions;
+use pageserver_client_grpc::PageserverClientAggregateMetrics;
+use pageserver_client_grpc::AuthInterceptor;
+
+use pageserver_client_grpc::client_cache::ChannelFactory;
+
+use tonic::{transport::{Channel}, Request};
+
+use rand::prelude::*;
+
+use pageserver_api::key::Key;
+
+use utils::lsn::Lsn;
+use utils::id::TenantTimelineId;
+
+use futures::stream::FuturesOrdered;
+use futures::StreamExt;
+// use chrono
+use chrono::Utc;
+
+use pageserver_page_api::{GetPageClass, GetPageResponse};
+use pageserver_page_api::proto;
+#[derive(Clone)]
+struct KeyRange {
+    timeline: TenantTimelineId,
+    timeline_lsn: Lsn,
+    start: i128,
+    end: i128,
+}
+
+impl KeyRange {
+    fn len(&self) -> i128 {
+        self.end - self.start
+    }
+}
+
+#[tokio::main]
+async fn main() {
+    // 1) configure the client‐pool behavior
+    let client_cache_options = ClientCacheOptions {
+        max_delay_ms:       0,
+        drop_rate:          0.0,
+        hang_rate:          0.0,
+        connect_timeout:    Duration::from_secs(10),
+        connect_backoff:    Duration::from_millis(200),
+        max_consumers:      64,
+        error_threshold:    10,
+        max_idle_duration:  Duration::from_secs(60),
+        max_total_connections: 12,
+    };
+
+    // 2) metrics collector (we assume Default is implemented)
+    let metrics = Arc::new(PageserverClientAggregateMetrics::new());
+    let pool = ConnectionPool::<StreamReturner>::new(
+        Arc::new(MockStreamFactory::new(
+        )),
+        client_cache_options.connect_timeout,
+        client_cache_options.connect_backoff,
+        client_cache_options.max_consumers,
+        client_cache_options.error_threshold,
+        client_cache_options.max_idle_duration,
+        client_cache_options.max_total_connections,
+        Some(Arc::clone(&metrics)),
+    );
+
+    // -----------
+    // There is no mock for the unary connection pool, so for now just
+    // don't use this pool
+    //
+    let channel_fact : Arc<dyn PooledItemFactory<Channel> + Send + Sync> = Arc::new(ChannelFactory::new(
+        "".to_string(),
+        client_cache_options.max_delay_ms,
+        client_cache_options.drop_rate,
+        client_cache_options.hang_rate,
+    ));
+    let unary_pool: Arc<ConnectionPool<Channel>> = ConnectionPool::new(
+        Arc::clone(&channel_fact),
+        client_cache_options.connect_timeout,
+        client_cache_options.connect_backoff,
+        client_cache_options.max_consumers,
+        client_cache_options.error_threshold,
+        client_cache_options.max_idle_duration,
+        client_cache_options.max_total_connections,
+        Some(Arc::clone(&metrics)),
+    );
+
+    // -----------
+    // Dummy auth interceptor. This is not used in this test.
+    let auth_interceptor = AuthInterceptor::new("dummy_tenant_id",
+                                                "dummy_timeline_id",
+                                                None);
+    let mut tracker = RequestTracker::new(
+        pool,
+        unary_pool,
+        auth_interceptor,
+    );
+
+    // 4) fire off 10 000 requests in parallel
+    let mut handles = FuturesOrdered::new();
+    for i in 0..500000 {
+
+            let mut rng = rand::thread_rng();
+            let r = 0..=1000000i128;
+            let key: i128 = rng.gen_range(r.clone());
+            let key = Key::from_i128(key);
+            let (rel_tag, block_no) = key
+                .to_rel_block()
+                .expect("we filter non-rel-block keys out above");
+
+            let req2 = proto::GetPageRequest {
+                request_id: 0,
+                request_class: proto::GetPageClass::Normal as i32,
+                read_lsn: Some(proto::ReadLsn {
+                    request_lsn: if rng.gen_bool(0.5) {
+                        u64::from(Lsn::MAX)
+                    } else {
+                        10000
+                    },
+                    not_modified_since_lsn: 10000,
+                }),
+                rel: Some(rel_tag.into()),
+                block_number: vec![block_no],
+            };
+        let req_model = pageserver_page_api::GetPageRequest::try_from(req2.clone());
+
+        // RequestTracker is Clone, so we can share it
+        let mut tr = tracker.clone();
+        let fut = async move {
+            let resp = tr.send_getpage_request(req_model.unwrap()).await.unwrap();
+            // sanity‐check: the mock echo returns the same request_id
+            assert!(resp.request_id > 0);
+        };
+        handles.push_back(fut);
+
+        // empty future
+        let fut = async move {};
+        fut.await;
+    }
+
+    // print timestamp
+    println!("Starting 5000000 requests at: {}", chrono::Utc::now());
+    // 5) wait for them all
+    for i in 0..500000 {
+        handles.next().await.expect("Failed to get next handle");
+    }
+
+    // print timestamp
+    println!("Finished 5000000 requests at: {}", chrono::Utc::now());
+
+    println!("✅ All 100000 requests completed successfully");
+}

pageserver/client_grpc/src/client_cache.rs

@@ -0,0 +1,741 @@
+use std::{
+    collections::HashMap,
+    io::{self, Error, ErrorKind},
+    sync::Arc,
+    time::{Duration, Instant},
+};
+
+use priority_queue::PriorityQueue;
+
+use tokio::{
+    io::{AsyncRead, AsyncWrite, ReadBuf},
+    net::TcpStream,
+    sync::{Mutex, OwnedSemaphorePermit, Semaphore},
+    time::sleep,
+};
+use tonic::transport::{Channel, Endpoint};
+
+use uuid;
+
+use std::{
+    pin::Pin,
+    task::{Context, Poll},
+};
+
+use futures::future;
+use rand::{Rng, SeedableRng, rngs::StdRng};
+
+use bytes::BytesMut;
+use http::Uri;
+use hyper_util::rt::TokioIo;
+use tower::service_fn;
+
+use tokio_util::sync::CancellationToken;
+use async_trait::async_trait;
+
+//
+// The "TokioTcp" is flakey TCP network for testing purposes, in order
+// to simulate network errors and delays.
+//
+
+/// Wraps a `TcpStream`, buffers incoming data, and injects a random delay per fresh read/write.
+pub struct TokioTcp {
+    tcp: TcpStream,
+    /// Maximum randomized delay in milliseconds
+    delay_ms: u64,
+
+    /// Next deadline instant for delay
+    deadline: Instant,
+    /// Internal buffer of previously-read data
+    buffer: BytesMut,
+}
+
+impl TokioTcp {
+    /// Create a new wrapper with given max delay (ms)
+    pub fn new(stream: TcpStream, delay_ms: u64) -> Self {
+        let initial = if delay_ms > 0 {
+            rand::thread_rng().gen_range(0..delay_ms)
+        } else {
+            0
+        };
+        let deadline = Instant::now() + Duration::from_millis(initial);
+        TokioTcp {
+            tcp: stream,
+            delay_ms,
+            deadline,
+            buffer: BytesMut::new(),
+        }
+    }
+}
+
+impl AsyncRead for TokioTcp {
+    fn poll_read(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        buf: &mut ReadBuf<'_>,
+    ) -> Poll<io::Result<()>> {
+        // Safe because TokioTcp is Unpin
+        let this = self.get_mut();
+
+        // 1) Drain any buffered data
+        if !this.buffer.is_empty() {
+            let to_copy = this.buffer.len().min(buf.remaining());
+            buf.put_slice(&this.buffer.split_to(to_copy));
+            return Poll::Ready(Ok(()));
+        }
+
+        // 2) If we're still before the deadline, schedule a wake and return Pending
+        let now = Instant::now();
+        if this.delay_ms > 0 && now < this.deadline {
+            let waker = cx.waker().clone();
+            let wait = this.deadline - now;
+            tokio::spawn(async move {
+                sleep(wait).await;
+                waker.wake_by_ref();
+            });
+            return Poll::Pending;
+        }
+
+        // 3) Past deadline: compute next random deadline
+        if this.delay_ms > 0 {
+            let next_ms = rand::thread_rng().gen_range(0..=this.delay_ms);
+            this.deadline = Instant::now() + Duration::from_millis(next_ms);
+        }
+
+        // 4) Perform actual read into a temporary buffer
+        let mut tmp = [0u8; 4096];
+        let mut rb = ReadBuf::new(&mut tmp);
+        match Pin::new(&mut this.tcp).poll_read(cx, &mut rb) {
+            Poll::Pending => Poll::Pending,
+            Poll::Ready(Ok(())) => {
+                let filled = rb.filled();
+                if filled.is_empty() {
+                    // EOF or zero bytes
+                    Poll::Ready(Ok(()))
+                } else {
+                    this.buffer.extend_from_slice(filled);
+                    let to_copy = this.buffer.len().min(buf.remaining());
+                    buf.put_slice(&this.buffer.split_to(to_copy));
+                    Poll::Ready(Ok(()))
+                }
+            }
+            Poll::Ready(Err(e)) => Poll::Ready(Err(e)),
+        }
+    }
+}
+
+impl AsyncWrite for TokioTcp {
+    fn poll_write(
+        self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        data: &[u8],
+    ) -> Poll<io::Result<usize>> {
+        let this = self.get_mut();
+
+        // 1) If before deadline, schedule wake and return Pending
+        let now = Instant::now();
+        if this.delay_ms > 0 && now < this.deadline {
+            let waker = cx.waker().clone();
+            let wait = this.deadline - now;
+            tokio::spawn(async move {
+                sleep(wait).await;
+                waker.wake_by_ref();
+            });
+            return Poll::Pending;
+        }
+
+        // 2) Past deadline: compute next random deadline
+        if this.delay_ms > 0 {
+            let next_ms = rand::thread_rng().gen_range(0..=this.delay_ms);
+            this.deadline = Instant::now() + Duration::from_millis(next_ms);
+        }
+
+        // 3) Actual write
+        Pin::new(&mut this.tcp).poll_write(cx, data)
+    }
+
+    fn poll_flush(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        let this = self.get_mut();
+        Pin::new(&mut this.tcp).poll_flush(cx)
+    }
+
+    fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        let this = self.get_mut();
+        Pin::new(&mut this.tcp).poll_shutdown(cx)
+    }
+}
+
+#[async_trait]
+pub trait PooledItemFactory<T>: Send + Sync + 'static {
+    /// Create a new pooled item.
+    async fn create(&self, connect_timeout: Duration) ->  Result<Result<T, tonic::Status>, tokio::time::error::Elapsed>;
+}
+
+pub struct ChannelFactory {
+    endpoint: String,
+    max_delay_ms: u64,
+    drop_rate: f64,
+    hang_rate: f64,
+}
+
+
+impl ChannelFactory {
+    pub fn new(
+        endpoint: String,
+        max_delay_ms: u64,
+        drop_rate: f64,
+        hang_rate: f64,
+    ) -> Self {
+        ChannelFactory {
+            endpoint,
+            max_delay_ms,
+            drop_rate,
+            hang_rate,
+        }
+    }
+}
+
+#[async_trait]
+impl PooledItemFactory<Channel> for ChannelFactory {
+    async fn create(&self, connect_timeout: Duration) -> Result<Result<Channel, tonic::Status>, tokio::time::error::Elapsed> {
+        let max_delay_ms = self.max_delay_ms;
+        let drop_rate = self.drop_rate;
+        let hang_rate = self.hang_rate;
+
+        // This is a custom connector that inserts delays and errors, for
+        // testing purposes. It would normally be disabled by the config.
+        let connector = service_fn(move |uri: Uri| {
+            let drop_rate = drop_rate;
+            let hang_rate = hang_rate;
+            async move {
+                let mut rng = StdRng::from_entropy();
+                // Simulate an indefinite hang
+                if hang_rate > 0.0 && rng.gen_bool(hang_rate) {
+                    // never completes, to test timeout
+                    return future::pending::<Result<TokioIo<TokioTcp>, std::io::Error>>().await;
+                }
+
+                // Random drop (connect error)
+                if drop_rate > 0.0 && rng.gen_bool(drop_rate) {
+                    return Err(std::io::Error::new(
+                        std::io::ErrorKind::Other,
+                        "simulated connect drop",
+                    ));
+                }
+
+                // Otherwise perform real TCP connect
+                let addr = match (uri.host(), uri.port()) {
+                    // host + explicit port
+                    (Some(host), Some(port)) => format!("{}:{}", host, port.as_str()),
+                    // host only (no port)
+                    (Some(host), None) => host.to_string(),
+                    // neither? error out
+                    _ => return Err(Error::new(ErrorKind::InvalidInput, "no host or port")),
+                };
+
+                let tcp = TcpStream::connect(addr).await?;
+                let tcpwrapper = TokioTcp::new(tcp, max_delay_ms);
+                Ok(TokioIo::new(tcpwrapper))
+            }
+        });
+
+
+        let attempt = tokio::time::timeout(
+            connect_timeout,
+            Endpoint::from_shared(self.endpoint.clone())
+                .expect("invalid endpoint")
+                .timeout(connect_timeout)
+                .connect_with_connector(connector),
+        )
+            .await;
+        match attempt {
+            Ok(Ok(channel)) => {
+                // Connection succeeded
+                Ok(Ok(channel))
+            }
+            Ok(Err(e)) => {
+                Ok(Err(tonic::Status::new(
+                    tonic::Code::Unavailable,
+                    format!("Failed to connect: {}", e),
+                )))
+            }
+            Err(e) => {
+                Err(e)
+            }
+        }
+    }
+}
+
+
+/// A pooled gRPC client with capacity tracking and error handling.
+pub struct ConnectionPool<T> {
+    inner: Mutex<Inner<T>>,
+
+    fact: Arc<dyn PooledItemFactory<T> + Send + Sync>,
+
+    connect_timeout: Duration,
+    connect_backoff: Duration,
+    /// The maximum number of consumers that can use a single connection.
+    max_consumers: usize,
+    /// The number of consecutive errors before a connection is removed from the pool.
+    error_threshold: usize,
+    /// The maximum duration a connection can be idle before being removed.
+    max_idle_duration: Duration,
+    max_total_connections: usize,
+
+    channel_semaphore: Arc<Semaphore>,
+
+    shutdown_token: CancellationToken,
+    aggregate_metrics: Option<Arc<crate::PageserverClientAggregateMetrics>>,
+}
+
+struct Inner<T> {
+    entries: HashMap<uuid::Uuid, ConnectionEntry<T>>,
+    pq: PriorityQueue<uuid::Uuid, usize>,
+    // This is updated when a connection is dropped, or we fail
+    // to create a new connection.
+    last_connect_failure: Option<Instant>,
+    waiters: usize,
+    in_progress: usize,
+}
+struct ConnectionEntry<T> {
+    channel: T,
+    active_consumers: usize,
+    consecutive_errors: usize,
+    last_used: Instant,
+}
+
+/// A client borrowed from the pool.
+pub struct PooledClient<T> {
+    pub channel: T,
+    pool: Arc<ConnectionPool<T>>,
+    is_ok: bool,
+    id: uuid::Uuid,
+    permit: OwnedSemaphorePermit,
+}
+
+impl<T: Clone + Send + 'static> ConnectionPool<T> {
+    pub fn new(
+        fact: Arc<dyn PooledItemFactory<T> + Send + Sync>,
+        connect_timeout: Duration,
+        connect_backoff: Duration,
+        max_consumers: usize,
+        error_threshold: usize,
+        max_idle_duration: Duration,
+        max_total_connections: usize,
+        aggregate_metrics: Option<Arc<crate::PageserverClientAggregateMetrics>>,
+    ) -> Arc<Self> {
+        let shutdown_token = CancellationToken::new();
+        let pool = Arc::new(Self {
+            inner: Mutex::new(Inner::<T> {
+                entries: HashMap::new(),
+                pq: PriorityQueue::new(),
+                last_connect_failure: None,
+                waiters: 0,
+                in_progress: 0,
+            }),
+            fact: Arc::clone(&fact),
+            connect_timeout,
+            connect_backoff,
+            max_consumers,
+            error_threshold,
+            max_idle_duration,
+            max_total_connections,
+            channel_semaphore: Arc::new(Semaphore::new(0)),
+            shutdown_token: shutdown_token.clone(),
+            aggregate_metrics: aggregate_metrics.clone(),
+        });
+
+        // Cancelable background task to sweep idle connections
+        let sweeper_token = shutdown_token.clone();
+        let sweeper_pool = Arc::clone(&pool);
+        tokio::spawn(async move {
+            loop {
+                tokio::select! {
+                    _ = sweeper_token.cancelled() => break,
+                    _ = async {
+                        sweeper_pool.sweep_idle_connections().await;
+                        sleep(Duration::from_secs(5)).await;
+                    } => {}
+                }
+            }
+        });
+
+        pool
+    }
+
+    pub async fn shutdown(self: Arc<Self>) {
+        self.shutdown_token.cancel();
+
+        loop {
+            let all_idle = {
+                let inner = self.inner.lock().await;
+                inner.entries.values().all(|e| e.active_consumers == 0)
+            };
+            if all_idle {
+                break;
+            }
+            sleep(Duration::from_millis(100)).await;
+        }
+
+        // 4. Remove all entries
+        let mut inner = self.inner.lock().await;
+        inner.entries.clear();
+    }
+
+    /// Sweep and remove idle connections safely, burning their permits.
+    async fn sweep_idle_connections(self: &Arc<Self>) {
+        let mut ids_to_remove = Vec::new();
+        let now = Instant::now();
+
+        // Remove idle entries. First collect permits for those connections so that
+        // no consumer will reserve them, then remove them from the pool.
+        {
+            let mut inner = self.inner.lock().await;
+            inner.entries.retain(|id, entry| {
+                if entry.active_consumers == 0
+                    && now.duration_since(entry.last_used) > self.max_idle_duration
+                {
+                    // metric
+                    match self.aggregate_metrics {
+                        Some(ref metrics) => {
+                            metrics
+                                .retry_counters
+                                .with_label_values(&["connection_swept"])
+                                .inc();
+                        }
+                        None => {}
+                    }
+                    ids_to_remove.push(*id);
+                    return false; // remove this entry
+                }
+                true
+            });
+            // Remove the entries from the priority queue
+            for id in ids_to_remove {
+                inner.pq.remove(&id);
+            }
+        }
+    }
+
+    // If we have a permit already, get a connection out of the heap
+    async fn get_conn_with_permit(
+        self: Arc<Self>,
+        permit: OwnedSemaphorePermit,
+    ) -> Option<PooledClient<T>> {
+        let mut inner = self.inner.lock().await;
+
+        // Pop the highest-active-consumers connection. There are no connections
+        // in the heap that have more than max_consumers active consumers.
+        if let Some((id, _cons)) = inner.pq.pop() {
+            let entry = inner
+                .entries
+                .get_mut(&id)
+                .expect("pq and entries got out of sync");
+
+            let mut active_consumers = entry.active_consumers;
+            entry.active_consumers += 1;
+            entry.last_used = Instant::now();
+
+            let client = PooledClient::<T> {
+                channel: entry.channel.clone(),
+                pool: Arc::clone(&self),
+                is_ok: true,
+                id,
+                permit: permit,
+            };
+
+            // re‐insert with updated priority
+            active_consumers += 1;
+            if active_consumers < self.max_consumers {
+                inner.pq.push(id, active_consumers as usize);
+            }
+            return Some(client);
+        } else {
+            // If there is no connection to take, it is because permits for a connection
+            // need to drain. This can happen if a connection is removed because it has
+            // too many errors. It is taken out of the heap/hash table in this case, but
+            // we can't remove it's permits until now.
+            //
+            // Just forget the permit and retry.
+            permit.forget();
+            return None;
+        }
+    }
+
+    pub async fn get_client(self: Arc<Self>) -> Result<PooledClient<T>, tonic::Status> {
+        // The pool is shutting down. Don't accept new connections.
+        if self.shutdown_token.is_cancelled() {
+            return Err(tonic::Status::unavailable("Pool is shutting down"));
+        }
+
+        // A loop is necessary because when a connection is draining, we have to return
+        // a permit and retry.
+        loop {
+            let self_clone = Arc::clone(&self);
+            let mut semaphore = Arc::clone(&self_clone.channel_semaphore);
+
+            match semaphore.try_acquire_owned() {
+                Ok(permit_) => {
+                    // We got a permit, so check the heap for a connection
+                    // we can use.
+                    let pool_conn = self_clone.get_conn_with_permit(permit_).await;
+                    match pool_conn {
+                        Some(pool_conn_) => {
+                            return Ok(pool_conn_);
+                        }
+                        None => {
+                            // No connection available. Forget the permit and retry.
+                            continue;
+                        }
+                    }
+                }
+                Err(_) => {
+                    match self_clone.aggregate_metrics {
+                        Some(ref metrics) => {
+                            metrics
+                                .retry_counters
+                                .with_label_values(&["sema_acquire_failed"])
+                                .inc();
+                        }
+                        None => {}
+                    }
+
+                    {
+                        //
+                        // This is going to generate enough connections to handle a burst,
+                        // but it may generate up to twice the number of connections needed
+                        // in the worst case. Extra connections will go idle and be cleaned
+                        // up.
+                        //
+                        let mut inner = self_clone.inner.lock().await;
+                        inner.waiters += 1;
+                        if inner.waiters > (inner.in_progress * self_clone.max_consumers) {
+                            if (inner.entries.len() + inner.in_progress) < self_clone.max_total_connections {
+
+                                let self_clone_spawn = Arc::clone(&self_clone);
+                                tokio::task::spawn(async move {
+                                    self_clone_spawn.create_connection().await;
+                                });
+                                inner.in_progress += 1;
+                            }
+
+                        }
+                    }
+                    // Wait for a connection to become available, either because it
+                    // was created or because a connection was returned to the pool
+                    // by another consumer.
+                    semaphore = Arc::clone(&self_clone.channel_semaphore);
+                    let conn_permit = semaphore.acquire_owned().await.unwrap();
+                    {
+                        let mut inner = self_clone.inner.lock().await;
+                        inner.waiters -= 1;
+                    }
+                    // We got a permit, check the heap for a connection.
+                    let pool_conn = self_clone.get_conn_with_permit(conn_permit).await;
+                    match pool_conn {
+                        Some(pool_conn_) => {
+                            return Ok(pool_conn_);
+                        }
+                        None => {
+                            // No connection was found, forget the permit and retry.
+                            continue;
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    async fn create_connection(&self) -> () {
+
+        // Generate a random backoff to add some jitter so that connections
+        // don't all retry at the same time.
+        let mut backoff_delay = Duration::from_millis(
+            rand::thread_rng().gen_range(0..=self.connect_backoff.as_millis() as u64),
+        );
+
+        loop {
+            if self.shutdown_token.is_cancelled() {
+                return;
+            }
+
+            // Back off.
+            // Loop because failure can occur while we are sleeping, so wait
+            // until the failure stopped for at least one backoff period. Backoff
+            // period includes some jitter, so that if multiple connections are
+            // failing, they don't all retry at the same time.
+            loop {
+                if let Some(delay) = {
+                    let inner = self.inner.lock().await;
+                    inner.last_connect_failure.and_then(|at| {
+                        (at.elapsed() < backoff_delay).then(|| backoff_delay - at.elapsed())
+                    })
+                } {
+                    sleep(delay).await;
+                } else {
+                    break; // No delay, so we can create a connection
+                }
+            }
+
+            //
+            // Create a new connection.
+            //
+            // The connect timeout is also the timeout for an individual gRPC request
+            // on this connection. (Requests made later on this channel will time out
+            // with the same timeout.)
+            //
+            match self.aggregate_metrics {
+                Some(ref metrics) => {
+                    metrics
+                        .retry_counters
+                        .with_label_values(&["connection_attempt"])
+                        .inc();
+                }
+                None => {}
+            }
+
+            let attempt = self.fact
+                .create(self.connect_timeout)
+                .await;
+
+            match attempt {
+                // Connection succeeded
+                Ok(Ok(channel)) => {
+                    {
+                        match self.aggregate_metrics {
+                            Some(ref metrics) => {
+                                metrics
+                                    .retry_counters
+                                    .with_label_values(&["connection_success"])
+                                    .inc();
+                            }
+                            None => {}
+                        }
+                        let mut inner = self.inner.lock().await;
+                        let id = uuid::Uuid::new_v4();
+                        inner.entries.insert(
+                            id,
+                            ConnectionEntry::<T> {
+                                channel: channel.clone(),
+                                active_consumers: 0,
+                                consecutive_errors: 0,
+                                last_used: Instant::now(),
+                            },
+                        );
+                        inner.pq.push(id, 0);
+                        inner.in_progress -= 1;
+                        self.channel_semaphore.add_permits(self.max_consumers);
+                        return;
+                    };
+                }
+                // Connection failed, back off and retry
+                Ok(Err(_)) | Err(_) => {
+                    match self.aggregate_metrics {
+                        Some(ref metrics) => {
+                            metrics
+                                .retry_counters
+                                .with_label_values(&["connect_failed"])
+                                .inc();
+                        }
+                        None => {}
+                    }
+                    let mut inner = self.inner.lock().await;
+                    inner.last_connect_failure = Some(Instant::now());
+                    // Add some jitter so that every connection doesn't retry at once
+                    let jitter = rand::thread_rng().gen_range(0..=backoff_delay.as_millis() as u64);
+                    backoff_delay =
+                        Duration::from_millis(backoff_delay.as_millis() as u64 + jitter);
+
+                    // Do not backoff longer than one minute
+                    if backoff_delay > Duration::from_secs(60) {
+                        backoff_delay = Duration::from_secs(60);
+                    }
+                    // continue the loop to retry
+                }
+            }
+        }
+    }
+
+    /// Return client to the pool, indicating success or error.
+    pub async fn return_client(&self, id: uuid::Uuid, success: bool, permit: OwnedSemaphorePermit) {
+        let mut inner = self.inner.lock().await;
+        if let Some(entry) = inner.entries.get_mut(&id) {
+            entry.last_used = Instant::now();
+            if entry.active_consumers <= 0 {
+                panic!("A consumer completed when active_consumers was zero!")
+            }
+            entry.active_consumers = entry.active_consumers - 1;
+            if success {
+                if entry.consecutive_errors < self.error_threshold {
+                    entry.consecutive_errors = 0;
+                }
+            } else {
+                entry.consecutive_errors += 1;
+                if entry.consecutive_errors == self.error_threshold {
+                    match self.aggregate_metrics {
+                        Some(ref metrics) => {
+                            metrics
+                                .retry_counters
+                                .with_label_values(&["connection_dropped"])
+                                .inc();
+                        }
+                        None => {}
+                    }
+                }
+            }
+
+            //
+            // Too many errors on this connection. If there are no active users,
+            // remove it. Otherwise just wait for active_consumers to go to zero.
+            // This connection will not be selected for new consumers.
+            //
+            let active_consumers = entry.active_consumers;
+            if entry.consecutive_errors >= self.error_threshold {
+                // too many errors, remove the connection permanently. Once it drains,
+                // it will be dropped.
+                if inner.pq.get_priority(&id).is_some() {
+                    inner.pq.remove(&id);
+                }
+
+                // remove from entries
+                // check if entry is in inner
+                if inner.entries.contains_key(&id) {
+                    inner.entries.remove(&id);
+                }
+                inner.last_connect_failure = Some(Instant::now());
+
+                // The connection has been removed, it's permits will be
+                // drained because if we look for a connection and it's not there
+                // we just forget the permit. However, this process can be a little
+                // bit faster if we just forget permits as the connections are returned.
+                permit.forget();
+            } else {
+                // update its priority in the queue
+                if inner.pq.get_priority(&id).is_some() {
+                    inner.pq.change_priority(&id, active_consumers);
+                } else {
+                    // This connection is not in the heap, but it has space
+                    // for more consumers. Put it back in the heap.
+                    if active_consumers < self.max_consumers {
+                        inner.pq.push(id, active_consumers);
+                    }
+                }
+            }
+        }
+    }
+}
+
+impl<T: Clone + Send + 'static> PooledClient<T> {
+    pub fn channel(&self) -> T {
+        return self.channel.clone();
+    }
+    pub async fn finish(mut self, result: Result<(), tonic::Status>) {
+        self.is_ok = result.is_ok();
+        self.pool.return_client(
+            self.id,
+            self.is_ok,
+            self.permit,
+        ).await;
+    }
+}

pageserver/client_grpc/src/lib.rs

@@ -0,0 +1,456 @@
+//! Pageserver Data API client
+//!
+//! - Manage connections to pageserver
+//! - Send requests to correct shards
+//!
+use std::collections::HashMap;
+use std::sync::Arc;
+use std::sync::RwLock;
+use std::time::Duration;
+
+use bytes::Bytes;
+use futures::{Stream, StreamExt};
+use thiserror::Error;
+use tonic::metadata::AsciiMetadataValue;
+
+use pageserver_page_api::proto;
+use pageserver_page_api::*;
+
+use pageserver_page_api::proto::PageServiceClient;
+use utils::shard::ShardIndex;
+
+use std::fmt::Debug;
+pub mod client_cache;
+pub mod request_tracker;
+use tonic::transport::Channel;
+
+use metrics::{IntCounterVec, core::Collector};
+use crate::client_cache::{PooledItemFactory};
+
+use tokio::sync::mpsc;
+use async_trait::async_trait;
+
+
+#[derive(Error, Debug)]
+pub enum PageserverClientError {
+    #[error("could not connect to service: {0}")]
+    ConnectError(#[from] tonic::transport::Error),
+    #[error("could not perform request: {0}`")]
+    RequestError(#[from] tonic::Status),
+    #[error("protocol error: {0}")]
+    ProtocolError(#[from] ProtocolError),
+
+    #[error("could not perform request: {0}`")]
+    InvalidUri(#[from] http::uri::InvalidUri),
+
+    #[error("could not perform request: {0}`")]
+    Other(String),
+}
+
+#[derive(Clone, Debug)]
+pub struct PageserverClientAggregateMetrics {
+    pub request_counters: IntCounterVec,
+    pub retry_counters: IntCounterVec,
+}
+impl PageserverClientAggregateMetrics {
+    pub fn new() -> Self {
+        let request_counters = IntCounterVec::new(
+            metrics::core::Opts::new(
+                "backend_requests_total",
+                "Number of requests from backends.",
+            ),
+            &["request_kind"],
+        )
+        .unwrap();
+
+        let retry_counters = IntCounterVec::new(
+            metrics::core::Opts::new(
+                "backend_requests_retries_total",
+                "Number of retried requests from backends.",
+            ),
+            &["request_kind"],
+        )
+        .unwrap();
+        Self {
+            request_counters,
+            retry_counters,
+        }
+    }
+
+    pub fn collect(&self) -> Vec<metrics::proto::MetricFamily> {
+        let mut metrics = Vec::new();
+        metrics.append(&mut self.request_counters.collect());
+        metrics.append(&mut self.retry_counters.collect());
+        metrics
+    }
+}
+
+pub struct PageserverClient {
+    _tenant_id: String,
+    _timeline_id: String,
+
+    _auth_token: Option<String>,
+
+    shard_map: HashMap<ShardIndex, String>,
+
+    channels: RwLock<HashMap<ShardIndex, Arc<client_cache::ConnectionPool<Channel>>>>,
+
+    auth_interceptor: AuthInterceptor,
+
+    client_cache_options: ClientCacheOptions,
+
+    aggregate_metrics: Option<Arc<PageserverClientAggregateMetrics>>,
+}
+#[derive(Clone)]
+pub struct ClientCacheOptions {
+    pub max_consumers: usize,
+    pub error_threshold: usize,
+    pub connect_timeout: Duration,
+    pub connect_backoff: Duration,
+    pub max_idle_duration: Duration,
+    pub max_total_connections: usize,
+    pub max_delay_ms: u64,
+    pub drop_rate: f64,
+    pub hang_rate: f64,
+}
+
+impl PageserverClient {
+    /// TODO: this doesn't currently react to changes in the shard map.
+    pub fn new(
+        tenant_id: &str,
+        timeline_id: &str,
+        auth_token: &Option<String>,
+        shard_map: HashMap<ShardIndex, String>,
+    ) -> Self {
+        let options = ClientCacheOptions {
+            max_consumers: 5000,
+            error_threshold: 5,
+            connect_timeout: Duration::from_secs(5),
+            connect_backoff: Duration::from_secs(1),
+            max_idle_duration: Duration::from_secs(60),
+            max_total_connections: 100000,
+            max_delay_ms: 0,
+            drop_rate: 0.0,
+            hang_rate: 0.0,
+        };
+        Self::new_with_config(tenant_id, timeline_id, auth_token, shard_map, options, None)
+    }
+    pub fn new_with_config(
+        tenant_id: &str,
+        timeline_id: &str,
+        auth_token: &Option<String>,
+        shard_map: HashMap<ShardIndex, String>,
+        options: ClientCacheOptions,
+        metrics: Option<Arc<PageserverClientAggregateMetrics>>,
+    ) -> Self {
+        Self {
+            _tenant_id: tenant_id.to_string(),
+            _timeline_id: timeline_id.to_string(),
+            _auth_token: auth_token.clone(),
+            shard_map,
+            channels: RwLock::new(HashMap::new()),
+            auth_interceptor: AuthInterceptor::new(tenant_id, timeline_id, auth_token.as_deref()),
+            client_cache_options: options,
+            aggregate_metrics: metrics,
+        }
+    }
+    pub async fn process_check_rel_exists_request(
+        &self,
+        request: CheckRelExistsRequest,
+    ) -> Result<bool, PageserverClientError> {
+        // Current sharding model assumes that all metadata is present only at shard 0.
+        let shard = ShardIndex::unsharded();
+        let pooled_client = self.get_client(shard).await;
+        let chan = pooled_client.channel();
+
+        let mut client =
+            PageServiceClient::with_interceptor(chan, self.auth_interceptor.for_shard(shard));
+
+        let request = proto::CheckRelExistsRequest::from(request);
+        let response = client.check_rel_exists(tonic::Request::new(request)).await;
+
+        match response {
+            Err(status) => {
+                pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                return Err(PageserverClientError::RequestError(status));
+            }
+            Ok(resp) => {
+                pooled_client.finish(Ok(())).await; // Pass success to finish
+                return Ok(resp.get_ref().exists);
+            }
+        }
+    }
+
+    pub async fn process_get_rel_size_request(
+        &self,
+        request: GetRelSizeRequest,
+    ) -> Result<u32, PageserverClientError> {
+        // Current sharding model assumes that all metadata is present only at shard 0.
+        let shard = ShardIndex::unsharded();
+        let pooled_client = self.get_client(shard).await;
+        let chan = pooled_client.channel();
+
+        let mut client =
+            PageServiceClient::with_interceptor(chan, self.auth_interceptor.for_shard(shard));
+
+        let request = proto::GetRelSizeRequest::from(request);
+        let response = client.get_rel_size(tonic::Request::new(request)).await;
+
+        match response {
+            Err(status) => {
+                pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                return Err(PageserverClientError::RequestError(status));
+            }
+            Ok(resp) => {
+                pooled_client.finish(Ok(())).await; // Pass success to finish
+                return Ok(resp.get_ref().num_blocks);
+            }
+        }
+    }
+
+    // Request a single batch of pages
+    //
+    // TODO: This opens a new gRPC stream for every request, which is extremely inefficient
+    pub async fn get_page(
+        &self,
+        request: GetPageRequest,
+    ) -> Result<Vec<Bytes>, PageserverClientError> {
+        // FIXME: calculate the shard number correctly
+        let shard = ShardIndex::unsharded();
+        let pooled_client = self.get_client(shard).await;
+        let chan = pooled_client.channel();
+
+        let mut client =
+            PageServiceClient::with_interceptor(chan, self.auth_interceptor.for_shard(shard));
+
+        let request = proto::GetPageRequest::from(request);
+
+        let request_stream = futures::stream::once(std::future::ready(request));
+
+        let mut response_stream = client
+            .get_pages(tonic::Request::new(request_stream))
+            .await?
+            .into_inner();
+
+        let Some(response) = response_stream.next().await else {
+            return Err(PageserverClientError::Other(
+                "no response received for getpage request".to_string(),
+            ));
+        };
+
+        match self.aggregate_metrics {
+            Some(ref metrics) => {
+                metrics
+                    .request_counters
+                    .with_label_values(&["get_page"])
+                    .inc();
+            }
+            None => {}
+        }
+
+        match response {
+            Err(status) => {
+                pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                return Err(PageserverClientError::RequestError(status));
+            }
+            Ok(resp) => {
+                pooled_client.finish(Ok(())).await; // Pass success to finish
+                let response: GetPageResponse = resp.into();
+                return Ok(response.page_images.to_vec());
+            }
+        }
+    }
+
+    // Open a stream for requesting pages
+    //
+    // TODO: This is a pretty low level interface, the caller should not need to be concerned
+    // with streams. But 'get_page' is currently very naive and inefficient.
+    pub async fn get_pages(
+        &self,
+        requests: impl Stream<Item = proto::GetPageRequest> + Send + 'static,
+    ) -> std::result::Result<
+        tonic::Response<tonic::codec::Streaming<proto::GetPageResponse>>,
+        PageserverClientError,
+    > {
+        // FIXME: calculate the shard number correctly
+        let shard = ShardIndex::unsharded();
+        let pooled_client = self.get_client(shard).await;
+        let chan = pooled_client.channel();
+
+        let mut client =
+            PageServiceClient::with_interceptor(chan, self.auth_interceptor.for_shard(shard));
+
+        let response = client.get_pages(tonic::Request::new(requests)).await;
+
+        match response {
+            Err(status) => {
+                pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                return Err(PageserverClientError::RequestError(status));
+            }
+            Ok(resp) => {
+                return Ok(resp);
+            }
+        }
+    }
+
+    /// Process a request to get the size of a database.
+    pub async fn process_get_dbsize_request(
+        &self,
+        request: GetDbSizeRequest,
+    ) -> Result<u64, PageserverClientError> {
+        // Current sharding model assumes that all metadata is present only at shard 0.
+        let shard = ShardIndex::unsharded();
+        let pooled_client = self.get_client(shard).await;
+        let chan = pooled_client.channel();
+
+        let mut client =
+            PageServiceClient::with_interceptor(chan, self.auth_interceptor.for_shard(shard));
+
+        let request = proto::GetDbSizeRequest::from(request);
+        let response = client.get_db_size(tonic::Request::new(request)).await;
+
+        match response {
+            Err(status) => {
+                pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                return Err(PageserverClientError::RequestError(status));
+            }
+            Ok(resp) => {
+                pooled_client.finish(Ok(())).await; // Pass success to finish
+                return Ok(resp.get_ref().num_bytes);
+            }
+        }
+    }
+    /// Process a request to get the size of a database.
+    pub async fn get_base_backup(
+        &self,
+        request: GetBaseBackupRequest,
+        gzip: bool,
+    ) -> std::result::Result<
+        tonic::Response<tonic::codec::Streaming<proto::GetBaseBackupResponseChunk>>,
+        PageserverClientError,
+    > {
+        // Current sharding model assumes that all metadata is present only at shard 0.
+        let shard = ShardIndex::unsharded();
+        let pooled_client = self.get_client(shard).await;
+        let chan = pooled_client.channel();
+
+        let mut client =
+            PageServiceClient::with_interceptor(chan, self.auth_interceptor.for_shard(shard));
+
+        if gzip {
+            client = client.accept_compressed(tonic::codec::CompressionEncoding::Gzip);
+        }
+
+        let request = proto::GetBaseBackupRequest::from(request);
+        let response = client.get_base_backup(tonic::Request::new(request)).await;
+
+        match response {
+            Err(status) => {
+                pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                return Err(PageserverClientError::RequestError(status));
+            }
+            Ok(resp) => {
+                pooled_client.finish(Ok(())).await; // Pass success to finish
+                return Ok(resp);
+            }
+        }
+    }
+    /// Get a client for given shard
+    ///
+    /// Get a client from the pool for this shard, also creating the pool if it doesn't exist.
+    ///
+    async fn get_client(&self, shard: ShardIndex) -> client_cache::PooledClient<Channel> {
+        let reused_pool: Option<Arc<client_cache::ConnectionPool<Channel>>> = {
+            let channels = self.channels.read().unwrap();
+            channels.get(&shard).cloned()
+        };
+
+        let usable_pool: Arc<client_cache::ConnectionPool<Channel>>;
+        match reused_pool {
+            Some(pool) => {
+                let pooled_client = pool.get_client().await.unwrap();
+                return pooled_client;
+            }
+            None => {
+                // Create a new pool using client_cache_options
+                // declare new_pool
+
+                let new_pool: Arc<client_cache::ConnectionPool<Channel>>;
+                let channel_fact = Arc::new(client_cache::ChannelFactory::new(
+                    self.shard_map.get(&shard).unwrap().clone(),
+                    self.client_cache_options.max_delay_ms,
+                    self.client_cache_options.drop_rate,
+                    self.client_cache_options.hang_rate,
+                ));
+                new_pool = client_cache::ConnectionPool::new(
+                    channel_fact,
+                    self.client_cache_options.connect_timeout,
+                    self.client_cache_options.connect_backoff,
+                    self.client_cache_options.max_consumers,
+                    self.client_cache_options.error_threshold,
+                    self.client_cache_options.max_idle_duration,
+                    self.client_cache_options.max_total_connections,
+                    self.aggregate_metrics.clone(),
+                );
+                let mut write_pool = self.channels.write().unwrap();
+                write_pool.insert(shard, new_pool.clone());
+                usable_pool = new_pool.clone();
+            }
+        }
+
+        let pooled_client = usable_pool.get_client().await.unwrap();
+        return pooled_client;
+    }
+}
+
+/// Inject tenant_id, timeline_id and authentication token to all pageserver requests.
+#[derive(Clone)]
+pub struct AuthInterceptor {
+    tenant_id: AsciiMetadataValue,
+    shard_id: Option<AsciiMetadataValue>,
+    timeline_id: AsciiMetadataValue,
+
+    auth_header: Option<AsciiMetadataValue>, // including "Bearer " prefix
+}
+
+impl AuthInterceptor {
+    pub fn new(tenant_id: &str, timeline_id: &str, auth_token: Option<&str>) -> Self {
+        Self {
+            tenant_id: tenant_id.parse().expect("could not parse tenant id"),
+            shard_id: None,
+            timeline_id: timeline_id.parse().expect("could not parse timeline id"),
+            auth_header: auth_token
+                .map(|t| format!("Bearer {t}"))
+                .map(|t| t.parse().expect("could not parse auth token")),
+        }
+    }
+
+    fn for_shard(&self, shard_id: ShardIndex) -> Self {
+        let mut with_shard = self.clone();
+        with_shard.shard_id = Some(
+            shard_id
+                .to_string()
+                .parse()
+                .expect("could not parse shard id"),
+        );
+        with_shard
+    }
+}
+
+impl tonic::service::Interceptor for AuthInterceptor {
+    fn call(&mut self, mut req: tonic::Request<()>) -> Result<tonic::Request<()>, tonic::Status> {
+        req.metadata_mut()
+            .insert("neon-tenant-id", self.tenant_id.clone());
+        if let Some(shard_id) = &self.shard_id {
+            req.metadata_mut().insert("neon-shard-id", shard_id.clone());
+        }
+        req.metadata_mut()
+            .insert("neon-timeline-id", self.timeline_id.clone());
+        if let Some(auth_header) = &self.auth_header {
+            req.metadata_mut()
+                .insert("authorization", auth_header.clone());
+        }
+
+        Ok(req)
+    }
+}

pageserver/client_grpc/src/request_tracker.rs

@@ -0,0 +1,590 @@
+
+//
+// API Visible to the spawner, just a function call that is async
+//
+use std::sync::Arc;
+use crate::client_cache;
+use pageserver_page_api::GetPageRequest;
+use pageserver_page_api::GetPageResponse;
+use pageserver_page_api::*;
+use pageserver_page_api::proto;
+use crate::client_cache::ConnectionPool;
+use crate::client_cache::ChannelFactory;
+use crate::AuthInterceptor;
+use tonic::{transport::{Channel}, Request};
+use crate::ClientCacheOptions;
+use crate::PageserverClientAggregateMetrics;
+use tokio::sync::Mutex;
+use std::sync::atomic::{AtomicU64, Ordering};
+
+use utils::shard::ShardIndex;
+
+use tokio_stream::wrappers::ReceiverStream;
+use pageserver_page_api::proto::PageServiceClient;
+
+use tonic::{
+    Status,
+    Code,
+};
+
+use async_trait::async_trait;
+use std::time::Duration;
+
+use client_cache::PooledItemFactory;
+//use tracing::info;
+//
+// A mock stream pool that just returns a sending channel, and whenever a GetPageRequest
+// comes in on that channel, it randomly sleeps before sending a GetPageResponse
+//
+
+#[derive(Clone)]
+pub struct StreamReturner {
+    sender: tokio::sync::mpsc::Sender<proto::GetPageRequest>,
+    sender_hashmap: Arc<Mutex<std::collections::HashMap<u64, tokio::sync::mpsc::Sender<Result<proto::GetPageResponse, Status>>>>>,
+}
+pub struct MockStreamFactory {
+}
+
+impl MockStreamFactory {
+    pub fn new() -> Self {
+        MockStreamFactory {
+        }
+    }
+}
+#[async_trait]
+impl PooledItemFactory<StreamReturner> for MockStreamFactory {
+    async fn create(&self, _connect_timeout: Duration) -> Result<Result<StreamReturner, tonic::Status>, tokio::time::error::Elapsed> {
+        let (sender, mut receiver) = tokio::sync::mpsc::channel::<proto::GetPageRequest>(1000);
+        // Create a StreamReturner that will send requests to the receiver channel
+        let stream_returner = StreamReturner {
+            sender: sender.clone(),
+            sender_hashmap: Arc::new(Mutex::new(std::collections::HashMap::new())),
+        };
+
+        let map : Arc<Mutex<std::collections::HashMap<u64, tokio::sync::mpsc::Sender<Result<proto::GetPageResponse, _>>>>>
+            = Arc::clone(&stream_returner.sender_hashmap);
+        tokio::spawn(async move {
+            while let Some(request) = receiver.recv().await {
+
+                // Break out of the loop with 1% chance
+                if rand::random::<f32>() < 0.001 {
+                    break;
+                }
+                // Generate a random number between 0 and 100
+                // Simulate some processing time
+                let mapclone = Arc::clone(&map);
+                tokio::spawn(async move {
+                    let sleep_ms = rand::random::<u64>() % 100;
+                    tokio::time::sleep(tokio::time::Duration::from_millis(sleep_ms)).await;
+                    let response = proto::GetPageResponse {
+                        request_id: request.request_id,
+                        ..Default::default()
+                    };
+                    // look up stream in hash map
+                    let mut hashmap = mapclone.lock().await;
+                    if let Some(sender) = hashmap.get(&request.request_id) {
+                        // Send the response to the original request sender
+                        if let Err(e) = sender.send(Ok(response.clone())).await {
+                            eprintln!("Failed to send response: {}", e);
+                        }
+                        hashmap.remove(&request.request_id);
+                    } else {
+                        eprintln!("No sender found for request ID: {}", request.request_id);
+                    }
+                });
+            }
+            // Close every sender stream in the hashmap
+            let hashmap = map.lock().await;
+            for sender in hashmap.values() {
+                let error = Status::new(Code::Unknown, "Stream closed");
+                if let Err(e) = sender.send(Err(error)).await {
+                    eprintln!("Failed to send close response: {}", e);
+                }
+            }
+        });
+
+        Ok(Ok(stream_returner))
+    }
+}
+
+
+pub struct StreamFactory {
+    connection_pool: Arc<client_cache::ConnectionPool<Channel>>,
+    auth_interceptor: AuthInterceptor,
+    shard: ShardIndex,
+}
+
+impl StreamFactory {
+    pub fn new(
+        connection_pool: Arc<ConnectionPool<Channel>>,
+        auth_interceptor: AuthInterceptor,
+        shard: ShardIndex,
+    ) -> Self {
+        StreamFactory {
+            connection_pool,
+            auth_interceptor,
+            shard,
+        }
+    }
+}
+
+#[async_trait]
+impl PooledItemFactory<StreamReturner> for StreamFactory {
+    async fn create(&self, _connect_timeout: Duration) ->
+    Result<Result<StreamReturner, tonic::Status>, tokio::time::error::Elapsed>
+    {
+        let pool_clone : Arc<ConnectionPool<Channel>> = Arc::clone(&self.connection_pool);
+        let pooled_client = pool_clone.get_client().await;
+        let channel = pooled_client.unwrap().channel();
+        let mut client =
+            PageServiceClient::with_interceptor(channel, self.auth_interceptor.for_shard(self.shard));
+
+        let (sender, receiver) = tokio::sync::mpsc::channel::<proto::GetPageRequest>(1000);
+        let outbound = ReceiverStream::new(receiver);
+
+        let client_resp = client
+            .get_pages(Request::new(outbound))
+            .await;
+
+        match client_resp {
+            Err(status) => {
+                // TODO: Convert this error correctly
+                Ok(Err(tonic::Status::new(
+                    status.code(),
+                    format!("Failed to connect to pageserver: {}", status.message()),
+                )))
+            }
+            Ok(resp) => {
+                let stream_returner = StreamReturner {
+                    sender: sender.clone(),
+                    sender_hashmap: Arc::new(Mutex::new(std::collections::HashMap::new())),
+                };
+                let map : Arc<Mutex<std::collections::HashMap<u64, tokio::sync::mpsc::Sender<Result<proto::GetPageResponse, _>>>>>
+                    = Arc::clone(&stream_returner.sender_hashmap);
+
+                tokio::spawn(async move {
+
+                    let map_clone = Arc::clone(&map);
+                    let mut inner = resp.into_inner();
+                    loop {
+
+                        let resp = inner.message().await;
+                        if !resp.is_ok() {
+                            break; // Exit the loop if no more messages
+                        }
+                        let response = resp.unwrap().unwrap();
+
+                        // look up stream in hash map
+                        let mut hashmap = map_clone.lock().await;
+                        if let Some(sender) = hashmap.get(&response.request_id) {
+                            // Send the response to the original request sender
+                            if let Err(e) = sender.send(Ok(response.clone())).await {
+                                eprintln!("Failed to send response: {}", e);
+                            }
+                            hashmap.remove(&response.request_id);
+                        } else {
+                            eprintln!("No sender found for request ID: {}", response.request_id);
+                        }
+                    }
+                    // Close every sender stream in the hashmap
+                    let hashmap = map_clone.lock().await;
+                    for sender in hashmap.values() {
+                        let error = Status::new(Code::Unknown, "Stream closed");
+                        if let Err(e) = sender.send(Err(error)).await {
+                            eprintln!("Failed to send close response: {}", e);
+                        }
+                    }
+                });
+
+                Ok(Ok(stream_returner))
+            }
+        }
+    }
+}
+
+#[derive(Clone)]
+pub struct RequestTracker {
+    cur_id: Arc<AtomicU64>,
+    stream_pool: Arc<ConnectionPool<StreamReturner>>,
+    unary_pool: Arc<ConnectionPool<Channel>>,
+    auth_interceptor: AuthInterceptor,
+    shard: ShardIndex,
+}
+
+impl RequestTracker {
+    pub fn new(stream_pool: Arc<ConnectionPool<StreamReturner>>,
+                unary_pool: Arc<ConnectionPool<Channel>>,
+                auth_interceptor: AuthInterceptor,
+                shard: ShardIndex,
+    ) -> Self {
+        let cur_id = Arc::new(AtomicU64::new(0));
+
+        RequestTracker {
+            cur_id: cur_id.clone(),
+            stream_pool: stream_pool,
+            unary_pool: unary_pool,
+            auth_interceptor: auth_interceptor,
+            shard: shard.clone()
+        }
+    }
+
+    pub async fn send_process_check_rel_exists_request(
+        &self,
+        req: CheckRelExistsRequest,
+    ) -> Result<bool, tonic::Status> {
+        loop {
+            let unary_pool = Arc::clone(&self.unary_pool);
+            let pooled_client = unary_pool.get_client().await.unwrap();
+            let channel = pooled_client.channel();
+            let mut ps_client = PageServiceClient::with_interceptor(channel, self.auth_interceptor.for_shard(self.shard));
+            let request = proto::CheckRelExistsRequest::from(req.clone());
+            let response = ps_client.check_rel_exists(tonic::Request::new(request)).await;
+
+            match response {
+                Err(status) => {
+                    pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                    continue;
+                }
+                Ok(resp) => {
+                    pooled_client.finish(Ok(())).await; // Pass success to finish
+                    return Ok(resp.get_ref().exists);
+                }
+            }
+        }
+    }
+
+    pub async fn send_process_get_rel_size_request(
+        &self,
+        req: GetRelSizeRequest,
+    ) -> Result<u32, tonic::Status> {
+        loop {
+            // Current sharding model assumes that all metadata is present only at shard 0.
+            let unary_pool = Arc::clone(&self.unary_pool);
+            let pooled_client = unary_pool.get_client().await.unwrap();
+            let channel = pooled_client.channel();
+            let mut ps_client = PageServiceClient::with_interceptor(channel, self.auth_interceptor.for_shard(self.shard));
+
+            let request = proto::GetRelSizeRequest::from(req.clone());
+            let response = ps_client.get_rel_size(tonic::Request::new(request)).await;
+
+            match response {
+                Err(status) => {
+                    pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                    continue;
+                }
+                Ok(resp) => {
+                    pooled_client.finish(Ok(())).await; // Pass success to finish
+                    return Ok(resp.get_ref().num_blocks);
+                }
+            }
+
+        }
+    }
+
+    pub async fn send_process_get_dbsize_request(
+        &self,
+        req: GetDbSizeRequest,
+    ) -> Result<u64, tonic::Status> {
+        loop {
+            // Current sharding model assumes that all metadata is present only at shard 0.
+            let unary_pool = Arc::clone(&self.unary_pool);
+            let pooled_client = unary_pool.get_client().await.unwrap();let channel = pooled_client.channel();
+            let mut ps_client = PageServiceClient::with_interceptor(channel, self.auth_interceptor.for_shard(self.shard));
+
+            let request = proto::GetDbSizeRequest::from(req.clone());
+            let response = ps_client.get_db_size(tonic::Request::new(request)).await;
+
+            match response {
+                Err(status) => {
+                    pooled_client.finish(Err(status.clone())).await; // Pass error to finish
+                    continue;
+                }
+                Ok(resp) => {
+                    pooled_client.finish(Ok(())).await; // Pass success to finish
+                    return Ok(resp.get_ref().num_bytes);
+                }
+            }
+
+        }
+    }
+
+    pub async fn send_getpage_request(
+        &mut self,
+        req: GetPageRequest,
+    ) -> Result<GetPageResponse, tonic::Status> {
+        loop {
+            let mut request = req.clone();
+            // Increment cur_id
+            //let request_id = self.cur_id.fetch_add(1, Ordering::SeqCst) + 1;
+            let request_id = request.request_id;
+            let response_sender: tokio::sync::mpsc::Sender<Result<proto::GetPageResponse, Status>>;
+            let mut response_receiver: tokio::sync::mpsc::Receiver<Result<proto::GetPageResponse, Status>>;
+
+            (response_sender, response_receiver) = tokio::sync::mpsc::channel(1);
+            //request.request_id = request_id;
+
+            // Get a stream from the stream pool
+            let pool_clone = Arc::clone(&self.stream_pool);
+            let sender_stream_pool = pool_clone.get_client().await;
+            let stream_returner = match sender_stream_pool {
+                Ok(stream_ret) => stream_ret,
+                Err(_e) => {
+                    // retry
+                    continue;
+                }
+            };
+            let returner = stream_returner.channel();
+            let map = returner.sender_hashmap.clone();
+            // Insert the response sender into the hashmap
+            {
+                let mut map_inner = map.lock().await;
+                map_inner.insert(request_id, response_sender);
+            }
+            let sent = returner.sender.send(proto::GetPageRequest::from(request))
+                .await;
+
+            if let Err(_e) = sent {
+                // Remove the request from the map if sending failed
+                {
+                    let mut map_inner = map.lock().await;
+                    // remove from hashmap
+                    map_inner.remove(&request_id);
+                }
+                stream_returner.finish(Err(Status::new(Code::Unknown,
+                                                       "Failed to send request"))).await;
+                continue;
+            }
+
+            let response: Option<Result<proto::GetPageResponse, Status>>;
+            response = response_receiver.recv().await;
+            match response {
+                Some (resp) => {
+                    match resp {
+                        Err(_status) => {
+                            // Handle the case where the response was not received
+                            stream_returner.finish(Err(Status::new(Code::Unknown,
+                                                                   "Failed to receive response"))).await;
+                            continue;
+                        },
+                        Ok(resp) => {
+                            stream_returner.finish(Result::Ok(())).await;
+                            return Ok(resp.clone().into());
+                        }
+                    }
+                }
+                None => {
+                    // Handle the case where the response channel was closed
+                    stream_returner.finish(Err(Status::new(Code::Unknown,
+                                                           "Response channel closed"))).await;
+                    continue;
+                }
+            }
+        }
+    }
+}
+
+struct ShardedRequestTrackerInner {
+    // Hashmap of shard index to RequestTracker
+    trackers: std::collections::HashMap<ShardIndex, RequestTracker>,
+}
+pub struct ShardedRequestTracker {
+    inner: Arc<Mutex<ShardedRequestTrackerInner>>,
+    tcp_client_cache_options: ClientCacheOptions,
+    stream_client_cache_options: ClientCacheOptions,
+}
+
+//
+// TODO: Functions in the ShardedRequestTracker should be able to timeout and
+// cancel a reqeust. The request should return an error if it is cancelled.
+//
+impl ShardedRequestTracker {
+    pub fn new() -> Self {
+        //
+        // Default configuration for the client. These could be added to a config file
+        //
+        let tcp_client_cache_options = ClientCacheOptions {
+            max_delay_ms:       0,
+            drop_rate:          0.0,
+            hang_rate:          0.0,
+            connect_timeout:    Duration::from_secs(1),
+            connect_backoff:    Duration::from_millis(100),
+            max_consumers:      8, // Streams per connection
+            error_threshold:    10,
+            max_idle_duration:  Duration::from_secs(5),
+            max_total_connections: 8,
+        };
+        let stream_client_cache_options = ClientCacheOptions {
+            max_delay_ms:       0,
+            drop_rate:          0.0,
+            hang_rate:          0.0,
+            connect_timeout:    Duration::from_secs(1),
+            connect_backoff:    Duration::from_millis(100),
+            max_consumers:      64, // Requests per stream
+            error_threshold:    10,
+            max_idle_duration:  Duration::from_secs(5),
+            max_total_connections: 64, // Total allowable number of streams
+        };
+        ShardedRequestTracker {
+            inner: Arc::new(Mutex::new(ShardedRequestTrackerInner {
+                trackers: std::collections::HashMap::new(),
+            })),
+            tcp_client_cache_options,
+            stream_client_cache_options,
+        }
+    }
+
+    pub async fn update_shard_map(&self,
+                            shard_urls: std::collections::HashMap<ShardIndex, String>,
+                            metrics: Option<Arc<PageserverClientAggregateMetrics>>,
+                            tenant_id: String, timeline_id: String, auth_str: Option<&str>) {
+
+
+       let mut trackers = std::collections::HashMap::new();
+        for (shard, endpoint_url) in shard_urls {
+            //
+            // Create a pool of streams for streaming get_page requests
+            //
+            let channel_fact : Arc<dyn PooledItemFactory<Channel> + Send + Sync> = Arc::new(ChannelFactory::new(
+                endpoint_url.clone(),
+                self.tcp_client_cache_options.max_delay_ms,
+                self.tcp_client_cache_options.drop_rate,
+                self.tcp_client_cache_options.hang_rate,
+            ));
+            let new_pool: Arc<ConnectionPool<Channel>>;
+            new_pool = ConnectionPool::new(
+                Arc::clone(&channel_fact),
+                self.tcp_client_cache_options.connect_timeout,
+                self.tcp_client_cache_options.connect_backoff,
+                self.tcp_client_cache_options.max_consumers,
+                self.tcp_client_cache_options.error_threshold,
+                self.tcp_client_cache_options.max_idle_duration,
+                self.tcp_client_cache_options.max_total_connections,
+                metrics.clone(),
+            );
+
+            let auth_interceptor = AuthInterceptor::new(tenant_id.as_str(),
+                                                        timeline_id.as_str(),
+                                                        auth_str);
+
+            let stream_pool = ConnectionPool::<StreamReturner>::new(
+                Arc::new(StreamFactory::new(new_pool.clone(),
+                                            auth_interceptor.clone(), ShardIndex::unsharded())),
+                self.stream_client_cache_options.connect_timeout,
+                self.stream_client_cache_options.connect_backoff,
+                self.stream_client_cache_options.max_consumers,
+                self.stream_client_cache_options.error_threshold,
+                self.stream_client_cache_options.max_idle_duration,
+                self.stream_client_cache_options.max_total_connections,
+                metrics.clone(),
+            );
+
+            //
+            // Create a client pool for unary requests
+            //
+
+            let unary_pool: Arc<ConnectionPool<Channel>>;
+            unary_pool = ConnectionPool::new(
+                Arc::clone(&channel_fact),
+                self.tcp_client_cache_options.connect_timeout,
+                self.tcp_client_cache_options.connect_backoff,
+                self.tcp_client_cache_options.max_consumers,
+                self.tcp_client_cache_options.error_threshold,
+                self.tcp_client_cache_options.max_idle_duration,
+                self.tcp_client_cache_options.max_total_connections,
+                metrics.clone()
+            );
+            //
+            // Create a new RequestTracker for this shard
+            //
+            let new_tracker = RequestTracker::new(stream_pool, unary_pool, auth_interceptor, shard);
+            trackers.insert(shard, new_tracker);
+        }
+        let mut inner = self.inner.lock().await;
+        inner.trackers = trackers;
+    }
+
+    pub async fn get_page(
+        &self,
+        req: GetPageRequest,
+    ) -> Result<GetPageResponse, tonic::Status> {
+
+        // Get shard index from the request
+        let shard_index = ShardIndex::unsharded();
+        let inner = self.inner.lock().await;
+        let mut tracker : RequestTracker;
+        if let Some(t) = inner.trackers.get(&shard_index) {
+            tracker = t.clone();
+        } else {
+            return Err(tonic::Status::not_found(format!("Shard {} not found", shard_index)));
+        }
+        drop(inner);
+        // Call the send_getpage_request method on the tracker
+        let response = tracker.send_getpage_request(req).await;
+        match response {
+            Ok(resp) => Ok(resp),
+            Err(e) => Err(tonic::Status::unknown(format!("Failed to get page: {}", e))),
+        }
+    }
+    pub async fn process_get_dbsize_request(
+        &self,
+        request: GetDbSizeRequest,
+    ) -> Result<u64, tonic::Status> {
+        let shard_index = ShardIndex::unsharded();
+        let inner = self.inner.lock().await;
+        let mut tracker: RequestTracker;
+        if let Some(t) = inner.trackers.get(&shard_index) {
+            tracker = t.clone();
+        } else {
+            return Err(tonic::Status::not_found(format!("Shard {} not found", shard_index)));
+        }
+        drop(inner); // Release the lock before calling send_process_get_dbsize_request
+        // Call the send_process_get_dbsize_request method on the tracker
+        let response = tracker.send_process_get_dbsize_request(request).await;
+        match response {
+            Ok(resp) => Ok(resp),
+            Err(e) => Err(e),
+        }
+    }
+
+    pub async fn process_get_rel_size_request(
+        &self,
+        request: GetRelSizeRequest,
+    ) -> Result<u32, tonic::Status> {
+        let shard_index = ShardIndex::unsharded();
+        let inner = self.inner.lock().await;
+        let mut tracker: RequestTracker;
+        if let Some(t) = inner.trackers.get(&shard_index) {
+            tracker = t.clone();
+        } else {
+            return Err(tonic::Status::not_found(format!("Shard {} not found", shard_index)));
+        }
+        drop(inner); // Release the lock before calling send_process_get_rel_size_request
+        // Call the send_process_get_rel_size_request method on the tracker
+        let response = tracker.send_process_get_rel_size_request(request).await;
+        match response {
+            Ok(resp) => Ok(resp),
+            Err(e) => Err(e),
+        }
+    }
+
+    pub async fn process_check_rel_exists_request(
+        &self,
+        request: CheckRelExistsRequest,
+    ) -> Result<bool, tonic::Status> {
+        let shard_index = ShardIndex::unsharded();
+        let inner = self.inner.lock().await;
+        let mut tracker: RequestTracker;
+        if let Some(t) = inner.trackers.get(&shard_index) {
+            tracker = t.clone();
+        } else {
+            return Err(tonic::Status::not_found(format!("Shard {} not found", shard_index)));
+        }
+        drop(inner); // Release the lock before calling send_process_check_rel_exists_request
+        // Call the send_process_check_rel_exists_request method on the tracker
+        let response = tracker.send_process_check_rel_exists_request(request).await;
+        match response {
+            Ok(resp) => Ok(resp),
+            Err(e) => Err(e),
+        }
+    }
+}

pageserver/ctl/src/main.rs

@@ -176,11 +176,9 @@ async fn main() -> anyhow::Result<()> {
             let config = RemoteStorageConfig::from_toml_str(&cmd.config_toml_str)?;
             let storage = remote_storage::GenericRemoteStorage::from_config(&config).await;
             let cancel = CancellationToken::new();
-            // Complexity limit: as we are running this command locally, we should have a lot of memory available, and we do not
-            // need to limit the number of versions we are going to delete.
             storage
                 .unwrap()
-                .time_travel_recover(Some(&prefix), timestamp, done_if_after, &cancel, None)
+                .time_travel_recover(Some(&prefix), timestamp, done_if_after, &cancel)
                 .await?;
         }
         Commands::Key(dkc) => dkc.execute(),

pageserver/page_api/proto/page_service.proto

@@ -104,8 +104,8 @@ message CheckRelExistsResponse {
 
 // Requests a base backup at a given LSN.
 message GetBaseBackupRequest {
-  // The LSN to fetch a base backup at.
-  ReadLsn read_lsn = 1;
+  // The LSN to fetch a base backup at. 0 or absent means the latest LSN known to the Pageserver.
+  uint64 lsn = 1;
   // If true, logical replication slots will not be created.
   bool replica = 2;
 }

pageserver/page_api/src/model.rs

@@ -185,30 +185,25 @@ impl From<CheckRelExistsResponse> for proto::CheckRelExistsResponse {
 /// Requests a base backup at a given LSN.
 #[derive(Clone, Copy, Debug)]
 pub struct GetBaseBackupRequest {
-    /// The LSN to fetch a base backup at.
-    pub read_lsn: ReadLsn,
+    /// The LSN to fetch a base backup at. If None, uses the latest LSN known to the Pageserver.
+    pub lsn: Option<Lsn>,
     /// If true, logical replication slots will not be created.
     pub replica: bool,
 }
 
-impl TryFrom<proto::GetBaseBackupRequest> for GetBaseBackupRequest {
-    type Error = ProtocolError;
-
-    fn try_from(pb: proto::GetBaseBackupRequest) -> Result<Self, Self::Error> {
-        Ok(Self {
-            read_lsn: pb
-                .read_lsn
-                .ok_or(ProtocolError::Missing("read_lsn"))?
-                .try_into()?,
+impl From<proto::GetBaseBackupRequest> for GetBaseBackupRequest {
+    fn from(pb: proto::GetBaseBackupRequest) -> Self {
+        Self {
+            lsn: (pb.lsn != 0).then_some(Lsn(pb.lsn)),
             replica: pb.replica,
-        })
+        }
     }
 }
 
 impl From<GetBaseBackupRequest> for proto::GetBaseBackupRequest {
     fn from(request: GetBaseBackupRequest) -> Self {
         Self {
-            read_lsn: Some(request.read_lsn.into()),
+            lsn: request.lsn.unwrap_or_default().0,
             replica: request.replica,
         }
     }
@@ -487,6 +482,7 @@ impl From<GetPageStatusCode> for i32 {
 
 // Fetches the size of a relation at a given LSN, as # of blocks. Only valid on shard 0, other
 // shards will error.
+#[derive(Clone)]
 pub struct GetRelSizeRequest {
     pub read_lsn: ReadLsn,
     pub rel: RelTag,

pageserver/pagebench/Cargo.toml

@@ -24,9 +24,13 @@ tracing.workspace = true
 tokio.workspace = true
 tokio-stream.workspace = true
 tokio-util.workspace = true
+axum.workspace = true
+http.workspace = true
+metrics.workspace = true
 tonic.workspace = true
 
 pageserver_client.workspace = true
+pageserver_client_grpc.workspace = true
 pageserver_api.workspace = true
 pageserver_page_api.workspace = true
 utils = { path = "../../libs/utils/" }

pageserver/pagebench/src/cmd/basebackup.rs

@@ -9,12 +9,15 @@ use anyhow::Context;
 use pageserver_api::shard::TenantShardId;
 use pageserver_client::mgmt_api::ForceAwaitLogicalSize;
 use pageserver_client::page_service::BasebackupRequest;
+use pageserver_page_api::{GetBaseBackupRequest, ReadLsn};
+
 use rand::prelude::*;
 use tokio::sync::Barrier;
 use tokio::task::JoinSet;
 use tracing::{info, instrument};
 use utils::id::TenantTimelineId;
 use utils::lsn::Lsn;
+use utils::shard::ShardIndex;
 
 use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
 use crate::util::{request_stats, tokio_thread_local_stats};
@@ -22,6 +25,8 @@ use crate::util::{request_stats, tokio_thread_local_stats};
 /// basebackup@LatestLSN
 #[derive(clap::Parser)]
 pub(crate) struct Args {
+    #[clap(long, default_value = "false")]
+    grpc: bool,
     #[clap(long, default_value = "http://localhost:9898")]
     mgmt_api_endpoint: String,
     #[clap(long, default_value = "postgres://postgres@localhost:64000")]
@@ -52,7 +57,7 @@ impl LiveStats {
 
 struct Target {
     timeline: TenantTimelineId,
-    lsn_range: Option<Range<Lsn>>,
+    lsn_range: Range<Lsn>,
 }
 
 #[derive(serde::Serialize)]
@@ -105,7 +110,7 @@ async fn main_impl(
                 anyhow::Ok(Target {
                     timeline,
                     // TODO: support lsn_range != latest LSN
-                    lsn_range: Some(info.last_record_lsn..(info.last_record_lsn + 1)),
+                    lsn_range: info.last_record_lsn..(info.last_record_lsn + 1),
                 })
             }
         });
@@ -149,14 +154,27 @@ async fn main_impl(
     for tl in &timelines {
         let (sender, receiver) = tokio::sync::mpsc::channel(1); // TODO: not sure what the implications of this are
         work_senders.insert(tl, sender);
-        tasks.push(tokio::spawn(client(
-            args,
-            *tl,
-            Arc::clone(&start_work_barrier),
-            receiver,
-            Arc::clone(&all_work_done_barrier),
-            Arc::clone(&live_stats),
-        )));
+
+        let client_task = if args.grpc {
+            tokio::spawn(client_grpc(
+                args,
+                *tl,
+                Arc::clone(&start_work_barrier),
+                receiver,
+                Arc::clone(&all_work_done_barrier),
+                Arc::clone(&live_stats),
+            ))
+        } else {
+            tokio::spawn(client(
+                args,
+                *tl,
+                Arc::clone(&start_work_barrier),
+                receiver,
+                Arc::clone(&all_work_done_barrier),
+                Arc::clone(&live_stats),
+            ))
+        };
+        tasks.push(client_task);
     }
 
     let work_sender = async move {
@@ -165,7 +183,7 @@ async fn main_impl(
             let (timeline, work) = {
                 let mut rng = rand::thread_rng();
                 let target = all_targets.choose(&mut rng).unwrap();
-                let lsn = target.lsn_range.clone().map(|r| rng.gen_range(r));
+                let lsn = rng.gen_range(target.lsn_range.clone());
                 (
                     target.timeline,
                     Work {
@@ -215,7 +233,7 @@ async fn main_impl(
 
 #[derive(Copy, Clone)]
 struct Work {
-    lsn: Option<Lsn>,
+    lsn: Lsn,
     gzip: bool,
 }
 
@@ -240,7 +258,7 @@ async fn client(
             .basebackup(&BasebackupRequest {
                 tenant_id: timeline.tenant_id,
                 timeline_id: timeline.timeline_id,
-                lsn,
+                lsn: Some(lsn),
                 gzip,
             })
             .await
@@ -270,3 +288,71 @@ async fn client(
 
     all_work_done_barrier.wait().await;
 }
+
+#[instrument(skip_all)]
+async fn client_grpc(
+    args: &'static Args,
+    timeline: TenantTimelineId,
+    start_work_barrier: Arc<Barrier>,
+    mut work: tokio::sync::mpsc::Receiver<Work>,
+    all_work_done_barrier: Arc<Barrier>,
+    live_stats: Arc<LiveStats>,
+) {
+    let shard_map = HashMap::from([(
+        ShardIndex::unsharded(),
+        args.page_service_connstring.clone(),
+    )]);
+    let client = pageserver_client_grpc::PageserverClient::new(
+        &timeline.tenant_id.to_string(),
+        &timeline.timeline_id.to_string(),
+        &None,
+        shard_map,
+    );
+
+    start_work_barrier.wait().await;
+
+    while let Some(Work { lsn, gzip }) = work.recv().await {
+        let start = Instant::now();
+
+        //tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+
+        info!("starting get_base_backup");
+        let mut basebackup_stream = client
+            .get_base_backup(
+                GetBaseBackupRequest {
+                    lsn: Some(lsn),
+                    replica: false,
+                },
+                gzip,
+            )
+            .await
+            .with_context(|| format!("start basebackup for {timeline}"))
+            .unwrap()
+            .into_inner();
+
+        info!("starting receive");
+        use futures::StreamExt;
+        let mut size = 0;
+        let mut nchunks = 0;
+        while let Some(chunk) = basebackup_stream.next().await {
+            let chunk = chunk
+                .with_context(|| format!("error during basebackup"))
+                .unwrap();
+            size += chunk.chunk.len();
+            nchunks += 1;
+        }
+
+        info!(
+            "basebackup size is {} bytes, avg chunk size {} bytes",
+            size,
+            size as f32 / nchunks as f32
+        );
+        let elapsed = start.elapsed();
+        live_stats.inc();
+        STATS.with(|stats| {
+            stats.borrow().lock().unwrap().observe(elapsed).unwrap();
+        });
+    }
+
+    all_work_done_barrier.wait().await;
+}

pageserver/pagebench/src/cmd/getpage_latest_lsn.rs

@@ -1,10 +1,11 @@
-use std::collections::{HashMap, HashSet, VecDeque};
+use std::collections::{HashSet, HashMap, VecDeque};
 use std::future::Future;
 use std::num::NonZeroUsize;
 use std::pin::Pin;
 use std::sync::atomic::{AtomicU64, Ordering};
 use std::sync::{Arc, Mutex};
 use std::time::{Duration, Instant};
+use std::io::Error;
 
 use anyhow::Context;
 use async_trait::async_trait;
@@ -23,6 +24,20 @@ use tracing::info;
 use utils::id::TenantTimelineId;
 use utils::lsn::Lsn;
 
+use tonic::transport::Channel;
+
+use axum::Router;
+use axum::body::Body;
+use axum::extract::State;
+use axum::response::Response;
+
+use http::StatusCode;
+use http::header::CONTENT_TYPE;
+
+use metrics;
+use metrics::proto::MetricFamily;
+use metrics::{Encoder, TextEncoder};
+
 use crate::util::tokio_thread_local_stats::AllThreadLocalStats;
 use crate::util::{request_stats, tokio_thread_local_stats};
 
@@ -35,6 +50,10 @@ enum Protocol {
 /// GetPage@LatestLSN, uniformly distributed across the compute-accessible keyspace.
 #[derive(clap::Parser)]
 pub(crate) struct Args {
+    #[clap(long, default_value = "false")]
+    grpc: bool,
+    #[clap(long, default_value = "false")]
+    grpc_stream: bool,
     #[clap(long, default_value = "http://localhost:9898")]
     mgmt_api_endpoint: String,
     #[clap(long, default_value = "postgres://postgres@localhost:64000")]
@@ -73,6 +92,9 @@ pub(crate) struct Args {
     #[clap(long)]
     set_io_mode: Option<pageserver_api::models::virtual_file::IoMode>,
 
+    #[clap(long)]
+    only_relnode: Option<u32>,
+
     /// Queue depth generated in each client.
     #[clap(long, default_value = "1")]
     queue_depth: NonZeroUsize,
@@ -87,10 +109,31 @@ pub(crate) struct Args {
     #[clap(long, default_value = "1")]
     batch_size: NonZeroUsize,
 
-    #[clap(long)]
-    only_relnode: Option<u32>,
-
     targets: Option<Vec<TenantTimelineId>>,
+
+    #[clap(long, default_value = "100")]
+    pool_max_consumers: NonZeroUsize,
+
+    #[clap(long, default_value = "5")]
+    pool_error_threshold: NonZeroUsize,
+
+    #[clap(long, default_value = "5000")]
+    pool_connect_timeout: NonZeroUsize,
+
+    #[clap(long, default_value = "1000")]
+    pool_connect_backoff: NonZeroUsize,
+
+    #[clap(long, default_value = "60000")]
+    pool_max_idle_duration: NonZeroUsize,
+
+    #[clap(long, default_value = "0")]
+    max_delay_ms: usize,
+
+    #[clap(long, default_value = "0")]
+    percent_drops: usize,
+
+    #[clap(long, default_value = "0")]
+    percent_hangs: usize,
 }
 
 /// State shared by all clients
@@ -147,6 +190,37 @@ pub(crate) fn main(args: Args) -> anyhow::Result<()> {
         main_impl(args, thread_local_stats)
     })
 }
+async fn get_metrics(
+    State(state): State<Arc<pageserver_client_grpc::PageserverClientAggregateMetrics>>,
+) -> Response {
+    let metrics = state.collect();
+
+    info!("metrics: {metrics:?}");
+    // When we call TextEncoder::encode() below, it will immediately return an
+    // error if a metric family has no metrics, so we need to preemptively
+    // filter out metric families with no metrics.
+    let metrics = metrics
+        .into_iter()
+        .filter(|m| !m.get_metric().is_empty())
+        .collect::<Vec<MetricFamily>>();
+
+    let encoder = TextEncoder::new();
+    let mut buffer = vec![];
+
+    if let Err(e) = encoder.encode(&metrics, &mut buffer) {
+        Response::builder()
+            .status(StatusCode::INTERNAL_SERVER_ERROR)
+            .header(CONTENT_TYPE, "application/text")
+            .body(Body::from(e.to_string()))
+            .unwrap()
+    } else {
+        Response::builder()
+            .status(StatusCode::OK)
+            .header(CONTENT_TYPE, encoder.format_type())
+            .body(Body::from(buffer))
+            .unwrap()
+    }
+}
 
 async fn main_impl(
     args: Args,
@@ -154,6 +228,24 @@ async fn main_impl(
 ) -> anyhow::Result<()> {
     let args: &'static Args = Box::leak(Box::new(args));
 
+    // Vector of pageserver clients
+    let client_metrics = Arc::new(pageserver_client_grpc::PageserverClientAggregateMetrics::new());
+
+    use axum::routing::get;
+    let app = Router::new()
+        .route("/metrics", get(get_metrics))
+        .with_state(client_metrics.clone());
+
+    // TODO: make configurable. Or listen on unix domain socket?
+    let listener = tokio::net::TcpListener::bind("127.0.0.1:9090")
+        .await
+        .unwrap();
+
+    tokio::spawn(async {
+        tracing::info!("metrics listener spawned");
+        axum::serve(listener, app).await.unwrap()
+    });
+
     let mgmt_api_client = Arc::new(pageserver_client::mgmt_api::Client::new(
         reqwest::Client::new(), // TODO: support ssl_ca_file for https APIs in pagebench.
         args.mgmt_api_endpoint.clone(),
@@ -312,6 +404,7 @@ async fn main_impl(
     let rps_period = args
         .per_client_rate
         .map(|rps_limit| Duration::from_secs_f64(1.0 / (rps_limit as f64)));
+
     let make_worker: &dyn Fn(WorkerId) -> Pin<Box<dyn Send + Future<Output = ()>>> = &|worker_id| {
         let ss = shared_state.clone();
         let cancel = cancel.clone();
@@ -337,6 +430,7 @@ async fn main_impl(
                         .await
                         .unwrap(),
                 ),
+
             };
             run_worker(args, client, ss, cancel, rps_period, ranges, weights).await
         })
@@ -604,6 +698,7 @@ impl Client for LibpqClient {
 struct GrpcClient {
     req_tx: tokio::sync::mpsc::Sender<proto::GetPageRequest>,
     resp_rx: tonic::Streaming<proto::GetPageResponse>,
+    start_times: Vec<Instant>,
 }
 
 impl GrpcClient {
@@ -627,6 +722,7 @@ impl GrpcClient {
         Ok(Self {
             req_tx,
             resp_rx: resp_stream,
+            start_times: Vec::new(),
         })
     }
 }
@@ -651,6 +747,7 @@ impl Client for GrpcClient {
             rel: Some(rel.into()),
             block_number: blks,
         };
+        self.start_times.push(Instant::now());
         self.req_tx.send(req).await?;
         Ok(())
     }
@@ -665,3 +762,4 @@ impl Client for GrpcClient {
         Ok((resp.request_id, resp.page_image))
     }
 }
+

pageserver/src/basebackup_cache.rs

@@ -1,6 +1,5 @@
 use std::{collections::HashMap, sync::Arc};
 
-use anyhow::Context;
 use async_compression::tokio::write::GzipEncoder;
 use camino::{Utf8Path, Utf8PathBuf};
 use metrics::core::{AtomicU64, GenericCounter};
@@ -168,17 +167,14 @@ impl BasebackupCache {
             .join(Self::entry_filename(tenant_id, timeline_id, lsn))
     }
 
-    fn tmp_dir(&self) -> Utf8PathBuf {
-        self.data_dir.join("tmp")
-    }
-
     fn entry_tmp_path(
         &self,
         tenant_id: TenantId,
         timeline_id: TimelineId,
         lsn: Lsn,
     ) -> Utf8PathBuf {
-        self.tmp_dir()
+        self.data_dir
+            .join("tmp")
             .join(Self::entry_filename(tenant_id, timeline_id, lsn))
     }
 
@@ -198,18 +194,15 @@ impl BasebackupCache {
         Some((tenant_id, timeline_id, lsn))
     }
 
-    // Recreate the tmp directory to clear all files in it.
-    async fn clean_tmp_dir(&self) -> anyhow::Result<()> {
-        let tmp_dir = self.tmp_dir();
-        if tmp_dir.exists() {
-            tokio::fs::remove_dir_all(&tmp_dir).await?;
-        }
-        tokio::fs::create_dir_all(&tmp_dir).await?;
-        Ok(())
-    }
-
     async fn cleanup(&self) -> anyhow::Result<()> {
-        self.clean_tmp_dir().await?;
+        // Cleanup tmp directory.
+        let tmp_dir = self.data_dir.join("tmp");
+        let mut tmp_dir = tokio::fs::read_dir(&tmp_dir).await?;
+        while let Some(dir_entry) = tmp_dir.next_entry().await? {
+            if let Err(e) = tokio::fs::remove_file(dir_entry.path()).await {
+                tracing::warn!("Failed to remove basebackup cache tmp file: {:#}", e);
+            }
+        }
 
         // Remove outdated entries.
         let entries_old = self.entries.lock().unwrap().clone();
@@ -248,14 +241,16 @@ impl BasebackupCache {
     }
 
     async fn on_startup(&self) -> anyhow::Result<()> {
-        // Create data_dir if it does not exist.
-        tokio::fs::create_dir_all(&self.data_dir)
+        // Create data_dir and tmp directory if they do not exist.
+        tokio::fs::create_dir_all(&self.data_dir.join("tmp"))
             .await
-            .context("Failed to create basebackup cache data directory")?;
-
-        self.clean_tmp_dir()
-            .await
-            .context("Failed to clean tmp directory")?;
+            .map_err(|e| {
+                anyhow::anyhow!(
+                    "Failed to create basebackup cache data_dir {:?}: {:?}",
+                    self.data_dir,
+                    e
+                )
+            })?;
 
         // Read existing entries from the data_dir and add them to in-memory state.
         let mut entries = HashMap::new();
@@ -413,19 +408,6 @@ impl BasebackupCache {
             .tenant_manager
             .get_attached_tenant_shard(tenant_shard_id)?;
 
-        let feature_flag = tenant
-            .feature_resolver
-            .evaluate_boolean("enable-basebackup-cache", tenant_shard_id.tenant_id);
-
-        if feature_flag.is_err() {
-            tracing::info!(
-                tenant_id = %tenant_shard_id.tenant_id,
-                "Basebackup cache is disabled for tenant by feature flag, skipping basebackup",
-            );
-            self.prepare_skip_count.inc();
-            return Ok(());
-        }
-
         let tenant_state = tenant.current_state();
         if tenant_state != TenantState::Active {
             anyhow::bail!(
@@ -469,11 +451,6 @@ impl BasebackupCache {
         }
 
         // Move the tmp file to the final location atomically.
-        // The tmp file is fsynced, so it's guaranteed that we will not have a partial file
-        // in the main directory.
-        // It's not necessary to fsync the inode after renaming, because the worst case is that
-        // the rename operation will be rolled back on the disk failure, the entry will disappear
-        // from the main directory, and the entry access will cause a cache miss.
         let entry_path = self.entry_path(tenant_shard_id.tenant_id, timeline_id, req_lsn);
         tokio::fs::rename(&entry_tmp_path, &entry_path).await?;
 
@@ -491,17 +468,16 @@ impl BasebackupCache {
     }
 
     /// Prepares a basebackup in a temporary file.
-    /// Guarantees that the tmp file is fsynced before returning.
     async fn prepare_basebackup_tmp(
         &self,
-        entry_tmp_path: &Utf8Path,
+        emptry_tmp_path: &Utf8Path,
         timeline: &Arc<Timeline>,
         req_lsn: Lsn,
     ) -> anyhow::Result<()> {
         let ctx = RequestContext::new(TaskKind::BasebackupCache, DownloadBehavior::Download);
         let ctx = ctx.with_scope_timeline(timeline);
 
-        let file = tokio::fs::File::create(entry_tmp_path).await?;
+        let file = tokio::fs::File::create(emptry_tmp_path).await?;
         let mut writer = BufWriter::new(file);
 
         let mut encoder = GzipEncoder::with_quality(

pageserver/src/bin/pageserver.rs

@@ -23,7 +23,6 @@ use pageserver::deletion_queue::DeletionQueue;
 use pageserver::disk_usage_eviction_task::{self, launch_disk_usage_global_eviction_task};
 use pageserver::feature_resolver::FeatureResolver;
 use pageserver::metrics::{STARTUP_DURATION, STARTUP_IS_LOADING};
-use pageserver::page_service::GrpcPageServiceHandler;
 use pageserver::task_mgr::{
     BACKGROUND_RUNTIME, COMPUTE_REQUEST_RUNTIME, MGMT_REQUEST_RUNTIME, WALRECEIVER_RUNTIME,
 };
@@ -573,8 +572,7 @@ fn start_pageserver(
         tokio::sync::mpsc::unbounded_channel();
     let deletion_queue_client = deletion_queue.new_client();
     let background_purges = mgr::BackgroundPurges::default();
-
-    let tenant_manager = mgr::init(
+    let tenant_manager = BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(
         conf,
         background_purges.clone(),
         TenantSharedResources {
@@ -585,10 +583,10 @@ fn start_pageserver(
             basebackup_prepare_sender,
             feature_resolver,
         },
+        order,
         shutdown_pageserver.clone(),
-    );
+    ))?;
     let tenant_manager = Arc::new(tenant_manager);
-    BACKGROUND_RUNTIME.block_on(mgr::init_tenant_mgr(tenant_manager.clone(), order))?;
 
     let basebackup_cache = BasebackupCache::spawn(
         BACKGROUND_RUNTIME.handle(),
@@ -816,7 +814,7 @@ fn start_pageserver(
     // necessary?
     let mut page_service_grpc = None;
     if let Some(grpc_listener) = grpc_listener {
-        page_service_grpc = Some(GrpcPageServiceHandler::spawn(
+        page_service_grpc = Some(page_service::spawn_grpc(
             tenant_manager.clone(),
             grpc_auth,
             otel_guard.as_ref().map(|g| g.dispatch.clone()),

pageserver/src/controller_upcall_client.rs

@@ -195,6 +195,8 @@ impl StorageControllerUpcallApi for StorageControllerUpcallClient {
                         node_id: conf.id,
                         listen_pg_addr: m.postgres_host,
                         listen_pg_port: m.postgres_port,
+                        listen_grpc_addr: m.grpc_host,
+                        listen_grpc_port: m.grpc_port,
                         listen_http_addr: m.http_host,
                         listen_http_port: m.http_port,
                         listen_https_port: m.https_port,

pageserver/src/feature_resolver.rs

@@ -1,6 +1,5 @@
 use std::{collections::HashMap, sync::Arc, time::Duration};
 
-use pageserver_api::config::NodeMetadata;
 use posthog_client_lite::{
     CaptureEvent, FeatureResolverBackgroundLoop, PostHogClientConfig, PostHogEvaluationError,
     PostHogFlagFilterPropertyValue,
@@ -87,35 +86,7 @@ impl FeatureResolver {
                         }
                     }
                 }
-                // TODO: move this to a background task so that we don't block startup in case of slow disk
-                let metadata_path = conf.metadata_path();
-                match std::fs::read_to_string(&metadata_path) {
-                    Ok(metadata_str) => match serde_json::from_str::<NodeMetadata>(&metadata_str) {
-                        Ok(metadata) => {
-                            properties.insert(
-                                "hostname".to_string(),
-                                PostHogFlagFilterPropertyValue::String(metadata.http_host),
-                            );
-                            if let Some(cplane_region) = metadata.other.get("region_id") {
-                                if let Some(cplane_region) = cplane_region.as_str() {
-                                    // This region contains the cell number
-                                    properties.insert(
-                                        "neon_region".to_string(),
-                                        PostHogFlagFilterPropertyValue::String(
-                                            cplane_region.to_string(),
-                                        ),
-                                    );
-                                }
-                            }
-                        }
-                        Err(e) => {
-                            tracing::warn!("Failed to parse metadata.json: {}", e);
-                        }
-                    },
-                    Err(e) => {
-                        tracing::warn!("Failed to read metadata.json: {}", e);
-                    }
-                }
+                // TODO: add pageserver URL.
                 Arc::new(properties)
             };
             let fake_tenants = {

pageserver/src/http/routes.rs

@@ -73,7 +73,6 @@ use crate::tenant::remote_timeline_client::{
 use crate::tenant::secondary::SecondaryController;
 use crate::tenant::size::ModelInputs;
 use crate::tenant::storage_layer::{IoConcurrency, LayerAccessStatsReset, LayerName};
-use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 use crate::tenant::timeline::offload::{OffloadError, offload_timeline};
 use crate::tenant::timeline::{
     CompactFlags, CompactOptions, CompactRequest, CompactionError, MarkInvisibleRequest, Timeline,
@@ -1452,10 +1451,7 @@ async fn timeline_layer_scan_disposable_keys(
     let ctx = RequestContext::new(TaskKind::MgmtRequest, DownloadBehavior::Download)
         .with_scope_timeline(&timeline);
 
-    let guard = timeline
-        .layers
-        .read(LayerManagerLockHolder::GetLayerMapInfo)
-        .await;
+    let guard = timeline.layers.read().await;
     let Some(layer) = guard.try_get_from_key(&layer_name.clone().into()) else {
         return Err(ApiError::NotFound(
             anyhow::anyhow!("Layer {tenant_shard_id}/{timeline_id}/{layer_name} not found").into(),

pageserver/src/metrics.rs

@@ -1053,15 +1053,6 @@ pub(crate) static TENANT_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
     .expect("Failed to register pageserver_tenant_states_count metric")
 });
 
-pub(crate) static TIMELINE_STATE_METRIC: Lazy<UIntGaugeVec> = Lazy::new(|| {
-    register_uint_gauge_vec!(
-        "pageserver_timeline_states_count",
-        "Count of timelines per state",
-        &["state"]
-    )
-    .expect("Failed to register pageserver_timeline_states_count metric")
-});
-
 /// A set of broken tenants.
 ///
 /// These are expected to be so rare that a set is fine. Set as in a new timeseries per each broken
@@ -3334,8 +3325,6 @@ impl TimelineMetrics {
                 &timeline_id,
             );
 
-        TIMELINE_STATE_METRIC.with_label_values(&["active"]).inc();
-
         TimelineMetrics {
             tenant_id,
             shard_id,
@@ -3490,8 +3479,6 @@ impl TimelineMetrics {
             return;
         }
 
-        TIMELINE_STATE_METRIC.with_label_values(&["active"]).dec();
-
         let tenant_id = &self.tenant_id;
         let timeline_id = &self.timeline_id;
         let shard_id = &self.shard_id;

pageserver/src/page_service.rs

@@ -14,7 +14,7 @@ use std::{io, str};
 
 use anyhow::{Context as _, anyhow, bail};
 use async_compression::tokio::write::GzipEncoder;
-use bytes::{Buf, BytesMut};
+use bytes::{Buf, BufMut as _, BytesMut};
 use futures::future::BoxFuture;
 use futures::{FutureExt, Stream};
 use itertools::Itertools;
@@ -169,6 +169,99 @@ pub fn spawn(
     Listener { cancel, task }
 }
 
+/// Spawns a gRPC server for the page service.
+///
+/// TODO: move this onto GrpcPageServiceHandler::spawn().
+/// TODO: this doesn't support TLS. We need TLS reloading via ReloadingCertificateResolver, so we
+/// need to reimplement the TCP+TLS accept loop ourselves.
+pub fn spawn_grpc(
+    tenant_manager: Arc<TenantManager>,
+    auth: Option<Arc<SwappableJwtAuth>>,
+    perf_trace_dispatch: Option<Dispatch>,
+    get_vectored_concurrent_io: GetVectoredConcurrentIo,
+    listener: std::net::TcpListener,
+) -> anyhow::Result<CancellableTask> {
+    let cancel = CancellationToken::new();
+    let ctx = RequestContextBuilder::new(TaskKind::PageRequestHandler)
+        .download_behavior(DownloadBehavior::Download)
+        .perf_span_dispatch(perf_trace_dispatch)
+        .detached_child();
+    let gate = Gate::default();
+
+    // Set up the TCP socket. We take a preconfigured TcpListener to bind the
+    // port early during startup.
+    let incoming = {
+        let _runtime = COMPUTE_REQUEST_RUNTIME.enter(); // required by TcpListener::from_std
+        listener.set_nonblocking(true)?;
+        tonic::transport::server::TcpIncoming::from(tokio::net::TcpListener::from_std(listener)?)
+            .with_nodelay(Some(GRPC_TCP_NODELAY))
+            .with_keepalive(Some(GRPC_TCP_KEEPALIVE_TIME))
+    };
+
+    // Set up the gRPC server.
+    //
+    // TODO: consider tuning window sizes.
+    let mut server = tonic::transport::Server::builder()
+        .http2_keepalive_interval(Some(GRPC_HTTP2_KEEPALIVE_INTERVAL))
+        .http2_keepalive_timeout(Some(GRPC_HTTP2_KEEPALIVE_TIMEOUT))
+        .max_concurrent_streams(Some(GRPC_MAX_CONCURRENT_STREAMS));
+
+    // Main page service stack. Uses a mix of Tonic interceptors and Tower layers:
+    //
+    // * Interceptors: can inspect and modify the gRPC request. Sync code only, runs before service.
+    //
+    // * Layers: allow async code, can run code after the service response. However, only has access
+    //   to the raw HTTP request/response, not the gRPC types.
+    let page_service_handler = GrpcPageServiceHandler {
+        tenant_manager,
+        ctx,
+        gate_guard: gate.enter().expect("gate was just created"),
+        get_vectored_concurrent_io,
+    };
+
+    let observability_layer = ObservabilityLayer;
+    let mut tenant_interceptor = TenantMetadataInterceptor;
+    let mut auth_interceptor = TenantAuthInterceptor::new(auth);
+
+    let page_service = tower::ServiceBuilder::new()
+        // Create tracing span and record request start time.
+        .layer(observability_layer)
+        // Intercept gRPC requests.
+        .layer(tonic::service::InterceptorLayer::new(move |mut req| {
+            // Extract tenant metadata.
+            req = tenant_interceptor.call(req)?;
+            // Authenticate tenant JWT token.
+            req = auth_interceptor.call(req)?;
+            Ok(req)
+        }))
+        .service(proto::PageServiceServer::new(page_service_handler));
+    let server = server.add_service(page_service);
+
+    // Reflection service for use with e.g. grpcurl.
+    let reflection_service = tonic_reflection::server::Builder::configure()
+        .register_encoded_file_descriptor_set(proto::FILE_DESCRIPTOR_SET)
+        .build_v1()?;
+    let server = server.add_service(reflection_service);
+
+    // Spawn server task.
+    let task_cancel = cancel.clone();
+    let task = COMPUTE_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
+        "grpc listener",
+        async move {
+            let result = server
+                .serve_with_incoming_shutdown(incoming, task_cancel.cancelled())
+                .await;
+            if result.is_ok() {
+                // TODO: revisit shutdown logic once page service is implemented.
+                gate.close().await;
+            }
+            result
+        },
+    ));
+
+    Ok(CancellableTask { task, cancel })
+}
+
 impl Listener {
     pub async fn stop_accepting(self) -> Connections {
         self.cancel.cancel();
@@ -3273,101 +3366,6 @@ pub struct GrpcPageServiceHandler {
 }
 
 impl GrpcPageServiceHandler {
-    /// Spawns a gRPC server for the page service.
-    ///
-    /// TODO: this doesn't support TLS. We need TLS reloading via ReloadingCertificateResolver, so we
-    /// need to reimplement the TCP+TLS accept loop ourselves.
-    pub fn spawn(
-        tenant_manager: Arc<TenantManager>,
-        auth: Option<Arc<SwappableJwtAuth>>,
-        perf_trace_dispatch: Option<Dispatch>,
-        get_vectored_concurrent_io: GetVectoredConcurrentIo,
-        listener: std::net::TcpListener,
-    ) -> anyhow::Result<CancellableTask> {
-        let cancel = CancellationToken::new();
-        let ctx = RequestContextBuilder::new(TaskKind::PageRequestHandler)
-            .download_behavior(DownloadBehavior::Download)
-            .perf_span_dispatch(perf_trace_dispatch)
-            .detached_child();
-        let gate = Gate::default();
-
-        // Set up the TCP socket. We take a preconfigured TcpListener to bind the
-        // port early during startup.
-        let incoming = {
-            let _runtime = COMPUTE_REQUEST_RUNTIME.enter(); // required by TcpListener::from_std
-            listener.set_nonblocking(true)?;
-            tonic::transport::server::TcpIncoming::from(tokio::net::TcpListener::from_std(
-                listener,
-            )?)
-            .with_nodelay(Some(GRPC_TCP_NODELAY))
-            .with_keepalive(Some(GRPC_TCP_KEEPALIVE_TIME))
-        };
-
-        // Set up the gRPC server.
-        //
-        // TODO: consider tuning window sizes.
-        let mut server = tonic::transport::Server::builder()
-            .http2_keepalive_interval(Some(GRPC_HTTP2_KEEPALIVE_INTERVAL))
-            .http2_keepalive_timeout(Some(GRPC_HTTP2_KEEPALIVE_TIMEOUT))
-            .max_concurrent_streams(Some(GRPC_MAX_CONCURRENT_STREAMS));
-
-        // Main page service stack. Uses a mix of Tonic interceptors and Tower layers:
-        //
-        // * Interceptors: can inspect and modify the gRPC request. Sync code only, runs before service.
-        //
-        // * Layers: allow async code, can run code after the service response. However, only has access
-        //   to the raw HTTP request/response, not the gRPC types.
-        let page_service_handler = GrpcPageServiceHandler {
-            tenant_manager,
-            ctx,
-            gate_guard: gate.enter().expect("gate was just created"),
-            get_vectored_concurrent_io,
-        };
-
-        let observability_layer = ObservabilityLayer;
-        let mut tenant_interceptor = TenantMetadataInterceptor;
-        let mut auth_interceptor = TenantAuthInterceptor::new(auth);
-
-        let page_service = tower::ServiceBuilder::new()
-            // Create tracing span and record request start time.
-            .layer(observability_layer)
-            // Intercept gRPC requests.
-            .layer(tonic::service::InterceptorLayer::new(move |mut req| {
-                // Extract tenant metadata.
-                req = tenant_interceptor.call(req)?;
-                // Authenticate tenant JWT token.
-                req = auth_interceptor.call(req)?;
-                Ok(req)
-            }))
-            // Run the page service.
-            .service(proto::PageServiceServer::new(page_service_handler));
-        let server = server.add_service(page_service);
-
-        // Reflection service for use with e.g. grpcurl.
-        let reflection_service = tonic_reflection::server::Builder::configure()
-            .register_encoded_file_descriptor_set(proto::FILE_DESCRIPTOR_SET)
-            .build_v1()?;
-        let server = server.add_service(reflection_service);
-
-        // Spawn server task.
-        let task_cancel = cancel.clone();
-        let task = COMPUTE_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
-            "grpc listener",
-            async move {
-                let result = server
-                    .serve_with_incoming_shutdown(incoming, task_cancel.cancelled())
-                    .await;
-                if result.is_ok() {
-                    // TODO: revisit shutdown logic once page service is implemented.
-                    gate.close().await;
-                }
-                result
-            },
-        ));
-
-        Ok(CancellableTask { task, cancel })
-    }
-
     /// Errors if the request is executed on a non-zero shard. Only shard 0 has a complete view of
     /// relations and their sizes, as well as SLRU segments and similar data.
     #[allow(clippy::result_large_err)]
@@ -3608,35 +3606,38 @@ impl proto::PageService for GrpcPageServiceHandler {
         if timeline.is_archived() == Some(true) {
             return Err(tonic::Status::failed_precondition("timeline is archived"));
         }
-        let req: page_api::GetBaseBackupRequest = req.into_inner().try_into()?;
+        let req: page_api::GetBaseBackupRequest = req.into_inner().into();
 
-        span_record!(lsn=%req.read_lsn);
+        span_record!(lsn=?req.lsn);
 
-        let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
-        timeline
-            .wait_lsn(
-                req.read_lsn.request_lsn,
-                WaitLsnWaiter::PageService,
-                WaitLsnTimeout::Default,
-                &ctx,
-            )
-            .await?;
-        timeline
-            .check_lsn_is_in_scope(req.read_lsn.request_lsn, &latest_gc_cutoff_lsn)
-            .map_err(|err| {
-                tonic::Status::invalid_argument(format!("invalid basebackup LSN: {err}"))
-            })?;
+        if let Some(lsn) = req.lsn {
+            let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
+            timeline
+                .wait_lsn(
+                    lsn,
+                    WaitLsnWaiter::PageService,
+                    WaitLsnTimeout::Default,
+                    &ctx,
+                )
+                .await?;
+            timeline
+                .check_lsn_is_in_scope(lsn, &latest_gc_cutoff_lsn)
+                .map_err(|err| {
+                    tonic::Status::invalid_argument(format!("invalid basebackup LSN: {err}"))
+                })?;
+        }
 
         // Spawn a task to run the basebackup.
         //
-        // TODO: do we need to support full base backups, for debugging?
+        // TODO: do we need to support full base backups, for debugging? This also requires passing
+        // the prev_lsn parameter.
         let span = Span::current();
         let (mut simplex_read, mut simplex_write) = tokio::io::simplex(CHUNK_SIZE);
         let jh = tokio::spawn(async move {
             let result = basebackup::send_basebackup_tarball(
                 &mut simplex_write,
                 &timeline,
-                Some(req.read_lsn.request_lsn),
+                req.lsn,
                 None,
                 false,
                 req.replica,
@@ -3652,20 +3653,21 @@ impl proto::PageService for GrpcPageServiceHandler {
 
         // Emit chunks of size CHUNK_SIZE.
         let chunks = async_stream::try_stream! {
-            let mut chunk = BytesMut::with_capacity(CHUNK_SIZE);
             loop {
-                let n = simplex_read.read_buf(&mut chunk).await.map_err(|err| {
-                    tonic::Status::internal(format!("failed to read basebackup chunk: {err}"))
-                })?;
-
-                // If we read 0 bytes, either the chunk is full or the stream is closed.
-                if n == 0 {
-                    if chunk.is_empty() {
-                        break;
+                let mut chunk = BytesMut::with_capacity(CHUNK_SIZE).limit(CHUNK_SIZE);
+                loop {
+                    let n = simplex_read.read_buf(&mut chunk).await.map_err(|err| {
+                        tonic::Status::internal(format!("failed to read basebackup chunk: {err}"))
+                    })?;
+                    if n == 0 {
+                        break; // full chunk or closed stream
                     }
-                    yield proto::GetBaseBackupResponseChunk::from(chunk.clone().freeze());
-                    chunk.clear();
                 }
+                let chunk = chunk.into_inner().freeze();
+                if chunk.is_empty() {
+                    break;
+                }
+                yield proto::GetBaseBackupResponseChunk::from(chunk);
             }
             // Wait for the basebackup task to exit and check for errors.
             jh.await.map_err(|err| {

pageserver/src/tenant.rs

@@ -51,7 +51,6 @@ use secondary::heatmap::{HeatMapTenant, HeatMapTimeline};
 use storage_broker::BrokerClientChannel;
 use timeline::compaction::{CompactionOutcome, GcCompactionQueue};
 use timeline::import_pgdata::ImportingTimeline;
-use timeline::layer_manager::LayerManagerLockHolder;
 use timeline::offload::{OffloadError, offload_timeline};
 use timeline::{
     CompactFlags, CompactOptions, CompactionError, PreviousHeatmap, ShutdownMode, import_pgdata,
@@ -90,8 +89,7 @@ use crate::l0_flush::L0FlushGlobalState;
 use crate::metrics::{
     BROKEN_TENANTS_SET, CIRCUIT_BREAKERS_BROKEN, CIRCUIT_BREAKERS_UNBROKEN, CONCURRENT_INITDBS,
     INITDB_RUN_TIME, INITDB_SEMAPHORE_ACQUISITION_TIME, TENANT, TENANT_OFFLOADED_TIMELINES,
-    TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC, TIMELINE_STATE_METRIC,
-    remove_tenant_metrics,
+    TENANT_STATE_METRIC, TENANT_SYNTHETIC_SIZE_METRIC, remove_tenant_metrics,
 };
 use crate::task_mgr::TaskKind;
 use crate::tenant::config::LocationMode;
@@ -546,28 +544,6 @@ pub struct OffloadedTimeline {
 
     /// Part of the `OffloadedTimeline` object's lifecycle: this needs to be set before we drop it
     pub deleted_from_ancestor: AtomicBool,
-
-    _metrics_guard: OffloadedTimelineMetricsGuard,
-}
-
-/// Increases the offloaded timeline count metric when created, and decreases when dropped.
-struct OffloadedTimelineMetricsGuard;
-
-impl OffloadedTimelineMetricsGuard {
-    fn new() -> Self {
-        TIMELINE_STATE_METRIC
-            .with_label_values(&["offloaded"])
-            .inc();
-        Self
-    }
-}
-
-impl Drop for OffloadedTimelineMetricsGuard {
-    fn drop(&mut self) {
-        TIMELINE_STATE_METRIC
-            .with_label_values(&["offloaded"])
-            .dec();
-    }
 }
 
 impl OffloadedTimeline {
@@ -600,8 +576,6 @@ impl OffloadedTimeline {
 
             delete_progress: timeline.delete_progress.clone(),
             deleted_from_ancestor: AtomicBool::new(false),
-
-            _metrics_guard: OffloadedTimelineMetricsGuard::new(),
         })
     }
     fn from_manifest(tenant_shard_id: TenantShardId, manifest: &OffloadedTimelineManifest) -> Self {
@@ -621,7 +595,6 @@ impl OffloadedTimeline {
             archived_at,
             delete_progress: TimelineDeleteProgress::default(),
             deleted_from_ancestor: AtomicBool::new(false),
-            _metrics_guard: OffloadedTimelineMetricsGuard::new(),
         }
     }
     fn manifest(&self) -> OffloadedTimelineManifest {
@@ -1316,7 +1289,7 @@ impl TenantShard {
                         ancestor.is_some()
                             || timeline
                                 .layers
-                                .read(LayerManagerLockHolder::LoadLayerMap)
+                                .read()
                                 .await
                                 .layer_map()
                                 .expect(
@@ -2644,7 +2617,7 @@ impl TenantShard {
         }
         let layer_names = tline
             .layers
-            .read(LayerManagerLockHolder::Testing)
+            .read()
             .await
             .layer_map()
             .unwrap()
@@ -3159,12 +3132,7 @@ impl TenantShard {
 
         for timeline in &compact {
             // Collect L0 counts. Can't await while holding lock above.
-            if let Ok(lm) = timeline
-                .layers
-                .read(LayerManagerLockHolder::Compaction)
-                .await
-                .layer_map()
-            {
+            if let Ok(lm) = timeline.layers.read().await.layer_map() {
                 l0_counts.insert(timeline.timeline_id, lm.level0_deltas().len());
             }
         }
@@ -4906,7 +4874,7 @@ impl TenantShard {
         }
         let layer_names = tline
             .layers
-            .read(LayerManagerLockHolder::Testing)
+            .read()
             .await
             .layer_map()
             .unwrap()
@@ -6976,7 +6944,7 @@ mod tests {
             .await?;
         make_some_layers(tline.as_ref(), Lsn(0x20), &ctx).await?;
 
-        let layer_map = tline.layers.read(LayerManagerLockHolder::Testing).await;
+        let layer_map = tline.layers.read().await;
         let level0_deltas = layer_map
             .layer_map()?
             .level0_deltas()
@@ -7212,7 +7180,7 @@ mod tests {
         let lsn = Lsn(0x10);
         let inserted = bulk_insert_compact_gc(&tenant, &tline, &ctx, lsn, 50, 10000).await?;
 
-        let guard = tline.layers.read(LayerManagerLockHolder::Testing).await;
+        let guard = tline.layers.read().await;
         let lm = guard.layer_map()?;
 
         lm.dump(true, &ctx).await?;
@@ -8240,23 +8208,12 @@ mod tests {
             tline.freeze_and_flush().await?; // force create a delta layer
         }
 
-        let before_num_l0_delta_files = tline
-            .layers
-            .read(LayerManagerLockHolder::Testing)
-            .await
-            .layer_map()?
-            .level0_deltas()
-            .len();
+        let before_num_l0_delta_files =
+            tline.layers.read().await.layer_map()?.level0_deltas().len();
 
         tline.compact(&cancel, EnumSet::default(), &ctx).await?;
 
-        let after_num_l0_delta_files = tline
-            .layers
-            .read(LayerManagerLockHolder::Testing)
-            .await
-            .layer_map()?
-            .level0_deltas()
-            .len();
+        let after_num_l0_delta_files = tline.layers.read().await.layer_map()?.level0_deltas().len();
 
         assert!(
             after_num_l0_delta_files < before_num_l0_delta_files,

pageserver/src/tenant/config.rs

@@ -61,8 +61,8 @@ pub(crate) struct LocationConf {
     /// The detailed shard identity.  This structure is already scoped within
     /// a TenantShardId, but we need the full ShardIdentity to enable calculating
     /// key->shard mappings.
-    // TODO(vlad): Remove this default once all configs have a shard identity on disk.
     #[serde(default = "ShardIdentity::unsharded")]
+    #[serde(skip_serializing_if = "ShardIdentity::is_unsharded")]
     pub(crate) shard: ShardIdentity,
 
     /// The pan-cluster tenant configuration, the same on all locations
@@ -149,12 +149,7 @@ impl LocationConf {
     /// For use when attaching/re-attaching: update the generation stored in this
     /// structure.  If we were in a secondary state, promote to attached (posession
     /// of a fresh generation implies this).
-    pub(crate) fn attach_in_generation(
-        &mut self,
-        mode: AttachmentMode,
-        generation: Generation,
-        stripe_size: ShardStripeSize,
-    ) {
+    pub(crate) fn attach_in_generation(&mut self, mode: AttachmentMode, generation: Generation) {
         match &mut self.mode {
             LocationMode::Attached(attach_conf) => {
                 attach_conf.generation = generation;
@@ -168,8 +163,6 @@ impl LocationConf {
                 })
             }
         }
-
-        self.shard.stripe_size = stripe_size;
     }
 
     pub(crate) fn try_from(conf: &'_ models::LocationConfig) -> anyhow::Result<Self> {

pageserver/src/tenant/remote_timeline_client/upload.rs

@@ -1,7 +1,6 @@
 //! Helper functions to upload files to remote storage with a RemoteStorage
 
 use std::io::{ErrorKind, SeekFrom};
-use std::num::NonZeroU32;
 use std::time::SystemTime;
 
 use anyhow::{Context, bail};
@@ -229,25 +228,11 @@ pub(crate) async fn time_travel_recover_tenant(
         let timelines_path = super::remote_timelines_path(tenant_shard_id);
         prefixes.push(timelines_path);
     }
-
-    // Limit the number of versions deletions, mostly so that we don't
-    // keep requesting forever if the list is too long, as we'd put the
-    // list in RAM.
-    // Building a list of 100k entries that reaches the limit roughly takes
-    // 40 seconds, and roughly corresponds to tenants of 2 TiB physical size.
-    const COMPLEXITY_LIMIT: Option<NonZeroU32> = NonZeroU32::new(100_000);
-
     for prefix in &prefixes {
         backoff::retry(
             || async {
                 storage
-                    .time_travel_recover(
-                        Some(prefix),
-                        timestamp,
-                        done_if_after,
-                        cancel,
-                        COMPLEXITY_LIMIT,
-                    )
+                    .time_travel_recover(Some(prefix), timestamp, done_if_after, cancel)
                     .await
             },
             |e| !matches!(e, TimeTravelError::Other(_)),

pageserver/src/tenant/storage_layer/delta_layer.rs

@@ -1635,7 +1635,6 @@ pub(crate) mod test {
     use crate::tenant::disk_btree::tests::TestDisk;
     use crate::tenant::harness::{TIMELINE_ID, TenantHarness};
     use crate::tenant::storage_layer::{Layer, ResidentLayer};
-    use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
     use crate::tenant::{TenantShard, Timeline};
 
     /// Construct an index for a fictional delta layer and and then
@@ -2003,7 +2002,7 @@ pub(crate) mod test {
 
         let initdb_layer = timeline
             .layers
-            .read(crate::tenant::timeline::layer_manager::LayerManagerLockHolder::Testing)
+            .read()
             .await
             .likely_resident_layers()
             .next()
@@ -2079,7 +2078,7 @@ pub(crate) mod test {
 
         let new_layer = timeline
             .layers
-            .read(LayerManagerLockHolder::Testing)
+            .read()
             .await
             .likely_resident_layers()
             .find(|&x| x != &initdb_layer)

pageserver/src/tenant/storage_layer/layer/tests.rs

@@ -10,7 +10,6 @@ use super::*;
 use crate::context::DownloadBehavior;
 use crate::tenant::harness::{TenantHarness, test_img};
 use crate::tenant::storage_layer::{IoConcurrency, LayerVisibilityHint};
-use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 
 /// Used in tests to advance a future to wanted await point, and not futher.
 const ADVANCE: std::time::Duration = std::time::Duration::from_secs(3600);
@@ -60,7 +59,7 @@ async fn smoke_test() {
     // there to avoid the timeline being illegally empty
     let (layer, dummy_layer) = {
         let mut layers = {
-            let layers = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+            let layers = timeline.layers.read().await;
             layers.likely_resident_layers().cloned().collect::<Vec<_>>()
         };
 
@@ -216,7 +215,7 @@ async fn smoke_test() {
 
     // Simulate GC removing our test layer.
     {
-        let mut g = timeline.layers.write(LayerManagerLockHolder::Testing).await;
+        let mut g = timeline.layers.write().await;
 
         let layers = &[layer];
         g.open_mut().unwrap().finish_gc_timeline(layers);
@@ -262,7 +261,7 @@ async fn evict_and_wait_on_wanted_deleted() {
 
     let layer = {
         let mut layers = {
-            let layers = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+            let layers = timeline.layers.read().await;
             layers.likely_resident_layers().cloned().collect::<Vec<_>>()
         };
 
@@ -306,7 +305,7 @@ async fn evict_and_wait_on_wanted_deleted() {
     // assert that once we remove the `layer` from the layer map and drop our reference,
     // the deletion of the layer in remote_storage happens.
     {
-        let mut layers = timeline.layers.write(LayerManagerLockHolder::Testing).await;
+        let mut layers = timeline.layers.write().await;
         layers.open_mut().unwrap().finish_gc_timeline(&[layer]);
     }
 
@@ -348,7 +347,7 @@ fn read_wins_pending_eviction() {
 
         let layer = {
             let mut layers = {
-                let layers = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+                let layers = timeline.layers.read().await;
                 layers.likely_resident_layers().cloned().collect::<Vec<_>>()
             };
 
@@ -481,7 +480,7 @@ fn multiple_pending_evictions_scenario(name: &'static str, in_order: bool) {
 
         let layer = {
             let mut layers = {
-                let layers = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+                let layers = timeline.layers.read().await;
                 layers.likely_resident_layers().cloned().collect::<Vec<_>>()
             };
 
@@ -656,7 +655,7 @@ async fn cancelled_get_or_maybe_download_does_not_cancel_eviction() {
 
     let layer = {
         let mut layers = {
-            let layers = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+            let layers = timeline.layers.read().await;
             layers.likely_resident_layers().cloned().collect::<Vec<_>>()
         };
 
@@ -742,7 +741,7 @@ async fn evict_and_wait_does_not_wait_for_download() {
 
     let layer = {
         let mut layers = {
-            let layers = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+            let layers = timeline.layers.read().await;
             layers.likely_resident_layers().cloned().collect::<Vec<_>>()
         };
 
@@ -863,7 +862,7 @@ async fn eviction_cancellation_on_drop() {
 
     let (evicted_layer, not_evicted) = {
         let mut layers = {
-            let mut guard = timeline.layers.write(LayerManagerLockHolder::Testing).await;
+            let mut guard = timeline.layers.write().await;
             let layers = guard.likely_resident_layers().cloned().collect::<Vec<_>>();
             // remove the layers from layermap
             guard.open_mut().unwrap().finish_gc_timeline(&layers);

pageserver/src/tenant/timeline.rs

@@ -35,11 +35,7 @@ use fail::fail_point;
 use futures::stream::FuturesUnordered;
 use futures::{FutureExt, StreamExt};
 use handle::ShardTimelineId;
-use layer_manager::{
-    LayerManagerLockHolder, LayerManagerReadGuard, LayerManagerWriteGuard, LockedLayerManager,
-    Shutdown,
-};
-
+use layer_manager::Shutdown;
 use offload::OffloadError;
 use once_cell::sync::Lazy;
 use pageserver_api::config::tenant_conf_defaults::DEFAULT_PITR_INTERVAL;
@@ -86,6 +82,7 @@ use wal_decoder::serialized_batch::{SerializedValueBatch, ValueMeta};
 use self::delete::DeleteTimelineFlow;
 pub(super) use self::eviction_task::EvictionTaskTenantState;
 use self::eviction_task::EvictionTaskTimelineState;
+use self::layer_manager::LayerManager;
 use self::logical_size::LogicalSize;
 use self::walreceiver::{WalReceiver, WalReceiverConf};
 use super::remote_timeline_client::RemoteTimelineClient;
@@ -184,13 +181,13 @@ impl std::fmt::Display for ImageLayerCreationMode {
 
 /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
 /// Can be removed after all refactors are done.
-fn drop_layer_manager_rlock(rlock: LayerManagerReadGuard<'_>) {
+fn drop_rlock<T>(rlock: tokio::sync::RwLockReadGuard<T>) {
     drop(rlock)
 }
 
 /// Temporary function for immutable storage state refactor, ensures we are dropping mutex guard instead of other things.
 /// Can be removed after all refactors are done.
-fn drop_layer_manager_wlock(rlock: LayerManagerWriteGuard<'_>) {
+fn drop_wlock<T>(rlock: tokio::sync::RwLockWriteGuard<'_, T>) {
     drop(rlock)
 }
 
@@ -244,7 +241,7 @@ pub struct Timeline {
     ///
     /// In the future, we'll be able to split up the tuple of LayerMap and `LayerFileManager`,
     /// so that e.g. on-demand-download/eviction, and layer spreading, can operate just on `LayerFileManager`.
-    pub(crate) layers: LockedLayerManager,
+    pub(crate) layers: tokio::sync::RwLock<LayerManager>,
 
     last_freeze_at: AtomicLsn,
     // Atomic would be more appropriate here.
@@ -1058,8 +1055,8 @@ pub(crate) enum WaitLsnWaiter<'a> {
 /// Argument to [`Timeline::shutdown`].
 #[derive(Debug, Clone, Copy)]
 pub(crate) enum ShutdownMode {
-    /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk. This method can
-    /// take multiple seconds for a busy timeline.
+    /// Graceful shutdown, may do a lot of I/O as we flush any open layers to disk and then
+    /// also to remote storage.  This method can easily take multiple seconds for a busy timeline.
     ///
     /// While we are flushing, we continue to accept read I/O for LSNs ingested before
     /// the call to [`Timeline::shutdown`].
@@ -1538,10 +1535,7 @@ impl Timeline {
     /// This method makes no distinction between local and remote layers.
     /// Hence, the result **does not represent local filesystem usage**.
     pub(crate) async fn layer_size_sum(&self) -> u64 {
-        let guard = self
-            .layers
-            .read(LayerManagerLockHolder::GetLayerMapInfo)
-            .await;
+        let guard = self.layers.read().await;
         guard.layer_size_sum()
     }
 
@@ -1851,7 +1845,7 @@ impl Timeline {
         // time, and this was missed.
         // if write_guard.is_none() { return; }
 
-        let Ok(layers_guard) = self.layers.try_read(LayerManagerLockHolder::TryFreezeLayer) else {
+        let Ok(layers_guard) = self.layers.try_read() else {
             // Don't block if the layer lock is busy
             return;
         };
@@ -2164,7 +2158,7 @@ impl Timeline {
         if let ShutdownMode::FreezeAndFlush = mode {
             let do_flush = if let Some((open, frozen)) = self
                 .layers
-                .read(LayerManagerLockHolder::Shutdown)
+                .read()
                 .await
                 .layer_map()
                 .map(|lm| (lm.open_layer.is_some(), lm.frozen_layers.len()))
@@ -2268,10 +2262,7 @@ impl Timeline {
             // Allow any remaining in-memory layers to do cleanup -- until that, they hold the gate
             // open.
             let mut write_guard = self.write_lock.lock().await;
-            self.layers
-                .write(LayerManagerLockHolder::Shutdown)
-                .await
-                .shutdown(&mut write_guard);
+            self.layers.write().await.shutdown(&mut write_guard);
         }
 
         // Finally wait until any gate-holders are complete.
@@ -2374,10 +2365,7 @@ impl Timeline {
         &self,
         reset: LayerAccessStatsReset,
     ) -> Result<LayerMapInfo, layer_manager::Shutdown> {
-        let guard = self
-            .layers
-            .read(LayerManagerLockHolder::GetLayerMapInfo)
-            .await;
+        let guard = self.layers.read().await;
         let layer_map = guard.layer_map()?;
         let mut in_memory_layers = Vec::with_capacity(layer_map.frozen_layers.len() + 1);
         if let Some(open_layer) = &layer_map.open_layer {
@@ -3244,7 +3232,7 @@ impl Timeline {
 
     /// Initialize with an empty layer map. Used when creating a new timeline.
     pub(super) fn init_empty_layer_map(&self, start_lsn: Lsn) {
-        let mut layers = self.layers.try_write(LayerManagerLockHolder::Init).expect(
+        let mut layers = self.layers.try_write().expect(
             "in the context where we call this function, no other task has access to the object",
         );
         layers
@@ -3264,10 +3252,7 @@ impl Timeline {
         use init::Decision::*;
         use init::{Discovered, DismissedLayer};
 
-        let mut guard = self
-            .layers
-            .write(LayerManagerLockHolder::LoadLayerMap)
-            .await;
+        let mut guard = self.layers.write().await;
 
         let timer = self.metrics.load_layer_map_histo.start_timer();
 
@@ -3884,10 +3869,7 @@ impl Timeline {
         &self,
         layer_name: &LayerName,
     ) -> Result<Option<Layer>, layer_manager::Shutdown> {
-        let guard = self
-            .layers
-            .read(LayerManagerLockHolder::GetLayerMapInfo)
-            .await;
+        let guard = self.layers.read().await;
         let layer = guard
             .layer_map()?
             .iter_historic_layers()
@@ -3920,10 +3902,7 @@ impl Timeline {
             return None;
         }
 
-        let guard = self
-            .layers
-            .read(LayerManagerLockHolder::GenerateHeatmap)
-            .await;
+        let guard = self.layers.read().await;
 
         // Firstly, if there's any heatmap left over from when this location
         // was a secondary, take that into account. Keep layers that are:
@@ -4021,10 +4000,7 @@ impl Timeline {
     }
 
     pub(super) async fn generate_unarchival_heatmap(&self, end_lsn: Lsn) -> PreviousHeatmap {
-        let guard = self
-            .layers
-            .read(LayerManagerLockHolder::GenerateHeatmap)
-            .await;
+        let guard = self.layers.read().await;
 
         let now = SystemTime::now();
         let mut heatmap_layers = Vec::default();
@@ -4366,7 +4342,7 @@ impl Timeline {
         query: &VersionedKeySpaceQuery,
     ) -> Result<LayerFringe, GetVectoredError> {
         let mut fringe = LayerFringe::new();
-        let guard = self.layers.read(LayerManagerLockHolder::GetPage).await;
+        let guard = self.layers.read().await;
 
         match query {
             VersionedKeySpaceQuery::Uniform { keyspace, lsn } => {
@@ -4469,7 +4445,7 @@ impl Timeline {
             // required for correctness, but avoids visiting extra layers
             // which turns out to be a perf bottleneck in some cases.
             if !unmapped_keyspace.is_empty() {
-                let guard = timeline.layers.read(LayerManagerLockHolder::GetPage).await;
+                let guard = timeline.layers.read().await;
                 guard.update_search_fringe(&unmapped_keyspace, cont_lsn, &mut fringe)?;
 
                 // It's safe to drop the layer map lock after planning the next round of reads.
@@ -4579,10 +4555,7 @@ impl Timeline {
         _guard: &tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
         ctx: &RequestContext,
     ) -> anyhow::Result<Arc<InMemoryLayer>> {
-        let mut guard = self
-            .layers
-            .write(LayerManagerLockHolder::GetLayerForWrite)
-            .await;
+        let mut guard = self.layers.write().await;
 
         let last_record_lsn = self.get_last_record_lsn();
         ensure!(
@@ -4624,10 +4597,7 @@ impl Timeline {
         write_lock: &mut tokio::sync::MutexGuard<'_, Option<TimelineWriterState>>,
     ) -> Result<u64, FlushLayerError> {
         let frozen = {
-            let mut guard = self
-                .layers
-                .write(LayerManagerLockHolder::TryFreezeLayer)
-                .await;
+            let mut guard = self.layers.write().await;
             guard
                 .open_mut()?
                 .try_freeze_in_memory_layer(at, &self.last_freeze_at, write_lock, &self.metrics)
@@ -4668,12 +4638,7 @@ impl Timeline {
         ctx: &RequestContext,
     ) {
         // Subscribe to L0 delta layer updates, for compaction backpressure.
-        let mut watch_l0 = match self
-            .layers
-            .read(LayerManagerLockHolder::FlushLoop)
-            .await
-            .layer_map()
-        {
+        let mut watch_l0 = match self.layers.read().await.layer_map() {
             Ok(lm) => lm.watch_level0_deltas(),
             Err(Shutdown) => return,
         };
@@ -4710,7 +4675,7 @@ impl Timeline {
 
                 // Fetch the next layer to flush, if any.
                 let (layer, l0_count, frozen_count, frozen_size) = {
-                    let layers = self.layers.read(LayerManagerLockHolder::FlushLoop).await;
+                    let layers = self.layers.read().await;
                     let Ok(lm) = layers.layer_map() else {
                         info!("dropping out of flush loop for timeline shutdown");
                         return;
@@ -5006,10 +4971,7 @@ impl Timeline {
         // in-memory layer from the map now. The flushed layer is stored in
         // the mapping in `create_delta_layer`.
         {
-            let mut guard = self
-                .layers
-                .write(LayerManagerLockHolder::FlushFrozenLayer)
-                .await;
+            let mut guard = self.layers.write().await;
 
             guard.open_mut()?.finish_flush_l0_layer(
                 delta_layer_to_add.as_ref(),
@@ -5224,7 +5186,7 @@ impl Timeline {
     async fn time_for_new_image_layer(&self, partition: &KeySpace, lsn: Lsn) -> bool {
         let threshold = self.get_image_creation_threshold();
 
-        let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
+        let guard = self.layers.read().await;
         let Ok(layers) = guard.layer_map() else {
             return false;
         };
@@ -5642,7 +5604,7 @@ impl Timeline {
             if let ImageLayerCreationMode::Force = mode {
                 // When forced to create image layers, we might try and create them where they already
                 // exist.  This mode is only used in tests/debug.
-                let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
+                let layers = self.layers.read().await;
                 if layers.contains_key(&PersistentLayerKey {
                     key_range: img_range.clone(),
                     lsn_range: PersistentLayerDesc::image_layer_lsn_range(lsn),
@@ -5767,7 +5729,7 @@ impl Timeline {
 
         let image_layers = batch_image_writer.finish(self, ctx).await?;
 
-        let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
+        let mut guard = self.layers.write().await;
 
         // FIXME: we could add the images to be uploaded *before* returning from here, but right
         // now they are being scheduled outside of write lock; current way is inconsistent with
@@ -5775,7 +5737,7 @@ impl Timeline {
         guard
             .open_mut()?
             .track_new_image_layers(&image_layers, &self.metrics);
-        drop_layer_manager_wlock(guard);
+        drop_wlock(guard);
         let duration = timer.stop_and_record();
 
         // Creating image layers may have caused some previously visible layers to be covered
@@ -6145,7 +6107,7 @@ impl Timeline {
         layers_to_remove: &[Layer],
     ) -> Result<(), CompactionError> {
         let mut guard = tokio::select! {
-            guard = self.layers.write(LayerManagerLockHolder::Compaction) => guard,
+            guard = self.layers.write() => guard,
             _ = self.cancel.cancelled() => {
                 return Err(CompactionError::ShuttingDown);
             }
@@ -6194,7 +6156,7 @@ impl Timeline {
         self.remote_client
             .schedule_compaction_update(&remove_layers, new_deltas)?;
 
-        drop_layer_manager_wlock(guard);
+        drop_wlock(guard);
 
         Ok(())
     }
@@ -6204,7 +6166,7 @@ impl Timeline {
         mut replace_layers: Vec<(Layer, ResidentLayer)>,
         mut drop_layers: Vec<Layer>,
     ) -> Result<(), CompactionError> {
-        let mut guard = self.layers.write(LayerManagerLockHolder::Compaction).await;
+        let mut guard = self.layers.write().await;
 
         // Trim our lists in case our caller (compaction) raced with someone else (GC) removing layers: we want
         // to avoid double-removing, and avoid rewriting something that was removed.
@@ -6555,10 +6517,7 @@ impl Timeline {
         // 5. newer on-disk image layers cover the layer's whole key range
         //
         // TODO holding a write lock is too agressive and avoidable
-        let mut guard = self
-            .layers
-            .write(LayerManagerLockHolder::GarbageCollection)
-            .await;
+        let mut guard = self.layers.write().await;
         let layers = guard.layer_map()?;
         'outer: for l in layers.iter_historic_layers() {
             result.layers_total += 1;
@@ -6860,10 +6819,7 @@ impl Timeline {
         use pageserver_api::models::DownloadRemoteLayersTaskState;
 
         let remaining = {
-            let guard = self
-                .layers
-                .read(LayerManagerLockHolder::GetLayerMapInfo)
-                .await;
+            let guard = self.layers.read().await;
             let Ok(lm) = guard.layer_map() else {
                 // technically here we could look into iterating accessible layers, but downloading
                 // all layers of a shutdown timeline makes no sense regardless.
@@ -6969,7 +6925,7 @@ impl Timeline {
 impl Timeline {
     /// Returns non-remote layers for eviction.
     pub(crate) async fn get_local_layers_for_disk_usage_eviction(&self) -> DiskUsageEvictionInfo {
-        let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
+        let guard = self.layers.read().await;
         let mut max_layer_size: Option<u64> = None;
 
         let resident_layers = guard
@@ -7070,7 +7026,7 @@ impl Timeline {
         let image_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
         info!("force created image layer {}", image_layer.local_path());
         {
-            let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
+            let mut guard = self.layers.write().await;
             guard
                 .open_mut()
                 .unwrap()
@@ -7133,7 +7089,7 @@ impl Timeline {
         let delta_layer = Layer::finish_creating(self.conf, self, desc, &path)?;
         info!("force created delta layer {}", delta_layer.local_path());
         {
-            let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
+            let mut guard = self.layers.write().await;
             guard
                 .open_mut()
                 .unwrap()
@@ -7228,7 +7184,7 @@ impl Timeline {
 
         // Link the layer to the layer map
         {
-            let mut guard = self.layers.write(LayerManagerLockHolder::Testing).await;
+            let mut guard = self.layers.write().await;
             let layer_map = guard.open_mut().unwrap();
             layer_map.force_insert_in_memory_layer(Arc::new(layer));
         }
@@ -7245,7 +7201,7 @@ impl Timeline {
         io_concurrency: IoConcurrency,
     ) -> anyhow::Result<Vec<(Key, Bytes)>> {
         let mut all_data = Vec::new();
-        let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
+        let guard = self.layers.read().await;
         for layer in guard.layer_map()?.iter_historic_layers() {
             if !layer.is_delta() && layer.image_layer_lsn() == lsn {
                 let layer = guard.get_from_desc(&layer);
@@ -7274,7 +7230,7 @@ impl Timeline {
         self: &Arc<Timeline>,
     ) -> anyhow::Result<Vec<super::storage_layer::PersistentLayerKey>> {
         let mut layers = Vec::new();
-        let guard = self.layers.read(LayerManagerLockHolder::Testing).await;
+        let guard = self.layers.read().await;
         for layer in guard.layer_map()?.iter_historic_layers() {
             layers.push(layer.key());
         }
@@ -7386,7 +7342,7 @@ impl TimelineWriter<'_> {
         let l0_count = self
             .tl
             .layers
-            .read(LayerManagerLockHolder::GetLayerMapInfo)
+            .read()
             .await
             .layer_map()?
             .level0_deltas()
@@ -7605,7 +7561,6 @@ mod tests {
     use crate::tenant::harness::{TenantHarness, test_img};
     use crate::tenant::layer_map::LayerMap;
     use crate::tenant::storage_layer::{Layer, LayerName, LayerVisibilityHint};
-    use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
     use crate::tenant::timeline::{DeltaLayerTestDesc, EvictionError};
     use crate::tenant::{PreviousHeatmap, Timeline};
 
@@ -7713,7 +7668,7 @@ mod tests {
         // Evict all the layers and stash the old heatmap in the timeline.
         // This simulates a migration to a cold secondary location.
 
-        let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+        let guard = timeline.layers.read().await;
         let mut all_layers = Vec::new();
         let forever = std::time::Duration::from_secs(120);
         for layer in guard.likely_resident_layers() {
@@ -7835,7 +7790,7 @@ mod tests {
             })));
 
         // Evict all the layers in the previous heatmap
-        let guard = timeline.layers.read(LayerManagerLockHolder::Testing).await;
+        let guard = timeline.layers.read().await;
         let forever = std::time::Duration::from_secs(120);
         for layer in guard.likely_resident_layers() {
             layer.evict_and_wait(forever).await.unwrap();
@@ -7898,10 +7853,7 @@ mod tests {
     }
 
     async fn find_some_layer(timeline: &Timeline) -> Layer {
-        let layers = timeline
-            .layers
-            .read(LayerManagerLockHolder::GetLayerMapInfo)
-            .await;
+        let layers = timeline.layers.read().await;
         let desc = layers
             .layer_map()
             .unwrap()

pageserver/src/tenant/timeline/analysis.rs

@@ -4,7 +4,6 @@ use std::ops::Range;
 use utils::lsn::Lsn;
 
 use super::Timeline;
-use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 
 #[derive(serde::Serialize)]
 pub(crate) struct RangeAnalysis {
@@ -25,10 +24,7 @@ impl Timeline {
 
         let num_of_l0;
         let all_layer_files = {
-            let guard = self
-                .layers
-                .read(LayerManagerLockHolder::GetLayerMapInfo)
-                .await;
+            let guard = self.layers.read().await;
             num_of_l0 = guard.layer_map().unwrap().level0_deltas().len();
             guard.all_persistent_layers()
         };

pageserver/src/tenant/timeline/compaction.rs

@@ -9,7 +9,7 @@ use std::ops::{Deref, Range};
 use std::sync::Arc;
 use std::time::{Duration, Instant};
 
-use super::layer_manager::{LayerManagerLockHolder, LayerManagerReadGuard};
+use super::layer_manager::LayerManager;
 use super::{
     CompactFlags, CompactOptions, CompactionError, CreateImageLayersError, DurationRecorder,
     GetVectoredError, ImageLayerCreationMode, LastImageLayerCreationStatus, RecordedDuration,
@@ -62,7 +62,7 @@ use crate::tenant::storage_layer::{
 use crate::tenant::tasks::log_compaction_error;
 use crate::tenant::timeline::{
     DeltaLayerWriter, ImageLayerCreationOutcome, ImageLayerWriter, IoConcurrency, Layer,
-    ResidentLayer, drop_layer_manager_rlock,
+    ResidentLayer, drop_rlock,
 };
 use crate::tenant::{DeltaLayer, MaybeOffloaded};
 use crate::virtual_file::{MaybeFatalIo, VirtualFile};
@@ -314,10 +314,7 @@ impl GcCompactionQueue {
             .unwrap_or(Lsn::INVALID);
 
         let layers = {
-            let guard = timeline
-                .layers
-                .read(LayerManagerLockHolder::GetLayerMapInfo)
-                .await;
+            let guard = timeline.layers.read().await;
             let layer_map = guard.layer_map()?;
             layer_map.iter_historic_layers().collect_vec()
         };
@@ -411,10 +408,7 @@ impl GcCompactionQueue {
         timeline: &Arc<Timeline>,
         lsn: Lsn,
     ) -> Result<u64, CompactionError> {
-        let guard = timeline
-            .layers
-            .read(LayerManagerLockHolder::GetLayerMapInfo)
-            .await;
+        let guard = timeline.layers.read().await;
         let layer_map = guard.layer_map()?;
         let layers = layer_map.iter_historic_layers().collect_vec();
         let mut size = 0;
@@ -857,7 +851,7 @@ impl KeyHistoryRetention {
         }
         let layer_generation;
         {
-            let guard = tline.layers.read(LayerManagerLockHolder::Compaction).await;
+            let guard = tline.layers.read().await;
             if !guard.contains_key(key) {
                 return false;
             }
@@ -1288,10 +1282,7 @@ impl Timeline {
             // We do the repartition on the L0-L1 boundary. All data below the boundary
             // are compacted by L0 with low read amplification, thus making the `repartition`
             // function run fast.
-            let guard = self
-                .layers
-                .read(LayerManagerLockHolder::GetLayerMapInfo)
-                .await;
+            let guard = self.layers.read().await;
             guard
                 .all_persistent_layers()
                 .iter()
@@ -1470,7 +1461,7 @@ impl Timeline {
         let latest_gc_cutoff = self.get_applied_gc_cutoff_lsn();
         let pitr_cutoff = self.gc_info.read().unwrap().cutoffs.time;
 
-        let layers = self.layers.read(LayerManagerLockHolder::Compaction).await;
+        let layers = self.layers.read().await;
         let layers_iter = layers.layer_map()?.iter_historic_layers();
         let (layers_total, mut layers_checked) = (layers_iter.len(), 0);
         for layer_desc in layers_iter {
@@ -1731,10 +1722,7 @@ impl Timeline {
         // are implicitly left visible, because LayerVisibilityHint's default is Visible, and we never modify it here.
         // Note that L0 deltas _can_ be covered by image layers, but we consider them 'visible' because we anticipate that
         // they will be subject to L0->L1 compaction in the near future.
-        let layer_manager = self
-            .layers
-            .read(LayerManagerLockHolder::GetLayerMapInfo)
-            .await;
+        let layer_manager = self.layers.read().await;
         let layer_map = layer_manager.layer_map()?;
 
         let readable_points = {
@@ -1787,7 +1775,7 @@ impl Timeline {
             };
 
             let begin = tokio::time::Instant::now();
-            let phase1_layers_locked = self.layers.read(LayerManagerLockHolder::Compaction).await;
+            let phase1_layers_locked = self.layers.read().await;
             let now = tokio::time::Instant::now();
             stats.read_lock_acquisition_micros =
                 DurationRecorder::Recorded(RecordedDuration(now - begin), now);
@@ -1815,7 +1803,7 @@ impl Timeline {
     /// Level0 files first phase of compaction, explained in the [`Self::compact_legacy`] comment.
     async fn compact_level0_phase1<'a>(
         self: &'a Arc<Self>,
-        guard: LayerManagerReadGuard<'a>,
+        guard: tokio::sync::RwLockReadGuard<'a, LayerManager>,
         mut stats: CompactLevel0Phase1StatsBuilder,
         target_file_size: u64,
         force_compaction_ignore_threshold: bool,
@@ -2041,7 +2029,7 @@ impl Timeline {
             holes
         };
         stats.read_lock_held_compute_holes_micros = stats.read_lock_held_key_sort_micros.till_now();
-        drop_layer_manager_rlock(guard);
+        drop_rlock(guard);
 
         if self.cancel.is_cancelled() {
             return Err(CompactionError::ShuttingDown);
@@ -2481,7 +2469,7 @@ impl Timeline {
 
         // Find the top of the historical layers
         let end_lsn = {
-            let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
+            let guard = self.layers.read().await;
             let layers = guard.layer_map()?;
 
             let l0_deltas = layers.level0_deltas();
@@ -3020,7 +3008,7 @@ impl Timeline {
         }
         split_key_ranges.sort();
         let all_layers = {
-            let guard = self.layers.read(LayerManagerLockHolder::Compaction).await;
+            let guard = self.layers.read().await;
             let layer_map = guard.layer_map()?;
             layer_map.iter_historic_layers().collect_vec()
         };
@@ -3124,12 +3112,12 @@ impl Timeline {
                 .await?;
             let jobs_len = jobs.len();
             for (idx, job) in jobs.into_iter().enumerate() {
-                let sub_compaction_progress = format!("{}/{}", idx + 1, jobs_len);
+                info!(
+                    "running enhanced gc bottom-most compaction, sub-compaction {}/{}",
+                    idx + 1,
+                    jobs_len
+                );
                 self.compact_with_gc_inner(cancel, job, ctx, yield_for_l0)
-                    .instrument(info_span!(
-                        "sub_compaction",
-                        sub_compaction_progress = sub_compaction_progress
-                    ))
                     .await?;
             }
             if jobs_len == 0 {
@@ -3197,10 +3185,7 @@ impl Timeline {
         // 1. If a layer is in the selection, all layers below it are in the selection.
         // 2. Inferred from (1), for each key in the layer selection, the value can be reconstructed only with the layers in the layer selection.
         let job_desc = {
-            let guard = self
-                .layers
-                .read(LayerManagerLockHolder::GarbageCollection)
-                .await;
+            let guard = self.layers.read().await;
             let layers = guard.layer_map()?;
             let gc_info = self.gc_info.read().unwrap();
             let mut retain_lsns_below_horizon = Vec::new();
@@ -3971,10 +3956,7 @@ impl Timeline {
 
         // First, do a sanity check to ensure the newly-created layer map does not contain overlaps.
         let all_layers = {
-            let guard = self
-                .layers
-                .read(LayerManagerLockHolder::GarbageCollection)
-                .await;
+            let guard = self.layers.read().await;
             let layer_map = guard.layer_map()?;
             layer_map.iter_historic_layers().collect_vec()
         };
@@ -4038,10 +4020,7 @@ impl Timeline {
             let update_guard = self.gc_compaction_layer_update_lock.write().await;
             // Acquiring the update guard ensures current read operations end and new read operations are blocked.
             // TODO: can we use `latest_gc_cutoff` Rcu to achieve the same effect?
-            let mut guard = self
-                .layers
-                .write(LayerManagerLockHolder::GarbageCollection)
-                .await;
+            let mut guard = self.layers.write().await;
             guard
                 .open_mut()?
                 .finish_gc_compaction(&layer_selection, &compact_to, &self.metrics);
@@ -4109,11 +4088,7 @@ impl TimelineAdaptor {
 
     pub async fn flush_updates(&mut self) -> Result<(), CompactionError> {
         let layers_to_delete = {
-            let guard = self
-                .timeline
-                .layers
-                .read(LayerManagerLockHolder::Compaction)
-                .await;
+            let guard = self.timeline.layers.read().await;
             self.layers_to_delete
                 .iter()
                 .map(|x| guard.get_from_desc(x))
@@ -4158,11 +4133,7 @@ impl CompactionJobExecutor for TimelineAdaptor {
     ) -> anyhow::Result<Vec<OwnArc<PersistentLayerDesc>>> {
         self.flush_updates().await?;
 
-        let guard = self
-            .timeline
-            .layers
-            .read(LayerManagerLockHolder::Compaction)
-            .await;
+        let guard = self.timeline.layers.read().await;
         let layer_map = guard.layer_map()?;
 
         let result = layer_map
@@ -4201,11 +4172,7 @@ impl CompactionJobExecutor for TimelineAdaptor {
         // this is a lot more complex than a simple downcast...
         if layer.is_delta() {
             let l = {
-                let guard = self
-                    .timeline
-                    .layers
-                    .read(LayerManagerLockHolder::Compaction)
-                    .await;
+                let guard = self.timeline.layers.read().await;
                 guard.get_from_desc(layer)
             };
             let result = l.download_and_keep_resident(ctx).await?;

pageserver/src/tenant/timeline/detach_ancestor.rs

@@ -19,7 +19,7 @@ use utils::id::TimelineId;
 use utils::lsn::Lsn;
 use utils::sync::gate::GateError;
 
-use super::layer_manager::{LayerManager, LayerManagerLockHolder};
+use super::layer_manager::LayerManager;
 use super::{FlushLayerError, Timeline};
 use crate::context::{DownloadBehavior, RequestContext};
 use crate::task_mgr::TaskKind;
@@ -199,10 +199,7 @@ pub(crate) async fn generate_tombstone_image_layer(
     let image_lsn = ancestor_lsn;
 
     {
-        let layers = detached
-            .layers
-            .read(LayerManagerLockHolder::DetachAncestor)
-            .await;
+        let layers = detached.layers.read().await;
         for layer in layers.all_persistent_layers() {
             if !layer.is_delta
                 && layer.lsn_range.start == image_lsn
@@ -426,7 +423,7 @@ pub(super) async fn prepare(
         // we do not need to start from our layers, because they can only be layers that come
         // *after* ancestor_lsn
         let layers = tokio::select! {
-            guard = ancestor.layers.read(LayerManagerLockHolder::DetachAncestor) => guard,
+            guard = ancestor.layers.read() => guard,
             _ = detached.cancel.cancelled() => {
                 return Err(ShuttingDown);
             }
@@ -872,12 +869,7 @@ async fn remote_copy(
 
                 // Double check that the file is orphan (probably from an earlier attempt), then delete it
                 let key = file_name.clone().into();
-                if adoptee
-                    .layers
-                    .read(LayerManagerLockHolder::DetachAncestor)
-                    .await
-                    .contains_key(&key)
-                {
+                if adoptee.layers.read().await.contains_key(&key) {
                     // We are supposed to filter out such cases before coming to this function
                     return Err(Error::Prepare(anyhow::anyhow!(
                         "layer file {file_name} already present and inside layer map"

pageserver/src/tenant/timeline/eviction_task.rs

@@ -33,7 +33,6 @@ use crate::tenant::size::CalculateSyntheticSizeError;
 use crate::tenant::storage_layer::LayerVisibilityHint;
 use crate::tenant::tasks::{BackgroundLoopKind, BackgroundLoopSemaphorePermit, sleep_random};
 use crate::tenant::timeline::EvictionError;
-use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 use crate::tenant::{LogicalSizeCalculationCause, TenantShard};
 
 #[derive(Default)]
@@ -209,7 +208,7 @@ impl Timeline {
 
         let mut js = tokio::task::JoinSet::new();
         {
-            let guard = self.layers.read(LayerManagerLockHolder::Eviction).await;
+            let guard = self.layers.read().await;
 
             guard
                 .likely_resident_layers()

pageserver/src/tenant/timeline/import_pgdata.rs

@@ -15,7 +15,6 @@ use super::{Timeline, TimelineDeleteProgress};
 use crate::context::RequestContext;
 use crate::controller_upcall_client::{StorageControllerUpcallApi, StorageControllerUpcallClient};
 use crate::tenant::metadata::TimelineMetadata;
-use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 
 mod flow;
 mod importbucket_client;
@@ -164,10 +163,7 @@ async fn prepare_import(
     info!("wipe the slate clean");
     {
         // TODO: do we need to hold GC lock for this?
-        let mut guard = timeline
-            .layers
-            .write(LayerManagerLockHolder::ImportPgData)
-            .await;
+        let mut guard = timeline.layers.write().await;
         assert!(
             guard.layer_map()?.open_layer.is_none(),
             "while importing, there should be no in-memory layer" // this just seems like a good place to assert it

pageserver/src/tenant/timeline/import_pgdata/flow.rs

@@ -56,7 +56,6 @@ use crate::pgdatadir_mapping::{
 };
 use crate::task_mgr::TaskKind;
 use crate::tenant::storage_layer::{AsLayerDesc, ImageLayerWriter, Layer};
-use crate::tenant::timeline::layer_manager::LayerManagerLockHolder;
 
 pub async fn run(
     timeline: Arc<Timeline>,
@@ -985,10 +984,7 @@ impl ChunkProcessingJob {
             let (desc, path) = writer.finish(ctx).await?;
 
             {
-                let guard = timeline
-                    .layers
-                    .read(LayerManagerLockHolder::ImportPgData)
-                    .await;
+                let guard = timeline.layers.read().await;
                 let existing_layer = guard.try_get_from_key(&desc.key());
                 if let Some(layer) = existing_layer {
                     if layer.metadata().generation == timeline.generation {
@@ -1011,10 +1007,7 @@ impl ChunkProcessingJob {
         // certain that the existing layer is identical to the new one, so in that case
         // we replace the old layer with the one we just generated.
 
-        let mut guard = timeline
-            .layers
-            .write(LayerManagerLockHolder::ImportPgData)
-            .await;
+        let mut guard = timeline.layers.write().await;
 
         let existing_layer = guard
             .try_get_from_key(&resident_layer.layer_desc().key())
@@ -1043,7 +1036,7 @@ impl ChunkProcessingJob {
             }
         }
 
-        crate::tenant::timeline::drop_layer_manager_wlock(guard);
+        crate::tenant::timeline::drop_wlock(guard);
 
         timeline
             .remote_client

pageserver/src/tenant/timeline/layer_manager.rs

@@ -1,8 +1,5 @@
 use std::collections::HashMap;
-use std::mem::ManuallyDrop;
-use std::ops::{Deref, DerefMut};
 use std::sync::Arc;
-use std::time::Duration;
 
 use anyhow::{Context, bail, ensure};
 use itertools::Itertools;
@@ -23,155 +20,6 @@ use crate::tenant::storage_layer::{
     PersistentLayerKey, ReadableLayerWeak, ResidentLayer,
 };
 
-/// Warn if the lock was held for longer than this threshold.
-/// It's very generous and we should bring this value down over time.
-const LAYER_MANAGER_LOCK_WARN_THRESHOLD: Duration = Duration::from_secs(5);
-const LAYER_MANAGER_LOCK_READ_WARN_THRESHOLD: Duration = Duration::from_secs(30);
-
-/// Describes the operation that is holding the layer manager lock
-#[derive(Debug, Clone, Copy, strum_macros::Display)]
-#[strum(serialize_all = "kebab_case")]
-pub(crate) enum LayerManagerLockHolder {
-    GetLayerMapInfo,
-    GenerateHeatmap,
-    GetPage,
-    Init,
-    LoadLayerMap,
-    GetLayerForWrite,
-    TryFreezeLayer,
-    FlushFrozenLayer,
-    FlushLoop,
-    Compaction,
-    GarbageCollection,
-    Shutdown,
-    ImportPgData,
-    DetachAncestor,
-    Eviction,
-    #[cfg(test)]
-    Testing,
-}
-
-/// Wrapper for the layer manager that tracks the amount of time during which
-/// it was held under read or write lock
-#[derive(Default)]
-pub(crate) struct LockedLayerManager {
-    locked: tokio::sync::RwLock<LayerManager>,
-}
-
-pub(crate) struct LayerManagerReadGuard<'a> {
-    guard: ManuallyDrop<tokio::sync::RwLockReadGuard<'a, LayerManager>>,
-    acquired_at: std::time::Instant,
-    holder: LayerManagerLockHolder,
-}
-
-pub(crate) struct LayerManagerWriteGuard<'a> {
-    guard: ManuallyDrop<tokio::sync::RwLockWriteGuard<'a, LayerManager>>,
-    acquired_at: std::time::Instant,
-    holder: LayerManagerLockHolder,
-}
-
-impl Drop for LayerManagerReadGuard<'_> {
-    fn drop(&mut self) {
-        // Drop the lock first, before potentially warning if it was held for too long.
-        // SAFETY: ManuallyDrop in Drop implementation
-        unsafe { ManuallyDrop::drop(&mut self.guard) };
-
-        let held_for = self.acquired_at.elapsed();
-        if held_for >= LAYER_MANAGER_LOCK_READ_WARN_THRESHOLD {
-            tracing::warn!(
-                holder=%self.holder,
-                "Layer manager read lock held for {}s",
-                held_for.as_secs_f64(),
-            );
-        }
-    }
-}
-
-impl Drop for LayerManagerWriteGuard<'_> {
-    fn drop(&mut self) {
-        // Drop the lock first, before potentially warning if it was held for too long.
-        // SAFETY: ManuallyDrop in Drop implementation
-        unsafe { ManuallyDrop::drop(&mut self.guard) };
-
-        let held_for = self.acquired_at.elapsed();
-        if held_for >= LAYER_MANAGER_LOCK_WARN_THRESHOLD {
-            tracing::warn!(
-                holder=%self.holder,
-                "Layer manager write lock held for {}s",
-                held_for.as_secs_f64(),
-            );
-        }
-    }
-}
-
-impl Deref for LayerManagerReadGuard<'_> {
-    type Target = LayerManager;
-
-    fn deref(&self) -> &Self::Target {
-        self.guard.deref()
-    }
-}
-
-impl Deref for LayerManagerWriteGuard<'_> {
-    type Target = LayerManager;
-
-    fn deref(&self) -> &Self::Target {
-        self.guard.deref()
-    }
-}
-
-impl DerefMut for LayerManagerWriteGuard<'_> {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        self.guard.deref_mut()
-    }
-}
-
-impl LockedLayerManager {
-    pub(crate) async fn read(&self, holder: LayerManagerLockHolder) -> LayerManagerReadGuard {
-        let guard = ManuallyDrop::new(self.locked.read().await);
-        LayerManagerReadGuard {
-            guard,
-            acquired_at: std::time::Instant::now(),
-            holder,
-        }
-    }
-
-    pub(crate) fn try_read(
-        &self,
-        holder: LayerManagerLockHolder,
-    ) -> Result<LayerManagerReadGuard, tokio::sync::TryLockError> {
-        let guard = ManuallyDrop::new(self.locked.try_read()?);
-
-        Ok(LayerManagerReadGuard {
-            guard,
-            acquired_at: std::time::Instant::now(),
-            holder,
-        })
-    }
-
-    pub(crate) async fn write(&self, holder: LayerManagerLockHolder) -> LayerManagerWriteGuard {
-        let guard = ManuallyDrop::new(self.locked.write().await);
-        LayerManagerWriteGuard {
-            guard,
-            acquired_at: std::time::Instant::now(),
-            holder,
-        }
-    }
-
-    pub(crate) fn try_write(
-        &self,
-        holder: LayerManagerLockHolder,
-    ) -> Result<LayerManagerWriteGuard, tokio::sync::TryLockError> {
-        let guard = ManuallyDrop::new(self.locked.try_write()?);
-
-        Ok(LayerManagerWriteGuard {
-            guard,
-            acquired_at: std::time::Instant::now(),
-            holder,
-        })
-    }
-}
-
 /// Provides semantic APIs to manipulate the layer map.
 pub(crate) enum LayerManager {
     /// Open as in not shutdown layer manager; we still have in-memory layers and we can manipulate

pgxn/neon/Makefile

@@ -1,10 +1,10 @@
 # pgxs/neon/Makefile
 
-
 MODULE_big = neon
 OBJS = \
 	$(WIN32RES) \
 	communicator.o \
+	communicator_new.o \
 	extension_server.o \
 	file_cache.o \
 	hll.o \
@@ -22,11 +22,13 @@ OBJS = \
 	walproposer.o \
 	walproposer_pg.o \
 	control_plane_connector.o \
-	walsender_hooks.o
+	walsender_hooks.o \
+	$(LIBCOMMUNICATOR_PATH)/libcommunicator.a
 
 PG_CPPFLAGS = -I$(libpq_srcdir)
 SHLIB_LINK_INTERNAL = $(libpq)
 SHLIB_LINK = -lcurl
+SHLIB_LINK += -framework Security -framework CoreFoundation -framework CoreServices -framework IOKit -framework SystemConfiguration
 
 EXTENSION = neon
 DATA = \

pgxn/neon/communicator.c

@@ -1092,15 +1092,13 @@ communicator_prefetch_register_bufferv(BufferTag tag, neon_request_lsns *frlsns,
 		   MyPState->ring_last <= ring_index);
 }
 
-/* Internal version. Returns the ring index of the last block (result of this function is used only
-*  when nblocks==1)
-*/
+/* internal version. Returns the ring index */
 static uint64
 prefetch_register_bufferv(BufferTag tag, neon_request_lsns *frlsns,
 						  BlockNumber nblocks, const bits8 *mask,
 						  bool is_prefetch)
 {
-	uint64		last_ring_index;
+	uint64		min_ring_index;
 	PrefetchRequest hashkey;
 #ifdef USE_ASSERT_CHECKING
 	bool		any_hits = false;
@@ -1124,12 +1122,13 @@ Retry:
 		MyPState->ring_unused - MyPState->ring_receive;
 	MyNeonCounters->getpage_prefetches_buffered =
 		MyPState->n_responses_buffered;
-	last_ring_index = UINT64_MAX;
 
+	min_ring_index = UINT64_MAX;
 	for (int i = 0; i < nblocks; i++)
 	{
 		PrefetchRequest *slot = NULL;
 		PrfHashEntry *entry = NULL;
+		uint64		ring_index;
 		neon_request_lsns *lsns;
 
 		if (PointerIsValid(mask) && BITMAP_ISSET(mask, i))
@@ -1153,12 +1152,12 @@ Retry:
 		if (entry != NULL)
 		{
 			slot = entry->slot;
-			last_ring_index = slot->my_ring_index;
-			Assert(slot == GetPrfSlot(last_ring_index));
+			ring_index = slot->my_ring_index;
+			Assert(slot == GetPrfSlot(ring_index));
 
 			Assert(slot->status != PRFS_UNUSED);
-			Assert(MyPState->ring_last <= last_ring_index &&
-				   last_ring_index < MyPState->ring_unused);
+			Assert(MyPState->ring_last <= ring_index &&
+				   ring_index < MyPState->ring_unused);
 			Assert(BufferTagsEqual(&slot->buftag, &hashkey.buftag));
 
 			/*
@@ -1170,9 +1169,9 @@ Retry:
 				if (!neon_prefetch_response_usable(lsns, slot))
 				{
 					/* Wait for the old request to finish and discard it */
-					if (!prefetch_wait_for(last_ring_index))
+					if (!prefetch_wait_for(ring_index))
 						goto Retry;
-					prefetch_set_unused(last_ring_index);
+					prefetch_set_unused(ring_index);
 					entry = NULL;
 					slot = NULL;
 					pgBufferUsage.prefetch.expired += 1;
@@ -1189,12 +1188,13 @@ Retry:
 				 */
 				if (slot->status == PRFS_TAG_REMAINS)
 				{
-					prefetch_set_unused(last_ring_index);
+					prefetch_set_unused(ring_index);
 					entry = NULL;
 					slot = NULL;
 				}
 				else
 				{
+					min_ring_index = Min(min_ring_index, ring_index);
 					/* The buffered request is good enough, return that index */
 					if (is_prefetch)
 						pgBufferUsage.prefetch.duplicates++;
@@ -1283,12 +1283,12 @@ Retry:
 		 * The next buffer pointed to by `ring_unused` is now definitely empty, so
 		 * we can insert the new request to it.
 		 */
-		last_ring_index = MyPState->ring_unused;
+		ring_index = MyPState->ring_unused;
 
-		Assert(MyPState->ring_last <= last_ring_index &&
-			   last_ring_index <= MyPState->ring_unused);
+		Assert(MyPState->ring_last <= ring_index &&
+			   ring_index <= MyPState->ring_unused);
 
-		slot = GetPrfSlotNoCheck(last_ring_index);
+		slot = GetPrfSlotNoCheck(ring_index);
 
 		Assert(slot->status == PRFS_UNUSED);
 
@@ -1298,9 +1298,11 @@ Retry:
 		 */
 		slot->buftag = hashkey.buftag;
 		slot->shard_no = get_shard_number(&tag);
-		slot->my_ring_index = last_ring_index;
+		slot->my_ring_index = ring_index;
 		slot->flags = 0;
 
+		min_ring_index = Min(min_ring_index, ring_index);
+
 		if (is_prefetch)
 			MyNeonCounters->getpage_prefetch_requests_total++;
 		else
@@ -1313,12 +1315,11 @@ Retry:
 		MyPState->ring_unused - MyPState->ring_receive;
 
 	Assert(any_hits);
-	Assert(last_ring_index != UINT64_MAX);
 
-	Assert(GetPrfSlot(last_ring_index)->status == PRFS_REQUESTED ||
-		   GetPrfSlot(last_ring_index)->status == PRFS_RECEIVED);
-	Assert(MyPState->ring_last <= last_ring_index &&
-		   last_ring_index < MyPState->ring_unused);
+	Assert(GetPrfSlot(min_ring_index)->status == PRFS_REQUESTED ||
+		   GetPrfSlot(min_ring_index)->status == PRFS_RECEIVED);
+	Assert(MyPState->ring_last <= min_ring_index &&
+		   min_ring_index < MyPState->ring_unused);
 
 	if (flush_every_n_requests > 0 &&
 		MyPState->ring_unused - MyPState->ring_flush >= flush_every_n_requests)
@@ -1334,7 +1335,7 @@ Retry:
 		MyPState->ring_flush = MyPState->ring_unused;
 	}
 
-	return last_ring_index;
+	return min_ring_index;
 }
 
 static bool

pgxn/neon/communicator/Cargo.lock

@@ -0,0 +1,372 @@
+# This file is automatically @generated by Cargo.
+# It is not intended for manual editing.
+version = 4
+
+[[package]]
+name = "addr2line"
+version = "0.24.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dfbe277e56a376000877090da837660b4427aad530e3028d44e0bffe4f89a1c1"
+dependencies = [
+ "gimli",
+]
+
+[[package]]
+name = "adler2"
+version = "2.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "512761e0bb2578dd7380c6baaa0f4ce03e84f95e960231d1dec8bf4d7d6e2627"
+
+[[package]]
+name = "backtrace"
+version = "0.3.74"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8d82cb332cdfaed17ae235a638438ac4d4839913cc2af585c3c6746e8f8bee1a"
+dependencies = [
+ "addr2line",
+ "cfg-if",
+ "libc",
+ "miniz_oxide",
+ "object",
+ "rustc-demangle",
+ "windows-targets",
+]
+
+[[package]]
+name = "base64"
+version = "0.22.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "72b3254f16251a8381aa12e40e3c4d2f0199f8c6508fbecb9d91f575e0fbb8c6"
+
+[[package]]
+name = "bytes"
+version = "1.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d71b6127be86fdcfddb610f7182ac57211d4b18a3e9c82eb2d17662f2227ad6a"
+
+[[package]]
+name = "cfg-if"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd"
+
+[[package]]
+name = "communicator"
+version = "0.1.0"
+dependencies = [
+ "tonic",
+]
+
+[[package]]
+name = "fnv"
+version = "1.0.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3f9eec918d3f24069decb9af1554cad7c880e2da24a9afd88aca000531ab82c1"
+
+[[package]]
+name = "futures-core"
+version = "0.3.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "05f29059c0c2090612e8d742178b0580d2dc940c837851ad723096f87af6663e"
+
+[[package]]
+name = "gimli"
+version = "0.31.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07e28edb80900c19c28f1072f2e8aeca7fa06b23cd4169cefe1af5aa3260783f"
+
+[[package]]
+name = "http"
+version = "1.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f4a85d31aea989eead29a3aaf9e1115a180df8282431156e533de47660892565"
+dependencies = [
+ "bytes",
+ "fnv",
+ "itoa",
+]
+
+[[package]]
+name = "http-body"
+version = "1.0.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1efedce1fb8e6913f23e0c92de8e62cd5b772a67e7b3946df930a62566c93184"
+dependencies = [
+ "bytes",
+ "http",
+]
+
+[[package]]
+name = "http-body-util"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b021d93e26becf5dc7e1b75b1bed1fd93124b374ceb73f43d4d4eafec896a64a"
+dependencies = [
+ "bytes",
+ "futures-core",
+ "http",
+ "http-body",
+ "pin-project-lite",
+]
+
+[[package]]
+name = "itoa"
+version = "1.0.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4a5f13b858c8d314ee3e8f639011f7ccefe71f97f96e50151fb991f267928e2c"
+
+[[package]]
+name = "libc"
+version = "0.2.171"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c19937216e9d3aa9956d9bb8dfc0b0c8beb6058fc4f7a4dc4d850edf86a237d6"
+
+[[package]]
+name = "memchr"
+version = "2.7.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "78ca9ab1a0babb1e7d5695e3530886289c18cf2f87ec19a575a0abdce112e3a3"
+
+[[package]]
+name = "miniz_oxide"
+version = "0.8.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ff70ce3e48ae43fa075863cef62e8b43b71a4f2382229920e0df362592919430"
+dependencies = [
+ "adler2",
+]
+
+[[package]]
+name = "object"
+version = "0.36.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "62948e14d923ea95ea2c7c86c71013138b66525b86bdc08d2dcc262bdb497b87"
+dependencies = [
+ "memchr",
+]
+
+[[package]]
+name = "once_cell"
+version = "1.21.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "42f5e15c9953c5e4ccceeb2e7382a716482c34515315f7b03532b8b4e8393d2d"
+
+[[package]]
+name = "percent-encoding"
+version = "2.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e3148f5046208a5d56bcfc03053e3ca6334e51da8dfb19b6cdc8b306fae3283e"
+
+[[package]]
+name = "pin-project"
+version = "1.1.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "677f1add503faace112b9f1373e43e9e054bfdd22ff1a63c1bc485eaec6a6a8a"
+dependencies = [
+ "pin-project-internal",
+]
+
+[[package]]
+name = "pin-project-internal"
+version = "1.1.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6e918e4ff8c4549eb882f14b3a4bc8c8bc93de829416eacf579f1207a8fbf861"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "pin-project-lite"
+version = "0.2.16"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3b3cff922bd51709b605d9ead9aa71031d81447142d828eb4a6eba76fe619f9b"
+
+[[package]]
+name = "proc-macro2"
+version = "1.0.94"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a31971752e70b8b2686d7e46ec17fb38dad4051d94024c88df49b667caea9c84"
+dependencies = [
+ "unicode-ident",
+]
+
+[[package]]
+name = "quote"
+version = "1.0.40"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1885c039570dc00dcb4ff087a89e185fd56bae234ddc7f056a945bf36467248d"
+dependencies = [
+ "proc-macro2",
+]
+
+[[package]]
+name = "rustc-demangle"
+version = "0.1.24"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "719b953e2095829ee67db738b3bfa9fa368c94900df327b3f07fe6e794d2fe1f"
+
+[[package]]
+name = "syn"
+version = "2.0.100"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b09a44accad81e1ba1cd74a32461ba89dee89095ba17b32f5d03683b1b1fc2a0"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "unicode-ident",
+]
+
+[[package]]
+name = "tokio"
+version = "1.44.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6b88822cbe49de4185e3a4cbf8321dd487cf5fe0c5c65695fef6346371e9c48"
+dependencies = [
+ "backtrace",
+ "pin-project-lite",
+]
+
+[[package]]
+name = "tokio-stream"
+version = "0.1.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "eca58d7bba4a75707817a2c44174253f9236b2d5fbd055602e9d5c07c139a047"
+dependencies = [
+ "futures-core",
+ "pin-project-lite",
+ "tokio",
+]
+
+[[package]]
+name = "tonic"
+version = "0.13.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "85839f0b32fd242bb3209262371d07feda6d780d16ee9d2bc88581b89da1549b"
+dependencies = [
+ "base64",
+ "bytes",
+ "http",
+ "http-body",
+ "http-body-util",
+ "percent-encoding",
+ "pin-project",
+ "tokio-stream",
+ "tower-layer",
+ "tower-service",
+ "tracing",
+]
+
+[[package]]
+name = "tower-layer"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "121c2a6cda46980bb0fcd1647ffaf6cd3fc79a013de288782836f6df9c48780e"
+
+[[package]]
+name = "tower-service"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
+
+[[package]]
+name = "tracing"
+version = "0.1.41"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "784e0ac535deb450455cbfa28a6f0df145ea1bb7ae51b821cf5e7927fdcfbdd0"
+dependencies = [
+ "pin-project-lite",
+ "tracing-attributes",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-attributes"
+version = "0.1.28"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "395ae124c09f9e6918a2310af6038fba074bcf474ac352496d5910dd59a2226d"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
+[[package]]
+name = "tracing-core"
+version = "0.1.33"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e672c95779cf947c5311f83787af4fa8fffd12fb27e4993211a84bdfd9610f9c"
+dependencies = [
+ "once_cell",
+]
+
+[[package]]
+name = "unicode-ident"
+version = "1.0.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5a5f39404a5da50712a4c1eecf25e90dd62b613502b7e925fd4e4d19b5c96512"
+
+[[package]]
+name = "windows-targets"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973"
+dependencies = [
+ "windows_aarch64_gnullvm",
+ "windows_aarch64_msvc",
+ "windows_i686_gnu",
+ "windows_i686_gnullvm",
+ "windows_i686_msvc",
+ "windows_x86_64_gnu",
+ "windows_x86_64_gnullvm",
+ "windows_x86_64_msvc",
+]
+
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
+
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
+
+[[package]]
+name = "windows_i686_gnu"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
+
+[[package]]
+name = "windows_i686_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
+
+[[package]]
+name = "windows_i686_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
+
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
+
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
+
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.52.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"