Repository navigation
Conversation
…le branch
Two criterion benches under nativelink-util/benches/ and three additional
security tests in fs_util.rs that together prove (a) the clonefile fast
path on macOS is faster than per-file hardlinks and (b) the
download_to_directory concurrency cap is not a regression.
Results on Apple M4 Max / APFS (full table in BENCHMARKS.md):
- hardlink_directory_tree, 1978 files / 466 MB: 150 ms vs 590 ms = 3.93x
- hardlink_directory_tree, 635 files / 180 MB: 49 ms vs 181 ms = 3.70x
- download_to_directory_concurrency, n=1978: 893 ms vs 887 ms = 1.01x
(perf-neutral on single-process, as expected)
The treatment-vs-baseline comparison uses a new #[doc(hidden)]
hardlink_directory_tree_perfile helper that forces the per-file path so
both can be measured on the same host.
Security tests added (all green):
- test_clonefile_preserves_internal_symlinks (symlinks inside the tree
are cloned as symlinks, not followed)
- test_clonefile_nofollow_on_top_level_symlink_src (top-level symlink
src is cloned as a symlink, not followed)
- test_dst_under_file_parent_errors_cleanly (clean error, no half-
materialized tree on bad dst parent)
erneestoc
force-pushed
the
ec/macos-clonefile-optimizations-benchmarks
branch
2 times, most recently
from
May 16, 2026 19:26
59daf65 to
9e357d8
Compare
erneestoc
added a commit
that referenced
this pull request
May 22, 2026
Reverts the bounded-concurrency construct on this POC branch. After #1 makes construct metadata-only, intra-tree parallelism is bounded by APFS metadata serialization; on a busy worker (inter-action concurrency already saturates the box) 64-wide spawn_blocking fan-out risks oversubscription and stealing cycles from the compiles. Keeping #1/#2/TraceMachina#5. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
erneestoc
added a commit
that referenced
this pull request
Jul 2, 2026
…achina#2366) * util: make permission walks symlink-safe The recursive permission walk `set_perms_recursive_impl` (driving both `set_readonly_recursive` and `set_dir_writable_recursive`) used `fs::metadata` (stat), which follows symlinks. On input trees containing symlinks - e.g. `.venv/bin/python3` produced by rules_python / rules_apple venv tooling - this had two failure modes: * A symlink to a directory reported `is_dir() == true`, so the walk recursed *through* the link, escaping the materialized tree or descending into an unrelated directory. * A symlink was passed to `set_permissions`; `chmod` follows symlinks, so it mutated the link's target. When the target did not exist (a dangling link - common when a venv points outside the action's input set) the `chmod` returned ENOENT and failed the entire walk. That ENOENT failure surfaced as `set_readonly_recursive` erroring inside `DirectoryCache::get_or_create`, which made `prepare_action_inputs` log "Directory cache failed, falling back to traditional download" and take the slow `download_to_directory` path. Fix: `set_perms_recursive_impl` now uses `symlink_metadata` (lstat) and returns early on symlink entries - it never chmods a symlink and never recurses through one. Regular files keep their existing read-only (0o555) treatment, so the CAS-hardlinked-inode hermeticity contract (PR TraceMachina#2347) is unchanged. `hardlink_directory_tree_recursive` already recreated symlinks as symlinks; its symlink branch is reordered ahead of the `is_dir()` / `is_file()` branches to make the symlink-first intent explicit and robust. Adds regression tests covering set-readonly, set-dir-writable, and hardlink/clone walks over a tree containing a symlink to an in-tree file, a dangling relative symlink, and a symlink to an in-tree directory, asserting each walk succeeds and the symlinks are preserved with their targets intact. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * worker: make directory-cache entries already-writable `DirectoryCache` locks each cache entry down with `set_readonly_recursive` after construction. Previously that helper made the entire entry tree mode 0o555 — directories included — so every materialization had to follow up with a separate `set_dir_writable_recursive` recursive chmod walk in `prepare_action_inputs` to re-add write permission to directories (Bazel actions declare outputs at paths nested inside input subdirectories). That post-walk is redundant work. Directories are not hardlink-shared between cache entries — only file content inodes are — so directory mode can safely be made writable once, at the cache entry, instead of on every materialization. `set_readonly_recursive` now locks a tree down as a cache entry by making only FILES read-only (0o555) and leaving DIRECTORIES writable (0o755). Both materialization paths then produce a directly-usable tree: - macOS `clonefile(2)` copies the source's modes verbatim, so the clone's directories are writable and its files read-only. - The Linux per-file hardlink walk creates fresh directories (writable) and hardlinks files (which keep the source inode's read-only mode). Files stay read-only on both paths, so the hermeticity contract and the CAS-hardlink shared-inode invariant (PR TraceMachina#2347) are preserved. With the materialized tree already correct, the `set_dir_writable_recursive` call is removed from `prepare_action_inputs`. `set_dir_writable_recursive` itself is unchanged and still used by the cache eviction cleanup path. Tests: - fs_util: `test_set_readonly_recursive` now also asserts directories stay writable; the macOS clonefile tests assert cloned subdirs are writable and that a nested output can be created with no `set_dir_writable_recursive` walk; `test_set_dir_writable_recursive_walks_nested_dirs` keeps covering the eviction-cleanup helper. - directory_cache: new `test_materialized_tree_dirs_writable_files_readonly` builds a nested tree and asserts that, after `get_or_create` on both the fresh-materialize and cache-hit paths, every directory is writable and every file is read-only, with no separate chmod walk. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * worker: hardlink CAS blobs in directory-cache construct `DirectoryCache::construct_directory` previously materialized every file by fetching the whole blob into RAM (`get_part_unchunked`) and writing a full copy (`fs::write`). For a cache that exists to avoid re-fetching from the CAS, this is the dominant cost on a miss. Switch the cache-entry file build to hardlink the FilesystemStore CAS blob directly into the cache entry — zero-copy, metadata-only — exactly the way `download_to_directory` already does on the fallback path: `populate_fast_store` then `get_file_entry_for_digest` / `get_file_path_locked` / `fs::hard_link`. Correctness: * A hardlinked CAS blob shares its inode with the CAS store and every other action that hardlinked the same blob, so it must never be chmod'd (the inode-corruption bug PR TraceMachina#2347 fixed). Executable files (`FileNode.is_executable`) therefore get their own private inode via fetch+write and are chmod'd 0o555 on that unshared copy — never hardlinked. * When the blob is not locally hardlinkable (the fast tier is not a FilesystemStore, or the blob is absent / evicted from it), the file falls back to fetch+write rather than failing the build. * Zero-byte files keep their existing direct-write special case. * The post-construction lockdown switches from `set_readonly_recursive` (which chmods files, and would corrupt the shared CAS inode) to `set_dir_writable_recursive`, which only touches directories. `DirectoryCache::new` now takes the worker's `Arc<FastSlowStore>` so it can reach `populate_fast_store` and downcast the fast tier to FilesystemStore. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * worker: drop the two redundant full-tree walks in directory-cache build After `construct_directory`, the cache-miss path walked the materialized tree twice more: `calculate_directory_size` (an `fs::metadata` per file) to compute the LRU size, and a recursive permission pass to normalize directory modes. Both are now folded into construction itself. Size: `construct_directory` returns the total tree size, accumulated from `FileNode.digest.size_bytes` in the `Directory` protos it already decodes. This is also more correct than the old filesystem walk — it counts each file once by its CAS size and never follows symlinks into possibly-shared or external targets. Symlinks contribute nothing. Directory mode: each cache-entry directory is chmod'd 0o755 the moment it is created (`create_dir_writable`), umask-independent. The directory is writable while it is populated and that is its stable final mode, so the separate post-construction `set_dir_writable_recursive` walk is gone. Cache-entry files are still never chmod'd here — they may be CAS-blob hardlinks (OPT #1) and mutating their mode would corrupt the shared inode. Reconciliation with PR TraceMachina#2357: that PR reworks `set_readonly_recursive` so the recursive walk leaves dirs 0o755 / files 0o555. This commit removes the directory-cache build's dependence on any such recursive walk entirely — modes are set at creation. Whichever lands second, the rebase is a straight delete of the now-unused call site; there is no semantic conflict because both converge on 0o755 directories, and TraceMachina#2357's file handling is irrelevant here since the cache build no longer touches file modes at all. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * worker: narrow the directory-cache lock and single-flight construction The cache write lock was held across syscall-heavy I/O, serializing every concurrent `get_or_create`: * On the cache-hit paths, `cache.write()` was held across the whole `hardlink_directory_tree` (clonefile / per-file hardlink) materialization. * `evict_lru` ran `set_dir_writable_recursive` + `remove_dir_all` on the evicted tree under the write lock during a cache miss. Lock narrowing: * `acquire_entry` / `release_entry` take the write lock only to bump and drop a `ref_count` pin and snapshot the entry path; the `hardlink_directory_tree` materialization runs fully unlocked. The pin is what makes this safe — `evict_lru` never selects an entry with `ref_count > 0`, so the cache tree cannot be deleted mid-hardlink. The newly constructed entry is likewise inserted pre-pinned (`ref_count: 1`) and unpinned only after its destination hardlink completes; otherwise a concurrent miss for an unrelated digest could evict the brand-new entry (its `last_access` is recent but it is the only unpinned one) while this caller is still hardlinking from it. * `evict_if_needed` / `evict_lru` are now pure in-memory: they select victims and remove them from the map under the lock, returning the victim paths. `dispatch_evictions` then performs the chmod + removal on a `background_spawn` task, off the lock. Single-flight: the existing per-digest construction mutex already ensures a digest is constructed once while N callers wait; this commit additionally unmaps the per-digest mutex (`forget_construction_lock`) once construction finishes so `construction_locks` no longer grows unbounded over the worker's lifetime. Unmapping is race-free: a waiter has already cloned the `Arc<Mutex>` before blocking, and a late arrival that creates a fresh mutex still re-checks the cache, finds the entry, and takes the fast hardlink path — never a redundant construct. `ref_count` / `CachedDirectoryMetadata` semantics are unchanged; the hit/miss return contract is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * address merge interactions with read-only CAS * remove inode stat in test * update dependencies to 1.3.1 * worker: materialize executable inputs by hardlink to a created-once 0o555 variant (fix ETXTBSY) Materializing an executable input via a per-action `std::fs::copy` opened a writable fd in the worker's hot prepare path. Under fork-heavy concurrency a sibling action's forked child could inherit that fd, and a concurrent `execve` of the executable then failed with `ETXTBSY` ("Text file busy", os error 26) — seen on Linux RBE (k8) building rules_go's `builder`. macOS was largely shielded because its directory-cache path uses APFS `clonefile(2)` (a distinct COW inode per action), but the per-file `download_to_directory` fallback hardlinks on both platforms, so the regression spanned both. Fix (keep the hot path hardlink-only — no writable fd): - nativelink-store: add `FilesystemStore::get_executable_hardlink_source`. The CAS blob is read-only 0o444 and shared by hardlink, so it cannot carry +x and must never be chmod'd (TraceMachina#2347). This creates a per-digest 0o555 variant exactly once (single-flight), copy -> chmod -> fsync -> atomic rename, so the writer fd is closed before the inode is ever hardlinked or executed. Stored in a sibling `{content_path}.exec` dir (ignored by the content/temp scan + prune) and cleared on startup. On APFS the copy is itself a `clonefile`. - download_to_directory: executables now hardlink that shared 0o555 variant and non-executables hardlink the 0o444 CAS blob. A private copy is used only for the rare custom unix_mode / mtime case, applied to a private inode. The macOS `clonefile` materialization (`hardlink_directory_tree`, TraceMachina#2349) and the directory cache's executable handling are left untouched, preserving the macOS speedup. Test: executable_hardlink_source_created_once_and_readonly asserts the variant is 0o555, a separate inode from the 0o444 blob, stable across calls, leaves the blob untouched, and hardlinks into an executable. nativelink-store 243/0, nativelink-worker 88/0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * filesystem_store: use nativelink spawn_blocking! macro (clippy disallowed_methods) tokio::task::spawn_blocking is banned by clippy.toml in favor of nativelink-util's spawn_blocking! macro (adds the tracing span + JoinHandleDropGuard). Fixes the -D clippy::disallowed-methods CI failure on get_executable_hardlink_source's executable-variant creation. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * filesystem_store: gate executable-variant machinery to unix (fix Windows build) The executable 0o555 variant (and its single-flight map, variant path, .exec dir, and spawn_blocking copy) only exists to carry the unix executable bit and dodge the unix ETXTBSY race. On Windows it was dead code, failing the build under -D warnings (unused import spawn_blocking, never-read executable_locks, never-used executable_variant_path). Gate all of it (and the HashMap / Mutex / EXECUTABLE_DIR_SUFFIX it pulls in) behind #[cfg(unix)]; the existing #[cfg(not(unix))] get_executable_hardlink_source just hardlinks the CAS blob. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Ernesto Cambuston <e.cambuston@gmail.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why this branch exists
Our CI Bazel builds on the macOS Nativelink RBE workers spend more wall time materializing action inputs than actually running the compile. From a real CI run:
Input-side phases total ~7,400s vs ~4,900s for actual compiles. The worker is spending 1.5x more wall time setting up inputs than running them.
What's being shipped
The base branch (
ec/macos-clonefile-optimizations) extracts four focused commits from TraceMachina/nativelink#2243 — a 156-commit PR we did NOT want to merge wholesale because it's entangled with unrelated refactors (Moka, Redis, QUIC, scheduler):clonefile(2)fast path inhardlink_directory_tree(macOS only)CloneMethodtelemetry so production logs can confirm the fast path is engageddownload_to_directory(defends against APFS metadata-lock thrash)directory_cache(fixes the ~30% of cache constructions that fail on empty marker files)Linux / Windows paths are untouched.
What this PR adds
Benchmarks and security tests proving the base branch is safe and fast. Two criterion benches under
nativelink-util/benches/and three new fs_util tests. Full numbers + acceptance verdicts innativelink-util/BENCHMARKS.md.Performance (Apple M4 Max, APFS)
hardlink_directory_treeon shapes that mirror real SwiftCompile inputs:small_flatpcm_clustermedium_flatlarge_flatThe 64-cap on
download_to_directoryis performance-neutral (0.98x–1.01x) across all input counts on a single process — proves "no regression," which is the important security claim for that change.Caveat (worth knowing before merge)
The original PR claimed ~10x on the largest shapes. We see ~4x because
hardlink_directory_treedoes an O(N)set_readwrite_recursivechmod walk after the O(1) clonefile (the cache stores files 0o555/0o444 and actions need 0o755/0o644 to drop outputs). 4x is still a 440ms-per-action saving on the p95 SwiftCompile shape (~6 min/build across 814 such actions). Parallelizing or replacing the chmod walk is a known follow-up that should unlock the remaining 2-3x.Security tests
10/10 fs_util tests green, including the 3 new ones:
/etc/passwd)