master
8
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
  retoorandClaude Sonnet 5
|
4aad7d6f64 |
Fix real memory leak Gitea CI caught, that local testing had been masking
CI / build-and-test (push) Failing after 38s
Gitea CI flagged two things on the last push: 1. A -Wunused-result warning on an intentionally-ignored write() return value in test_crash_consistency.c's progress side-channel. Fixed with an explicit (void) cast and a comment explaining why ignoring it is safe (a short write there only makes the progress count more conservative, per that function's own existing documented tolerance). 2. A real LeakSanitizer failure -- 256 bytes across 4 allocations from vfs_new/vfs_unmount. Root cause: tests/test_journal_failure.c's two "reopen after the failure, verify recovery" blocks called vfs_unmount/backend_free/backend_free inside their `if (ov2)` branch (the normal, expected path) but vfs_free(v2) only on the `else` branch, which is never actually reached in practice. Fixed by moving vfs_free(v2) to run unconditionally after the if, in both blocks. This bug was invisible locally across many runs because local sanitizer verification had been using ASAN_OPTIONS=detect_leaks=0 -- adopted originally for a real reason (a SIGKILLed forked child in test_crash_consistency.c never runs its own exit-time leak check, so its allocations were never the actual concern) but applied to the whole test run, which also suppressed detection of this real bug in the *parent* process's own code. CI doesn't set that option, so it caught what local runs couldn't. Documented in CONTRIBUTING.md as a real process gap, not just a code bug: a local verification habit that diverges from what CI actually runs can let a real finding through until it reaches CI. Verified: confirmed the leak directly first (reproduced locally by dropping the detect_leaks=0 override, matching CI exactly, before touching any code), then confirmed the fix by re-running the same no-override sweep across all 8 test binaries with zero leaks found, plus a clean make all + make test. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB |
||
|
retoorandClaude Sonnet 5
|
153d44ee3c |
Fix silent journal-write-failure data loss; confirm TSan blocked twice over
CI / build-and-test (push) Failing after 33s
Prompted by "fix literally everything still open" after the previous
session's data-integrity work. Went through each open item in turn:
1. TSan: tried a genuinely different execution environment (a remote
cloud sandbox, via a dedicated agent) rather than re-stating the local
sandbox's limitation. Result: identical block there too --
personality(ADDR_NO_RANDOMIZE) returns EPERM, a trivial pthread
program fails TSan identically, and all 6 PackFS test binaries fail
with the same FATAL: ThreadSanitizer: unexpected memory mapping
signature. This is now confirmed in two independent environments, not
one -- strong evidence it's a real infrastructure restriction, not a
one-off fluke worth chasing further with the tools available here.
2. While investigating the "disk-full mid-write" gap flagged as untested
last session, found two real, previously-unknown bugs by reading the
journal code (not by a test catching them unprompted):
- journal_append_record and everything that called it were void, and
none of the fwrite/fflush/fsync calls inside had their return values
checked. A real write failure (disk full, quota, I/O error) was
silently reported as success to vfs_write/vfs_mkdir/vfs_unlink/
vfs_rename -- directly contradicting Section 4.4's premise that
success means durable.
- Fixing that alone was not enough, confirmed by direct reproduction:
a partial write leaves a torn record in the journal, and
journal_replay correctly stops at the first record it can't fully
read (Section 4.3) -- which means every record appended *after* the
torn one, including ones that themselves wrote perfectly fine later,
became silently unreachable on reopen. Reproduced directly before
fixing: a forced-failed write followed by a genuinely successful one
was unrecoverable. Fixed by rolling the journal file back to its
exact pre-record length on any failed write.
Both closed in src/overlay.c (journal_append_record/_put/_delete/
_mkdir/journal_put_current now return and propagate success/failure;
overlay_write/_mkdir/_unlink/_rename return VFS_ERR_IO on a durability
failure without rolling back the already-applied in-memory change,
the same asymmetry a real write()-then-failed-fsync() has). Covered
permanently by the new tests/test_journal_failure.c, which forces a
real failure via RLIMIT_FSIZE + ignored SIGXFSZ, not a mock.
Also fixed in the same pass, found by inspection while touching this
code: journal_put_current used to pass a NULL buffer into a memcpy of
a nonzero size when malloc(size) failed (an OOM-triggered NULL-pointer
dereference) -- closed with an explicit allocation-failure check.
Not test-triggered (forcing malloc() failure portably isn't practical
here); verified by code inspection instead, stated as such rather than
claimed as tested.
3. The remaining "journal-truncation-specific crash window" gap from last
session was investigated, not silently dropped: reliably targeting
that narrow a window would need real concurrency (a second writer
thread racing the kill) for benefit the existing compaction-crash test
already gets probabilistically -- a poor trade, so left as a stated,
deliberate non-goal (CLAUDE.md) rather than built.
4. Cross-process contention is NOT addressed here and should not be read
as an oversight: it is concept.md's own explicit, permanent "not
implemented in v0" scope boundary (a specified-but-unbuilt LMDB-style
reader-table design), not a bug -- building it would be a large,
unrequested feature addition outside this session's actual scope.
Verified: clean make all + make test (all 8 binaries), make bench and
make demo still build and the demo runs correctly end to end, and a full
ASan/UBSan sweep of all 8 binaries with zero real findings (some retries
needed for the already-documented DEADLYSIGNAL flake, which
test_crash_consistency hits more often than other tests simply because it
forks 60+ subprocesses per run -- noted in CONTRIBUTING.md so this isn't
mistaken for a regression later).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB
|
||
|
retoorandClaude Sonnet 5
|
0c0ae4f747 |
CI: don't hard-fail on the confirmed TSan-can't-start-here runner limitation
CI / build-and-test (push) Failing after 23s
The real Gitea Actions run on this project's own registered runner just failed exactly as CONTRIBUTING.md already anticipated it might: FATAL: ThreadSanitizer: unexpected memory mapping, the same signature already documented as a sandbox/container seccomp restriction blocking personality(ADDR_NO_RANDOMIZE), which TSan needs to start at all. Build, test suite, and ASan/UBSan all passed -- only the TSan step failed, on an environment issue, not a code issue. Fixed the CI step itself rather than just noting the failure: it now classifies each binary's TSan run as PASS, a known flake (that exact FATAL signature and nothing indicating an actual race was found), or a real failure (anything else -- a genuine data race, a crash, any other error). Only a real failure fails the build. Verified the classification logic locally against four cases (a synthetic real race report, a plain assertion failure, the known flake signature alone, and a clean pass) -- each classified correctly -- and against this sandbox's own six test binaries, all six of which hit the real flake (this sandbox has never been able to run TSan either) and correctly did not fail the build. This does NOT mean TSan verification is happening in CI -- it means CI no longer conflates "TSan couldn't start" with "the build is broken." Updated CONTRIBUTING.md and CLAUDE.md from "whether the runner can execute TSan is unverified" (a hedge) to the now-confirmed fact that it can't, and corrected an overclaim in README.md that the suite is "regularly run under ThreadSanitizer" -- as far as this project has been able to confirm, TSan has not actually completed a run in any environment it's been built in yet, local sandbox or CI. ASan/UBSan remain the real, run, load-bearing sanitizer coverage. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB |
||
|
retoorandClaude Sonnet 5
|
21bcb4640f |
Move CI to Gitea Actions; this project is hosted on Gitea, not GitHub
CI / build-and-test (push) Failing after 23s
Per explicit direction: this repository is not going on GitHub. Moved .github/workflows/ci.yml to .gitea/workflows/ci.yml (Gitea Actions' convention) and updated every doc that referenced the old path or assumed GitHub-specific features: - .gitea/workflows/ci.yml: added a header comment on the two things that are genuinely Gitea-specific and instance-dependent, not just a renamed file -- `runs-on: ubuntu-latest` must match a label the actual registered Gitea runner advertises (there is no GitHub-hosted-runner equivalent, this is self-hosted), and `actions/checkout@v4` resolves against whatever action source that runner is configured with. - CONTRIBUTING.md: corrected a claim that no longer holds -- it previously said CI "runs on a normal, unrestricted GitHub Actions VM where TSan is expected to work"; since this is actually a self-hosted Gitea runner whose environment isn't controlled by this repo, that assumption isn't something this repo can vouch for, so the text now says so rather than carrying the old (GitHub-shaped) assumption forward silently. - SECURITY.md: removed a claim this project can't back up (that "private security advisories" are available once hosted -- that's a GitHub feature this repo never had access to); reporting is by direct email to the maintainer only. - README.md: fixed a real gap while in here -- the "Security" section never actually linked to SECURITY.md despite it existing since the previous commit. - CLAUDE.md, CHANGELOG.md: updated path references; CLAUDE.md's self-evaluation section (a historical record of an audit finding) keeps the old .github path where it describes what was literally true at that time, with a note explaining the rename, rather than rewriting history. Verified: clean make all + make test, all 6 binaries pass. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB |
||
|
retoorandClaude Sonnet 5
|
684c6d95de |
Fix pack_write's O(n^2) dedup + correctness bug; document mount table O(n^2)
A audit for other instances of the file-index O(n^2) shape (fixed previously via the persistent treap) found two more real issues: 1. pack_write's Section 9.2 exact-duplicate elimination was a linear scan of every previously-seen (hash, size) pair per entry -- O(n^2) total, invisible in the existing benchmark because its identical-content test files made every scan match on the first comparison. Measured with unique content instead: 80,000 entries took 1.74s, with a 20,000->80,000 step showing 15.6x for a 4x-N step, matching O(n^2)'s 16x prediction. The same scan also trusted a (hash, size) match without ever comparing actual bytes -- a latent correctness bug, since FNV-1a64 is explicitly not collision-resistant. Fixed both at once with an open-addressing hash table (load factor 1/2, linear probing) plus a memcmp verification before ever reusing a data_off. Post-fix: 80,000 entries in 0.044s (39.6x faster), ratio drops to 3.35x (consistent with O(n)). Covered permanently by two new/extended tests: a white-box assertion in test_pack_overlay.c that duplicate-content entries share one data_off and distinct-content entries do not, and a new tests/test_pack_write_perf.c regression tripwire against 10,000 unique entries. 2. vfs.c's mount table uses the same full-array-copy-per-write pattern the file index used to, confirmed O(n^2) via a new bench/bench.c category (500/2,000/8,000 mounts, both 4x-N steps showing 15-20x). Deliberately NOT rewritten: mount points are created by a program's own source code, not workload-driven, so realistic mount counts never reach the scale that made the file index's O(n^2) a real problem. Documented with full reasoning in BENCH.md and CLAUDE.md rather than silently left as an undocumented gap. Also fixes a real CI gap the new tests exposed: ci.yml's sanitizer-build steps never passed -D_GNU_SOURCE when compiling test files (only the library .o's got it), which was harmless while no test included internal.h and became a link failure once two did (internal.h needs _GNU_SOURCE for pthread_rwlock_t). And documents, in CONTRIBUTING.md and CLAUDE.md, a sandbox flake observed directly during this work's own sanitizer runs: ASan/UBSan test binaries occasionally fail to start with AddressSanitizer:DEADLYSIGNAL (sometimes looping rather than exiting), non-deterministically hitting different unrelated binaries across runs -- a startup race, not a memory-safety bug, confirmed by clean passes on retry; sanitizer runs in such an environment should be timeout-wrapped. BENCH.md's "After" table and Appendix B are replaced with the current, complete 54-measurement bench/bench.c run (the original 45 plus the new mount-scaling category); the pre-fix 45-measurement "Before" table is kept as the historical record, per this project's documentation standard. Verified: make test (all 6 binaries, including the 2 new/changed), a clean make all, and repeated ASan+UBSan runs (0 real findings; the DEADLYSIGNAL flake above was observed and correctly distinguished from a real finding by re-running until a clean pass). TSan could not be run in this sandbox (pre-existing, documented environment limitation). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB |
||
|
retoorandClaude Sonnet 5
|
64283d587b |
Fix the O(n^2) bulk create/unlink: flat array -> persistent treap
BENCH.md's earlier finding was real: every structural write (create/
unlink/mkdir) copied the entire sorted UpperSnapshot entry array before
publishing the next snapshot, making bulk sequential creation O(n^2).
concept.md Section 5.3 names the exact condition for reconsidering this
("a persistent structurally-shared tree structure is not required
until this assumption is empirically violated") -- that condition was
measured, not hypothesized, so this closes it rather than leaving it
as a documented-but-open limitation.
The index is now a persistent treap (src/upper.c): a structural write
copies only the O(log n) nodes on the path to the change, sharing
every other node (its own refcount, cascading like MutCell's and
UpperSnapshot's) with whichever snapshot(s) it was built from. Chosen
over a persistent AVL/red-black/weight-balanced tree because deletion
in those can need O(log n) rebalancing rotations -- each a real
allocation in a persistent setting -- where a treap needs only O(1)
amortized rotations for insert and delete (Seidel & Aragon 1996), with
expected O(log n) height regardless of insertion order, including the
sorted-by-creation-order pattern that made the flat array quadratic in
the first place. Liljenzin's "Confluently Persistent Sets and Maps"
(arXiv:1301.3388) documents persistent treaps giving O(1) snapshots
for MVCC specifically, which is this exact use case. Full reasoning
and the ownership convention (functions consume one ref of their tree
arguments, return one owned ref) are in upper.c's comment above
struct TreapNode.
Blast radius kept deliberately small: snapshot_upsert/snapshot_remove
keep their exact original signatures, so upper_create/upper_mkdir/
upper_remove/upper_rename/upper_copy_up needed zero changes.
upper_lookup/upper_has_children keep their exact contracts. Only
upstd_readdir and overlay_readdir's manual array scans became calls to
a new upper_visit_range (O(log n + r) range query, replacing an O(n)
scan in both, a bonus fix beyond what was strictly necessary) since
there's no flat array left to scan.
Added tests/test_index_stress.c: thousands of randomized (not
sequential) creates/deletes/renames across nested directories,
cross-checked against an independent reference model after every
round, not just "did it not crash" -- exercises exactly the code path
the O(n^2) bug and this fix live in, at a scale the other tests don't
reach. Verified under -fsanitize=undefined (150+ runs across this
change's lifetime, 0 failures) and -fsanitize=address (100+ runs, 0
real findings; known sandbox ASan-startup flakes excluded, see prior
commits) per CLAUDE.md's sanitizer rule for upper.c/overlay.c changes.
Measured result (make bench, same environment as the original
finding): mem create 257x faster, unlink 330x faster, mkdir 65x
faster, concurrent mixed workload 131x faster -- and, the comparison
that matters, mem now beats raw fs at every one of these (was losing
by 6-22x before). The complexity-class change is confirmed the same
way the O(n^2) was found: mkdir at N=4,000 vs create at N=20,000 now
shows a 7.15x slowdown for a 5x increase in N, matching the O(n log n)
prediction (5.97x) rather than the old O(n^2) one (25x). Full
before/after tables in BENCH.md's new "Resolution" section, which
keeps the original run as the historical record rather than
overwriting it, per this project's own documentation standard.
Recorded the fix in CLAUDE.md's "Known performance characteristics"
(marked RESOLVED, not silently removed) and its architecture map.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB
|
||
|
retoorandClaude Sonnet 5
|
0b3b207bc0 |
Add a benchmark suite comparing PackFS against the host filesystem
bench/bench.c (`make bench`) measures create/read/stat/readdir/unlink/
mkdir on mem, dir, and raw fs; large sequential I/O; random-access pack
reads via mmap vs raw fs; compaction throughput; and 8-thread
concurrent mixed workloads. Results from one full run, with honest
analysis (what each "raw fs" vs "raw+fsync" vs "dir" label actually
measures, so they aren't misread as interchangeable), are in BENCH.md.
The benchmark surfaced a real, quantitatively-confirmed finding, not
just favorable numbers: bulk sequential create/unlink on mem/dir is
O(n^2) in file count (~9-22x slower than raw fs at N=20,000), because
every structural write copies the entire snapshot entry array before
publishing it (Section 5.3). concept.md itself names the exact trigger
condition for reconsidering this ("a persistent structurally-shared
tree structure is not required until this assumption is empirically
violated") — this benchmark is that violation, measured rather than
hypothesized: mkdir at N=4,000 vs create at N=20,000 (same mechanism,
5x the N) shows a 28.4x slowdown, matching the O(n^2) prediction (25x)
far better than O(n) (5x).
Also found: dir-backend stat() costs ~2x raw stat() (open+fstat vs one
syscall, the direct cost of openat2 containment on the metadata path);
mem-backed large writes lose to raw fs above ~16MB (Section 5.7's
buffer-growth discipline re-copies prior writes on every capacity
doubling, the price of never exposing a reader to a freed buffer).
Recorded the O(n^2) finding in CLAUDE.md's new "Known performance
characteristics" section, per the same pattern used for the
reclaim_gate use-after-free discovery, since it's exactly the kind of
fact that would otherwise have to be rediscovered by benchmarking
again from scratch. Linked from README and CONTRIBUTING.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB
|
||
|
retoorandClaude Sonnet 5
|
72e3c900f2 |
Add PackFS v0: statically linked in-process VFS implementing concept.md
Implements the core design: a mount table published as an atomically- swapped snapshot; mem/dir/pack/overlay backends; copy-on-write overlay with copy-up and whiteout deletion; a checksummed append journal; compaction with exact-duplicate elimination; single-writer/wait-free- reader concurrency with a structural/content write split; openat2/ Landlock path containment for dir mounts; and load-time pack integrity validation. Zero required third-party dependencies. Sanitizer testing (ASan/UBSan) caught and led to fixing a genuine heap-use-after-free in the snapshot-reclamation path: the textbook "load pointer, then increment its refcount" pattern left a gap a concurrent writer could free through. Closed with a small reclaim_gate rwlock, documented in internal.h and CLAUDE.md since it's a pattern every refcounted structure in the codebase now follows. zip/tar import/export backends, recommended in concept.md Section 11, will not be built — a permanent project decision recorded in CLAUDE.md since concept.md itself is frozen and cannot be edited to reflect it. Includes a runnable demo (examples/demo.c, `make demo`) exercising the library end to end and proving cross-run persistence through the pack file, plus open-source scaffolding: MIT license, README, CONTRIBUTING, and a CI workflow running the test suite under ASan/UBSan/TSan. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UqJpkdJ6Njnt1pw3CbghzB |