Repository navigation
e2e: TestXoverStalledClientDoesNotHoldGroupLock covers the whole stall - #375
Merged
Merged
Conversation
Old (#274): after A's XOVER 1-40000 status line the test slept 700 ms "for A's reply to fill the socket", then measured B's ARTICLE store (bound 1.5 s). If A's session had not reached the records yet (slow or loaded machine), B took the lock first and passed whatever the code did with the lock. New: after the warm-up, wait until the daemon has no child, so that A's session is its only child; after A's 224, poll /proc/<pid>/syscall until the session sleeps in sendto/sendmsg/write (state S), then start B. Linux only now (skips elsewhere, and if /proc/<pid>/syscall can't be read); the code under test is the same on every system. The warm cache and the 1.5 s bound stay. Measured (12 cores): A blocked in send() 5-27 ms after the request, 128-142 ms with -race; B waited 0 s in all 5 + 5 runs. Mutants (scratch only), NVNewsgroup::printoverdb_stored: 1. flush_chunk() before guard.release() (each chunk written with the shared lock held): new test fails, 3/3: the second session waited 3.042975562s for the group lock while the first client stopped reading; ... (the old test fails too, with 2.31 s, as the stall had begun 0.7 s before B). 2. the same plus usleep(1 s) before the first chunk (a slow machine): the old test passes ("second session waited 0s"); the new one waits for the block (1.032 s) and fails with 2.99 s. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The previous version waited for /proc/<pid>/syscall to show the first session in send() and skipped where that file doesn't exist, FreeBSD included, where main still ran the test. Now the second client stores one uncached article after the other (each needs the group's exclusive lock) from the first client's 224 until the first session has ended. That session reads nothing and ends only when a send times out, ClientTimeout after it blocked, so the stores cover the whole stall on any system: with the lock held across the write one of them waits until the session ends. The session is found with daemonChildren (/proc or pgrep). Mutants (scratch builds, not committed), NVNewsgroup::printoverdb_stored: flush_chunk() before guard.release() fails 3/3 (waits 3.01-3.04 s); the same plus usleep(1 s) first, a slow machine, fails 3/3 (3.01-3.05 s), also with -race and with /proc hidden (pgrep). main passes, longest wait 10-90 ms (20-220 ms with -race). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
go-while
changed the title
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
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.
Old (#274): after A's XOVER 1-40000 status line the test slept 700 ms
"for A's reply to fill the socket", then measured one store by B
(bound 1.5 s). If A's session had not reached the records yet (slow or
loaded machine), B took the lock first and passed whatever the code did
with the lock.
New: after A's 224, B stores one uncached article after the other (each
store needs the group's exclusive lock) until A's session has ended. A
reads nothing, so its session ends only when a send times out,
ClientTimeout after it blocked: B's stores cover the whole stall, and
with the lock held across the write one of them waits until A's session
ends. The longest wait must stay within 1.5 s, as before. A's session is
found with
daemonChildren(/proc, or pgrep -P without /proc, from #369),so the test runs on FreeBSD and macOS too; it skips only where neither
exists. The stores are 2 ms apart (pacing only: at most about 1500 of
the group's 40000 articles in A's 3 s), and the test fails clearly if it
runs out of uncached articles.
An earlier head of this PR waited for /proc//syscall to show A's
session in send() and skipped elsewhere (Terra's finding: FreeBSD, a
required check, lost the test). A client-side check (A's receive queue
via FIONREAD stops growing) was tried as the fallback and rejected: it
only shows that A's TCP window closed, while the server still fills its
own send buffer, and mutant 2 below passed with it 3/3.
Measured (12 cores), 3 runs each:
Mutants (scratch only, not committed), NVNewsgroup::printoverdb_stored:
shared lock held.
the old test passed with it ("second session waited 0s").
Part of the review of weakened tests that go-while asked for. Shown to fail with the guarded defect (scratch mutant; not committed).
NEWS:
changes/fix-xover-stall-test-wait.md.🤖 Generated with Claude Code