Repository navigation
Human action 337
Procedure for issue #337. The issue is the record of ownership, status, and completion evidence; this page is the execution instruction.
Owner role: the repository owner, or anyone willing to post to the Git mailing list under their own name. An agent does not send outward posts.
Why: causet's generated-history corpus (#308, PR #333) generated a commit
whose committer date overflows, 99999999999999999999. On Git 2.56.0, git gc
then spun at 100% CPU for 8.5 hours, until it was killed, in
write_commit_graph → compute_reachable_generation_numbers. Git reads such
commits (git fsck reports badDateOverflow, but every command still reads
them), so a repository holding one can hang every git gc or git maintenance
run. causet now avoids generating that date on a committer line and stops any
process after 5 minutes, so nothing in causet waits on this. Reporting it helps
Git, which is still causet's substrate.
- A POSIX shell with
git,mktempandtimeout(Linux, or Git Bash on Windows), and the newest Git release you can install. Check withgit --version. - A mail client that can send plain-text mail (no HTML). The Git list
rejects HTML mail.
git send-emailworks too, but isn't needed for a bug report. - No subscription is needed to post to git@vger.kernel.org.
-
Check for an existing report. Search the list archive for
commit-graph generation overflowandcompute_reachable_generation_numbersat https://lore.kernel.org/git/. If a report or a fix already covers this, link it on #337 and skip to Completion. -
Run the reproduction on the newest Git you have. Save this as
git-hang-repro.shin any directory and runsh git-hang-repro.sh. It works in a temporary directory and removes it.#!/bin/sh # git commit-graph write loops forever on a committer date that overflows set -e dir=$(mktemp -d) git init -q --bare "$dir/repo.git" cd "$dir/repo.git" tree=$(git hash-object -t tree -w --stdin </dev/null) commit() { # $1 = parent line or empty, $2 = committer date printf 'tree %s\n%sauthor A <a@example.com> 1700000000 +0000\ncommitter C <c@example.com> %s +0000\n\nm\n' \ "$tree" "$1" "$2" | git hash-object -t commit -w --literally --stdin } c1=$(commit "" 99999999999999999999) c2=$(commit "parent $c1 " 1700000001) c3=$(commit "parent $c2 " 1700000002) git update-ref refs/heads/main "$c3" git --version if timeout 20 git commit-graph write --reachable; then echo "commit-graph write finished" else echo "commit-graph write did not finish in 20 s (exit $?)" fi rm -rf "$dir"
Expected on an affected Git:
git version 2.56.0 commit-graph write did not finish in 20 s (exit 124)If it prints
commit-graph write finished, the bug is fixed in that version: record the version on #337 and skip to Completion. -
Send the report. Send a plain-text mail to
git@vger.kernel.org. Replace<VERSION>with the version from step 2 and<OS>with your system:Subject: commit-graph write never ends on a commit whose committer date overflows Hi, `git commit-graph write --reachable`, and so `git gc`, loops forever when a commit's committer date overflows timestamp_t and the commit has at least two generations of descendants. It reproduces on git <VERSION> (<OS>): <paste the script from step 2> The script's output: <paste the output of step 2> A stack sampled three times during a hung `git gc` (2.56.0) was the same each time: commit_graph_data_at get_generation_from_graph_data compute_reachable_generation_numbers write_commit_graph write_commit_graph_reachable cmd_gc With only one descendant there is no hang, and neither a date of 2^63 nor an overflowing author date causes one. I'd guess the corrected commit date computed from the saturated date wraps, so a generation never leaves GENERATION_NUMBER_ZERO and the commit is pushed again forever, but I haven't confirmed that. `git fsck` reports the commit as badDateOverflow, but such a commit can arrive by fetch from a repository that doesn't check, and then every gc or maintenance run hangs. We found this with generated test histories in causet (https://github.com/jwh3times/causet/issues/337). Thanks, <your name>
- Pass: within about an hour, the mail appears in https://lore.kernel.org/git/ (search for the subject). Record its archive link on #337.
- Not yet: if it hasn't appeared after a day, check that it was sent as plain text and that your mail server didn't reject it, then send it again.
- If the list rejects the mail, the bounce names the reason. HTML parts and attachments are the usual ones. Fix it and send again.
- If a reply asks for more detail, the copy in causet's checkout is
test-support/generated-histories.js(ident); the corpus history that found it was index 13 modulo 16 of seedcauset-n1before the change. The reproduction above is self-contained. - Nothing is created outside the script's temporary directory, so there is nothing to clean up.
On #337, record:
- the reproduction's output and Git version;
- the archive link to the report, or to the existing report or fix that covers it;
- any reply that changes what causet should do, for example the Git version that fixes it, after which the corpus could draw the date for committers again.
Then remove human-action-required, close #337, set its board Status to Done,
and mark this page's row on Human TODO Done with the link.