Re: Bug report: Strange behavior with `git gc` and `reference-transaction` hook

On Wed, Nov 17, 2021 at 4:42 PM Waleed Khan [off-list ref] wrote:

Hi all,

I'm seeing unusual behavior on Git built from source at
cd3e606211bb1cf8bc57f7d76bab98cc17a150bc, but which appears to extend
back to Git v2.29.

This is a repro script:

#!/bin/sh

temp_dir=$(mktemp -d)
echo "git dir is: $temp_dir"
mkdir -p "$temp_dir"
trap "rm -rf $temp_dir" EXIT

cd "$temp_dir" || exit 1
git init -q
date='Thu, 07 Apr 2005 22:13:13 +0200'
GIT_AUTHOR_DATE="$date" GIT_COMMITTER_DATE="$date" git commit -q
--allow-empty -m 'Initial commit'

mkdir -p '.git/hooks'
cat >.git/hooks/reference-transaction <<'EOT'
#!/bin/sh

[[ "$1" != 'committed' ]] && exit 0
echo 'New reference-transaction invocation'

while read old new ref; do
  echo "  old: $old, new: $new, ref: $ref"
done
EOT
chmod +x .git/hooks/reference-transaction

git gc --prune=now -q
git show-ref refs/heads/master

And this is the output it produces:

git dir is: /var/folders/gn/gdp9z_g968b9nx7c9lvgy8y00000gp/T/tmp.b3Jc6qnb
New reference-transaction invocation
  old: 0000000000000000000000000000000000000000, new:
e197d18c017d4038418be8b1cd38f4503e289165, ref: refs/heads/master
New reference-transaction invocation
  old: e197d18c017d4038418be8b1cd38f4503e289165, new:
0000000000000000000000000000000000000000, ref: refs/heads/master
e197d18c017d4038418be8b1cd38f4503e289165 refs/heads/master

These hooks are invoked a few milliseconds one after another.

There are two built-in "ref backends" that Git uses out of the box: A
packed backend, which manages the "packed-refs" file, and a loose
backend, which manages other files under "$GIT_DIR/refs". What you're
seeing is a reference transaction for the packed backend which is
adding the new value for "refs/heads/master" to "packed-refs" (packed
backend reference-transactions rarely, if ever, include an actual old
hash, as far as I can tell), and then a second reference transaction
for the loose backend to delete the loose "refs/heads/master" file,
now that it's packed.

The expected behavior would be that the latest reference transaction
hook refers to the state of the references on disk. That is, either
`master` should point to 0 (be deleted), or it should have said that
`master` pointed to `e197d1`.

But if we actually examine `master`, it's set to `e197d1`, just as you
would expect. The GC should have been a no-op overall.

One of the subtasks of "git gc" is "git pack-refs". If you inspect in
more detail, I suspect you'll find that "refs/heads/master" was loose
before "git gc" ran (as in, there was a file
"$GIT_DIR/refs/heads/master") and "packed-refs" either didn't have a
"refs/heads/master" entry or had a different hash. (Loose refs always
"win" over packed, since ref updates only write loose refs.)

Hope this helps,
Bryan Turner

Best,
Waleed