commit 8b23132d5874e1d4245e9cab4c279395f56328c3
author Martin Fick <mfick@nvidia.com> Mon Sep 15 14:48:51 2025 -0700
committer Matthias Sohn <matthias.sohn@sap.com> Thu Nov 06 15:41:43 2025 +0100
tree 9527000ed34457d27f4f8ae130cea708f58c2e57
parent 39409b6c04836c0c71dbef80b0f393d2f26df843

Avoid reading packed-refs concurrently sometimes

If the packed-refs file has not changed since a thread has started
reading and parsing it, do not bother to re-read and re-parse it from
another thread as this will just create extra load and likely slow down
the current reading/parsing thread only to in theory return the same
result. Of course, jgit uses timestamps to identify that a file has not
changed since the current thread has started refreshing it, without
rereading it, so this optimization only comes into play if some form of
trusting the packed-refs stats is enabled, otherwise the behavior before
and after should be similar.

This duplicate parsing does not synchronize the refresh, and it still
allows for more then one thread to read/parse the packed-refs file in
parallel, but only when the file has changed since the current refresh
started. This change does however synchronize creating a
PackedRefsRefresher which for just long enough to ensure that only one
is created when only one is needed. This change should result in better
performance all while maintaining the same data consistency with respect
to requests as before.

Change-Id: Ic486e0975a67c6ea493493083411255953fba72e
Signed-off-by: Martin Fick <mfick@nvidia.com>