git-update-ref does not make a new reflog entry if the ref is not
actually being changed.
This can mean that `dgit build' might fail to find a good cache entry,
but also fail to generate one that push will accept.
In this case, generate a dummy reflog update. We make a spurious
child of the $dgitview, as that's straightforward. Giving it the same
tree and a parent means that human use of reflog commands will produce
empty results rather than lots of confusing output.
Signed-off-by: Ian Jackson <ijackson@chiark.greenend.org.uk>
Closes:#841100.
* split brain cache: Fix a wrong implicit reference to $_.
Closes:#841383.
Closes:#841100.
* split brain cache: Fix a wrong implicit reference to $_.
Closes:#841383.
+ * split brain cache: Make sure to write reflog entries for cache updates
+ even if the eventual tree (and therefore commit) is the same.
+ Otherwise, after updating dgit, the cache might have the right answer
+ but not be refreshed even by a build.
ensuredir ".git/logs/refs/dgit-intern";
my $makelogfh = new IO::File ".git/logs/refs/$splitbraincache", '>>'
or die $!;
ensuredir ".git/logs/refs/dgit-intern";
my $makelogfh = new IO::File ".git/logs/refs/$splitbraincache", '>>'
or die $!;
+
+ my $oldcache = git_get_ref "refs/$splitbraincache";
+ if ($oldcache eq $dgitview) {
+ my $tree = cmdoutput qw(git rev-parse), "$dgitview:";
+ # git update-ref doesn't always update, in this case. *sigh*
+ my $dummy = make_commit_text <<END;
+tree $tree
+parent $dgitview
+author Dgit <dgit\@example.com> 1000000000 +0000
+committer Dgit <dgit\@example.com> 1000000000 +0000
+
+Dummy commit - do not use
+END
+ runcmd @git, qw(update-ref -m), "dgit $our_version - dummy",
+ "refs/$splitbraincache", $dummy;
+ }
runcmd @git, qw(update-ref -m), $cachekey, "refs/$splitbraincache",
$dgitview;
runcmd @git, qw(update-ref -m), $cachekey, "refs/$splitbraincache",
$dgitview;