"missing referenced chk root" while merging gcc into gcc-linaro

[Michael Hope]

I get the attached crash when merging the GCC 4.7 release branch into the Linaro GCC release branch. This causes significant pain each month as we need to do a manual, error prone diff/patch instead.

To reproduce:
 * mkdir gcc; cd gcc
 * bzr init-repo .
 * bzr branch lp:gcc-linaro/4.7
 * bzr branch --standalone lp:gcc/4.7 gcc-4.7
 * bzr branch --hardlink 4.7 merge-from-fsf
 * cd merge-from-fsf
 * bzr merge ../gcc-4.7

bzr also fails with the shared repo unless you add the --standalone.

This process works fine on 4.6 and 4.5.

[Michael Hope (michaelh1)]

jelmer talked with ams about this on IRC: http://irclogs.ubuntu.com/2012/03/14/%23bzr.txt

[Joey Stanford (joey)]

<jelmer> there is invalid data in one of the two branches caused by an old bzr-svn bug

<jelmer> joey: if you don't care about the actual history of gcc-linaro, then the simplest thing to do would be to take the last revision from lp:gcc that was merged into lp:gcc-linaro, "rsync --delete -avz" over the existing contents from lp:gcc-linaro and commit it in a single commit

<joey> hmm I'd have to talk to michael hope about that. It may be possible for us to start a new series as part of normal ops and then fix it that way

[Joey Stanford (joey)]

<joey> jelmer: gotcha. If we can't resync because we need history, is there another method we can use?

<jelmer> joey: I think you would have to replay the history and make sure text revisions are recorded consistently with the way they are in lp:gcc

[Jelmer Vernooij (jelmer)]

So, it turns out we do already have *a* way of fixing broken repositories like this by replaying all data and recording the text revisions properly. This is fairly well hidden, and can be done by running "bzr reconcile --canonicalize-chks" in the affected repository(s).

Unfortunately this command is quite slow - on my machine it took 4 hours for the gcc repo just to get to rev 100, which is only a tiny fraction of the 115544 revisions in the repository overall.

[Martin Packman (gz)]

Seems like 'quite slow' is something of an understatement. Do we want a followup bug for making it less pathological?

[Vincent Ladeuil (vila)]

> Seems like 'quite slow' is something of an understatement. Do we want a followup bug for making it less pathological?

Yes.

[Jelmer Vernooij (jelmer)]

In addition, it looks like 'bzr reconcile --canonicalize-chks' doesn't seem to fix all instances of this problem. Just confirmed this with glyph for the twisted repository.

[Michael Hope (michaelh1)]

I'm afraid the current fix isn't usable if it's that slow. Jelmer's numbers suggest the repair would take six months.