bzr crashed with ShortReadvError in _seek_and_read()

As seen in the ML thread, this is caused by an empty pack file.

The attached script will help restore the *repository* in a working state (by getting rid of the reference to the empty pack file) but will not fix any branch that refer to it (nor restore its content which is lost for good :-/).

Fixing the branch itself can be achieved while keeping the changes in the working tree and committing again which will create a new pack file.

To get there:
- create a branch based on a valid revision,
- copy the 'last-revision' file in '.bzr/branch' from the good branch to the broken one
- 'bzr st' and 'bzr diff' should show you the uncommitted changes

Rescue script that should be integrated into bzr with proper tests.

Thanks, the script seems to have done as advertised, but I can't seem to create a new branch. bzr branch foo bar -r N always returns the following error:
bzr: ERROR: Revision {<email address hidden>} not present in "CHKInventoryRepository('file:///path/to/repository/.bzr/repository/')".

The other problem is that I don't know which the last working revision is.

Right, finding the last working revision is the tricky bit, try searching your '.bzr.log' file for messages like:

  201.324 Committed revid <email address hidden> as revno 60.

You should find one for '<email address hidden>' which has been lost (I dunno if you edited this one to inject '<email address hidden>' but the date '20101204' seem to match your empty pack file so it's probably the lost revision), and a previous one which should be working.

Once you have that one you can use it as -rrevid:<the_good_one> in your branch command.

Hmmm, I don't seem to have a .bzr.log file anywhere I can see (not ~/, not repository/, not /var/log/, not ~/.bazaar/). Should I have enabled a switch somewhere to turn on the logging?

'bzr version' should tell you, on unices it's $HOME/.bzr.log

There is no switch to activate it, but you can redefine it by setting BZR_LOG... did you do that ?

Its set to ~/.bzr.log, but that file doesn't exist. Could this be another manifestation of the pack problem?

Ultimately, its not a catastrophe as the code is all there, but I would like to get to the bottom of it so it doesn't happen again.

Nope, no relation at all with the pack problem (well, no obvious one that is).
Even issuing 'bzr version' should create it, so you have *another* problem here... pretty weird though, we've never had report of the .bzr.log file not being created, are you sure you're not tricked by using some utility that doesn't reveal the hidden files (.bzr.log starts with a dot to be considered an hidden file and not being shown by default).

Doing:

  cd
  ls -al

should reveal it.

Otherwise please post the output of 'bzr version'.

So it seems ~/.bzr.log was there. Not quite sure how I missed it (apologies for being a fool).

Anyway, I was having trouble with getting a sensible internal state for the code using the suggested method, so I branched the tree to an old commit and then replaced the .bzr directory in my working tree with the .bzr from the new branch. A few commits later and it all seems to be working fine.

Thanks for all the help. Is the recommendation then that ext4 should not be used?

Ha good thing it was there and that you're now out of trouble.

No we don't recommend *against* ext4, this kind of issue have only been reported a few times and mostly on irc, so it's still unclear which kind of crash can lead to such situations. I use ext4 myself.

I'll keep this bug open to track the integration of the script with proper tests into bzr-core.

There's a different case possible here, which is in bug 413430, where the pack file is fine, but the index file is truncated. (Probably ext4 again.)