Re: [PATCH v3 0/3] Make update refs more atomic

Ronnie Sahlberg [off-list ref] writes:

refs.c:ref_transaction_commit() intermingles doing updates and checks with
actually applying changes to the refs in loops that abort on error.
This is done one ref at a time and means that if an error is detected that
will fail the operation partway through the list of refs to update we
will end up with some changes applied to disk and others not.

Without having transaction support from the filesystem, it is hard to
make an update that involves multiple refs to guarantee atomicity, but we
can do a somewhat better than we currently do.

These patches change the update and delete functions to use a three
call pattern of

1, lock
2, update, or flag for deletion
3, apply on disk  (rename() or unlink())

When a transaction is commited we first do all the locking, preparations
and most of the error checking before we actually start applying any changes
to the filesystem store.

This means that more of the error cases that will fail the commit
will trigger before we start doing any changes to the actual files.


This should make the changes of refs in refs_transaction_commit slightly
more atomic.

Hmph.  At least 5801 9350 and 9300 seem to fail with these three
queued on top of mh/ref-transaction series.