summaryrefslogtreecommitdiff
path: root/doc/bugs/git_mail_notification_race.mdwn
blob: ebe158fb9c728c567c1820447dde0b811774fafb (plain)

I was suprised to receive two mails from ikiwiki about one web edit:

 1   F Oct 30 To joey+ikiwiki update of ikiwiki's plugins/contrib/gallery.mdwn by http://arpitjain11.myopenid.com/
 1   F Oct 30 To joey+ikiwiki update of ikiwiki's plugins/contrib/gallery.mdwn by http://arpitjain11.myopenid.com/

The first of these had the correct diff for the changes made by the web edit (00259020061577316895370ee04cf00b634db98a).

But the second had a diff for modifications I made to ikiwiki code around the same time (2a6e353c205a6c2c8b8e2eaf85fe9c585c1af0cd).

I'm now fairly sure a race is involved. Note that the name of the modified page is correct, while the diff was not. The git rcs_commit code first gets the list of modified page(s) and then gets the diff, and apparently my other commit happened in the meantime, so HEAD changed.

Seems that the rcs_notify for git assumes that HEAD is the commit that triggered the ikiwiki run, and that it won't change while the function is running. Both assumptions are bad. Git does no locking and another commit can come in at any time.

Notice that the equivilant code for subversion is careful to look at $REV. git's post-update hook doesn't have an equivilant of $REV. Its post-receive hook does, but I'm not sure if that's an appropriate hook for ikiwiki to be using for this. Switching existing wikis to use a different hook would be tricky..

I've avoided part of the race; now the code does:

  1. Get commit info for HEAD commit.
  2. Extract file list from that.
  3. Extract sha1 of commit from it too.
  4. git diff sha1^ sha1

Still seems like there can be a race if another commit comes in before step #1. In this case, two post-receive hooks could run at the same time, and both see the same sha1 as HEAD, so two diffs might be sent for second commit and no diff for the first commit.

Ikiwiki's own locking prevents this from happenning if both commits are web edits. At least one of the two commits has to be a non-web commit.