git-annex/doc/bugs/cyclic_drop.mdwn
Joey Hess d3e1a3619f safer inannex checking
git-annex-shell inannex now returns always 0, 1, or 100 (the last when
it's unclear if content is currently in the index due to it currently being
moved or dropped).

(Actual locking code still not yet written.)
2011-11-09 18:33:15 -04:00

79 lines
3.1 KiB
Markdown

drop's verification that a remote still has content can fail
if the remote is also dropping the content at the same time. Each
repository checks that the other still has the content, and then both
drop it. Could also happen with larger cycles of repositories.
---
Fixing this requires locking. (Well, there are other ways, like moving the
content to a holding area when checking if it's safe to drop, but they
seem complicated, and would be hard to implement for move --from.)
Add per-content lock files. An exclusive lock is held on content when
it's in the process of being dropped, or moved. The lock is taken
nonblocking; if it cannot be obtained, something else is acting on the
content and git-annex should refuse to do anything.
Then when checking inannex, try to take a shared lock. Note that to avoid
deadlock, this must be a nonblocking lock. If it fails, the status of
the content is unknown, so inannex should fail. Note that this failure
needs to be distinguishable from "not in annex".
> Thinking about these lock files, this would be a lot more files,
> and would possibly break some assumptions that everything in
> `.git/annex/objects` is a key's content. (Or would need lots more
> directories to put the lock files elsewhere.) There would be more
> overhead to manage these and have them on disk.
>
> What if it just locked the actual content file? The obvious limitation
> is only content that was already inannex could be locked, but that
> happens to be exactly what's needed here; if content is not present,
> it's not going to get dropped or moved.
>
> Of course, if some consumer of a file locked it, then it could prevent it
> from being dropped or moved. This could be considered a bug, or a feature. :)
>
> However, this would mean that such a hypothetical consumer could also
> make inannex checks fail.
---
drop --from could also cycle. Locking should fix.
---
move --to can also be included in the cycle, since it can drop data.
Consider move to a remote that already has the content and
is at the same time doing a drop (or a move). The remote
verifies the content is present on the movee, and removes its copy.
The movee removes its copy.
So move --to needs to take the content lock on start. Then the inannex
will fail.
This is why it's important for inannex to fail in a way that is
distinguishable from "not in annex". Otherwise, move --to
would see the cycle as the remote not having content, and try to
redundantly send it, drop it locally, and still race.
--
move --from is similar. Consider a case where both the local and the remote
are doing a move --from. Both have the content, and confirm the other does,
via inannex checks. Then both run git-annex-shell dropkey, removing both
copies.
So move --from needs to take the content lock on start, so the inannex will
fail. NB: If the content is not locally present, don't take the lock.
---
Another cycle might be running move --to and move --from on the same file,
locally. The exclusivity of the content lock solves this, as only one can
run at a time.
---
Another cycle might involve move --from and drop, both run on the same
file, locally. Again, the exclusive lock solves this.