git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH v2 00/12] Fix some reference-related races

From
Jeff King <peff@peff.net>
Date
Jun 20, 2013, 11:52 UTC
Message-ID
<20130620115221.GA773@sigill.intra.peff.net>
In-Reply-To
<51C2C6DD.90107@alum.mit.edu>
On Thu, Jun 20, 2013 at 11:09:49AM +0200, Michael Haggerty wrote:
Show 13 quoted lines
> If the packed-refs file is already locked by another process (and there
> is no reason why that cannot be, and there is only one attempt to
> acquire the lock), then repack_without_ref() emits an error and returns
> with an error code.  delete_ref() passes the error along, but doesn't
> restore the loose ref.
>
> [...]
>
> I think this problem would also be fixed by the locking scheme that I
> proposed earlier [1]: to acquire and hold the packed-refs lock across
> the modification of *both* files, and to rewrite the packed-refs file
> *before* deleting the loose-refs file (because rewriting the packed-refs
> file without the to-be-deleted reference is a logical NOP).

Yeah, I agree. You could also "roll back" the loose deletion, but I'd rather just try to do it atomically.

I don't think this increases lock contention, since delete_ref would need to lock the packed-refs file anyway. However, there is the related change that we should probably lock the packed-refs file before checking "is this ref in the packed-refs file?" in repack_without_ref. Which does increase lock contention, but is more correct.

We should also consider deadlock issues. If the order is always "acquire packed-refs lock, then acquire loose locks", we are fine. If this does loose-then-packed, we are also fine with the current code, as git-pack-refs does not prune the loose refs while under the packed-refs lock. But I seem to recall discussion of pruning them under the packed-refs lock, which would deadlock if repack_without_ref does loose-then-packed.

But I guess we do not actually block on locks, but rather just die (and release our locks), so deadlock is not an option for us.

-Peff
Previous: Michael Haggerty
Message 24 of 24 in “Fix some reference-related races”
  1. 00/12 Fix some reference-related racesMichael Haggerty, Jun 19, 2013
  2. 01/12 repack_without_ref(): split list curation and entry writingMichael Haggerty, Jun 19, 2013
  3. 02/12 pack_refs(): split creation of packed refs and entry writingMichael Haggerty, Jun 19, 2013
  4. 03/12 refs: wrap the packed refs cache in a level of indirectionMichael Haggerty, Jun 19, 2013
  5. 04/12 refs: implement simple transactions for the packed-refs fileMichael Haggerty, Jun 19, 2013
  6. Junio C HamanoJun 19, 2013
  7. Michael HaggertyJun 20, 2013
  8. Jeff KingJun 20, 2013
  9. Michael HaggertyJun 20, 2013
  10. Jeff KingJun 20, 2013
  11. Junio C HamanoJun 20, 2013
  12. Michael HaggertyJun 20, 2013
  13. Junio C HamanoJun 20, 2013
  14. 05/12 refs: manage lifetime of packed refs cache via reference countingMichael Haggerty, Jun 19, 2013
  15. 06/12 do_for_each_entry(): increment the packed refs cache refcountMichael Haggerty, Jun 19, 2013
  16. 07/12 packed_ref_cache: increment refcount when lockedMichael Haggerty, Jun 19, 2013
  17. 08/12 Extract a struct stat_data from cache_entryMichael Haggerty, Jun 19, 2013
  18. 09/12 add a stat_validity structMichael Haggerty, Jun 19, 2013
  19. 10/12 get_packed_ref_cache: reload packed-refs file when it changesMichael Haggerty, Jun 19, 2013
  20. 11/12 for_each_ref: load all loose refs before packed refsMichael Haggerty, Jun 19, 2013
  21. 12/12 refs: do not invalidate the packed-refs cache unnecessarilyMichael Haggerty, Jun 19, 2013
  22. Jeff KingJun 19, 2013
  23. Michael HaggertyJun 20, 2013
  24. Jeff KingJun 20, 2013

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.