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

[PATCH 0/16] enabling GIT_REF_PARANOIA by default

From
Jeff King <peff@peff.net>
Date
Sep 24, 2021, 18:30 UTC
Message-ID
<YU4ZOF9+ubmoItmK@coredump.intra.peff.net>

I recently ran into a situation where dealing with a corrupted repository was more confusing than necessary, because Git by default ignores corrupted refs in many commands.

A while ago we introduced GIT_REF_PARANOIA, which works by including broken refs in iteration, which then typically causes later operations to fail (e.g., during repacking, you'd prefer to barf loudly when trying to access the missing object rather than incorrectly assume the objects from the broken ref aren't reachable).

I think this is a better default for Git to have in general, not just for a few select operations (we turn it on by default for pruning and some repacks). We shouldn't see corruptions in general, and complaining loudly when we do is the safest option. The reason we held back when the knob was introduced was mostly out of deference to the historical behavior.

So this series started as a patch to just flip that default, but I found some interesting things:

 - there are a couple of tests that get confused. IMHO this is
   vindicating the idea of flipping the default, beacuse in each case
   these tests were poorly written (either corruptions they didn't
   realize they had, or doing questionable operations on an incomplete
   set of refs)
 - the existing GIT_REF_PARANOIA is over-eager to complain about
   dangling symrefs, even though they're perfectly fine
 - as usual, there was some obvious cleanup along the way. ;)

Even if you don't buy the argument that we should flip the default, I think everything up through patch 11 is a worthwhile cleanup on its own.

Note that this conflicts with jt/no-abuse-alternate-odb-for-submodules, since it is touching the innards of DO_FOR_EACH_REF_INCLUDE_BROKEN, too. I left a note on that series about how I think that could be reconciled (i.e., the conflict is just around how the code is written, and not inherent to the goals).

In the end I left GIT_REF_PARANOIA as a knob, just defaulting to "1". I think it's possibly useful as an escape hatch when dealing with a corrupt repo. But we _could_ go all the way and basically drop DO_FOR_EACH_REF_INCLUDE_BROKEN's do-we-have-the-object check entirely. That would totally sever the relationship between the ref store and the object store, which would make things conceptually a lot simpler (and I saw was discussed in some of those earlier threads).

Just a breakdown of the series:
  [01/16]: t7900: clean up some more broken refs
  [02/16]: t5516: don't use HEAD ref for invalid ref-deletion tests
  [03/16]: t5600: provide detached HEAD for corruption failures
  [04/16]: t5312: drop "verbose" helper
  [05/16]: t5312: create bogus ref as necessary
  [06/16]: t5312: test non-destructive repack
  [07/16]: t5312: be more assertive about command failure
     Test cleanups. Necessary for the default flip, but I think each
     stands on its own.
  [08/16]: refs-internal.h: move DO_FOR_EACH_* flags next to each other
  [09/16]: refs-internal.h: reorganize DO_FOR_EACH_* flag documentation
     Cleanup of existing features.
  [10/16]: refs: add DO_FOR_EACH_OMIT_DANGLING_SYMREFS flag
  [11/16]: refs: omit dangling symrefs when using GIT_REF_PARANOIA
     Fixing the current over-eager behavior of GIT_REF_PARANOIA.
  [12/16]: refs: turn on GIT_REF_PARANOIA by default
     The actual flip.
  [13/16]: repack, prune: drop GIT_REF_PARANOIA settings
  [14/16]: ref-filter: stop setting FILTER_REFS_INCLUDE_BROKEN
  [15/16]: ref-filter: drop broken-ref code entirely
  [16/16]: refs: drop "broken" flag from for_each_fullref_in()
     Some small cleanups we can do as a result.
 Documentation/git.txt         | 19 ++++++------
 builtin/branch.c              |  2 +-
 builtin/for-each-ref.c        |  2 +-
 builtin/prune.c               |  1 -
 builtin/repack.c              |  3 --
 builtin/rev-parse.c           |  4 +--
 cache.h                       |  8 -----
 environment.c                 |  1 -
 ls-refs.c                     |  2 +-
 ref-filter.c                  | 22 ++++++--------
 ref-filter.h                  |  1 -
 refs.c                        | 42 +++++++++++++-------------
 refs.h                        |  9 ++----
 refs/files-backend.c          |  5 ++++
 refs/refs-internal.h          | 56 ++++++++++++++++++++++-------------
 revision.c                    |  2 +-
 t/t1430-bad-ref-name.sh       |  2 +-
 t/t5312-prune-corruption.sh   | 48 ++++++++++++++++++++++--------
 t/t5516-fetch-push.sh         | 19 ++++++------
 t/t5600-clone-fail-cleanup.sh |  4 ++-
 t/t7900-maintenance.sh        |  6 +++-
 21 files changed, 142 insertions(+), 116 deletions(-)
-Peff
Next: Jeff King
Message 1 of 22 in “enabling GIT_REF_PARANOIA by default”
  1. 0/16 enabling GIT_REF_PARANOIA by defaultJeff King, Sep 24, 2021
  2. 01/16 t7900: clean up some more broken refsJeff King, Sep 24, 2021
  3. Jonathan TanSep 27, 2021
  4. Jeff KingSep 27, 2021
  5. 02/16 t5516: don't use HEAD ref for invalid ref-deletion testsJeff King, Sep 24, 2021
  6. 03/16 t5600: provide detached HEAD for corruption failuresJeff King, Sep 24, 2021
  7. 04/16 t5312: drop "verbose" helperJeff King, Sep 24, 2021
  8. 05/16 t5312: create bogus ref as necessaryJeff King, Sep 24, 2021
  9. 06/16 t5312: test non-destructive repackJeff King, Sep 24, 2021
  10. 07/16 t5312: be more assertive about command failureJeff King, Sep 24, 2021
  11. 08/16 refs-internal.h: move DO_FOR_EACH_* flags next to each otherJeff King, Sep 24, 2021
  12. 09/16 refs-internal.h: reorganize DO_FOR_EACH_* flag documentationJeff King, Sep 24, 2021
  13. 10/16 refs: add DO_FOR_EACH_OMIT_DANGLING_SYMREFS flagJeff King, Sep 24, 2021
  14. 11/16 refs: omit dangling symrefs when using GIT_REF_PARANOIAJeff King, Sep 24, 2021
  15. 12/16 refs: turn on GIT_REF_PARANOIA by defaultJeff King, Sep 24, 2021
  16. Jonathan TanSep 27, 2021
  17. 13/16 repack, prune: drop GIT_REF_PARANOIA settingsJeff King, Sep 24, 2021
  18. 14/16 ref-filter: stop setting FILTER_REFS_INCLUDE_BROKENJeff King, Sep 24, 2021
  19. 15/16 ref-filter: drop broken-ref code entirelyJeff King, Sep 24, 2021
  20. 16/16 refs: drop "broken" flag from for_each_fullref_in()Jeff King, Sep 24, 2021
  21. Jonathan TanSep 27, 2021
  22. Junio C HamanoSep 24, 2021

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.