Re: [PATCH 2/3] do not write null sha1s to on-disk index

From: Jeff King <peff@peff.net>
Date: Dec 29, 2012, 11:05 UTC
Message-ID: <20121229110541.GA1408@sigill.intra.peff.net>
In-Reply-To: <20121229103430.GG18903@elie.Belkin>

On Sat, Dec 29, 2012 at 02:34:30AM -0800, Jonathan Nieder wrote:

Show 11 quoted lines

> >>   $ git am --abort
> >>   Unstaged changes after reset:
> >>   M       sound/usb/midi.c
> >
> > What does your index look like afterwards? Does it have a null sha1 in
> > it (check "ls-files -s")?
> 
> $ git diff-index --abbrev HEAD
> :100644 100644 eeefbce3873c... 000000000000... M	sound/usb/midi.c
> $ git ls-files --abbrev -s sound/usb/midi.c
> 100644 eeefbce3873c 0	sound/usb/midi.c

Hmm. It looks like "am --abort" overwrites the index again after the read-tree which complains. If I downgrade the error in write_index to a warning, like this:

diff --git a/read-cache.c b/read-cache.c
index fda78bc..70a6d86 100644
--- a/read-cache.c
+++ b/read-cache.c
@@ -1797,7 +1797,7 @@ int write_index(struct index_state *istate, int newfd)
 		if (!ce_uptodate(ce) && is_racy_timestamp(istate, ce))
 			ce_smudge_racily_clean_entry(ce);
 		if (is_null_sha1(ce->sha1))
-			return error("cache entry has null sha1: %s", ce->name);
+			warning("cache entry has null sha1: %s", ce->name);
 		if (ce_write_entry(&c, newfd, ce, previous_name) < 0)
 			return -1;
 	}

and then just run this:

  [clear state from last run]
  $ rm -rf .git/rebase-apply
  $ git reset --hard

  [apply the patch; we get a conflict]
  $ git am -3sc queue-3.2/alsa-usb-audio-fix-missing-autopm-for-midi-input.patch

  [now run just the read-tree from "am --abort"]
  $ git.compile read-tree --reset -u HEAD ORIG_HEAD
  warning: cache entry has null sha1: sound/usb/midi.c

  [and now check our index]
  $ git ls-files -s sound/usb/midi.c
  100644 0000000000000000000000000000000000000000 0 sound/usb/midi.c

  [yes, this index is bogus]
  $ git write-tree
  error: invalid object 100644 0000000000000000000000000000000000000000 for 'sound/usb/midi.c'
  fatal: git-write-tree: error building trees

So I think this check may actually be finding a real bug. I have seen
these null sha1s in the wild, but I was never able to track down the
actual cause. Maybe this will give us a clue. Now we just need to work
backwards and figure out who is putting it in the in-memory index and
why.

I'll try to work on it tomorrow, but please don't let that stop you if
you want to keep digging in the meantime.

-Peff