Re: Git 2.7.0 gitignore behaviour regression

On Fri, Jan 08, 2016 at 02:41:25AM +0000, brian m. carlson wrote:

On Fri, Jan 08, 2016 at 07:38:58AM +0700, Duy Nguyen wrote:

quoted

On Fri, Jan 8, 2016 at 6:44 AM, brian m. carlson
[off-list ref] wrote:

quoted

I think there's still a bug in the code here.  If you do

  git init
  mkdir -p base/a/
  printf 'base/a/\n!base/a/b.txt\n' >.gitignore

Here we have the ignore rule "base/a/", but gitignore.txt, section
NOTES mentions this

 - The rules to exclude the parent directory must not end with a
   trailing slash.

The text here says, "To re-include files or directories when their
parent directory is excluded, the following conditions must be met".  In
other words, the text implies that it's required for re-inclusion to
work, not exclusion.

quoted

quoted

  git add .gitignore
  git commit -m 'Add .gitignore'
  >base/a/b.txt
  git add base/a/b.txt
  git commit -m 'Add base/a/b.txt'
  >base/a/c.txt
  git status --porcelain

git status outputs base/a/c.txt as unknown, when it should be ignored.
We saw this in a repository at $DAYJOB.

If I delete that trailing slash, c.txt is ignored. So it's known
limitation. I think we can make trailing slash case work too, but if I
remember correctly it would involve a lot more changes, so I didn't do
it (there are other conditions to follow anyway to make it work).

The case I'm seeing is that b.txt was already checked into the
repository before being re-added, and c.txt was not.  So it didn't
affect us that b.txt was ignored (as it was already in the repo), but
c.txt not being ignored broke a whole bunch of scripts that checked that
the repository was clean, simply because we upgraded Git.

I think regardless of whether b.txt is re-included, c.txt should be
ignored.  If it isn't possible to re-include b.txt, that's fine, since
that isn't a regression, but ignored files should remain ignored.

Thanks for clarification. I looked at this the wrong way. I agree it
is a regression.

The following should fix it. It looks correct (and does fix your test
case), but I will have to look harder over the weekend before sending
a proper patch.

-- 8< --

diff --git a/dir.c b/dir.c
index d2a8f06..7934e87 100644
--- a/dir.c
+++ b/dir.c

@@ -1008,6 +1008,7 @@ static struct exclude *last_exclude_matching_from_list(const char *pathname,
 	if (exc &&
 	    !(exc->flags & EXC_FLAG_NEGATIVE) &&
 	    !(exc->flags & EXC_FLAG_NODIR) &&
+	    !(exc->flags & EXC_FLAG_MUSTBEDIR) &&
 	    matched_negative_path)
 		exc = NULL;
 	return exc;
-- 8< --
--

Duy