On 2019-08-15 at 22:10:29, Junio C Hamano wrote:
> Junio C Hamano <gitster@pobox.com> writes:
> 
> > "brian m. carlson" <sandals@crustytoothpaste.net> writes:
> >
> >> When applying multiple patches with git am, or when rebasing using the
> >> am backend, it's possible that one of our patches has updated a
> >> gitattributes file. Currently, we cache this information, so if a
> >> file in a subsequent patch has attributes applied, the file will be
> >> written out with the attributes in place as of the time we started the
> >> rebase or am operation, not with the attributes applied by the previous
> >> patch. This problem does not occur when using the -m or -i flags to
> >> rebase.
> > ...
> > "rebase -m" and "rebase -i" are not repeated run_command() calls
> > that invoke "git cherry-pick" or "git merge" these days, either, so
> > I am somewhat curious how they avoid fallilng into the same trap.
> >
> > Thanks for the fix.  Will queue.
> 
> Actually there still is one more thing I wasn't clear about the
> change.
> 
> > To ensure we write the correct data into the working tree, expire the
> > cache after each patch that touches a path ending in ".gitattributes".
> > ...
> > +			if (!flush_attributes && patch->new_name &&
> > +			    ends_with_path_components(patch->new_name, GITATTRIBUTES_FILE))
> > +				flush_attributes = 1;
> 
> When an attribute file is removed by a patch, we should forget what
> we read earlier from the file before it got removed.  Would such a
> case, where patch->new_name would be NULL, be handled correctly?

That's a good question. I don't think we handle that case, so I'll
include that (and an appropriate test) with my reroll.

I suspect we probably want something like:

  if ((patch->new_name && ends_with_path_components(patch->new_name, GITATTRIBUTES_FILE)) ||
      (patch->old_name && ends_with_path_components(patch->old_name, GITATTRIBUTES_FILE)))

This should cause at most one flush per patch, and I think this should
cover both cases, as well as any we haven't thought of. It's also
possible that we could get a copy or rename that causes a false
positive, but considering the contents of a .gitattributes file, that
seems unlikely enough in practice that it's probably not worth worrying
about for now.

> The call to ends_with_path_components() is almost no cost, and I
> would suspect that this call is easier to reason about without the
> "!flush_attributes &&" in the conditional part, by the way.

Yeah, it probably is. It was added to avoid the allocation, but now that
we don't have one, it shouldn't be a problem.
-- 
brian m. carlson: Houston, Texas, US
OpenPGP: https://keybase.io/bk2204