diff mbox series

fixup! propagate errno from failing read

Message ID pull.1068.git.git.1629203489546.gitgitgadget@gmail.com (mailing list archive)
State New, archived
Headers show
Series fixup! propagate errno from failing read | expand

Commit Message

Han-Wen Nienhuys Aug. 17, 2021, 12:31 p.m. UTC
From: Han-Wen Nienhuys <hanwen@google.com>

This fixes a crash triggered by the BUG() statement.

This can occur with symlinked .git/refs. To check availability,
refs_verify_refname_available will run refs_read_raw_ref() on each prefix,
leading to a read() from .git/refs (which is a directory).

When handling the symlink case, it is probably more robust to re-issue the
lstat() as a normal stat(), in which case, we would fall back to the directory
case.

For now, propagating errno from strbuf_read() is simpler and avoids the crash.

Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
---
    fixup! propagate errno from failing read
    
    This fixes a crash triggered by the BUG() statement.
    
    This can occur with symlinked .git/refs. To check availability,
    refs_verify_refname_available will run refs_read_raw_ref() on each
    prefix, leading to a read() from .git/refs (which is a directory).
    
    When handling the symlink case, it is probably more robust to re-issue
    the lstat() as a normal stat(), in which case, we would fall back to the
    directory case.
    
    For now, propagating errno from strbuf_read() is simpler and avoids the
    crash.
    
    Signed-off-by: Han-Wen Nienhuys hanwen@google.com

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1068%2Fhanwen%2Ffiles-fixup-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1068/hanwen/files-fixup-v1
Pull-Request: https://github.com/git/git/pull/1068

 refs/files-backend.c |  1 +
 t/t3200-branch.sh    | 14 ++++++++++++++
 2 files changed, 15 insertions(+)


base-commit: f000ecbed922c727382651490e75014f003c89ca

Comments

Jonathan Tan Aug. 17, 2021, 4:14 p.m. UTC | #1
> diff --git a/refs/files-backend.c b/refs/files-backend.c
> index 0d96eeba61b..f546cc3cc3d 100644
> --- a/refs/files-backend.c
> +++ b/refs/files-backend.c
> @@ -454,6 +454,7 @@ stat_ref:
>  	}
>  	strbuf_reset(&sb_contents);
>  	if (strbuf_read(&sb_contents, fd, 256) < 0) {
> +		myerr = errno;
>  		close(fd);
>  		goto out;
>  	}

Reviewed-by: Jonathan Tan <jonathantanmy@google.com>

Thanks - a straightforward fixup. (I don't think we need the errno from
close() in this case.)
Junio C Hamano Aug. 17, 2021, 10:38 p.m. UTC | #2
"Han-Wen Nienhuys via GitGitGadget" <gitgitgadget@gmail.com> writes:

> From: Han-Wen Nienhuys <hanwen@google.com>
>
> This fixes a crash triggered by the BUG() statement.
>
> This can occur with symlinked .git/refs. To check availability,
> refs_verify_refname_available will run refs_read_raw_ref() on each prefix,
> leading to a read() from .git/refs (which is a directory).
>
> When handling the symlink case, it is probably more robust to re-issue the
> lstat() as a normal stat(), in which case, we would fall back to the directory
> case.
>
> For now, propagating errno from strbuf_read() is simpler and avoids the crash.
>
> Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
> ---
>     fixup! propagate errno from failing read

Hmph, I do not see a commit with "propagate errno from failing read"
in its title anywhere in 'seen'.

I think the convention to assign errno to myerr in this codepath
originates in a0731250 (refs: explicitly return failure_errno from
parse_loose_ref_contents, 2021-07-20), and it forgot the part of the
code being fixed with this patch.  The commit being fixed is already
is in 'next' as part of the hn/refs-errno-cleanup topic.

Usually, a flaw in a topic that is already in 'next' is corrected by
a follow-up patch, but then they won't say "fixup!" (none of our
bugfix patches do).  But a post-release is a special time, as we
will soon be rewinding 'next', restarting it from the latest release
and we have a choice to tentatively eject a topic, fix it up or
even replace it, before merging the corrected topic to 'next'.

Do you mean that you want me to squash this change into that commit
before the topic graduates to 'master' during the new development
cycle?  If so please be a bit more explicit next time.  Using the
title of the commit after "fixup!" would be a good starting point.

Thanks.
Han-Wen Nienhuys Aug. 18, 2021, 9 a.m. UTC | #3
On Wed, Aug 18, 2021 at 12:38 AM Junio C Hamano <gitster@pobox.com> wrote:
>
> "Han-Wen Nienhuys via GitGitGadget" <gitgitgadget@gmail.com> writes:
>
> > From: Han-Wen Nienhuys <hanwen@google.com>
> >
> > This fixes a crash triggered by the BUG() statement.
> >
> > This can occur with symlinked .git/refs. To check availability,
> > refs_verify_refname_available will run refs_read_raw_ref() on each prefix,
> > leading to a read() from .git/refs (which is a directory).
> >
> > When handling the symlink case, it is probably more robust to re-issue the
> > lstat() as a normal stat(), in which case, we would fall back to the directory
> > case.
> >
> > For now, propagating errno from strbuf_read() is simpler and avoids the crash.
> >
> > Signed-off-by: Han-Wen Nienhuys <hanwen@google.com>
> > ---
> >     fixup! propagate errno from failing read
>
> Hmph, I do not see a commit with "propagate errno from failing read"
> in its title anywhere in 'seen'.

> I think the convention to assign errno to myerr in this codepath
> originates in a0731250 (refs: explicitly return failure_errno from
> parse_loose_ref_contents, 2021-07-20), and it forgot the part of the
> code being fixed with this patch.  The commit being fixed is already
> is in 'next' as part of the hn/refs-errno-cleanup topic.
>
> Usually, a flaw in a topic that is already in 'next' is corrected by
> a follow-up patch, but then they won't say "fixup!" (none of our
> bugfix patches do).  But a post-release is a special time, as we
> will soon be rewinding 'next', restarting it from the latest release
> and we have a choice to tentatively eject a topic, fix it up or
> even replace it, before merging the corrected topic to 'next'.
>
> Do you mean that you want me to squash this change into that commit
> before the topic graduates to 'master' during the new development
> cycle?  If so please be a bit more explicit next time.  Using the
> title of the commit after "fixup!" would be a good starting point.

The problem fixed here affects anyone who uses git-repo (ie. does
Android development) and runs "git-branch -m", which is a large group
of people, so I think it should not be allowed to get into a release.

So it could be squashed into commit a0731250, or put on top of next as
a separate commit (probably with 'fixup!' removed).

Note that, even though commit a0731250 originates from a branch called
"hn/XXX" and has me as Author, the BUG() call causing the crash was
actually introduced by AEvar when he reworked the series.
Junio C Hamano Aug. 18, 2021, 6:30 p.m. UTC | #4
Han-Wen Nienhuys <hanwen@google.com> writes:

>> I think the convention to assign errno to myerr in this codepath
>> originates in a0731250 (refs: explicitly return failure_errno from
>> parse_loose_ref_contents, 2021-07-20), and it forgot the part of the
>> code being fixed with this patch.  The commit being fixed is already
>> is in 'next' as part of the hn/refs-errno-cleanup topic.
>>
>> Usually, a flaw in a topic that is already in 'next' is corrected by
>> a follow-up patch, but then they won't say "fixup!" (none of our
>> bugfix patches do).  But a post-release is a special time, as we
>> will soon be rewinding 'next', restarting it from the latest release
>> and we have a choice to tentatively eject a topic, fix it up or
>> even replace it, before merging the corrected topic to 'next'.
>>
>> Do you mean that you want me to squash this change into that commit
>> before the topic graduates to 'master' during the new development
>> cycle?  If so please be a bit more explicit next time.  Using the
>> title of the commit after "fixup!" would be a good starting point.
>
> The problem fixed here affects anyone who uses git-repo (ie. does
> Android development) and runs "git-branch -m", which is a large group
> of people, so I think it should not be allowed to get into a release.

OK.  The problem already is in 'next' and we want to make sure it
won't graduate to 'master' for the next release as-is.  I agree with
that ;-)

> So it could be squashed into commit a0731250, or put on top of next as
> a separate commit (probably with 'fixup!' removed).

I'd try the former first and will fall back on the latter, then.

> Note that, even though commit a0731250 originates from a branch called
> "hn/XXX" and has me as Author, the BUG() call causing the crash was
> actually introduced by AEvar when he reworked the series.

Yup, I see his Sob after yours and it is quite understandable if a
new bug was introduced by his changes. It also would be
understandable if his change was only to add a call to BUG() in
order to assert that the original patch used myerr consistently, and
it uncovered a bug in the original version he took from you.

I do not care too much about how exactly the bug was introduced and
uncovered---it matters more that the end result has one fewer bug
thanks to the team effort.
Jonathan Nieder Aug. 18, 2021, 10:19 p.m. UTC | #5
Hi,

Jonathan Tan wrote:
> Han-Wen Nienhuys wrote:

>> diff --git a/refs/files-backend.c b/refs/files-backend.c
>> index 0d96eeba61b..f546cc3cc3d 100644
>> --- a/refs/files-backend.c
>> +++ b/refs/files-backend.c
>> @@ -454,6 +454,7 @@ stat_ref:
>>  	}
>>  	strbuf_reset(&sb_contents);
>>  	if (strbuf_read(&sb_contents, fd, 256) < 0) {
>> +		myerr = errno;
>>  		close(fd);
>>  		goto out;
>>  	}
>
> Reviewed-by: Jonathan Tan <jonathantanmy@google.com>
>
> Thanks - a straightforward fixup. (I don't think we need the errno from
> close() in this case.)

This looks good as far as it goes, but how do we know this has covered
all the code paths?  Let's see.

The only nonobvious paths are

 stat_ref:
	/*
	 * We might have to loop back here to avoid a race
	 * condition: first we lstat() the file, then we try
	 * to read it as a link or as a file.  But if somebody
	 * changes the type of the file (file <-> directory
	 * <-> symlink) between the lstat() and reading, then
	 * we don't want to report that as an error but rather
	 * try again starting with the lstat().
	 *
	 * We'll keep a count of the retries, though, just to avoid
	 * any confusing situation sending us into an infinite loop.
	 */

	if (remaining_retries-- <= 0)
		goto out;

and

	ret = parse_loose_ref_contents(buf, oid, referent, type, &myerr);

 out:
	if (ret && !myerr)
		BUG("returning non-zero %d, should have set myerr!", ret);

For the 'stat_ref' case, we have to check that whenever we 'goto
stat_ref', we set myerr moments before.  Fortunately, that is the
case.

For the 'fall through into out' case, we have the check the
parse_loose_ref_contents always sets *failure_errno on error.  That is
also the case.

So this indeed covers all our cases, and the BUG now correctly
reflects an invariant we can count on.  Thanks for the fix, and thanks
for looking it over.

Sincerely,
Jonathan
Junio C Hamano Aug. 19, 2021, 12:11 a.m. UTC | #6
Junio C Hamano <gitster@pobox.com> writes:

>> So it could be squashed into commit a0731250, or put on top of next as
>> a separate commit (probably with 'fixup!' removed).
>
> I'd try the former first and will fall back on the latter, then.

I've reverted a0731250 out of 'next', squashed the fix in, rebuilt
the topic and merged it back to 'next'.  Thanks.
Jeff King Aug. 19, 2021, 3:55 a.m. UTC | #7
On Tue, Aug 17, 2021 at 12:31:29PM +0000, Han-Wen Nienhuys via GitGitGadget wrote:

> diff --git a/t/t3200-branch.sh b/t/t3200-branch.sh
> index cc4b10236e2..dd17718160a 100755
> --- a/t/t3200-branch.sh
> +++ b/t/t3200-branch.sh
> @@ -109,6 +109,20 @@ test_expect_success 'git branch -m n/n n should work' '
>  	git reflog exists refs/heads/n
>  '
>  
> +test_expect_success 'git branch -m with symlinked .git/refs' '
> +	git init subdir &&
> +	test_when_finished "rm -rf subdir" &&
> +	(cd subdir &&
> +		for d in refs objects packed-refs ; do
> +		rm -rf .git/$d &&
> +		ln -s ../../.git/$d .git/$d ; done ) &&
> +	git --git-dir subdir/.git/ branch rename-src &&
> +	expect=$(git rev-parse rename-src) &&
> +	git --git-dir subdir/.git/ branch -m rename-src rename-dest &&
> +	test $(git rev-parse rename-dest) = "$expect" &&
> +	git branch -D rename-dest
> +'

This test presumably needs the SYMLINKS prerequisite. I noticed that the
Windows CI for "next" is now failing.

-Peff
Han-Wen Nienhuys Aug. 19, 2021, 9:01 a.m. UTC | #8
On Wed, Aug 18, 2021 at 8:30 PM Junio C Hamano <gitster@pobox.com> wrote:

> I do not care too much about how exactly the bug was introduced and
> uncovered---it matters more that the end result has one fewer bug
> thanks to the team effort.

agreed. Just wanted to give context on my bungled "fixup!" subject.
diff mbox series

Patch

diff --git a/refs/files-backend.c b/refs/files-backend.c
index 0d96eeba61b..f546cc3cc3d 100644
--- a/refs/files-backend.c
+++ b/refs/files-backend.c
@@ -454,6 +454,7 @@  stat_ref:
 	}
 	strbuf_reset(&sb_contents);
 	if (strbuf_read(&sb_contents, fd, 256) < 0) {
+		myerr = errno;
 		close(fd);
 		goto out;
 	}
diff --git a/t/t3200-branch.sh b/t/t3200-branch.sh
index cc4b10236e2..dd17718160a 100755
--- a/t/t3200-branch.sh
+++ b/t/t3200-branch.sh
@@ -109,6 +109,20 @@  test_expect_success 'git branch -m n/n n should work' '
 	git reflog exists refs/heads/n
 '
 
+test_expect_success 'git branch -m with symlinked .git/refs' '
+	git init subdir &&
+	test_when_finished "rm -rf subdir" &&
+	(cd subdir &&
+		for d in refs objects packed-refs ; do
+		rm -rf .git/$d &&
+		ln -s ../../.git/$d .git/$d ; done ) &&
+	git --git-dir subdir/.git/ branch rename-src &&
+	expect=$(git rev-parse rename-src) &&
+	git --git-dir subdir/.git/ branch -m rename-src rename-dest &&
+	test $(git rev-parse rename-dest) = "$expect" &&
+	git branch -D rename-dest
+'
+
 # The topmost entry in reflog for branch bbb is about branch creation.
 # Hence, we compare bbb@{1} (instead of bbb@{0}) with aaa@{0}.