Re: git-core: try_to_follow_renames(): git killed by SIGSEGV

From: SZEDER Gábor <szeder.dev@gmail.com>
Date: Mar 10, 2020, 16:57 UTC
Message-ID: <20200310165723.GB3122@szeder.dev>
In-Reply-To: <3c722d21-ee57-7d20-81fb-0399f02f1bc7@syntevo.com>

On Fri, Mar 06, 2020 at 03:44:34PM +0100, Alexandr Miloslavskiy wrote:

Show 5 quoted lines

> Since I like studying crashes and noone else replied, I decided to have a
> look.
> 
> The problem is easy to reproduce with this (replace 1.c with any file):
>   git log --follow -L 1,1:1.c -- 1.c

Don't do this.  In particular:

  - Don't use line-level log with a pathspec, because the
    documentation of 'git log -L' explicitly told you not to do so
    ("You may not give any pathspec limiters.").  This should have
    errored out since the beginning, but, unfortunately, has never
    been enforced.

  - Don't use '-L' with '--follow'.  On one hand, line-level log on
    its own already follows file renames, even multiple files at once,
    there is no need for an additional '--follow' (which can only
    follow one file).  OTOH, you shouldn't be able to use '-L' and
    '--follow' together, because the former forbids a pathspec, while
    the latter requires one.

In any case, '--follow' has always been an ugly hack on top of the revision walking machinery, while line-level log is a rather poorly integrated bolt-on. They simply weren't designed to work together, as evidenced by their contradicting requirements about the pathspec.

Show 46 quoted lines

> It occurs because `opt->pathspec.items` gets cleaned here:
>     clear_pathspec
>     queue_diffs
>         /* must look at the full tree diff to detect renames */
>         clear_pathspec(&opt->pathspec);
>         DIFF_QUEUE_CLEAR(&diff_queued_diff);
>     process_ranges_ordinary_commit
>     process_ranges_arbitrary_commit
>     line_log_filter
>     prepare_revision_walk
>     cmd_log_walk
>     cmd_log
> 
> And on next iteration it crashes in 'try_to_follow_renames' on this line:
>     diff_opts.single_follow = opt->pathspec.items[0].match;
> 
> I think that bug comes from commit:
>     a2bb801f by SZEDER Gábor, 2019-08-21 13:04:24
>     line-log: avoid unnecessary full tree diffs
> 
> @szeder could you please look into that?
> 
> On 27.02.2020 13:56, Ondrej Pohorelsky wrote:
> >Hi,
> >
> >there is a SIGSEGV appearing in Fedora[0] with Git 2.24.1
> >
> >This bug started to appear after update to Git 2.24.1.
> >Bug reporter said that Git crashed on him while running VS Code with
> >Git Lens extension[1]
> >I have tried to reproduce this bug with my own compiled Git with debug
> >flags, but sadly SIGSEGV never appeared.
> >
> >To me it seems like there is a problem in commit a2bb801f6a[2] which
> >changes move_diff_queue() function. This function calls
> >diff_tree_oid() that calls try_to_follow_renames(). In the last two
> >functions there are no arguments checks.
> >
> >Best regards,
> >Ondřej Pohořelský
> >
> >[0] https://retrace.fedoraproject.org/faf/problems/bthash/?bth=25aa7d7267ab5de548ffca337115cb68f7b65105
> >[1] https://bugzilla.redhat.com/show_bug.cgi?id=1791810
> >[2] https://git.kernel.org/pub/scm/git/git.git/commit/?id=a2bb801f6a430f6049e5c9729a8f3bf9097d9b34
> >
>