Re: What's cooking in git.git (Nov 2016, #06; Mon, 28)

[Jeff King <peff@peff.net>]

On Wed, Nov 30, 2016 at 04:06:05PM -0800, Brandon Williams wrote:

> On 11/30, Jeff King wrote:
> > So I think there is some other chdir(). I'm not sure if there is an easy
> > way to get a backtrace on every call to chdir() in every thread. I'm
> > sure somebody more clever than me could figure out how to make gdb do it
> > automatically, but it might be workable manually. I think the chdir was
> > in the main thread.
> > 
> > -Peff
> 
> Yeah maybe we're missing something else...
> 
> How did you run strace with your stress script?

It's hidden in the patch I sent a moment ago, but basically just "strace
-o foo.out" will dump the trace in the trash directory. After the stress
script runs, you can "cat fail/trash*/foo.out".

> > (gdb) bt
> > #0  chdir () at ../sysdeps/unix/syscall-template.S:84
> > #1  0x00005555555fe259 in real_path_internal (path=0x5555559f6b30 "su:b/../.git/modules/su:b", die_on_error=1)
> >     at abspath.c:84
> > #2  0x00005555555fe48a in real_path (path=0x5555559f6b30 "su:b/../.git/modules/su:b") at abspath.c:135
> > #3  0x00005555556d09e6 in read_gitfile_gently (path=0x5555559f6ac0 "su:b/.git", return_error_code=0x0)
> >     at setup.c:555
> > #4  0x00005555556d19cf in resolve_gitdir (suspect=0x5555559f6ac0 "su:b/.git") at setup.c:1021
> > #5  0x00005555556e7e34 in is_submodule_populated (path=0x5555559f5ec8 "su:b") at submodule.c:244
> > #6  0x00005555555a0f05 in grep_submodule (opt=0x7fffffffd8b0, sha1=0x0, filename=0x5555559f5ec8 "su:b", 
> >     path=0x5555559f5ec8 "su:b") at builtin/grep.c:619
> > #7  0x00005555555a12ac in grep_cache (opt=0x7fffffffd8b0, pathspec=0x7fffffffd880, cached=0) at builtin/grep.c:700
> > #8  0x00005555555a36cb in cmd_grep (argc=0, argv=0x7fffffffdf40, prefix=0x0) at builtin/grep.c:1257
> > #9  0x000055555556603b in run_builtin (p=0x5555559b3ad8 <commands+984>, argc=4, argv=0x7fffffffdf40) at git.c:373
> > #10 0x00005555555662bc in handle_builtin (argc=4, argv=0x7fffffffdf40) at git.c:572
> > #11 0x000055555556641a in run_argv (argcp=0x7fffffffddfc, argv=0x7fffffffddf0) at git.c:630
> > #12 0x00005555555665a8 in cmd_main (argc=4, argv=0x7fffffffdf40) at git.c:702
> > #13 0x00005555555fde47 in main (argc=7, argv=0x7fffffffdf28) at common-main.c:40
> > 
> > So is_submodule_populated() needs to take a lock. But what's really
> > gross is that the _other_ threads need to lock just to call lstat().
> > Presumably it could be done as a reader/writer type of lock where many
> > "reader" threads can take the "I need to lstat()" lock simultaneously,
> > but block when an "I'm going to chdir()" writer holds it.
> 
> Oh interesting, I wonder if there is a way to not have to perform a
> chdir since taking a lock to lstat wouldn't be ideal.

I don't think so.  It comes from real_path(), which needs to either
chdir(), or start interpreting symbolic links itself (and madness that
way lies).

I think with a reader/writer lock as I described it wouldn't be too bad.
The common case would pay only the locking cost and not ever block,
since submodules are rare (and they're super-heavyweight to descend into
anyway).

I think putting it at the individual lstat() would be way too low, but
probably you could do it right before calling grep_source(). It may even
be possible to do some of the submodule work ahead of time while holding
grep_lock().

> Thanks for helping out with this!

I wasn't planning on it, but this turned into an intriguing puzzle. ;)

-Peff