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

[Stefan Beller <sbeller@google.com>]

On Wed, Nov 30, 2016 at 3:40 PM, Jeff King <peff@peff.net> wrote:
> On Wed, Nov 30, 2016 at 06:32:04PM -0500, Jeff King wrote:
>
>> On Wed, Nov 30, 2016 at 03:28:23PM -0800, Brandon Williams wrote:
>>
>> > So I couldn't find a race condition in the code.  I tracked the problem
>> > to grep_source_load_file which attempts to run lstat on the file so that
>> > it can read it into a buffer.  The lstat call fails with ENOENT (which
>> > conveniently is skipped by the if statement which calls error_errno).  So
>> > for some reason the file cannot be found and read into memory resulting
>> > in nothing being grep'ed for that particular file (since the buffer is
>> > NULL).
>>
>> That's definitely weird. Is it possible that any of the underlying calls
>> from another thread are using chdir()? I think realpath() make do that
>> behind the scenes, and there may be others.
>>
>> A full strace from a failing case would be interesting reading. In
>> theory we should be able to get that by running the stress script for
>> long enough. :)
>
> Actually, it failed pretty much immediately. I guess the extra stracing
> changes the timing to make the problem _more_ likely.
>
> And indeed, I see:
>
> 20867 lstat("fi:le",  <unfinished ...>
> 20813 <... read resumed> "", 232)       = 0
> 20871 futex(0x558cdec8b164, FUTEX_WAIT_PRIVATE, 7, NULL <unfinished ...>
> 20813 close(7 <unfinished ...>
> 20870 <... futex resumed> )             = 0
> 20869 lstat(".gitmodules",  <unfinished ...>
> 20813 <... close resumed> )             = 0
> 20865 set_robust_list(0x7f1df92579e0, 24 <unfinished ...>
> 20813 lstat("su:b/../.git/modules/su:b/commondir", 0x7ffecc8b3ac0) = -1 ENOENT (No such file or directory)
> 20865 <... set_robust_list resumed> )   = 0
> 20868 set_robust_list(0x7f1df7a549e0, 24 <unfinished ...>
> 20813 access("su:b/../.git/modules/su:b/objects", X_OK) = 0
> 20813 access("su:b/../.git/modules/su:b/refs", X_OK) = 0
> 20813 stat("su:b/../.git/modules/su:b", {st_mode=S_IFDIR|0755, st_size=280, ...}) = 0
> 20813 getcwd("/var/ram/git-stress/root-4/trash directory.t7814-grep-recurse-submodules/parent", 129) = 80
> 20869 <... lstat resumed> {st_mode=S_IFREG|0644, st_size=47, ...}) = 0
> 20813 chdir("su:b/../.git/modules/su:b") = 0
> 20869 open(".gitmodules", O_RDONLY <unfinished ...>
> 20813 getcwd( <unfinished ...>
> 20867 <... lstat resumed> 0x7f1df8254cf0) = -1 ENOENT (No such file or directory)
>
> where 20813 and 20867 are two threads of the main process. One is doing
> the lstat and the other calls chdir at the same moment.
>

Lessons learned here:
The run-command API is not thread safe when used with setting the directory.
If you need to run a thing in a threaded environment run
git -C <dir> ... such that the child chdirs.

Are there any other threaded environments that run things with .dir set?