bug#21559: 25.0.50; auto-revert-mode breaks git rebase

[Mitchel Humpherys <mitch.special@gmail.com>]

On Sat, Feb 06 2016 at 09:03:35 PM, Mitchel Humpherys <mitch.special@gmail.com> wrote:
> On Sat, Feb 06 2016 at 05:34:49 PM, Mitchel Humpherys <mitch.special@gmail.com> wrote:
>> I took a quick look at the git source and the main source of this
>> particular error seems to be from the `git update-index' command.  I see
>> that vc-git.el is calling `git update-index' in a few places but my
>> attempts at instrumenting the code to track down where it was coming
>> from were fruitless.  I tried:
>>
>>     (setq vc-command-messages t)
>>
>> as well as:
>>
>>     (defun vc-git-command (buffer okstatus file-or-list &rest flags)
>>       "A wrapper around `vc-do-command' for use in vc-git.el.
>>     The difference to vc-do-command is that this function always invokes
>>      `vc-git-program'."
>>        (let ((coding-system-for-read vc-git-commits-coding-system)
>>      	(coding-system-for-write vc-git-commits-coding-system))
>>     +    (message "git: %s %s" file-or-list flags)
>>
>> Any ideas on how we can trace every git command that vc-git.el is
>> running?  I'm suspicious that we're calling `git update-index' in the
>> auto revert path somewhere...
>
> Sorry, I take it all back.  `git update-index' is the source of that one
> particular index.lock error message, but it's certainly not the only
> thing holding index.lock...  And actually it looks like `git ls-files'
> does take index.lock:
>
>     $ cd /path/to/emacs/
>     $ git ls-files # in another shell after running the following inotifywait:
>     $ inotifywait -m .git | grep index.lock
>     Setting up watches.
>     Watches established.
>     .git/ CREATE index.lock
>     .git/ OPEN index.lock
>     .git/ CLOSE_WRITE,CLOSE index.lock
>     .git/ DELETE index.lock
>     .git/ CREATE index.lock
>     .git/ OPEN index.lock
>     .git/ CLOSE_WRITE,CLOSE index.lock
>     .git/ DELETE index.lock
>
> I did a bit of git debugging and it looks like someone must be
> registering an atexit handler or something that takes index.lock,
> because it doesn't get taken in the ls-files code itself.  It's taken
> sometime after the `exit' function gets called in `handle_builtin' in
> git.c.
>
> Anyways, I'm more inclined now to agree that this a git bug.  I don't
> see why `git ls-files' would be taking index.lock...  So we're back to
> the question of "how can Emacs handle this more gracefully".

SORRY!  I take it all back AGAIN.  It was actually my shell's PS1 that
was locking index.lock (due to a `git status'), not `git ls-files'!
That's why it looked like it was happening when the program exited.
Because it really was after the program exited :).  That's
embarassing...

After re-running my experiment above in a non-fancy shell I don't see
index.lock being taken at all when I run `git ls-files'.

So we're back to instrumenting vc{,-git}.el to see exactly which git
commands are being run when auto-revert refreshes a buffer.