Re: [BUG] t7063-status-untracked-cache flaky?

From: Junio C Hamano <gitster@pobox.com>
Date: Oct 16, 2015, 20:34 UTC
Message-ID: <xmqqfv1aeena.fsf@gitster.mtv.corp.google.com>
In-Reply-To: <1445026298.20887.28.camel@twopensource.com>

David Turner <dturner@twopensource.com> writes:

Show 20 quoted lines

> The problem is:
>
> trash directory.t7063-status-untracked-cache$ diff trace trace.expect 
> 3,4c3,4
> < directory invalidation: 1
> < opendir: 1
> ---
>> directory invalidation: 2
>> opendir: 2
>
>
> I can repro on a SSD.
>
> I had to try many times to reproduce (I think even more the second
> time). I just ran the test in a while loop until it failed.
>
> I suspect that the kernel might be a bit slow to update the mtime on the
> directory, but I have not yet been able to repro, and I don't understand
> why it only happens in this one test, since sparseness should be
> completely unrelated.

If you are invalidating your cache based on mtime of the directory, an operation within mtime granularity can cause you to miss it, unless you try to be conservative and do the "assume invalid if racy". And if you do "assume invalid if racy", the result will become timing dependent. You may end up invalidating more than absolutely necessary, i.e. the test writer may have known that at least 1 must be definitely stale at that point in the test, but if another directory was iffy and the code played safe, you would see one more invalidation than expected (which would lead to scanning the directory, hence one more opendir).

Is that what is going on here?