Re: [PATCH v5 0/6] fs/dcache: Track & limit # of negative dentries

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Mon, 2018-07-02 at 14:18 -0700, Andrew Morton wrote:
> On Mon, 2 Jul 2018 12:34:00 -0700 Linus Torvalds <torvalds@linux-foun
> dation.org> wrote:
> 
> > On Sun, Jul 1, 2018 at 10:52 PM Waiman Long <longman@redhat.com>
> > wrote:
> > > 
> > > A rogue application can potentially create a large number of
> > > negative
> > > dentries in the system consuming most of the memory available if
> > > it
> > > is not under the direct control of a memory controller that
> > > enforce
> > > kernel memory limit.
> > 
> > I certainly don't mind the patch series, but I would like it to be
> > accompanied with some actual example numbers, just to make it all a
> > bit more concrete.
> > 
> > Maybe even performance numbers showing "look, I've filled the
> > dentry
> > lists with nasty negative dentries, now it's all slower because we
> > walk those less interesting entries".
> > 
> 
> (Please cc linux-mm@kvack.org on this work)
> 
> Yup.  The description of the user-visible impact of current behavior
> is far too vague.
> 
> In the [5/6] changelog it is mentioned that a large number of -ve
> dentries can lead to oom-killings.  This sounds bad - -ve dentries
> should be trivially reclaimable and we shouldn't be oom-killing in
> such a situation.

If you're old enough, it's déjà vu; Andrea went on a negative dentry
rampage about 15 years ago:

https://lkml.org/lkml/2002/5/24/71

I think the summary of the thread is that it's not worth it because
dentries are a clean cache, so they're immediately shrinkable.

> Dumb question: do we know that negative dentries are actually
> worthwhile?  Has anyone checked in the past couple of
> decades?  Perhaps our lookups are so whizzy nowadays that we don't
> need them?

There are still a lot of applications that keep looking up non-existent 
files, so I think it's still beneficial to keep them.  Apparently
apache still looks for a .htaccess file in every directory it
traverses, for instance.  Round tripping every one of these to disk
instead of caching it as a negative dentry would seem to be a
performance loser here.

However, actually measuring this again might be useful.

James