Re: [lttng-dev] [-stable 3.8.1 performance regression] madvise POSIX_FADV_DONTNEED

[Andrew Morton <akpm@linux-foundation.org>]

On Mon, 17 Jun 2013 17:39:36 -0400 Rapha__l Beamonte <raphael.beamonte@gmail.com> wrote:

> 2013/6/17 Andrew Morton <akpm@linux-foundation.org>
> 
> > That change wasn't terribly efficient - if there are any unpopulated
> > pages in the range (which is quite likely), fadvise() will now always
> > call invalidate_mapping_pages() a second time.
> >
> > Perhaps this is fixable.  Say, make lru_add_drain_all() return a
> > success code, or even teach lru_add_drain_all() to return a code
> > indicating that one of the spilled pages was (or might have been) on a
> > particular mapping.
> >
> 
> Following our tests results, that was the call to lru_add_drain_all() that
> causes the problem. The second call to invalidate_mapping_pages() isn't
> really important. We tried to compile a kernel with the commit introducing
> this change but with the "lru_add_drain_all()" line removed, and the
> problem disappeared, even if we called two times invalidate_mapping_pages()
> (as the rest of the commit was still here).

Ah, OK, schedule_on_each_cpu() could certainly do that - it has to wait
for every CPU to context switch and schedule the worker function.

There's a lot we could do here.  Such as not doing the schedule_work()
at all for a cpu which has an empty lru_add_pvec.  Or even pass down
the address_space and only schedule the work for CPUs which have a page
from *this mapping* in their lru_add_pvec.  That will all be highly
racy, but as long as the failure mode is "flushed unnecessarily" then
that's OK.