Re: [PATCH RFC] x86: avoid atomic operation in test_and_set_bit_lock if possible

[Jack Steiner <steiner@sgi.com>]

On Fri, Mar 25, 2011 at 09:29:34AM -0700, Linus Torvalds wrote:
> On Fri, Mar 25, 2011 at 3:06 AM, Jan Beulich <JBeulich@novell.com> wrote:
> >
> > The problem was observed with __lock_page() (in a variant not
> > upstream for reasons not known to me), and prefixing e.g.
> > trylock_page() with an extra PageLocked() check yielded the
> > below quoted improvements.
> 
> Ok. __lock_page() _definitely_ should do the test_bit() thing first,
> because it's normally called from lock_page() that has already tested
> the bit.
> 
> But it already seems to do that, so I'm wondering what your variant is.
> 
> I'm also a bit surprised that lock_page() is that hot (unless your
> _lock_page() variant is simply too broken and ends up spinning?).
> Maybe we have some path that takes the page lock unnecessarily? What's
> the load?

We see the problem primarily on launching very large MPI applications.
The master process rapidly forks a large number (1 per cpu) of processes.
Each faults in a large number of text pages.

The text pages are resident in the page cache. No IO is involved but
the page lock quickly becomes a very hot contended cacheline.

Note also that this is observed in a 2.6.32 distro kernel that has a
different implementation of __lock_page. I think a similar problem
exists in the upstream kernel but have not had a chance to investigate.

We also see a similar problem during boot when a large number of udevd
processes are created.


--- jack