Re: Performance regression from switching lock to rw-sem for anon-vma tree

[Tim Chen <tim.c.chen@linux.intel.com>]

On Mon, 2013-08-12 at 20:52 +0200, Ingo Molnar wrote:
> * Tim Chen <tim.c.chen@linux.intel.com> wrote:
> 
> > On Tue, 2013-08-06 at 16:55 -0700, Davidlohr Bueso wrote:
> > 
> > > I got good numbers, recovering the performance drop I noticed with the
> > > i_mmap_mutex to rwsem patches.
> > 
> > That's good.  I remembered that the earlier version of the patch not 
> > only recovered the performance drop, but also provide some boost when 
> > you switch from i_mmap_mutex to rwsem for aim7.  Do you see similar 
> > boost with this version?
> > 
> > >  Looking forward to a more upstreamable
> > > patchset that deals with this work, including the previous patches.
> > > 
> > > One thing that's bugging me about this series though is the huge amount
> > > of duplicated code being introduced to rwsems from mutexes. We can share
> > > common functionality such as mcs locking (perhaps in a new file under
> > > lib/), can_spin_on_owner() and owner_running(), perhaps moving those
> > > functions into sheduler code, were AFAIK they were originally.
> > 
> > I think that MCS locking is worth breaking out as its
> > own library.  After we've done that, the rest of
> > the duplication are minimal. It is easier
> > to keep them separate as there are some rwsem 
> > specific logic that may require tweaking
> > to can_spin_on_owner and owner_running.  
> 
> That's what I would strongly suggest to be the approach of these patches: 
> first the MCS locking factoring out, then changes in rwsem behavior.
> 
> I'd suggest the librarization should be done using inlines or so, so that 
> we don't touch the current (pretty good) mutex.o code generation. I.e. 
> code library only on the source code level.
> 
> Done that way we could also apply the librarization first, without having 
> to worry about performance aspects. Having the code shared will also make 
> sure that an improvement to the mutex slowpaths automatically carries over 
> into rwems and vice versa.

Ingo and Davidlohr,

Thanks for your feedbacks.  I'll spin off a set of new patches to
incorporate your suggestions later.

Tim

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>