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

[Tim Chen]

On Tue, 2013-07-30 at 21:24 +0200, Ingo Molnar wrote:

> > 
> > Signed-off-by: Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx>
> 
> > +config RWSEM_SPIN_ON_WRITE_OWNER
> > +	bool "Optimistic spin write acquisition for writer owned rw-sem"
> > +	default n
> > +	depends on SMP
> > +	help
> > +	  Allows a writer to perform optimistic spinning if another writer own
> > +	  the read write semaphore.  If the lock owner is running, it is likely
> > +	  to release the lock soon. Spinning gives a greater chance for writer to
> > +	  acquire a semaphore before putting it to sleep.
> 
> The way you've worded this new Kconfig option makes it 
> sound as if it's not just for testing, and I'm not a big 
> believer in extra Kconfig degrees of freedom for 
> scalability features of core locking primitives like 
> rwsems, in production kernels ...
> 
> So the bad news is that such scalability optimizations 
> really need to work for everyone.
> 
> The good news is that I don't think there's anything 
> particularly controversial about making the rwsem write 
> side perform just as well as mutexes - it would in fact be 
> a very nice quality of implementation feature: it gives 
> people freedom to switch between mutexes and rwsems without 
> having to worry about scalability differences too much.
> 

Sorry for replying to your email late as I was pulled to
some other tasks.

Ingo, any objection if I make the optimistic writer spin the
default for SMP without an extra config? This will make 
the rw_semaphore structure grow a bit to accommodate the
owner and spin_mlock field.

Thanks.

Tim
> Once readers are mixed into the workload can we keep the 
> XFS assumptions, if they are broken in any way?
> 
> We are spinning here so we have full awareness about the 
> state of the lock and we can react to a new reader in very 
> short order - so at a quick glance I don't see any 
> fundamental difficulty in being able to resolve it - beyond 
> the SMOP aspect that is ... :-)
> 
> Thanks,
> 
> 	Ingo


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/