Re: [PATCH] mmu notifiers #v5

[Andrea Arcangeli]

On Tue, Feb 05, 2008 at 03:10:52PM -0800, Christoph Lameter wrote:
> On Tue, 5 Feb 2008, Andrea Arcangeli wrote:
> 
> > > You can avoid the page-pin and the pt lock completely by zapping the 
> > > mappings at _start and then holding off new references until _end.
> > 
> > "holding off new references until _end" = per-range mutex less scalar
> > and more expensive than the PT lock that has to be taken anyway.
> 
> You can of course setup a 2M granularity lock to get the same granularity 
> as the pte lock. That would even work for the cases where you have to page 
> pin now.

If you set a 2M granularity lock, the _start callback would need to
do:

	for_each_2m_lock()
		mutex_lock()

so you'd run zillon of mutex_lock in a row, you're the one with the
million of operations argument.

> The size of the mmap is relevant if you have to perform callbacks on 
> every mapped page that involved take mmu specific locks. That seems to be 
> the case with this approach.

mmap should never trigger any range_start/_end callback unless it's
overwriting an older mapping which is definitely not the interesting
workload for those apps including kvm.

> Optimizing do_exit by taking a single lock to zap all external references 
> instead of 1 mio callbacks somehow leads to slowdown?

It can if the application runs for more than a couple of seconds,
i.e. not a fork flood in which you care about do_exit speed. Keep in
mind if you had 1mio invalidate_pages callback it means you previously
called follow_page 1 mio of times too...
--
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/