Re: [PATCH] KVM: try __get_user_pages_fast even if not in atomic context

[Andrea Arcangeli <aarcange@redhat.com>]

Hello,

On Mon, Aug 06, 2018 at 01:44:49PM +0200, Paolo Bonzini wrote:
> On 06/08/2018 09:51, Xiao Guangrong wrote:
> > 
> > 
> > On 07/27/2018 11:46 PM, Paolo Bonzini wrote:
> >> We are currently cutting hva_to_pfn_fast short if we do not want an
> >> immediate exit, which is represented by !async && !atomic.  However,
> >> this is unnecessary, and __get_user_pages_fast is *much* faster
> >> because the regular get_user_pages takes pmd_lock/pte_lock.
> >> In fact, when many CPUs take a nested vmexit at the same time
> >> the contention on those locks is visible, and this patch removes
> >> about 25% (compared to 4.18) from vmexit.flat on a 16 vCPU
> >> nested guest.
> >>
> > 
> > Nice improvement.
> > 
> > Then after that, we will unconditionally try hva_to_pfn_fast(), does
> > it hurt the case that the mappings in the host's page tables have not
> > been present yet?
> 
> I don't think so, because that's quite slow anyway.

There will be a minimal impact, but it's worth it.

The reason it's worth is that we shouldn't be calling
get_user_pages_unlocked in hva_to_pfn_slow if we could pass
FOLL_HWPOISON to get_user_pages_fast.

And get_user_pages_fast is really just __get_user_pages_fast +
get_user_pages_unlocked with just a difference (see below).

Reviewed-by: Andrea Arcangeli <aarcange@redhat.com>

> 
> > Can we apply this tech to other places using gup or even squash it
> > into  get_user_pages()?
> 
> That may make sense.  Andrea, do you have an idea?

About further improvements looking at commit
5b65c4677a57a1d4414212f9995aa0e46a21ff80 it looks like it may be worth
adding a new gup variant __get_user_pages_fast_irq_enabled to make our
slow path "__get_user_pages_fast_irq_enabled +
get_user_pages_unlocked" really as fast as get_user_pages_fast (which
we can't call in the atomic case and can't take the foll flags, making
it take the foll flags would also make it somewhat slower by adding
branches).

If I understand correctly the commit header Before refers to when
get_user_pages_fast was calling __get_user_pages_fast, and After is
the optimized version without local_irq_save/restore but instead using
local_irq_disable/enable.

So we'd need to call a new __get_user_pages_fast_irq_enabled instead
of __get_user_pages_fast that would only safe to call when irq are
enabled and that's always the case for KVM also for the atomic case
(KVM's atomic case is atomic only because of the spinlock, not because
irqs are disabled). Such new method would then also be ok to be called
from interrupts as long as irq are enabled when it is being called.

Such change would also contribute to reduce the minimal impact to the
_slow case. x86 would be sure fine with the generic version and it's
trivial to implement, I haven't checked other arch details.

Thanks,
Andrea