Re: [PATCH v2 3/5] KVM: Conditionally reschedule when resetting the dirty ring

[Sean Christopherson <seanjc@google.com>]

On Mon, May 12, 2025, James Houghton wrote:
> On Thu, May 8, 2025 at 7:11 AM Sean Christopherson <seanjc@google.com> wrote:
> > ---
> >  virt/kvm/dirty_ring.c | 10 ++++++++++
> >  1 file changed, 10 insertions(+)
> >
> > diff --git a/virt/kvm/dirty_ring.c b/virt/kvm/dirty_ring.c
> > index e844e869e8c7..97cca0c02fd1 100644
> > --- a/virt/kvm/dirty_ring.c
> > +++ b/virt/kvm/dirty_ring.c
> > @@ -134,6 +134,16 @@ int kvm_dirty_ring_reset(struct kvm *kvm, struct kvm_dirty_ring *ring,
> >
> >                 ring->reset_index++;
> >                 (*nr_entries_reset)++;
> > +
> > +               /*
> > +                * While the size of each ring is fixed, it's possible for the
> > +                * ring to be constantly re-dirtied/harvested while the reset
> > +                * is in-progress (the hard limit exists only to guard against
> > +                * wrapping the count into negative space).
> > +                */
> > +               if (!first_round)
> > +                       cond_resched();
> 
> Should we be dropping slots_lock here?

Could we?  Yes.  Should we?  Eh.  I don't see any value in doing so, because in
practice, it's extremely unlikely anything will be waiting on slots_lock.

kvm_vm_ioctl_reset_dirty_pages() operates on all vCPUs, i.e. there won't be
competing calls to reset other rings.  A well-behaved userspace won't be modifying
memslots or dirty logs, and won't be toggling nx_huge_pages.

That leaves kvm_vm_ioctl_set_mem_attributes(), kvm_inhibit_apic_access_page(),
kvm_assign_ioeventfd(), snp_launch_update(), and coalesced IO/MMIO (un)registration.
Except for snp_launch_update(), those are all brutally slow paths, e.g. require
SRCU synchronization and/or zapping of SPTEs.  And snp_launch_update() is probably
fairly slow too.

And dropping slots_lock only makes any sense for non-preemptible kernels, because
preemptible kernels include an equivalent check in KVM_MMU_UNLOCK().

It's also possible that dropping slots_lock in this case could be a net negative.
I don't think it's likely, but I don't think it's any more or less likely that
droppings slots_lock is a net positive.  Without performance data to guide us,
it'd be little more than a guess, and I really, really don't want to set a
precedence of dropping a mutex on cond_resched() without a very strong reason
for doing so.

> It seems like we need to be holding slots_lock to call kvm_reset_dirty_gfn(),
> but that's it. Userspace can already change the memslots after enabling the
> dirty ring, so `entry->slot` can already be stale, so dropping slots_lock for
> the cond_resched() seems harmless (and better than not dropping it).