Re: [PATCH V5 0/2] selftests: KVM: Add a test for eager page splitting

[Paolo Bonzini]

On 3/15/23 20:22, Sean Christopherson wrote:
> On Wed, Mar 15, 2023, Paolo Bonzini wrote:
> > On 3/15/23 13:24, Paolo Bonzini wrote:
> > > On Tue, Mar 14, 2023 at 5:00 PM David Matlack <dmatlack@xxxxxxxxxx> wrote:
> > > > I wonder if pages are getting swapped, especially if running on a
> > > > workstation. If so, mlock()ing all guest memory VMAs might be
> > > > necessary to be able to assert exact page counts.
> > >
> > > I don't think so, it's 100% reproducible and the machine is idle and
> > > only accessed via network. Also has 64 GB of RAM. :)
> >
> > It also reproduces on Intel with pml=0 and eptad=0; the reason is due
> > to the different semantics of dirty bits for page-table pages on AMD
> > and Intel.  Both AMD and eptad=0 Intel treat those as writes, therefore
> > more pages are dropped before the repopulation phase when dirty logging
> > is disabled.
> >
> > The "missing" page had been included in the population phase because it
> > hosts the page tables for vcpu_args, but repopulation does not need it.
> >
> > This fixes it:
> >
> > -------------------- 8< ---------------
> > From: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> > Subject: [PATCH] selftests: KVM: perform the same memory accesses on every memstress iteration
> >
> > Perform the same memory accesses including the initialization steps
> > that read from args and vcpu_args.  This ensures that the state of
> > KVM's page tables is the same after every iteration, including the
> > pages that host the guest page tables for args and vcpu_args.
> >
> > This fixes a failure of dirty_log_page_splitting_test on AMD machines,
> > as well as on Intel if PML and EPT A/D bits are both disabled.
> >
> > Signed-off-by: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> >
> > diff --git a/tools/testing/selftests/kvm/lib/memstress.c b/tools/testing/selftests/kvm/lib/memstress.c
> > index 3632956c6bcf..8a429f4c86db 100644
> > --- a/tools/testing/selftests/kvm/lib/memstress.c
> > +++ b/tools/testing/selftests/kvm/lib/memstress.c
> > @@ -56,15 +56,15 @@ void memstress_guest_code(uint32_t vcpu_idx)
> >   	uint64_t page;
> >   	int i;
> > -	rand_state = new_guest_random_state(args->random_seed + vcpu_idx);
> > +	while (true) {
> > +		rand_state = new_guest_random_state(args->random_seed + vcpu_idx);
>
> Doesn't this partially defeat the randomization that some tests like want?  E.g.
> a test that wants to heavily randomize state will get the same pRNG for every
> iteration.  Seems like we should have a knob to control whether or not each
> iteration needs to be identical.

Yes, this wasn't really a full patch, just to prove what the bug is.

One possibility to avoid adding a new knob is to do something like:

unsigned iteration = 0;
rand_state = new_guest_random_state(args->random_seed
	+ vcpu_idx + iteration++);

Paolo