On Thu, Dec 19, 2024, Maxim Levitsky wrote:
> On Wed, 2024-12-18 at 18:00 -0800, Sean Christopherson wrote:
> > On Tue, Dec 17, 2024, Maxim Levitsky wrote:
> > > On Fri, 2024-12-13 at 17:07 -0800, Sean Christopherson wrote:
> > > > Now that the vCPU doesn't dirty every page on the first iteration for
> > > > architectures that support the dirty ring, honor vcpu_stop in the dirty
> > > > ring's vCPU worker, i.e. stop when the main thread says "stop". This will
> > > > allow plumbing vcpu_stop into the guest so that the vCPU doesn't need to
> > > > periodically exit to userspace just to see if it should stop.
> > >
> > > This is very misleading - by the very nature of this test it all runs in
> > > userspace, so every time KVM_RUN ioctl exits, it is by definition an
> > > userspace VM exit.
> >
> > I honestly don't see how being more precise is misleading.
>
> "Exit to userspace" is misleading - the *whole test* is userspace.
No, the test has a guest component. Just because the host portion of the test
only runs in userspace doesn't make KVM go away. If this were pure emulation,
then I would completely agree, but there multiple distinct components here, one
of which is host userspace.
> You treat vCPU worker thread as if it not userspace, but it is *userspace* by
> the definition of how KVM works.
By simply "vCPU" I am strictly referring to the guest entity. I refered to the
host side worker as "vCPU woker" to try to distinguish between the two.
> Right way to say it is something like 'don't pause the vCPU worker thread
> when its not needed' or something like that.
That's inaccurate though. GUEST_SYNC() doesn't pause the vCPU, it forces it to
exit to userspace. The test forces the vCPU to exit to check to see if it needs
to pause/stop, which I'm contending is wasteful and unnecessarily complex. The
vCPU can instead check to see if it needs to stop simply by reading the global
variable.
If vcpu_sync_stop_requested is false, the worker thread immediated resumes the
vCPU.
/* Should only be called after a GUEST_SYNC */
static void vcpu_handle_sync_stop(void)
{
if (atomic_read(&vcpu_sync_stop_requested)) {
/* It means main thread is sleeping waiting */
atomic_set(&vcpu_sync_stop_requested, false);
sem_post(&sem_vcpu_stop);
sem_wait_until(&sem_vcpu_cont);
}
}
The future cleanup is to change the guest loop to keep running _in the guest_
until a stop is requested. Whereas the current code exits to userspace every
4096 writes to see if it should stop. But as above, the vCPU doesn't actually
stop on each exit.
@@ -112,7 +111,7 @@ static void guest_code(void)
#endif
while (true) {
- for (i = 0; i < TEST_PAGES_PER_LOOP; i++) {
+ while (!READ_ONCE(vcpu_stop)) {
addr = guest_test_virt_mem;
addr += (guest_random_u64(&guest_rng) % guest_num_pages)
* guest_page_size;
