On Mon, Nov 13, 2017 at 11:46:34AM +0100, Peter Zijlstra wrote:
> On Mon, Nov 13, 2017 at 04:26:57PM +0800, Wanpeng Li wrote:
> > 2017-11-13 16:04 GMT+08:00 Peter Zijlstra <peterz@xxxxxxxxxxxxx>:
>
> > > So if at this point a vCPU gets preempted we'll still spin-wait for it,
> > > which is sub-optimal.
> > >
> > > I think we can come up with something to get around that 'problem' if
> > > indeed it is a problem. But we can easily do that as follow up patches.
> > > Just let me know if you think its worth spending more time on.
> >
> > You can post your idea, it is always smart. :) Then we can evaluate
> > the complexity and gains.
>
> I'm not sure I have a fully baked idea just yet, but the general idea
> would be something like:
>
> - switch (back) to a dedicated TLB invalidate IPI
Just for PV that is; the !PV code can continue doing what it does today.
> - introduce KVM_VCPU_IPI_PENDING
>
> - change flush_tlb_others() into something like:
>
> for_each_cpu(cpu, flushmask) {
> src = &per_cpu(steal_time, cpu);
> state = READ_ONCE(src->preempted);
> do {
> if (state & KVM_VCPU_PREEMPTED) {
> if (try_cmpxchg(&src->preempted, &state,
> state | KVM_VCPU_SHOULD_FLUSH)) {
> __cpumask_clear_cpu(cpu, flushmask);
> break;
> }
> }
> } while (!try_cmpxchg(&src->preempted, &state,
> state | KVM_VCPU_IPI_PENDING));
That can be written like:
do {
if (state & KVM_VCPU_PREEMPTED)
new_state = state | KVM_VCPU_SHOULD_FLUSH;
else
new_state = state | KVM_VCPU_IPI_PENDING;
} while (!try_cmpxchg(&src->preempted, state, new_state);
if (new_state & KVM_VCPU_IPI_PENDING)
__cpumask_clear_cpu(cpu, flushmask);
> }
>
> apic->send_IPI_mask(flushmask, CALL_TLB_VECTOR);
>
> for_each_cpu(cpu, flushmask) {
> src = &per_cpu(steal_time, cpu);
/*
* The ACQUIRE pairs with the cmpxchg clearing IPI_PENDING,
* which is either the TLB IPI handler, or the VMEXIT path.
* It ensure that the invalidate happens-before.
*/
> smp_cond_load_acquire(&src->preempted, !(VAL & KVM_VCPU_IPI_PENDING);
> }
And here we wait for completion of the invalidate; but because of the
VMEXIT change below, this will never stall on a !running vCPU.
Note that PLE will not help (much) here, without this extra IPI_PENDING
state and the VMEXIT transferring it to SHOULD_FLUSH this vCPU's progress
will be held up until all vCPU's you've IPI'd will have ran the IPI
handler, which in the worst case is still a very long time.
> - have the TLB invalidate handler do something like:
>
> state = READ_ONCE(src->preempted);
> if (!(state & KVM_VCPU_IPI_PENDING))
> return;
>
> local_flush_tlb();
>
> do {
> } while (!try_cmpxchg(&src->preempted, &state,
> state & ~KVM_VCPU_IPI_PENDING));
That needs to be:
/*
* Clear KVM_VCPU_IPI_PENDING to 'complete' flush_tlb_others().
*/
do {
/*
* VMEXIT could have cleared this for us, in which case
* we're done.
*/
if (!(state & KVM_VCPU_IPI_PENDING))
return;
} while (!try_cmpxchg(&src->preempted, state,
state & ~KVM_VCPU_IPI_PENDING));
> - then at VMEXIT time do something like:
>
/*
* If we have IPI_PENDING set at VMEXIT time, transfer it to
* SHOULD_FLUSH. Clearing IPI_PENDING here allows the
* flush_others() vCPU to continue while the SHOULD_FLUSH
* guarantees this vCPU will flush TLBs before it continues
* execution.
*/
> state = READ_ONCE(src->preempted);
> do {
> if (!(state & KVM_VCPU_IPI_PENDING))
> break;
> } while (!try_cmpxchg(&src->preempted, state,
> (state & ~KVM_VCPU_IPI_PENDING) |
> KVM_VCPU_SHOULD_FLUSH));
>
> and clear any possible pending TLB_VECTOR in the guest state to avoid
> raising that IPI spuriously on enter again.
>
