On 9/10/19 6:56 AM, Wanpeng Li wrote: > On Mon, 9 Sep 2019 at 18:56, Waiman Long <longman@xxxxxxxxxx> wrote: >> On 9/9/19 2:40 AM, Wanpeng Li wrote: >>> From: Wanpeng Li <wanpengli@xxxxxxxxxxx> >>> >>> This patch reverts commit 75437bb304b20 (locking/pvqspinlock: Don't wait if >>> vCPU is preempted), we found great regression caused by this commit. >>> >>> Xeon Skylake box, 2 sockets, 40 cores, 80 threads, three VMs, each is 80 vCPUs. >>> The score of ebizzy -M can reduce from 13000-14000 records/s to 1700-1800 >>> records/s with this commit. >>> >>> Host Guest score >>> >>> vanilla + w/o kvm optimizes vanilla 1700-1800 records/s >>> vanilla + w/o kvm optimizes vanilla + revert 13000-14000 records/s >>> vanilla + w/ kvm optimizes vanilla 4500-5000 records/s >>> vanilla + w/ kvm optimizes vanilla + revert 14000-15500 records/s >>> >>> Exit from aggressive wait-early mechanism can result in yield premature and >>> incur extra scheduling latency in over-subscribe scenario. >>> >>> kvm optimizes: >>> [1] commit d73eb57b80b (KVM: Boost vCPUs that are delivering interrupts) >>> [2] commit 266e85a5ec9 (KVM: X86: Boost queue head vCPU to mitigate lock waiter preemption) >>> >>> Tested-by: loobinliu@xxxxxxxxxxx >>> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx> >>> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx> >>> Cc: Ingo Molnar <mingo@xxxxxxxxxx> >>> Cc: Waiman Long <longman@xxxxxxxxxx> >>> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx> >>> Cc: Radim Krčmář <rkrcmar@xxxxxxxxxx> >>> Cc: loobinliu@xxxxxxxxxxx >>> Cc: stable@xxxxxxxxxxxxxxx >>> Fixes: 75437bb304b20 (locking/pvqspinlock: Don't wait if vCPU is preempted) >>> Signed-off-by: Wanpeng Li <wanpengli@xxxxxxxxxxx> >>> --- >>> kernel/locking/qspinlock_paravirt.h | 2 +- >>> 1 file changed, 1 insertion(+), 1 deletion(-) >>> >>> diff --git a/kernel/locking/qspinlock_paravirt.h b/kernel/locking/qspinlock_paravirt.h >>> index 89bab07..e84d21a 100644 >>> --- a/kernel/locking/qspinlock_paravirt.h >>> +++ b/kernel/locking/qspinlock_paravirt.h >>> @@ -269,7 +269,7 @@ pv_wait_early(struct pv_node *prev, int loop) >>> if ((loop & PV_PREV_CHECK_MASK) != 0) >>> return false; >>> >>> - return READ_ONCE(prev->state) != vcpu_running || vcpu_is_preempted(prev->cpu); >>> + return READ_ONCE(prev->state) != vcpu_running; >>> } >>> >>> /* >> There are several possibilities for this performance regression: >> >> 1) Multiple vcpus calling vcpu_is_preempted() repeatedly may cause some >> cacheline contention issue depending on how that callback is implemented. >> >> 2) KVM may set the preempt flag for a short period whenver an vmexit >> happens even if a vmenter is executed shortly after. In this case, we >> may want to use a more durable vcpu suspend flag that indicates the vcpu >> won't get a real vcpu back for a longer period of time. >> >> Perhaps you can add a lock event counter to count the number of >> wait_early events caused by vcpu_is_preempted() being true to see if it >> really cause a lot more wait_early than without the vcpu_is_preempted() >> call. > pv_wait_again:1:179 > pv_wait_early:1:189429 > pv_wait_head:1:263 > pv_wait_node:1:189429 > pv_vcpu_is_preempted:1:45588 > =========sleep 5============ > pv_wait_again:1:181 > pv_wait_early:1:202574 > pv_wait_head:1:267 > pv_wait_node:1:202590 > pv_vcpu_is_preempted:1:46336 > > The sampling period is 5s, 6% of wait_early events caused by > vcpu_is_preempted() being true. 6% isn't that high. However, when one vCPU voluntarily releases its vCPU, all the subsequently waiters in the queue will do the same. It is a cascading effect. Perhaps we wait early too aggressive with the original patch. I also look up the email chain of the original commit. The patch submitter did not provide any performance data to support this change. The patch just looked reasonable at that time. So there was no objection. Given that we now have hard evidence that this was not a good idea. I think we should revert it. Reviewed-by: Waiman Long <longman@xxxxxxxxxx> Thanks, Longman
