| Message ID | 20221104062105.4119003-1-usama.arif@bytedance.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | KVM: arm64: implement vcpu_is_preempted check | expand |
On Fri, 04 Nov 2022 06:20:59 +0000, Usama Arif <usama.arif@bytedance.com> wrote: > > This patchset adds support for vcpu_is_preempted in arm64, which allows the guest > to check if a vcpu was scheduled out, which is useful to know incase it was > holding a lock. vcpu_is_preempted can be used to improve > performance in locking (see owner_on_cpu usage in mutex_spin_on_owner, > mutex_can_spin_on_owner, rtmutex_spin_on_owner and osq_lock) and scheduling > (see available_idle_cpu which is used in several places in kernel/sched/fair.c > for e.g. in wake_affine to determine which CPU can run soonest): Please refrain from reposting a series only two days after the initial one. One week is a minimum, and only if there is enough review comments to justify a respin (there were no valuable comments so far). Reposting more often only results in the review process being exponentially delayed. Thanks, M.
On Fri, 04 Nov 2022 06:20:59 +0000, Usama Arif <usama.arif@bytedance.com> wrote: > > This patchset adds support for vcpu_is_preempted in arm64, which > allows the guest to check if a vcpu was scheduled out, which is > useful to know incase it was holding a lock. vcpu_is_preempted can > be used to improve performance in locking (see owner_on_cpu usage in > mutex_spin_on_owner, mutex_can_spin_on_owner, rtmutex_spin_on_owner > and osq_lock) and scheduling (see available_idle_cpu which is used > in several places in kernel/sched/fair.c for e.g. in wake_affine to > determine which CPU can run soonest): [...] > pvcy shows a smaller overall improvement (50%) compared to > vcpu_is_preempted (277%). Host side flamegraph analysis shows that > ~60% of the host time when using pvcy is spent in kvm_handle_wfx, > compared with ~1.5% when using vcpu_is_preempted, hence > vcpu_is_preempted shows a larger improvement. And have you worked out *why* we spend so much time handling WFE? M.
On 06/11/2022 16:35, Marc Zyngier wrote: > On Fri, 04 Nov 2022 06:20:59 +0000, > Usama Arif <usama.arif@bytedance.com> wrote: >> >> This patchset adds support for vcpu_is_preempted in arm64, which >> allows the guest to check if a vcpu was scheduled out, which is >> useful to know incase it was holding a lock. vcpu_is_preempted can >> be used to improve performance in locking (see owner_on_cpu usage in >> mutex_spin_on_owner, mutex_can_spin_on_owner, rtmutex_spin_on_owner >> and osq_lock) and scheduling (see available_idle_cpu which is used >> in several places in kernel/sched/fair.c for e.g. in wake_affine to >> determine which CPU can run soonest): > > [...] > >> pvcy shows a smaller overall improvement (50%) compared to >> vcpu_is_preempted (277%). Host side flamegraph analysis shows that >> ~60% of the host time when using pvcy is spent in kvm_handle_wfx, >> compared with ~1.5% when using vcpu_is_preempted, hence >> vcpu_is_preempted shows a larger improvement. > > And have you worked out *why* we spend so much time handling WFE? > > M. Its from the following change in pvcy patchset: diff --git a/arch/arm64/kvm/handle_exit.c b/arch/arm64/kvm/handle_exit.c index e778eefcf214..915644816a85 100644 --- a/arch/arm64/kvm/handle_exit.c +++ b/arch/arm64/kvm/handle_exit.c @@ -118,7 +118,12 @@ static int kvm_handle_wfx(struct kvm_vcpu *vcpu) } if (esr & ESR_ELx_WFx_ISS_WFE) { - kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); + int state; + while ((state = kvm_pvcy_check_state(vcpu)) == 0) + schedule(); + + if (state == -1) + kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); } else { if (esr & ESR_ELx_WFx_ISS_WFxT) vcpu_set_flag(vcpu, IN_WFIT); If my understanding is correct of the pvcy changes, whenever pvcy returns an unchanged vcpu state, we would schedule to another vcpu. And its the constant scheduling where the time is spent. I guess the affects are much higher when the lock contention is very high. This can be seem from the pvcy host side flamegraph as well with (~67% of the time spent in the schedule() call in kvm_handle_wfx), For reference, I have put the graph at: https://uarif1.github.io/pvlock/perf_host_pvcy_nmi.svg Thanks, Usama >
Hi Usama, Usama Arif <usama.arif@bytedance.com> writes: > This patchset adds support for vcpu_is_preempted in arm64, which allows the guest > to check if a vcpu was scheduled out, which is useful to know incase it was > holding a lock. vcpu_is_preempted can be used to improve > performance in locking (see owner_on_cpu usage in mutex_spin_on_owner, > mutex_can_spin_on_owner, rtmutex_spin_on_owner and osq_lock) and scheduling > (see available_idle_cpu which is used in several places in kernel/sched/fair.c > for e.g. in wake_affine to determine which CPU can run soonest): > > This patchset shows improvement on overcommitted hosts (vCPUs > pCPUS), as waiting > for preempted vCPUs reduces performance. > > This patchset is inspired from the para_steal_clock implementation and from the > work originally done by Zengruan Ye: > https://lore.kernel.org/linux-arm-kernel/20191226135833.1052-1-yezengruan@huawei.com/. > > All the results in the below experiments are done on an aws r6g.metal instance > which has 64 pCPUs. > > The following table shows the index results of UnixBench running on a 128 vCPU VM > with (6.0.0+vcpu_is_preempted) and without (6.0.0 base) the patchset. > TestName 6.0.0 base 6.0.0+vcpu_is_preempted % improvement for vcpu_is_preempted > Dhrystone 2 using register variables 187761 191274.7 1.871368389 > Double-Precision Whetstone 96743.6 98414.4 1.727039308 > Execl Throughput 689.3 10426 1412.548963 > File Copy 1024 bufsize 2000 maxblocks 549.5 3165 475.978162 > File Copy 256 bufsize 500 maxblocks 400.7 2084.7 420.2645371 > File Copy 4096 bufsize 8000 maxblocks 894.3 5003.2 459.4543218 > Pipe Throughput 76819.5 78601.5 2.319723508 > Pipe-based Context Switching 3444.8 13414.5 289.4130283 > Process Creation 301.1 293.4 -2.557289937 > Shell Scripts (1 concurrent) 1248.1 28300.6 2167.494592 > Shell Scripts (8 concurrent) 781.2 26222.3 3256.669227 > System Call Overhead 3426 3729.4 8.855808523 > > System Benchmarks Index Score 3053 11534 277.7923354 > > This shows a 277% overall improvement using these patches. > > The biggest improvement is in the shell scripts benchmark, which forks a lot of processes. > This acquires rwsem lock where a large chunk of time is spent in base 6.0.0 kernel. > This can be seen from one of the callstack of the perf output of the shell > scripts benchmark on 6.0.0 base (pseudo NMI enabled for perf numbers below): > - 33.79% el0_svc > - 33.43% do_el0_svc > - 33.43% el0_svc_common.constprop.3 > - 33.30% invoke_syscall > - 17.27% __arm64_sys_clone > - 17.27% __do_sys_clone > - 17.26% kernel_clone > - 16.73% copy_process > - 11.91% dup_mm > - 11.82% dup_mmap > - 9.15% down_write > - 8.87% rwsem_down_write_slowpath > - 8.48% osq_lock > > Just under 50% of the total time in the shell script benchmarks ends up being > spent in osq_lock in the base 6.0.0 kernel: > Children Self Command Shared Object Symbol > 17.19% 10.71% sh [kernel.kallsyms] [k] osq_lock > 6.17% 4.04% sort [kernel.kallsyms] [k] osq_lock > 4.20% 2.60% multi. [kernel.kallsyms] [k] osq_lock > 3.77% 2.47% grep [kernel.kallsyms] [k] osq_lock > 3.50% 2.24% expr [kernel.kallsyms] [k] osq_lock > 3.41% 2.23% od [kernel.kallsyms] [k] osq_lock > 3.36% 2.15% rm [kernel.kallsyms] [k] osq_lock > 3.28% 2.12% tee [kernel.kallsyms] [k] osq_lock > 3.16% 2.02% wc [kernel.kallsyms] [k] osq_lock > 0.21% 0.13% looper [kernel.kallsyms] [k] osq_lock > 0.01% 0.00% Run [kernel.kallsyms] [k] osq_lock > > and this comes down to less than 1% total with 6.0.0+vcpu_is_preempted kernel: > Children Self Command Shared Object Symbol > 0.26% 0.21% sh [kernel.kallsyms] [k] osq_lock > 0.10% 0.08% multi. [kernel.kallsyms] [k] osq_lock > 0.04% 0.04% sort [kernel.kallsyms] [k] osq_lock > 0.02% 0.01% grep [kernel.kallsyms] [k] osq_lock > 0.02% 0.02% od [kernel.kallsyms] [k] osq_lock > 0.01% 0.01% tee [kernel.kallsyms] [k] osq_lock > 0.01% 0.00% expr [kernel.kallsyms] [k] osq_lock > 0.01% 0.01% looper [kernel.kallsyms] [k] osq_lock > 0.00% 0.00% wc [kernel.kallsyms] [k] osq_lock > 0.00% 0.00% rm [kernel.kallsyms] [k] osq_lock > > To make sure, there is no change in performance when vCPUs < pCPUs, UnixBench > was run on a 32 CPU VM. The kernel with vcpu_is_preempted implemented > performed 0.9% better overall than base kernel, and the individual benchmarks > were within +/-2% improvement over 6.0.0 base. > Hence the patches have no negative affect when vCPUs < pCPUs. > > > The other method discussed in https://lore.kernel.org/linux-arm-kernel/20191226135833.1052-1-yezengruan@huawei.com/ > was pv conditional yield by Marc Zyngier and Will Deacon to reduce vCPU reschedule > if the vCPU will exit immediately. > (https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pvcy). > The patches were ported to 6.0.0 kernel and tested with UnixBench with a 128 vCPU VM: > > TestName 6.0.0 base 6.0.0+pvcy % improvement for pvcy > Dhrystone 2 using register variables 187761 183128 -2.467498575 > Double-Precision Whetstone 96743.6 96645 -0.101918887 > Execl Throughput 689.3 1019.8 47.9471928 > File Copy 1024 bufsize 2000 maxblocks 549.5 2029.7 269.3721565 > File Copy 256 bufsize 500 maxblocks 400.7 1439.4 259.2213626 > File Copy 4096 bufsize 8000 maxblocks 894.3 3434.1 283.9986582 > Pipe Throughput 76819.5 74268.8 -3.320380893 > Pipe-based Context Switching 3444.8 5963.3 73.11019508 > Process Creation 301.1 163.2 -45.79873796 > Shell Scripts (1 concurrent) 1248.1 1859.7 49.00248378 > Shell Scripts (8 concurrent) 781.2 1171 49.89759345 > System Call Overhead 3426 3194.4 -6.760070053 > > System Benchmarks Index Score 3053 4605 50.83524402 > > pvcy shows a smaller overall improvement (50%) compared to vcpu_is_preempted (277%). > Host side flamegraph analysis shows that ~60% of the host time when using pvcy > is spent in kvm_handle_wfx, compared with ~1.5% when using vcpu_is_preempted, > hence vcpu_is_preempted shows a larger improvement. > > It might be that pvcy can be used in combination with vcpu_is_preempted, but this > series is to start a discussion on vcpu_is_preempted as it shows a much bigger > improvement in performance and its much easier to review vcpu_is_preempted standalone. Looking at both the patchsets - this one and the pvcy, it looks to me that vcpu_is_preempted() and the pvcy patches are somewhat orthogonal. The former is optimizing mutex and rwsem in their optimistic spinning phase while the latter is going after spinlocks (via wfe). Unless I'm missing something the features are not necessarily comparable on the same workloads - unixbench is probably mutex heavy and doesn't show as much benefit with just the pvcy changes. I wonder if it's easy to have both the features enabled to see this in effect. I've left some comments on the patches; but no need to respin just yet. Let's see if there is any other feedback. Thanks, Punit [...]
On 07/11/2022 18:24, Punit Agrawal wrote: > Hi Usama, > > Usama Arif <usama.arif@bytedance.com> writes: > >> This patchset adds support for vcpu_is_preempted in arm64, which allows the guest >> to check if a vcpu was scheduled out, which is useful to know incase it was >> holding a lock. vcpu_is_preempted can be used to improve >> performance in locking (see owner_on_cpu usage in mutex_spin_on_owner, >> mutex_can_spin_on_owner, rtmutex_spin_on_owner and osq_lock) and scheduling >> (see available_idle_cpu which is used in several places in kernel/sched/fair.c >> for e.g. in wake_affine to determine which CPU can run soonest): >> >> This patchset shows improvement on overcommitted hosts (vCPUs > pCPUS), as waiting >> for preempted vCPUs reduces performance. >> >> This patchset is inspired from the para_steal_clock implementation and from the >> work originally done by Zengruan Ye: >> https://lore.kernel.org/linux-arm-kernel/20191226135833.1052-1-yezengruan@huawei.com/. >> >> All the results in the below experiments are done on an aws r6g.metal instance >> which has 64 pCPUs. >> >> The following table shows the index results of UnixBench running on a 128 vCPU VM >> with (6.0.0+vcpu_is_preempted) and without (6.0.0 base) the patchset. >> TestName 6.0.0 base 6.0.0+vcpu_is_preempted % improvement for vcpu_is_preempted >> Dhrystone 2 using register variables 187761 191274.7 1.871368389 >> Double-Precision Whetstone 96743.6 98414.4 1.727039308 >> Execl Throughput 689.3 10426 1412.548963 >> File Copy 1024 bufsize 2000 maxblocks 549.5 3165 475.978162 >> File Copy 256 bufsize 500 maxblocks 400.7 2084.7 420.2645371 >> File Copy 4096 bufsize 8000 maxblocks 894.3 5003.2 459.4543218 >> Pipe Throughput 76819.5 78601.5 2.319723508 >> Pipe-based Context Switching 3444.8 13414.5 289.4130283 >> Process Creation 301.1 293.4 -2.557289937 >> Shell Scripts (1 concurrent) 1248.1 28300.6 2167.494592 >> Shell Scripts (8 concurrent) 781.2 26222.3 3256.669227 >> System Call Overhead 3426 3729.4 8.855808523 >> >> System Benchmarks Index Score 3053 11534 277.7923354 >> >> This shows a 277% overall improvement using these patches. >> >> The biggest improvement is in the shell scripts benchmark, which forks a lot of processes. >> This acquires rwsem lock where a large chunk of time is spent in base 6.0.0 kernel. >> This can be seen from one of the callstack of the perf output of the shell >> scripts benchmark on 6.0.0 base (pseudo NMI enabled for perf numbers below): >> - 33.79% el0_svc >> - 33.43% do_el0_svc >> - 33.43% el0_svc_common.constprop.3 >> - 33.30% invoke_syscall >> - 17.27% __arm64_sys_clone >> - 17.27% __do_sys_clone >> - 17.26% kernel_clone >> - 16.73% copy_process >> - 11.91% dup_mm >> - 11.82% dup_mmap >> - 9.15% down_write >> - 8.87% rwsem_down_write_slowpath >> - 8.48% osq_lock >> >> Just under 50% of the total time in the shell script benchmarks ends up being >> spent in osq_lock in the base 6.0.0 kernel: >> Children Self Command Shared Object Symbol >> 17.19% 10.71% sh [kernel.kallsyms] [k] osq_lock >> 6.17% 4.04% sort [kernel.kallsyms] [k] osq_lock >> 4.20% 2.60% multi. [kernel.kallsyms] [k] osq_lock >> 3.77% 2.47% grep [kernel.kallsyms] [k] osq_lock >> 3.50% 2.24% expr [kernel.kallsyms] [k] osq_lock >> 3.41% 2.23% od [kernel.kallsyms] [k] osq_lock >> 3.36% 2.15% rm [kernel.kallsyms] [k] osq_lock >> 3.28% 2.12% tee [kernel.kallsyms] [k] osq_lock >> 3.16% 2.02% wc [kernel.kallsyms] [k] osq_lock >> 0.21% 0.13% looper [kernel.kallsyms] [k] osq_lock >> 0.01% 0.00% Run [kernel.kallsyms] [k] osq_lock >> >> and this comes down to less than 1% total with 6.0.0+vcpu_is_preempted kernel: >> Children Self Command Shared Object Symbol >> 0.26% 0.21% sh [kernel.kallsyms] [k] osq_lock >> 0.10% 0.08% multi. [kernel.kallsyms] [k] osq_lock >> 0.04% 0.04% sort [kernel.kallsyms] [k] osq_lock >> 0.02% 0.01% grep [kernel.kallsyms] [k] osq_lock >> 0.02% 0.02% od [kernel.kallsyms] [k] osq_lock >> 0.01% 0.01% tee [kernel.kallsyms] [k] osq_lock >> 0.01% 0.00% expr [kernel.kallsyms] [k] osq_lock >> 0.01% 0.01% looper [kernel.kallsyms] [k] osq_lock >> 0.00% 0.00% wc [kernel.kallsyms] [k] osq_lock >> 0.00% 0.00% rm [kernel.kallsyms] [k] osq_lock >> >> To make sure, there is no change in performance when vCPUs < pCPUs, UnixBench >> was run on a 32 CPU VM. The kernel with vcpu_is_preempted implemented >> performed 0.9% better overall than base kernel, and the individual benchmarks >> were within +/-2% improvement over 6.0.0 base. >> Hence the patches have no negative affect when vCPUs < pCPUs. >> >> >> The other method discussed in https://lore.kernel.org/linux-arm-kernel/20191226135833.1052-1-yezengruan@huawei.com/ >> was pv conditional yield by Marc Zyngier and Will Deacon to reduce vCPU reschedule >> if the vCPU will exit immediately. >> (https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pvcy). >> The patches were ported to 6.0.0 kernel and tested with UnixBench with a 128 vCPU VM: >> >> TestName 6.0.0 base 6.0.0+pvcy % improvement for pvcy >> Dhrystone 2 using register variables 187761 183128 -2.467498575 >> Double-Precision Whetstone 96743.6 96645 -0.101918887 >> Execl Throughput 689.3 1019.8 47.9471928 >> File Copy 1024 bufsize 2000 maxblocks 549.5 2029.7 269.3721565 >> File Copy 256 bufsize 500 maxblocks 400.7 1439.4 259.2213626 >> File Copy 4096 bufsize 8000 maxblocks 894.3 3434.1 283.9986582 >> Pipe Throughput 76819.5 74268.8 -3.320380893 >> Pipe-based Context Switching 3444.8 5963.3 73.11019508 >> Process Creation 301.1 163.2 -45.79873796 >> Shell Scripts (1 concurrent) 1248.1 1859.7 49.00248378 >> Shell Scripts (8 concurrent) 781.2 1171 49.89759345 >> System Call Overhead 3426 3194.4 -6.760070053 >> >> System Benchmarks Index Score 3053 4605 50.83524402 >> >> pvcy shows a smaller overall improvement (50%) compared to vcpu_is_preempted (277%). >> Host side flamegraph analysis shows that ~60% of the host time when using pvcy >> is spent in kvm_handle_wfx, compared with ~1.5% when using vcpu_is_preempted, >> hence vcpu_is_preempted shows a larger improvement. >> >> It might be that pvcy can be used in combination with vcpu_is_preempted, but this >> series is to start a discussion on vcpu_is_preempted as it shows a much bigger >> improvement in performance and its much easier to review vcpu_is_preempted standalone. > > Looking at both the patchsets - this one and the pvcy, it looks to me > that vcpu_is_preempted() and the pvcy patches are somewhat > orthogonal. The former is optimizing mutex and rwsem in their optimistic > spinning phase while the latter is going after spinlocks (via wfe). > > Unless I'm missing something the features are not necessarily comparable > on the same workloads - unixbench is probably mutex heavy and doesn't > show as much benefit with just the pvcy changes. I wonder if it's easy > to have both the features enabled to see this in effect. > > I've left some comments on the patches; but no need to respin just > yet. Let's see if there is any other feedback. > > Thanks, > Punit > There was a small bug in v2, where pv_lock_init was called too early in the boot in setup_arch, hence pvlock_vcpu_state was not initialized for vCPU 0 (the state was initialized for vCPUs 1-127 during secondary core boot, hence the rest of the vCPUs were using pvlock correctly). I will send the next revision making it an early_initcall along with addressing Punits' comments, but will wait for further comments on v2 before sending it. I have tested it with early_initcall and didnt see a significant change in performance (which is expected as only 1 out of 128 vCPUs wasnt using pvlock correctly). I tried pvcy+vcpu_is_preempted patches together and I see a slight reduction in performance over pvcy only. As a summary, with the above changes to move to early_initcall included the overall Unixbench score improvements are: Change over 6.0.0 base kernel % improvement over base vcpu_is_preempted 279% pvcy 51% pvcy+vcpu_is_preempted 37% Thanks, Usama > [...] >
On Mon, 07 Nov 2022 12:00:44 +0000, Usama Arif <usama.arif@bytedance.com> wrote: > > > > On 06/11/2022 16:35, Marc Zyngier wrote: > > On Fri, 04 Nov 2022 06:20:59 +0000, > > Usama Arif <usama.arif@bytedance.com> wrote: > >> > >> This patchset adds support for vcpu_is_preempted in arm64, which > >> allows the guest to check if a vcpu was scheduled out, which is > >> useful to know incase it was holding a lock. vcpu_is_preempted can > >> be used to improve performance in locking (see owner_on_cpu usage in > >> mutex_spin_on_owner, mutex_can_spin_on_owner, rtmutex_spin_on_owner > >> and osq_lock) and scheduling (see available_idle_cpu which is used > >> in several places in kernel/sched/fair.c for e.g. in wake_affine to > >> determine which CPU can run soonest): > > > > [...] > > > >> pvcy shows a smaller overall improvement (50%) compared to > >> vcpu_is_preempted (277%). Host side flamegraph analysis shows that > >> ~60% of the host time when using pvcy is spent in kvm_handle_wfx, > >> compared with ~1.5% when using vcpu_is_preempted, hence > >> vcpu_is_preempted shows a larger improvement. > > > > And have you worked out *why* we spend so much time handling WFE? > > > > M. > > Its from the following change in pvcy patchset: > > diff --git a/arch/arm64/kvm/handle_exit.c b/arch/arm64/kvm/handle_exit.c > index e778eefcf214..915644816a85 100644 > --- a/arch/arm64/kvm/handle_exit.c > +++ b/arch/arm64/kvm/handle_exit.c > @@ -118,7 +118,12 @@ static int kvm_handle_wfx(struct kvm_vcpu *vcpu) > } > > if (esr & ESR_ELx_WFx_ISS_WFE) { > - kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); > + int state; > + while ((state = kvm_pvcy_check_state(vcpu)) == 0) > + schedule(); > + > + if (state == -1) > + kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); > } else { > if (esr & ESR_ELx_WFx_ISS_WFxT) > vcpu_set_flag(vcpu, IN_WFIT); > > > If my understanding is correct of the pvcy changes, whenever pvcy > returns an unchanged vcpu state, we would schedule to another > vcpu. And its the constant scheduling where the time is spent. I guess > the affects are much higher when the lock contention is very > high. This can be seem from the pvcy host side flamegraph as well with > (~67% of the time spent in the schedule() call in kvm_handle_wfx), For > reference, I have put the graph at: > https://uarif1.github.io/pvlock/perf_host_pvcy_nmi.svg The real issue here is that we don't try to pick the right vcpu to run, and strictly rely on schedule() to eventually pick something that can run. An interesting to do would be to try and fit the directed yield mechanism there. It would be a lot more interesting than the one-off vcpu_is_preempted hack, as it gives us a low-level primitive on which to construct things (pvcy is effectively a mwait-like primitive). M.
On 18/11/2022 00:20, Marc Zyngier wrote: > On Mon, 07 Nov 2022 12:00:44 +0000, > Usama Arif <usama.arif@bytedance.com> wrote: >> >> >> >> On 06/11/2022 16:35, Marc Zyngier wrote: >>> On Fri, 04 Nov 2022 06:20:59 +0000, >>> Usama Arif <usama.arif@bytedance.com> wrote: >>>> >>>> This patchset adds support for vcpu_is_preempted in arm64, which >>>> allows the guest to check if a vcpu was scheduled out, which is >>>> useful to know incase it was holding a lock. vcpu_is_preempted can >>>> be used to improve performance in locking (see owner_on_cpu usage in >>>> mutex_spin_on_owner, mutex_can_spin_on_owner, rtmutex_spin_on_owner >>>> and osq_lock) and scheduling (see available_idle_cpu which is used >>>> in several places in kernel/sched/fair.c for e.g. in wake_affine to >>>> determine which CPU can run soonest): >>> >>> [...] >>> >>>> pvcy shows a smaller overall improvement (50%) compared to >>>> vcpu_is_preempted (277%). Host side flamegraph analysis shows that >>>> ~60% of the host time when using pvcy is spent in kvm_handle_wfx, >>>> compared with ~1.5% when using vcpu_is_preempted, hence >>>> vcpu_is_preempted shows a larger improvement. >>> >>> And have you worked out *why* we spend so much time handling WFE? >>> >>> M. >> >> Its from the following change in pvcy patchset: >> >> diff --git a/arch/arm64/kvm/handle_exit.c b/arch/arm64/kvm/handle_exit.c >> index e778eefcf214..915644816a85 100644 >> --- a/arch/arm64/kvm/handle_exit.c >> +++ b/arch/arm64/kvm/handle_exit.c >> @@ -118,7 +118,12 @@ static int kvm_handle_wfx(struct kvm_vcpu *vcpu) >> } >> >> if (esr & ESR_ELx_WFx_ISS_WFE) { >> - kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); >> + int state; >> + while ((state = kvm_pvcy_check_state(vcpu)) == 0) >> + schedule(); >> + >> + if (state == -1) >> + kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); >> } else { >> if (esr & ESR_ELx_WFx_ISS_WFxT) >> vcpu_set_flag(vcpu, IN_WFIT); >> >> >> If my understanding is correct of the pvcy changes, whenever pvcy >> returns an unchanged vcpu state, we would schedule to another >> vcpu. And its the constant scheduling where the time is spent. I guess >> the affects are much higher when the lock contention is very >> high. This can be seem from the pvcy host side flamegraph as well with >> (~67% of the time spent in the schedule() call in kvm_handle_wfx), For >> reference, I have put the graph at: >> https://uarif1.github.io/pvlock/perf_host_pvcy_nmi.svg > > The real issue here is that we don't try to pick the right vcpu to > run, and strictly rely on schedule() to eventually pick something that > can run. > > An interesting to do would be to try and fit the directed yield > mechanism there. It would be a lot more interesting than the one-off > vcpu_is_preempted hack, as it gives us a low-level primitive on which > to construct things (pvcy is effectively a mwait-like primitive). We could use kvm_vcpu_yield_to to yield to a specific vcpu, but how would we determine which vcpu to yield to? IMO vcpu_is_preempted is very well integrated in a lot of core kernel code, i.e. mutex, rtmutex, rwsem and osq_lock. It is also used in scheduler to determine better which vCPU we can run on soonest, select idle core, etc. I am not sure if all of these cases will be optimized by pvcy? Also, with vcpu_is_preempted, some of the lock heavy benchmarks come down from spending around 50% of the time in lock to less than 1% (so not sure how much more room is there for improvement). We could also use vcpu_is_preempted to optimize IPI performance (along with directed yield to target IPI vCPU) similar to how its done in x86 (https://lore.kernel.org/all/1560255830-8656-2-git-send-email-wanpengli@tencent.com/). This case definitely wont be covered by pvcy. Considering all the above, i.e. the core kernel integration already present and possible future usecases of vcpu_is_preempted, maybe its worth making vcpu_is_preempted work on arm independently of pvcy? Thanks, Usama > > M. >
On 24/11/2022 13:55, Usama Arif wrote: > > > On 18/11/2022 00:20, Marc Zyngier wrote: >> On Mon, 07 Nov 2022 12:00:44 +0000, >> Usama Arif <usama.arif@bytedance.com> wrote: >>> >>> >>> >>> On 06/11/2022 16:35, Marc Zyngier wrote: >>>> On Fri, 04 Nov 2022 06:20:59 +0000, >>>> Usama Arif <usama.arif@bytedance.com> wrote: >>>>> >>>>> This patchset adds support for vcpu_is_preempted in arm64, which >>>>> allows the guest to check if a vcpu was scheduled out, which is >>>>> useful to know incase it was holding a lock. vcpu_is_preempted can >>>>> be used to improve performance in locking (see owner_on_cpu usage in >>>>> mutex_spin_on_owner, mutex_can_spin_on_owner, rtmutex_spin_on_owner >>>>> and osq_lock) and scheduling (see available_idle_cpu which is used >>>>> in several places in kernel/sched/fair.c for e.g. in wake_affine to >>>>> determine which CPU can run soonest): >>>> >>>> [...] >>>> >>>>> pvcy shows a smaller overall improvement (50%) compared to >>>>> vcpu_is_preempted (277%). Host side flamegraph analysis shows that >>>>> ~60% of the host time when using pvcy is spent in kvm_handle_wfx, >>>>> compared with ~1.5% when using vcpu_is_preempted, hence >>>>> vcpu_is_preempted shows a larger improvement. >>>> >>>> And have you worked out *why* we spend so much time handling WFE? >>>> >>>> M. >>> >>> Its from the following change in pvcy patchset: >>> >>> diff --git a/arch/arm64/kvm/handle_exit.c b/arch/arm64/kvm/handle_exit.c >>> index e778eefcf214..915644816a85 100644 >>> --- a/arch/arm64/kvm/handle_exit.c >>> +++ b/arch/arm64/kvm/handle_exit.c >>> @@ -118,7 +118,12 @@ static int kvm_handle_wfx(struct kvm_vcpu *vcpu) >>> } >>> >>> if (esr & ESR_ELx_WFx_ISS_WFE) { >>> - kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); >>> + int state; >>> + while ((state = kvm_pvcy_check_state(vcpu)) == 0) >>> + schedule(); >>> + >>> + if (state == -1) >>> + kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); >>> } else { >>> if (esr & ESR_ELx_WFx_ISS_WFxT) >>> vcpu_set_flag(vcpu, IN_WFIT); >>> >>> >>> If my understanding is correct of the pvcy changes, whenever pvcy >>> returns an unchanged vcpu state, we would schedule to another >>> vcpu. And its the constant scheduling where the time is spent. I guess >>> the affects are much higher when the lock contention is very >>> high. This can be seem from the pvcy host side flamegraph as well with >>> (~67% of the time spent in the schedule() call in kvm_handle_wfx), For >>> reference, I have put the graph at: >>> https://uarif1.github.io/pvlock/perf_host_pvcy_nmi.svg >> >> The real issue here is that we don't try to pick the right vcpu to >> run, and strictly rely on schedule() to eventually pick something that >> can run. >> >> An interesting to do would be to try and fit the directed yield >> mechanism there. It would be a lot more interesting than the one-off >> vcpu_is_preempted hack, as it gives us a low-level primitive on which >> to construct things (pvcy is effectively a mwait-like primitive). > > We could use kvm_vcpu_yield_to to yield to a specific vcpu, but how > would we determine which vcpu to yield to? > > IMO vcpu_is_preempted is very well integrated in a lot of core kernel > code, i.e. mutex, rtmutex, rwsem and osq_lock. It is also used in > scheduler to determine better which vCPU we can run on soonest, select > idle core, etc. I am not sure if all of these cases will be optimized by > pvcy? Also, with vcpu_is_preempted, some of the lock heavy benchmarks > come down from spending around 50% of the time in lock to less than 1% > (so not sure how much more room is there for improvement). > > We could also use vcpu_is_preempted to optimize IPI performance (along > with directed yield to target IPI vCPU) similar to how its done in x86 > (https://lore.kernel.org/all/1560255830-8656-2-git-send-email-wanpengli@tencent.com/). > This case definitely wont be covered by pvcy. > > Considering all the above, i.e. the core kernel integration already > present and possible future usecases of vcpu_is_preempted, maybe its > worth making vcpu_is_preempted work on arm independently of pvcy? > Hi, Just wanted to check if there are any comments on above? I can send a v3 with the doc and code fixes suggested in the earlier reviews if it makes sense? Thanks, Usama > Thanks, > Usama > >> >> M. >>
On 05/12/2022 13:43, Usama Arif wrote: > > > On 24/11/2022 13:55, Usama Arif wrote: >> >> >> On 18/11/2022 00:20, Marc Zyngier wrote: >>> On Mon, 07 Nov 2022 12:00:44 +0000, >>> Usama Arif <usama.arif@bytedance.com> wrote: >>>> >>>> >>>> >>>> On 06/11/2022 16:35, Marc Zyngier wrote: >>>>> On Fri, 04 Nov 2022 06:20:59 +0000, >>>>> Usama Arif <usama.arif@bytedance.com> wrote: >>>>>> >>>>>> This patchset adds support for vcpu_is_preempted in arm64, which >>>>>> allows the guest to check if a vcpu was scheduled out, which is >>>>>> useful to know incase it was holding a lock. vcpu_is_preempted can >>>>>> be used to improve performance in locking (see owner_on_cpu usage in >>>>>> mutex_spin_on_owner, mutex_can_spin_on_owner, rtmutex_spin_on_owner >>>>>> and osq_lock) and scheduling (see available_idle_cpu which is used >>>>>> in several places in kernel/sched/fair.c for e.g. in wake_affine to >>>>>> determine which CPU can run soonest): >>>>> >>>>> [...] >>>>> >>>>>> pvcy shows a smaller overall improvement (50%) compared to >>>>>> vcpu_is_preempted (277%). Host side flamegraph analysis shows that >>>>>> ~60% of the host time when using pvcy is spent in kvm_handle_wfx, >>>>>> compared with ~1.5% when using vcpu_is_preempted, hence >>>>>> vcpu_is_preempted shows a larger improvement. >>>>> >>>>> And have you worked out *why* we spend so much time handling WFE? >>>>> >>>>> M. >>>> >>>> Its from the following change in pvcy patchset: >>>> >>>> diff --git a/arch/arm64/kvm/handle_exit.c >>>> b/arch/arm64/kvm/handle_exit.c >>>> index e778eefcf214..915644816a85 100644 >>>> --- a/arch/arm64/kvm/handle_exit.c >>>> +++ b/arch/arm64/kvm/handle_exit.c >>>> @@ -118,7 +118,12 @@ static int kvm_handle_wfx(struct kvm_vcpu *vcpu) >>>> } >>>> >>>> if (esr & ESR_ELx_WFx_ISS_WFE) { >>>> - kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); >>>> + int state; >>>> + while ((state = kvm_pvcy_check_state(vcpu)) == 0) >>>> + schedule(); >>>> + >>>> + if (state == -1) >>>> + kvm_vcpu_on_spin(vcpu, vcpu_mode_priv(vcpu)); >>>> } else { >>>> if (esr & ESR_ELx_WFx_ISS_WFxT) >>>> vcpu_set_flag(vcpu, IN_WFIT); >>>> >>>> >>>> If my understanding is correct of the pvcy changes, whenever pvcy >>>> returns an unchanged vcpu state, we would schedule to another >>>> vcpu. And its the constant scheduling where the time is spent. I guess >>>> the affects are much higher when the lock contention is very >>>> high. This can be seem from the pvcy host side flamegraph as well with >>>> (~67% of the time spent in the schedule() call in kvm_handle_wfx), For >>>> reference, I have put the graph at: >>>> https://uarif1.github.io/pvlock/perf_host_pvcy_nmi.svg >>> >>> The real issue here is that we don't try to pick the right vcpu to >>> run, and strictly rely on schedule() to eventually pick something that >>> can run. >>> >>> An interesting to do would be to try and fit the directed yield >>> mechanism there. It would be a lot more interesting than the one-off >>> vcpu_is_preempted hack, as it gives us a low-level primitive on which >>> to construct things (pvcy is effectively a mwait-like primitive). >> >> We could use kvm_vcpu_yield_to to yield to a specific vcpu, but how >> would we determine which vcpu to yield to? >> >> IMO vcpu_is_preempted is very well integrated in a lot of core kernel >> code, i.e. mutex, rtmutex, rwsem and osq_lock. It is also used in >> scheduler to determine better which vCPU we can run on soonest, select >> idle core, etc. I am not sure if all of these cases will be optimized >> by pvcy? Also, with vcpu_is_preempted, some of the lock heavy >> benchmarks come down from spending around 50% of the time in lock to >> less than 1% (so not sure how much more room is there for improvement). >> >> We could also use vcpu_is_preempted to optimize IPI performance (along >> with directed yield to target IPI vCPU) similar to how its done in x86 >> (https://lore.kernel.org/all/1560255830-8656-2-git-send-email-wanpengli@tencent.com/). >> This case definitely wont be covered by pvcy. >> >> Considering all the above, i.e. the core kernel integration already >> present and possible future usecases of vcpu_is_preempted, maybe its >> worth making vcpu_is_preempted work on arm independently of pvcy? >> > > Hi, > > Just wanted to check if there are any comments on above? I can send a v3 > with the doc and code fixes suggested in the earlier reviews if it makes > sense? > > Thanks, > Usama > >> Thanks, >> Usama >> Hi, The discussion on the patches had died down around November. I have sent v3 of the patches (https://lore.kernel.org/all/20230117102930.1053337-1-usama.arif@bytedance.com/) to hopefully restart it as I think that there is a significant performance improvement to be had with vcpu_is_preempted being implemented in arm64 which is well integrated in mutex, rtmutex, rwsem, osq_lock and scheduler, and could potentially be used to improve the IPI performance in the future. Thanks, Usama >>> >>> M. >>>
This patchset adds support for vcpu_is_preempted in arm64, which allows the guest to check if a vcpu was scheduled out, which is useful to know incase it was holding a lock. vcpu_is_preempted can be used to improve performance in locking (see owner_on_cpu usage in mutex_spin_on_owner, mutex_can_spin_on_owner, rtmutex_spin_on_owner and osq_lock) and scheduling (see available_idle_cpu which is used in several places in kernel/sched/fair.c for e.g. in wake_affine to determine which CPU can run soonest): This patchset shows improvement on overcommitted hosts (vCPUs > pCPUS), as waiting for preempted vCPUs reduces performance. This patchset is inspired from the para_steal_clock implementation and from the work originally done by Zengruan Ye: https://lore.kernel.org/linux-arm-kernel/20191226135833.1052-1-yezengruan@huawei.com/. All the results in the below experiments are done on an aws r6g.metal instance which has 64 pCPUs. The following table shows the index results of UnixBench running on a 128 vCPU VM with (6.0.0+vcpu_is_preempted) and without (6.0.0 base) the patchset. TestName 6.0.0 base 6.0.0+vcpu_is_preempted % improvement for vcpu_is_preempted Dhrystone 2 using register variables 187761 191274.7 1.871368389 Double-Precision Whetstone 96743.6 98414.4 1.727039308 Execl Throughput 689.3 10426 1412.548963 File Copy 1024 bufsize 2000 maxblocks 549.5 3165 475.978162 File Copy 256 bufsize 500 maxblocks 400.7 2084.7 420.2645371 File Copy 4096 bufsize 8000 maxblocks 894.3 5003.2 459.4543218 Pipe Throughput 76819.5 78601.5 2.319723508 Pipe-based Context Switching 3444.8 13414.5 289.4130283 Process Creation 301.1 293.4 -2.557289937 Shell Scripts (1 concurrent) 1248.1 28300.6 2167.494592 Shell Scripts (8 concurrent) 781.2 26222.3 3256.669227 System Call Overhead 3426 3729.4 8.855808523 System Benchmarks Index Score 3053 11534 277.7923354 This shows a 277% overall improvement using these patches. The biggest improvement is in the shell scripts benchmark, which forks a lot of processes. This acquires rwsem lock where a large chunk of time is spent in base 6.0.0 kernel. This can be seen from one of the callstack of the perf output of the shell scripts benchmark on 6.0.0 base (pseudo NMI enabled for perf numbers below): - 33.79% el0_svc - 33.43% do_el0_svc - 33.43% el0_svc_common.constprop.3 - 33.30% invoke_syscall - 17.27% __arm64_sys_clone - 17.27% __do_sys_clone - 17.26% kernel_clone - 16.73% copy_process - 11.91% dup_mm - 11.82% dup_mmap - 9.15% down_write - 8.87% rwsem_down_write_slowpath - 8.48% osq_lock Just under 50% of the total time in the shell script benchmarks ends up being spent in osq_lock in the base 6.0.0 kernel: Children Self Command Shared Object Symbol 17.19% 10.71% sh [kernel.kallsyms] [k] osq_lock 6.17% 4.04% sort [kernel.kallsyms] [k] osq_lock 4.20% 2.60% multi. [kernel.kallsyms] [k] osq_lock 3.77% 2.47% grep [kernel.kallsyms] [k] osq_lock 3.50% 2.24% expr [kernel.kallsyms] [k] osq_lock 3.41% 2.23% od [kernel.kallsyms] [k] osq_lock 3.36% 2.15% rm [kernel.kallsyms] [k] osq_lock 3.28% 2.12% tee [kernel.kallsyms] [k] osq_lock 3.16% 2.02% wc [kernel.kallsyms] [k] osq_lock 0.21% 0.13% looper [kernel.kallsyms] [k] osq_lock 0.01% 0.00% Run [kernel.kallsyms] [k] osq_lock and this comes down to less than 1% total with 6.0.0+vcpu_is_preempted kernel: Children Self Command Shared Object Symbol 0.26% 0.21% sh [kernel.kallsyms] [k] osq_lock 0.10% 0.08% multi. [kernel.kallsyms] [k] osq_lock 0.04% 0.04% sort [kernel.kallsyms] [k] osq_lock 0.02% 0.01% grep [kernel.kallsyms] [k] osq_lock 0.02% 0.02% od [kernel.kallsyms] [k] osq_lock 0.01% 0.01% tee [kernel.kallsyms] [k] osq_lock 0.01% 0.00% expr [kernel.kallsyms] [k] osq_lock 0.01% 0.01% looper [kernel.kallsyms] [k] osq_lock 0.00% 0.00% wc [kernel.kallsyms] [k] osq_lock 0.00% 0.00% rm [kernel.kallsyms] [k] osq_lock To make sure, there is no change in performance when vCPUs < pCPUs, UnixBench was run on a 32 CPU VM. The kernel with vcpu_is_preempted implemented performed 0.9% better overall than base kernel, and the individual benchmarks were within +/-2% improvement over 6.0.0 base. Hence the patches have no negative affect when vCPUs < pCPUs. The other method discussed in https://lore.kernel.org/linux-arm-kernel/20191226135833.1052-1-yezengruan@huawei.com/ was pv conditional yield by Marc Zyngier and Will Deacon to reduce vCPU reschedule if the vCPU will exit immediately. (https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pvcy). The patches were ported to 6.0.0 kernel and tested with UnixBench with a 128 vCPU VM: TestName 6.0.0 base 6.0.0+pvcy % improvement for pvcy Dhrystone 2 using register variables 187761 183128 -2.467498575 Double-Precision Whetstone 96743.6 96645 -0.101918887 Execl Throughput 689.3 1019.8 47.9471928 File Copy 1024 bufsize 2000 maxblocks 549.5 2029.7 269.3721565 File Copy 256 bufsize 500 maxblocks 400.7 1439.4 259.2213626 File Copy 4096 bufsize 8000 maxblocks 894.3 3434.1 283.9986582 Pipe Throughput 76819.5 74268.8 -3.320380893 Pipe-based Context Switching 3444.8 5963.3 73.11019508 Process Creation 301.1 163.2 -45.79873796 Shell Scripts (1 concurrent) 1248.1 1859.7 49.00248378 Shell Scripts (8 concurrent) 781.2 1171 49.89759345 System Call Overhead 3426 3194.4 -6.760070053 System Benchmarks Index Score 3053 4605 50.83524402 pvcy shows a smaller overall improvement (50%) compared to vcpu_is_preempted (277%). Host side flamegraph analysis shows that ~60% of the host time when using pvcy is spent in kvm_handle_wfx, compared with ~1.5% when using vcpu_is_preempted, hence vcpu_is_preempted shows a larger improvement. It might be that pvcy can be used in combination with vcpu_is_preempted, but this series is to start a discussion on vcpu_is_preempted as it shows a much bigger improvement in performance and its much easier to review vcpu_is_preempted standalone. The respective QEMU change to test this is at https://github.com/uarif1/qemu/commit/2da2c2927ae8de8f03f439804a0dad9cf68501b6, Looking forward to your response! Thanks, Usama --- RFC->v2 - Fixed table and code referencing in pvlock documentation - Switched to using a single hypercall similar to ptp_kvm and made check for has_kvm_pvlock simpler Usama Arif (6): KVM: arm64: Document PV-lock interface KVM: arm64: Add SMCCC paravirtualised lock calls KVM: arm64: Support pvlock preempted via shared structure KVM: arm64: Provide VCPU attributes for PV lock KVM: arm64: Support the VCPU preemption check KVM: selftests: add tests for PV time specific hypercall Documentation/virt/kvm/arm/hypercalls.rst | 3 + Documentation/virt/kvm/arm/index.rst | 1 + Documentation/virt/kvm/arm/pvlock.rst | 52 ++++++++ Documentation/virt/kvm/devices/vcpu.rst | 25 ++++ arch/arm64/include/asm/kvm_host.h | 25 ++++ arch/arm64/include/asm/paravirt.h | 2 + arch/arm64/include/asm/pvlock-abi.h | 17 +++ arch/arm64/include/asm/spinlock.h | 16 ++- arch/arm64/include/uapi/asm/kvm.h | 3 + arch/arm64/kernel/paravirt.c | 112 ++++++++++++++++++ arch/arm64/kernel/setup.c | 3 + arch/arm64/kvm/Makefile | 2 +- arch/arm64/kvm/arm.c | 8 ++ arch/arm64/kvm/guest.c | 9 ++ arch/arm64/kvm/hypercalls.c | 8 ++ arch/arm64/kvm/pvlock.c | 100 ++++++++++++++++ include/linux/arm-smccc.h | 8 ++ include/linux/cpuhotplug.h | 1 + include/uapi/linux/kvm.h | 2 + tools/arch/arm64/include/uapi/asm/kvm.h | 1 + tools/include/linux/arm-smccc.h | 8 ++ .../selftests/kvm/aarch64/hypercalls.c | 2 + 22 files changed, 406 insertions(+), 2 deletions(-) create mode 100644 Documentation/virt/kvm/arm/pvlock.rst create mode 100644 arch/arm64/include/asm/pvlock-abi.h create mode 100644 arch/arm64/kvm/pvlock.c