Re: [PATCH 4/5] kvm: Replace vcpu->swait with rcuwait
On 23/04/20 11:19, Peter Zijlstra wrote: >> >> 1) Davidlohr, please post only patches 1-3 to "equalize" the swait and >> rcuwait APIs. >> >> 2) Peter, please prepare a topic branch for those, or provide Acked-by > I don't think I have anything that conflicts with this, so sure, take > the whole thing through KVM. > > For 1-3 (and I'll send a small niggle for 3 right after this): > > Acked-by: Peter Zijlstra (Intel) Great thanks. I assume you have no issue with rcuwait_active either. Paolo > I'll keep 5 as it is unrelated. > ___ kvmarm mailing list kvmarm@lists.cs.columbia.edu https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Re: [PATCH 4/5] kvm: Replace vcpu->swait with rcuwait
On Thu, Apr 23, 2020 at 10:57:57AM +0200, Paolo Bonzini wrote: > On 23/04/20 10:41, Marc Zyngier wrote: > >> > >> - if (swait_active(kvm_arch_vcpu_wq(vcpu))) > >> + if (rcu_dereference(kvm_arch_vpu_get_wait(vcpu)) != NULL) > > This doesn't compile (wrong function name, and rcu_dereference takes a > > variable). But whatever it would do if we fixed it looks dodgy. it isn't > > the rcuwait structure that you want to dereference, but rcuwait->task > > (we are checking whether we are called because we are blocking or being > > preempted). > > > > Yes, I agree. Replacing swait with rcuwait is all good, but please make > the API look the same first. Just like you added prepare_to_rcuwait and > finish_rcuwait, let's add rcuwait_active as well. > > Actually let's do it like this: > > 1) Davidlohr, please post only patches 1-3 to "equalize" the swait and > rcuwait APIs. > > 2) Peter, please prepare a topic branch for those, or provide Acked-by I don't think I have anything that conflicts with this, so sure, take the whole thing through KVM. For 1-3 (and I'll send a small niggle for 3 right after this): Acked-by: Peter Zijlstra (Intel) I'll keep 5 as it is unrelated. ___ kvmarm mailing list kvmarm@lists.cs.columbia.edu https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Re: [PATCH 4/5] kvm: Replace vcpu->swait with rcuwait
On 23/04/20 10:41, Marc Zyngier wrote: >> >> -if (swait_active(kvm_arch_vcpu_wq(vcpu))) >> +if (rcu_dereference(kvm_arch_vpu_get_wait(vcpu)) != NULL) > This doesn't compile (wrong function name, and rcu_dereference takes a > variable). But whatever it would do if we fixed it looks dodgy. it isn't > the rcuwait structure that you want to dereference, but rcuwait->task > (we are checking whether we are called because we are blocking or being > preempted). > Yes, I agree. Replacing swait with rcuwait is all good, but please make the API look the same first. Just like you added prepare_to_rcuwait and finish_rcuwait, let's add rcuwait_active as well. Actually let's do it like this: 1) Davidlohr, please post only patches 1-3 to "equalize" the swait and rcuwait APIs. 2) Peter, please prepare a topic branch for those, or provide Acked-by 3) let's get everything else through the KVM tree. Thanks, Paolo ___ kvmarm mailing list kvmarm@lists.cs.columbia.edu https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Re: [PATCH 4/5] kvm: Replace vcpu->swait with rcuwait
On Tue, 21 Apr 2020 21:07:38 -0700 Davidlohr Bueso wrote: > The use of any sort of waitqueue (simple or regular) for > wait/waking vcpus has always been an overkill and semantically > wrong. Because this is per-vcpu (which is blocked) there is > only ever a single waiting vcpu, thus no need for any sort of > queue. > > As such, make use of the rcuwait primitive, with the following > considerations: > > - rcuwait already provides the proper barriers that serialize > concurrent waiter and waker. > > - Task wakeup is done in rcu read critical region, with a > stable task pointer. > > - Because there is no concurrency among waiters, we need > not worry about rcuwait_wait_event() calls corrupting > the wait->task. As a consequence, this saves the locking > done in swait when modifying the queue. This also applies > to per-vcore wait for powerpc kvm-hv. > > The x86 tscdeadline_latency test mentioned in 8577370fb0cb > ("KVM: Use simple waitqueue for vcpu->wq") shows that, on avg, > latency is reduced by around 15-20% with this change. > > Cc: Paul Mackerras > Cc: kvmarm@lists.cs.columbia.edu > Cc: linux-m...@vger.kernel.org > Signed-off-by: Davidlohr Bueso > --- > arch/mips/kvm/mips.c | 6 ++ > arch/powerpc/include/asm/kvm_book3s.h | 2 +- > arch/powerpc/include/asm/kvm_host.h | 2 +- > arch/powerpc/kvm/book3s_hv.c | 22 -- > arch/powerpc/kvm/powerpc.c| 2 +- > arch/x86/kvm/lapic.c | 2 +- > include/linux/kvm_host.h | 10 +- > virt/kvm/arm/arch_timer.c | 2 +- > virt/kvm/arm/arm.c| 9 + > virt/kvm/async_pf.c | 3 +-- > virt/kvm/kvm_main.c | 19 +-- > 11 files changed, 35 insertions(+), 44 deletions(-) [...] I should have tested it *before* acking it, really. > diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c > index 93bd59b46848..b2805105bbe5 100644 > --- a/virt/kvm/arm/arch_timer.c > +++ b/virt/kvm/arm/arch_timer.c > @@ -593,7 +593,7 @@ void kvm_timer_vcpu_put(struct kvm_vcpu *vcpu) > if (map.emul_ptimer) > soft_timer_cancel(&map.emul_ptimer->hrtimer); > > - if (swait_active(kvm_arch_vcpu_wq(vcpu))) > + if (rcu_dereference(kvm_arch_vpu_get_wait(vcpu)) != NULL) This doesn't compile (wrong function name, and rcu_dereference takes a variable). But whatever it would do if we fixed it looks dodgy. it isn't the rcuwait structure that you want to dereference, but rcuwait->task (we are checking whether we are called because we are blocking or being preempted). > kvm_timer_blocking(vcpu); > > /* > diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c > index 48d0ec44ad77..f94a10bb1251 100644 > --- a/virt/kvm/arm/arm.c > +++ b/virt/kvm/arm/arm.c > @@ -579,16 +579,17 @@ void kvm_arm_resume_guest(struct kvm *kvm) > > kvm_for_each_vcpu(i, vcpu, kvm) { > vcpu->arch.pause = false; > - swake_up_one(kvm_arch_vcpu_wq(vcpu)); > + rcuwait_wake_up(kvm_arch_vcpu_get_wait(vcpu)); > } > } > > static void vcpu_req_sleep(struct kvm_vcpu *vcpu) > { > - struct swait_queue_head *wq = kvm_arch_vcpu_wq(vcpu); > + struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); > > - swait_event_interruptible_exclusive(*wq, ((!vcpu->arch.power_off) && > -(!vcpu->arch.pause))); > + rcuwait_wait_event(*wait, > +(!vcpu->arch.power_off) &&(!vcpu->arch.pause), > +TASK_INTERRUPTIBLE); As noticed by the kbuild robot, this doesn't compile either. I fixed it as follow, and it survived a very basic test run in a model (more testing later). Thanks, M. diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c index b2805105bbe56..2dbd14dcae9fb 100644 --- a/virt/kvm/arm/arch_timer.c +++ b/virt/kvm/arm/arch_timer.c @@ -569,6 +569,7 @@ bool kvm_timer_should_notify_user(struct kvm_vcpu *vcpu) void kvm_timer_vcpu_put(struct kvm_vcpu *vcpu) { + struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); struct arch_timer_cpu *timer = vcpu_timer(vcpu); struct timer_map map; @@ -593,7 +594,7 @@ void kvm_timer_vcpu_put(struct kvm_vcpu *vcpu) if (map.emul_ptimer) soft_timer_cancel(&map.emul_ptimer->hrtimer); - if (rcu_dereference(kvm_arch_vpu_get_wait(vcpu)) != NULL) + if (rcu_dereference(wait->task)) kvm_timer_blocking(vcpu); /* diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c index f94a10bb1251b..479f36d02418d 100644 --- a/virt/kvm/arm/arm.c +++ b/virt/kvm/arm/arm.c @@ -587,7 +587,7 @@ static void vcpu_req_sleep(struct kvm_vcpu *vcpu) { struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); - rcuwait_wait_event(*wait, + rcuwait_wait_event(wait, (!vcpu->arch.power_
Re: [PATCH 4/5] kvm: Replace vcpu->swait with rcuwait
On 2020-04-22 05:07, Davidlohr Bueso wrote: The use of any sort of waitqueue (simple or regular) for wait/waking vcpus has always been an overkill and semantically wrong. Because this is per-vcpu (which is blocked) there is only ever a single waiting vcpu, thus no need for any sort of queue. As such, make use of the rcuwait primitive, with the following considerations: - rcuwait already provides the proper barriers that serialize concurrent waiter and waker. - Task wakeup is done in rcu read critical region, with a stable task pointer. - Because there is no concurrency among waiters, we need not worry about rcuwait_wait_event() calls corrupting the wait->task. As a consequence, this saves the locking done in swait when modifying the queue. This also applies to per-vcore wait for powerpc kvm-hv. The x86 tscdeadline_latency test mentioned in 8577370fb0cb ("KVM: Use simple waitqueue for vcpu->wq") shows that, on avg, latency is reduced by around 15-20% with this change. Cc: Paul Mackerras Cc: kvmarm@lists.cs.columbia.edu Cc: linux-m...@vger.kernel.org Signed-off-by: Davidlohr Bueso Reviewed-by: Marc Zyngier Thanks, M. -- Who you jivin' with that Cosmik Debris? ___ kvmarm mailing list kvmarm@lists.cs.columbia.edu https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
[PATCH 4/5] kvm: Replace vcpu->swait with rcuwait
The use of any sort of waitqueue (simple or regular) for wait/waking vcpus has always been an overkill and semantically wrong. Because this is per-vcpu (which is blocked) there is only ever a single waiting vcpu, thus no need for any sort of queue. As such, make use of the rcuwait primitive, with the following considerations: - rcuwait already provides the proper barriers that serialize concurrent waiter and waker. - Task wakeup is done in rcu read critical region, with a stable task pointer. - Because there is no concurrency among waiters, we need not worry about rcuwait_wait_event() calls corrupting the wait->task. As a consequence, this saves the locking done in swait when modifying the queue. This also applies to per-vcore wait for powerpc kvm-hv. The x86 tscdeadline_latency test mentioned in 8577370fb0cb ("KVM: Use simple waitqueue for vcpu->wq") shows that, on avg, latency is reduced by around 15-20% with this change. Cc: Paul Mackerras Cc: kvmarm@lists.cs.columbia.edu Cc: linux-m...@vger.kernel.org Signed-off-by: Davidlohr Bueso --- arch/mips/kvm/mips.c | 6 ++ arch/powerpc/include/asm/kvm_book3s.h | 2 +- arch/powerpc/include/asm/kvm_host.h | 2 +- arch/powerpc/kvm/book3s_hv.c | 22 -- arch/powerpc/kvm/powerpc.c| 2 +- arch/x86/kvm/lapic.c | 2 +- include/linux/kvm_host.h | 10 +- virt/kvm/arm/arch_timer.c | 2 +- virt/kvm/arm/arm.c| 9 + virt/kvm/async_pf.c | 3 +-- virt/kvm/kvm_main.c | 19 +-- 11 files changed, 35 insertions(+), 44 deletions(-) diff --git a/arch/mips/kvm/mips.c b/arch/mips/kvm/mips.c index 8f05dd0a0f4e..fad6acce46e4 100644 --- a/arch/mips/kvm/mips.c +++ b/arch/mips/kvm/mips.c @@ -284,8 +284,7 @@ static enum hrtimer_restart kvm_mips_comparecount_wakeup(struct hrtimer *timer) kvm_mips_callbacks->queue_timer_int(vcpu); vcpu->arch.wait = 0; - if (swq_has_sleeper(&vcpu->wq)) - swake_up_one(&vcpu->wq); + rcuwait_wake_up(&vcpu->wait); return kvm_mips_count_timeout(vcpu); } @@ -511,8 +510,7 @@ int kvm_vcpu_ioctl_interrupt(struct kvm_vcpu *vcpu, dvcpu->arch.wait = 0; - if (swq_has_sleeper(&dvcpu->wq)) - swake_up_one(&dvcpu->wq); + rcuwait_wake_up(&dvcpu->wait); return 0; } diff --git a/arch/powerpc/include/asm/kvm_book3s.h b/arch/powerpc/include/asm/kvm_book3s.h index 506e4df2d730..6e5d85ba588d 100644 --- a/arch/powerpc/include/asm/kvm_book3s.h +++ b/arch/powerpc/include/asm/kvm_book3s.h @@ -78,7 +78,7 @@ struct kvmppc_vcore { struct kvm_vcpu *runnable_threads[MAX_SMT_THREADS]; struct list_head preempt_list; spinlock_t lock; - struct swait_queue_head wq; + struct rcuwait wait; spinlock_t stoltb_lock; /* protects stolen_tb and preempt_tb */ u64 stolen_tb; u64 preempt_tb; diff --git a/arch/powerpc/include/asm/kvm_host.h b/arch/powerpc/include/asm/kvm_host.h index 1dc63101ffe1..337047ba4a56 100644 --- a/arch/powerpc/include/asm/kvm_host.h +++ b/arch/powerpc/include/asm/kvm_host.h @@ -751,7 +751,7 @@ struct kvm_vcpu_arch { u8 irq_pending; /* Used by XIVE to signal pending guest irqs */ u32 last_inst; - struct swait_queue_head *wqp; + struct rcuwait *waitp; struct kvmppc_vcore *vcore; int ret; int trap; diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c index 93493f0cbfe8..b8d42f523ca7 100644 --- a/arch/powerpc/kvm/book3s_hv.c +++ b/arch/powerpc/kvm/book3s_hv.c @@ -230,13 +230,11 @@ static bool kvmppc_ipi_thread(int cpu) static void kvmppc_fast_vcpu_kick_hv(struct kvm_vcpu *vcpu) { int cpu; - struct swait_queue_head *wqp; + struct rcuwait *wait; - wqp = kvm_arch_vcpu_wq(vcpu); - if (swq_has_sleeper(wqp)) { - swake_up_one(wqp); + wait = kvm_arch_vcpu_get_wait(vcpu); + if (rcuwait_wake_up(wait)) ++vcpu->stat.halt_wakeup; - } cpu = READ_ONCE(vcpu->arch.thread_cpu); if (cpu >= 0 && kvmppc_ipi_thread(cpu)) @@ -2125,7 +2123,7 @@ static struct kvmppc_vcore *kvmppc_vcore_create(struct kvm *kvm, int id) spin_lock_init(&vcore->lock); spin_lock_init(&vcore->stoltb_lock); - init_swait_queue_head(&vcore->wq); + rcuwait_init(&vcore->wait); vcore->preempt_tb = TB_NIL; vcore->lpcr = kvm->arch.lpcr; vcore->first_vcpuid = id; @@ -3784,7 +3782,6 @@ static void kvmppc_vcore_blocked(struct kvmppc_vcore *vc) ktime_t cur, start_poll, start_wait; int do_sleep = 1; u64 block_ns; - DECLARE_SWAITQUEUE(wait); /* Poll for pending exceptions and ceded state */ cur = start_poll = ktime_get(); @@ -3812,10 +3809,7 @@ static void kvmppc_vcore_blocked(struct kvm
Re: [PATCH 4/5] kvm: Replace vcpu->swait with rcuwait
On 22/04/20 06:07, Davidlohr Bueso wrote: > The use of any sort of waitqueue (simple or regular) for > wait/waking vcpus has always been an overkill and semantically > wrong. Because this is per-vcpu (which is blocked) there is > only ever a single waiting vcpu, thus no need for any sort of > queue. > > As such, make use of the rcuwait primitive, with the following > considerations: > > - rcuwait already provides the proper barriers that serialize > concurrent waiter and waker. > > - Task wakeup is done in rcu read critical region, with a > stable task pointer. > > - Because there is no concurrency among waiters, we need > not worry about rcuwait_wait_event() calls corrupting > the wait->task. As a consequence, this saves the locking > done in swait when modifying the queue. This also applies > to per-vcore wait for powerpc kvm-hv. > > The x86 tscdeadline_latency test mentioned in 8577370fb0cb > ("KVM: Use simple waitqueue for vcpu->wq") shows that, on avg, > latency is reduced by around 15-20% with this change. > > Cc: Paul Mackerras > Cc: kvmarm@lists.cs.columbia.edu > Cc: linux-m...@vger.kernel.org > Signed-off-by: Davidlohr Bueso Reviewed-by: Paolo Bonzini > --- > arch/mips/kvm/mips.c | 6 ++ > arch/powerpc/include/asm/kvm_book3s.h | 2 +- > arch/powerpc/include/asm/kvm_host.h | 2 +- > arch/powerpc/kvm/book3s_hv.c | 22 -- > arch/powerpc/kvm/powerpc.c| 2 +- > arch/x86/kvm/lapic.c | 2 +- > include/linux/kvm_host.h | 10 +- > virt/kvm/arm/arch_timer.c | 2 +- > virt/kvm/arm/arm.c| 9 + > virt/kvm/async_pf.c | 3 +-- > virt/kvm/kvm_main.c | 19 +-- > 11 files changed, 35 insertions(+), 44 deletions(-) > > diff --git a/arch/mips/kvm/mips.c b/arch/mips/kvm/mips.c > index 8f05dd0a0f4e..fad6acce46e4 100644 > --- a/arch/mips/kvm/mips.c > +++ b/arch/mips/kvm/mips.c > @@ -284,8 +284,7 @@ static enum hrtimer_restart > kvm_mips_comparecount_wakeup(struct hrtimer *timer) > kvm_mips_callbacks->queue_timer_int(vcpu); > > vcpu->arch.wait = 0; > - if (swq_has_sleeper(&vcpu->wq)) > - swake_up_one(&vcpu->wq); > + rcuwait_wake_up(&vcpu->wait); > > return kvm_mips_count_timeout(vcpu); > } > @@ -511,8 +510,7 @@ int kvm_vcpu_ioctl_interrupt(struct kvm_vcpu *vcpu, > > dvcpu->arch.wait = 0; > > - if (swq_has_sleeper(&dvcpu->wq)) > - swake_up_one(&dvcpu->wq); > + rcuwait_wake_up(&dvcpu->wait); > > return 0; > } > diff --git a/arch/powerpc/include/asm/kvm_book3s.h > b/arch/powerpc/include/asm/kvm_book3s.h > index 506e4df2d730..6e5d85ba588d 100644 > --- a/arch/powerpc/include/asm/kvm_book3s.h > +++ b/arch/powerpc/include/asm/kvm_book3s.h > @@ -78,7 +78,7 @@ struct kvmppc_vcore { > struct kvm_vcpu *runnable_threads[MAX_SMT_THREADS]; > struct list_head preempt_list; > spinlock_t lock; > - struct swait_queue_head wq; > + struct rcuwait wait; > spinlock_t stoltb_lock; /* protects stolen_tb and preempt_tb */ > u64 stolen_tb; > u64 preempt_tb; > diff --git a/arch/powerpc/include/asm/kvm_host.h > b/arch/powerpc/include/asm/kvm_host.h > index 1dc63101ffe1..337047ba4a56 100644 > --- a/arch/powerpc/include/asm/kvm_host.h > +++ b/arch/powerpc/include/asm/kvm_host.h > @@ -751,7 +751,7 @@ struct kvm_vcpu_arch { > u8 irq_pending; /* Used by XIVE to signal pending guest irqs */ > u32 last_inst; > > - struct swait_queue_head *wqp; > + struct rcuwait *waitp; > struct kvmppc_vcore *vcore; > int ret; > int trap; > diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c > index 93493f0cbfe8..b8d42f523ca7 100644 > --- a/arch/powerpc/kvm/book3s_hv.c > +++ b/arch/powerpc/kvm/book3s_hv.c > @@ -230,13 +230,11 @@ static bool kvmppc_ipi_thread(int cpu) > static void kvmppc_fast_vcpu_kick_hv(struct kvm_vcpu *vcpu) > { > int cpu; > - struct swait_queue_head *wqp; > + struct rcuwait *wait; > > - wqp = kvm_arch_vcpu_wq(vcpu); > - if (swq_has_sleeper(wqp)) { > - swake_up_one(wqp); > + wait = kvm_arch_vcpu_get_wait(vcpu); > + if (rcuwait_wake_up(wait)) > ++vcpu->stat.halt_wakeup; > - } > > cpu = READ_ONCE(vcpu->arch.thread_cpu); > if (cpu >= 0 && kvmppc_ipi_thread(cpu)) > @@ -2125,7 +2123,7 @@ static struct kvmppc_vcore *kvmppc_vcore_create(struct > kvm *kvm, int id) > > spin_lock_init(&vcore->lock); > spin_lock_init(&vcore->stoltb_lock); > - init_swait_queue_head(&vcore->wq); > + rcuwait_init(&vcore->wait); > vcore->preempt_tb = TB_NIL; > vcore->lpcr = kvm->arch.lpcr; > vcore->first_vcpuid = id; > @@ -3784,7 +3782,6 @@ static void kvmppc_vcore_blocked(struct kvmppc_vcore > *vc) > ktime_t cur, start_poll, start_wait; >