Skip to contentSkip to conversation

KVM: Yield to IPI target if necessary

PATCH Superseded Wanpeng Li 4 participants 18 messages May 30, 2019 – Jun 28, 2019

Subsystems (5)
Documentation x86 KVM KVM x86 +1 more subsystems

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Wanpeng Li

author
#0 lore raw

v3 cover KVM: Yield to IPI target if necessary

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

v2 -> v3:
* add bounds-check on dest_id

v1 -> v2:
* check map is not NULL
* check map->phys_map[dest_id] is not NULL
* make kvm_sched_yield static
* change dest_id to unsinged long

Wanpeng Li (3):
KVM: X86: Yield to IPI target if necessary
KVM: X86: Implement PV sched yield hypercall
KVM: X86: Expose PV_SCHED_YIELD CPUID feature bit to guest

 Documentation/virtual/kvm/cpuid.txt      |  4 ++++
 Documentation/virtual/kvm/hypercalls.txt | 11 +++++++++++
 arch/x86/include/uapi/asm/kvm_para.h     |  1 +
 arch/x86/kernel/kvm.c                    | 21 +++++++++++++++++++++
 arch/x86/kvm/cpuid.c                     |  3 ++-
 arch/x86/kvm/x86.c                       | 26 ++++++++++++++++++++++++++
 include/uapi/linux/kvm_para.h            |  1 +
 7 files changed, 66 insertions(+), 1 deletion(-)
Signature · 1 line
2.7.4
author
#1 lore raw

PATCH v3 1/3 KVM: X86: Yield to IPI target if necessary

From: Wanpeng Li <[email protected]>

When sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted, we just select the first preempted target vCPU which we found since the state of target vCPUs can change underneath and to avoid race conditions.

Cc Paolo Bonzini <[email protected]>
Cc Radim Krčmář <[email protected]>
Cc Liran Alon <[email protected]>
---
 Documentation/virtual/kvm/hypercalls.txt | 11 +++++++++++
 arch/x86/include/uapi/asm/kvm_para.h     |  1 +
 arch/x86/kernel/kvm.c                    | 21 +++++++++++++++++++++
 include/uapi/linux/kvm_para.h            |  1 +
 4 files changed, 34 insertions(+)
Documentation/virtual/kvm/hypercalls.txt +11 -0
index da24c13..da21065 100644
@@ -141,3 +141,14 @@ a0 corresponds to the APIC ID in the third argument (a2), bit 1
 corresponds to the APIC ID a2+1, and so on.
 
 Returns the number of CPUs to which the IPIs were delivered successfully.
+
+7. KVM_HC_SCHED_YIELD
+------------------------
+Architecture: x86
+Status: active
+Purpose: Hypercall used to yield if the IPI target vCPU is preempted
+
+a0: destination APIC ID
+
+Usage example: When sending a call-function IPI-many to vCPUs, yield if
+any of the IPI target vCPUs was preempted.
arch/x86/include/uapi/asm/kvm_para.h +1 -0
index 19980ec..d0bf77c 100644
@@ -29,6 +29,7 @@
 #define KVM_FEATURE_PV_TLB_FLUSH	9
 #define KVM_FEATURE_ASYNC_PF_VMEXIT	10
 #define KVM_FEATURE_PV_SEND_IPI	11
+#define KVM_FEATURE_PV_SCHED_YIELD	12
 
 #define KVM_HINTS_REALTIME      0
 
arch/x86/kernel/kvm.c +21 -0
index 3f0cc82..54400c2 100644
@@ -540,6 +540,21 @@ static void kvm_setup_pv_ipi(void)
 	pr_info("KVM setup pv IPIs\n");
 }
 
+static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
+{
+	int cpu;
+
+	native_send_call_func_ipi(mask);
+
+	/* Make sure other vCPUs get a chance to run if they need to. */
+	for_each_cpu(cpu, mask) {
+		if (vcpu_is_preempted(cpu)) {
+			kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
+			break;
+		}
+	}
+}
+
 static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
 {
 	native_smp_prepare_cpus(max_cpus);
@@ -651,6 +666,12 @@ static void __init kvm_guest_init(void)
 #ifdef CONFIG_SMP
 	smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
 	smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
+	if (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
+	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
+	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
+		smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
+		pr_info("KVM setup pv sched yield\n");
+	}
 	if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 				      kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 		pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
include/uapi/linux/kvm_para.h +1 -0
index 6c0ce49..8b86609 100644
@@ -28,6 +28,7 @@
 #define KVM_HC_MIPS_CONSOLE_OUTPUT	8
 #define KVM_HC_CLOCK_PAIRING		9
 #define KVM_HC_SEND_IPI		10
+#define KVM_HC_SCHED_YIELD		11
 
 /*
  * hypercalls use architecture specific
Signature · 1 line
2.7.4
author
#2 lore raw

PATCH v3 2/3 KVM: X86: Implement PV sched yield hypercall

In reply to Wanpeng Li

From: Wanpeng Li <[email protected]>

The target vCPUs are in runnable state after vcpu_kick and suitable as a yield target. This patch implements the sched yield hypercall.

17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Cc Paolo Bonzini <[email protected]>
Cc Radim Krčmář <[email protected]>
Cc Liran Alon <[email protected]>
---
 arch/x86/kvm/x86.c | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)
arch/x86/kvm/x86.c +26 -0
index e7e57de..8575b36 100644
@@ -7172,6 +7172,28 @@ void kvm_vcpu_deactivate_apicv(struct kvm_vcpu *vcpu)
 	kvm_x86_ops->refresh_apicv_exec_ctrl(vcpu);
 }
 
+static void kvm_sched_yield(struct kvm *kvm, unsigned long dest_id)
+{
+	struct kvm_vcpu *target = NULL;
+	struct kvm_apic_map *map = NULL;
+
+	rcu_read_lock();
+	map = rcu_dereference(kvm->arch.apic_map);
+
+	if (unlikely(!map) || dest_id > map->max_apic_id)
+		goto out;
+
+	if (map->phys_map[dest_id]->vcpu) {
+		target = map->phys_map[dest_id]->vcpu;
+		rcu_read_unlock();
+		kvm_vcpu_yield_to(target);
+	}
+
+out:
+	if (!target)
+		rcu_read_unlock();
+}
+
 int kvm_emulate_hypercall(struct kvm_vcpu *vcpu)
 {
 	unsigned long nr, a0, a1, a2, a3, ret;
@@ -7218,6 +7240,10 @@ int kvm_emulate_hypercall(struct kvm_vcpu *vcpu)
 	case KVM_HC_SEND_IPI:
 		ret = kvm_pv_send_ipi(vcpu->kvm, a0, a1, a2, a3, op_64_bit);
 		break;
+	case KVM_HC_SCHED_YIELD:
+		kvm_sched_yield(vcpu->kvm, a0);
+		ret = 0;
+		break;
 	default:
 		ret = -KVM_ENOSYS;
 		break;
Signature · 1 line
2.7.4
#3 lore raw

Re: [PATCH v3 2/3] KVM: X86: Implement PV sched yield hypercall

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 22 lines

From: Wanpeng Li <[email protected]>

The target vCPUs are in runnable state after vcpu_kick and suitable as a yield target. This patch implements the sched yield hypercall.

17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Cc Paolo Bonzini <[email protected]>
Cc Radim Krčmář <[email protected]>
Cc Liran Alon <[email protected]>

---

arch/x86/kvm/x86.c +2 -0
@@ -7172,6 +7172,28 @@ void kvm_vcpu_deactivate_apicv(struct kvm_vcpu *vcpu)
 	kvm_x86_ops->refresh_apicv_exec_ctrl(vcpu);
 }
 
+static void kvm_sched_yield(struct kvm *kvm, unsigned long dest_id)
+{

arch/x86/kvm/x86.c void kvm_vcpu_deactivate_apicv(struct kvm_vcpu *vcpu) · line 7186

Quoted 10 lines
arch/x86/kvm/x86.c +10 -0
+	struct kvm_vcpu *target = NULL;
+	struct kvm_apic_map *map = NULL;
+
+	rcu_read_lock();
+	map = rcu_dereference(kvm->arch.apic_map);
+
+	if (unlikely(!map) || dest_id > map->max_apic_id)
+		goto out;
+
+	if (map->phys_map[dest_id]->vcpu) {

This should check for map->phys_map[dest_id].

arch/x86/kvm/x86.c void kvm_vcpu_deactivate_apicv(struct kvm_vcpu *vcpu) · line 7194

Quoted 8 lines
arch/x86/kvm/x86.c +8 -0
+		target = map->phys_map[dest_id]->vcpu;
+		rcu_read_unlock();
+		kvm_vcpu_yield_to(target);
+	}
+
+out:
+	if (!target)
+		rcu_read_unlock();

Also, I find the following logic clearer

  {
  	struct kvm_vcpu *target = NULL;
  	struct kvm_apic_map *map;
  	
  	rcu_read_lock();
  	map = rcu_dereference(kvm->arch.apic_map);
  	
  	if (likely(map) && dest_id <= map->max_apic_id && map->phys_map[dest_id])
  		target = map->phys_map[dest_id]->vcpu;
  	
  	rcu_read_unlock();
  	
  	if (target)
  		kvm_vcpu_yield_to(target);
  }

thanks.

author
#4 lore raw

Re: [PATCH v3 2/3] KVM: X86: Implement PV sched yield hypercall

Quoted 22 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 20 lines

From: Wanpeng Li <[email protected]>

The target vCPUs are in runnable state after vcpu_kick and suitable as a yield target. This patch implements the sched yield hypercall.

17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Cc Paolo Bonzini <[email protected]>
Cc Radim Krčmář <[email protected]>
Cc Liran Alon <[email protected]>

---

arch/x86/kvm/x86.c +0 -0
@@ -7172,6 +7172,28 @@ void kvm_vcpu_deactivate_apicv(struct kvm_vcpu *vcpu)
      kvm_x86_ops->refresh_apicv_exec_ctrl(vcpu);
 }
 
Quoted 14 lines

arch/x86/kvm/x86.c void kvm_vcpu_deactivate_apicv(struct kvm_vcpu *vcpu) · line 7186

Quoted 12 lines
arch/x86/kvm/x86.c +12 -0
+static void kvm_sched_yield(struct kvm *kvm, unsigned long dest_id)
+{
+     struct kvm_vcpu *target = NULL;
+     struct kvm_apic_map *map = NULL;
+
+     rcu_read_lock();
+     map = rcu_dereference(kvm->arch.apic_map);
+
+     if (unlikely(!map) || dest_id > map->max_apic_id)
+             goto out;
+
+     if (map->phys_map[dest_id]->vcpu) {

This should check for map->phys_map[dest_id].

Yeah, make a mistake here.

Quoted 26 lines

arch/x86/kvm/x86.c void kvm_vcpu_deactivate_apicv(struct kvm_vcpu *vcpu) · line 7194

Quoted 9 lines
arch/x86/kvm/x86.c +8 -0
+             target = map->phys_map[dest_id]->vcpu;
+             rcu_read_unlock();
+             kvm_vcpu_yield_to(target);
+     }
+
+out:
+     if (!target)
+             rcu_read_unlock();

Also, I find the following logic clearer

  {
        struct kvm_vcpu *target = NULL;
        struct kvm_apic_map *map;
        rcu_read_lock();
        map = rcu_dereference(kvm->arch.apic_map);

        if (likely(map) && dest_id <= map->max_apic_id && map->phys_map[dest_id])
                target = map->phys_map[dest_id]->vcpu;

        rcu_read_unlock();

        if (target)
                kvm_vcpu_yield_to(target);
Quoted 1 line

}

More better, thanks.

Regards,
Wanpeng Li

author
#5 lore raw

PATCH v3 3/3 KVM: X86: Expose PV_SCHED_YIELD CPUID feature bit to guest

In reply to Wanpeng Li

From: Wanpeng Li <[email protected]>

Expose PV_SCHED_YIELD feature bit to guest, the guest can check this feature bit before using paravirtualized sched yield.

Cc Paolo Bonzini <[email protected]>
Cc Radim Krčmář <[email protected]>
Cc Liran Alon <[email protected]>
---
 Documentation/virtual/kvm/cpuid.txt | 4 ++++
 arch/x86/kvm/cpuid.c                | 3 ++-
 2 files changed, 6 insertions(+), 1 deletion(-)
Documentation/virtual/kvm/cpuid.txt +4 -0
index 97ca194..1c39683 100644
@@ -66,6 +66,10 @@ KVM_FEATURE_PV_SEND_IPI            ||    11 || guest checks this feature bit
                                    ||       || before using paravirtualized
                                    ||       || send IPIs.
 ------------------------------------------------------------------------------
+KVM_FEATURE_PV_SHED_YIELD          ||    12 || guest checks this feature bit
+                                   ||       || before using paravirtualized
+                                   ||       || sched yield.
+------------------------------------------------------------------------------
 KVM_FEATURE_CLOCKSOURCE_STABLE_BIT ||    24 || host will warn if no guest-side
                                    ||       || per-cpu warps are expected in
                                    ||       || kvmclock.
arch/x86/kvm/cpuid.c +2 -1
index e18a9f9..c018fc8 100644
@@ -643,7 +643,8 @@ static inline int __do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,
 			     (1 << KVM_FEATURE_PV_UNHALT) |
 			     (1 << KVM_FEATURE_PV_TLB_FLUSH) |
 			     (1 << KVM_FEATURE_ASYNC_PF_VMEXIT) |
-			     (1 << KVM_FEATURE_PV_SEND_IPI);
+			     (1 << KVM_FEATURE_PV_SEND_IPI) |
+			     (1 << KVM_FEATURE_PV_SCHED_YIELD);
 
 		if (sched_info_on())
 			entry->eax |= (1 << KVM_FEATURE_STEAL_TIME);
Signature · 1 line
2.7.4
author
#6 lore raw
In reply to Wanpeng Li

ping, :)

Quoted 34 lines · Wanpeng Li

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

v2 -> v3:
* add bounds-check on dest_id

v1 -> v2:
* check map is not NULL
* check map->phys_map[dest_id] is not NULL
* make kvm_sched_yield static
* change dest_id to unsinged long

Wanpeng Li (3):
KVM: X86: Yield to IPI target if necessary
KVM: X86: Implement PV sched yield hypercall
KVM: X86: Expose PV_SCHED_YIELD CPUID feature bit to guest

 Documentation/virtual/kvm/cpuid.txt      |  4 ++++
 Documentation/virtual/kvm/hypercalls.txt | 11 +++++++++++
 arch/x86/include/uapi/asm/kvm_para.h     |  1 +
 arch/x86/kernel/kvm.c                    | 21 +++++++++++++++++++++
 arch/x86/kvm/cpuid.c                     |  3 ++-
 arch/x86/kvm/x86.c                       | 26 ++++++++++++++++++++++++++
 include/uapi/linux/kvm_para.h            |  1 +
 7 files changed, 66 insertions(+), 1 deletion(-)

--
2.7.4

#7 lore raw
In reply to Wanpeng Li

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a
complication when splitting the target mask into several PV IPI hypercalls.)

Thanks.

#8 lore raw
Quoted 8 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Quoted 7 lines

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a
complication when splitting the target mask into several PV IPI hypercalls.)

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

author
#9 lore raw
Quoted 18 lines · Sean Christopherson
Quoted 8 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Quoted 7 lines

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a
complication when splitting the target mask into several PV IPI hypercalls.)

Quoted 4 lines

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Not sure why shorthand is not used for native x2apic mode.

Regards,
Wanpeng Li

#10 lore raw
Quoted 21 lines · Wanpeng Li
Quoted 17 lines · Sean Christopherson
Quoted 15 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a complication when splitting the target mask into several PV IPI hypercalls.)

Quoted 6 lines
Quoted 4 lines

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Not sure why shorthand is not used for native x2apic mode.

Why do you say so? native_send_call_func_ipi() checks if allbutself shorthand should be used and does so (even though the check can be more efficient - I’m looking at that code right now…)

author
#11 lore raw
Quoted 27 lines · Nadav Amit
Quoted 27 lines · Wanpeng Li
Quoted 21 lines · Sean Christopherson
Quoted 8 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Quoted 7 lines

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a complication when splitting the target mask into several PV IPI hypercalls.)

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Quoted 5 lines
Quoted 1 line

Not sure why shorthand is not used for native x2apic mode.

Why do you say so? native_send_call_func_ipi() checks if allbutself shorthand should be used and does so (even though the check can be more efficient - I’m looking at that code right now…)

Please continue to follow the apic/x2apic driver. Just apic_flat set APIC_DEST_ALLBUT/APIC_DEST_ALLINC to ICR.

Regards,
Wanpeng Li

#12 lore raw
Quoted 29 lines · Wanpeng Li
Quoted 26 lines · Nadav Amit
Quoted 26 lines · Wanpeng Li
Quoted 21 lines · Sean Christopherson
Quoted 8 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Quoted 7 lines

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a complication when splitting the target mask into several PV IPI hypercalls.)

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Not sure why shorthand is not used for native x2apic mode.

Why do you say so? native_send_call_func_ipi() checks if allbutself shorthand should be used and does so (even though the check can be more efficient - I’m looking at that code right now…)

Please continue to follow the apic/x2apic driver. Just apic_flat set APIC_DEST_ALLBUT/APIC_DEST_ALLINC to ICR.

Indeed - I was sure by the name that it does it correctly. That’s stupid.

I’ll add it to the patch-set I am working on (TLB shootdown improvements), if you don’t mind.

author
#13 lore raw
Quoted 36 lines · Nadav Amit
Quoted 36 lines · Wanpeng Li
Quoted 31 lines · Nadav Amit
Quoted 21 lines · Wanpeng Li
Quoted 17 lines · Sean Christopherson
Quoted 15 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a complication when splitting the target mask into several PV IPI hypercalls.)

Quoted 6 lines
Quoted 4 lines

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Not sure why shorthand is not used for native x2apic mode.

Why do you say so? native_send_call_func_ipi() checks if allbutself shorthand should be used and does so (even though the check can be more efficient - I’m looking at that code right now…)

Quoted 7 lines
Quoted 2 lines

Please continue to follow the apic/x2apic driver. Just apic_flat set APIC_DEST_ALLBUT/APIC_DEST_ALLINC to ICR.

Indeed - I was sure by the name that it does it correctly. That’s stupid.

I’ll add it to the patch-set I am working on (TLB shootdown improvements), if you don’t mind.

Original for hotplug cpu safe.
https://lwn.net/Articles/138365/
https://lwn.net/Articles/138368/
Not sure shortcut native support is acceptable, I will play my kvm_send_ipi_allbutself and kvm_send_ipi_all. :)

Regards,
Wanpeng Li

#14 lore raw
Quoted 43 lines · Wanpeng Li
Quoted 40 lines · Nadav Amit
Quoted 29 lines · Wanpeng Li
Quoted 26 lines · Nadav Amit
Quoted 26 lines · Wanpeng Li
Quoted 21 lines · Sean Christopherson
Quoted 8 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Quoted 7 lines

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a complication when splitting the target mask into several PV IPI hypercalls.)

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Not sure why shorthand is not used for native x2apic mode.

Why do you say so? native_send_call_func_ipi() checks if allbutself shorthand should be used and does so (even though the check can be more efficient - I’m looking at that code right now…)

Please continue to follow the apic/x2apic driver. Just apic_flat set APIC_DEST_ALLBUT/APIC_DEST_ALLINC to ICR.

Indeed - I was sure by the name that it does it correctly. That’s stupid.

Quoted 8 lines
Quoted 2 lines

I’ll add it to the patch-set I am working on (TLB shootdown improvements), if you don’t mind.

Original for hotplug cpu safe.
https://lwn.net/Articles/138365/
https://lwn.net/Articles/138368/
Not sure shortcut native support is acceptable, I will play my kvm_send_ipi_allbutself and kvm_send_ipi_all. :)

Yes, I saw these threads before. But I think the test in native_send_call_func_ipi() should take care of it.

I’ll recheck.

author
#15 lore raw
Quoted 47 lines · Nadav Amit
Quoted 47 lines · Wanpeng Li
Quoted 42 lines · Nadav Amit
Quoted 29 lines · Wanpeng Li
Quoted 26 lines · Nadav Amit
Quoted 26 lines · Wanpeng Li
Quoted 21 lines · Sean Christopherson
Quoted 8 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Quoted 7 lines

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a complication when splitting the target mask into several PV IPI hypercalls.)

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Quoted 8 lines
Quoted 5 lines
Quoted 1 line

Not sure why shorthand is not used for native x2apic mode.

Why do you say so? native_send_call_func_ipi() checks if allbutself shorthand should be used and does so (even though the check can be more efficient - I’m looking at that code right now…)

Please continue to follow the apic/x2apic driver. Just apic_flat set APIC_DEST_ALLBUT/APIC_DEST_ALLINC to ICR.

Indeed - I was sure by the name that it does it correctly. That’s stupid.

I’ll add it to the patch-set I am working on (TLB shootdown improvements), if you don’t mind.

Quoted 8 lines
Quoted 5 lines

Original for hotplug cpu safe.
https://lwn.net/Articles/138365/
https://lwn.net/Articles/138368/
Not sure shortcut native support is acceptable, I will play my kvm_send_ipi_allbutself and kvm_send_ipi_all. :)

Yes, I saw these threads before. But I think the test in native_send_call_func_ipi() should take care of it.

Good news, https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/log/?h=WIP.x86/ipi
Thomas who also is the hotplug state machine author introduces
shorthands support to native kernel now, I will add the support to kvm_send_ipi_allbutself() and kvm_send_ipi_all() after his work
complete.

Regards,
Wanpeng Li

author
#16 lore raw
Quoted 55 lines · Wanpeng Li
Quoted 53 lines · Nadav Amit
Quoted 52 lines · Wanpeng Li
Quoted 35 lines · Nadav Amit
Quoted 35 lines · Wanpeng Li
Quoted 31 lines · Nadav Amit
Quoted 21 lines · Wanpeng Li
Quoted 17 lines · Sean Christopherson
Quoted 15 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It would allow us to skip the VM entry/exit overhead on the caller. (The benefit of that might be negligible and it also poses a complication when splitting the target mask into several PV IPI hypercalls.)

Quoted 6 lines
Quoted 4 lines

Tangetially related to splitting PV IPI hypercalls, are there any major hurdles to supporting shorthand? Not having to generate the mask for ->send_IPI_allbutself and ->kvm_send_ipi_all seems like an easy to way shave cycles for affected flows.

Not sure why shorthand is not used for native x2apic mode.

Why do you say so? native_send_call_func_ipi() checks if allbutself shorthand should be used and does so (even though the check can be more efficient - I’m looking at that code right now…)

Quoted 7 lines
Quoted 2 lines

Please continue to follow the apic/x2apic driver. Just apic_flat set APIC_DEST_ALLBUT/APIC_DEST_ALLINC to ICR.

Indeed - I was sure by the name that it does it correctly. That’s stupid.

I’ll add it to the patch-set I am working on (TLB shootdown improvements), if you don’t mind.

Original for hotplug cpu safe.
https://lwn.net/Articles/138365/
https://lwn.net/Articles/138368/
Not sure shortcut native support is acceptable, I will play my kvm_send_ipi_allbutself and kvm_send_ipi_all. :)

Quoted 8 lines
Quoted 2 lines

Yes, I saw these threads before. But I think the test in native_send_call_func_ipi() should take care of it.

Good news, https://git.kernel.org/pub/scm/linux/kernel/git/tip/tip.git/log/?h=WIP.x86/ipi
Thomas who also is the hotplug state machine author introduces
shorthands support to native kernel now, I will add the support to kvm_send_ipi_allbutself() and kvm_send_ipi_all() after his work
complete.

Hmm, should fallback to native shorthands when support.

Regards,
Wanpeng Li

author
#17 lore raw
In reply to Radim Krčmář
Quoted 11 lines · Radim Krčmář

2019-05-30 09:05+0800, Wanpeng Li:

Quoted 6 lines

The idea is from Xen, when sending a call-function IPI-many to vCPUs, yield if any of the IPI target vCPUs was preempted. 17% performance increasement of ebizzy benchmark can be observed in an over-subscribe environment. (w/ kvm-pv-tlb disabled, testing TLB flush call-function IPI-many since call-function is not easy to be trigged by userspace workload).

Have you checked if we could gain performance by having the yield as an extension to our PV IPI call?

It will extend irq disabled time in __send_ipi_mask(). In addition, sched yield can be used to optimize other synchronization primitives in guest I think.

Regards,
Wanpeng Li