Skip to contentSkip to conversation

kvm: irqfd: avoid update unmodified entries of the routing

PATCH Applied Longpeng(Mike) 2 participants 3 messages Aug 27, 2021 – Sep 23, 2021

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kvm: irqfd: avoid update unmodified entries of the routing

All of the irqfds would to be updated when update the irq routing, it's too expensive if there're too many irqfds.

However we can reduce the cost by avoid some unnecessary updates. For irqs of MSI type on X86, the update can be saved if the msi values are not change.

The vfio migration could receives benefit from this optimi- zaiton. The test VM has 128 vcpus and 8 VF (with 65 vectors enabled), so the VM has more than 520 irqfds. We mesure the cost of the vfio_msix_enable (in QEMU, it would set routing for each irqfd) for each VF, and we can see the total cost can be significantly reduced.

                Origin         Apply this Patch
1st             8              4
2nd             15             5
3rd             22             6
4th             24             6
5th             36             7
6th             44             7
7th             51             8
8th             58             8
Total           258ms          51ms

We're also tring to optimize the QEMU part [1], but it's still worth to optimize the KVM to gain more benefits.

[1] https://lists.gnu.org/archive/html/qemu-devel/2021-08/msg04215.html

---
 arch/x86/kvm/x86.c       |  9 +++++++++
 include/linux/kvm_host.h |  2 ++
 virt/kvm/eventfd.c       | 15 ++++++++++++++-
 3 files changed, 25 insertions(+), 1 deletion(-)
arch/x86/kvm/x86.c +9 -0
index e5d5c5e..22cf20e 100644
@@ -12023,6 +12023,15 @@ int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
 	return static_call(kvm_x86_update_pi_irte)(kvm, host_irq, guest_irq, set);
 }
 
+bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new)
+{
+	if (new->type != KVM_IRQ_ROUTING_MSI)
+		return true;
+
+	return !!memcmp(&old->msi, &new->msi, sizeof(new->msi));
+}
+
 bool kvm_vector_hashing_enabled(void)
 {
 	return vector_hashing;
include/linux/kvm_host.h +2 -0
index ae7735b..c0954ae 100644
@@ -1621,6 +1621,8 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *,
 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *);
 int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
 				  uint32_t guest_irq, bool set);
+bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *,
+				  struct kvm_kernel_irq_routing_entry *);
 #endif /* CONFIG_HAVE_KVM_IRQ_BYPASS */
 
 #ifdef CONFIG_HAVE_KVM_INVALID_WAKEUPS
virt/kvm/eventfd.c +14 -1
index e996989..2ad013b 100644
@@ -281,6 +281,13 @@ int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 {
 	return 0;
 }
+
+bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
+				struct kvm_kernel_irq_routing_entry *old,
+				struct kvm_kernel_irq_routing_entry *new)
+{
+	return true;
+}
 #endif
 
 static int
@@ -615,10 +622,16 @@ void kvm_irq_routing_update(struct kvm *kvm)
 	spin_lock_irq(&kvm->irqfds.lock);
 
 	list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
+#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
+		/* Under irqfds.lock, so can read irq_entry safely */
+		struct kvm_kernel_irq_routing_entry old = irqfd->irq_entry;
+#endif
+
 		irqfd_update(kvm, irqfd);
 
 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
-		if (irqfd->producer) {
+		if (irqfd->producer &&
+		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
 			int ret = kvm_arch_update_irqfd_routing(
 					irqfd->kvm, irqfd->producer->irq,
 					irqfd->gsi, 1);
Signature · 1 line
1.8.3.1
#1 lore raw

Hi guys,

Do you have any suggestions ? Thanks.

Quoted 107 lines · Longpeng(Mike)

All of the irqfds would to be updated when update the irq routing, it's too expensive if there're too many irqfds.

However we can reduce the cost by avoid some unnecessary updates. For irqs of MSI type on X86, the update can be saved if the msi values are not change.

The vfio migration could receives benefit from this optimi- zaiton. The test VM has 128 vcpus and 8 VF (with 65 vectors enabled), so the VM has more than 520 irqfds. We mesure the cost of the vfio_msix_enable (in QEMU, it would set routing for each irqfd) for each VF, and we can see the total cost can be significantly reduced.

                Origin         Apply this Patch
1st             8              4
2nd             15             5
3rd             22             6
4th             24             6
5th             36             7
6th             44             7
7th             51             8
8th             58             8
Total           258ms          51ms

We're also tring to optimize the QEMU part [1], but it's still worth to optimize the KVM to gain more benefits.

[1] https://lists.gnu.org/archive/html/qemu-devel/2021-08/msg04215.html

---
 arch/x86/kvm/x86.c       |  9 +++++++++
 include/linux/kvm_host.h |  2 ++
 virt/kvm/eventfd.c       | 15 ++++++++++++++-
 3 files changed, 25 insertions(+), 1 deletion(-)
arch/x86/kvm/x86.c +9 -0
index e5d5c5e..22cf20e 100644
@@ -12023,6 +12023,15 @@ int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
 	return static_call(kvm_x86_update_pi_irte)(kvm, host_irq, guest_irq, set);
 }
 
+bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new)
+{
+	if (new->type != KVM_IRQ_ROUTING_MSI)
+		return true;
+
+	return !!memcmp(&old->msi, &new->msi, sizeof(new->msi));
+}
+
 bool kvm_vector_hashing_enabled(void)
 {
 	return vector_hashing;
include/linux/kvm_host.h +2 -0
index ae7735b..c0954ae 100644
@@ -1621,6 +1621,8 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *,
 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *);
 int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
 				  uint32_t guest_irq, bool set);
+bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *,
+				  struct kvm_kernel_irq_routing_entry *);
 #endif /* CONFIG_HAVE_KVM_IRQ_BYPASS */
 
 #ifdef CONFIG_HAVE_KVM_INVALID_WAKEUPS
virt/kvm/eventfd.c +14 -1
index e996989..2ad013b 100644
@@ -281,6 +281,13 @@ int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
 {
 	return 0;
 }
+
+bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
+				struct kvm_kernel_irq_routing_entry *old,
+				struct kvm_kernel_irq_routing_entry *new)
+{
+	return true;
+}
 #endif
 
 static int
@@ -615,10 +622,16 @@ void kvm_irq_routing_update(struct kvm *kvm)
 	spin_lock_irq(&kvm->irqfds.lock);
 
 	list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
+#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
+		/* Under irqfds.lock, so can read irq_entry safely */
+		struct kvm_kernel_irq_routing_entry old = irqfd->irq_entry;
+#endif
+
 		irqfd_update(kvm, irqfd);
 
 #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
-		if (irqfd->producer) {
+		if (irqfd->producer &&
+		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
 			int ret = kvm_arch_update_irqfd_routing(
 					irqfd->kvm, irqfd->producer->irq,
 					irqfd->gsi, 1);
maintainer Applied
#2 lore raw
In reply to Longpeng(Mike)
Quoted 97 lines · Longpeng(Mike)

All of the irqfds would to be updated when update the irq routing, it's too expensive if there're too many irqfds.

However we can reduce the cost by avoid some unnecessary updates. For irqs of MSI type on X86, the update can be saved if the msi values are not change.

The vfio migration could receives benefit from this optimi- zaiton. The test VM has 128 vcpus and 8 VF (with 65 vectors enabled), so the VM has more than 520 irqfds. We mesure the cost of the vfio_msix_enable (in QEMU, it would set routing for each irqfd) for each VF, and we can see the total cost can be significantly reduced.

Origin Apply this Patch

1st             8              4
2nd             15             5
3rd             22             6
4th             24             6
5th             36             7
6th             44             7
7th             51             8
8th             58             8
Total           258ms          51ms

We're also tring to optimize the QEMU part [1], but it's still worth to optimize the KVM to gain more benefits.

[1] https://lists.gnu.org/archive/html/qemu-devel/2021-08/msg04215.html

---
  arch/x86/kvm/x86.c       |  9 +++++++++
  include/linux/kvm_host.h |  2 ++
  virt/kvm/eventfd.c       | 15 ++++++++++++++-
  3 files changed, 25 insertions(+), 1 deletion(-)
arch/x86/kvm/x86.c +9 -0
index e5d5c5e..22cf20e 100644
@@ -12023,6 +12023,15 @@ int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
  	return static_call(kvm_x86_update_pi_irte)(kvm, host_irq, guest_irq, set);
  }
 
+bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *old,
+				  struct kvm_kernel_irq_routing_entry *new)
+{
+	if (new->type != KVM_IRQ_ROUTING_MSI)
+		return true;
+
+	return !!memcmp(&old->msi, &new->msi, sizeof(new->msi));
+}
+
  bool kvm_vector_hashing_enabled(void)
  {
  	return vector_hashing;
include/linux/kvm_host.h +2 -0
index ae7735b..c0954ae 100644
@@ -1621,6 +1621,8 @@ void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *,
  void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *);
  int kvm_arch_update_irqfd_routing(struct kvm *kvm, unsigned int host_irq,
  				  uint32_t guest_irq, bool set);
+bool kvm_arch_irqfd_route_changed(struct kvm_kernel_irq_routing_entry *,
+				  struct kvm_kernel_irq_routing_entry *);
  #endif /* CONFIG_HAVE_KVM_IRQ_BYPASS */
 
  #ifdef CONFIG_HAVE_KVM_INVALID_WAKEUPS
virt/kvm/eventfd.c +11 -0
index e996989..2ad013b 100644
@@ -281,6 +281,13 @@ int  __attribute__((weak)) kvm_arch_update_irqfd_routing(
  {
  	return 0;
  }
+
+bool __attribute__((weak)) kvm_arch_irqfd_route_changed(
+				struct kvm_kernel_irq_routing_entry *old,
+				struct kvm_kernel_irq_routing_entry *new)
+{
+	return true;
+}
  #endif
 
  static int
@@ -615,10 +622,16 @@ void kvm_irq_routing_update(struct kvm *kvm)
  	spin_lock_irq(&kvm->irqfds.lock);
 
  	list_for_each_entry(irqfd, &kvm->irqfds.items, list) {
+#ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
+		/* Under irqfds.lock, so can read irq_entry safely */
+		struct kvm_kernel_irq_routing_entry old = irqfd->irq_entry;
+#endif

virt/kvm/eventfd.c void kvm_irq_routing_update(struct kvm *kvm) · line 637

Quoted 10 lines
virt/kvm/eventfd.c +3 -1
+
  		irqfd_update(kvm, irqfd);
 
  #ifdef CONFIG_HAVE_KVM_IRQ_BYPASS
-		if (irqfd->producer) {
+		if (irqfd->producer &&
+		    kvm_arch_irqfd_route_changed(&old, &irqfd->irq_entry)) {
  			int ret = kvm_arch_update_irqfd_routing(
  					irqfd->kvm, irqfd->producer->irq,
  					irqfd->gsi, 1);

Queued, thanks.

Paolo