KVM: arm64: Allow to limit number of PMU counters
Subsystems (4)
PATCH v2 1/1 KVM: arm64: Allow to limit number of PMU counters
We currently pass through the number of PMU counters that we have available in hardware to guests. So if my host supports 10 concurrently active PMU counters, my guest will be able to spawn 10 counters as well.
This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
To help that case, let's add a PMU attr that allows us to limit the number of PMU counters that we expose. With this patch in place, user space can keep some counters free for host use.
---
Because this patch touches the same code paths as the vPMU filtering one and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.
v1 -> v2:
- Add documentation
- Add read support
--- Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32 ++++++++++++++++++++++++++++++++ arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 + 5 files changed, 67 insertions(+), 3 deletions(-)
Documentation/virt/kvm/devices/vcpu.rst +25 -0
@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering the cycle counter is possible using event 0x11 (CPU_CYCLES). +1.4 ATTRIBUTE: KVM_ARM_VCPU_PMU_V3_NUM_EVENTS +--------------------------------------------- + +:Parameters: in kvm_device_attr.addr the address for the limit of concurrent + events is a pointer to an int + +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events + ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest can monitor. +This number is directly exposed as part of the PMCR_EL0 register. + +On vcpu creation, this attribute is set to the hardware limit of the current +platform. If you need to determine the hardware limit, you can read this +attribute before setting it. + +Restrictions: The default value for this property is the number of hardware +supported events. Only values that are smaller than the hardware limit can +be set. + 2. GROUP: KVM_ARM_VCPU_TIMER_CTRL =================================
arch/arm64/include/uapi/asm/kvm.h +4 -3
@@ -342,9 +342,10 @@ struct kvm_vcpu_events {/* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3#define KVM_ARM_VCPU_TIMER_CTRL 1#define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0#define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1
arch/arm64/kvm/pmu-emul.c +32 -0
@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)for (i =0 ; i <ARMV8_PMU_MAX_COUNTERS ; i++) pmu->pmc[i].idx = i; + + pmu->num_events =perf_num_counters () -1 ; }/**
@@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)return 0 ; } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +u64 mask =ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (get_user (num_events, uaddr)) +return -EFAULT ; + +if (num_events >=perf_num_counters ()) +return -EINVAL ; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<=ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) &= ~mask; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) |= num_events; + +return 0 ; + }case KVM_ARM_VCPU_PMU_V3_INIT :return kvm_arm_pmu_v3_init (vcpu); }
@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num;return put_user (irq, uaddr); } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (!test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features)) +return -ENODEV ; + + num_events = vcpu->arch.pmu.num_events; +return put_user (num_events, uaddr); + } }return -ENXIO ;
@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case KVM_ARM_VCPU_PMU_V3_IRQ :case KVM_ARM_VCPU_PMU_V3_INIT :case KVM_ARM_VCPU_PMU_V3_FILTER : +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS :if (kvm_arm_support_pmu_v3 () &&test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features))return 0 ;
arch/arm64/kvm/sys_regs.c +5 -0
@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r) | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) & (~ARMV8_PMU_PMCR_E);if (!system_supports_32bit_el0 ()) val |=ARMV8_PMU_PMCR_LC ; + +/* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ); + val |= (u32 )vcpu->arch.pmu.num_events <<ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu, r->reg) = val; }
include/kvm/arm_pmu.h +1 -0
@@ -27,6 +27,7 @@ struct kvm_pmu {bool ready;bool created;bool irq_level; +u8 num_events; };#define kvm_arm_pmu_v3_ready (v) ((v)->arch.pmu.ready)
Signature · 12 lines
2.16.4 Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
Quoted 11 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available in hardware to guests. So if my host supports 10 concurrently active PMU counters, my guest will be able to spawn 10 counters as well.
This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
To help that case, let's add a PMU attr that allows us to limit the number of PMU counters that we expose. With this patch in place, user space can keep some counters free for host use.
Hi Alex,
Is there any reason to use the device API instead of just giving the user control over the necessary PMCR_EL0 bits through set/get-one-reg?
Thanks,
drew
Quoted 183 lines
Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.
v1 -> v2:
- Add documentation
- Add read support--- Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32 ++++++++++++++++++++++++++++++++ arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 + 5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +25
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering the cycle counter is possible using event 0x11 (CPU_CYCLES). +1.4 ATTRIBUTE: KVM_ARM_VCPU_PMU_V3_NUM_EVENTS +--------------------------------------------- + +:Parameters: in kvm_device_attr.addr the address for the limit of concurrent + events is a pointer to an int + +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events + ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest can monitor. +This number is directly exposed as part of the PMCR_EL0 register. + +On vcpu creation, this attribute is set to the hardware limit of the current +platform. If you need to determine the hardware limit, you can read this +attribute before setting it. + +Restrictions: The default value for this property is the number of hardware +supported events. Only values that are smaller than the hardware limit can +be set. + 2. GROUP: KVM_ARM_VCPU_TIMER_CTRL =================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events {/* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3#define KVM_ARM_VCPU_TIMER_CTRL 1#define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0#define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +32
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)for (i =0 ; i <ARMV8_PMU_MAX_COUNTERS ; i++) pmu->pmc[i].idx = i; + + pmu->num_events =perf_num_counters () -1 ; }/** @@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)return 0 ; } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +u64 mask =ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (get_user (num_events, uaddr)) +return -EFAULT ; + +if (num_events >=perf_num_counters ()) +return -EINVAL ; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<=ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) &= ~mask; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) |= num_events; + +return 0 ; + }case KVM_ARM_VCPU_PMU_V3_INIT :return kvm_arm_pmu_v3_init (vcpu); }@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num;return put_user (irq, uaddr); } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (!test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features)) +return -ENODEV ; + + num_events = vcpu->arch.pmu.num_events; +return put_user (num_events, uaddr); + } }return -ENXIO ;@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case KVM_ARM_VCPU_PMU_V3_IRQ :case KVM_ARM_VCPU_PMU_V3_INIT :case KVM_ARM_VCPU_PMU_V3_FILTER : +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS :if (kvm_arm_support_pmu_v3 () &&test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features))return 0 ;arch/arm64/kvm/sys_regs.c +5
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r) | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) & (~ARMV8_PMU_PMCR_E);if (!system_supports_32bit_el0 ()) val |=ARMV8_PMU_PMCR_LC ; + +/* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ); + val |= (u32 )vcpu->arch.pmu.num_events <<ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu, r->reg) = val; }include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu {bool ready;bool created;bool irq_level; +u8 num_events; };#define kvm_arm_pmu_v3_ready (v) ((v)->arch.pmu.ready)Signature · 12 lines
2.16.4 Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Hey Drew!
Quoted 10 lines · Andrew Jones
Quoted 8 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available in hardware to guests. So if my host supports 10 concurrently active PMU counters, my guest will be able to spawn 10 counters as well.
This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
Quoted 8 lines
Quoted 3 lines
To help that case, let's add a PMU attr that allows us to limit the number of PMU counters that we expose. With this patch in place, user space can keep some counters free for host use.
Hi Alex,
Is there any reason to use the device API instead of just giving the user control over the necessary
PMCR_EL0bits through set/get-one-reg?
I mostly used the attr interface because I was in that particular mental mode after looking at the filtering bits :).
Today, the PMCR_EL0 register gets reset implicitly on every vcpu reset call. How would we persist the counter field across resets? Would we in the first place?
I'm slightly hazy how the ONE_REG API would look like here. Do you have recommendations?
Alex
Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879
Quoted 25 lines · Alexander Graf
Hey Drew!
Quoted 10 lines · Andrew Jones
Quoted 8 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available in hardware to guests. So if my host supports 10 concurrently active PMU counters, my guest will be able to spawn 10 counters as well.
This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
Quoted 8 lines
Quoted 3 lines
To help that case, let's add a PMU attr that allows us to limit the number of PMU counters that we expose. With this patch in place, user space can keep some counters free for host use.
Hi Alex,
Is there any reason to use the device API instead of just giving the user control over the necessary
PMCR_EL0bits through set/get-one-reg?I mostly used the attr interface because I was in that particular mental mode after looking at the filtering bits :).
Quoted 7 lines
Today, the
PMCR_EL0register gets reset implicitly on every vcpu reset call. How would we persist the counter field across resets? Would we in the first place?I'm slightly hazy how the
ONE_REGAPI would look like here. Do you have recommendations?
Using the set/get_user hooks of the sysreg table we can accept a user input PMCR_EL0. We would only accept one that matches what the hardware and KVM supports though (EINVAL otherwise). We'll need to modify reset to use the value selected by the user too, which we can store in 'val' of the sysreg table.
Since userspace will likely get before set in order to know what's valid, we'll need to provide the current reset state on get until it has been set. I'm not sure how to track whether it has been set or not. Maybe new state is needed or an initial val=0 or val=~0 may work.
Thanks,
drew
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Hi Alex,
Quoted 7 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available in hardware to guests. So if my host supports 10 concurrently active PMU counters, my guest will be able to spawn 10 counters as well.
This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
but don't we have a trap and emulate approach as opposed to current SPE implementation?
Quoted 35 lines
To help that case, let's add a PMU attr that allows us to limit the number of PMU counters that we expose. With this patch in place, user space can keep some counters free for host use.
Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.
v1 -> v2:
- Add documentation
- Add read support--- Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32 ++++++++++++++++++++++++++++++++ arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 + 5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +2
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering the cycle counter is possible using event 0x11 (CPU_CYCLES). +1.4 ATTRIBUTE: KVM_ARM_VCPU_PMU_V3_NUM_EVENTS +---------------------------------------------
Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 116
Quoted 10 lines
Documentation/virt/kvm/devices/vcpu.rst +10
-0+ +:Parameters: in kvm_device_attr.addr the address for the limit of concurrent + events is a pointer to an int + +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events
s/events/event counters
I see that in perf code indeed num_events is used for that but I think for the end-user the event counter terminology is better as it fits the ARM spec.
Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 119
Quoted 3 lines
Documentation/virt/kvm/devices/vcpu.rst +3
-0+ ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest can monitor.
here also
Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 120
Quoted 1 line
Documentation/virt/kvm/devices/vcpu.rst +1
-0+This number is directly exposed as part of the PMCR_EL0 register.
Maybe quote the "N" field
Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 127
Quoted 7 lines
Documentation/virt/kvm/devices/vcpu.rst +7
-0+ +On vcpu creation, this attribute is set to the hardware limit of the current +platform. If you need to determine the hardware limit, you can read this +attribute before setting it. + +Restrictions: The default value for this property is the number of hardware +supported events. Only values that are smaller than the hardware limit can
event counters
Quoted 109 lines
Documentation/virt/kvm/devices/vcpu.rst +2
-0+be set. + 2. GROUP: KVM_ARM_VCPU_TIMER_CTRL =================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events {/* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3#define KVM_ARM_VCPU_TIMER_CTRL 1#define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0#define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +32
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)for (i =0 ; i <ARMV8_PMU_MAX_COUNTERS ; i++) pmu->pmc[i].idx = i; + + pmu->num_events =perf_num_counters () -1 ; }/** @@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)return 0 ; } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +u64 mask =ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (get_user (num_events, uaddr)) +return -EFAULT ; + +if (num_events >=perf_num_counters ()) +return -EINVAL ; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<=ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) &= ~mask; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) |= num_events; + +return 0 ; + }case KVM_ARM_VCPU_PMU_V3_INIT :return kvm_arm_pmu_v3_init (vcpu); }@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num;return put_user (irq, uaddr); } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (!test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features)) +return -ENODEV ; + + num_events = vcpu->arch.pmu.num_events; +return put_user (num_events, uaddr); + } }return -ENXIO ;@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case KVM_ARM_VCPU_PMU_V3_IRQ :case KVM_ARM_VCPU_PMU_V3_INIT :case KVM_ARM_VCPU_PMU_V3_FILTER : +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS :if (kvm_arm_support_pmu_v3 () &&test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features))return 0 ;arch/arm64/kvm/sys_regs.c +5
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r) | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) & (~ARMV8_PMU_PMCR_E);if (!system_supports_32bit_el0 ()) val |=ARMV8_PMU_PMCR_LC ; + +/* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ); + val |= (u32 )vcpu->arch.pmu.num_events <<ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu, r->reg) = val; }include/kvm/arm_pmu.h +0
-0
include/kvm/arm_pmu.h struct kvm_pmu { · line 33
Quoted 7 lines
include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu {bool ready;bool created;bool irq_level; +u8 num_events; };#define kvm_arm_pmu_v3_ready (v) ((v)->arch.pmu.ready)
Thanks
Eric
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Hi Alex,
Quoted 7 lines · Auger Eric
Hi Alex,
Quoted 4 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available in hardware to guests. So if my host supports 10 concurrently active PMU counters, my guest will be able to spawn 10 counters as well.
Quoted 5 lines
Quoted 3 lines
This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
but don't we have a trap and emulate approach as opposed to current SPE implementation?
Looking again at the switch code (__pmu_switch_to_guest), I see we just enable/disable the event counters owned by either the host/guest. I thought this was more involved. So now it is clear.
Thanks
Eric
Quoted 178 lines
Quoted 35 lines
To help that case, let's add a PMU attr that allows us to limit the number of PMU counters that we expose. With this patch in place, user space can keep some counters free for host use.
Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.
v1 -> v2:
- Add documentation
- Add read support--- Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32 ++++++++++++++++++++++++++++++++ arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 + 5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +2
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering the cycle counter is possible using event 0x11 (CPU_CYCLES). +1.4 ATTRIBUTE: KVM_ARM_VCPU_PMU_V3_NUM_EVENTS +---------------------------------------------Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 116
Quoted 10 lines
Documentation/virt/kvm/devices/vcpu.rst +10
-0+ +:Parameters: in kvm_device_attr.addr the address for the limit of concurrent + events is a pointer to an int + +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of eventss/events/event counters
I see that in perf code indeed
num_eventsis used for that but I think for the end-user the event counter terminology is better as it fits the ARM spec.Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 119
Quoted 3 lines
Documentation/virt/kvm/devices/vcpu.rst +3
-0+ ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest can monitor.here also
Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 120
Quoted 1 line
Documentation/virt/kvm/devices/vcpu.rst +1
-0+This number is directly exposed as part of the PMCR_EL0 register.Maybe quote the "N" field
Documentation/virt/kvm/devices/vcpu.rst isn't strictly speaking an event. Filtering the cycle counter is possible · line 127
Quoted 7 lines
Documentation/virt/kvm/devices/vcpu.rst +7
-0+ +On vcpu creation, this attribute is set to the hardware limit of the current +platform. If you need to determine the hardware limit, you can read this +attribute before setting it. + +Restrictions: The default value for this property is the number of hardware +supported events. Only values that are smaller than the hardware limit canevent counters
Quoted 114 lines
Documentation/virt/kvm/devices/vcpu.rst +2
-0+be set. + 2. GROUP: KVM_ARM_VCPU_TIMER_CTRL =================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events {/* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3#define KVM_ARM_VCPU_TIMER_CTRL 1#define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0#define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +32
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)for (i =0 ; i <ARMV8_PMU_MAX_COUNTERS ; i++) pmu->pmc[i].idx = i; + + pmu->num_events =perf_num_counters () -1 ; }/** @@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)return 0 ; } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +u64 mask =ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (get_user (num_events, uaddr)) +return -EFAULT ; + +if (num_events >=perf_num_counters ()) +return -EINVAL ; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<=ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) &= ~mask; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) |= num_events; + +return 0 ; + }case KVM_ARM_VCPU_PMU_V3_INIT :return kvm_arm_pmu_v3_init (vcpu); }@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num;return put_user (irq, uaddr); } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (!test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features)) +return -ENODEV ; + + num_events = vcpu->arch.pmu.num_events; +return put_user (num_events, uaddr); + } }return -ENXIO ;@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case KVM_ARM_VCPU_PMU_V3_IRQ :case KVM_ARM_VCPU_PMU_V3_INIT :case KVM_ARM_VCPU_PMU_V3_FILTER : +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS :if (kvm_arm_support_pmu_v3 () &&test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features))return 0 ;arch/arm64/kvm/sys_regs.c +5
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r) | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) & (~ARMV8_PMU_PMCR_E);if (!system_supports_32bit_el0 ()) val |=ARMV8_PMU_PMCR_LC ; + +/* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ); + val |= (u32 )vcpu->arch.pmu.num_events <<ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu, r->reg) = val; }include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu {bool ready;bool created;bool irq_level; +u8 num_events; };
Quoted 7 lines
include/kvm/arm_pmu.h struct kvm_pmu { · line 33
Quoted 2 lines
include/kvm/arm_pmu.h +0
-0#define kvm_arm_pmu_v3_ready (v) ((v)->arch.pmu.ready)Thanks
Eric
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Marc Zyngier
maintainerQuoted 185 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available in hardware to guests. So if my host supports 10 concurrently active PMU counters, my guest will be able to spawn 10 counters as well.
This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
To help that case, let's add a PMU attr that allows us to limit the number of PMU counters that we expose. With this patch in place, user space can keep some counters free for host use.
Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.
v1 -> v2:
- Add documentation
- Add read support--- Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32 ++++++++++++++++++++++++++++++++ arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 + 5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +0
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering thecycle counter is possible
using event0x11(CPU_CYCLES).+1.4 ATTRIBUTE:
KVM_ARM_VCPU_PMU_V3_NUM_EVENTS
+---------------------------------------------
+
+:Parameters: in kvm_device_attr.addr the address for the limit of concurrentDocumentation/virt/kvm/devices/vcpu.rst +11
-0+ events is a pointer to an int + +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events + ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest canmonitor.
Documentation/virt/kvm/devices/vcpu.rst +3
-0+This number is directly exposed as part of the PMCR_EL0 register. + +On vcpu creation, this attribute is set to the hardware limit of thecurrent
+platform. If you need to determine the hardware limit, you can read thisDocumentation/virt/kvm/devices/vcpu.rst +3
-0+attribute before setting it. + +Restrictions: The default value for this property is the number ofhardware
+supported events. Only values that are smaller than the hardware limit canDocumentation/virt/kvm/devices/vcpu.rst +2
-0+be set. + 2. GROUP: KVM_ARM_VCPU_TIMER_CTRL =================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events {/* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3#define KVM_ARM_VCPU_TIMER_CTRL 1#define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0#define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +21
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)for (i =0 ; i <ARMV8_PMU_MAX_COUNTERS ; i++) pmu->pmc[i].idx = i; + + pmu->num_events =perf_num_counters () -1 ; }/** @@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)return 0 ; } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +u64 mask =ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (get_user (num_events, uaddr)) +return -EFAULT ; + +if (num_events >=perf_num_counters ()) +return -EINVAL ; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<=ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) &= ~mask; +__vcpu_sys_reg (vcpu,SYS_PMCR_EL0 ) |= num_events; + +return 0 ; + }case KVM_ARM_VCPU_PMU_V3_INIT :return kvm_arm_pmu_v3_init (vcpu);arch/arm64/kvm/pmu-emul.c +10
-0}@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num;return put_user (irq, uaddr); } +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS : { +int __user *uaddr = (int __user *)(long )attr->addr; +u32 num_events; + +if (!test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features)) +return -ENODEV ; + + num_events = vcpu->arch.pmu.num_events; +return put_user (num_events, uaddr); + } }arch/arm64/kvm/pmu-emul.c +1
-0return -ENXIO ;@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case KVM_ARM_VCPU_PMU_V3_IRQ :case KVM_ARM_VCPU_PMU_V3_INIT :case KVM_ARM_VCPU_PMU_V3_FILTER : +case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS :if (kvm_arm_support_pmu_v3 () &&test_bit (KVM_ARM_VCPU_PMU_V3 , vcpu->arch.features))arch/arm64/kvm/pmu-emul.c +0
-0return 0 ;arch/arm64/kvm/sys_regs.c +5
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r) | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) & (~ARMV8_PMU_PMCR_E);if (!system_supports_32bit_el0 ()) val |=ARMV8_PMU_PMCR_LC ; + +/* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ); + val |= (u32 )vcpu->arch.pmu.num_events <<ARMV8_PMU_PMCR_N_SHIFT ; +__vcpu_sys_reg (vcpu, r->reg) = val; }include/kvm/arm_pmu.h +0
-0
include/kvm/arm_pmu.h struct kvm_pmu { · line 33
Quoted 7 lines
include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu {bool ready;bool created;bool irq_level; +u8 num_events; };#define kvm_arm_pmu_v3_ready (v) ((v)->arch.pmu.ready)
I see several problems with this approach:
- userspace doesn't really have a good way to retrieve the number of counters.
- Limiting the number of counters for the guest doesn't mean anything when it comes to the actual use of the HW counters, given that we don't allocate them ourselves (it's all perf doing the actual work).
- If you want to "pin" counters for the host, why don't you just do that before starting the guest?
I think you need to look at the bigger picture: how to limit the use of physical counter usage for a given userspace task. This needs to happen in perf itself, and not in KVM.
M.
Signature · 5 lines
Jazz is not dead. It just smells funny... _______________________________________________ kvmarm mailing list [email protected] https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Quoted 199 lines · Marc Zyngier
Quoted 197 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available
in hardware to guests. So if my host supports 10 concurrently active PMU
counters, my guest will be able to spawn 10 counters as well.This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
To help that case, let's add a PMU attr that allows us to limit the number
of PMU counters that we expose. With this patch in place, user space can
keep some counters free for host use.Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one
and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.v1 -> v2:
- Add documentation
- Add read support
---
Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32
++++++++++++++++++++++++++++++++
arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 +5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +0
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering thecycle counter is possible
using event0x11(CPU_CYCLES).+1.4 ATTRIBUTE:
KVM_ARM_VCPU_PMU_V3_NUM_EVENTS
+---------------------------------------------
+
+:Parameters: in kvm_device_attr.addr the address for the limit of concurrent
+ events is a pointer to an intDocumentation/virt/kvm/devices/vcpu.rst +10
-0+ +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events + ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest canmonitor.
Documentation/virt/kvm/devices/vcpu.rst +3
-0+This number is directly exposed as part of the PMCR_EL0 register. + +On vcpu creation, this attribute is set to the hardware limit of thecurrent
+platform. If you need to determine the hardware limit, you can read thisDocumentation/virt/kvm/devices/vcpu.rst +3
-0+attribute before setting it. + +Restrictions: The default value for this property is the number ofhardware
+supported events. Only values that are smaller than the hardware limit canDocumentation/virt/kvm/devices/vcpu.rst +2
-0+be set. +2. GROUP:
KVM_ARM_VCPU_TIMER_CTRL
=================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events { /* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3 #define KVM_ARM_VCPU_TIMER_CTRL 1 #define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0 #define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +21
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu) for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) pmu->pmc[i].idx = i; + + pmu->num_events = perf_num_counters() - 1; } /**@@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) return 0; } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + u64 mask = ARMV8_PMU_PMCR_N_MASK << ARMV8_PMU_PMCR_N_SHIFT; + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (get_user(num_events, uaddr)) + return -EFAULT; + + if (num_events >= perf_num_counters()) + return -EINVAL; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<= ARMV8_PMU_PMCR_N_SHIFT; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) &= ~mask; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) |= num_events; + + return 0; + }case
KVM_ARM_VCPU_PMU_V3_INIT:
returnkvm_arm_pmu_v3_init(vcpu);
}arch/arm64/kvm/pmu-emul.c +10
-0@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num; return put_user(irq, uaddr); } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) + return -ENODEV; + + num_events = vcpu->arch.pmu.num_events; + return put_user(num_events, uaddr); + } }return
-ENXIO;arch/arm64/kvm/pmu-emul.c +0
-0@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case
KVM_ARM_VCPU_PMU_V3_IRQ:
caseKVM_ARM_VCPU_PMU_V3_INIT:
caseKVM_ARM_VCPU_PMU_V3_FILTER:
+ caseKVM_ARM_VCPU_PMU_V3_NUM_EVENTS:
if (kvm_arm_support_pmu_v3()&&test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) return 0;arch/arm64/kvm/sys_regs.c +5
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r) | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) & (~ARMV8_PMU_PMCR_E); if (!system_supports_32bit_el0()) val |= ARMV8_PMU_PMCR_LC; + + /* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK << ARMV8_PMU_PMCR_N_SHIFT); + val |= (u32)vcpu->arch.pmu.num_events << ARMV8_PMU_PMCR_N_SHIFT; + __vcpu_sys_reg(vcpu, r->reg) = val; }include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu { bool ready; bool created; bool irq_level; + u8 num_events; };
Quoted 6 lines
include/kvm/arm_pmu.h struct kvm_pmu { · line 33
Quoted 1 line
include/kvm/arm_pmu.h +0
-0#define kvm_arm_pmu_v3_ready(v) ((v)->arch.pmu.ready)I see several problems with this approach:
- userspace doesn't really have a good way to retrieve the number of counters.
It does with v2, because it can then just read the register ;). I agree that it's clunky though.
Quoted 4 lines
- Limiting the number of counters for the guest doesn't mean anything when it comes to the actual use of the HW counters, given that we don't allocate them ourselves (it's all perf doing the actual work).
We do cap the number of actively requestable counters via perf by the PMCR.N limit. So in a way, it does mean something.
Quoted 2 lines
- If you want to "pin" counters for the host, why don't you just do that before starting the guest?
You can do that. Imagine I have 10 counters. I pin 4 of them to the host. I still tell my guest that it can use 6. That means perf will then time slice and juggle 10 guest event counters on those remaining 6 hardware counters. That juggling heavily reduces accuracy.
Quoted 3 lines
I think you need to look at the bigger picture: how to limit the use of physical counter usage for a given userspace task. This needs to happen in perf itself, and not in KVM.
That's definitely another way to look at it that I agree with.
What we really want is to expose the number of counters the guest has available, not the number of counters hardware can support at maximum.
So in theory it would be enough to ask perf how many counters it does have free for me to consume without overcommitting. But that would potentially change between multiple invocations of KVM and thus break things like live migration, no?
Maybe what we really want is an interface to perf from user space to say "how many counters can you dedicate to me?" and "reserve them for me". Then user space could reserve them as dedicated counters and KVM would just need to either probe for the reservation or get told by user space what to expose via ONE_REG as Drew suggested. It'd be up to user space to ensure that the reservation matches the number of exposed counters then.
Thoughts?
Alex
Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879
Quoted 227 lines · Alexander Graf
Quoted 200 lines · Marc Zyngier
Quoted 198 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available
in hardware to guests. So if my host supports 10 concurrently active PMU
counters, my guest will be able to spawn 10 counters as well.This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
To help that case, let's add a PMU attr that allows us to limit the number
of PMU counters that we expose. With this patch in place, user space can
keep some counters free for host use.Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one
and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.v1 -> v2:
- Add documentation
- Add read support
---
Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32
++++++++++++++++++++++++++++++++
arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 +5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +0
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering thecycle counter is possible
using event0x11(CPU_CYCLES).+1.4 ATTRIBUTE:
KVM_ARM_VCPU_PMU_V3_NUM_EVENTS
+---------------------------------------------
+
+:Parameters: in kvm_device_attr.addr the address for the limit of concurrent
+ events is a pointer to an intDocumentation/virt/kvm/devices/vcpu.rst +10
-0+ +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events + ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest canmonitor.
Documentation/virt/kvm/devices/vcpu.rst +3
-0+This number is directly exposed as part of the PMCR_EL0 register. + +On vcpu creation, this attribute is set to the hardware limit of thecurrent
+platform. If you need to determine the hardware limit, you can read thisDocumentation/virt/kvm/devices/vcpu.rst +3
-0+attribute before setting it. + +Restrictions: The default value for this property is the number ofhardware
+supported events. Only values that are smaller than the hardware limit canDocumentation/virt/kvm/devices/vcpu.rst +2
-0+be set. +2. GROUP:
KVM_ARM_VCPU_TIMER_CTRL
=================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events { /* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3 #define KVM_ARM_VCPU_TIMER_CTRL 1 #define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0 #define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +21
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu) for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) pmu->pmc[i].idx = i; + + pmu->num_events = perf_num_counters() - 1; } /**@@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) return 0; } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + u64 mask = ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (get_user(num_events, uaddr)) + return -EFAULT; + + if (num_events >= perf_num_counters()) + return -EINVAL; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<= ARMV8_PMU_PMCR_N_SHIFT; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) &= ~mask; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) |= num_events; + + return 0; + }case
KVM_ARM_VCPU_PMU_V3_INIT:
returnkvm_arm_pmu_v3_init(vcpu);
}arch/arm64/kvm/pmu-emul.c +10
-0@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num; return put_user(irq, uaddr); } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) + return -ENODEV; + + num_events = vcpu->arch.pmu.num_events; + return put_user(num_events, uaddr); + } }return
-ENXIO;arch/arm64/kvm/pmu-emul.c +0
-0@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case
KVM_ARM_VCPU_PMU_V3_IRQ:
caseKVM_ARM_VCPU_PMU_V3_INIT:
caseKVM_ARM_VCPU_PMU_V3_FILTER:
+ caseKVM_ARM_VCPU_PMU_V3_NUM_EVENTS:
if (kvm_arm_support_pmu_v3()&&test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) return 0;arch/arm64/kvm/sys_regs.c +5
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r) | (ARMV8_PMU_PMCR_MASK & 0xdecafbad)) & (~ARMV8_PMU_PMCR_E); if (!system_supports_32bit_el0()) val |= ARMV8_PMU_PMCR_LC; + + /* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK << ARMV8_PMU_PMCR_N_SHIFT); + val |= (u32)vcpu->arch.pmu.num_events << ARMV8_PMU_PMCR_N_SHIFT; + __vcpu_sys_reg(vcpu, r->reg) = val; }include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu { bool ready; bool created; bool irq_level; + u8 num_events; };Quoted 6 lines
include/kvm/arm_pmu.h struct kvm_pmu { · line 33
Quoted 1 line
include/kvm/arm_pmu.h +0
-0#define kvm_arm_pmu_v3_ready(v) ((v)->arch.pmu.ready)I see several problems with this approach:
- userspace doesn't really have a good way to retrieve the number of counters.
It does with v2, because it can then just read the register ;). I agree that it's clunky though.
Quoted 4 lines
- Limiting the number of counters for the guest doesn't mean anything when it comes to the actual use of the HW counters, given that we don't allocate them ourselves (it's all perf doing the actual work).
We do cap the number of actively requestable counters via perf by the PMCR.N limit. So in a way, it does mean something.
Quoted 2 lines
- If you want to "pin" counters for the host, why don't you just do that before starting the guest?
You can do that. Imagine I have 10 counters. I pin 4 of them to the host. I still tell my guest that it can use 6. That means perf will then time slice and juggle 10 guest event counters on those remaining 6 hardware counters. That juggling heavily reduces accuracy.
Quoted 3 lines
I think you need to look at the bigger picture: how to limit the use of physical counter usage for a given userspace task. This needs to happen in perf itself, and not in KVM.
Quoted 7 lines
That's definitely another way to look at it that I agree with.
What we really want is to expose the number of counters the guest has available, not the number of counters hardware can support at maximum.
So in theory it would be enough to ask perf how many counters it does have free for me to consume without overcommitting. But that would potentially change between multiple invocations of KVM and thus break things like live migration, no?
Maybe what we really want is an interface to perf from user space to say "how many counters can you dedicate to me?" and "reserve them for me". Then user space could reserve them as dedicated counters and KVM would just need to either probe for the reservation or get told by user space what to expose via
ONE_REGas Drew suggested. It'd be up to user space to ensure that the reservation matches the number of exposed counters then.
Note that if the aim is to avoid the guest seeing unexpectedly weird behaviour, then it's not just the *number* of counters that matters, but the underlying physical allocation too, thanks to the possibility of chained events.
Robin.
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Quoted 252 lines · Robin Murphy
Quoted 243 lines · Alexander Graf
Quoted 201 lines · Marc Zyngier
Quoted 199 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available
in hardware to guests. So if my host supports 10 concurrently active PMU
counters, my guest will be able to spawn 10 counters as well.This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
To help that case, let's add a PMU attr that allows us to limit the number
of PMU counters that we expose. With this patch in place, user space can
keep some counters free for host use.Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one
and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.v1 -> v2:
- Add documentation
- Add read support
---
Documentation/virt/kvm/devices/vcpu.rst | 25 +++++++++++++++++++++++++ arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32
++++++++++++++++++++++++++++++++
arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 +5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +0
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering thecycle counter is possible
using event0x11(CPU_CYCLES).+1.4 ATTRIBUTE:
KVM_ARM_VCPU_PMU_V3_NUM_EVENTS
+---------------------------------------------
+
+:Parameters: in kvm_device_attr.addr the address for the limit of concurrent
+ events is a pointer to an intDocumentation/virt/kvm/devices/vcpu.rst +10
-0+ +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events + ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest canmonitor.
Documentation/virt/kvm/devices/vcpu.rst +3
-0+This number is directly exposed as part of the PMCR_EL0 register. + +On vcpu creation, this attribute is set to the hardware limit of thecurrent
+platform. If you need to determine the hardware limit, you can read thisDocumentation/virt/kvm/devices/vcpu.rst +3
-0+attribute before setting it. + +Restrictions: The default value for this property is the number ofhardware
+supported events. Only values that are smaller than the hardware limit canDocumentation/virt/kvm/devices/vcpu.rst +2
-0+be set. +2. GROUP:
KVM_ARM_VCPU_TIMER_CTRL
=================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events { /* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3 #define KVM_ARM_VCPU_TIMER_CTRL 1 #define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0 #define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +21
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu) for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) pmu->pmc[i].idx = i; + + pmu->num_events = perf_num_counters() - 1; } /**@@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) return 0; } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + u64 mask = ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (get_user(num_events, uaddr)) + return -EFAULT; + + if (num_events >= perf_num_counters()) + return -EINVAL; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<= ARMV8_PMU_PMCR_N_SHIFT; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) &= ~mask; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) |= num_events; + + return 0; + }case
KVM_ARM_VCPU_PMU_V3_INIT:
returnkvm_arm_pmu_v3_init(vcpu);
}arch/arm64/kvm/pmu-emul.c +10
-0@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num; return put_user(irq, uaddr); } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) + return -ENODEV; + + num_events = vcpu->arch.pmu.num_events; + return put_user(num_events, uaddr); + } }return
-ENXIO;arch/arm64/kvm/pmu-emul.c +0
-0@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case
KVM_ARM_VCPU_PMU_V3_IRQ:
caseKVM_ARM_VCPU_PMU_V3_INIT:
caseKVM_ARM_VCPU_PMU_V3_FILTER:
+ caseKVM_ARM_VCPU_PMU_V3_NUM_EVENTS:
if (kvm_arm_support_pmu_v3()&&test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) return 0;arch/arm64/kvm/sys_regs.c +0
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r)| (
ARMV8_PMU_PMCR_MASK&0xdecafbad)) & (~ARMV8_PMU_PMCR_E);
if (!system_supports_32bit_el0())
val |=ARMV8_PMU_PMCR_LC;arch/arm64/kvm/sys_regs.c +5
-0+ + /* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK << ARMV8_PMU_PMCR_N_SHIFT); + val |= (u32)vcpu->arch.pmu.num_events << ARMV8_PMU_PMCR_N_SHIFT; +
__vcpu_sys_reg(vcpu, r->reg)= val;
}include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu { bool ready; bool created; bool irq_level; + u8 num_events; };Quoted 6 lines
include/kvm/arm_pmu.h struct kvm_pmu { · line 33
Quoted 1 line
include/kvm/arm_pmu.h +0
-0#define kvm_arm_pmu_v3_ready(v) ((v)->arch.pmu.ready)I see several problems with this approach:
- userspace doesn't really have a good way to retrieve the number of counters.
It does with v2, because it can then just read the register ;). I agree that it's clunky though.
Quoted 4 lines
- Limiting the number of counters for the guest doesn't mean anything when it comes to the actual use of the HW counters, given that we don't allocate them ourselves (it's all perf doing the actual work).
We do cap the number of actively requestable counters via perf by the PMCR.N limit. So in a way, it does mean something.
Quoted 2 lines
- If you want to "pin" counters for the host, why don't you just do that before starting the guest?
You can do that. Imagine I have 10 counters. I pin 4 of them to the host. I still tell my guest that it can use 6. That means perf will then time slice and juggle 10 guest event counters on those remaining 6 hardware counters. That juggling heavily reduces accuracy.
Quoted 3 lines
I think you need to look at the bigger picture: how to limit the use of physical counter usage for a given userspace task. This needs to happen in perf itself, and not in KVM.
That's definitely another way to look at it that I agree with.
What we really want is to expose the number of counters the guest has available, not the number of counters hardware can support at maximum.
So in theory it would be enough to ask perf how many counters it does have free for me to consume without overcommitting. But that would potentially change between multiple invocations of KVM and thus break things like live migration, no?
Quoted 6 lines
Maybe what we really want is an interface to perf from user space to say "how many counters can you dedicate to me?" and "reserve them for me". Then user space could reserve them as dedicated counters and KVM would just need to either probe for the reservation or get told by user space what to expose via
ONE_REGas Drew suggested. It'd be up to user space to ensure that the reservation matches the number of exposed counters then.
Quoted 4 lines
Note that if the aim is to avoid the guest seeing unexpectedly weird behaviour, then it's not just the *number* of counters that matters, but the underlying physical allocation too, thanks to the possibility of chained events.
Wouldn't ideally guest chaining propagate into host chaining as well? I'd have to double check if it does, but in my naive thinking if I reserve 4 hardware counters for the guest and the guest ends up using 4 hardware counters regardless of their chaining attributes, I'd still be able to fit them all?
Alex
Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
Quoted 257 lines · Alexander Graf
Quoted 254 lines · Robin Murphy
Quoted 251 lines · Alexander Graf
Quoted 203 lines · Marc Zyngier
Quoted 201 lines · Alexander Graf
We currently pass through the number of PMU counters that we have available
in hardware to guests. So if my host supports 10 concurrently active PMU
counters, my guest will be able to spawn 10 counters as well.This is undesireable if we also want to use the PMU on the host for monitoring. In that case, we want to split the PMU between guest and host.
To help that case, let's add a PMU attr that allows us to limit the number
of PMU counters that we expose. With this patch in place, user space can
keep some counters free for host use.Signed-off-by Alexander Graf <[email protected]>---
Because this patch touches the same code paths as the vPMU filtering one
and the vPMU filtering generalized a few conditions in the attr path, I've based it on top. Please let me know if you want it independent instead.v1 -> v2:
- Add documentation
- Add read support
---
Documentation/virt/kvm/devices/vcpu.rst | 25
+++++++++++++++++++++++++
arch/arm64/include/uapi/asm/kvm.h | 7 ++++--- arch/arm64/kvm/pmu-emul.c | 32
++++++++++++++++++++++++++++++++
arch/arm64/kvm/sys_regs.c | 5 +++++ include/kvm/arm_pmu.h | 1 +5 files changed, 67 insertions(+), 3 deletions(-)Documentation/virt/kvm/devices/vcpu.rst +0
-0@@ -102,6 +102,31 @@ isn't strictly speaking an event. Filtering thecycle counter is possible
using event0x11(CPU_CYCLES).+1.4 ATTRIBUTE:
KVM_ARM_VCPU_PMU_V3_NUM_EVENTS
+---------------------------------------------
+
+:Parameters: in kvm_device_attr.addr the address for the limit of concurrent
+ events is a pointer to an intDocumentation/virt/kvm/devices/vcpu.rst +10
-0+ +:Returns: + + ======= ====================================================== + -ENODEV: PMUv3 not supported + -EBUSY: PMUv3 already initialized + -EINVAL: Too large number of events + ======= ====================================================== + +Reconfigure the limit of concurrent PMU events that the guest canmonitor.
Documentation/virt/kvm/devices/vcpu.rst +3
-0+This number is directly exposed as part of the PMCR_EL0 register. + +On vcpu creation, this attribute is set to the hardware limit of thecurrent
+platform. If you need to determine the hardware limit, you can read thisDocumentation/virt/kvm/devices/vcpu.rst +3
-0+attribute before setting it. + +Restrictions: The default value for this property is the number ofhardware
+supported events. Only values that are smaller than the hardware limit
canDocumentation/virt/kvm/devices/vcpu.rst +2
-0+be set. +2. GROUP:
KVM_ARM_VCPU_TIMER_CTRL
=================================arch/arm64/include/uapi/asm/kvm.h +4
-3@@ -342,9 +342,10 @@ struct kvm_vcpu_events { /* Device Control API on vcpu fd */ #define KVM_ARM_VCPU_PMU_V3_CTRL 0 -#define KVM_ARM_VCPU_PMU_V3_IRQ 0 -#define KVM_ARM_VCPU_PMU_V3_INIT 1 -#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_IRQ 0 +#define KVM_ARM_VCPU_PMU_V3_INIT 1 +#define KVM_ARM_VCPU_PMU_V3_FILTER 2 +#define KVM_ARM_VCPU_PMU_V3_NUM_EVENTS 3 #define KVM_ARM_VCPU_TIMER_CTRL 1 #define KVM_ARM_VCPU_TIMER_IRQ_VTIMER 0 #define KVM_ARM_VCPU_TIMER_IRQ_PTIMER 1arch/arm64/kvm/pmu-emul.c +21
-0@@ -253,6 +253,8 @@ void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu) for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) pmu->pmc[i].idx = i; + + pmu->num_events = perf_num_counters() - 1; } /**@@ -978,6 +980,25 @@ int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) return 0; } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + u64 mask = ARMV8_PMU_PMCR_N_MASK <<ARMV8_PMU_PMCR_N_SHIFT ; + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (get_user(num_events, uaddr)) + return -EFAULT; + + if (num_events >= perf_num_counters()) + return -EINVAL; + + vcpu->arch.pmu.num_events = num_events; + + num_events <<= ARMV8_PMU_PMCR_N_SHIFT; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) &= ~mask; + __vcpu_sys_reg(vcpu, SYS_PMCR_EL0) |= num_events; + + return 0; + }case
KVM_ARM_VCPU_PMU_V3_INIT:
returnkvm_arm_pmu_v3_init(vcpu);
}arch/arm64/kvm/pmu-emul.c +10
-0@@ -1004,6 +1025,16 @@ int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr) irq = vcpu->arch.pmu.irq_num; return put_user(irq, uaddr); } + case KVM_ARM_VCPU_PMU_V3_NUM_EVENTS: { + int __user *uaddr = (int __user *)(long)attr->addr; + u32 num_events; + + if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) + return -ENODEV; + + num_events = vcpu->arch.pmu.num_events; + return put_user(num_events, uaddr); + } }return
-ENXIO;arch/arm64/kvm/pmu-emul.c +0
-0@@ -1015,6 +1046,7 @@ int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)case
KVM_ARM_VCPU_PMU_V3_IRQ:
caseKVM_ARM_VCPU_PMU_V3_INIT:
caseKVM_ARM_VCPU_PMU_V3_FILTER:
+ caseKVM_ARM_VCPU_PMU_V3_NUM_EVENTS:
if (kvm_arm_support_pmu_v3()&&test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features)) return 0;arch/arm64/kvm/sys_regs.c +0
-0@@ -672,6 +672,11 @@ static void reset_pmcr(struct kvm_vcpu *vcpu, const struct sys_reg_desc *r)| (
ARMV8_PMU_PMCR_MASK&0xdecafbad)) & (~ARMV8_PMU_PMCR_E);
if (!system_supports_32bit_el0())
val |=ARMV8_PMU_PMCR_LC;arch/arm64/kvm/sys_regs.c +5
-0+ + /* Override number of event selectors */ + val &= ~(ARMV8_PMU_PMCR_N_MASK << ARMV8_PMU_PMCR_N_SHIFT); + val |= (u32)vcpu->arch.pmu.num_events << ARMV8_PMU_PMCR_N_SHIFT; +
__vcpu_sys_reg(vcpu, r->reg)= val;
}include/kvm/arm_pmu.h +1
-0@@ -27,6 +27,7 @@ struct kvm_pmu { bool ready; bool created; bool irq_level; + u8 num_events; };Quoted 6 lines
include/kvm/arm_pmu.h struct kvm_pmu { · line 33
Quoted 1 line
include/kvm/arm_pmu.h +0
-0#define kvm_arm_pmu_v3_ready(v) ((v)->arch.pmu.ready)I see several problems with this approach:
- userspace doesn't really have a good way to retrieve the number of counters.
It does with v2, because it can then just read the register ;). I agree that it's clunky though.
Quoted 4 lines
- Limiting the number of counters for the guest doesn't mean anything when it comes to the actual use of the HW counters, given that we don't allocate them ourselves (it's all perf doing the actual work).
We do cap the number of actively requestable counters via perf by the PMCR.N limit. So in a way, it does mean something.
Quoted 2 lines
- If you want to "pin" counters for the host, why don't you just do that before starting the guest?
You can do that. Imagine I have 10 counters. I pin 4 of them to the host. I still tell my guest that it can use 6. That means perf will then time slice and juggle 10 guest event counters on those remaining 6 hardware counters. That juggling heavily reduces accuracy.
Quoted 3 lines
I think you need to look at the bigger picture: how to limit the use of physical counter usage for a given userspace task. This needs to happen in perf itself, and not in KVM.
That's definitely another way to look at it that I agree with.
What we really want is to expose the number of counters the guest has available, not the number of counters hardware can support at maximum.
So in theory it would be enough to ask perf how many counters it does have free for me to consume without overcommitting. But that would potentially change between multiple invocations of KVM and thus break things like live migration, no?
Maybe what we really want is an interface to perf from user space to say "how many counters can you dedicate to me?" and "reserve them for me". Then user space could reserve them as dedicated counters and KVM would just need to either probe for the reservation or get told by user space what to expose via
ONE_REGas Drew suggested. It'd be up to user space to ensure that the reservation matches the number of exposed counters then.
Quoted 6 lines
Quoted 4 lines
Note that if the aim is to avoid the guest seeing unexpectedly weird behaviour, then it's not just the *number* of counters that matters, but the underlying physical allocation too, thanks to the possibility of chained events.
Wouldn't ideally guest chaining propagate into host chaining as well? I'd have to double check if it does, but in my naive thinking if I reserve 4 hardware counters for the guest and the guest ends up using 4 hardware counters regardless of their chaining attributes, I'd still be able to fit them all?
It depends what you mean by "reserve" - if you merely tell the guest that the vPMU only has 4 counters, and at any given time your (n-4) host events happen to have ended up scheduled such that no physical even/odd pair of counters is free, then even a single chained event from the guest might trigger context rotation despite there being 4 or more counters free in total.
If you want truly deterministic results then you'll probably need some brain surgery on the PMU driver to actually partition the physical counters rather than simply relying on perf's scheduling.
Robin.
_______________________________________________
kvmarm mailing list
[email protected]
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm