In a KVM guest on arm64, performance counters interrupts have an
unnecessary overhead which slows down execution when using the "perf
record" command and limits the "perf record" sampling period.
The problem is that when a guest VM disables counters by clearing the
PMCR_EL0.E bit (bit 0), KVM will disable all counters defined in
PMCR_EL0 even if they are not enabled in PMCNTENSET_EL0.
KVM disables a counter by calling into the perf framework, in particular
by calling perf_event_create_kernel_counter() which is a time consuming
operation. So, for example, with a Neoverse N1 CPU core which has 6 event
counters and one cycle counter, KVM will always disable all 7 counters
even if only one is enabled.
This typically happens when using the "perf record" command in a guest
VM: perf will disable all event counters with PMCNTENTSET_EL0 and only
uses the cycle counter. And when using the "perf record" -F option with
a high profiling frequency, the overhead of KVM disabling all counters
instead of one on every counter interrupt becomes very noticeable.
The problem is fixed by having KVM disable only counters which are
enabled in PMCNTENSET_EL0. If a counter is not enabled in PMCNTENSET_EL0
then KVM will not enable it when setting PMCR_EL0.E and it will remain
disabled as long as it is not enabled in PMCNTENSET_EL0. So there is
effectively no need to disable a counter when clearing PMCR_EL0.E if it
is not enabled PMCNTENSET_EL0.
Signed-off-by: Alexandre Chartre <redacted>
---
The patch is based on https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pmu/reset-values
arch/arm64/kvm/pmu-emul.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
From: Robin Murphy <robin.murphy@arm.com> Date: 2021-07-12 15:45:57
On 2021-07-12 16:17, Alexandre Chartre wrote:
quoted hunk
In a KVM guest on arm64, performance counters interrupts have an
unnecessary overhead which slows down execution when using the "perf
record" command and limits the "perf record" sampling period.
The problem is that when a guest VM disables counters by clearing the
PMCR_EL0.E bit (bit 0), KVM will disable all counters defined in
PMCR_EL0 even if they are not enabled in PMCNTENSET_EL0.
KVM disables a counter by calling into the perf framework, in particular
by calling perf_event_create_kernel_counter() which is a time consuming
operation. So, for example, with a Neoverse N1 CPU core which has 6 event
counters and one cycle counter, KVM will always disable all 7 counters
even if only one is enabled.
This typically happens when using the "perf record" command in a guest
VM: perf will disable all event counters with PMCNTENTSET_EL0 and only
uses the cycle counter. And when using the "perf record" -F option with
a high profiling frequency, the overhead of KVM disabling all counters
instead of one on every counter interrupt becomes very noticeable.
The problem is fixed by having KVM disable only counters which are
enabled in PMCNTENSET_EL0. If a counter is not enabled in PMCNTENSET_EL0
then KVM will not enable it when setting PMCR_EL0.E and it will remain
disabled as long as it is not enabled in PMCNTENSET_EL0. So there is
effectively no need to disable a counter when clearing PMCR_EL0.E if it
is not enabled PMCNTENSET_EL0.
Signed-off-by: Alexandre Chartre <redacted>
---
The patch is based on https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pmu/reset-values
arch/arm64/kvm/pmu-emul.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
This looks suspiciously opposite of what it replaces; however did we
even need to do a bitwise operation here in the first place? Couldn't we
skip the cycle counter by just limiting the for_each_set_bit iteration
below to 31 bits?
Robin.
for_each_set_bit(i, &mask, 32)
kvm_pmu_set_counter_value(vcpu, i, 0);
}
base-commit: 83f870a663592797c576846db3611e0a1664eda2
From: Alexandru Elisei <hidden> Date: 2021-07-12 15:52:17
Hi Robin,
On 7/12/21 4:44 PM, Robin Murphy wrote:
On 2021-07-12 16:17, Alexandre Chartre wrote:
quoted
In a KVM guest on arm64, performance counters interrupts have an
unnecessary overhead which slows down execution when using the "perf
record" command and limits the "perf record" sampling period.
The problem is that when a guest VM disables counters by clearing the
PMCR_EL0.E bit (bit 0), KVM will disable all counters defined in
PMCR_EL0 even if they are not enabled in PMCNTENSET_EL0.
KVM disables a counter by calling into the perf framework, in particular
by calling perf_event_create_kernel_counter() which is a time consuming
operation. So, for example, with a Neoverse N1 CPU core which has 6 event
counters and one cycle counter, KVM will always disable all 7 counters
even if only one is enabled.
This typically happens when using the "perf record" command in a guest
VM: perf will disable all event counters with PMCNTENTSET_EL0 and only
uses the cycle counter. And when using the "perf record" -F option with
a high profiling frequency, the overhead of KVM disabling all counters
instead of one on every counter interrupt becomes very noticeable.
The problem is fixed by having KVM disable only counters which are
enabled in PMCNTENSET_EL0. If a counter is not enabled in PMCNTENSET_EL0
then KVM will not enable it when setting PMCR_EL0.E and it will remain
disabled as long as it is not enabled in PMCNTENSET_EL0. So there is
effectively no need to disable a counter when clearing PMCR_EL0.E if it
is not enabled PMCNTENSET_EL0.
Signed-off-by: Alexandre Chartre <redacted>
---
The patch is based on
https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pmu/reset-values
arch/arm64/kvm/pmu-emul.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
This looks suspiciously opposite of what it replaces;
It always sets the bit, which goes against the architecture and the code it was
replacing, yes.
however did we even need to do a bitwise operation here in the first place?
Couldn't we skip the cycle counter by just limiting the for_each_set_bit
iteration below to 31 bits?
To quote myself [1]:
"Entertained the idea of restricting the number of bits in for_each_set_bit() to
31 since Linux (and the architecture, to some degree) treats the cycle count
register as the 32nd event counter. Settled on this approach because I think it's
clearer."
To expand on that, incorrectly resetting the cycle counter was introduced by a
refactoring, so I preferred making it very clear that PMCR_EL0.P is not supposed
to clear the cycle counter.
[1] https://lore.kernel.org/kvmarm/20210618105139.83795-1-alexandru.elisei@arm.com/
Thanks,
Alex
Robin.
quoted
for_each_set_bit(i, &mask, 32)
kvm_pmu_set_counter_value(vcpu, i, 0);
}
base-commit: 83f870a663592797c576846db3611e0a1664eda2
In a KVM guest on arm64, performance counters interrupts have an
unnecessary overhead which slows down execution when using the "perf
record" command and limits the "perf record" sampling period.
The problem is that when a guest VM disables counters by clearing the
PMCR_EL0.E bit (bit 0), KVM will disable all counters defined in
PMCR_EL0 even if they are not enabled in PMCNTENSET_EL0.
KVM disables a counter by calling into the perf framework, in particular
by calling perf_event_create_kernel_counter() which is a time consuming
operation. So, for example, with a Neoverse N1 CPU core which has 6 event
counters and one cycle counter, KVM will always disable all 7 counters
even if only one is enabled.
This typically happens when using the "perf record" command in a guest
VM: perf will disable all event counters with PMCNTENTSET_EL0 and only
uses the cycle counter. And when using the "perf record" -F option with
a high profiling frequency, the overhead of KVM disabling all counters
instead of one on every counter interrupt becomes very noticeable.
The problem is fixed by having KVM disable only counters which are
enabled in PMCNTENSET_EL0. If a counter is not enabled in PMCNTENSET_EL0
then KVM will not enable it when setting PMCR_EL0.E and it will remain
disabled as long as it is not enabled in PMCNTENSET_EL0. So there is
effectively no need to disable a counter when clearing PMCR_EL0.E if it
is not enabled PMCNTENSET_EL0.
Signed-off-by: Alexandre Chartre <redacted>
---
The patch is based on
https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pmu/reset-values
arch/arm64/kvm/pmu-emul.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
This looks suspiciously opposite of what it replaces;
It always sets the bit, which goes against the architecture and the code it was
replacing, yes.
My bad, I screw up and I dropped the ~. I will resend.
Sorry,
alex.
quoted
however did we even need to do a bitwise operation here in the first place?
Couldn't we skip the cycle counter by just limiting the for_each_set_bit
iteration below to 31 bits?
To quote myself [1]:
"Entertained the idea of restricting the number of bits in for_each_set_bit() to
31 since Linux (and the architecture, to some degree) treats the cycle count
register as the 32nd event counter. Settled on this approach because I think it's
clearer."
To expand on that, incorrectly resetting the cycle counter was introduced by a
refactoring, so I preferred making it very clear that PMCR_EL0.P is not supposed
to clear the cycle counter.
[1] https://lore.kernel.org/kvmarm/20210618105139.83795-1-alexandru.elisei@arm.com/
Thanks,
Alex
quoted
Robin.
quoted
for_each_set_bit(i, &mask, 32)
kvm_pmu_set_counter_value(vcpu, i, 0);
}
base-commit: 83f870a663592797c576846db3611e0a1664eda2
From: Robin Murphy <robin.murphy@arm.com> Date: 2021-07-12 16:04:34
On 2021-07-12 16:51, Alexandru Elisei wrote:
Hi Robin,
On 7/12/21 4:44 PM, Robin Murphy wrote:
quoted
On 2021-07-12 16:17, Alexandre Chartre wrote:
quoted
In a KVM guest on arm64, performance counters interrupts have an
unnecessary overhead which slows down execution when using the "perf
record" command and limits the "perf record" sampling period.
The problem is that when a guest VM disables counters by clearing the
PMCR_EL0.E bit (bit 0), KVM will disable all counters defined in
PMCR_EL0 even if they are not enabled in PMCNTENSET_EL0.
KVM disables a counter by calling into the perf framework, in particular
by calling perf_event_create_kernel_counter() which is a time consuming
operation. So, for example, with a Neoverse N1 CPU core which has 6 event
counters and one cycle counter, KVM will always disable all 7 counters
even if only one is enabled.
This typically happens when using the "perf record" command in a guest
VM: perf will disable all event counters with PMCNTENTSET_EL0 and only
uses the cycle counter. And when using the "perf record" -F option with
a high profiling frequency, the overhead of KVM disabling all counters
instead of one on every counter interrupt becomes very noticeable.
The problem is fixed by having KVM disable only counters which are
enabled in PMCNTENSET_EL0. If a counter is not enabled in PMCNTENSET_EL0
then KVM will not enable it when setting PMCR_EL0.E and it will remain
disabled as long as it is not enabled in PMCNTENSET_EL0. So there is
effectively no need to disable a counter when clearing PMCR_EL0.E if it
is not enabled PMCNTENSET_EL0.
Signed-off-by: Alexandre Chartre <redacted>
---
The patch is based on
https://git.kernel.org/pub/scm/linux/kernel/git/maz/arm-platforms.git/log/?h=kvm-arm64/pmu/reset-values
arch/arm64/kvm/pmu-emul.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
This looks suspiciously opposite of what it replaces;
It always sets the bit, which goes against the architecture and the code it was
replacing, yes.
quoted
however did we even need to do a bitwise operation here in the first place?
Couldn't we skip the cycle counter by just limiting the for_each_set_bit
iteration below to 31 bits?
To quote myself [1]:
"Entertained the idea of restricting the number of bits in for_each_set_bit() to
31 since Linux (and the architecture, to some degree) treats the cycle count
register as the 32nd event counter.
FWIW I wouldn't say there's any degree to it - we're iterating over the
bits in a register where the cycle counter enable is unequivocally the
32nd bit.
Settled on this approach because I think it's
clearer."
To expand on that, incorrectly resetting the cycle counter was introduced by a
refactoring, so I preferred making it very clear that PMCR_EL0.P is not supposed
to clear the cycle counter.
Fair enough, but if this has turned out to be a contentious hot path
then masking the bit to zero and then deliberately iterating to see if
it's set (find_next_bit() isn't exactly free) adds up to more overhead
than a comment ;)
Robin.