| Message ID | 1519431578-11995-2-git-send-email-skannan@codeaurora.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 02/23/2018 04:19 PM, Saravana Kannan wrote: > Some PMUs events can be read from any CPU. So allow the PMU to mark > events as such. For these events, we don't need to reject reads or > make smp calls to the event's CPU and cause unnecessary wake ups. > > Good examples of such events would be events from caches shared across > all CPUs. > > Signed-off-by: Saravana Kannan <skannan@codeaurora.org> > --- > include/linux/perf_event.h | 3 +++ > kernel/events/core.c | 10 ++++++++-- > 2 files changed, 11 insertions(+), 2 deletions(-) > > Ugh! Didn't mean to chain these two emails. This one is independent of the other email. -Saravana
On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: > Some PMUs events can be read from any CPU. So allow the PMU to mark > events as such. For these events, we don't need to reject reads or > make smp calls to the event's CPU and cause unnecessary wake ups. > > Good examples of such events would be events from caches shared across > all CPUs. So why would the existing ACTIVE_PKG not work for you? Because clearly your example does not cross a package. > Signed-off-by: Saravana Kannan <skannan@codeaurora.org> > --- > include/linux/perf_event.h | 3 +++ > kernel/events/core.c | 10 ++++++++-- > 2 files changed, 11 insertions(+), 2 deletions(-) > > diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h > index 7546822..ee8978f 100644 > --- a/include/linux/perf_event.h > +++ b/include/linux/perf_event.h > @@ -510,9 +510,12 @@ typedef void (*perf_overflow_handler_t)(struct perf_event *, > * PERF_EV_CAP_SOFTWARE: Is a software event. > * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read > * from any CPU in the package where it is active. > + * PERF_EV_CAP_READ_ANY_CPU: A CPU event (or cgroup event) that can be read > + * from any CPU. > */ > #define PERF_EV_CAP_SOFTWARE BIT(0) > #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1) > +#define PERF_EV_CAP_READ_ANY_CPU BIT(2) > > #define SWEVENT_HLIST_BITS 8 > #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS) > diff --git a/kernel/events/core.c b/kernel/events/core.c > index 5d3df58..570187b 100644 > --- a/kernel/events/core.c > +++ b/kernel/events/core.c > @@ -3484,6 +3484,10 @@ static int __perf_event_read_cpu(struct perf_event *event, int event_cpu) > { > u16 local_pkg, event_pkg; > > + if (event->group_caps & PERF_EV_CAP_READ_ANY_CPU) { > + return smp_processor_id(); > + } > + > if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) { > int local_cpu = smp_processor_id(); > > @@ -3575,6 +3579,7 @@ int perf_event_read_local(struct perf_event *event, u64 *value, > { > unsigned long flags; > int ret = 0; > + bool is_any_cpu = !!(event->group_caps & PERF_EV_CAP_READ_ANY_CPU); > > /* > * Disabling interrupts avoids all counter scheduling (context > @@ -3600,7 +3605,8 @@ int perf_event_read_local(struct perf_event *event, u64 *value, > > /* If this is a per-CPU event, it must be for this CPU */ > if (!(event->attach_state & PERF_ATTACH_TASK) && > - event->cpu != smp_processor_id()) { > + event->cpu != smp_processor_id() && > + !is_any_cpu) { > ret = -EINVAL; > goto out; > } > @@ -3610,7 +3616,7 @@ int perf_event_read_local(struct perf_event *event, u64 *value, > * or local to this CPU. Furthermore it means its ACTIVE (otherwise > * oncpu == -1). > */ > - if (event->oncpu == smp_processor_id()) > + if (event->oncpu == smp_processor_id() || is_any_cpu) > event->pmu->read(event); > > *value = local64_read(&event->count); And why are you modifying read_local for this? That didn't support ACTIVE_PKG, so why should it support this? And again, where are the users?
On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: > Some PMUs events can be read from any CPU. So allow the PMU to mark > events as such. For these events, we don't need to reject reads or > make smp calls to the event's CPU and cause unnecessary wake ups. > > Good examples of such events would be events from caches shared across > all CPUs. I think that if we need to generalize PERF_EV_CAP_READ_ACTIVE_PKG, it would be better to give events a pointer to a cpumask. That could then cover all cases quite trivially: static int __perf_event_read_cpu(struct perf_event *event, int event_cpu) { int local_cpu = smp_processor_id(); if (event->read_mask && cpumask_test_cpu(local_cpu, event->read_mask)) event_cpu = local_cpu; return event_cpu; } ... in the PERF_EV_CAP_READ_ACTIVE_PKG case, we can use the exiting(?) package masks, and more generally we can re-use the PMU's affinit mask if it has one. That said, I see that many pmu::read() implementations have side-effects on hwc->prev_count, and event->count, so I worry that this won't be sfe in general (e.g. if we race with the IRQ handler on another CPU). Thanks, Mark.
On 2018-02-24 00:41, Peter Zijlstra wrote: > On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: >> Some PMUs events can be read from any CPU. So allow the PMU to mark >> events as such. For these events, we don't need to reject reads or >> make smp calls to the event's CPU and cause unnecessary wake ups. >> >> Good examples of such events would be events from caches shared across >> all CPUs. > > So why would the existing ACTIVE_PKG not work for you? Because clearly > your example does not cross a package. Because based on testing it on hardware, it looks like the two clusters in an ARM DynamIQ design are not considered part of the same "package". When I say clusters, I using the more common interpretation of "homogeneous CPUs running on the same clock"/CPUs in a cpufreq policy and not ARM's new redefinition of cluster. So, on a SoC with 4 little and 4 big cores, it'll still trigger a lot of unnecessary smp calls/IPIs that cause unnecessary wakeups. Although, I like Mark's suggestion of just giving a cpumask for every event and using that instead. Because the meaning of "active package" is very ambiguous. For example if a SoC has 2 DynamIQ blocks (not sure if that's possible), what's considered a package? CPUs that are sitting on one L3 can't read the PMU counters of a different L3. In that case, neither "Any CPU" nor "Active Package" is correct/usable for reducing IPIs. > >> Signed-off-by: Saravana Kannan <skannan@codeaurora.org> >> --- >> include/linux/perf_event.h | 3 +++ >> kernel/events/core.c | 10 ++++++++-- >> 2 files changed, 11 insertions(+), 2 deletions(-) >> >> diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h >> index 7546822..ee8978f 100644 >> --- a/include/linux/perf_event.h >> +++ b/include/linux/perf_event.h >> @@ -510,9 +510,12 @@ typedef void (*perf_overflow_handler_t)(struct >> perf_event *, >> * PERF_EV_CAP_SOFTWARE: Is a software event. >> * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that >> can be read >> * from any CPU in the package where it is active. >> + * PERF_EV_CAP_READ_ANY_CPU: A CPU event (or cgroup event) that can >> be read >> + * from any CPU. >> */ >> #define PERF_EV_CAP_SOFTWARE BIT(0) >> #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1) >> +#define PERF_EV_CAP_READ_ANY_CPU BIT(2) >> >> #define SWEVENT_HLIST_BITS 8 >> #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS) >> diff --git a/kernel/events/core.c b/kernel/events/core.c >> index 5d3df58..570187b 100644 >> --- a/kernel/events/core.c >> +++ b/kernel/events/core.c >> @@ -3484,6 +3484,10 @@ static int __perf_event_read_cpu(struct >> perf_event *event, int event_cpu) >> { >> u16 local_pkg, event_pkg; >> >> + if (event->group_caps & PERF_EV_CAP_READ_ANY_CPU) { >> + return smp_processor_id(); >> + } >> + >> if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) { >> int local_cpu = smp_processor_id(); >> > >> @@ -3575,6 +3579,7 @@ int perf_event_read_local(struct perf_event >> *event, u64 *value, >> { >> unsigned long flags; >> int ret = 0; >> + bool is_any_cpu = !!(event->group_caps & PERF_EV_CAP_READ_ANY_CPU); >> >> /* >> * Disabling interrupts avoids all counter scheduling (context >> @@ -3600,7 +3605,8 @@ int perf_event_read_local(struct perf_event >> *event, u64 *value, >> >> /* If this is a per-CPU event, it must be for this CPU */ >> if (!(event->attach_state & PERF_ATTACH_TASK) && >> - event->cpu != smp_processor_id()) { >> + event->cpu != smp_processor_id() && >> + !is_any_cpu) { >> ret = -EINVAL; >> goto out; >> } >> @@ -3610,7 +3616,7 @@ int perf_event_read_local(struct perf_event >> *event, u64 *value, >> * or local to this CPU. Furthermore it means its ACTIVE (otherwise >> * oncpu == -1). >> */ >> - if (event->oncpu == smp_processor_id()) >> + if (event->oncpu == smp_processor_id() || is_any_cpu) >> event->pmu->read(event); >> >> *value = local64_read(&event->count); > > And why are you modifying read_local for this? That didn't support > ACTIVE_PKG, so why should it support this? Maybe I'll make a separate patch to first have perf_event_read_local() also handle ACTIVE_PACKAGE? Because in those cases, the smp call made by read_local is unnecessary too. > > And again, where are the users? The DynamIQ PMU driver would be the user. -Saravana
On 2018-02-25 06:38, Mark Rutland wrote: > On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: >> Some PMUs events can be read from any CPU. So allow the PMU to mark >> events as such. For these events, we don't need to reject reads or >> make smp calls to the event's CPU and cause unnecessary wake ups. >> >> Good examples of such events would be events from caches shared across >> all CPUs. > > I think that if we need to generalize PERF_EV_CAP_READ_ACTIVE_PKG, it > would be > better to give events a pointer to a cpumask. That could then cover all > cases > quite trivially: > > static int __perf_event_read_cpu(struct perf_event *event, int > event_cpu) > { > int local_cpu = smp_processor_id(); > > if (event->read_mask && > cpumask_test_cpu(local_cpu, event->read_mask)) > event_cpu = local_cpu; > > return event_cpu; > } > This is a good improvement on my attempt. If I send a patch for this, is that something you'd be willing to incorporate into your patch set and make sure the DSU pmu driver handles it correctly? > ... in the PERF_EV_CAP_READ_ACTIVE_PKG case, we can use the exiting(?) > package > masks, and more generally we can re-use the PMU's affinit mask if it > has one. > > That said, I see that many pmu::read() implementations have > side-effects on > hwc->prev_count, and event->count, so I worry that this won't be sfe in > general > (e.g. if we race with the IRQ handler on another CPU). > Yeah, this doesn't have to be mandatory. It can be an optional mask the PMU can set up during perf event init. Peter, Is this something that's acceptable to you? Thanks, Saravana
On Mon, Feb 26, 2018 at 06:11:45PM -0800, skannan@codeaurora.org wrote: > On 2018-02-25 06:38, Mark Rutland wrote: > > On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: > > > Some PMUs events can be read from any CPU. So allow the PMU to mark > > > events as such. For these events, we don't need to reject reads or > > > make smp calls to the event's CPU and cause unnecessary wake ups. > > > > > > Good examples of such events would be events from caches shared across > > > all CPUs. > > > > I think that if we need to generalize PERF_EV_CAP_READ_ACTIVE_PKG, it > > would be > > better to give events a pointer to a cpumask. That could then cover all > > cases > > quite trivially: > > > > static int __perf_event_read_cpu(struct perf_event *event, int > > event_cpu) > > { > > int local_cpu = smp_processor_id(); > > > > if (event->read_mask && > > cpumask_test_cpu(local_cpu, event->read_mask)) > > event_cpu = local_cpu; > > > > return event_cpu; > > } > > This is a good improvement on my attempt. If I send a patch for this, is > that something you'd be willing to incorporate into your patch set and make > sure the DSU pmu driver handles it correctly? As I commented, I don't think that willl work without more invasive changes as the DSU PMU's pmu::read() function has side effects on hwc->prev_count and event_count, and could race with an IRQ handler on another CPU. Is the IPI really a problem in practice? Thanks, Mark.
On Mon, Feb 26, 2018 at 05:53:57PM -0800, skannan@codeaurora.org wrote: > On 2018-02-24 00:41, Peter Zijlstra wrote: > > On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: > > > Some PMUs events can be read from any CPU. So allow the PMU to mark > > > events as such. For these events, we don't need to reject reads or > > > make smp calls to the event's CPU and cause unnecessary wake ups. > > > > > > Good examples of such events would be events from caches shared across > > > all CPUs. > > > > So why would the existing ACTIVE_PKG not work for you? Because clearly > > your example does not cross a package. > > Because based on testing it on hardware, it looks like the two clusters in > an ARM DynamIQ design are not considered part of the same "package". I don't think we should consider the topology masks at all for system PMU affinity. Due to the number of ways these can be integrated, and the lack of a standard(ish) topology across arm platforms. IIUC, there's ongoing work to try to clean that up, but that won't give us anything meaningful for PMU affinity. If we need a mask, that should be something the FW description of the PMU provides, and the PMU driver provides to the core code. Thanks, Mark.
On 2018-02-27 03:43, Mark Rutland wrote: > On Mon, Feb 26, 2018 at 06:11:45PM -0800, skannan@codeaurora.org wrote: >> On 2018-02-25 06:38, Mark Rutland wrote: >> > On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: >> > > Some PMUs events can be read from any CPU. So allow the PMU to mark >> > > events as such. For these events, we don't need to reject reads or >> > > make smp calls to the event's CPU and cause unnecessary wake ups. >> > > >> > > Good examples of such events would be events from caches shared across >> > > all CPUs. >> > >> > I think that if we need to generalize PERF_EV_CAP_READ_ACTIVE_PKG, it >> > would be >> > better to give events a pointer to a cpumask. That could then cover all >> > cases >> > quite trivially: >> > >> > static int __perf_event_read_cpu(struct perf_event *event, int >> > event_cpu) >> > { >> > int local_cpu = smp_processor_id(); >> > >> > if (event->read_mask && >> > cpumask_test_cpu(local_cpu, event->read_mask)) >> > event_cpu = local_cpu; >> > >> > return event_cpu; >> > } >> >> This is a good improvement on my attempt. If I send a patch for this, >> is >> that something you'd be willing to incorporate into your patch set and >> make >> sure the DSU pmu driver handles it correctly? > > As I commented, I don't think that willl work without more invasive > changes as the DSU PMU's pmu::read() function has side effects on > hwc->prev_count and event_count, and could race with an IRQ handler on > another CPU. > > Is the IPI really a problem in practice? > There are a bunch of cases, but the simplest one is if you try to collect DSU stats (for analysis) while measuring power, it completely messes up the power measurements. Thanks, Saravana
On Mon, Feb 26, 2018 at 05:53:57PM -0800, skannan@codeaurora.org wrote: > On 2018-02-24 00:41, Peter Zijlstra wrote: > > On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: > > > Some PMUs events can be read from any CPU. So allow the PMU to mark > > > events as such. For these events, we don't need to reject reads or > > > make smp calls to the event's CPU and cause unnecessary wake ups. > > > > > > Good examples of such events would be events from caches shared across > > > all CPUs. > > > > So why would the existing ACTIVE_PKG not work for you? Because clearly > > your example does not cross a package. > > Because based on testing it on hardware, it looks like the two clusters in > an ARM DynamIQ design are not considered part of the same "package". When I > say clusters, I using the more common interpretation of "homogeneous CPUs > running on the same clock"/CPUs in a cpufreq policy and not ARM's new > redefinition of cluster. So, on a SoC with 4 little and 4 big cores, it'll > still trigger a lot of unnecessary smp calls/IPIs that cause unnecessary > wakeups. arch/arm64/include/asm/topology.h:#define topology_physical_package_id(cpu) (cpu_topology[cpu].cluster_id) *sigh*... that's just broken...
Hi, On 03/03/2018 09:41 AM, Peter Zijlstra wrote: > On Mon, Feb 26, 2018 at 05:53:57PM -0800, skannan@codeaurora.org wrote: >> On 2018-02-24 00:41, Peter Zijlstra wrote: >>> On Fri, Feb 23, 2018 at 04:19:38PM -0800, Saravana Kannan wrote: >>>> Some PMUs events can be read from any CPU. So allow the PMU to mark >>>> events as such. For these events, we don't need to reject reads or >>>> make smp calls to the event's CPU and cause unnecessary wake ups. >>>> >>>> Good examples of such events would be events from caches shared across >>>> all CPUs. >>> >>> So why would the existing ACTIVE_PKG not work for you? Because clearly >>> your example does not cross a package. >> >> Because based on testing it on hardware, it looks like the two clusters in >> an ARM DynamIQ design are not considered part of the same "package". When I >> say clusters, I using the more common interpretation of "homogeneous CPUs >> running on the same clock"/CPUs in a cpufreq policy and not ARM's new >> redefinition of cluster. So, on a SoC with 4 little and 4 big cores, it'll >> still trigger a lot of unnecessary smp calls/IPIs that cause unnecessary >> wakeups. > > arch/arm64/include/asm/topology.h:#define topology_physical_package_id(cpu) (cpu_topology[cpu].cluster_id) > > *sigh*... that's just broken... > Its being reworked in the PPTT (currently v7) patch set. For ACPI systems (and hopefully DT machines with the package property set) topology_physical_package and core siblings represent the socket as one would expect. Thanks,
diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h index 7546822..ee8978f 100644 --- a/include/linux/perf_event.h +++ b/include/linux/perf_event.h @@ -510,9 +510,12 @@ typedef void (*perf_overflow_handler_t)(struct perf_event *, * PERF_EV_CAP_SOFTWARE: Is a software event. * PERF_EV_CAP_READ_ACTIVE_PKG: A CPU event (or cgroup event) that can be read * from any CPU in the package where it is active. + * PERF_EV_CAP_READ_ANY_CPU: A CPU event (or cgroup event) that can be read + * from any CPU. */ #define PERF_EV_CAP_SOFTWARE BIT(0) #define PERF_EV_CAP_READ_ACTIVE_PKG BIT(1) +#define PERF_EV_CAP_READ_ANY_CPU BIT(2) #define SWEVENT_HLIST_BITS 8 #define SWEVENT_HLIST_SIZE (1 << SWEVENT_HLIST_BITS) diff --git a/kernel/events/core.c b/kernel/events/core.c index 5d3df58..570187b 100644 --- a/kernel/events/core.c +++ b/kernel/events/core.c @@ -3484,6 +3484,10 @@ static int __perf_event_read_cpu(struct perf_event *event, int event_cpu) { u16 local_pkg, event_pkg; + if (event->group_caps & PERF_EV_CAP_READ_ANY_CPU) { + return smp_processor_id(); + } + if (event->group_caps & PERF_EV_CAP_READ_ACTIVE_PKG) { int local_cpu = smp_processor_id(); @@ -3575,6 +3579,7 @@ int perf_event_read_local(struct perf_event *event, u64 *value, { unsigned long flags; int ret = 0; + bool is_any_cpu = !!(event->group_caps & PERF_EV_CAP_READ_ANY_CPU); /* * Disabling interrupts avoids all counter scheduling (context @@ -3600,7 +3605,8 @@ int perf_event_read_local(struct perf_event *event, u64 *value, /* If this is a per-CPU event, it must be for this CPU */ if (!(event->attach_state & PERF_ATTACH_TASK) && - event->cpu != smp_processor_id()) { + event->cpu != smp_processor_id() && + !is_any_cpu) { ret = -EINVAL; goto out; } @@ -3610,7 +3616,7 @@ int perf_event_read_local(struct perf_event *event, u64 *value, * or local to this CPU. Furthermore it means its ACTIVE (otherwise * oncpu == -1). */ - if (event->oncpu == smp_processor_id()) + if (event->oncpu == smp_processor_id() || is_any_cpu) event->pmu->read(event); *value = local64_read(&event->count);
Some PMUs events can be read from any CPU. So allow the PMU to mark events as such. For these events, we don't need to reject reads or make smp calls to the event's CPU and cause unnecessary wake ups. Good examples of such events would be events from caches shared across all CPUs. Signed-off-by: Saravana Kannan <skannan@codeaurora.org> --- include/linux/perf_event.h | 3 +++ kernel/events/core.c | 10 ++++++++-- 2 files changed, 11 insertions(+), 2 deletions(-)