[PATCH 0/2] perf/x86: Package residency counter improvements

[PATCH 0/2] perf/x86: Package residency counter improvements

[Tero Kristo]

Hello,

Following two patches address a couple of different issues with Intel
C-state package residency counters, which are used to track power saving
state usage on Intel chips. The patches don't have any dependencies to
each other and can be applied separately.

1) The residency counters are always read from the first CPU of the
   package via an SMP call, even if they are available from any CPU.
   This causes extra latency. Patch #1 fixes this issue by flagging
   the perf event properly and allowing to execute it on any CPU of
   the package.
2) The residency counters are completely impossible to read from a BPF
   program running on any other than the first CPU of the package, as the
   SMP call is not available in this context. Patch #2 addresses this
   issue by allowing the read of the perf event from the local CPU,
   similar to what is done in perf_read_event(). This patch also allows
   reading any other package scope perf events (RAPL/UNCORE) from
   arbitrary CPUs on a package via BPF.

-Tero

[PATCH 1/2] perf/x86/cstate: Allow reading the package statistics from local CPU

[Tero Kristo]

The MSR registers for reading the package residency counters are
available on every CPU of the package. To avoid doing unnecessary SMP
calls to read the values for these from the various CPUs inside a
package, allow reading them from any CPU of the package.

Signed-off-by: Tero Kristo <tero.kristo@linux.intel.com>
Suggested-by: Kan Liang <kan.liang@intel.com>
---
 arch/x86/events/intel/cstate.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/arch/x86/events/intel/cstate.c b/arch/x86/events/intel/cstate.c
index 835862c548cc..5cfe021edc65 100644
--- a/arch/x86/events/intel/cstate.c
+++ b/arch/x86/events/intel/cstate.c
@@ -336,6 +336,9 @@ static int cstate_pmu_event_init(struct perf_event *event)
 		cfg = array_index_nospec((unsigned long)cfg, PERF_CSTATE_PKG_EVENT_MAX);
 		if (!(pkg_msr_mask & (1 << cfg)))
 			return -EINVAL;
+
+		event->event_caps |= PERF_EV_CAP_READ_ACTIVE_PKG;
+
 		event->hw.event_base = pkg_msr[cfg].msr;
 		cpu = cpumask_any_and(&cstate_pkg_cpu_mask,
 				      topology_die_cpumask(event->cpu));
-- 
2.25.1

[PATCH 2/2] perf/core: Allow reading package events from perf_event_read_local

[Tero Kristo]

Per-package perf events are typically registered with a single CPU only,
however they can be read across all the CPUs within the package.
Currently perf_event_read maps the event CPU according to the topology
information to avoid an unnecessary SMP call, however
perf_event_read_local deals with hard values and rejects a read with a
failure if the CPU is not the one exactly registered. Allow similar
mapping within the perf_event_read_local if the perf event in question
can support this.

This allows users like BPF code to read the package perf events properly
across different CPUs within a package.

Signed-off-by: Tero Kristo <tero.kristo@linux.intel.com>
---
 kernel/events/core.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 78ae7b6f90fd..37db7c003b79 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -4528,6 +4528,7 @@ int perf_event_read_local(struct perf_event *event, u64 *value,
 {
 	unsigned long flags;
 	int ret = 0;
+	int event_cpu;
 
 	/*
 	 * Disabling interrupts avoids all counter scheduling (context
@@ -4551,15 +4552,19 @@ int perf_event_read_local(struct perf_event *event, u64 *value,
 		goto out;
 	}
 
+	/* Allow reading a per-package perf-event from local CPU also */
+	event_cpu = READ_ONCE(event->oncpu);
+	event_cpu = __perf_event_read_cpu(event, event_cpu);
+
 	/* 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()) {
 		ret = -EINVAL;
 		goto out;
 	}
 
 	/* If this is a pinned event it must be running on this CPU */
-	if (event->attr.pinned && event->oncpu != smp_processor_id()) {
+	if (event->attr.pinned && event_cpu != smp_processor_id()) {
 		ret = -EBUSY;
 		goto out;
 	}
@@ -4569,7 +4574,7 @@ 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_cpu == smp_processor_id())
 		event->pmu->read(event);
 
 	*value = local64_read(&event->count);
-- 
2.25.1