[PATCH v3 0/3] perf: use hrtimer for event multiplexing

[PATCH v3 0/3] perf: use hrtimer for event multiplexing

[Stephane Eranian]

The current scheme of using the timer tick was fine
for per-thread events. However, it was causing
bias issues in system-wide mode (including for
uncore PMUs). Event groups would not get their
fair share of runtime on the PMU. With tickless
kernels, if a core is idle there is no timer tick,
and thus no event rotation (multiplexing). However,
there are events (especially uncore events) which do
count even though cores are asleep.

This patch changes the timer source for multiplexing.
It introduces a per-cpu hrtimer. The advantage is that
even when the core goes idle, it will come back to
service the hrtimer, thus multiplexing on system-wide
events works much better.

In order to minimize the impact of the hrtimer, it
is turned on and off on demand. When the PMU on
a CPU is overcommited, the hrtimer is activated.
It is stopped when the PMU is not overcommitted.

In order for this to work properly with HOTPLUG_CPU,
we had to change the order of initialization in
start_kernel() such that hrtimer_init() is run
before perf_event_init().

The second patch provide a sysctl control to
adjust the multiplexing interval. Unit is
milliseconds.

Here is a simple before/after example with
two event groups which do require multiplexing.
This is done in system-wide mode on an idle
system. What matters here is the scaling factor
in [] in not the total counts.

Before:

# perf stat -a -e ref-cycles,ref-cycles sleep 10
 Performance counter stats for 'sleep 10':
 34,319,545 ref-cycles  [56.51%]
 31,917,229 ref-cycles  [43.50%]

 10.000827569 seconds time elapsed

After:
# perf stat -a -e ref-cycles,ref-cycles sleep 10
 Performance counter stats for 'sleep 10':
 11,144,822,193 ref-cycles [50.00%]
 11,103,760,513 ref-cycles [50.00%]

 10.000672946 seconds time elapsed

In this second version of the patchset, we now
have the hrtimer_interval per PMU instance. The
tunable is in /sys/devices/XXX/mux_interval_ms,
where XXX is the name of the PMU instance. Due
to initialization changes of each hrtimer, we
had to introduce hrtimer_init_cpu() to initialize
a hrtimer from another CPU.

In the 3rd version, we simplify the code a bit
by using hrtimer_active(). We stopped using
the rotation_list for perf_cpu_hrtimer_cancel().
We also fix an intialization problem.

Signed-off-by: Stephane Eranian <eranian@google.com>
---

Stephane Eranian (3):
  hrtimer: add hrtimer_init_cpu()
  perf: use hrtimer for event multiplexing
  perf: add sysfs entry to adjust multiplexing interval per PMU

 include/linux/hrtimer.h    |    2 +
 include/linux/perf_event.h |    5 +-
 init/main.c                |    2 +-
 kernel/events/core.c       |  166 +++++++++++++++++++++++++++++++++++++++++---
 kernel/hrtimer.c           |   17 +++--
 5 files changed, 176 insertions(+), 16 deletions(-)

-- 
1.7.5.4

[PATCH v3 1/2] perf: use hrtimer for event multiplexing

[Stephane Eranian]

The current scheme of using the timer tick was fine
for per-thread events. However, it was causing
bias issues in system-wide mode (including for
uncore PMUs). Event groups would not get their
fair share of runtime on the PMU. With tickless
kernels, if a core is idle there is no timer tick,
and thus no event rotation (multiplexing). However,
there are events (especially uncore events) which do
count even though cores are asleep.

This patch changes the timer source for multiplexing.
It introduces a per-PMU per-cpu hrtimer. The advantage
is that even when a core goes idle, it will come back
to service the hrtimer, thus multiplexing on system-wide
events works much better.

The per-PMU implementation (suggested by PeterZ) enables
adjusting the multiplexing interval per PMU. The preferred
interval is stashed into the struct pmu. If not set, it
will be forced to the default interval value.

In order to minimize the impact of the hrtimer, it
is turned on and off on demand. When the PMU on
a CPU is overcommited, the hrtimer is activated.
It is stopped when the PMU is not overcommitted.

In order for this to work properly, we had to change
the order of initialization in start_kernel() such
that hrtimer_init() is run before perf_event_init().

The default interval in milliseconds is set to a
timer tick just like with the old code. We will
provide a sysctl to tune this in another patch.

Signed-off-by: Stephane Eranian <eranian@google.com>
---
 include/linux/perf_event.h |    3 +-
 init/main.c                |    2 +-
 kernel/events/core.c       |  117 +++++++++++++++++++++++++++++++++++++++++---
 3 files changed, 113 insertions(+), 9 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index cc5e2cd..204e1dc 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -1106,8 +1106,9 @@ struct perf_cpu_context {
 	struct perf_event_context	*task_ctx;
 	int				active_oncpu;
 	int				exclusive;
+	struct hrtimer			hrtimer;
+	ktime_t				hrtimer_interval;
 	struct list_head		rotation_list;
-	int				jiffies_interval;
 	struct pmu			*active_pmu;
 	struct perf_cgroup		*cgrp;
 };
diff --git a/init/main.c b/init/main.c
index b286730..8ffe441 100644
--- a/init/main.c
+++ b/init/main.c
@@ -541,7 +541,6 @@ asmlinkage void __init start_kernel(void)
 		local_irq_disable();
 	}
 	idr_init_cache();
-	perf_event_init();
 	rcu_init();
 	radix_tree_init();
 	/* init some links before init_ISA_irqs() */
@@ -553,6 +552,7 @@ asmlinkage void __init start_kernel(void)
 	softirq_init();
 	timekeeping_init();
 	time_init();
+	perf_event_init();
 	profile_init();
 	call_function_init();
 	if (!irqs_disabled())
diff --git a/kernel/events/core.c b/kernel/events/core.c
index 7b9df35..cb9d811 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -168,6 +168,8 @@ int sysctl_perf_event_sample_rate __read_mostly = DEFAULT_MAX_SAMPLE_RATE;
 static int max_samples_per_tick __read_mostly =
 	DIV_ROUND_UP(DEFAULT_MAX_SAMPLE_RATE, HZ);
 
+static int perf_rotate_context(struct perf_cpu_context *cpuctx);
+
 int perf_proc_update_handler(struct ctl_table *table, int write,
 		void __user *buffer, size_t *lenp,
 		loff_t *ppos)
@@ -625,6 +627,98 @@ perf_cgroup_mark_enabled(struct perf_event *event,
 }
 #endif
 
+/*
+ * set default to be dependent on timer tick just
+ * like original code
+ */
+#define PERF_CPU_HRTIMER (1000 / HZ)
+static enum hrtimer_restart perf_cpu_hrtimer_handler(struct hrtimer *hr)
+{
+	struct perf_cpu_context *cpuctx;
+	enum hrtimer_restart ret = HRTIMER_NORESTART;
+	unsigned long flags;
+	int rotations = 0;
+
+	cpuctx = container_of(hr, struct perf_cpu_context, hrtimer);
+
+	local_irq_save(flags);
+
+	rotations = perf_rotate_context(cpuctx);
+
+	local_irq_restore(flags);
+
+	/*
+	 * arm timer if needed
+	 */
+	if (rotations) {
+		hrtimer_forward_now(hr, cpuctx->hrtimer_interval);
+		ret = HRTIMER_RESTART;
+	}
+
+	return ret;
+}
+
+/* CPU is going down */
+void perf_cpu_hrtimer_cancel(int cpu)
+{
+	struct perf_cpu_context *cpuctx;
+	struct pmu *pmu;
+	unsigned long flags;
+
+	if (WARN_ON(cpu != smp_processor_id()))
+		return;
+
+	local_irq_save(flags);
+
+	rcu_read_lock();
+
+	list_for_each_entry_rcu(pmu, &pmus, entry) {
+		cpuctx = this_cpu_ptr(pmu->pmu_cpu_context);
+
+		if (pmu->task_ctx_nr == perf_sw_context)
+			continue;
+
+		hrtimer_cancel(&cpuctx->hrtimer);
+	}
+
+	rcu_read_unlock();
+
+	local_irq_restore(flags);
+}
+
+static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
+{
+	struct hrtimer *hr = &cpuctx->hrtimer;
+	struct pmu *pmu = cpuctx->ctx.pmu;
+
+	/* no multiplexing needed for SW PMU */
+	if (pmu->task_ctx_nr == perf_sw_context)
+		return;
+
+	cpuctx->hrtimer_interval =
+		ns_to_ktime(NSEC_PER_MSEC * PERF_CPU_HRTIMER);
+
+	hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
+	hr->function = perf_cpu_hrtimer_handler;
+}
+
+static void perf_cpu_hrtimer_restart(struct perf_cpu_context *cpuctx)
+{
+	struct hrtimer *hr = &cpuctx->hrtimer;
+	struct pmu *pmu = cpuctx->ctx.pmu;
+
+	/* not for SW PMU */
+	if (pmu->task_ctx_nr == perf_sw_context)
+		return;
+
+	if (hrtimer_active(hr))
+		return;
+
+	if (!hrtimer_callback_running(hr))
+		__hrtimer_start_range_ns(hr, cpuctx->hrtimer_interval,
+					 0, HRTIMER_MODE_REL_PINNED, 0);
+}
+
 void perf_pmu_disable(struct pmu *pmu)
 {
 	int *count = this_cpu_ptr(pmu->pmu_disable_count);
@@ -1454,6 +1548,7 @@ group_sched_in(struct perf_event *group_event,
 
 	if (event_sched_in(group_event, cpuctx, ctx)) {
 		pmu->cancel_txn(pmu);
+		perf_cpu_hrtimer_restart(cpuctx);
 		return -EAGAIN;
 	}
 
@@ -1500,6 +1595,8 @@ group_sched_in(struct perf_event *group_event,
 
 	pmu->cancel_txn(pmu);
 
+	perf_cpu_hrtimer_restart(cpuctx);
+
 	return -EAGAIN;
 }
 
@@ -1755,8 +1852,10 @@ static int __perf_event_enable(void *info)
 		 * If this event can't go on and it's part of a
 		 * group, then the whole group has to come off.
 		 */
-		if (leader != event)
+		if (leader != event) {
 			group_sched_out(leader, cpuctx, ctx);
+			perf_cpu_hrtimer_restart(cpuctx);
+		}
 		if (leader->attr.pinned) {
 			update_group_times(leader);
 			leader->state = PERF_EVENT_STATE_ERROR;
@@ -2503,7 +2602,7 @@ static void rotate_ctx(struct perf_event_context *ctx)
  * because they're strictly cpu affine and rotate_start is called with IRQs
  * disabled, while rotate_context is called from IRQ context.
  */
-static void perf_rotate_context(struct perf_cpu_context *cpuctx)
+static int perf_rotate_context(struct perf_cpu_context *cpuctx)
 {
 	struct perf_event_context *ctx = NULL;
 	int rotate = 0, remove = 1;
@@ -2542,6 +2641,8 @@ static void perf_rotate_context(struct perf_cpu_context *cpuctx)
 done:
 	if (remove)
 		list_del_init(&cpuctx->rotation_list);
+
+	return rotate;
 }
 
 void perf_event_task_tick(void)
@@ -2563,10 +2664,6 @@ void perf_event_task_tick(void)
 		ctx = cpuctx->task_ctx;
 		if (ctx)
 			perf_adjust_freq_unthr_context(ctx, throttled);
-
-		if (cpuctx->jiffies_interval == 1 ||
-				!(jiffies % cpuctx->jiffies_interval))
-			perf_rotate_context(cpuctx);
 	}
 }
 
@@ -5996,7 +6093,9 @@ int perf_pmu_register(struct pmu *pmu, char *name, int type)
 		lockdep_set_class(&cpuctx->ctx.lock, &cpuctx_lock);
 		cpuctx->ctx.type = cpu_context;
 		cpuctx->ctx.pmu = pmu;
-		cpuctx->jiffies_interval = 1;
+
+		__perf_cpu_hrtimer_init(cpuctx, cpu);
+
 		INIT_LIST_HEAD(&cpuctx->rotation_list);
 		cpuctx->active_pmu = pmu;
 	}
@@ -7375,6 +7474,10 @@ perf_cpu_notify(struct notifier_block *self, unsigned long action, void *hcpu)
 	case CPU_DOWN_PREPARE:
 		perf_event_exit_cpu(cpu);
 		break;
+	case CPU_DYING:
+		/* must be run on actual cpu */
+		perf_cpu_hrtimer_cancel(cpu);
+		break;
 
 	default:
 		break;
-- 
1.7.5.4

[PATCH v3 2/2] perf: add sysfs entry to adjust multiplexing interval per PMU

[Stephane Eranian]

This patch adds /sys/device/xxx/mux_interval_ms to ajust the
multiplexing interval per PMU. The unit is milliseconds. Value
has to be >= 1.

Signed-off-by: Stephane Eranian <eranian@google.com>
---
 include/linux/perf_event.h |    1 +
 kernel/events/core.c       |   54 ++++++++++++++++++++++++++++++++++++++++---
 2 files changed, 51 insertions(+), 4 deletions(-)

diff --git a/include/linux/perf_event.h b/include/linux/perf_event.h
index 204e1dc..ef1a6fe 100644
--- a/include/linux/perf_event.h
+++ b/include/linux/perf_event.h
@@ -784,6 +784,7 @@ struct pmu {
 	int * __percpu			pmu_disable_count;
 	struct perf_cpu_context * __percpu pmu_cpu_context;
 	int				task_ctx_nr;
+	int				hrtimer_interval_ms;
 
 	/*
 	 * Fully disable/enable this PMU, can be used to protect from the PMI
diff --git a/kernel/events/core.c b/kernel/events/core.c
index cb9d811..de0ef2b 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -690,13 +690,21 @@ static void __perf_cpu_hrtimer_init(struct perf_cpu_context *cpuctx, int cpu)
 {
 	struct hrtimer *hr = &cpuctx->hrtimer;
 	struct pmu *pmu = cpuctx->ctx.pmu;
+	int timer;
 
 	/* no multiplexing needed for SW PMU */
 	if (pmu->task_ctx_nr == perf_sw_context)
 		return;
 
-	cpuctx->hrtimer_interval =
-		ns_to_ktime(NSEC_PER_MSEC * PERF_CPU_HRTIMER);
+	/*
+	 * check default is sane, if not set then force to
+	 * default interval (1/tick)
+	 */
+	timer = pmu->hrtimer_interval_ms;
+	if (timer < 1)
+		timer = pmu->hrtimer_interval_ms = PERF_CPU_HRTIMER;
+
+	cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
 
 	hrtimer_init(hr, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
 	hr->function = perf_cpu_hrtimer_handler;
@@ -5993,9 +6001,47 @@ type_show(struct device *dev, struct device_attribute *attr, char *page)
 	return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->type);
 }
 
+static ssize_t
+mux_interval_ms_show(struct device *dev, struct device_attribute *attr,
+		     char *page)
+{
+	struct pmu *pmu = dev_get_drvdata(dev);
+
+	return snprintf(page, PAGE_SIZE-1, "%d\n", pmu->hrtimer_interval_ms);
+}
+
+static ssize_t
+mux_interval_ms_store(struct device *dev, struct device_attribute *attr,
+		      const char *buf, size_t count)
+{
+	struct pmu *pmu = dev_get_drvdata(dev);
+	int timer, cpu, ret;
+
+	ret = kstrtoint(buf, 0, &timer);
+	if (ret)
+		return ret;
+
+	if (timer < 1)
+		return -EINVAL;
+
+	pmu->hrtimer_interval_ms = timer;
+
+	/* update all cpuctx for this PMU */
+	for_each_possible_cpu(cpu) {
+		struct perf_cpu_context *cpuctx;
+		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
+		cpuctx->hrtimer_interval = ns_to_ktime(NSEC_PER_MSEC * timer);
+	}
+
+	return count;
+}
+
+#define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
+
 static struct device_attribute pmu_dev_attrs[] = {
-       __ATTR_RO(type),
-       __ATTR_NULL,
+	__ATTR_RO(type),
+	__ATTR_RW(mux_interval_ms),
+	__ATTR_NULL,
 };
 
 static int pmu_bus_running;
-- 
1.7.5.4

Re: [PATCH v3 0/3] perf: use hrtimer for event multiplexing

[Stephane Eranian]

Any comment on this patch set?
This is an important improvement for any system-wide measurement.

On Thu, Sep 13, 2012 at 4:10 PM, Stephane Eranian <eranian@google.com> wrote:
> The current scheme of using the timer tick was fine
> for per-thread events. However, it was causing
> bias issues in system-wide mode (including for
> uncore PMUs). Event groups would not get their
> fair share of runtime on the PMU. With tickless
> kernels, if a core is idle there is no timer tick,
> and thus no event rotation (multiplexing). However,
> there are events (especially uncore events) which do
> count even though cores are asleep.
>
> This patch changes the timer source for multiplexing.
> It introduces a per-cpu hrtimer. The advantage is that
> even when the core goes idle, it will come back to
> service the hrtimer, thus multiplexing on system-wide
> events works much better.
>
> In order to minimize the impact of the hrtimer, it
> is turned on and off on demand. When the PMU on
> a CPU is overcommited, the hrtimer is activated.
> It is stopped when the PMU is not overcommitted.
>
> In order for this to work properly with HOTPLUG_CPU,
> we had to change the order of initialization in
> start_kernel() such that hrtimer_init() is run
> before perf_event_init().
>
> The second patch provide a sysctl control to
> adjust the multiplexing interval. Unit is
> milliseconds.
>
> Here is a simple before/after example with
> two event groups which do require multiplexing.
> This is done in system-wide mode on an idle
> system. What matters here is the scaling factor
> in [] in not the total counts.
>
> Before:
>
> # perf stat -a -e ref-cycles,ref-cycles sleep 10
>  Performance counter stats for 'sleep 10':
>  34,319,545 ref-cycles  [56.51%]
>  31,917,229 ref-cycles  [43.50%]
>
>  10.000827569 seconds time elapsed
>
> After:
> # perf stat -a -e ref-cycles,ref-cycles sleep 10
>  Performance counter stats for 'sleep 10':
>  11,144,822,193 ref-cycles [50.00%]
>  11,103,760,513 ref-cycles [50.00%]
>
>  10.000672946 seconds time elapsed
>
> In this second version of the patchset, we now
> have the hrtimer_interval per PMU instance. The
> tunable is in /sys/devices/XXX/mux_interval_ms,
> where XXX is the name of the PMU instance. Due
> to initialization changes of each hrtimer, we
> had to introduce hrtimer_init_cpu() to initialize
> a hrtimer from another CPU.
>
> In the 3rd version, we simplify the code a bit
> by using hrtimer_active(). We stopped using
> the rotation_list for perf_cpu_hrtimer_cancel().
> We also fix an intialization problem.
>
> Signed-off-by: Stephane Eranian <eranian@google.com>
> ---
>
> Stephane Eranian (3):
>   hrtimer: add hrtimer_init_cpu()
>   perf: use hrtimer for event multiplexing
>   perf: add sysfs entry to adjust multiplexing interval per PMU
>
>  include/linux/hrtimer.h    |    2 +
>  include/linux/perf_event.h |    5 +-
>  init/main.c                |    2 +-
>  kernel/events/core.c       |  166 +++++++++++++++++++++++++++++++++++++++++---
>  kernel/hrtimer.c           |   17 +++--
>  5 files changed, 176 insertions(+), 16 deletions(-)
>
> --
> 1.7.5.4
>