[PATCH 0/2] perf: Scheduling optimizations

[PATCH 0/2] perf: Scheduling optimizations

[Alexander Shishkin]

Hi Peter,

These here two patches deal with rescheduling on active contexts. The common
idea is that we shouldn't reschedule groups of higher priority than the ones
that are being added or scheduled in. 1/1 takes on the scheduling path, where
we don't want to touch cpu flexible groups if the scheduling context doesn't
carry pinned groups. 1/2 deals with the enable/install cases, where we have
to consider all four types of groups.

Alexander Shishkin (2):
  perf: Don't re-schedule cpu flexible events needlessly
  perf: Optimize event rescheduling on active contexts

 kernel/events/core.c | 86 ++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 74 insertions(+), 12 deletions(-)

-- 
2.11.0

[PATCH 1/2] perf: Don't re-schedule cpu flexible events needlessly

[Alexander Shishkin]

In the sched-in path, we first remove cpu's flexible events in order to
give priority to task's pinned events. However, this step can be safely
skipped if the task doesn't have its own pinned events.

This patch implements this skipping.

Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reported-by: Adrian Hunter <adrian.hunter@intel.com>
---
 kernel/events/core.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index dedfd3078c..c196b9537f 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3137,8 +3137,12 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
 	 * We want to keep the following priority order:
 	 * cpu pinned (that don't need to move), task pinned,
 	 * cpu flexible, task flexible.
+	 *
+	 * However, if task's ctx is not carrying any pinned
+	 * events, no need to flip the cpuctx's events around.
 	 */
-	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
+	if (!list_empty(&ctx->pinned_groups))
+		cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
 	perf_event_sched_in(cpuctx, ctx, task);
 	perf_pmu_enable(ctx->pmu);
 	perf_ctx_unlock(cpuctx, ctx);
-- 
2.11.0

[PATCH 2/2] perf: Optimize event rescheduling on active contexts

[Alexander Shishkin]

When new events are added to an active context, we go and reschedule
all cpu groups and all task groups in order to preserve the priority
(cpu pinned, task pinned, cpu flexible, task flexible), but in
reality we only need to reschedule groups of the same priority as
that of the events being added, and below.

This patch changes the behavior so that only groups that need to be
rescheduled are rescheduled.

Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Reported-by: Adrian Hunter <adrian.hunter@intel.com>
---
 kernel/events/core.c | 80 ++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 69 insertions(+), 11 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index c196b9537f..0f0de3a64c 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -355,6 +355,8 @@ enum event_type_t {
 	EVENT_FLEXIBLE = 0x1,
 	EVENT_PINNED = 0x2,
 	EVENT_TIME = 0x4,
+	/* see ctx_resched() for details */
+	EVENT_CPU = 0x8,
 	EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
 };
 
@@ -1453,6 +1455,20 @@ static void update_group_times(struct perf_event *leader)
 		update_event_times(event);
 }
 
+static enum event_type_t get_event_type(struct perf_event *event)
+{
+	struct perf_event_context *ctx = event->ctx;
+	enum event_type_t event_type;
+
+	lockdep_assert_held(&ctx->lock);
+
+	event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE;
+	if (!ctx->task)
+		event_type |= EVENT_CPU;
+
+	return event_type;
+}
+
 static struct list_head *
 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
 {
@@ -2256,7 +2272,8 @@ ctx_sched_in(struct perf_event_context *ctx,
 	     struct task_struct *task);
 
 static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
-			       struct perf_event_context *ctx)
+			       struct perf_event_context *ctx,
+			       enum event_type_t event_type)
 {
 	if (!cpuctx->task_ctx)
 		return;
@@ -2264,7 +2281,7 @@ static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
 	if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
 		return;
 
-	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
+	ctx_sched_out(ctx, cpuctx, event_type);
 }
 
 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
@@ -2279,13 +2296,51 @@ static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
 		ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
 }
 
+/*
+ * We want to maintain the following priority of scheduling:
+ *  - cpu pinned (EVENT_CPU | EVENT_PINNED)
+ *  - task pinned (EVENT_PINNED)
+ *  - cpu flexible (EVENT_CPU | EVENT_FLEXIBLE)
+ *  - task flexible (EVENT_FLEXIBLE).
+ *
+ * In order to avoid unscheduling and scheduling back in everything every
+ * time an event is added, only do it for the groups of equal priority and
+ * below.
+ *
+ * This can be called after a batch operation on task events, in which case
+ * event_type is a bit mask of the types of events involved. For cpu events,
+ * event_type is only either EVENT_PINNED or EVENT_FLEXIBLE.
+ */
 static void ctx_resched(struct perf_cpu_context *cpuctx,
-			struct perf_event_context *task_ctx)
+			struct perf_event_context *task_ctx,
+			enum event_type_t event_type)
 {
+	enum event_type_t ctx_event_type = event_type & EVENT_ALL;
+	bool cpu_event = !!(event_type & EVENT_CPU);
+
+	/*
+	 * If pinned groups are involved, flexible groups also need to be
+	 * scheduled out.
+	 */
+	if (event_type & EVENT_PINNED)
+		event_type |= EVENT_FLEXIBLE;
+
 	perf_pmu_disable(cpuctx->ctx.pmu);
 	if (task_ctx)
-		task_ctx_sched_out(cpuctx, task_ctx);
-	cpu_ctx_sched_out(cpuctx, EVENT_ALL);
+		task_ctx_sched_out(cpuctx, task_ctx, event_type);
+
+	/*
+	 * Decide which cpu ctx groups to schedule out based on the types
+	 * of events that caused rescheduling:
+	 *  - EVENT_CPU: schedule out corresponding groups;
+	 *  - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
+	 *  - otherwise, do nothing more.
+	 */
+	if (cpu_event)
+		cpu_ctx_sched_out(cpuctx, ctx_event_type);
+	else if (ctx_event_type & EVENT_PINNED)
+		cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
+
 	perf_event_sched_in(cpuctx, task_ctx, current);
 	perf_pmu_enable(cpuctx->ctx.pmu);
 }
@@ -2333,7 +2388,7 @@ static int  __perf_install_in_context(void *info)
 	if (activate) {
 		ctx_sched_out(ctx, cpuctx, EVENT_TIME);
 		add_event_to_ctx(event, ctx);
-		ctx_resched(cpuctx, task_ctx);
+		ctx_resched(cpuctx, task_ctx, get_event_type(event));
 	} else {
 		add_event_to_ctx(event, ctx);
 	}
@@ -2473,7 +2528,7 @@ static void __perf_event_enable(struct perf_event *event,
 	if (ctx->task)
 		WARN_ON_ONCE(task_ctx != ctx);
 
-	ctx_resched(cpuctx, task_ctx);
+	ctx_resched(cpuctx, task_ctx, get_event_type(event));
 }
 
 /*
@@ -2900,7 +2955,7 @@ static void perf_event_context_sched_out(struct task_struct *task, int ctxn,
 
 	if (do_switch) {
 		raw_spin_lock(&ctx->lock);
-		task_ctx_sched_out(cpuctx, ctx);
+		task_ctx_sched_out(cpuctx, ctx, EVENT_ALL);
 		raw_spin_unlock(&ctx->lock);
 	}
 }
@@ -3457,6 +3512,7 @@ static int event_enable_on_exec(struct perf_event *event,
 static void perf_event_enable_on_exec(int ctxn)
 {
 	struct perf_event_context *ctx, *clone_ctx = NULL;
+	enum event_type_t event_type = 0;
 	struct perf_cpu_context *cpuctx;
 	struct perf_event *event;
 	unsigned long flags;
@@ -3470,15 +3526,17 @@ static void perf_event_enable_on_exec(int ctxn)
 	cpuctx = __get_cpu_context(ctx);
 	perf_ctx_lock(cpuctx, ctx);
 	ctx_sched_out(ctx, cpuctx, EVENT_TIME);
-	list_for_each_entry(event, &ctx->event_list, event_entry)
+	list_for_each_entry(event, &ctx->event_list, event_entry) {
 		enabled |= event_enable_on_exec(event, ctx);
+		event_type |= get_event_type(event);
+	}
 
 	/*
 	 * Unclone and reschedule this context if we enabled any event.
 	 */
 	if (enabled) {
 		clone_ctx = unclone_ctx(ctx);
-		ctx_resched(cpuctx, ctx);
+		ctx_resched(cpuctx, ctx, event_type);
 	}
 	perf_ctx_unlock(cpuctx, ctx);
 
@@ -10208,7 +10266,7 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 	 * in.
 	 */
 	raw_spin_lock_irq(&child_ctx->lock);
-	task_ctx_sched_out(__get_cpu_context(child_ctx), child_ctx);
+	task_ctx_sched_out(__get_cpu_context(child_ctx), child_ctx, EVENT_ALL);
 
 	/*
 	 * Now that the context is inactive, destroy the task <-> ctx relation
-- 
2.11.0

[tip:perf/core] perf/core: Don't re-schedule CPU flexible events needlessly

[tip-bot for Alexander Shishkin]

Commit-ID:  fe45bafbd0e1b5e828aa9d44d07e569df85869a2
Gitweb:     http://git.kernel.org/tip/fe45bafbd0e1b5e828aa9d44d07e569df85869a2
Author:     Alexander Shishkin <alexander.shishkin@linux.intel.com>
AuthorDate: Thu, 19 Jan 2017 18:43:29 +0200
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Mon, 30 Jan 2017 12:01:14 +0100

perf/core: Don't re-schedule CPU flexible events needlessly

In the sched-in path, we first remove a CPU's flexible events in order to
give priority to the task's pinned events. However, this step can be safely
skipped if the task doesn't have its own pinned events.

This patch implements this skipping.

Reported-by: Adrian Hunter <adrian.hunter@intel.com>
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: vince@deater.net
Link: http://lkml.kernel.org/r/20170119164330.22887-2-alexander.shishkin@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 kernel/events/core.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index d92c7ad..8c0b733 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -3122,8 +3122,12 @@ static void perf_event_context_sched_in(struct perf_event_context *ctx,
 	 * We want to keep the following priority order:
 	 * cpu pinned (that don't need to move), task pinned,
 	 * cpu flexible, task flexible.
+	 *
+	 * However, if task's ctx is not carrying any pinned
+	 * events, no need to flip the cpuctx's events around.
 	 */
-	cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
+	if (!list_empty(&ctx->pinned_groups))
+		cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
 	perf_event_sched_in(cpuctx, ctx, task);
 	perf_pmu_enable(ctx->pmu);
 	perf_ctx_unlock(cpuctx, ctx);

[tip:perf/core] perf/core: Optimize event rescheduling on active contexts

[tip-bot for Alexander Shishkin]

Commit-ID:  487f05e18aa4efacee6357480f293a5afe6593b5
Gitweb:     http://git.kernel.org/tip/487f05e18aa4efacee6357480f293a5afe6593b5
Author:     Alexander Shishkin <alexander.shishkin@linux.intel.com>
AuthorDate: Thu, 19 Jan 2017 18:43:30 +0200
Committer:  Ingo Molnar <mingo@kernel.org>
CommitDate: Mon, 30 Jan 2017 12:01:15 +0100

perf/core: Optimize event rescheduling on active contexts

When new events are added to an active context, we go and reschedule
all cpu groups and all task groups in order to preserve the priority
(cpu pinned, task pinned, cpu flexible, task flexible), but in
reality we only need to reschedule groups of the same priority as
that of the events being added, and below.

This patch changes the behavior so that only groups that need to be
rescheduled are rescheduled.

Reported-by: Adrian Hunter <adrian.hunter@intel.com>
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: vince@deater.net
Link: http://lkml.kernel.org/r/20170119164330.22887-3-alexander.shishkin@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
---
 kernel/events/core.c | 80 ++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 69 insertions(+), 11 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index 8c0b733..cbcee23 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -355,6 +355,8 @@ enum event_type_t {
 	EVENT_FLEXIBLE = 0x1,
 	EVENT_PINNED = 0x2,
 	EVENT_TIME = 0x4,
+	/* see ctx_resched() for details */
+	EVENT_CPU = 0x8,
 	EVENT_ALL = EVENT_FLEXIBLE | EVENT_PINNED,
 };
 
@@ -1442,6 +1444,20 @@ static void update_group_times(struct perf_event *leader)
 		update_event_times(event);
 }
 
+static enum event_type_t get_event_type(struct perf_event *event)
+{
+	struct perf_event_context *ctx = event->ctx;
+	enum event_type_t event_type;
+
+	lockdep_assert_held(&ctx->lock);
+
+	event_type = event->attr.pinned ? EVENT_PINNED : EVENT_FLEXIBLE;
+	if (!ctx->task)
+		event_type |= EVENT_CPU;
+
+	return event_type;
+}
+
 static struct list_head *
 ctx_group_list(struct perf_event *event, struct perf_event_context *ctx)
 {
@@ -2215,7 +2231,8 @@ ctx_sched_in(struct perf_event_context *ctx,
 	     struct task_struct *task);
 
 static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
-			       struct perf_event_context *ctx)
+			       struct perf_event_context *ctx,
+			       enum event_type_t event_type)
 {
 	if (!cpuctx->task_ctx)
 		return;
@@ -2223,7 +2240,7 @@ static void task_ctx_sched_out(struct perf_cpu_context *cpuctx,
 	if (WARN_ON_ONCE(ctx != cpuctx->task_ctx))
 		return;
 
-	ctx_sched_out(ctx, cpuctx, EVENT_ALL);
+	ctx_sched_out(ctx, cpuctx, event_type);
 }
 
 static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
@@ -2238,13 +2255,51 @@ static void perf_event_sched_in(struct perf_cpu_context *cpuctx,
 		ctx_sched_in(ctx, cpuctx, EVENT_FLEXIBLE, task);
 }
 
+/*
+ * We want to maintain the following priority of scheduling:
+ *  - CPU pinned (EVENT_CPU | EVENT_PINNED)
+ *  - task pinned (EVENT_PINNED)
+ *  - CPU flexible (EVENT_CPU | EVENT_FLEXIBLE)
+ *  - task flexible (EVENT_FLEXIBLE).
+ *
+ * In order to avoid unscheduling and scheduling back in everything every
+ * time an event is added, only do it for the groups of equal priority and
+ * below.
+ *
+ * This can be called after a batch operation on task events, in which case
+ * event_type is a bit mask of the types of events involved. For CPU events,
+ * event_type is only either EVENT_PINNED or EVENT_FLEXIBLE.
+ */
 static void ctx_resched(struct perf_cpu_context *cpuctx,
-			struct perf_event_context *task_ctx)
+			struct perf_event_context *task_ctx,
+			enum event_type_t event_type)
 {
+	enum event_type_t ctx_event_type = event_type & EVENT_ALL;
+	bool cpu_event = !!(event_type & EVENT_CPU);
+
+	/*
+	 * If pinned groups are involved, flexible groups also need to be
+	 * scheduled out.
+	 */
+	if (event_type & EVENT_PINNED)
+		event_type |= EVENT_FLEXIBLE;
+
 	perf_pmu_disable(cpuctx->ctx.pmu);
 	if (task_ctx)
-		task_ctx_sched_out(cpuctx, task_ctx);
-	cpu_ctx_sched_out(cpuctx, EVENT_ALL);
+		task_ctx_sched_out(cpuctx, task_ctx, event_type);
+
+	/*
+	 * Decide which cpu ctx groups to schedule out based on the types
+	 * of events that caused rescheduling:
+	 *  - EVENT_CPU: schedule out corresponding groups;
+	 *  - EVENT_PINNED task events: schedule out EVENT_FLEXIBLE groups;
+	 *  - otherwise, do nothing more.
+	 */
+	if (cpu_event)
+		cpu_ctx_sched_out(cpuctx, ctx_event_type);
+	else if (ctx_event_type & EVENT_PINNED)
+		cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
+
 	perf_event_sched_in(cpuctx, task_ctx, current);
 	perf_pmu_enable(cpuctx->ctx.pmu);
 }
@@ -2291,7 +2346,7 @@ static int  __perf_install_in_context(void *info)
 	if (reprogram) {
 		ctx_sched_out(ctx, cpuctx, EVENT_TIME);
 		add_event_to_ctx(event, ctx);
-		ctx_resched(cpuctx, task_ctx);
+		ctx_resched(cpuctx, task_ctx, get_event_type(event));
 	} else {
 		add_event_to_ctx(event, ctx);
 	}
@@ -2458,7 +2513,7 @@ static void __perf_event_enable(struct perf_event *event,
 	if (ctx->task)
 		WARN_ON_ONCE(task_ctx != ctx);
 
-	ctx_resched(cpuctx, task_ctx);
+	ctx_resched(cpuctx, task_ctx, get_event_type(event));
 }
 
 /*
@@ -2885,7 +2940,7 @@ unlock:
 
 	if (do_switch) {
 		raw_spin_lock(&ctx->lock);
-		task_ctx_sched_out(cpuctx, ctx);
+		task_ctx_sched_out(cpuctx, ctx, EVENT_ALL);
 		raw_spin_unlock(&ctx->lock);
 	}
 }
@@ -3442,6 +3497,7 @@ static int event_enable_on_exec(struct perf_event *event,
 static void perf_event_enable_on_exec(int ctxn)
 {
 	struct perf_event_context *ctx, *clone_ctx = NULL;
+	enum event_type_t event_type = 0;
 	struct perf_cpu_context *cpuctx;
 	struct perf_event *event;
 	unsigned long flags;
@@ -3455,15 +3511,17 @@ static void perf_event_enable_on_exec(int ctxn)
 	cpuctx = __get_cpu_context(ctx);
 	perf_ctx_lock(cpuctx, ctx);
 	ctx_sched_out(ctx, cpuctx, EVENT_TIME);
-	list_for_each_entry(event, &ctx->event_list, event_entry)
+	list_for_each_entry(event, &ctx->event_list, event_entry) {
 		enabled |= event_enable_on_exec(event, ctx);
+		event_type |= get_event_type(event);
+	}
 
 	/*
 	 * Unclone and reschedule this context if we enabled any event.
 	 */
 	if (enabled) {
 		clone_ctx = unclone_ctx(ctx);
-		ctx_resched(cpuctx, ctx);
+		ctx_resched(cpuctx, ctx, event_type);
 	}
 	perf_ctx_unlock(cpuctx, ctx);
 
@@ -10224,7 +10282,7 @@ static void perf_event_exit_task_context(struct task_struct *child, int ctxn)
 	 * in.
 	 */
 	raw_spin_lock_irq(&child_ctx->lock);
-	task_ctx_sched_out(__get_cpu_context(child_ctx), child_ctx);
+	task_ctx_sched_out(__get_cpu_context(child_ctx), child_ctx, EVENT_ALL);
 
 	/*
 	 * Now that the context is inactive, destroy the task <-> ctx relation