OSDN Git Service

perf/core: Optimize perf_rotate_context() event scheduling
authorPeter Zijlstra <peterz@infradead.org>
Fri, 9 Mar 2018 13:56:27 +0000 (14:56 +0100)
committerIngo Molnar <mingo@kernel.org>
Mon, 12 Mar 2018 14:28:50 +0000 (15:28 +0100)
The event schedule order (as per perf_event_sched_in()) is:

 - cpu  pinned
 - task pinned
 - cpu  flexible
 - task flexible

But perf_rotate_context() will unschedule cpu-flexible even if it
doesn't need a rotation.

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
kernel/events/core.c

index 460e485..f98c0f8 100644 (file)
@@ -430,7 +430,7 @@ static void update_perf_cpu_limits(void)
        WRITE_ONCE(perf_sample_allowed_ns, tmp);
 }
 
-static int perf_rotate_context(struct perf_cpu_context *cpuctx);
+static bool 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,
@@ -1041,7 +1041,7 @@ list_update_cgroup_event(struct perf_event *event,
 static enum hrtimer_restart perf_mux_hrtimer_handler(struct hrtimer *hr)
 {
        struct perf_cpu_context *cpuctx;
-       int rotations = 0;
+       bool rotations;
 
        lockdep_assert_irqs_disabled();
 
@@ -3600,52 +3600,66 @@ static void rotate_ctx(struct perf_event_context *ctx, struct perf_event *event)
        perf_event_groups_insert(&ctx->flexible_groups, event);
 }
 
-static int perf_rotate_context(struct perf_cpu_context *cpuctx)
+static inline struct perf_event *
+ctx_first_active(struct perf_event_context *ctx)
 {
-       struct perf_event *ctx_event = NULL, *cpuctx_event = NULL;
+       return list_first_entry_or_null(&ctx->flexible_active,
+                                       struct perf_event, active_list);
+}
+
+static bool perf_rotate_context(struct perf_cpu_context *cpuctx)
+{
+       struct perf_event *cpu_event = NULL, *task_event = NULL;
+       bool cpu_rotate = false, task_rotate = false;
        struct perf_event_context *ctx = NULL;
-       int rotate = 0;
+
+       /*
+        * Since we run this from IRQ context, nobody can install new
+        * events, thus the event count values are stable.
+        */
 
        if (cpuctx->ctx.nr_events) {
                if (cpuctx->ctx.nr_events != cpuctx->ctx.nr_active)
-                       rotate = 1;
+                       cpu_rotate = true;
        }
 
        ctx = cpuctx->task_ctx;
        if (ctx && ctx->nr_events) {
                if (ctx->nr_events != ctx->nr_active)
-                       rotate = 1;
+                       task_rotate = true;
        }
 
-       if (!rotate)
-               goto done;
+       if (!(cpu_rotate || task_rotate))
+               return false;
 
        perf_ctx_lock(cpuctx, cpuctx->task_ctx);
        perf_pmu_disable(cpuctx->ctx.pmu);
 
-       cpuctx_event = list_first_entry_or_null(&cpuctx->ctx.flexible_active,
-                                               struct perf_event, active_list);
-       if (ctx) {
-               ctx_event = list_first_entry_or_null(&ctx->flexible_active,
-                                                    struct perf_event, active_list);
-       }
+       if (task_rotate)
+               task_event = ctx_first_active(ctx);
+       if (cpu_rotate)
+               cpu_event = ctx_first_active(&cpuctx->ctx);
 
-       cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
-       if (ctx)
+       /*
+        * As per the order given at ctx_resched() first 'pop' task flexible
+        * and then, if needed CPU flexible.
+        */
+       if (task_event || (ctx && cpu_event))
                ctx_sched_out(ctx, cpuctx, EVENT_FLEXIBLE);
+       if (cpu_event)
+               cpu_ctx_sched_out(cpuctx, EVENT_FLEXIBLE);
 
-       if (cpuctx_event)
-               rotate_ctx(&cpuctx->ctx, cpuctx_event);
-       if (ctx_event)
-               rotate_ctx(ctx, ctx_event);
+       if (task_event)
+               rotate_ctx(ctx, task_event);
+       if (cpu_event)
+               rotate_ctx(&cpuctx->ctx, cpu_event);
 
        perf_event_sched_in(cpuctx, ctx, current);
 
        perf_pmu_enable(cpuctx->ctx.pmu);
        perf_ctx_unlock(cpuctx, cpuctx->task_ctx);
-done:
 
-       return rotate;
+       return true;
 }
 
 void perf_event_task_tick(void)