perf_counter: rework context time
authorPeter Zijlstra <a.p.zijlstra@chello.nl>
Mon, 6 Apr 2009 09:45:10 +0000 (11:45 +0200)
committerIngo Molnar <mingo@elte.hu>
Tue, 7 Apr 2009 08:49:00 +0000 (10:49 +0200)
Since perf_counter_context is switched along with tasks, we can
maintain the context time without using the task runtime clock.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <20090406094518.353552838@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
include/linux/perf_counter.h
kernel/perf_counter.c

index 0f5a4005048f5eda7c00ac79ab28acc41f1ecb48..7f5d353d78ac595ad9a5ec63497174c4a0766f66 100644 (file)
@@ -477,14 +477,10 @@ struct perf_counter_context {
        struct task_struct      *task;
 
        /*
-        * time_now is the current time in nanoseconds since an arbitrary
-        * point in the past.  For per-task counters, this is based on the
-        * task clock, and for per-cpu counters it is based on the cpu clock.
-        * time_lost is an offset from the task/cpu clock, used to make it
-        * appear that time only passes while the context is scheduled in.
+        * Context clock, runs when context enabled.
         */
-       u64                     time_now;
-       u64                     time_lost;
+       u64                     time;
+       u64                     timestamp;
 #endif
 };
 
index 8c8eaf0625f9a433f6e7be3fa2dabc6f1fee27e2..84d85ab4e1611b4cb4cb3b063a1ec170e14be2b5 100644 (file)
@@ -117,7 +117,7 @@ counter_sched_out(struct perf_counter *counter,
                return;
 
        counter->state = PERF_COUNTER_STATE_INACTIVE;
-       counter->tstamp_stopped = ctx->time_now;
+       counter->tstamp_stopped = ctx->time;
        counter->hw_ops->disable(counter);
        counter->oncpu = -1;
 
@@ -253,27 +253,20 @@ retry:
        spin_unlock_irq(&ctx->lock);
 }
 
-/*
- * Get the current time for this context.
- * If this is a task context, we use the task's task clock,
- * or for a per-cpu context, we use the cpu clock.
- */
-static u64 get_context_time(struct perf_counter_context *ctx, int update)
+static inline u64 perf_clock(void)
 {
-       struct task_struct *curr = ctx->task;
-
-       if (!curr)
-               return cpu_clock(smp_processor_id());
-
-       return __task_delta_exec(curr, update) + curr->se.sum_exec_runtime;
+       return cpu_clock(smp_processor_id());
 }
 
 /*
  * Update the record of the current time in a context.
  */
-static void update_context_time(struct perf_counter_context *ctx, int update)
+static void update_context_time(struct perf_counter_context *ctx)
 {
-       ctx->time_now = get_context_time(ctx, update) - ctx->time_lost;
+       u64 now = perf_clock();
+
+       ctx->time += now - ctx->timestamp;
+       ctx->timestamp = now;
 }
 
 /*
@@ -284,15 +277,17 @@ static void update_counter_times(struct perf_counter *counter)
        struct perf_counter_context *ctx = counter->ctx;
        u64 run_end;
 
-       if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
-               counter->total_time_enabled = ctx->time_now -
-                       counter->tstamp_enabled;
-               if (counter->state == PERF_COUNTER_STATE_INACTIVE)
-                       run_end = counter->tstamp_stopped;
-               else
-                       run_end = ctx->time_now;
-               counter->total_time_running = run_end - counter->tstamp_running;
-       }
+       if (counter->state < PERF_COUNTER_STATE_INACTIVE)
+               return;
+
+       counter->total_time_enabled = ctx->time - counter->tstamp_enabled;
+
+       if (counter->state == PERF_COUNTER_STATE_INACTIVE)
+               run_end = counter->tstamp_stopped;
+       else
+               run_end = ctx->time;
+
+       counter->total_time_running = run_end - counter->tstamp_running;
 }
 
 /*
@@ -332,7 +327,7 @@ static void __perf_counter_disable(void *info)
         * If it is in error state, leave it in error state.
         */
        if (counter->state >= PERF_COUNTER_STATE_INACTIVE) {
-               update_context_time(ctx, 1);
+               update_context_time(ctx);
                update_counter_times(counter);
                if (counter == counter->group_leader)
                        group_sched_out(counter, cpuctx, ctx);
@@ -426,7 +421,7 @@ counter_sched_in(struct perf_counter *counter,
                return -EAGAIN;
        }
 
-       counter->tstamp_running += ctx->time_now - counter->tstamp_stopped;
+       counter->tstamp_running += ctx->time - counter->tstamp_stopped;
 
        if (!is_software_counter(counter))
                cpuctx->active_oncpu++;
@@ -493,9 +488,9 @@ static void add_counter_to_ctx(struct perf_counter *counter,
        list_add_counter(counter, ctx);
        ctx->nr_counters++;
        counter->prev_state = PERF_COUNTER_STATE_OFF;
-       counter->tstamp_enabled = ctx->time_now;
-       counter->tstamp_running = ctx->time_now;
-       counter->tstamp_stopped = ctx->time_now;
+       counter->tstamp_enabled = ctx->time;
+       counter->tstamp_running = ctx->time;
+       counter->tstamp_stopped = ctx->time;
 }
 
 /*
@@ -522,7 +517,7 @@ static void __perf_install_in_context(void *info)
 
        curr_rq_lock_irq_save(&flags);
        spin_lock(&ctx->lock);
-       update_context_time(ctx, 1);
+       update_context_time(ctx);
 
        /*
         * Protect the list operation against NMI by disabling the
@@ -648,13 +643,13 @@ static void __perf_counter_enable(void *info)
 
        curr_rq_lock_irq_save(&flags);
        spin_lock(&ctx->lock);
-       update_context_time(ctx, 1);
+       update_context_time(ctx);
 
        counter->prev_state = counter->state;
        if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
                goto unlock;
        counter->state = PERF_COUNTER_STATE_INACTIVE;
-       counter->tstamp_enabled = ctx->time_now - counter->total_time_enabled;
+       counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
 
        /*
         * If the counter is in a group and isn't the group leader,
@@ -737,8 +732,8 @@ static void perf_counter_enable(struct perf_counter *counter)
         */
        if (counter->state == PERF_COUNTER_STATE_OFF) {
                counter->state = PERF_COUNTER_STATE_INACTIVE;
-               counter->tstamp_enabled = ctx->time_now -
-                       counter->total_time_enabled;
+               counter->tstamp_enabled =
+                       ctx->time - counter->total_time_enabled;
        }
  out:
        spin_unlock_irq(&ctx->lock);
@@ -778,7 +773,7 @@ void __perf_counter_sched_out(struct perf_counter_context *ctx,
        ctx->is_active = 0;
        if (likely(!ctx->nr_counters))
                goto out;
-       update_context_time(ctx, 0);
+       update_context_time(ctx);
 
        flags = hw_perf_save_disable();
        if (ctx->nr_active) {
@@ -883,12 +878,7 @@ __perf_counter_sched_in(struct perf_counter_context *ctx,
        if (likely(!ctx->nr_counters))
                goto out;
 
-       /*
-        * Add any time since the last sched_out to the lost time
-        * so it doesn't get included in the total_time_enabled and
-        * total_time_running measures for counters in the context.
-        */
-       ctx->time_lost = get_context_time(ctx, 0) - ctx->time_now;
+       ctx->timestamp = perf_clock();
 
        flags = hw_perf_save_disable();
 
@@ -1043,8 +1033,8 @@ int perf_counter_task_enable(void)
                if (counter->state > PERF_COUNTER_STATE_OFF)
                        continue;
                counter->state = PERF_COUNTER_STATE_INACTIVE;
-               counter->tstamp_enabled = ctx->time_now -
-                       counter->total_time_enabled;
+               counter->tstamp_enabled =
+                       ctx->time - counter->total_time_enabled;
                counter->hw_event.disabled = 0;
        }
        hw_perf_restore(perf_flags);
@@ -1113,7 +1103,7 @@ static void __read(void *info)
 
        curr_rq_lock_irq_save(&flags);
        if (ctx->is_active)
-               update_context_time(ctx, 1);
+               update_context_time(ctx);
        counter->hw_ops->read(counter);
        update_counter_times(counter);
        curr_rq_unlock_irq_restore(&flags);