[PATCH] sched: improve migration accuracy
authorMike Galbraith <efault@gmx.de>
Sun, 10 Dec 2006 10:20:31 +0000 (02:20 -0800)
committerLinus Torvalds <torvalds@woody.osdl.org>
Sun, 10 Dec 2006 17:55:43 +0000 (09:55 -0800)
Co-opt rq->timestamp_last_tick to maintain a cache_hot_time evaluation
reference timestamp at both tick and sched times to prevent said reference,
formerly rq->timestamp_last_tick, from being behind task->last_ran at
evaluation time, and to move said reference closer to current time on the
remote processor, intent being to improve cache hot evaluation and
timestamp adjustment accuracy for task migration.

Fix minor sched_time double accounting error which occurs when a task
passing through schedule() does not schedule off, and takes the next timer
tick.

[kenneth.w.chen@intel.com: cleanup]
Signed-off-by: Mike Galbraith <efault@gmx.de>
Acked-by: Ingo Molnar <mingo@elte.hu>
Acked-by: Ken Chen <kenneth.w.chen@intel.com>
Cc: Don Mullis <dwm@meer.net>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
kernel/sched.c

index 2b2b780939c9818013324f4c5ad65af10bd5ed35..15ce772a471a25b20ab1d0c8602bbbf9c54a08b5 100644 (file)
@@ -225,7 +225,8 @@ struct rq {
        unsigned long nr_uninterruptible;
 
        unsigned long expired_timestamp;
-       unsigned long long timestamp_last_tick;
+       /* Cached timestamp set by update_cpu_clock() */
+       unsigned long long most_recent_timestamp;
        struct task_struct *curr, *idle;
        unsigned long next_balance;
        struct mm_struct *prev_mm;
@@ -944,8 +945,8 @@ static void activate_task(struct task_struct *p, struct rq *rq, int local)
        if (!local) {
                /* Compensate for drifting sched_clock */
                struct rq *this_rq = this_rq();
-               now = (now - this_rq->timestamp_last_tick)
-                       + rq->timestamp_last_tick;
+               now = (now - this_rq->most_recent_timestamp)
+                       + rq->most_recent_timestamp;
        }
 #endif
 
@@ -1689,8 +1690,8 @@ void fastcall wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
                 * Not the local CPU - must adjust timestamp. This should
                 * get optimised away in the !CONFIG_SMP case.
                 */
-               p->timestamp = (p->timestamp - this_rq->timestamp_last_tick)
-                                       + rq->timestamp_last_tick;
+               p->timestamp = (p->timestamp - this_rq->most_recent_timestamp)
+                                       + rq->most_recent_timestamp;
                __activate_task(p, rq);
                if (TASK_PREEMPTS_CURR(p, rq))
                        resched_task(rq->curr);
@@ -2068,8 +2069,8 @@ static void pull_task(struct rq *src_rq, struct prio_array *src_array,
        set_task_cpu(p, this_cpu);
        inc_nr_running(p, this_rq);
        enqueue_task(p, this_array);
-       p->timestamp = (p->timestamp - src_rq->timestamp_last_tick)
-                               + this_rq->timestamp_last_tick;
+       p->timestamp = (p->timestamp - src_rq->most_recent_timestamp)
+                               + this_rq->most_recent_timestamp;
        /*
         * Note that idle threads have a prio of MAX_PRIO, for this test
         * to be always true for them.
@@ -2105,10 +2106,15 @@ int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
         * 2) too many balance attempts have failed.
         */
 
-       if (sd->nr_balance_failed > sd->cache_nice_tries)
+       if (sd->nr_balance_failed > sd->cache_nice_tries) {
+#ifdef CONFIG_SCHEDSTATS
+               if (task_hot(p, rq->most_recent_timestamp, sd))
+                       schedstat_inc(sd, lb_hot_gained[idle]);
+#endif
                return 1;
+       }
 
-       if (task_hot(p, rq->timestamp_last_tick, sd))
+       if (task_hot(p, rq->most_recent_timestamp, sd))
                return 0;
        return 1;
 }
@@ -2206,11 +2212,6 @@ skip_queue:
                goto skip_bitmap;
        }
 
-#ifdef CONFIG_SCHEDSTATS
-       if (task_hot(tmp, busiest->timestamp_last_tick, sd))
-               schedstat_inc(sd, lb_hot_gained[idle]);
-#endif
-
        pull_task(busiest, array, tmp, this_rq, dst_array, this_cpu);
        pulled++;
        rem_load_move -= tmp->load_weight;
@@ -2971,7 +2972,8 @@ EXPORT_PER_CPU_SYMBOL(kstat);
 static inline void
 update_cpu_clock(struct task_struct *p, struct rq *rq, unsigned long long now)
 {
-       p->sched_time += now - max(p->timestamp, rq->timestamp_last_tick);
+       p->sched_time += now - p->last_ran;
+       p->last_ran = rq->most_recent_timestamp = now;
 }
 
 /*
@@ -2984,8 +2986,7 @@ unsigned long long current_sched_time(const struct task_struct *p)
        unsigned long flags;
 
        local_irq_save(flags);
-       ns = max(p->timestamp, task_rq(p)->timestamp_last_tick);
-       ns = p->sched_time + sched_clock() - ns;
+       ns = p->sched_time + sched_clock() - p->last_ran;
        local_irq_restore(flags);
 
        return ns;
@@ -3176,8 +3177,6 @@ void scheduler_tick(void)
 
        update_cpu_clock(p, rq, now);
 
-       rq->timestamp_last_tick = now;
-
        if (p == rq->idle)
                /* Task on the idle queue */
                wake_priority_sleeper(rq);
@@ -5032,8 +5031,8 @@ static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
                 * afterwards, and pretending it was a local activate.
                 * This way is cleaner and logically correct.
                 */
-               p->timestamp = p->timestamp - rq_src->timestamp_last_tick
-                               + rq_dest->timestamp_last_tick;
+               p->timestamp = p->timestamp - rq_src->most_recent_timestamp
+                               + rq_dest->most_recent_timestamp;
                deactivate_task(p, rq_src);
                __activate_task(p, rq_dest);
                if (TASK_PREEMPTS_CURR(p, rq_dest))