[PATCH] sched: improve load balancing pinned tasks
authorNick Piggin <nickpiggin@yahoo.com.au>
Sat, 25 Jun 2005 21:57:07 +0000 (14:57 -0700)
committerLinus Torvalds <torvalds@ppc970.osdl.org>
Sat, 25 Jun 2005 23:24:40 +0000 (16:24 -0700)
John Hawkes explained the problem best:

A large number of processes that are pinned to a single CPU results
in every other CPU's load_balance() seeing this overloaded CPU as
"busiest", yet move_tasks() never finds a task to pull-migrate.  This
condition occurs during module unload, but can also occur as a
denial-of-service using sys_sched_setaffinity().  Several hundred
CPUs performing this fruitless load_balance() will livelock on the
busiest CPU's runqueue lock.  A smaller number of CPUs will livelock
if the pinned task count gets high.

Expanding slightly on John's patch, this one attempts to work out whether the
balancing failure has been due to too many tasks pinned on the runqueue.  This
allows it to be basically invisible to the regular blancing paths (ie.  when
there are no pinned tasks).  We can use this extra knowledge to shut down the
balancing faster, and ensure the migration threads don't start running which
is another problem observed in the wild.

Signed-off-by: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
kernel/sched.c

index 86be13ee500675c27d1792359cdf94d13850fbc6..2794c79b91977c0980d849b73905e1e644296937 100644 (file)
@@ -1632,7 +1632,7 @@ void pull_task(runqueue_t *src_rq, prio_array_t *src_array, task_t *p,
  */
 static inline
 int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
-                    struct sched_domain *sd, enum idle_type idle)
+            struct sched_domain *sd, enum idle_type idle, int *all_pinned)
 {
        /*
         * We do not migrate tasks that are:
@@ -1640,10 +1640,12 @@ int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
         * 2) cannot be migrated to this CPU due to cpus_allowed, or
         * 3) are cache-hot on their current CPU.
         */
-       if (task_running(rq, p))
-               return 0;
        if (!cpu_isset(this_cpu, p->cpus_allowed))
                return 0;
+       *all_pinned = 0;
+
+       if (task_running(rq, p))
+               return 0;
 
        /*
         * Aggressive migration if:
@@ -1656,7 +1658,7 @@ int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
                return 1;
 
        if (task_hot(p, rq->timestamp_last_tick, sd))
-                       return 0;
+               return 0;
        return 1;
 }
 
@@ -1669,16 +1671,18 @@ int can_migrate_task(task_t *p, runqueue_t *rq, int this_cpu,
  */
 static int move_tasks(runqueue_t *this_rq, int this_cpu, runqueue_t *busiest,
                      unsigned long max_nr_move, struct sched_domain *sd,
-                     enum idle_type idle)
+                     enum idle_type idle, int *all_pinned)
 {
        prio_array_t *array, *dst_array;
        struct list_head *head, *curr;
-       int idx, pulled = 0;
+       int idx, pulled = 0, pinned = 0;
        task_t *tmp;
 
-       if (max_nr_move <= 0 || busiest->nr_running <= 1)
+       if (max_nr_move == 0)
                goto out;
 
+       pinned = 1;
+
        /*
         * We first consider expired tasks. Those will likely not be
         * executed in the near future, and they are most likely to
@@ -1717,7 +1721,7 @@ skip_queue:
 
        curr = curr->prev;
 
-       if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle)) {
+       if (!can_migrate_task(tmp, busiest, this_cpu, sd, idle, &pinned)) {
                if (curr != head)
                        goto skip_queue;
                idx++;
@@ -1746,6 +1750,9 @@ out:
         * inside pull_task().
         */
        schedstat_add(sd, lb_gained[idle], pulled);
+
+       if (all_pinned)
+               *all_pinned = pinned;
        return pulled;
 }
 
@@ -1917,7 +1924,8 @@ static int load_balance(int this_cpu, runqueue_t *this_rq,
        struct sched_group *group;
        runqueue_t *busiest;
        unsigned long imbalance;
-       int nr_moved;
+       int nr_moved, all_pinned;
+       int active_balance = 0;
 
        spin_lock(&this_rq->lock);
        schedstat_inc(sd, lb_cnt[idle]);
@@ -1956,9 +1964,15 @@ static int load_balance(int this_cpu, runqueue_t *this_rq,
                 */
                double_lock_balance(this_rq, busiest);
                nr_moved = move_tasks(this_rq, this_cpu, busiest,
-                                               imbalance, sd, idle);
+                                               imbalance, sd, idle,
+                                               &all_pinned);
                spin_unlock(&busiest->lock);
+
+               /* All tasks on this runqueue were pinned by CPU affinity */
+               if (unlikely(all_pinned))
+                       goto out_balanced;
        }
+
        spin_unlock(&this_rq->lock);
 
        if (!nr_moved) {
@@ -1966,16 +1980,15 @@ static int load_balance(int this_cpu, runqueue_t *this_rq,
                sd->nr_balance_failed++;
 
                if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
-                       int wake = 0;
 
                        spin_lock(&busiest->lock);
                        if (!busiest->active_balance) {
                                busiest->active_balance = 1;
                                busiest->push_cpu = this_cpu;
-                               wake = 1;
+                               active_balance = 1;
                        }
                        spin_unlock(&busiest->lock);
-                       if (wake)
+                       if (active_balance)
                                wake_up_process(busiest->migration_thread);
 
                        /*
@@ -1984,18 +1997,21 @@ static int load_balance(int this_cpu, runqueue_t *this_rq,
                         */
                        sd->nr_balance_failed = sd->cache_nice_tries;
                }
-
-               /*
-                * We were unbalanced, but unsuccessful in move_tasks(),
-                * so bump the balance_interval to lessen the lock contention.
-                */
-               if (sd->balance_interval < sd->max_interval)
-                       sd->balance_interval++;
-       } else {
+       } else
                sd->nr_balance_failed = 0;
 
+       if (likely(!active_balance)) {
                /* We were unbalanced, so reset the balancing interval */
                sd->balance_interval = sd->min_interval;
+       } else {
+               /*
+                * If we've begun active balancing, start to back off. This
+                * case may not be covered by the all_pinned logic if there
+                * is only 1 task on the busy runqueue (because we don't call
+                * move_tasks).
+                */
+               if (sd->balance_interval < sd->max_interval)
+                       sd->balance_interval *= 2;
        }
 
        return nr_moved;
@@ -2047,7 +2063,7 @@ static int load_balance_newidle(int this_cpu, runqueue_t *this_rq,
 
        schedstat_add(sd, lb_imbalance[NEWLY_IDLE], imbalance);
        nr_moved = move_tasks(this_rq, this_cpu, busiest,
-                                       imbalance, sd, NEWLY_IDLE);
+                                       imbalance, sd, NEWLY_IDLE, NULL);
        if (!nr_moved)
                schedstat_inc(sd, lb_failed[NEWLY_IDLE]);
 
@@ -2126,7 +2142,7 @@ static void active_load_balance(runqueue_t *busiest_rq, int busiest_cpu)
                                /* move a task from busiest_rq to target_rq */
                                double_lock_balance(busiest_rq, target_rq);
                                if (move_tasks(target_rq, cpu, busiest_rq,
-                                               1, sd, SCHED_IDLE)) {
+                                               1, sd, SCHED_IDLE, NULL)) {
                                        schedstat_inc(sd, alb_pushed);
                                } else {
                                        schedstat_inc(sd, alb_failed);