sched: optimize RT affinity
authorGregory Haskins <ghaskins@novell.com>
Fri, 25 Jan 2008 20:08:11 +0000 (21:08 +0100)
committerIngo Molnar <mingo@elte.hu>
Fri, 25 Jan 2008 20:08:11 +0000 (21:08 +0100)
The current code base assumes a relatively flat CPU/core topology and will
route RT tasks to any CPU fairly equally.  In the real world, there are
various toplogies and affinities that govern where a task is best suited to
run with the smallest amount of overhead.  NUMA and multi-core CPUs are
prime examples of topologies that can impact cache performance.

Fortunately, linux is already structured to represent these topologies via
the sched_domains interface.  So we change our RT router to consult a
combination of topology and affinity policy to best place tasks during
migration.

Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
kernel/sched_rt.c

index ac7d06786454aae2f48b8ece928d66a1e835fc7e..a9d7d4408160c405f7562913adaad4a10a1e344c 100644 (file)
@@ -281,35 +281,111 @@ static struct task_struct *pick_next_highest_task_rt(struct rq *rq,
 }
 
 static DEFINE_PER_CPU(cpumask_t, local_cpu_mask);
+static DEFINE_PER_CPU(cpumask_t, valid_cpu_mask);
 
-static int find_lowest_rq(struct task_struct *task)
+static int find_lowest_cpus(struct task_struct *task, cpumask_t *lowest_mask)
 {
-       int cpu;
-       cpumask_t *cpu_mask = &__get_cpu_var(local_cpu_mask);
-       struct rq *lowest_rq = NULL;
+       int       cpu;
+       cpumask_t *valid_mask = &__get_cpu_var(valid_cpu_mask);
+       int       lowest_prio = -1;
+       int       ret         = 0;
 
-       cpus_and(*cpu_mask, cpu_online_map, task->cpus_allowed);
+       cpus_clear(*lowest_mask);
+       cpus_and(*valid_mask, cpu_online_map, task->cpus_allowed);
 
        /*
         * Scan each rq for the lowest prio.
         */
-       for_each_cpu_mask(cpu, *cpu_mask) {
+       for_each_cpu_mask(cpu, *valid_mask) {
                struct rq *rq = cpu_rq(cpu);
 
                /* We look for lowest RT prio or non-rt CPU */
                if (rq->rt.highest_prio >= MAX_RT_PRIO) {
-                       lowest_rq = rq;
-                       break;
+                       if (ret)
+                               cpus_clear(*lowest_mask);
+                       cpu_set(rq->cpu, *lowest_mask);
+                       return 1;
                }
 
                /* no locking for now */
-               if (rq->rt.highest_prio > task->prio &&
-                   (!lowest_rq || rq->rt.highest_prio > lowest_rq->rt.highest_prio)) {
-                       lowest_rq = rq;
+               if ((rq->rt.highest_prio > task->prio)
+                   && (rq->rt.highest_prio >= lowest_prio)) {
+                       if (rq->rt.highest_prio > lowest_prio) {
+                               /* new low - clear old data */
+                               lowest_prio = rq->rt.highest_prio;
+                               cpus_clear(*lowest_mask);
+                       }
+                       cpu_set(rq->cpu, *lowest_mask);
+                       ret = 1;
+               }
+       }
+
+       return ret;
+}
+
+static inline int pick_optimal_cpu(int this_cpu, cpumask_t *mask)
+{
+       int first;
+
+       /* "this_cpu" is cheaper to preempt than a remote processor */
+       if ((this_cpu != -1) && cpu_isset(this_cpu, *mask))
+               return this_cpu;
+
+       first = first_cpu(*mask);
+       if (first != NR_CPUS)
+               return first;
+
+       return -1;
+}
+
+static int find_lowest_rq(struct task_struct *task)
+{
+       struct sched_domain *sd;
+       cpumask_t *lowest_mask = &__get_cpu_var(local_cpu_mask);
+       int this_cpu = smp_processor_id();
+       int cpu      = task_cpu(task);
+
+       if (!find_lowest_cpus(task, lowest_mask))
+               return -1;
+
+       /*
+        * At this point we have built a mask of cpus representing the
+        * lowest priority tasks in the system.  Now we want to elect
+        * the best one based on our affinity and topology.
+        *
+        * We prioritize the last cpu that the task executed on since
+        * it is most likely cache-hot in that location.
+        */
+       if (cpu_isset(cpu, *lowest_mask))
+               return cpu;
+
+       /*
+        * Otherwise, we consult the sched_domains span maps to figure
+        * out which cpu is logically closest to our hot cache data.
+        */
+       if (this_cpu == cpu)
+               this_cpu = -1; /* Skip this_cpu opt if the same */
+
+       for_each_domain(cpu, sd) {
+               if (sd->flags & SD_WAKE_AFFINE) {
+                       cpumask_t domain_mask;
+                       int       best_cpu;
+
+                       cpus_and(domain_mask, sd->span, *lowest_mask);
+
+                       best_cpu = pick_optimal_cpu(this_cpu,
+                                                   &domain_mask);
+                       if (best_cpu != -1)
+                               return best_cpu;
                }
        }
 
-       return lowest_rq ? lowest_rq->cpu : -1;
+       /*
+        * And finally, if there were no matches within the domains
+        * just give the caller *something* to work with from the compatible
+        * locations.
+        */
+       return pick_optimal_cpu(this_cpu, lowest_mask);
 }
 
 /* Will lock the rq it finds */