Merge tag 'mac80211-for-davem-2016-09-13' of git://git.kernel.org/pub/scm/linux/kerne...
[cascardo/linux.git] / kernel / sched / core.c
index 5c883fe..2a906f2 100644 (file)
@@ -74,6 +74,7 @@
 #include <linux/context_tracking.h>
 #include <linux/compiler.h>
 #include <linux/frame.h>
+#include <linux/prefetch.h>
 
 #include <asm/switch_to.h>
 #include <asm/tlb.h>
@@ -2971,6 +2972,23 @@ DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
 EXPORT_PER_CPU_SYMBOL(kstat);
 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
 
+/*
+ * The function fair_sched_class.update_curr accesses the struct curr
+ * and its field curr->exec_start; when called from task_sched_runtime(),
+ * we observe a high rate of cache misses in practice.
+ * Prefetching this data results in improved performance.
+ */
+static inline void prefetch_curr_exec_start(struct task_struct *p)
+{
+#ifdef CONFIG_FAIR_GROUP_SCHED
+       struct sched_entity *curr = (&p->se)->cfs_rq->curr;
+#else
+       struct sched_entity *curr = (&task_rq(p)->cfs)->curr;
+#endif
+       prefetch(curr);
+       prefetch(&curr->exec_start);
+}
+
 /*
  * Return accounted runtime for the task.
  * In case the task is currently running, return the runtime plus current's
@@ -3005,6 +3023,7 @@ unsigned long long task_sched_runtime(struct task_struct *p)
         * thread, breaking clock_gettime().
         */
        if (task_current(rq, p) && task_on_rq_queued(p)) {
+               prefetch_curr_exec_start(p);
                update_rq_clock(rq);
                p->sched_class->update_curr(rq);
        }