Merge branch 'sched-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip

* 'sched-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip: (46 commits)
  sched: Add comments to find_busiest_group() function
  sched: Refactor the power savings balance code
  sched: Optimize the !power_savings_balance during fbg()
  sched: Create a helper function to calculate imbalance
  sched: Create helper to calculate small_imbalance in fbg()
  sched: Create a helper function to calculate sched_domain stats for fbg()
  sched: Define structure to store the sched_domain statistics for fbg()
  sched: Create a helper function to calculate sched_group stats for fbg()
  sched: Define structure to store the sched_group statistics for fbg()
  sched: Fix indentations in find_busiest_group() using gotos
  sched: Simple helper functions for find_busiest_group()
  sched: remove unused fields from struct rq
  sched: jiffies not printed per CPU
  sched: small optimisation of can_migrate_task()
  sched: fix typos in documentation
  sched: add avg_overlap decay
  x86, sched_clock(): mark variables read-mostly
  sched: optimize ttwu vs group scheduling
  sched: TIF_NEED_RESCHED -> need_reshed() cleanup
  sched: don't rebalance if attached on NULL domain
  ...

kernel/latencytop.c

@@ -9,6 +9,44 @@
  * as published by the Free Software Foundation; version 2
  * of the License.
  */
+
+/*
+ * CONFIG_LATENCYTOP enables a kernel latency tracking infrastructure that is
+ * used by the "latencytop" userspace tool. The latency that is tracked is not
+ * the 'traditional' interrupt latency (which is primarily caused by something
+ * else consuming CPU), but instead, it is the latency an application encounters
+ * because the kernel sleeps on its behalf for various reasons.
+ *
+ * This code tracks 2 levels of statistics:
+ * 1) System level latency
+ * 2) Per process latency
+ *
+ * The latency is stored in fixed sized data structures in an accumulated form;
+ * if the "same" latency cause is hit twice, this will be tracked as one entry
+ * in the data structure. Both the count, total accumulated latency and maximum
+ * latency are tracked in this data structure. When the fixed size structure is
+ * full, no new causes are tracked until the buffer is flushed by writing to
+ * the /proc file; the userspace tool does this on a regular basis.
+ *
+ * A latency cause is identified by a stringified backtrace at the point that
+ * the scheduler gets invoked. The userland tool will use this string to
+ * identify the cause of the latency in human readable form.
+ *
+ * The information is exported via /proc/latency_stats and /proc/<pid>/latency.
+ * These files look like this:
+ *
+ * Latency Top version : v0.1
+ * 70 59433 4897 i915_irq_wait drm_ioctl vfs_ioctl do_vfs_ioctl sys_ioctl
+ * |    |    |    |
+ * |    |    |    +----> the stringified backtrace
+ * |    |    +---------> The maximum latency for this entry in microseconds
+ * |    +--------------> The accumulated latency for this entry (microseconds)
+ * +-------------------> The number of times this entry is hit
+ *
+ * (note: the average latency is the accumulated latency divided by the number
+ * of times)
+ */
+
 #include <linux/latencytop.h>
 #include <linux/kallsyms.h>
 #include <linux/seq_file.h>
@@ -72,7 +110,7 @@ account_global_scheduler_latency(struct task_struct *tsk, struct latency_record
 				firstnonnull = i;
 			continue;
 		}
-		for (q = 0 ; q < LT_BACKTRACEDEPTH ; q++) {
+		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
 			unsigned long record = lat->backtrace[q];
 
 			if (latency_record[i].backtrace[q] != record) {
@@ -101,31 +139,52 @@ account_global_scheduler_latency(struct task_struct *tsk, struct latency_record
 	memcpy(&latency_record[i], lat, sizeof(struct latency_record));
 }
 
-static inline void store_stacktrace(struct task_struct *tsk, struct latency_record *lat)
+/*
+ * Iterator to store a backtrace into a latency record entry
+ */
+static inline void store_stacktrace(struct task_struct *tsk,
+					struct latency_record *lat)
 {
 	struct stack_trace trace;
 
 	memset(&trace, 0, sizeof(trace));
 	trace.max_entries = LT_BACKTRACEDEPTH;
 	trace.entries = &lat->backtrace[0];
-	trace.skip = 0;
 	save_stack_trace_tsk(tsk, &trace);
 }
 
+/**
+ * __account_scheduler_latency - record an occured latency
+ * @tsk - the task struct of the task hitting the latency
+ * @usecs - the duration of the latency in microseconds
+ * @inter - 1 if the sleep was interruptible, 0 if uninterruptible
+ *
+ * This function is the main entry point for recording latency entries
+ * as called by the scheduler.
+ *
+ * This function has a few special cases to deal with normal 'non-latency'
+ * sleeps: specifically, interruptible sleep longer than 5 msec is skipped
+ * since this usually is caused by waiting for events via select() and co.
+ *
+ * Negative latencies (caused by time going backwards) are also explicitly
+ * skipped.
+ */
 void __sched
-account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
+__account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
 {
 	unsigned long flags;
 	int i, q;
 	struct latency_record lat;
 
-	if (!latencytop_enabled)
-		return;
-
 	/* Long interruptible waits are generally user requested... */
 	if (inter && usecs > 5000)
 		return;
 
+	/* Negative sleeps are time going backwards */
+	/* Zero-time sleeps are non-interesting */
+	if (usecs <= 0)
+		return;
+
 	memset(&lat, 0, sizeof(lat));
 	lat.count = 1;
 	lat.time = usecs;
@@ -143,12 +202,12 @@ account_scheduler_latency(struct task_struct *tsk, int usecs, int inter)
 	if (tsk->latency_record_count >= LT_SAVECOUNT)
 		goto out_unlock;
 
-	for (i = 0; i < LT_SAVECOUNT ; i++) {
+	for (i = 0; i < LT_SAVECOUNT; i++) {
 		struct latency_record *mylat;
 		int same = 1;
 
 		mylat = &tsk->latency_record[i];
-		for (q = 0 ; q < LT_BACKTRACEDEPTH ; q++) {
+		for (q = 0; q < LT_BACKTRACEDEPTH; q++) {
 			unsigned long record = lat.backtrace[q];
 
 			if (mylat->backtrace[q] != record) {
@@ -186,7 +245,7 @@ static int lstats_show(struct seq_file *m, void *v)
 	for (i = 0; i < MAXLR; i++) {
 		if (latency_record[i].backtrace[0]) {
 			int q;
-			seq_printf(m, "%i %li %li ",
+			seq_printf(m, "%i %lu %lu ",
 				latency_record[i].count,
 				latency_record[i].time,
 				latency_record[i].max);
@@ -223,7 +282,7 @@ static int lstats_open(struct inode *inode, struct file *filp)
 	return single_open(filp, lstats_show, NULL);
 }
 
-static struct file_operations lstats_fops = {
+static const struct file_operations lstats_fops = {
 	.open		= lstats_open,
 	.read		= seq_read,
 	.write		= lstats_write,
@@ -236,4 +295,4 @@ static int __init init_lstats_procfs(void)
 	proc_create("latency_stats", 0644, NULL, &lstats_fops);
 	return 0;
 }
-__initcall(init_lstats_procfs);
+device_initcall(init_lstats_procfs);

kernel/sched_clock.c

@@ -24,11 +24,11 @@
  * The clock: sched_clock_cpu() is monotonic per cpu, and should be somewhat
  * consistent between cpus (never more than 2 jiffies difference).
  */
-#include <linux/sched.h>
-#include <linux/percpu.h>
 #include <linux/spinlock.h>
-#include <linux/ktime.h>
 #include <linux/module.h>
+#include <linux/percpu.h>
+#include <linux/ktime.h>
+#include <linux/sched.h>
 
 /*
  * Scheduler clock - returns current time in nanosec units.
@@ -43,6 +43,7 @@ unsigned long long __attribute__((weak)) sched_clock(void)
 static __read_mostly int sched_clock_running;
 
 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
+__read_mostly int sched_clock_stable;
 
 struct sched_clock_data {
 	/*
@@ -87,7 +88,7 @@ void sched_clock_init(void)
 }
 
 /*
- * min,max except they take wrapping into account
+ * min, max except they take wrapping into account
  */
 
 static inline u64 wrap_min(u64 x, u64 y)
@@ -111,15 +112,13 @@ static u64 __update_sched_clock(struct sched_clock_data *scd, u64 now)
 	s64 delta = now - scd->tick_raw;
 	u64 clock, min_clock, max_clock;
 
-	WARN_ON_ONCE(!irqs_disabled());
-
 	if (unlikely(delta < 0))
 		delta = 0;
 
 	/*
 	 * scd->clock = clamp(scd->tick_gtod + delta,
-	 * 		      max(scd->tick_gtod, scd->clock),
-	 * 		      scd->tick_gtod + TICK_NSEC);
+	 *		      max(scd->tick_gtod, scd->clock),
+	 *		      scd->tick_gtod + TICK_NSEC);
 	 */
 
 	clock = scd->tick_gtod + delta;
@@ -148,12 +147,13 @@ static void lock_double_clock(struct sched_clock_data *data1,
 
 u64 sched_clock_cpu(int cpu)
 {
-	struct sched_clock_data *scd = cpu_sdc(cpu);
 	u64 now, clock, this_clock, remote_clock;
+	struct sched_clock_data *scd;
 
-	if (unlikely(!sched_clock_running))
-		return 0ull;
+	if (sched_clock_stable)
+		return sched_clock();
 
+	scd = cpu_sdc(cpu);
 	WARN_ON_ONCE(!irqs_disabled());
 	now = sched_clock();
 
@@ -195,14 +195,18 @@ u64 sched_clock_cpu(int cpu)
 
 void sched_clock_tick(void)
 {
-	struct sched_clock_data *scd = this_scd();
+	struct sched_clock_data *scd;
 	u64 now, now_gtod;
 
+	if (sched_clock_stable)
+		return;
+
 	if (unlikely(!sched_clock_running))
 		return;
 
 	WARN_ON_ONCE(!irqs_disabled());
 
+	scd = this_scd();
 	now_gtod = ktime_to_ns(ktime_get());
 	now = sched_clock();
 
@@ -250,7 +254,7 @@ u64 sched_clock_cpu(int cpu)
 	return sched_clock();
 }
 
-#endif
+#endif /* CONFIG_HAVE_UNSTABLE_SCHED_CLOCK */
 
 unsigned long long cpu_clock(int cpu)
 {

kernel/sched_debug.c

@@ -272,7 +272,6 @@ static void print_cpu(struct seq_file *m, int cpu)
 	P(nr_switches);
 	P(nr_load_updates);
 	P(nr_uninterruptible);
-	SEQ_printf(m, "  .%-30s: %lu\n", "jiffies", jiffies);
 	PN(next_balance);
 	P(curr->pid);
 	PN(clock);
@@ -287,9 +286,6 @@ static void print_cpu(struct seq_file *m, int cpu)
 #ifdef CONFIG_SCHEDSTATS
 #define P(n) SEQ_printf(m, "  .%-30s: %d\n", #n, rq->n);
 
-	P(yld_exp_empty);
-	P(yld_act_empty);
-	P(yld_both_empty);
 	P(yld_count);
 
 	P(sched_switch);
@@ -314,7 +310,7 @@ static int sched_debug_show(struct seq_file *m, void *v)
 	u64 now = ktime_to_ns(ktime_get());
 	int cpu;
 
-	SEQ_printf(m, "Sched Debug Version: v0.08, %s %.*s\n",
+	SEQ_printf(m, "Sched Debug Version: v0.09, %s %.*s\n",
 		init_utsname()->release,
 		(int)strcspn(init_utsname()->version, " "),
 		init_utsname()->version);
@@ -325,6 +321,7 @@ static int sched_debug_show(struct seq_file *m, void *v)
 	SEQ_printf(m, "  .%-40s: %Ld\n", #x, (long long)(x))
 #define PN(x) \
 	SEQ_printf(m, "  .%-40s: %Ld.%06ld\n", #x, SPLIT_NS(x))
+	P(jiffies);
 	PN(sysctl_sched_latency);
 	PN(sysctl_sched_min_granularity);
 	PN(sysctl_sched_wakeup_granularity);
@@ -397,6 +394,7 @@ void proc_sched_show_task(struct task_struct *p, struct seq_file *m)
 	PN(se.vruntime);
 	PN(se.sum_exec_runtime);
 	PN(se.avg_overlap);
+	PN(se.avg_wakeup);
 
 	nr_switches = p->nvcsw + p->nivcsw;
 

kernel/sched_fair.c

@@ -1314,16 +1314,63 @@ out:
 }
 #endif /* CONFIG_SMP */
 
-static unsigned long wakeup_gran(struct sched_entity *se)
+/*
+ * Adaptive granularity
+ *
+ * se->avg_wakeup gives the average time a task runs until it does a wakeup,
+ * with the limit of wakeup_gran -- when it never does a wakeup.
+ *
+ * So the smaller avg_wakeup is the faster we want this task to preempt,
+ * but we don't want to treat the preemptee unfairly and therefore allow it
+ * to run for at least the amount of time we'd like to run.
+ *
+ * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
+ *
+ * NOTE: we use *nr_running to scale with load, this nicely matches the
+ *       degrading latency on load.
+ */
+static unsigned long
+adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
+{
+	u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
+	u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
+	u64 gran = 0;
+
+	if (this_run < expected_wakeup)
+		gran = expected_wakeup - this_run;
+
+	return min_t(s64, gran, sysctl_sched_wakeup_granularity);
+}
+
+static unsigned long
+wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
 {
 	unsigned long gran = sysctl_sched_wakeup_granularity;
 
+	if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
+		gran = adaptive_gran(curr, se);
+
 	/*
-	 * More easily preempt - nice tasks, while not making it harder for
-	 * + nice tasks.
+	 * Since its curr running now, convert the gran from real-time
+	 * to virtual-time in his units.
 	 */
-	if (!sched_feat(ASYM_GRAN) || se->load.weight > NICE_0_LOAD)
-		gran = calc_delta_fair(sysctl_sched_wakeup_granularity, se);
+	if (sched_feat(ASYM_GRAN)) {
+		/*
+		 * By using 'se' instead of 'curr' we penalize light tasks, so
+		 * they get preempted easier. That is, if 'se' < 'curr' then
+		 * the resulting gran will be larger, therefore penalizing the
+		 * lighter, if otoh 'se' > 'curr' then the resulting gran will
+		 * be smaller, again penalizing the lighter task.
+		 *
+		 * This is especially important for buddies when the leftmost
+		 * task is higher priority than the buddy.
+		 */
+		if (unlikely(se->load.weight != NICE_0_LOAD))
+			gran = calc_delta_fair(gran, se);
+	} else {
+		if (unlikely(curr->load.weight != NICE_0_LOAD))
+			gran = calc_delta_fair(gran, curr);
+	}
 
 	return gran;
 }
@@ -1350,7 +1397,7 @@ wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
 	if (vdiff <= 0)
 		return -1;
 
-	gran = wakeup_gran(curr);
+	gran = wakeup_gran(curr, se);
 	if (vdiff > gran)
 		return 1;
 

kernel/sched_features.h

@@ -1,5 +1,6 @@
 SCHED_FEAT(NEW_FAIR_SLEEPERS, 1)
-SCHED_FEAT(NORMALIZED_SLEEPER, 1)
+SCHED_FEAT(NORMALIZED_SLEEPER, 0)
+SCHED_FEAT(ADAPTIVE_GRAN, 1)
 SCHED_FEAT(WAKEUP_PREEMPT, 1)
 SCHED_FEAT(START_DEBIT, 1)
 SCHED_FEAT(AFFINE_WAKEUPS, 1)

kernel/sched_stats.h

@@ -4,7 +4,7 @@
  * bump this up when changing the output format or the meaning of an existing
  * format, so that tools can adapt (or abort)
  */
-#define SCHEDSTAT_VERSION 14
+#define SCHEDSTAT_VERSION 15
 
 static int show_schedstat(struct seq_file *seq, void *v)
 {
@@ -26,9 +26,8 @@ static int show_schedstat(struct seq_file *seq, void *v)
 
 		/* runqueue-specific stats */
 		seq_printf(seq,
-		    "cpu%d %u %u %u %u %u %u %u %u %u %llu %llu %lu",
-		    cpu, rq->yld_both_empty,
-		    rq->yld_act_empty, rq->yld_exp_empty, rq->yld_count,
+		    "cpu%d %u %u %u %u %u %u %llu %llu %lu",
+		    cpu, rq->yld_count,
 		    rq->sched_switch, rq->sched_count, rq->sched_goidle,
 		    rq->ttwu_count, rq->ttwu_local,
 		    rq->rq_cpu_time,