[PATCH] sleep profiling

Implement prof=sleep profiling.  TASK_UNINTERRUPTIBLE sleeps will be taken
as a profile hit, and every millisecond spent sleeping causes a profile-hit
for the call site that initiated the sleep.

Sample readprofile output on i386:

   306 ps2_sendbyte                               1.3973
   432 call_usermodehelper_keys                   1.9548
   484 ps2_command                                0.6453
   790 __driver_attach                            4.7879
  1593 msleep                                    44.2500
  3976 sync_buffer                               64.1290
  4076 do_lookup                                 12.4648
  8587 sync_page                                122.6714
 20820 total                                      0.0067

(NOTE: architectures need to check whether get_wchan() can be called from
deep within the wakeup path.)

akpm: we need to mark more functions __sched.  lock_sock(), msleep(), others..

akpm: the contention in do_lookup() is a surprise.  Presumably doing disk
reads for directory contents while holding i_mutex.

[akpm@osdl.org: various fixes]
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

Documentation/kernel-parameters.txt

@@ -1294,6 +1294,7 @@ and is between 256 and 4096 characters. It is defined in the file
 			Param: "schedule" - profile schedule points.
 			Param: <number> - step/bucket size as a power of 2 for
 				statistical time based profiling.
+			Param: "sleep" - profile D-state sleeping (millisecs)
 
 	processor.max_cstate=	[HW,ACPI]
 			Limit processor to maximum C-state

include/linux/profile.h

@@ -6,10 +6,15 @@
 #include <linux/kernel.h>
 #include <linux/init.h>
 #include <linux/cpumask.h>
+#include <linux/cache.h>
+
 #include <asm/errno.h>
 
+extern int prof_on __read_mostly;
+
 #define CPU_PROFILING	1
 #define SCHED_PROFILING	2
+#define SLEEP_PROFILING	3
 
 struct proc_dir_entry;
 struct pt_regs;
@@ -18,7 +23,24 @@ struct notifier_block;
 /* init basic kernel profiler */
 void __init profile_init(void);
 void profile_tick(int);
-void profile_hit(int, void *);
+
+/*
+ * Add multiple profiler hits to a given address:
+ */
+void profile_hits(int, void *ip, unsigned int nr_hits);
+
+/*
+ * Single profiler hit:
+ */
+static inline void profile_hit(int type, void *ip)
+{
+	/*
+	 * Speedup for the common (no profiling enabled) case:
+	 */
+	if (unlikely(prof_on == type))
+		profile_hits(type, ip, 1);
+}
+
 #ifdef CONFIG_PROC_FS
 void create_prof_cpu_mask(struct proc_dir_entry *);
 #else

kernel/profile.c

@@ -40,7 +40,7 @@ int (*timer_hook)(struct pt_regs *) __read_mostly;
 
 static atomic_t *prof_buffer;
 static unsigned long prof_len, prof_shift;
-static int prof_on __read_mostly;
+int prof_on __read_mostly;
 static cpumask_t prof_cpu_mask = CPU_MASK_ALL;
 #ifdef CONFIG_SMP
 static DEFINE_PER_CPU(struct profile_hit *[2], cpu_profile_hits);
@@ -51,9 +51,19 @@ static DEFINE_MUTEX(profile_flip_mutex);
 static int __init profile_setup(char * str)
 {
 	static char __initdata schedstr[] = "schedule";
+	static char __initdata sleepstr[] = "sleep";
 	int par;
 
-	if (!strncmp(str, schedstr, strlen(schedstr))) {
+	if (!strncmp(str, sleepstr, strlen(sleepstr))) {
+		prof_on = SLEEP_PROFILING;
+		if (str[strlen(sleepstr)] == ',')
+			str += strlen(sleepstr) + 1;
+		if (get_option(&str, &par))
+			prof_shift = par;
+		printk(KERN_INFO
+			"kernel sleep profiling enabled (shift: %ld)\n",
+			prof_shift);
+	} else if (!strncmp(str, sleepstr, strlen(sleepstr))) {
 		prof_on = SCHED_PROFILING;
 		if (str[strlen(schedstr)] == ',')
 			str += strlen(schedstr) + 1;
@@ -204,7 +214,8 @@ EXPORT_SYMBOL_GPL(profile_event_unregister);
  * positions to which hits are accounted during short intervals (e.g.
  * several seconds) is usually very small. Exclusion from buffer
  * flipping is provided by interrupt disablement (note that for
- * SCHED_PROFILING profile_hit() may be called from process context).
+ * SCHED_PROFILING or SLEEP_PROFILING profile_hit() may be called from
+ * process context).
  * The hash function is meant to be lightweight as opposed to strong,
  * and was vaguely inspired by ppc64 firmware-supported inverted
  * pagetable hash functions, but uses a full hashtable full of finite
@@ -257,7 +268,7 @@ static void profile_discard_flip_buffers(void)
 	mutex_unlock(&profile_flip_mutex);
 }
 
-void profile_hit(int type, void *__pc)
+void profile_hits(int type, void *__pc, unsigned int nr_hits)
 {
 	unsigned long primary, secondary, flags, pc = (unsigned long)__pc;
 	int i, j, cpu;
@@ -274,21 +285,31 @@ void profile_hit(int type, void *__pc)
 		put_cpu();
 		return;
 	}
+	/*
+	 * We buffer the global profiler buffer into a per-CPU
+	 * queue and thus reduce the number of global (and possibly
+	 * NUMA-alien) accesses. The write-queue is self-coalescing:
+	 */
 	local_irq_save(flags);
 	do {
 		for (j = 0; j < PROFILE_GRPSZ; ++j) {
 			if (hits[i + j].pc == pc) {
-				hits[i + j].hits++;
+				hits[i + j].hits += nr_hits;
 				goto out;
 			} else if (!hits[i + j].hits) {
 				hits[i + j].pc = pc;
-				hits[i + j].hits = 1;
+				hits[i + j].hits = nr_hits;
 				goto out;
 			}
 		}
 		i = (i + secondary) & (NR_PROFILE_HIT - 1);
 	} while (i != primary);
-	atomic_inc(&prof_buffer[pc]);
+
+	/*
+	 * Add the current hit(s) and flush the write-queue out
+	 * to the global buffer:
+	 */
+	atomic_add(nr_hits, &prof_buffer[pc]);
 	for (i = 0; i < NR_PROFILE_HIT; ++i) {
 		atomic_add(hits[i].hits, &prof_buffer[hits[i].pc]);
 		hits[i].pc = hits[i].hits = 0;
@@ -356,14 +377,14 @@ static int __devinit profile_cpu_callback(struct notifier_block *info,
 #define profile_flip_buffers()		do { } while (0)
 #define profile_discard_flip_buffers()	do { } while (0)
 
-void profile_hit(int type, void *__pc)
+void profile_hits(int type, void *__pc, unsigned int nr_hits)
 {
 	unsigned long pc;
 
 	if (prof_on != type || !prof_buffer)
 		return;
 	pc = ((unsigned long)__pc - (unsigned long)_stext) >> prof_shift;
-	atomic_inc(&prof_buffer[min(pc, prof_len - 1)]);
+	atomic_add(nr_hits, &prof_buffer[min(pc, prof_len - 1)]);
 }
 #endif /* !CONFIG_SMP */
 

kernel/sched.c

@@ -948,6 +948,17 @@ static void activate_task(struct task_struct *p, struct rq *rq, int local)
 	}
 #endif
 
+	/*
+	 * Sleep time is in units of nanosecs, so shift by 20 to get a
+	 * milliseconds-range estimation of the amount of time that the task
+	 * spent sleeping:
+	 */
+	if (unlikely(prof_on == SLEEP_PROFILING)) {
+		if (p->state == TASK_UNINTERRUPTIBLE)
+			profile_hits(SLEEP_PROFILING, (void *)get_wchan(p),
+				     (now - p->timestamp) >> 20);
+	}
+
 	if (!rt_task(p))
 		p->prio = recalc_task_prio(p, now);