[patch] sched: implement cpu_clock(cpu) high-speed time source

[Ingo Molnar <mingo@elte.hu>]

* Jeremy Fitzhardinge <jeremy@goop.org> wrote:

> > yeah. But then it should not be using sched_clock() but CFS's new 
> > rq_clock() method - which does try to construct a globally valid 
> > timesource out of sched_clock(). [that fix is not backportable 
> > though]
> 
> Hm, that doesn't look quite right.  Doesn't rq_clock measure time 
> spent running?  Unstolen time includes idle time too (it just excludes 
> time in which a VCPU is runnable but not actually running).

generally rq_clock() also includes idle time, so it should work fine for 
this purpose. So, what do you think about the patch below - does it suit 
Xen's purposes?

	Ingo

-------------------->
Subject: sched: implement cpu_clock(cpu) high-speed time source
From: Ingo Molnar <mingo@elte.hu>

Implement the cpu_clock(cpu) interface for kernel-internal use:
high-speed (but slightly incorrect) per-cpu clock constructed from
sched_clock().

fix up blktrace and softlockup-watchdog to use this new interface.

Signed-off-by: Ingo Molnar <mingo@elte.hu>
---
 block/blktrace.c      |   20 ++++++++++----------
 include/linux/sched.h |    7 +++++++
 kernel/sched.c        |   17 +++++++++++++++++
 kernel/softlockup.c   |   10 ++++++----
 4 files changed, 40 insertions(+), 14 deletions(-)

Index: linux/block/blktrace.c
===================================================================
--- linux.orig/block/blktrace.c
+++ linux/block/blktrace.c
@@ -41,7 +41,7 @@ static void trace_note(struct blk_trace 
 		const int cpu = smp_processor_id();
 
 		t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
-		t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
+		t->time = cpu_clock(cpu) - per_cpu(blk_trace_cpu_offset, cpu);
 		t->device = bt->dev;
 		t->action = action;
 		t->pid = pid;
@@ -159,7 +159,7 @@ void __blk_add_trace(struct blk_trace *b
 
 		t->magic = BLK_IO_TRACE_MAGIC | BLK_IO_TRACE_VERSION;
 		t->sequence = ++(*sequence);
-		t->time = sched_clock() - per_cpu(blk_trace_cpu_offset, cpu);
+		t->time = cpu_clock(cpu) - per_cpu(blk_trace_cpu_offset, cpu);
 		t->sector = sector;
 		t->bytes = bytes;
 		t->action = what;
@@ -488,17 +488,17 @@ void blk_trace_shutdown(request_queue_t 
 }
 
 /*
- * Average offset over two calls to sched_clock() with a gettimeofday()
+ * Average offset over two calls to cpu_clock() with a gettimeofday()
  * in the middle
  */
-static void blk_check_time(unsigned long long *t)
+static void blk_check_time(unsigned long long *t, int this_cpu)
 {
 	unsigned long long a, b;
 	struct timeval tv;
 
-	a = sched_clock();
+	a = cpu_clock(this_cpu);
 	do_gettimeofday(&tv);
-	b = sched_clock();
+	b = cpu_clock(this_cpu);
 
 	*t = tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
 	*t -= (a + b) / 2;
@@ -510,16 +510,16 @@ static void blk_check_time(unsigned long
 static void blk_trace_check_cpu_time(void *data)
 {
 	unsigned long long *t;
-	int cpu = get_cpu();
+	int this_cpu = get_cpu();
 
-	t = &per_cpu(blk_trace_cpu_offset, cpu);
+	t = &per_cpu(blk_trace_cpu_offset, this_cpu);
 
 	/*
 	 * Just call it twice, hopefully the second call will be cache hot
 	 * and a little more precise
 	 */
-	blk_check_time(t);
-	blk_check_time(t);
+	blk_check_time(t, this_cpu);
+	blk_check_time(t, this_cpu);
 
 	put_cpu();
 }
Index: linux/include/linux/sched.h
===================================================================
--- linux.orig/include/linux/sched.h
+++ linux/include/linux/sched.h
@@ -1327,6 +1327,13 @@ static inline int set_cpus_allowed(struc
 #endif
 
 extern unsigned long long sched_clock(void);
+
+/*
+ * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
+ * clock constructed from sched_clock():
+ */
+extern unsigned long long cpu_clock(int cpu);
+
 extern unsigned long long
 task_sched_runtime(struct task_struct *task);
 
Index: linux/kernel/sched.c
===================================================================
--- linux.orig/kernel/sched.c
+++ linux/kernel/sched.c
@@ -365,6 +365,23 @@ static inline unsigned long long rq_cloc
 }
 
 /*
+ * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
+ * clock constructed from sched_clock():
+ */
+unsigned long long cpu_clock(int cpu)
+{
+	struct rq *rq = cpu_rq(cpu);
+	unsigned long long now;
+	unsigned long flags;
+
+	spin_lock_irqsave(&rq->lock, flags);
+	now = rq_clock(rq);
+	spin_unlock_irqrestore(&rq->lock, flags);
+
+	return now;
+}
+
+/*
  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
  * See detach_destroy_domains: synchronize_sched for details.
  *
Index: linux/kernel/softlockup.c
===================================================================
--- linux.orig/kernel/softlockup.c
+++ linux/kernel/softlockup.c
@@ -41,14 +41,16 @@ static struct notifier_block panic_block
  * resolution, and we don't need to waste time with a big divide when
  * 2^30ns == 1.074s.
  */
-static unsigned long get_timestamp(void)
+static unsigned long get_timestamp(int this_cpu)
 {
-	return sched_clock() >> 30;  /* 2^30 ~= 10^9 */
+	return cpu_clock(this_cpu) >> 30;  /* 2^30 ~= 10^9 */
 }
 
 void touch_softlockup_watchdog(void)
 {
-	__raw_get_cpu_var(touch_timestamp) = get_timestamp();
+	int this_cpu = raw_smp_processor_id();
+
+	per_cpu(touch_timestamp, this_cpu) = get_timestamp(this_cpu);
 }
 EXPORT_SYMBOL(touch_softlockup_watchdog);
 
@@ -94,7 +96,7 @@ void softlockup_tick(void)
 		return;
 	}
 
-	now = get_timestamp();
+	now = get_timestamp(this_cpu);
 
 	/* Wake up the high-prio watchdog task every second: */
 	if (now > (touch_timestamp + 1))