Re: [PATCH v5] perf: Use monotonic clock as a source for timestamps

[Peter Zijlstra <peterz-wEGCiKHe2LqWVfeAwA7xHQ@public.gmane.org>]

On Thu, Feb 12, 2015 at 11:28:14AM +0100, Peter Zijlstra wrote:
> > and you would have to check the clocksource is TSC.
> 
> It implicitly does that; it has that sched_clock_stable() thing, but
> yeah I suppose someone could change the clocksource even though the tsc
> is stable.
> 
> Not using TSC when its available is quite crazy though.. but sure.

Something like this on top then.. it might have a few header issues, the
whole asm/tsc.h vs clocksource.h thing looks like pain.

I haven't tried to compile it, maybe we can move cycle_t into types and
fwd declare struct clocksource or whatnot.

Of course, all this is quite horrible on the timekeeping side; it might
be tglx and/or jstutlz are having spasms just reading it :-)

---
--- a/arch/x86/kernel/cpu/perf_event.c
+++ b/arch/x86/kernel/cpu/perf_event.c
@@ -1967,17 +1967,19 @@ static void local_clock_user_time(struct
 	cyc2ns_read_end(data);
 }
 
-extern void notrace __ktime_get_mono_fast(u64 *offset, u32 *mult, u16 *shift);
+extern bool notrace __ktime_get_mono_fast(cycle_t (*read)(struct clocksource *cs),
+		u64 *offset, u32 *mult, u16 *shift);
 
 static void ktime_fast_mono_user_time(struct perf_event_mmap_page *userpg, u64 now)
 {
+	if (!__ktime_get_mono_fast(read_tsc, &userpg->time_zero,
+				   &userpg->time_mult,
+				   &userpg->time_shift))
+		return;
+
 	userpg->cap_user_time = 1;
 	userpg->cap_user_time_zero = 1;
 
-	__ktime_get_mono_fast(&userpg->time_zero,
-			      &userpg->time_mult,
-			      &userpg->time_shift);
-
 	userpg->offset = userpg->time_zero - now;
 }
 
--- a/kernel/time/timekeeping.c
+++ b/kernel/time/timekeeping.c
@@ -334,7 +334,8 @@ u64 notrace ktime_get_mono_fast_ns(void)
 }
 EXPORT_SYMBOL_GPL(ktime_get_mono_fast_ns);
 
-void notrace __ktime_get_mono_fast(u64 *offset, u32 *mult, u16 *shift)
+bool notrace __ktime_get_mono_fast(cycle_t (*read)(struct clocksource *),
+				   u64 *offset, u32 *mult, u16 *shift)
 {
 	struct tk_read_base *tkr;
 	unsigned int seq;
@@ -345,6 +346,9 @@ void notrace __ktime_get_mono_fast(u64 *
 		seq = raw_read_seqcount(&tk_fast_mono.seq);
 		tkr = tk_fast_mono.base + (seq & 0x01);
 
+		if (tkr->read != read)
+			return false;
+
 		cycle_now = tkr->read(tkr->clock);
 		delta = clocksource_delta(cycle_now, tkr->cycle_last, tkr->mask);
 
@@ -362,6 +366,8 @@ void notrace __ktime_get_mono_fast(u64 *
 		*offset = now - nsec;
 
 	} while (read_seqcount_retry(&tk_fast_mono.seq, seq));
+
+	return true;
 }
 
 #ifdef CONFIG_GENERIC_TIME_VSYSCALL_OLD
diff --git a/arch/x86/include/asm/tsc.h b/arch/x86/include/asm/tsc.h
index 94605c0e9cee..68e4039a58ea 100644
--- a/arch/x86/include/asm/tsc.h
+++ b/arch/x86/include/asm/tsc.h
@@ -32,6 +32,8 @@ static inline cycles_t get_cycles(void)
 	return ret;
 }
 
+extern void cycle_t read_tsc(struct clocksource *);
+
 static __always_inline cycles_t vget_cycles(void)
 {
 	/*
diff --git a/arch/x86/kernel/tsc.c b/arch/x86/kernel/tsc.c
index 505449700e0c..c580998f0160 100644
--- a/arch/x86/kernel/tsc.c
+++ b/arch/x86/kernel/tsc.c
@@ -965,7 +965,7 @@ static struct clocksource clocksource_tsc;
  * checking the result of read_tsc() - cycle_last for being negative.
  * That works because CLOCKSOURCE_MASK(64) does not mask out any bit.
  */
-static cycle_t read_tsc(struct clocksource *cs)
+cycle_t read_tsc(struct clocksource *cs)
 {
 	return (cycle_t)get_cycles();
 }