Re: [PATCH] optimized ktime_get[_ts] for GENERIC_TIME=y
From: Martin Schwidefsky
Date: Tue Jul 07 2009 - 05:27:44 EST
On Tue, 7 Jul 2009 10:08:56 +0200 (CEST)
Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> Martin,
>
> On Tue, 7 Jul 2009, Martin Schwidefsky wrote:
> > On Mon, 6 Jul 2009 22:31:39 +0200 (CEST)
> > Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >
> > > On Mon, 6 Jul 2009, Martin Schwidefsky wrote:
> > > > +ktime_t ktime_get(void)
> > > > +{
> > > > + cycle_t cycle_now, cycle_delta;
> > > > + struct timespec time;
> > > > + unsigned long seq;
> > > > + s64 nsecs;
> > > > +
> > > > + do {
> > > > + seq = read_seqbegin(&xtime_lock);
> > > > + time.tv_sec = xtime.tv_sec + wall_to_monotonic.tv_sec;
> > > > + time.tv_nsec = xtime.tv_nsec + wall_to_monotonic.tv_nsec;
> > >
> > > That's actually a violation of the timespec semantics. tv_nsec can end
> > > up greater than (10^9 - 1). Please use separate sec and nsec variables.
> >
> > Well the tv_sec/tv_nsec fields of a timespec are long values. But its
> > no problem to switch to local variables.
>
> Right. I'm not worried about an overflow of the variable. I was about
> to suggest using timespec_to_ktime() to fix the !SCALAR problem when I
> noticed that the timespec is possibly not normalized. Not using a
> timespec makes the code more obvious I think.
That makes sense. New patch:
--
Subject: [PATCH] optimized ktime_get[_ts] for GENERIC_TIME=y
From: Martin Schwidefsky <schwidefsky@xxxxxxxxxx>
The generic ktime_get function defined in kernel/hrtimer.c is suboptimial
for GENERIC_TIME=y:
0) | ktime_get() {
0) | ktime_get_ts() {
0) | getnstimeofday() {
0) | read_tod_clock() {
0) 0.601 us | }
0) 1.938 us | }
0) | set_normalized_timespec() {
0) 0.602 us | }
0) 4.375 us | }
0) 5.523 us | }
Overall there are two read_seqbegin/read_seqretry loops and a lot of
unnecessary struct timespec calculations. ktime_get returns a nano second
value which is the sum of xtime, wall_to_monotonic and the nano second
delta from the clock source.
ktime_get can be optimized for GENERIC_TIME=y. The new version only calls
clocksource_read:
0) | ktime_get() {
0) | read_tod_clock() {
0) 0.610 us | }
0) 1.977 us | }
It uses a single read_seqbegin/readseqretry loop and just adds everthing
to a nano second value.
ktime_get_ts is optimized in a similar fashion.
Cc: Ingo Molnar <mingo@xxxxxxx>
Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
Acked-by: john stultz <johnstul@xxxxxxxxxx>
Signed-off-by: Martin Schwidefsky <schwidefsky@xxxxxxxxxx>
---
kernel/hrtimer.c | 4 ++
kernel/time/timekeeping.c | 65 ++++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 69 insertions(+)
diff -urpN linux-2.6/kernel/hrtimer.c linux-2.6-patched/kernel/hrtimer.c
--- linux-2.6/kernel/hrtimer.c 2009-07-07 11:24:29.000000000 +0200
+++ linux-2.6-patched/kernel/hrtimer.c 2009-07-07 11:24:37.000000000 +0200
@@ -48,6 +48,7 @@
#include <asm/uaccess.h>
+#ifndef CONFIG_GENERIC_TIME
/**
* ktime_get - get the monotonic time in ktime_t format
*
@@ -62,6 +63,7 @@ ktime_t ktime_get(void)
return timespec_to_ktime(now);
}
EXPORT_SYMBOL_GPL(ktime_get);
+#endif
/**
* ktime_get_real - get the real (wall-) time in ktime_t format
@@ -106,6 +108,7 @@ DEFINE_PER_CPU(struct hrtimer_cpu_base,
}
};
+#ifndef CONFIG_GENERIC_TIME
/**
* ktime_get_ts - get the monotonic clock in timespec format
* @ts: pointer to timespec variable
@@ -130,6 +133,7 @@ void ktime_get_ts(struct timespec *ts)
ts->tv_nsec + tomono.tv_nsec);
}
EXPORT_SYMBOL_GPL(ktime_get_ts);
+#endif
/*
* Get the coarse grained time at the softirq based on xtime and
diff -urpN linux-2.6/kernel/time/timekeeping.c linux-2.6-patched/kernel/time/timekeeping.c
--- linux-2.6/kernel/time/timekeeping.c 2009-07-07 11:24:29.000000000 +0200
+++ linux-2.6-patched/kernel/time/timekeeping.c 2009-07-07 11:24:37.000000000 +0200
@@ -125,6 +125,71 @@ void getnstimeofday(struct timespec *ts)
EXPORT_SYMBOL(getnstimeofday);
+ktime_t ktime_get(void)
+{
+ cycle_t cycle_now, cycle_delta;
+ unsigned int seq;
+ s64 secs, nsecs;
+
+ do {
+ seq = read_seqbegin(&xtime_lock);
+ secs = xtime.tv_sec + wall_to_monotonic.tv_sec;
+ nsecs = xtime.tv_nsec + wall_to_monotonic.tv_nsec;
+
+ /* read clocksource: */
+ cycle_now = clocksource_read(clock);
+
+ /* calculate the delta since the last update_wall_time: */
+ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
+
+ /* convert to nanoseconds: */
+ nsecs += cyc2ns(clock, cycle_delta);
+
+ } while (read_seqretry(&xtime_lock, seq));
+ /*
+ * Use ktime_set/ktime_add_ns to create a proper ktime on
+ * 32-bit architectures without CONFIG_KTIME_SCALAR.
+ */
+ return ktime_add_ns(ktime_set(secs, 0), nsecs);
+}
+EXPORT_SYMBOL_GPL(ktime_get);
+
+/**
+ * ktime_get_ts - get the monotonic clock in timespec format
+ * @ts: pointer to timespec variable
+ *
+ * The function calculates the monotonic clock from the realtime
+ * clock and the wall_to_monotonic offset and stores the result
+ * in normalized timespec format in the variable pointed to by @ts.
+ */
+void ktime_get_ts(struct timespec *ts)
+{
+ cycle_t cycle_now, cycle_delta;
+ struct timespec tomono;
+ unsigned int seq;
+ s64 nsecs;
+
+ do {
+ seq = read_seqbegin(&xtime_lock);
+ *ts = xtime;
+ tomono = wall_to_monotonic;
+
+ /* read clocksource: */
+ cycle_now = clocksource_read(clock);
+
+ /* calculate the delta since the last update_wall_time: */
+ cycle_delta = (cycle_now - clock->cycle_last) & clock->mask;
+
+ /* convert to nanoseconds: */
+ nsecs = cyc2ns(clock, cycle_delta);
+
+ } while (read_seqretry(&xtime_lock, seq));
+
+ set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
+ ts->tv_nsec + tomono.tv_nsec + nsecs);
+}
+EXPORT_SYMBOL_GPL(ktime_get_ts);
+
/**
* do_gettimeofday - Returns the time of day in a timeval
* @tv: pointer to the timeval to be set
--
blue skies,
Martin.
"Reality continues to ruin my life." - Calvin.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/