Re: long delays (possibly infinite) intime_interpolator_get_counter

[Alex Williamson]

   Ok, here's an optimization that should help reduce contention on the
cmpxchg, has zero impact on the nojitter path, and doesn't require any
changes to fsys.  When a caller already had the xtime_lock write lock
there's no need to fight with other CPUs for the cmpxchg.  The other
"reader" CPUs will have to fetch it again since a seqlock write is in
progress.  Therefore we can simplify this path as shown below.  The
write is atomic, and we don't care if another CPU has changed last_cycle
since it can't return the value until the write lock is released.  This
has only been compile tested, but I'm interested to hear your opinion.
Thanks,

	Alex


diff -r cff8d3633e9c kernel/timer.c
--- a/kernel/timer.c	Fri Jul 29 22:01:15 2005
+++ b/kernel/timer.c	Sun Jul 31 10:25:41 2005
@@ -1452,10 +1452,34 @@
 		return time_interpolator_get_cycles(src);
 }
 
+static inline u64 time_interpolator_get_counter_locked(void)
+{
+	unsigned int src = time_interpolator->source;
+
+	if (time_interpolator->jitter)
+	{
+		u64 lcycle = time_interpolator->last_cycle;
+		u64 now = time_interpolator_get_cycles(src);
+
+		if (lcycle && time_after(lcycle, now))
+			return lcycle;
+
+		/*
+		 * This path is called when holding the xtime write lock.
+		 * This allows us to avoid the contention of the cmpxchg
+		 * in get_counter, and still ensures jitter protection.
+		 */
+		time_interpolator->last_cycle = now;
+		return now;
+	}
+	else
+		return time_interpolator_get_cycles(src);
+}
+
 void time_interpolator_reset(void)
 {
 	time_interpolator->offset = 0;
-	time_interpolator->last_counter = time_interpolator_get_counter();
+	time_interpolator->last_counter = time_interpolator_get_counter_locked();
 }
 
 #define GET_TI_NSECS(count,i) (((((count) - i->last_counter) & (i)->mask) * (i)->nsec_per_cyc) >> (i)->shift)
@@ -1490,7 +1514,7 @@
 	 * and the tuning logic insures that.
          */
 
-	counter = time_interpolator_get_counter();
+	counter = time_interpolator_get_counter_locked();
 	offset = time_interpolator->offset + GET_TI_NSECS(counter, time_interpolator);
 
 	if (delta_nsec < 0 || (unsigned long) delta_nsec < offset)


-
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/