[PATCH -mm 1/6] cpu_relax(): interrupt.h

[PATCH -mm 1/6] cpu_relax(): interrupt.h

[Andreas Mohr]

Add cpu_relax() to tasklet_unlock_wait() loop.


I kept barrier() since it is said to still be required, in various older
postings.
Took care of kernel style formatting.

Tested on 2.6.17-mm1.


This is the first patch within a set, which aims to improve P4 HT performance
a bit:
these SMT CPUs want a rep nop (pause) (as done by cpu_relax()) when
busy-polling, in order to free pipeline resources for the SMT CPU sibling.

During I/O polling cpu_relax() will help for non-P4 CPUs, too (reduce
power consumption there).

Signed-off-by: Andreas Mohr <andi@lisas.de>


diff -urN linux-2.6.17-mm1.orig/include/linux/interrupt.h linux-2.6.17-mm1.my/include/linux/interrupt.h
--- linux-2.6.17-mm1.orig/include/linux/interrupt.h	2006-06-21 14:28:19.000000000 +0200
+++ linux-2.6.17-mm1.my/include/linux/interrupt.h	2006-06-21 21:34:12.000000000 +0200
@@ -227,7 +227,10 @@
 
 static inline void tasklet_unlock_wait(struct tasklet_struct *t)
 {
-	while (test_bit(TASKLET_STATE_RUN, &(t)->state)) { barrier(); }
+	while (test_bit(TASKLET_STATE_RUN, &(t)->state)) {
+		barrier();
+		cpu_relax();
+	}
 }
 #else
 #define tasklet_trylock(t) 1

Re: [PATCH -mm 1/6] cpu_relax(): interrupt.h

[Arjan van de Ven]

On Wed, 2006-06-21 at 22:59 +0200, Andreas Mohr wrote:
> Add cpu_relax() to tasklet_unlock_wait() loop.
> 
> 
> I kept barrier() since it is said to still be required, in various older
> postings.

cpu_relax() includes a barrier always, per definition.
(simply because anywhere you would use cpu_relax() you would also put a
barrier)