Skip to contentSkip to conversation

timers: add_timer should never be called if pending

RFC PATCH In review Krishna Kumar 2 participants 3 messages Sep 22, 2008 High-resolution timers, timer wheel, clockevents

lore
Share & tools
lore
#0 lore raw

cover timers: add_timer should never be called if pending

From: Krishna Kumar <[email protected]>

(Hopefully this is not something that has been suggested earlier and rejected).

add_timer should never be called on a pending timer - such bugs should be caught by reports generated by debug kernels. It is the responsibility of the subsystem users to ensure that timers are not added twice. Using the add_timer_on interface reduces the code that checks if the timer is pending, the timer_base change, updating 'expires' two times, etc.

1. Single CPU, Single Timer: Add the same timer serially on one CPU - do this many times.

	Laptop (2 way Xeon): Saves 10.8%
		ORG: Time: 35359
		NEW: Time: 31527
	Server (4 way x86-64): Saves 7.9%
		ORG: Time: 4520
		NEW: Time: 4164
2. Single CPU, Multiple Timer: Add different timers serially on one CPU - do
   this many times.
	Laptop (2 way Xeon): Saves 7.7%
		ORG: Time: 133728
		NEW: Time: 144822
	Server (4 way x86-64): Saves 15.7%
		ORG: Time: 69012
		NEW: Time: 58186
3. Many CPU's, Single Timer: Add the same timer in parallel on all CPUs - do
   this many times.
	Laptop (2 way Xeon): Saves 14%
		ORG: Time: 69845
		NEW: Time: 60067
	Server (4 way x86-64): Saves 21.8%
		ORG: Time: 18047
		NEW: Time: 14116
4. Many CPU's, Multiple Timer: Add different timers in parallel on all CPUs -
   do this many times.
	Laptop (2 way Xeon): Saves 47.1%
		ORG: Time: 292173
		NEW: Time: 154485
	Server (4 way x86-64): Saves 8.23
		ORG: Time: 319129
		NEW: Time: 292842
#1 lore raw

RFC PATCH 1/1 : timers: Change add_timer to use add_timer_on

From: Krishna Kumar <[email protected]>

timer: Change add_timer to not check for existing timer.

---

 include/linux/timer.h |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
include/linux/timer.h +2 -2
diff -ruNp 2.6.27-rc7-org/include/linux/timer.h 2.6.27-rc7-new/include/linux/timer.h
@@ -4,6 +4,7 @@
 #include <linux/list.h>
 #include <linux/ktime.h>
 #include <linux/stddef.h>
+#include <linux/smp.h>
 #include <linux/debugobjects.h>
 
 struct tvec_base;
@@ -162,8 +163,7 @@ static inline void timer_stats_timer_cle
  */
 static inline void add_timer(struct timer_list *timer)
 {
-	BUG_ON(timer_pending(timer));
-	__mod_timer(timer, timer->expires);
+	add_timer_on(timer, raw_smp_processor_id());
 }
 
 #ifdef CONFIG_SMP
#2 lore raw

Re: [RFC] [PATCH 1/1]: timers: Change add_timer to use add_timer_on

(add Thomas)

Quoted 2 lines · Krishna Kumar

timer: Change add_timer to not check for existing timer.

Perhaps the changelog should explain why ;)

But, besides the removing BUG_ON(), the patch does something else.

include/linux/timer.h static inline void timer_stats_timer_cle · line 167

Quoted 6 lines

static inline void add_timer(struct timer_list *timer) {

include/linux/timer.h +1 -2
-	BUG_ON(timer_pending(timer));
-	__mod_timer(timer, timer->expires);
+	add_timer_on(timer, raw_smp_processor_id());
 }

This is racy wrt CPU hotplug. The CPU can be cpu_down()'ed right after raw_smp_processor_id() returns, and before we actually add the timer.

Another problem is that now it is add_timer() becomes unsafe wrt concurrent del_timer/del_timer_sync, because add_timer_on() does not lock the timer, it only locks the target base.

Oleg.