Skip to contentSkip to conversation

UNTESTED backport rcu_needs_cpu and call it from stop_hz_timer UNTESTED

PATCH In review Harry Butterworth 3 participants 4 messages Jun 9, 2006 – Jun 12, 2006

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 UNTESTED backport rcu_needs_cpu and call it from stop_hz_timer UNTESTED

THIS PATCH FOR REVIEW ONLY IT'S UNTESTED!!!!!! Ill test on Monday if no-one has done it by then.

There is a problem with the current implementation of stop_hz_timer in arch/i386/kernel/time-xen.c where the hz timer can be stopped on a CPU which has RCU callbacks pending.

This patch backports a new RCU API created to fix this problem for the s390 implementation of stop_hz_timer and also updates the time-xen.c implementation of stop_hz_timer to call the new API.

linux-2.6-xen-sparse/arch/i386/kernel/time-xen.c +9 -2
diff -r 8c64169a05d3 -r 8ef455f268b3 linux-2.6-xen-sparse/arch/i386/kernel/time-xen.c
@@ -978,12 +978,19 @@
 	unsigned int cpu = smp_processor_id();
 	unsigned long j;
 
-	/* We must do this /before/ checking rcu_pending(). */
 	cpu_set(cpu, nohz_cpu_mask);
+
+	/* See matching smp_mb in rcu_start_batch in rcupdate.c.  These mbs  */
+	/* ensure that if __rcu_pending (nested in rcu_needs_cpu) fetches a  */
+	/* value of rcp->cur that matches rdp->quiescbatch and allows us to  */
+	/* stop the hz timer then the cpumasks created for subsequent values */
+	/* of cur in rcu_start_batch are guaranteed to pick up the updated   */
+	/* nohz_cpu_mask and so will not depend on this cpu.                 */
+
 	smp_mb();
 
 	/* Leave ourselves in 'tick mode' if rcu or softirq pending. */
-	if (rcu_pending(cpu) || local_softirq_pending()) {
+	if (rcu_needs_cpu(cpu) || local_softirq_pending()) {
 		cpu_clear(cpu, nohz_cpu_mask);
 		j = jiffies + 1;
 	} else {
patches/linux-2.6.16.13/rcu_needs_cpu.patch new file +33 -0
diff -r 8c64169a05d3 -r 8ef455f268b3 patches/linux-2.6.16.13/rcu_needs_cpu.patch
@@ -0,0 +1,33 @@
+--- ../pristine-linux-2.6.16.13/kernel/rcupdate.c	2006-05-02 22:38:44.000000000 +0100
++++ ./kernel/rcupdate.c	2006-06-09 20:27:45.000000000 +0100
+@@ -485,6 +485,20 @@ int rcu_pending(int cpu)
+ 		__rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
+ }
+ 
++/*
++ * Check to see if any future RCU-related work will need to be done
++ * by the current CPU, even if none need be done immediately, returning
++ * 1 if so.  This function is part of the RCU implementation; it is -not-
++ * an exported member of the RCU API.
++ */
++int rcu_needs_cpu(int cpu)
++{
++	struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
++	struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
++
++	return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
++}
++
+ void rcu_check_callbacks(int cpu, int user)
+ {
+ 	if (user || 
+--- ../pristine-linux-2.6.16.13/include/linux/rcupdate.h	2006-05-02 22:38:44.000000000 +0100
++++ ./include/linux/rcupdate.h	2006-06-09 20:28:57.000000000 +0100
+@@ -134,6 +134,7 @@ static inline void rcu_bh_qsctr_inc(int 
+ }
+ 
+ extern int rcu_pending(int cpu);
++extern int rcu_needs_cpu(int cpu);
+ 
+ /**
+  * rcu_read_lock - mark the beginning of an RCU read-side critical section.

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#1 lore raw
Quoted 81 lines · Harry Butterworth

THIS PATCH FOR REVIEW ONLY IT'S UNTESTED!!!!!! Ill test on Monday if no-one has done it by then.

There is a problem with the current implementation of stop_hz_timer in arch/i386/kernel/time-xen.c where the hz timer can be stopped on a CPU which has RCU callbacks pending.

This patch backports a new RCU API created to fix this problem for the s390 implementation of stop_hz_timer and also updates the time-xen.c implementation of stop_hz_timer to call the new API.

------------------------------------------------------------------------

diff -r 8c64169a05d3 -r 8ef455f268b3 linux-2.6-xen-sparse/arch/i386/kernel/time-xen.c

linux-2.6-xen-sparse/arch/i386/kernel/time-xen.c +9 -2
@@ -978,12 +978,19 @@
 	unsigned int cpu = smp_processor_id();
 	unsigned long j;
 
-	/* We must do this /before/ checking rcu_pending(). */
 	cpu_set(cpu, nohz_cpu_mask);
+
+	/* See matching smp_mb in rcu_start_batch in rcupdate.c.  These mbs  */
+	/* ensure that if __rcu_pending (nested in rcu_needs_cpu) fetches a  */
+	/* value of rcp->cur that matches rdp->quiescbatch and allows us to  */
+	/* stop the hz timer then the cpumasks created for subsequent values */
+	/* of cur in rcu_start_batch are guaranteed to pick up the updated   */
+	/* nohz_cpu_mask and so will not depend on this cpu.                 */
+
 	smp_mb();
 
 	/* Leave ourselves in 'tick mode' if rcu or softirq pending. */
-	if (rcu_pending(cpu) || local_softirq_pending()) {
+	if (rcu_needs_cpu(cpu) || local_softirq_pending()) {
 		cpu_clear(cpu, nohz_cpu_mask);
 		j = jiffies + 1;
 	} else {

diff -r 8c64169a05d3 -r 8ef455f268b3 patches/linux-2.6.16.13/rcu_needs_cpu.patch

patches/linux-2.6.16.13/rcu_needs_cpu.patch new file +33 -1
@@ -0,0 +1,33 @@
+--- ../pristine-linux-2.6.16.13/kernel/rcupdate.c	2006-05-02 22:38:44.000000000 +0100
++++ ./kernel/rcupdate.c	2006-06-09 20:27:45.000000000 +0100
+@@ -485,6 +485,20 @@ int rcu_pending(int cpu)
+ 		__rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
+ }
+ 
++/*
++ * Check to see if any future RCU-related work will need to be done
++ * by the current CPU, even if none need be done immediately, returning
++ * 1 if so.  This function is part of the RCU implementation; it is -not-
++ * an exported member of the RCU API.
++ */
++int rcu_needs_cpu(int cpu)
++{
++	struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
++	struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
++
++	return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
++}
++
+ void rcu_check_callbacks(int cpu, int user)
+ {
+ 	if (user || 
+--- ../pristine-linux-2.6.16.13/include/linux/rcupdate.h	2006-05-02 22:38:44.000000000 +0100
++++ ./include/linux/rcupdate.h	2006-06-09 20:28:57.000000000 +0100
+@@ -134,6 +134,7 @@ static inline void rcu_bh_qsctr_inc(int 
+ }
+ 
+ extern int rcu_pending(int cpu);
++extern int rcu_needs_cpu(int cpu);
+ 
+ /**
+  * rcu_read_lock - mark the beginning of an RCU read-side critical section.
 
------------------------------------------------------------------------
Quoted 5 lines

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

Harry,

I am doing some quick testing of your patch. So far, the results are consistent with the mainline testing results.

Signature · 9 lines
Regards,

David F Barrera
Linux Technology Center
Systems and Technology Group, IBM

"The wisest men follow their own direction. "
	
                         Euripides
#2 lore raw

PATCH 1/1 backport rcu_needs_cpu and call it from stop_hz_timer

In reply to Harry Butterworth

I have now tested this against xen-testing 9735. Without the patch, and with a domU kernel configured with CONFIG_DEBUG_SOFTLOCKUP disabled (this debug code masks the problem), and smp and cpu hotplug enabled, a script which does lots of cpu hotplug/removal will hang (cpu hotplug/removal stresses the RCU synchronize_sched function). With the patch, the script doing cpu hotplug/removal runs for as long as I have let it so far (thousands of hotplugs).

Please apply.

Quoted 8 lines · Harry Butterworth

There is a problem with the current implementation of stop_hz_timer in arch/i386/kernel/time-xen.c where the hz timer can be stopped on a CPU which has RCU callbacks pending.

This patch backports a new RCU API created to fix this problem for the s390 implementation of stop_hz_timer and also updates the time-xen.c implementation of stop_hz_timer to call the new API.

Quoted 6 lines

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

linux-2.6-xen-sparse/arch/i386/kernel/time-xen.c +9 -2
diff -r 8c64169a05d3 -r 8ef455f268b3 linux-2.6-xen-sparse/arch/i386/kernel/time-xen.c
@@ -978,12 +978,19 @@
 	unsigned int cpu = smp_processor_id();
 	unsigned long j;
 
-	/* We must do this /before/ checking rcu_pending(). */
 	cpu_set(cpu, nohz_cpu_mask);
+
+	/* See matching smp_mb in rcu_start_batch in rcupdate.c.  These mbs  */
+	/* ensure that if __rcu_pending (nested in rcu_needs_cpu) fetches a  */
+	/* value of rcp->cur that matches rdp->quiescbatch and allows us to  */
+	/* stop the hz timer then the cpumasks created for subsequent values */
+	/* of cur in rcu_start_batch are guaranteed to pick up the updated   */
+	/* nohz_cpu_mask and so will not depend on this cpu.                 */
+
 	smp_mb();
 
 	/* Leave ourselves in 'tick mode' if rcu or softirq pending. */
-	if (rcu_pending(cpu) || local_softirq_pending()) {
+	if (rcu_needs_cpu(cpu) || local_softirq_pending()) {
 		cpu_clear(cpu, nohz_cpu_mask);
 		j = jiffies + 1;
 	} else {
patches/linux-2.6.16.13/rcu_needs_cpu.patch new file +33 -0
diff -r 8c64169a05d3 -r 8ef455f268b3 patches/linux-2.6.16.13/rcu_needs_cpu.patch
@@ -0,0 +1,33 @@
+--- ../pristine-linux-2.6.16.13/kernel/rcupdate.c	2006-05-02 22:38:44.000000000 +0100
++++ ./kernel/rcupdate.c	2006-06-09 20:27:45.000000000 +0100
+@@ -485,6 +485,20 @@ int rcu_pending(int cpu)
+ 		__rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu));
+ }
+ 
++/*
++ * Check to see if any future RCU-related work will need to be done
++ * by the current CPU, even if none need be done immediately, returning
++ * 1 if so.  This function is part of the RCU implementation; it is -not-
++ * an exported member of the RCU API.
++ */
++int rcu_needs_cpu(int cpu)
++{
++	struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
++	struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu);
++
++	return (!!rdp->curlist || !!rdp_bh->curlist || rcu_pending(cpu));
++}
++
+ void rcu_check_callbacks(int cpu, int user)
+ {
+ 	if (user || 
+--- ../pristine-linux-2.6.16.13/include/linux/rcupdate.h	2006-05-02 22:38:44.000000000 +0100
++++ ./include/linux/rcupdate.h	2006-06-09 20:28:57.000000000 +0100
+@@ -134,6 +134,7 @@ static inline void rcu_bh_qsctr_inc(int 
+ }
+ 
+ extern int rcu_pending(int cpu);
++extern int rcu_needs_cpu(int cpu);
+ 
+ /**
+  * rcu_read_lock - mark the beginning of an RCU read-side critical section.

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#3 lore raw

Re: [PATCH] backport rcu_needs_cpu and call it from stop_hz_timer

Quoted 7 lines · Harry Butterworth

I have now tested this against xen-testing 9735. Without the patch, and with a domU kernel configured with CONFIG_DEBUG_SOFTLOCKUP disabled (this debug code masks the problem), and smp and cpu hotplug enabled, a script which does lots of cpu hotplug/removal will hang (cpu hotplug/removal stresses the RCU synchronize_sched function). With the patch, the script doing cpu hotplug/removal runs for as long as I have let it so far (thousands of hotplugs).

Okay, the patch looks good to me. I'll apply to our trees. Thanks.

-- Keir