Skip to contentSkip to conversation

2.6.25+: Fix cpu hotplug in softirq code

PATCH In review Christian Borntraeger 3 participants 4 messages Apr 24, 2008 – Apr 29, 2008

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 2.6.25+: Fix cpu hotplug in softirq code

Hello Olof,

currently cpu hotplug (unplug) seems broken on s390 and likely others. On cpu unplug the system starts to behave very strange and hangs.

I bisected the problem to the following commit:

-----
commit 48f20a9a9488c432fc86df1ff4b7f4fa895d1183

Author: Olof Johansson <[email protected]>
Date:   Tue Mar 4 15:23:25 2008 -0800

tasklets: execute tasklets in the same order they were queued
-----

Reverting this patch seems to fix the problem. I looked into takeover_tasklet and it seems that there is a way to corrupt the tail pointer of the current cpu. If the tasklet list of the frozen cpu is empty, the tail pointer of the current cpu points to the address of the head pointer of the stopped cpu and not to the next pointer of a tasklet_struct.

This patch avoids the list splice of the list is empty and cpu hotplug seems to work as the tail pointer is not corrupted.
Olof, can you look into that patch and ACK/NACK it so Andrew can push this to Linus, if appropriate?
Please note that some lines are longer than 80 chars, but line-wrapping looked worse that this version.

---
 kernel/softirq.c |   20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)
kernel/softirq.c +12 -8
Index: kvm/kernel/softirq.c
===================================================================
@@ -589,16 +589,20 @@ static void takeover_tasklets(unsigned i
 	local_irq_disable();
 
 	/* Find end, append list for that CPU. */
-	*__get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).head;
-	__get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).tail;
-	per_cpu(tasklet_vec, cpu).head = NULL;
-	per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
+	if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
+		*(__get_cpu_var(tasklet_vec).tail) = per_cpu(tasklet_vec, cpu).head;
+		__get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).tail;
+		per_cpu(tasklet_vec, cpu).head = NULL;
+		per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
+	}
 	raise_softirq_irqoff(TASKLET_SOFTIRQ);
 
-	*__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).head;
-	__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).tail;
-	per_cpu(tasklet_hi_vec, cpu).head = NULL;
-	per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
+	if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
+		*__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).head;
+		__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).tail;
+		per_cpu(tasklet_hi_vec, cpu).head = NULL;
+		per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
+	}
 	raise_softirq_irqoff(HI_SOFTIRQ);
 
 	local_irq_enable();
#1 lore raw

[PATCH/resend] 2.6.25+: Fix cpu hotplug problem in softirq code

This is a resend of a patch, which fixes a bug in cpu hotplug introduced after 2.6.25. Andrew, Olof, any opinions on this patch?

Christian

--- old mail ---

Hello Olof,

currently cpu hotplug (unplug) seems broken on s390 and likely others. On cpu unplug the system starts to behave very strange and hangs.

I bisected the problem to the following commit:

-----
commit 48f20a9a9488c432fc86df1ff4b7f4fa895d1183
Author: Olof Johansson <[email protected]>
Date:   Tue Mar 4 15:23:25 2008 -0800
    tasklets: execute tasklets in the same order they were queued
-----

Reverting this patch seems to fix the problem. I looked into takeover_tasklet and it seems that there is a way to corrupt the tail pointer of the current cpu. If the tasklet list of the frozen cpu is empty, the tail pointer of the current cpu points to the address of the head pointer of the stopped cpu and not to the next pointer of a tasklet_struct.

This patch avoids the list splice of the list is empty and cpu hotplug seems to work as the tail pointer is not corrupted.
Olof, can you look into that patch and ACK/NACK it so Andrew can push this to Linus, if appropriate?
Please note that some lines are longer than 80 chars, but line-wrapping looked worse that this version.

---
 kernel/softirq.c |   20 ++++++++++++--------
 1 file changed, 12 insertions(+), 8 deletions(-)
kernel/softirq.c +12 -8
Index: kvm/kernel/softirq.c
===================================================================
@@ -589,16 +589,20 @@ static void takeover_tasklets(unsigned i
 	local_irq_disable();
 
 	/* Find end, append list for that CPU. */
-	*__get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).head;
-	__get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).tail;
-	per_cpu(tasklet_vec, cpu).head = NULL;
-	per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
+	if (&per_cpu(tasklet_vec, cpu).head != per_cpu(tasklet_vec, cpu).tail) {
+		*(__get_cpu_var(tasklet_vec).tail) = per_cpu(tasklet_vec, cpu).head;
+		__get_cpu_var(tasklet_vec).tail = per_cpu(tasklet_vec, cpu).tail;
+		per_cpu(tasklet_vec, cpu).head = NULL;
+		per_cpu(tasklet_vec, cpu).tail = &per_cpu(tasklet_vec, cpu).head;
+	}
 	raise_softirq_irqoff(TASKLET_SOFTIRQ);
 
-	*__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).head;
-	__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).tail;
-	per_cpu(tasklet_hi_vec, cpu).head = NULL;
-	per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
+	if (&per_cpu(tasklet_hi_vec, cpu).head != per_cpu(tasklet_hi_vec, cpu).tail) {
+		*__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).head;
+		__get_cpu_var(tasklet_hi_vec).tail = per_cpu(tasklet_hi_vec, cpu).tail;
+		per_cpu(tasklet_hi_vec, cpu).head = NULL;
+		per_cpu(tasklet_hi_vec, cpu).tail = &per_cpu(tasklet_hi_vec, cpu).head;
+	}
 	raise_softirq_irqoff(HI_SOFTIRQ);
 
 	local_irq_enable();
Acked-by
#2 lore raw

Re: [PATCH/resend] 2.6.25+: Fix cpu hotplug problem in softirq code

Quoted 2 lines · Christian Borntraeger

This is a resend of a patch, which fixes a bug in cpu hotplug introduced after 2.6.25. Andrew, Olof, any opinions on this patch?

hm, nice catch.

Ingo

Acked-by on v1 1/1
#3 lore raw
In reply to Christian Borntraeger
Quoted 20 lines · Christian Borntraeger

Hello Olof,

currently cpu hotplug (unplug) seems broken on s390 and likely others. On cpu unplug the system starts to behave very strange and hangs.

I bisected the problem to the following commit:

-----
commit 48f20a9a9488c432fc86df1ff4b7f4fa895d1183

Author: Olof Johansson <[email protected]>
Date:   Tue Mar 4 15:23:25 2008 -0800

tasklets: execute tasklets in the same order they were queued
-----

Reverting this patch seems to fix the problem. I looked into takeover_tasklet and it seems that there is a way to corrupt the tail pointer of the current cpu. If the tasklet list of the frozen cpu is empty, the tail pointer of the current cpu points to the address of the head pointer of the stopped cpu and not to the next pointer of a tasklet_struct.

Quoted 8 lines

This patch avoids the list splice of the list is empty and cpu hotplug seems to work as the tail pointer is not corrupted.
Olof, can you look into that patch and ACK/NACK it so Andrew can push this to Linus, if appropriate?
Please note that some lines are longer than 80 chars, but line-wrapping looked worse that this version.

I don't have a hotplug-capable system to test on, but the patch looks good to me. Good catch.

-Olof