Skip to contentSkip to conversation

x86: only call smp_processor_id in non-preempt cases

PATCH In review Don Zickus 4 participants 9 messages Nov 12, 2010 – Nov 19, 2010 x86

lore
Share & tools
lore

Don Zickus

author
#0 lore raw

PATCH 1/3 x86: only call smp_processor_id in non-preempt cases

There are some paths that walk the die_chain with preemption on. Make sure we are in an NMI call before we start doing anything.

---
 arch/x86/kernel/apic/hw_nmi.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)
arch/x86/kernel/apic/hw_nmi.c +2 -1
index 5c4f952..ef4755d 100644
@@ -49,7 +49,7 @@ arch_trigger_all_cpu_backtrace_handler(struct notifier_block *self,
 {
 	struct die_args *args = __args;
 	struct pt_regs *regs;
-	int cpu = smp_processor_id();
+	int cpu;
 
 	switch (cmd) {
 	case DIE_NMI:
@@ -60,6 +60,7 @@ arch_trigger_all_cpu_backtrace_handler(struct notifier_block *self,
 	}
 
 	regs = args->regs;
+	cpu = smp_processor_id();
 
 	if (cpumask_test_cpu(cpu, to_cpumask(backtrace_mask))) {
 		static arch_spinlock_t lock = __ARCH_SPIN_LOCK_UNLOCKED;
Signature · 1 line
1.7.2.3
author
#1 lore raw

PATCH 2/3 x86, hw_nmi: Move backtrace_mask declaration under ARCH_HAS_NMI_WATCHDOG.

From: Rakib Mullick <[email protected]>

backtrace_mask has been used under the code context of ARCH_HAS_NMI_WATCHDOG. So put it into that context. We were warned by the following warning:

arch/x86/kernel/apic/hw_nmi.c:21: warning: ‘backtrace_mask’ defined but not used

---
 arch/x86/kernel/apic/hw_nmi.c |    7 ++++---
 1 files changed, 4 insertions(+), 3 deletions(-)
arch/x86/kernel/apic/hw_nmi.c +4 -3
index ef4755d..f349647 100644
@@ -17,15 +17,16 @@
 #include <linux/nmi.h>
 #include <linux/module.h>
 
-/* For reliability, we're prepared to waste bits here. */
-static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
-
 u64 hw_nmi_get_sample_period(void)
 {
 	return (u64)(cpu_khz) * 1000 * 60;
 }
 
 #ifdef ARCH_HAS_NMI_WATCHDOG
+
+/* For reliability, we're prepared to waste bits here. */
+static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
+
 void arch_trigger_all_cpu_backtrace(void)
 {
 	int i;
Signature · 1 line
1.7.2.3
author
#2 lore raw

PATCH 3/3 x86: Avoid calling arch_trigger_all_cpu_backtrace() at the same time

In reply to Don Zickus

From: Dongdong Deng <[email protected]>

The spin_lock_debug/rcu_cpu_stall detector uses trigger_all_cpu_backtrace() to dump cpu backtrace. Therefore it is possible that trigger_all_cpu_backtrace() could be called at the same time on different CPUs, which triggers and 'unknown reason NMI' warning. The following case illustrates the problem:

CPU1 CPU2 ... CPU N
trigger_all_cpu_backtrace() set "backtrace_mask" to cpu mask
|
generate NMI interrupts generate NMI interrupts ...
\ | /
\ | /

The "backtrace_mask" will be cleaned by the first NMI interrupt at nmi_watchdog_tick(), then the following NMI interrupts generated by other cpus's arch_trigger_all_cpu_backtrace() will be taken as unknown reason NMI interrupts.

This patch uses a test_and_set to avoid the problem, and stop the arch_trigger_all_cpu_backtrace() from calling to avoid dumping a double cpu backtrace info when there is already a trigger_all_cpu_backtrace() in progress.

---
 arch/x86/kernel/apic/hw_nmi.c |   24 ++++++++++++++++++++++++
 1 files changed, 24 insertions(+), 0 deletions(-)
arch/x86/kernel/apic/hw_nmi.c +24 -0
index f349647..d892896 100644
@@ -27,9 +27,27 @@ u64 hw_nmi_get_sample_period(void)
 /* For reliability, we're prepared to waste bits here. */
 static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
 
+/* "in progress" flag of arch_trigger_all_cpu_backtrace */
+static unsigned long backtrace_flag;
+
 void arch_trigger_all_cpu_backtrace(void)
 {
 	int i;
+	unsigned long flags;
+
+	/*
+	 * Have to disable irq here, as the
+	 * arch_trigger_all_cpu_backtrace() could be
+	 * triggered by "spin_lock()" with irqs on.
+	 */
+	local_irq_save(flags);
+
+	if (test_and_set_bit(0, &backtrace_flag))
+		/*
+		 * If there is already a trigger_all_cpu_backtrace() in progress
+		 * (backtrace_flag == 1), don't output double cpu dump infos.
+		 */
+		goto out_restore_irq;
 
 	cpumask_copy(to_cpumask(backtrace_mask), cpu_online_mask);
 
@@ -42,6 +60,12 @@ void arch_trigger_all_cpu_backtrace(void)
 			break;
 		mdelay(1);
 	}
+
+	clear_bit(0, &backtrace_flag);
+	smp_mb__after_clear_bit();
+
+out_restore_irq:
+	local_irq_restore(flags);
 }
 
 static int __kprobes
Signature · 1 line
1.7.2.3
#3 lore raw

Re: [PATCH 3/3] x86: Avoid calling arch_trigger_all_cpu_backtrace() at the same time

Quoted 51 lines · Don Zickus

From: Dongdong Deng <[email protected]>

The spin_lock_debug/rcu_cpu_stall detector uses trigger_all_cpu_backtrace() to dump cpu backtrace. Therefore it is possible that trigger_all_cpu_backtrace() could be called at the same time on different CPUs, which triggers and 'unknown reason NMI' warning. The following case illustrates the problem:

CPU1 CPU2 ... CPU N
trigger_all_cpu_backtrace() set "backtrace_mask" to cpu mask
|
generate NMI interrupts generate NMI interrupts ...
\ | /
\ | /

The "backtrace_mask" will be cleaned by the first NMI interrupt at nmi_watchdog_tick(), then the following NMI interrupts generated by other cpus's arch_trigger_all_cpu_backtrace() will be taken as unknown reason NMI interrupts.

This patch uses a test_and_set to avoid the problem, and stop the arch_trigger_all_cpu_backtrace() from calling to avoid dumping a double cpu backtrace info when there is already a trigger_all_cpu_backtrace() in progress.

---
 arch/x86/kernel/apic/hw_nmi.c |   24 ++++++++++++++++++++++++
 1 files changed, 24 insertions(+), 0 deletions(-)
arch/x86/kernel/apic/hw_nmi.c +3 -0
index f349647..d892896 100644
@@ -27,9 +27,27 @@ u64 hw_nmi_get_sample_period(void)
 /* For reliability, we're prepared to waste bits here. */
 static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
 
+/* "in progress" flag of arch_trigger_all_cpu_backtrace */
+static unsigned long backtrace_flag;
+
 void arch_trigger_all_cpu_backtrace(void)

arch/x86/kernel/apic/hw_nmi.c u64 hw_nmi_get_sample_period(void) · line 43

Quoted 10 lines
arch/x86/kernel/apic/hw_nmi.c +8 -0
 {
 	int i;
+	unsigned long flags;
+
+	/*
+	 * Have to disable irq here, as the
+	 * arch_trigger_all_cpu_backtrace() could be
+	 * triggered by "spin_lock()" with irqs on.
+	 */
+	local_irq_save(flags);

I'm not sure I understand why you disable irqs here. It looks safe with the test_and_set_bit already.

Quoted 31 lines
arch/x86/kernel/apic/hw_nmi.c +13 -0
+
+	if (test_and_set_bit(0, &backtrace_flag))
+		/*
+		 * If there is already a trigger_all_cpu_backtrace() in progress
+		 * (backtrace_flag == 1), don't output double cpu dump infos.
+		 */
+		goto out_restore_irq;
 
 	cpumask_copy(to_cpumask(backtrace_mask), cpu_online_mask);
 
@@ -42,6 +60,12 @@ void arch_trigger_all_cpu_backtrace(void)
 			break;
 		mdelay(1);
 	}
+
+	clear_bit(0, &backtrace_flag);
+	smp_mb__after_clear_bit();
+
+out_restore_irq:
+	local_irq_restore(flags);
 }
 
 static int __kprobes
Signature · 7 lines
1.7.2.3

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
#4 lore raw

Re: [PATCH 3/3] x86: Avoid calling arch_trigger_all_cpu_backtrace() at the same time

Quoted 59 lines · Frederic Weisbecker
Quoted 58 lines · Don Zickus

From: Dongdong Deng <[email protected]>

The spin_lock_debug/rcu_cpu_stall detector uses trigger_all_cpu_backtrace() to dump cpu backtrace. Therefore it is possible that trigger_all_cpu_backtrace() could be called at the same time on different CPUs, which triggers and 'unknown reason NMI' warning. The following case illustrates the problem:

CPU1 CPU2 ... CPU N
trigger_all_cpu_backtrace() set "backtrace_mask" to cpu mask
|
generate NMI interrupts generate NMI interrupts ...
\ | /
\ | /

The "backtrace_mask" will be cleaned by the first NMI interrupt at nmi_watchdog_tick(), then the following NMI interrupts generated by other cpus's arch_trigger_all_cpu_backtrace() will be taken as unknown reason NMI interrupts.

This patch uses a test_and_set to avoid the problem, and stop the arch_trigger_all_cpu_backtrace() from calling to avoid dumping a double cpu backtrace info when there is already a trigger_all_cpu_backtrace() in progress.

---
 arch/x86/kernel/apic/hw_nmi.c |   24 ++++++++++++++++++++++++
 1 files changed, 24 insertions(+), 0 deletions(-)
arch/x86/kernel/apic/hw_nmi.c +8 -0
index f349647..d892896 100644
@@ -27,9 +27,27 @@ u64 hw_nmi_get_sample_period(void)
 /* For reliability, we're prepared to waste bits here. */
 static DECLARE_BITMAP(backtrace_mask, NR_CPUS) __read_mostly;
 
+/* "in progress" flag of arch_trigger_all_cpu_backtrace */
+static unsigned long backtrace_flag;
+
 void arch_trigger_all_cpu_backtrace(void)
 {
 	int i;
+	unsigned long flags;
+
+	/*
+	 * Have to disable irq here, as the
+	 * arch_trigger_all_cpu_backtrace() could be
Quoted 8 lines

arch/x86/kernel/apic/hw_nmi.c u64 hw_nmi_get_sample_period(void) · line 43

Quoted 3 lines
arch/x86/kernel/apic/hw_nmi.c +3 -0
+	 * triggered by "spin_lock()" with irqs on.
+	 */
+	local_irq_save(flags);

I'm not sure I understand why you disable irqs here. It looks safe with the test_and_set_bit already.

Hi Frederic,

Yep, after we use test_and_set_bit to replace spin_lock, the disable irqs ops obvious could be removed here.

I will redo this patch, and send it to Don.

Thanks,
Dongdong

Quoted 36 lines
Quoted 34 lines
arch/x86/kernel/apic/hw_nmi.c +13 -0
+
+	if (test_and_set_bit(0, &backtrace_flag))
+		/*
+		 * If there is already a trigger_all_cpu_backtrace() in progress
+		 * (backtrace_flag == 1), don't output double cpu dump infos.
+		 */
+		goto out_restore_irq;
 
 	cpumask_copy(to_cpumask(backtrace_mask), cpu_online_mask);
 
@@ -42,6 +60,12 @@ void arch_trigger_all_cpu_backtrace(void)
 			break;
 		mdelay(1);
 	}
+
+	clear_bit(0, &backtrace_flag);
+	smp_mb__after_clear_bit();
+
+out_restore_irq:
+	local_irq_restore(flags);
 }
 
 static int __kprobes
Signature · 7 lines
1.7.2.3

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/
#5 lore raw
In reply to Don Zickus
Quoted 1 line · Don Zickus

There are some paths that walk the die_chain with preemption on.

What are those codepaths? At minimum it's worth documenting them.

Thanks,

Ingo

author
#6 lore raw
Quoted 6 lines · Ingo Molnar
Quoted 1 line · Don Zickus

There are some paths that walk the die_chain with preemption on.

What are those codepaths? At minimum it's worth documenting them.

Well the one that caused the bug was do_general_protection which walks the die_chain with DIE_GPF.

I can document them, though it might be time consuming to audit them and hope they don't change. I guess my bigger question is, is it expected that anyone who calls the die_chain to have preemption disabled? If not, then does it matter if we document it?

Cheers,
Don

#7 lore raw
Quoted 6 lines · Don Zickus
Quoted 5 lines · Ingo Molnar
Quoted 1 line · Don Zickus

There are some paths that walk the die_chain with preemption on.

Quoted 7 lines
Quoted 1 line

What are those codepaths? At minimum it's worth documenting them.

Well the one that caused the bug was do_general_protection which walks the die_chain with DIE_GPF.

I can document them, though it might be time consuming to audit them and hope they don't change.

Listing one example is enough.

Quoted 3 lines

[...] I guess my bigger question is, is it expected that anyone who calls the die_chain to have preemption disabled? If not, then does it matter if we document it?

Yes, it might be a bug to call those handlers with preemption on (or even with irqs on). But if the code is fine as-is then documenting a single example would be nice.

Thanks,

Ingo

author
#8 lore raw
Quoted 17 lines · Ingo Molnar
Quoted 6 lines · Don Zickus
Quoted 5 lines · Ingo Molnar
Quoted 1 line · Don Zickus

There are some paths that walk the die_chain with preemption on.

Quoted 7 lines
Quoted 1 line

What are those codepaths? At minimum it's worth documenting them.

Well the one that caused the bug was do_general_protection which walks the die_chain with DIE_GPF.

I can document them, though it might be time consuming to audit them and hope they don't change.

Quoted 9 lines

Listing one example is enough.

Quoted 3 lines

[...] I guess my bigger question is, is it expected that anyone who calls the die_chain to have preemption disabled? If not, then does it matter if we document it?

Yes, it might be a bug to call those handlers with preemption on (or even with irqs on). But if the code is fine as-is then documenting a single example would be nice.

Is this better?

Cheers,
Don

------------------------------------->
From: Don Zickus <[email protected]>
Date: Mon, 1 Nov 2010 13:34:33 -0400
Subject: [PATCH 1/6] x86: only call smp_processor_id in non-preempt cases

There are some paths that walk the die_chain with preemption on. Make sure we are in an NMI call before we start doing anything.

This was triggered by do_general_protection calling notify_die with DIE_GPF.

---
 arch/x86/kernel/apic/hw_nmi.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)
arch/x86/kernel/apic/hw_nmi.c +2 -1
index 5c4f952..ef4755d 100644
@@ -49,7 +49,7 @@ arch_trigger_all_cpu_backtrace_handler(struct notifier_block *self,
 {
 	struct die_args *args = __args;
 	struct pt_regs *regs;
-	int cpu = smp_processor_id();
+	int cpu;
 
 	switch (cmd) {
 	case DIE_NMI:
@@ -60,6 +60,7 @@ arch_trigger_all_cpu_backtrace_handler(struct notifier_block *self,
 	}
 
 	regs = args->regs;
+	cpu = smp_processor_id();
 
 	if (cpumask_test_cpu(cpu, to_cpumask(backtrace_mask))) {
 		static arch_spinlock_t lock = __ARCH_SPIN_LOCK_UNLOCKED;
Signature · 1 line
1.7.3.2