Skip to contentSkip to conversation

x86, fix x86 fixup_irqs() error handling

PATCH In review Prarit Bhargava 4 participants 7 messages Mar 5, 2014 – Mar 11, 2014 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86, fix x86 fixup_irqs() error handling

fixup_irqs() calls chip->set_irq_affinity which eventually calls __assign_irq_vector(). Errors are not propogated back from this function call and this results in silent irq relocation failures. This patch fixes this issue and prints out a warning if there is a relocation failure.

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Rui Wang <[email protected]>
Cc Liu Ping Fan <[email protected]>
Cc Bjorn Helgaas <[email protected]>
Cc Yoshihiro YUNOMAE <[email protected]>
Cc Lv Zheng <[email protected]>
Cc Seiji Aguchi <[email protected]>
Cc Yang Zhang <[email protected]>
Cc Andi Kleen <[email protected]>
Cc "Steven Rostedt (Red Hat)" <[email protected]>
Cc Li Fei <[email protected]>
---
 arch/x86/kernel/apic/io_apic.c |   28 ++++++++++++++++++----------
 arch/x86/kernel/irq.c          |    9 +++++++--
 2 files changed, 25 insertions(+), 12 deletions(-)
arch/x86/kernel/apic/io_apic.c +18 -10
index 6ad4658..b4b21db 100644
@@ -2312,7 +2312,7 @@ int __ioapic_set_affinity(struct irq_data *data, const struct cpumask *mask,
 	int err;
 
 	if (!config_enabled(CONFIG_SMP))
-		return -1;
+		return -EPERM;
 
 	if (!cpumask_intersects(mask, cpu_online_mask))
 		return -EINVAL;
@@ -2343,7 +2343,7 @@ int native_ioapic_set_affinity(struct irq_data *data,
 	int ret;
 
 	if (!config_enabled(CONFIG_SMP))
-		return -1;
+		return -EPERM;
 
 	raw_spin_lock_irqsave(&ioapic_lock, flags);
 	ret = __ioapic_set_affinity(data, mask, &dest);
@@ -3075,9 +3075,11 @@ msi_set_affinity(struct irq_data *data, const struct cpumask *mask, bool force)
 	struct irq_cfg *cfg = data->chip_data;
 	struct msi_msg msg;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	__get_cached_msi_msg(data->msi_desc, &msg);
 
@@ -3177,9 +3179,11 @@ dmar_msi_set_affinity(struct irq_data *data, const struct cpumask *mask,
 	struct irq_cfg *cfg = data->chip_data;
 	unsigned int dest, irq = data->irq;
 	struct msi_msg msg;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	dmar_msi_read(irq, &msg);
 
@@ -3226,9 +3230,11 @@ static int hpet_msi_set_affinity(struct irq_data *data,
 	struct irq_cfg *cfg = data->chip_data;
 	struct msi_msg msg;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	hpet_msi_read(data->handler_data, &msg);
 
@@ -3295,9 +3301,11 @@ ht_set_affinity(struct irq_data *data, const struct cpumask *mask, bool force)
 {
 	struct irq_cfg *cfg = data->chip_data;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	target_ht_irq(data->irq, dest, cfg->vector);
 	return IRQ_SET_MASK_OK_NOCOPY;
arch/x86/kernel/irq.c +7 -2
index d99f31d..55fab61 100644
@@ -351,6 +351,7 @@ void fixup_irqs(void)
 	struct irq_desc *desc;
 	struct irq_data *data;
 	struct irq_chip *chip;
+	int ret;
 
 	for_each_irq_desc(irq, desc) {
 		int break_affinity = 0;
@@ -389,8 +390,12 @@ void fixup_irqs(void)
 		if (!irqd_can_move_in_process_context(data) && chip->irq_mask)
 			chip->irq_mask(data);
 
-		if (chip->irq_set_affinity)
-			chip->irq_set_affinity(data, affinity, true);
+		if (chip->irq_set_affinity) {
+			ret = chip->irq_set_affinity(data, affinity, true);
+			WARN(ret == -ENOSPC,
+			     "IRQ %d set affinity failed with %d.  The device assigned to this IRQ is unstable.\n",
+			     irq, ret);
+		}
 		else if (!(warned++))
 			set_affinity = 0;
 
Signature · 1 line
1.7.9.3
#1 lore raw
Quoted 13 lines · Prarit Bhargava
arch/x86/kernel/irq.c +1 -0
index d99f31d..55fab61 100644
@@ -351,6 +351,7 @@ void fixup_irqs(void)
 	struct irq_desc *desc;
 	struct irq_data *data;
 	struct irq_chip *chip;
+	int ret;
 
 	for_each_irq_desc(irq, desc) {
 		int break_affinity = 0;

arch/x86/kernel/irq.c void fixup_irqs(void) · line 397

Quoted 10 lines
arch/x86/kernel/irq.c +5 -2
@@ -389,8 +390,12 @@ void fixup_irqs(void)
 		if (!irqd_can_move_in_process_context(data) && chip->irq_mask)
 			chip->irq_mask(data);
 
-		if (chip->irq_set_affinity)
-			chip->irq_set_affinity(data, affinity, true);
+		if (chip->irq_set_affinity) {
+			ret = chip->irq_set_affinity(data, affinity, true);
+			WARN(ret == -ENOSPC,
+			     "IRQ %d set affinity failed with %d.  The device assigned to this IRQ is unstable.\n",
+			     irq, ret);

Should this be WARN_ON_ONCE() to avoid filling the kernel log instead?

It doesn't make much sense to print out the negative return value, maybe you meant to print -ret instead?

Quoted 4 lines
arch/x86/kernel/irq.c +1 -0
+		}
 		else if (!(warned++))
 			set_affinity = 0;
 
#2 lore raw
Quoted 18 lines · David Rientjes
Quoted 16 lines · Prarit Bhargava
arch/x86/kernel/irq.c +1 -0
index d99f31d..55fab61 100644
@@ -351,6 +351,7 @@ void fixup_irqs(void)
 	struct irq_desc *desc;
 	struct irq_data *data;
 	struct irq_chip *chip;
+	int ret;
 
 	for_each_irq_desc(irq, desc) {
 		int break_affinity = 0;
@@ -389,8 +390,12 @@ void fixup_irqs(void)
 		if (!irqd_can_move_in_process_context(data) && chip->irq_mask)
 			chip->irq_mask(data);
 
Quoted 9 lines

arch/x86/kernel/irq.c void fixup_irqs(void) · line 397

Quoted 7 lines
arch/x86/kernel/irq.c +5 -2
-		if (chip->irq_set_affinity)
-			chip->irq_set_affinity(data, affinity, true);
+		if (chip->irq_set_affinity) {
+			ret = chip->irq_set_affinity(data, affinity, true);
+			WARN(ret == -ENOSPC,
+			     "IRQ %d set affinity failed with %d.  The device assigned to this IRQ is unstable.\n",
+			     irq, ret);

Should this be WARN_ON_ONCE() to avoid filling the kernel log instead?

The problem is that it could hit multiple IRQs ... maybe pr_crit might be better here so we don't flood the log with an unnecessary stack trace; anyone with the source can figure out what the call path is.

Quoted 3 lines

It doesn't make much sense to print out the negative return value, maybe you meant to print -ret instead?

Heh :) You're right. I'll fix that too.

P.

Quoted 5 lines
Quoted 5 lines
arch/x86/kernel/irq.c +1 -0
+		}
 		else if (!(warned++))
 			set_affinity = 0;
 
#3 lore raw
In reply to David Rientjes
Quoted 28 lines · David Rientjes
Quoted 13 lines · Prarit Bhargava
arch/x86/kernel/irq.c +1 -0
index d99f31d..55fab61 100644
@@ -351,6 +351,7 @@ void fixup_irqs(void)
 	struct irq_desc *desc;
 	struct irq_data *data;
 	struct irq_chip *chip;
+	int ret;
 
 	for_each_irq_desc(irq, desc) {
 		int break_affinity = 0;

arch/x86/kernel/irq.c void fixup_irqs(void) · line 397

Quoted 10 lines
arch/x86/kernel/irq.c +5 -2
@@ -389,8 +390,12 @@ void fixup_irqs(void)
 		if (!irqd_can_move_in_process_context(data) && chip->irq_mask)
 			chip->irq_mask(data);
 
-		if (chip->irq_set_affinity)
-			chip->irq_set_affinity(data, affinity, true);
+		if (chip->irq_set_affinity) {
+			ret = chip->irq_set_affinity(data, affinity, true);
+			WARN(ret == -ENOSPC,
+			     "IRQ %d set affinity failed with %d.  The device assigned to this IRQ is unstable.\n",
+			     irq, ret);

Should this be WARN_ON_ONCE() to avoid filling the kernel log instead?

Quoted 2 lines

It doesn't make much sense to print out the negative return value, maybe you meant to print -ret instead?

Well, that does not make sense either. We only print if ret == -ENOSPC!

Quoted 5 lines
Quoted 4 lines
arch/x86/kernel/irq.c +1 -0
+		}
 		else if (!(warned++))
 			set_affinity = 0;
 
#4 lore raw

PATCH 1/1 x86, fix x86 fixup_irqs() error handling [v2]

fixup_irqs() calls chip->set_irq_affinity which eventually calls __assign_irq_vector(). Errors are not propogated back from this function call and this results in silent irq relocation failures. This patch fixes this issue and prints out a warning if there is a relocation failure.

[v2]: modified WARN to pr_crit and changed message

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Rui Wang <[email protected]>
Cc Liu Ping Fan <[email protected]>
Cc Bjorn Helgaas <[email protected]>
Cc Yoshihiro YUNOMAE <[email protected]>
Cc Lv Zheng <[email protected]>
Cc Seiji Aguchi <[email protected]>
Cc Yang Zhang <[email protected]>
Cc Andi Kleen <[email protected]>
Cc "Steven Rostedt (Red Hat)" <[email protected]>
Cc Li Fei <[email protected]>
---
 arch/x86/kernel/apic/io_apic.c |   28 ++++++++++++++++++----------
 arch/x86/kernel/irq.c          |    9 +++++++--
 2 files changed, 25 insertions(+), 12 deletions(-)
arch/x86/kernel/apic/io_apic.c +18 -10
index 6ad4658..b4b21db 100644
@@ -2312,7 +2312,7 @@ int __ioapic_set_affinity(struct irq_data *data, const struct cpumask *mask,
 	int err;
 
 	if (!config_enabled(CONFIG_SMP))
-		return -1;
+		return -EPERM;
 
 	if (!cpumask_intersects(mask, cpu_online_mask))
 		return -EINVAL;
@@ -2343,7 +2343,7 @@ int native_ioapic_set_affinity(struct irq_data *data,
 	int ret;
 
 	if (!config_enabled(CONFIG_SMP))
-		return -1;
+		return -EPERM;
 
 	raw_spin_lock_irqsave(&ioapic_lock, flags);
 	ret = __ioapic_set_affinity(data, mask, &dest);
@@ -3075,9 +3075,11 @@ msi_set_affinity(struct irq_data *data, const struct cpumask *mask, bool force)
 	struct irq_cfg *cfg = data->chip_data;
 	struct msi_msg msg;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	__get_cached_msi_msg(data->msi_desc, &msg);
 
@@ -3177,9 +3179,11 @@ dmar_msi_set_affinity(struct irq_data *data, const struct cpumask *mask,
 	struct irq_cfg *cfg = data->chip_data;
 	unsigned int dest, irq = data->irq;
 	struct msi_msg msg;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	dmar_msi_read(irq, &msg);
 
@@ -3226,9 +3230,11 @@ static int hpet_msi_set_affinity(struct irq_data *data,
 	struct irq_cfg *cfg = data->chip_data;
 	struct msi_msg msg;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	hpet_msi_read(data->handler_data, &msg);
 
@@ -3295,9 +3301,11 @@ ht_set_affinity(struct irq_data *data, const struct cpumask *mask, bool force)
 {
 	struct irq_cfg *cfg = data->chip_data;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	target_ht_irq(data->irq, dest, cfg->vector);
 	return IRQ_SET_MASK_OK_NOCOPY;
arch/x86/kernel/irq.c +7 -2
index d99f31d..32a351d 100644
@@ -351,6 +351,7 @@ void fixup_irqs(void)
 	struct irq_desc *desc;
 	struct irq_data *data;
 	struct irq_chip *chip;
+	int ret;
 
 	for_each_irq_desc(irq, desc) {
 		int break_affinity = 0;
@@ -389,8 +390,12 @@ void fixup_irqs(void)
 		if (!irqd_can_move_in_process_context(data) && chip->irq_mask)
 			chip->irq_mask(data);
 
-		if (chip->irq_set_affinity)
-			chip->irq_set_affinity(data, affinity, true);
+		if (chip->irq_set_affinity) {
+			ret = chip->irq_set_affinity(data, affinity, true);
+			if (ret == -ENOSPC)
+				pr_crit("IRQ %d set affinity failed because there are no available vectors.  The device assigned to this IRQ is unstable.\n",
+					irq);
+		}
 		else if (!(warned++))
 			set_affinity = 0;
 
Signature · 1 line
1.7.9.3
#5 lore raw

Re: [PATCH] x86, fix x86 fixup_irqs() error handling [v2]

Quoted 6 lines · Prarit Bhargava

fixup_irqs() calls chip->set_irq_affinity which eventually calls __assign_irq_vector(). Errors are not propogated back from this function call and this results in silent irq relocation failures. This patch fixes this issue and prints out a warning if there is a relocation failure.

[v2]: modified WARN to pr_crit and changed message

Please fix the changelog to conform to the standard changelog style:

- first describe the symptoms of the bug - how does a user notice?

- then describe how the code behaves today and how that is causing the bug

- and then only describe how it's fixed.

The first item is the most important one - yet it's missing from this changelog.

Thanks,

Ingo

#6 lore raw

PATCH 1/1 x86, fix x86 fixup_irqs() error handling [v3]

Several patches to fix cpu hotplug and the down'd cpu's irq relocations have been submitted in the past month or so. The patches should resolve the problems with cpu hotplug and irq relocation, however, there is always a possibility that a bug still exists. The big problem with debugging these irq reassignments is that the cpu down completes and then we get random stack traces from drivers for which irqs have not been properly assigned to a new cpu. The stack traces are a mix of storage, network, and other kernel subsystem (I once saw the serial port stop working ...) warnings and failures.

The problem with these failures is that they are difficult to diagnose. There is no warning in the cpu hotplug down path to indicate that an IRQ has failed to be assigned to a new cpu, and all we are left with is a stack trace from a driver, or a non-functional device. If we had some information on the console debugging these situations would be much easier; after all we can map an IRQ to a device by simply using lspci or /proc/interrupts.

The current code, fixup_irqs(), which migrates IRQs from the down'd cpu and is called close to the end of the cpu down path, calls chip->set_irq_affinity which eventually calls __assign_irq_vector(). Errors are not propogated back from this function call and this results in silent irq relocation failures.

This patch fixes this issue by returning the error codes up the call stack and prints out a warning if there is a relocation failure.

[v2]: modified WARN to pr_crit and changed message [v3]: updated changelog to conform to standard changelog style

Cc Thomas Gleixner <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc "H. Peter Anvin" <[email protected]>
Cc Rui Wang <[email protected]>
Cc Liu Ping Fan <[email protected]>
Cc Bjorn Helgaas <[email protected]>
Cc Yoshihiro YUNOMAE <[email protected]>
Cc Lv Zheng <[email protected]>
Cc Seiji Aguchi <[email protected]>
Cc Yang Zhang <[email protected]>
Cc Andi Kleen <[email protected]>
Cc "Steven Rostedt (Red Hat)" <[email protected]>
Cc Li Fei <[email protected]>
---
 arch/x86/kernel/apic/io_apic.c |   28 ++++++++++++++++++----------
 arch/x86/kernel/irq.c          |    9 +++++++--
 2 files changed, 25 insertions(+), 12 deletions(-)
arch/x86/kernel/apic/io_apic.c +18 -10
index 6ad4658..b4b21db 100644
@@ -2312,7 +2312,7 @@ int __ioapic_set_affinity(struct irq_data *data, const struct cpumask *mask,
 	int err;
 
 	if (!config_enabled(CONFIG_SMP))
-		return -1;
+		return -EPERM;
 
 	if (!cpumask_intersects(mask, cpu_online_mask))
 		return -EINVAL;
@@ -2343,7 +2343,7 @@ int native_ioapic_set_affinity(struct irq_data *data,
 	int ret;
 
 	if (!config_enabled(CONFIG_SMP))
-		return -1;
+		return -EPERM;
 
 	raw_spin_lock_irqsave(&ioapic_lock, flags);
 	ret = __ioapic_set_affinity(data, mask, &dest);
@@ -3075,9 +3075,11 @@ msi_set_affinity(struct irq_data *data, const struct cpumask *mask, bool force)
 	struct irq_cfg *cfg = data->chip_data;
 	struct msi_msg msg;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	__get_cached_msi_msg(data->msi_desc, &msg);
 
@@ -3177,9 +3179,11 @@ dmar_msi_set_affinity(struct irq_data *data, const struct cpumask *mask,
 	struct irq_cfg *cfg = data->chip_data;
 	unsigned int dest, irq = data->irq;
 	struct msi_msg msg;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	dmar_msi_read(irq, &msg);
 
@@ -3226,9 +3230,11 @@ static int hpet_msi_set_affinity(struct irq_data *data,
 	struct irq_cfg *cfg = data->chip_data;
 	struct msi_msg msg;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	hpet_msi_read(data->handler_data, &msg);
 
@@ -3295,9 +3301,11 @@ ht_set_affinity(struct irq_data *data, const struct cpumask *mask, bool force)
 {
 	struct irq_cfg *cfg = data->chip_data;
 	unsigned int dest;
+	int ret;
 
-	if (__ioapic_set_affinity(data, mask, &dest))
-		return -1;
+	ret = __ioapic_set_affinity(data, mask, &dest);
+	if (ret)
+		return ret;
 
 	target_ht_irq(data->irq, dest, cfg->vector);
 	return IRQ_SET_MASK_OK_NOCOPY;
arch/x86/kernel/irq.c +7 -2
index d99f31d..32a351d 100644
@@ -351,6 +351,7 @@ void fixup_irqs(void)
 	struct irq_desc *desc;
 	struct irq_data *data;
 	struct irq_chip *chip;
+	int ret;
 
 	for_each_irq_desc(irq, desc) {
 		int break_affinity = 0;
@@ -389,8 +390,12 @@ void fixup_irqs(void)
 		if (!irqd_can_move_in_process_context(data) && chip->irq_mask)
 			chip->irq_mask(data);
 
-		if (chip->irq_set_affinity)
-			chip->irq_set_affinity(data, affinity, true);
+		if (chip->irq_set_affinity) {
+			ret = chip->irq_set_affinity(data, affinity, true);
+			if (ret == -ENOSPC)
+				pr_crit("IRQ %d set affinity failed because there are no available vectors.  The device assigned to this IRQ is unstable.\n",
+					irq);
+		}
 		else if (!(warned++))
 			set_affinity = 0;
 
Signature · 1 line
1.7.9.3