Skip to contentSkip to conversation

perf/arm-dmc620: Fix dmc620_pmu_irqs_lock/cpu_hotplug_lock circular lock dependency

PATCH Superseded Waiman Long 3 participants 5 messages Aug 7, 2023 – Aug 10, 2023 Arm PMU profiling and debugging

v4 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 perf/arm-dmc620: Fix dmc620_pmu_irqs_lock/cpu_hotplug_lock circular lock dependency

The following circular locking dependency was reported when running cpus online/offline test on an arm64 system.

[   84.195923] Chain exists of:
                 dmc620_pmu_irqs_lock --> cpu_hotplug_lock --> cpuhp_state-down
[   84.207305]  Possible unsafe locking scenario:
[   84.213212]        CPU0                    CPU1
[   84.217729]        ----                    ----
[   84.222247]   lock(cpuhp_state-down);
[   84.225899]                                lock(cpu_hotplug_lock);
[   84.232068]                                lock(cpuhp_state-down);
[   84.238237]   lock(dmc620_pmu_irqs_lock);
[   84.242236]
                *** DEADLOCK ***

The problematic locking order seems to be

	lock(dmc620_pmu_irqs_lock) --> lock(cpu_hotplug_lock)

This locking order happens when dmc620_pmu_get_irq() calls cpuhp_state_add_instance_nocalls(). Since dmc620_pmu_irqs_lock is used for protecting the dmc620_pmu_irqs structure only, we don't actually need to hold the lock when adding a new instance to the CPU hotplug subsystem.

Fix this possible deadlock scenario by adding a new dmc620_pmu_get_irq_lock for protecting the call to __dmc620_pmu_get_irq() and taking dmc620_pmu_irqs_lock inside __dmc620_pmu_get_irq() only when dmc620_pmu_irqs is being searched or modified. As a result, cpuhp_state_add_instance_nocalls() won't be called with dmc620_pmu_irqs_lock held and cpu_hotplug_lock won't be acquired after dmc620_pmu_irqs_lock.

---
 drivers/perf/arm_dmc620_pmu.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)
drivers/perf/arm_dmc620_pmu.c +14 -4
index 9d0f01c4455a..895971915f2d 100644
@@ -68,6 +68,7 @@
 
 static LIST_HEAD(dmc620_pmu_irqs);
 static DEFINE_MUTEX(dmc620_pmu_irqs_lock);
+static DEFINE_MUTEX(dmc620_pmu_get_irq_lock);
 
 struct dmc620_pmu_irq {
 	struct hlist_node node;
@@ -421,11 +422,18 @@ static irqreturn_t dmc620_pmu_handle_irq(int irq_num, void *data)
 static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
 {
 	struct dmc620_pmu_irq *irq;
+	bool found = false;
 	int ret;
 
+	mutex_lock(&dmc620_pmu_irqs_lock);
 	list_for_each_entry(irq, &dmc620_pmu_irqs, irqs_node)
-		if (irq->irq_num == irq_num && refcount_inc_not_zero(&irq->refcount))
-			return irq;
+		if (irq->irq_num == irq_num && refcount_inc_not_zero(&irq->refcount)) {
+			found = true;
+			break;
+		}
+	mutex_unlock(&dmc620_pmu_irqs_lock);
+	if (found)
+		return irq;
 
 	irq = kzalloc(sizeof(*irq), GFP_KERNEL);
 	if (!irq)
@@ -452,7 +460,9 @@ static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
 		goto out_free_irq;
 
 	irq->irq_num = irq_num;
+	mutex_lock(&dmc620_pmu_irqs_lock);
 	list_add(&irq->irqs_node, &dmc620_pmu_irqs);
+	mutex_unlock(&dmc620_pmu_irqs_lock);
 
 	return irq;
 
@@ -467,9 +477,9 @@ static int dmc620_pmu_get_irq(struct dmc620_pmu *dmc620_pmu, int irq_num)
 {
 	struct dmc620_pmu_irq *irq;
 
-	mutex_lock(&dmc620_pmu_irqs_lock);
+	mutex_lock(&dmc620_pmu_get_irq_lock);
 	irq = __dmc620_pmu_get_irq(irq_num);
-	mutex_unlock(&dmc620_pmu_irqs_lock);
+	mutex_unlock(&dmc620_pmu_get_irq_lock);
 
 	if (IS_ERR(irq))
 		return PTR_ERR(irq);
Signature · 7 lines
2.31.1


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
#1 lore raw
Quoted 53 lines · Waiman Long

The following circular locking dependency was reported when running cpus online/offline test on an arm64 system.

[   84.195923] Chain exists of:
                  dmc620_pmu_irqs_lock --> cpu_hotplug_lock --> cpuhp_state-down
[   84.207305]  Possible unsafe locking scenario:
[   84.213212]        CPU0                    CPU1
[   84.217729]        ----                    ----
[   84.222247]   lock(cpuhp_state-down);
[   84.225899]                                lock(cpu_hotplug_lock);
[   84.232068]                                lock(cpuhp_state-down);
[   84.238237]   lock(dmc620_pmu_irqs_lock);
[   84.242236]
                 *** DEADLOCK ***

The problematic locking order seems to be

	lock(dmc620_pmu_irqs_lock) --> lock(cpu_hotplug_lock)

This locking order happens when dmc620_pmu_get_irq() calls cpuhp_state_add_instance_nocalls(). Since dmc620_pmu_irqs_lock is used for protecting the dmc620_pmu_irqs structure only, we don't actually need to hold the lock when adding a new instance to the CPU hotplug subsystem.

Fix this possible deadlock scenario by adding a new dmc620_pmu_get_irq_lock for protecting the call to __dmc620_pmu_get_irq() and taking dmc620_pmu_irqs_lock inside __dmc620_pmu_get_irq() only when dmc620_pmu_irqs is being searched or modified. As a result, cpuhp_state_add_instance_nocalls() won't be called with dmc620_pmu_irqs_lock held and cpu_hotplug_lock won't be acquired after dmc620_pmu_irqs_lock.

---
  drivers/perf/arm_dmc620_pmu.c | 18 ++++++++++++++----
  1 file changed, 14 insertions(+), 4 deletions(-)
drivers/perf/arm_dmc620_pmu.c +1 -0
index 9d0f01c4455a..895971915f2d 100644
@@ -68,6 +68,7 @@
 
  static LIST_HEAD(dmc620_pmu_irqs);
  static DEFINE_MUTEX(dmc620_pmu_irqs_lock);
+static DEFINE_MUTEX(dmc620_pmu_get_irq_lock);
 
  struct dmc620_pmu_irq {
  	struct hlist_node node;

drivers/perf/arm_dmc620_pmu.c static irqreturn_t dmc620_pmu_handle_irq(int irq_num, void *data) · line 428

Quoted 7 lines
drivers/perf/arm_dmc620_pmu.c +2 -0
@@ -421,11 +422,18 @@ static irqreturn_t dmc620_pmu_handle_irq(int irq_num, void *data)
  static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
  {
  	struct dmc620_pmu_irq *irq;
+	bool found = false;
  	int ret;
 
+	mutex_lock(&dmc620_pmu_irqs_lock);

Do we strictly need this? I'd hope that the outer release/acquire of dmc620_get_pmu_irqs_lock already means we can't observe an invalid value of irq->irq_num, and the refcount op should be atomic in itself, no? Fair enough if there's some other subtlety I'm missing - I do trust that you're more experienced in locking and barrier semantics than I am! - and if it comes to it I'd agree that simple extra locking is preferable to getting into explicit memory barriers here.

One other nit either way, could we clarify the names to be something like irqs_list_lock and irqs_users_lock? The split locking scheme doesn't exactly lend itself to being super-obvious, especially if we do end up nesting both locks, so I think naming them after what they semantically protect seems the most readable option. Otherwise, this does pretty much look like what I originally had in mind.

Cheers,
Robin.

Quoted 25 lines
drivers/perf/arm_dmc620_pmu.c +9 -2
  	list_for_each_entry(irq, &dmc620_pmu_irqs, irqs_node)
-		if (irq->irq_num == irq_num && refcount_inc_not_zero(&irq->refcount))
-			return irq;
+		if (irq->irq_num == irq_num && refcount_inc_not_zero(&irq->refcount)) {
+			found = true;
+			break;
+		}
+	mutex_unlock(&dmc620_pmu_irqs_lock);
+	if (found)
+		return irq;
 
  	irq = kzalloc(sizeof(*irq), GFP_KERNEL);
  	if (!irq)
@@ -452,7 +460,9 @@ static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
  		goto out_free_irq;
 
  	irq->irq_num = irq_num;
+	mutex_lock(&dmc620_pmu_irqs_lock);
  	list_add(&irq->irqs_node, &dmc620_pmu_irqs);
+	mutex_unlock(&dmc620_pmu_irqs_lock);
 
  	return irq;
 
@@ -467,9 +477,9 @@ static int dmc620_pmu_get_irq(struct dmc620_pmu *dmc620_pmu, int irq_num)
  {

drivers/perf/arm_dmc620_pmu.c static int dmc620_pmu_get_irq(struct dmc620_pmu *dmc620_pmu, int irq_num) · line 485

Quoted 10 lines
drivers/perf/arm_dmc620_pmu.c +2 -2
  	struct dmc620_pmu_irq *irq;
 
-	mutex_lock(&dmc620_pmu_irqs_lock);
+	mutex_lock(&dmc620_pmu_get_irq_lock);
  	irq = __dmc620_pmu_get_irq(irq_num);
-	mutex_unlock(&dmc620_pmu_irqs_lock);
+	mutex_unlock(&dmc620_pmu_get_irq_lock);
 
  	if (IS_ERR(irq))
  		return PTR_ERR(irq);

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

#2 lore raw
Quoted 53 lines · Robin Murphy
Quoted 52 lines · Waiman Long

The following circular locking dependency was reported when running cpus online/offline test on an arm64 system.

[   84.195923] Chain exists of:
                  dmc620_pmu_irqs_lock --> cpu_hotplug_lock --> cpuhp_state-down

[   84.207305]  Possible unsafe locking scenario:

[   84.213212]        CPU0                    CPU1 [   84.217729]        ----                    ---- [   84.222247]   lock(cpuhp_state-down);
[   84.225899] lock(cpu_hotplug_lock);
[   84.232068] lock(cpuhp_state-down);
[   84.238237]   lock(dmc620_pmu_irqs_lock); [   84.242236]
                 *** DEADLOCK ***

The problematic locking order seems to be

    lock(dmc620_pmu_irqs_lock) --> lock(cpu_hotplug_lock)

This locking order happens when dmc620_pmu_get_irq() calls cpuhp_state_add_instance_nocalls(). Since dmc620_pmu_irqs_lock is used for protecting the dmc620_pmu_irqs structure only, we don't actually need to hold the lock when adding a new instance to the CPU hotplug subsystem.

Fix this possible deadlock scenario by adding a new dmc620_pmu_get_irq_lock for protecting the call to __dmc620_pmu_get_irq() and taking dmc620_pmu_irqs_lock inside __dmc620_pmu_get_irq() only when dmc620_pmu_irqs is being searched or modified. As a result, cpuhp_state_add_instance_nocalls() won't be called with dmc620_pmu_irqs_lock held and cpu_hotplug_lock won't be acquired after dmc620_pmu_irqs_lock.

---
  drivers/perf/arm_dmc620_pmu.c | 18 ++++++++++++++----

  1 file changed, 14 insertions(+), 4 deletions(-)
drivers/perf/arm_dmc620_pmu.c +1 -0
index 9d0f01c4455a..895971915f2d 100644
@@ -68,6 +68,7 @@
     static LIST_HEAD(dmc620_pmu_irqs);
   static DEFINE_MUTEX(dmc620_pmu_irqs_lock);
+static DEFINE_MUTEX(dmc620_pmu_get_irq_lock);
     struct dmc620_pmu_irq {
       struct hlist_node node;
@@ -421,11 +422,18 @@ static irqreturn_t dmc620_pmu_handle_irq(int 
 irq_num, void *data)
Quoted 8 lines

drivers/perf/arm_dmc620_pmu.c static irqreturn_t dmc620_pmu_handle_irq(int irq_num, void *data)

Quoted 6 lines
drivers/perf/arm_dmc620_pmu.c +1 -0
   static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
   {
       struct dmc620_pmu_irq *irq;
+    bool found = false;
       int ret;
   +    mutex_lock(&dmc620_pmu_irqs_lock);

Do we strictly need this? I'd hope that the outer release/acquire of dmc620_get_pmu_irqs_lock already means we can't observe an invalid value of irq->irq_num, and the refcount op should be atomic in itself, no? Fair enough if there's some other subtlety I'm missing - I do trust that you're more experienced in locking and barrier semantics than I am! - and if it comes to it I'd agree that simple extra locking is preferable to getting into explicit memory barriers here. locking

I guess we can use rcu_read_lock/rcu_read_unlock and list_for_each_entry_rcu() to avoid taking dmc620_pmu_irqs_lock here. However, we also need to change the list_del(&irq->irqs_node) & list_add(&irq->irqs_node,...) to use their rcu equivalents to make it more fail-safe. The problem with RCU is that you have to think carefully before you can use it. Locking, on the other hand, don't need such serious thought. So it is easier for lazy people :-) So I still prefer the simple locking scheme.

Quoted 2 lines

One other nit either way, could we clarify the names to be something like irqs_list_lock and irqs_users_lock? The split locking scheme doesn't exactly lend itself to being super-obvious, especially if we do end up nesting both locks, so I think naming them after what they semantically protect seems the most readable option. Otherwise, this does pretty much look like what I originally had in mind.

I think it is a good to rename dmc620_pmu_irqs_lock to dmc620_pmu_irqs_list_lock. For the other lock, its purpose is to make sure that only one user can get to __dmc620_pmu_get_irq(), may be dmc620_irqs_get_lock. I can add some comment to clarify the nesting relationship.

Cheers,
Longman

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

maintainer
#3 lore raw
Quoted 75 lines · Waiman Long
Quoted 65 lines · Robin Murphy
Quoted 56 lines · Waiman Long

The following circular locking dependency was reported when running cpus online/offline test on an arm64 system.

[   84.195923] Chain exists of:
                  dmc620_pmu_irqs_lock --> cpu_hotplug_lock --> cpuhp_state-down

[   84.207305]  Possible unsafe locking scenario:

[   84.213212]        CPU0                    CPU1 [   84.217729]        ----                    ---- [   84.222247]   lock(cpuhp_state-down);
[   84.225899] lock(cpu_hotplug_lock);
[   84.232068] lock(cpuhp_state-down);
[   84.238237]   lock(dmc620_pmu_irqs_lock); [   84.242236]
                 *** DEADLOCK ***

The problematic locking order seems to be

    lock(dmc620_pmu_irqs_lock) --> lock(cpu_hotplug_lock)

This locking order happens when dmc620_pmu_get_irq() calls cpuhp_state_add_instance_nocalls(). Since dmc620_pmu_irqs_lock is used for protecting the dmc620_pmu_irqs structure only, we don't actually need
to hold the lock when adding a new instance to the CPU hotplug subsystem.

Fix this possible deadlock scenario by adding a new dmc620_pmu_get_irq_lock for protecting the call to __dmc620_pmu_get_irq()
and taking dmc620_pmu_irqs_lock inside __dmc620_pmu_get_irq() only when dmc620_pmu_irqs is being searched or modified. As a result, cpuhp_state_add_instance_nocalls() won't be called with dmc620_pmu_irqs_lock held and cpu_hotplug_lock won't be acquired after dmc620_pmu_irqs_lock.

---
  drivers/perf/arm_dmc620_pmu.c | 18 ++++++++++++++----

  1 file changed, 14 insertions(+), 4 deletions(-)
drivers/perf/arm_dmc620_pmu.c +1 -0
index 9d0f01c4455a..895971915f2d 100644
@@ -68,6 +68,7 @@
     static LIST_HEAD(dmc620_pmu_irqs);
   static DEFINE_MUTEX(dmc620_pmu_irqs_lock);
+static DEFINE_MUTEX(dmc620_pmu_get_irq_lock);
     struct dmc620_pmu_irq {
       struct hlist_node node;

drivers/perf/arm_dmc620_pmu.c static irqreturn_t dmc620_pmu_handle_irq(int · line 428

Quoted 7 lines
drivers/perf/arm_dmc620_pmu.c +1 -0
@@ -421,11 +422,18 @@ static irqreturn_t dmc620_pmu_handle_irq(int
 irq_num, void *data)
   static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
   {
       struct dmc620_pmu_irq *irq;
+    bool found = false;
       int ret;
   +    mutex_lock(&dmc620_pmu_irqs_lock);
Quoted 7 lines

Do we strictly need this? I'd hope that the outer release/acquire of dmc620_get_pmu_irqs_lock already means we can't observe an invalid value of irq->irq_num, and the refcount op should be atomic in itself, no? Fair enough if there's some other subtlety I'm missing - I do trust that you're more experienced in locking and barrier semantics than I am! - and if it comes to it I'd agree that simple extra locking is preferable to getting into explicit memory barriers here. locking

Quoted 2 lines

I guess we can use rcu_read_lock/rcu_read_unlock and list_for_each_entry_rcu() to avoid taking dmc620_pmu_irqs_lock here.

I thought we decided that we couldn't use RCU in:

https://lore.kernel.org/r/[email protected]

?

Quoted 7 lines
Quoted 6 lines

One other nit either way, could we clarify the names to be something like irqs_list_lock and irqs_users_lock? The split locking scheme doesn't exactly lend itself to being super-obvious, especially if we do end up nesting both locks, so I think naming them after what they semantically protect seems the most readable option. Otherwise, this does pretty much look like what I originally had in mind.

Quoted 5 lines

I think it is a good to rename dmc620_pmu_irqs_lock to dmc620_pmu_irqs_list_lock. For the other lock, its purpose is to make sure that only one user can get to __dmc620_pmu_get_irq(), may be dmc620_irqs_get_lock. I can add some comment to clarify the nesting relationship.

Please do that and I'll pick the patch up for 6.6.

Will

#4 lore raw
Quoted 73 lines · Will Deacon
Quoted 72 lines · Waiman Long
Quoted 71 lines · Robin Murphy
Quoted 56 lines · Waiman Long

The following circular locking dependency was reported when running cpus online/offline test on an arm64 system.

[   84.195923] Chain exists of:
                  dmc620_pmu_irqs_lock --> cpu_hotplug_lock -->
cpuhp_state-down

[   84.207305]  Possible unsafe locking scenario:

[   84.213212]        CPU0                    CPU1 [   84.217729]        ----                    ---- [   84.222247]   lock(cpuhp_state-down);
[   84.225899] lock(cpu_hotplug_lock);
[   84.232068] lock(cpuhp_state-down);
[   84.238237]   lock(dmc620_pmu_irqs_lock); [   84.242236]
                 *** DEADLOCK ***

The problematic locking order seems to be

    lock(dmc620_pmu_irqs_lock) --> lock(cpu_hotplug_lock)

This locking order happens when dmc620_pmu_get_irq() calls cpuhp_state_add_instance_nocalls(). Since dmc620_pmu_irqs_lock is used for protecting the dmc620_pmu_irqs structure only, we don't actually need
to hold the lock when adding a new instance to the CPU hotplug subsystem.

Fix this possible deadlock scenario by adding a new dmc620_pmu_get_irq_lock for protecting the call to __dmc620_pmu_get_irq()
and taking dmc620_pmu_irqs_lock inside __dmc620_pmu_get_irq() only when dmc620_pmu_irqs is being searched or modified. As a result, cpuhp_state_add_instance_nocalls() won't be called with dmc620_pmu_irqs_lock held and cpu_hotplug_lock won't be acquired after dmc620_pmu_irqs_lock.

---
   drivers/perf/arm_dmc620_pmu.c | 18 ++++++++++++++----
   1 file changed, 14 insertions(+), 4 deletions(-)
drivers/perf/arm_dmc620_pmu.c +1 -0
index 9d0f01c4455a..895971915f2d 100644
@@ -68,6 +68,7 @@
     static LIST_HEAD(dmc620_pmu_irqs);
   static DEFINE_MUTEX(dmc620_pmu_irqs_lock);
+static DEFINE_MUTEX(dmc620_pmu_get_irq_lock);
     struct dmc620_pmu_irq {
       struct hlist_node node;

drivers/perf/arm_dmc620_pmu.c static irqreturn_t dmc620_pmu_handle_irq(int · line 428

Quoted 7 lines
drivers/perf/arm_dmc620_pmu.c +1 -0
@@ -421,11 +422,18 @@ static irqreturn_t dmc620_pmu_handle_irq(int
 irq_num, void *data)
   static struct dmc620_pmu_irq *__dmc620_pmu_get_irq(int irq_num)
   {
       struct dmc620_pmu_irq *irq;
+    bool found = false;
       int ret;
   +    mutex_lock(&dmc620_pmu_irqs_lock);

Do we strictly need this? I'd hope that the outer release/acquire of dmc620_get_pmu_irqs_lock already means we can't observe an invalid value of irq->irq_num, and the refcount op should be atomic in itself, no? Fair enough if there's some other subtlety I'm missing - I do trust that you're more experienced in locking and barrier semantics than I am! - and if it comes to it I'd agree that simple extra locking is preferable to getting into explicit memory barriers here. locking

Quoted 7 lines
Quoted 2 lines

I guess we can use rcu_read_lock/rcu_read_unlock and list_for_each_entry_rcu() to avoid taking dmc620_pmu_irqs_lock here.

I thought we decided that we couldn't use RCU in:

https://lore.kernel.org/r/[email protected]

?

Right. I am not planning to use RCU anyway.

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines

One other nit either way, could we clarify the names to be something like irqs_list_lock and irqs_users_lock? The split locking scheme doesn't exactly lend itself to being super-obvious, especially if we do end up nesting both locks, so I think naming them after what they semantically protect seems the most readable option. Otherwise, this does pretty much look like what I originally had in mind.

Quoted 6 lines
Quoted 5 lines

I think it is a good to rename dmc620_pmu_irqs_lock to dmc620_pmu_irqs_list_lock. For the other lock, its purpose is to make sure that only one user can get to __dmc620_pmu_get_irq(), may be dmc620_irqs_get_lock. I can add some comment to clarify the nesting relationship.

Please do that and I'll pick the patch up for 6.6.

Will do.

Cheers,
Longman

_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel