Re: [PATCH v5] perf/arm-dmc620: Fix dmc620_pmu_irqs_lock/cpu_hotplug_lock circular lock dependency
From: Waiman Long
Date: Fri Aug 11 2023 - 14:27:30 EST
On 8/10/23 13:21, Robin Murphy wrote:
On 10/08/2023 4:41 pm, Waiman Long wrote:
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, 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_lock
for protecting the call to __dmc620_pmu_get_irq(). While at it, rename
dmc620_pmu_irqs_lock to dmc620_pmu_list_lock as it is now just
protecting
the iteration and modification of pmus_node and irqs_node lists.
As a result, cpuhp_state_add_instance_nocalls() won't be called with
reanemd dmc620_pmu_list_lock held and cpu_hotplug_lock won't be acquired
after dmc620_pmu_list_lock.
Suggested-by: Robin Murphy <robin.murphy@xxxxxxx>
Signed-off-by: Waiman Long <longman@xxxxxxxxxx>
---
drivers/perf/arm_dmc620_pmu.c | 34 ++++++++++++++++++++++------------
1 file changed, 22 insertions(+), 12 deletions(-)
diff --git a/drivers/perf/arm_dmc620_pmu.c
b/drivers/perf/arm_dmc620_pmu.c
index 9d0f01c4455a..a5bfc8f2e6ab 100644
--- a/drivers/perf/arm_dmc620_pmu.c
+++ b/drivers/perf/arm_dmc620_pmu.c
@@ -66,8 +66,14 @@
#define DMC620_PMU_COUNTERn_OFFSET(n) \
(DMC620_PMU_COUNTERS_BASE + 0x28 * (n))
+/*
+ * The allowable lock ordering is:
+ * - dmc620_pmu_get_lock (protects call to __dmc620_pmu_get_irq())
+ * - dmc620_pmu_list_lock (protects pmus_node & irqs_node lists)
Sorry, this isn't right: touching the irqs_node list *is* the aspect
of __dmc620_pmu_get_irq() which warrants globally locking. It's then
the pmus_node lists which want locking separately from that - those
could strictly be locked per dmc620_pmu_irq instance, but that would
be a big waste of space, so we can still combine them under a single
global lock. I just went too far in thinking I could get away with
(ab)using the same lock for both purposes since they didn't overlap :)
OK, you want separate locks for pmus_node list and irqs_node list. That
will be fine too. I can make the change.
Thanks,
Longman
+ */
+static DEFINE_MUTEX(dmc620_pmu_get_lock);
+static DEFINE_MUTEX(dmc620_pmu_list_lock);
static LIST_HEAD(dmc620_pmu_irqs);
-static DEFINE_MUTEX(dmc620_pmu_irqs_lock);
struct dmc620_pmu_irq {
struct hlist_node node;
@@ -423,9 +429,11 @@ static struct dmc620_pmu_irq
*__dmc620_pmu_get_irq(int irq_num)
struct dmc620_pmu_irq *irq;
int ret;
+ mutex_lock(&dmc620_pmu_list_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;
+ goto unlock_out;
+ mutex_unlock(&dmc620_pmu_list_lock);
irq = kzalloc(sizeof(*irq), GFP_KERNEL);
if (!irq)
@@ -452,8 +460,10 @@ 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_list_lock);
list_add(&irq->irqs_node, &dmc620_pmu_irqs);
-
+unlock_out:
+ mutex_unlock(&dmc620_pmu_list_lock);
return irq;
out_free_irq:
@@ -467,17 +477,17 @@ 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_lock);
irq = __dmc620_pmu_get_irq(irq_num);
- mutex_unlock(&dmc620_pmu_irqs_lock);
+ mutex_unlock(&dmc620_pmu_get_lock);
if (IS_ERR(irq))
return PTR_ERR(irq);
dmc620_pmu->irq = irq;
- mutex_lock(&dmc620_pmu_irqs_lock);
+ mutex_lock(&dmc620_pmu_list_lock);
list_add_rcu(&dmc620_pmu->pmus_node, &irq->pmus_node);
- mutex_unlock(&dmc620_pmu_irqs_lock);
+ mutex_unlock(&dmc620_pmu_list_lock);
return 0;
}
@@ -486,16 +496,16 @@ static void dmc620_pmu_put_irq(struct
dmc620_pmu *dmc620_pmu)
{
struct dmc620_pmu_irq *irq = dmc620_pmu->irq;
- mutex_lock(&dmc620_pmu_irqs_lock);
+ mutex_lock(&dmc620_pmu_list_lock);
list_del_rcu(&dmc620_pmu->pmus_node);
Ah, it might be the laziness in this function that's misled you.
Logically it ought to just be a case of dropping pmu_list_lock here
after removing from the pmus_node list, then taking pmu_get_lock
before the following list_del from the main global list (I think that
shouldn't *need* to cover the refcount operation as well, but equally
there's probably no harm if it does).
Thanks,
Robin.
if (!refcount_dec_and_test(&irq->refcount)) {
- mutex_unlock(&dmc620_pmu_irqs_lock);
+ mutex_unlock(&dmc620_pmu_list_lock);
return;
}
list_del(&irq->irqs_node);
- mutex_unlock(&dmc620_pmu_irqs_lock);
+ mutex_unlock(&dmc620_pmu_list_lock);
free_irq(irq->irq_num, irq);
cpuhp_state_remove_instance_nocalls(cpuhp_state_num, &irq->node);
@@ -638,10 +648,10 @@ static int dmc620_pmu_cpu_teardown(unsigned int
cpu,
return 0;
/* We're only reading, but this isn't the place to be
involving RCU */
- mutex_lock(&dmc620_pmu_irqs_lock);
+ mutex_lock(&dmc620_pmu_list_lock);
list_for_each_entry(dmc620_pmu, &irq->pmus_node, pmus_node)
perf_pmu_migrate_context(&dmc620_pmu->pmu, irq->cpu, target);
- mutex_unlock(&dmc620_pmu_irqs_lock);
+ mutex_unlock(&dmc620_pmu_list_lock);
WARN_ON(irq_set_affinity(irq->irq_num, cpumask_of(target)));
irq->cpu = target;