amdkfd: fix deadlock in set_perfcount
Subsystems (3)
PATCH 1/1 amdkfd: fix deadlock in set_perfcount
set_perfcount holds the DQM lock (dqm_lock(dqm)
while walking queues, but pqm_update_mqd
calls update_queue, which acquires the
same lock by dqm_lock(dqm), causing deadlocks.
This commit introduces update_queue_locked which assumes the dqm lock is already held, and update_queue calls it with proper lockings.
set_perfcount calls update_queue_locked instead of pqm_update_mqd, therefore the deadlock is fixed.
--- .../drm/amd/amdkfd/kfd_device_queue_manager.c | 44 ++++++++++++------- 1 file changed, 27 insertions(+), 17 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +27 -17
@@ -77,6 +77,9 @@ static struct queue *find_queue_by_doorbell_offset(struct device_queue_manager *static void set_queue_as_reset (struct device_queue_manager *dqm,struct queue *q,struct qcm_process_device *qpd);static int reset_queues_mes (struct device_queue_manager *dqm,struct queue *q); +static int update_queue_locked (struct device_queue_manager *dqm, +struct queue *q, +struct mqd_update_info *minfo);static inline enum KFD_MQD_TYPE get_mqd_type_from_queue_type (enum kfd_queue_type type)
@@ -341,8 +344,7 @@ static void set_perfcount(struct device_queue_manager *dqm, int enable)list_for_each_entry (cur, &dqm->queues, list) { qpd = cur->qpd;list_for_each_entry (q, &qpd->queues_list, list) { -pqm_update_mqd (qpd->pqm, q->properties.queue_id, - &minfo); +update_queue_locked (dqm, q, &minfo); } }dqm_unlock (dqm);
@@ -1059,8 +1061,9 @@ static int destroy_queue_nocpsch(struct device_queue_manager *dqm,return retval; } -static int update_queue (struct device_queue_manager *dqm,struct queue *q, -struct mqd_update_info *minfo) +static int update_queue_locked (struct device_queue_manager *dqm, +struct queue *q, +struct mqd_update_info *minfo) {int retval =0 ;struct device *dev = dqm->dev->adev->dev;
@@ -1068,12 +1071,11 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,struct kfd_process_device *pdd;bool prev_active = false; -dqm_lock (dqm); +lockdep_assert_held (&dqm->lock_hidden); + pdd =kfd_get_process_device_data (q->device, q->process); -if (!pdd) { - retval = -ENODEV ; - goto out_unlock; - } +if (!pdd) +return -ENODEV ; mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type ( q->properties.type)];
@@ -1089,14 +1091,12 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, retval =remove_queue_mes (dqm, q, &pdd->qpd);/* queue is reset so inaccessable */ -if (pdd->has_reset_queue) { - retval = -EACCES ; - goto out_unlock; - } +if (pdd->has_reset_queue) +return -EACCES ;if (retval) {dev_err (dev,"unmap queue failed\n" ); - goto out_unlock; +return retval; } } else if (prev_active && (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
@@ -1105,7 +1105,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,if (!dqm->sched_running) {WARN_ONCE (1 ,"Update non-HWS queue while stopped\n" ); - goto out_unlock; +return -EIO ; } retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
@@ -1115,7 +1115,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);if (retval) {dev_err (dev,"destroy mqd failed\n" ); - goto out_unlock; +return retval; } }
@@ -1163,8 +1163,18 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, &q->properties, current->mm); } -out_unlock: +return retval; +} + +static int update_queue (struct device_queue_manager *dqm,struct queue *q, +struct mqd_update_info *minfo) +{ +int retval; + +dqm_lock (dqm); + retval =update_queue_locked (dqm, q, minfo);dqm_unlock (dqm); +return retval; }
Signature · 1 line
2.53.0
[+Benjamin]
Quoted 11 lines · Zhu Lingshan
set_perfcountholds the DQM lock (dqm_lock(dqm)
while walking queues, butpqm_update_mqd
callsupdate_queue, which acquires the
same lock bydqm_lock(dqm), causing deadlocks.This commit introduces
update_queue_lockedwhich assumes the dqm lock is already held, andupdate_queuecalls it with proper lockings.
set_perfcountcallsupdate_queue_lockedinstead ofpqm_update_mqd, therefore the deadlock is fixed.
I think the patch looks fine to me. I'm just wondering how this bug was not noticed in developer testing of this feature. I see a very similar patch already exists on the DKMS branch but somehow it's missing in amd-staging-drm-next:
commit 5cd0710732a45527b6398f8b522289f9adca708b
Author: Benjamin Welton<[email protected]> AuthorDate: Mon Jan 6 10:30:01 2025 -0800 Commit: Chengjun Yao<[email protected]> CommitDate: Mon Jun 22 12:51:06 2026 +0800
amd/amdkfd: fix double lock aquisition in set_perfcount
Seperates out locking from update_queue to allow updating of queues by code already holding the mqd lock. Fixes a hang in set_perfcount. This change was in the original mailing list commit for set_perfcount but was not included in gerrit.
Benjamin, do you know what's going on here? I think the above patch just needs to be ported to amd-staging-drm-next. But then I'm wondering what else is missing. Or maybe the whole feature was submitted to amd-staging-drm-next by mistake.
Regards,
Felix
Quoted 115 lines
Signed-off-by Zhu Lingshan<[email protected]>--- .../drm/amd/amdkfd/kfd_device_queue_manager.c | 44 ++++++++++++------- 1 file changed, 27 insertions(+), 17 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +27
-17@@ -77,6 +77,9 @@ static struct queue *find_queue_by_doorbell_offset(struct device_queue_manager *static void set_queue_as_reset (struct device_queue_manager *dqm,struct queue *q,struct qcm_process_device *qpd);static int reset_queues_mes (struct device_queue_manager *dqm,struct queue *q); +static int update_queue_locked (struct device_queue_manager *dqm, +struct queue *q, +struct mqd_update_info *minfo);static inline enum KFD_MQD_TYPE get_mqd_type_from_queue_type (enum kfd_queue_type type)@@ -341,8 +344,7 @@ static void set_perfcount(struct device_queue_manager *dqm, int enable)list_for_each_entry (cur, &dqm->queues, list) { qpd = cur->qpd;list_for_each_entry (q, &qpd->queues_list, list) { -pqm_update_mqd (qpd->pqm, q->properties.queue_id, - &minfo); +update_queue_locked (dqm, q, &minfo); } }dqm_unlock (dqm);@@ -1059,8 +1061,9 @@ static int destroy_queue_nocpsch(struct device_queue_manager *dqm,return retval; } -static int update_queue (struct device_queue_manager *dqm,struct queue *q, -struct mqd_update_info *minfo) +static int update_queue_locked (struct device_queue_manager *dqm, +struct queue *q, +struct mqd_update_info *minfo) {int retval =0 ;struct device *dev = dqm->dev->adev->dev;@@ -1068,12 +1071,11 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,struct kfd_process_device *pdd;bool prev_active = false; -dqm_lock (dqm); +lockdep_assert_held (&dqm->lock_hidden); + pdd =kfd_get_process_device_data (q->device, q->process); -if (!pdd) { - retval = -ENODEV ; - goto out_unlock; - } +if (!pdd) +return -ENODEV ; mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type ( q->properties.type)];@@ -1089,14 +1091,12 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, retval =remove_queue_mes (dqm, q, &pdd->qpd);/* queue is reset so inaccessable */ -if (pdd->has_reset_queue) { - retval = -EACCES ; - goto out_unlock; - } +if (pdd->has_reset_queue) +return -EACCES ;if (retval) {dev_err (dev,"unmap queue failed\n" ); - goto out_unlock; +return retval; } } else if (prev_active && (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||@@ -1105,7 +1105,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,if (!dqm->sched_running) {WARN_ONCE (1 ,"Update non-HWS queue while stopped\n" ); - goto out_unlock; +return -EIO ; } retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,@@ -1115,7 +1115,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);if (retval) {dev_err (dev,"destroy mqd failed\n" ); - goto out_unlock; +return retval; } }@@ -1163,8 +1163,18 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, &q->properties, current->mm); } -out_unlock: +return retval; +} + +static int update_queue (struct device_queue_manager *dqm,struct queue *q, +struct mqd_update_info *minfo) +{ +int retval; + +dqm_lock (dqm); + retval =update_queue_locked (dqm, q, minfo);dqm_unlock (dqm); +return retval; }
Welton, Benjamin
Acked-by Kent RussellAMD General
Interesting, wonder how this was dropped. This has been present and been used by customers for a while (and this is an issue they would absolutely hit).
There was some weirdness in the migration if I remember correctly from DKMS -> Public when a migration for the PTL work took place and I suspect that maybe the change just wasn’t picked up or was dropped as part of that migration. What two branches do you need to check here?
Ben
From: Kuehling, Felix <[email protected]>
Sent: Tuesday, August 4, 2026 3:41 PM
To: Zhu, Lingshan <[email protected]>; Deucher, Alexander <[email protected]>; Koenig, Christian <[email protected]>; Welton, Benjamin <[email protected]>
Cc: Huang, Ray <[email protected]>; [email protected]
Subject: Re: [PATCH] amdkfd: fix deadlock in set_perfcount
[+Benjamin]
On 2026-07-31 05:51, Zhu Lingshan wrote:
set_perfcount holds the DQM lock (dqm_lock(dqm)
while walking queues, but pqm_update_mqd
calls update_queue, which acquires the
same lock by dqm_lock(dqm), causing deadlocks.
This commit introduces update_queue_locked which assumes
the dqm lock is already held, and update_queue calls it
with proper lockings.
set_perfcount calls update_queue_locked instead of pqm_update_mqd,
therefore the deadlock is fixed.
I think the patch looks fine to me. I'm just wondering how this bug was not noticed in developer testing of this feature. I see a very similar patch already exists on the DKMS branch but somehow it's missing in amd-staging-drm-next:
commit 5cd0710732a45527b6398f8b522289f9adca708b
Author: Benjamin Welton <[email protected]><mailto:[email protected]>
AuthorDate: Mon Jan 6 10:30:01 2025 -0800
Commit: Chengjun Yao <[email protected]><mailto:[email protected]>
CommitDate: Mon Jun 22 12:51:06 2026 +0800
amd/amdkfd: fix double lock aquisition in set_perfcount
Seperates out locking from update_queue to allow updating
of queues by code already holding the mqd lock. Fixes a
hang in set_perfcount. This change was in the original
mailing list commit for set_perfcount but was not included
in gerrit.
Benjamin, do you know what's going on here? I think the above patch just needs to be ported to amd-staging-drm-next. But then I'm wondering what else is missing. Or maybe the whole feature was submitted to amd-staging-drm-next by mistake.
Regards,
Felix
---
.../drm/amd/amdkfd/kfd_device_queue_manager.c | 44 ++++++++++++-------
1 file changed, 27 insertions(+), 17 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index 51ee9c39104b..c34c5f6a5541 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -77,6 +77,9 @@ static struct queue *find_queue_by_doorbell_offset(struct device_queue_manager *
static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q,
struct qcm_process_device *qpd);
static int reset_queues_mes(struct device_queue_manager *dqm, struct queue *q);
+static int update_queue_locked(struct device_queue_manager *dqm,
+ struct queue *q,
+ struct mqd_update_info *minfo);
static inline
enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
@@ -341,8 +344,7 @@ static void set_perfcount(struct device_queue_manager *dqm, int enable)
list_for_each_entry(cur, &dqm->queues, list) { qpd = cur->qpd;
list_for_each_entry(q, &qpd->queues_list, list) {- pqm_update_mqd(qpd->pqm, q->properties.queue_id,
- &minfo);
+ update_queue_locked(dqm, q, &minfo);
}
} dqm_unlock(dqm);
@@ -1059,8 +1061,9 @@ static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
return retval;
}
-static int update_queue(struct device_queue_manager *dqm, struct queue *q,
- struct mqd_update_info *minfo)
+static int update_queue_locked(struct device_queue_manager *dqm,
+ struct queue *q,
+ struct mqd_update_info *minfo)
{
int retval = 0; struct device *dev = dqm->dev->adev->dev;
@@ -1068,12 +1071,11 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,
struct kfd_process_device *pdd;
bool prev_active = false;
- dqm_lock(dqm);
+ lockdep_assert_held(&dqm->lock_hidden);
+
pdd = kfd_get_process_device_data(q->device, q->process);
- if (!pdd) {
- retval = -ENODEV;
- goto out_unlock;
- }
+ if (!pdd)
+ return -ENODEV;
mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
q->properties.type)];
@@ -1089,14 +1091,12 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,
retval = remove_queue_mes(dqm, q, &pdd->qpd);
/* queue is reset so inaccessable */- if (pdd->has_reset_queue) {
- retval = -EACCES;
- goto out_unlock;
- }
+ if (pdd->has_reset_queue)
+ return -EACCES;
if (retval) {
dev_err(dev, "unmap queue failed\n");- goto out_unlock;
+ return retval;
}
} else if (prev_active &&
(q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
@@ -1105,7 +1105,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,
if (!dqm->sched_running) {
WARN_ONCE(1, "Update non-HWS queue while stopped\n");- goto out_unlock;
+ return -EIO;
}
retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,@@ -1115,7 +1115,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,
KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
if (retval) {
dev_err(dev, "destroy mqd failed\n");- goto out_unlock;
+ return retval;
}
}
@@ -1163,8 +1163,18 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,
&q->properties, current->mm);
}
-out_unlock:
+ return retval;
+}
+
+static int update_queue(struct device_queue_manager *dqm, struct queue *q,
+ struct mqd_update_info *minfo)
+{
+ int retval;
+
+ dqm_lock(dqm);
+ retval = update_queue_locked(dqm, q, minfo);
dqm_unlock(dqm);
+
return retval;
}