[PATCH] drm/amdkfd: Fix a race between queue destroy and process termination
xinhui pan
xinhui.pan at amd.com
Tue Jun 15 08:02:33 UTC 2021
We call free_mqd without dqm lock hold, that causes double free of
mqd_mem_obj. Fix it by using a tmp pointer.
We need walk through the queues_list with dqm lock hold. Otherwise hit
list corruption.
Signed-off-by: xinhui pan <xinhui.pan at amd.com>
---
.../drm/amd/amdkfd/kfd_device_queue_manager.c | 17 +++++++++++++----
1 file changed, 13 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
index e6366b408420..575c895fc241 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -1484,6 +1484,7 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
struct mqd_manager *mqd_mgr;
uint64_t sdma_val = 0;
struct kfd_process_device *pdd = qpd_to_pdd(qpd);
+ void *ptr = NULL;
/* Get the SDMA queue stats */
if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
@@ -1543,10 +1544,12 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm,
pr_debug("Total of %d queues are accountable so far\n",
dqm->total_queue_count);
+ swap(ptr, q->mqd_mem_obj);
dqm_unlock(dqm);
/* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */
- mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
+ if (ptr)
+ mqd_mgr->free_mqd(mqd_mgr, q->mqd, ptr);
return retval;
@@ -1709,6 +1712,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
enum kfd_unmap_queues_filter filter =
KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
bool found = false;
+ void *ptr = NULL;
retval = 0;
@@ -1737,8 +1741,6 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
qpd->mapped_gws_queue = false;
}
}
-
- dqm->total_queue_count--;
}
/* Unregister process */
@@ -1770,13 +1772,20 @@ static int process_termination_cpsch(struct device_queue_manager *dqm,
/* Lastly, free mqd resources.
* Do free_mqd() after dqm_unlock to avoid circular locking.
*/
+ dqm_lock(dqm);
list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
q->properties.type)];
list_del(&q->list);
qpd->queue_count--;
- mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
+ dqm->total_queue_count--;
+ swap(ptr, q->mqd_mem_obj);
+ dqm_unlock(dqm);
+ if (ptr)
+ mqd_mgr->free_mqd(mqd_mgr, q->mqd, ptr);
+ dqm_lock(dqm);
}
+ dqm_unlock(dqm);
return retval;
}
--
2.25.1
More information about the amd-gfx
mailing list