I think this could also be fixed by not taking the process_info lock in
svm_range_restore_work and svm_range_set_attr. I'm not even sure why
we're taking this lock in the SVM code. I think that was copied from the
restore workers in amdgpu_amdkfd_gpuvm.c because there it's used to
protect the BO lists. But I think in the case of SVM, the equivalent
lists are protected by the svms->lock.
Regards,
Felix
On 2022-07-07 17:23, Philip Yang wrote:
Move process queues cleanup to deferred work kfd_process_wq_release, to
avoid potential deadlock circular locking warning:
WARNING: possible circular locking dependency detected
the existing dependency chain (in reverse order) is:
-> #2
((work_completion)(&svms->deferred_list_work)){+.+.}-{0:0}:
__flush_work+0x343/0x4a0
svm_range_list_lock_and_flush_work+0x39/0xc0
svm_range_set_attr+0xe8/0x1080 [amdgpu]
kfd_ioctl+0x19b/0x600 [amdgpu]
__x64_sys_ioctl+0x81/0xb0
do_syscall_64+0x34/0x80
entry_SYSCALL_64_after_hwframe+0x44/0xae
-> #1 (&info->lock#2){+.+.}-{3:3}:
__mutex_lock+0xa4/0x940
amdgpu_amdkfd_gpuvm_acquire_process_vm+0x2e3/0x590
kfd_process_device_init_vm+0x61/0x200 [amdgpu]
kfd_ioctl_acquire_vm+0x83/0xb0 [amdgpu]
kfd_ioctl+0x19b/0x600 [amdgpu]
__x64_sys_ioctl+0x81/0xb0
do_syscall_64+0x34/0x80
entry_SYSCALL_64_after_hwframe+0x44/0xae
-> #0 (&process->mutex){+.+.}-{3:3}:
__lock_acquire+0x1365/0x23d0
lock_acquire+0xc9/0x2e0
__mutex_lock+0xa4/0x940
kfd_process_notifier_release+0x96/0xe0 [amdgpu]
__mmu_notifier_release+0x94/0x210
exit_mmap+0x35/0x1f0
mmput+0x63/0x120
svm_range_deferred_list_work+0x177/0x2c0 [amdgpu]
process_one_work+0x2a4/0x600
worker_thread+0x39/0x3e0
kthread+0x16d/0x1a0
Possible unsafe locking scenario:
CPU0 CPU1
---- ----
lock((work_completion)(&svms->deferred_list_work));
lock(&info->lock#2);
lock((work_completion)(&svms->deferred_list_work));
lock(&process->mutex);
Signed-off-by: Philip Yang <Philip.Yang@xxxxxxx>
---
drivers/gpu/drm/amd/amdkfd/kfd_process.c | 21 +++++++++------------
1 file changed, 9 insertions(+), 12 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
index fc38a4d81420..3c9cf9bdb63f 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c
@@ -1115,6 +1115,15 @@ static void kfd_process_wq_release(struct work_struct *work)
struct kfd_process *p = container_of(work, struct kfd_process,
release_work);
+ kfd_process_dequeue_from_all_devices(p);
+ pqm_uninit(&p->pqm);
+
+ /* Signal the eviction fence after user mode queues are
+ * destroyed. This allows any BOs to be freed without
+ * triggering pointless evictions or waiting for fences.
+ */
+ dma_fence_signal(p->ef);
+
kfd_process_remove_sysfs(p);
kfd_iommu_unbind_process(p);
@@ -1179,20 +1188,8 @@ static void kfd_process_notifier_release(struct mmu_notifier *mn,
cancel_delayed_work_sync(&p->eviction_work);
cancel_delayed_work_sync(&p->restore_work);
- mutex_lock(&p->mutex);
-
- kfd_process_dequeue_from_all_devices(p);
- pqm_uninit(&p->pqm);
-
/* Indicate to other users that MM is no longer valid */
p->mm = NULL;
- /* Signal the eviction fence after user mode queues are
- * destroyed. This allows any BOs to be freed without
- * triggering pointless evictions or waiting for fences.
- */
- dma_fence_signal(p->ef);
-
- mutex_unlock(&p->mutex);
mmu_notifier_put(&p->mmu_notifier);
}
