Skip to contentSkip to conversation

drm/amdgpu: Fix one use-after-free of VM

PATCH In review xinhui pan 3 participants 4 messages Apr 14, 2022

Subsystems (2)
lore
Share & tools
lore

xinhui pan

author
#0 lore raw

PATCH v2 1/1 drm/amdgpu: Fix one use-after-free of VM

VM might already be freed when amdgpu_vm_tlb_seq_cb() is called. We see the calltrace below.

Fix it by keeping the last flush fence around and wait for it to signal

BUG kmalloc-4k (Not tainted): Poison overwritten

0xffff9c88630414e8-0xffff9c88630414e8 @offset=5352. First byte 0x6c instead of 0x6b Allocated in amdgpu_driver_open_kms+0x9d/0x360 [amdgpu] age=44 cpu=0 pid=2343

 __slab_alloc.isra.0+0x4f/0x90
 kmem_cache_alloc_trace+0x6b8/0x7a0
 amdgpu_driver_open_kms+0x9d/0x360 [amdgpu]
 drm_file_alloc+0x222/0x3e0 [drm]
 drm_open+0x11d/0x410 [drm]
Freed in amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu] age=22 cpu=1
pid=2485
 kfree+0x4a2/0x580
 amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu]
 drm_file_free+0x24e/0x3c0 [drm]
 drm_close_helper.isra.0+0x90/0xb0 [drm]
 drm_release+0x97/0x1a0 [drm]
 __fput+0xb6/0x280
 ____fput+0xe/0x10
 task_work_run+0x64/0xb0
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 17 ++++++++++++++---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  1 +
 2 files changed, 15 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +14 -3
index 645ce28277c2..cd5aa7edd451 100644
@@ -932,9 +932,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 
 	if (flush_tlb || params.table_freed) {
 		tlb_cb->vm = vm;
-		if (!fence || !*fence ||
-		    dma_fence_add_callback(*fence, &tlb_cb->cb,
-					   amdgpu_vm_tlb_seq_cb))
+		if (fence && *fence &&
+		    !dma_fence_add_callback(*fence, &tlb_cb->cb,
+					   amdgpu_vm_tlb_seq_cb)) {
+			dma_fence_put(vm->last_tlb_flush);
+			vm->last_tlb_flush = dma_fence_get(*fence);
+		} else
 			amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);
 		tlb_cb = NULL;
 	}
@@ -2094,6 +2097,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 		vm->update_funcs = &amdgpu_vm_sdma_funcs;
 	vm->last_update = NULL;
 	vm->last_unlocked = dma_fence_get_stub();
+	vm->last_tlb_flush = dma_fence_get_stub();
 
 	mutex_init(&vm->eviction_lock);
 	vm->evicting = false;
@@ -2132,6 +2136,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	vm->root.bo = NULL;
 
 error_free_delayed:
+	dma_fence_put(vm->last_tlb_flush);
 	dma_fence_put(vm->last_unlocked);
 	drm_sched_entity_destroy(&vm->delayed);
 
@@ -2248,6 +2253,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	struct amdgpu_bo_va_mapping *mapping, *tmp;
 	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
 	struct amdgpu_bo *root;
+	unsigned long flags;
 	int i;
 
 	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
@@ -2257,6 +2263,11 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	amdgpu_vm_set_pasid(adev, vm, 0);
 	dma_fence_wait(vm->last_unlocked, false);
 	dma_fence_put(vm->last_unlocked);
+	dma_fence_wait(vm->last_tlb_flush, false);
+	/* Make sure that all fence callbacks have completed */
+	spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
+	spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
+	dma_fence_put(vm->last_tlb_flush);
 
 	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
 		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
index 1a814fbffff8..6b06a214f05f 100644
@@ -286,6 +286,7 @@ struct amdgpu_vm {
 
 	/* Last finished delayed update */
 	atomic64_t		tlb_seq;
+	struct dma_fence	*last_tlb_flush;
 
 	/* Last unlocked submission to the scheduler entities */
 	struct dma_fence	*last_unlocked;
Signature · 1 line
2.25.1
#1 lore raw

Dear Xinhui,

Thank you for rerolling the patch.

Am 14.04.22 um 07:03 schrieb xinhui pan:

Quoted 4 lines

VM might already be freed when amdgpu_vm_tlb_seq_cb() is called. We see the calltrace below.

Fix it by keeping the last flush fence around and wait for it to signal

Nit: Please add a dot/period to the end of sentences.

Quoted 6 lines

BUG kmalloc-4k (Not tainted): Poison overwritten

0xffff9c88630414e8-0xffff9c88630414e8 @offset=5352. First byte 0x6c instead of 0x6b Allocated in amdgpu_driver_open_kms+0x9d/0x360 [amdgpu] age=44 cpu=0 pid=2343

Quoted 15 lines
  __slab_alloc.isra.0+0x4f/0x90
  kmem_cache_alloc_trace+0x6b8/0x7a0
  amdgpu_driver_open_kms+0x9d/0x360 [amdgpu]
  drm_file_alloc+0x222/0x3e0 [drm]
  drm_open+0x11d/0x410 [drm]
Freed in amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu] age=22 cpu=1
pid=2485
  kfree+0x4a2/0x580
  amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu]
  drm_file_free+0x24e/0x3c0 [drm]
  drm_close_helper.isra.0+0x90/0xb0 [drm]
  drm_release+0x97/0x1a0 [drm]
  __fput+0xb6/0x280
  ____fput+0xe/0x10
  task_work_run+0x64/0xb0

The v2 annotation is missing.

Quoted 15 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 17 ++++++++++++++---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  1 +
  2 files changed, 15 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 645ce28277c2..cd5aa7edd451 100644
@@ -932,9 +932,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 
  	if (flush_tlb || params.table_freed) {
  		tlb_cb->vm = vm;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 941

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -3
-		if (!fence || !*fence ||
-		    dma_fence_add_callback(*fence, &tlb_cb->cb,
-					   amdgpu_vm_tlb_seq_cb))
+		if (fence && *fence &&
+		    !dma_fence_add_callback(*fence, &tlb_cb->cb,
+					   amdgpu_vm_tlb_seq_cb)) {
+			dma_fence_put(vm->last_tlb_flush);
+			vm->last_tlb_flush = dma_fence_get(*fence);
+		} else
  			amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);

The Linux kernel coding style uses braces for all branches of a conditional statement, if one branch uses braces. [1]

Kind regards,

Paul

Quoted 43 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -0
  		tlb_cb = NULL;
  	}
@@ -2094,6 +2097,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  		vm->update_funcs = &amdgpu_vm_sdma_funcs;
  	vm->last_update = NULL;
  	vm->last_unlocked = dma_fence_get_stub();
+	vm->last_tlb_flush = dma_fence_get_stub();
 
  	mutex_init(&vm->eviction_lock);
  	vm->evicting = false;
@@ -2132,6 +2136,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  	vm->root.bo = NULL;
 
  error_free_delayed:
+	dma_fence_put(vm->last_tlb_flush);
  	dma_fence_put(vm->last_unlocked);
  	drm_sched_entity_destroy(&vm->delayed);
 
@@ -2248,6 +2253,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  	struct amdgpu_bo_va_mapping *mapping, *tmp;
  	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
  	struct amdgpu_bo *root;
+	unsigned long flags;
  	int i;
 
  	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
@@ -2257,6 +2263,11 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  	amdgpu_vm_set_pasid(adev, vm, 0);
  	dma_fence_wait(vm->last_unlocked, false);
  	dma_fence_put(vm->last_unlocked);
+	dma_fence_wait(vm->last_tlb_flush, false);
+	/* Make sure that all fence callbacks have completed */
+	spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
+	spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
+	dma_fence_put(vm->last_tlb_flush);
 
  	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
  		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +0 -0
index 1a814fbffff8..6b06a214f05f 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h struct amdgpu_vm { · line 292

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
@@ -286,6 +286,7 @@ struct amdgpu_vm {
 
  	/* Last finished delayed update */
  	atomic64_t		tlb_seq;
+	struct dma_fence	*last_tlb_flush;
 
  	/* Last unlocked submission to the scheduler entities */
  	struct dma_fence	*last_unlocked;

[1]: https://www.kernel.org/doc/html/latest/process/coding-style.html#placing-braces-and-spaces

#2 lore raw

Dear Xinhui,

Am 14.04.22 um 07:45 schrieb Paul Menzel:

Quoted 53 lines

Thank you for rerolling the patch.

Am 14.04.22 um 07:03 schrieb xinhui pan:

Quoted 4 lines

VM might already be freed when amdgpu_vm_tlb_seq_cb() is called. We see the calltrace below.

Fix it by keeping the last flush fence around and wait for it to signal

Nit: Please add a dot/period to the end of sentences.

Quoted 6 lines

BUG kmalloc-4k (Not tainted): Poison overwritten

0xffff9c88630414e8-0xffff9c88630414e8 @offset=5352. First byte 0x6c instead of 0x6b Allocated in amdgpu_driver_open_kms+0x9d/0x360 [amdgpu] age=44 cpu=0 pid=2343

Quoted 15 lines
  __slab_alloc.isra.0+0x4f/0x90
  kmem_cache_alloc_trace+0x6b8/0x7a0
  amdgpu_driver_open_kms+0x9d/0x360 [amdgpu]
  drm_file_alloc+0x222/0x3e0 [drm]
  drm_open+0x11d/0x410 [drm]
Freed in amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu] age=22 cpu=1
pid=2485
  kfree+0x4a2/0x580
  amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu]
  drm_file_free+0x24e/0x3c0 [drm]
  drm_close_helper.isra.0+0x90/0xb0 [drm]
  drm_release+0x97/0x1a0 [drm]
  __fput+0xb6/0x280
  ____fput+0xe/0x10
  task_work_run+0x64/0xb0

The v2 annotation is missing.

Quoted 19 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 17 ++++++++++++++--- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  1 +

  2 files changed, 15 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -3
index 645ce28277c2..cd5aa7edd451 100644
@@ -932,9 +932,12 @@ int amdgpu_vm_update_range(struct amdgpu_device 
 *adev, struct amdgpu_vm *vm,
       if (flush_tlb || params.table_freed) {
           tlb_cb->vm = vm;
-        if (!fence || !*fence ||
-            dma_fence_add_callback(*fence, &tlb_cb->cb,
-                       amdgpu_vm_tlb_seq_cb))
+        if (fence && *fence &&
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 941

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +5 -0
+            !dma_fence_add_callback(*fence, &tlb_cb->cb,
+                       amdgpu_vm_tlb_seq_cb)) {
+            dma_fence_put(vm->last_tlb_flush);
+            vm->last_tlb_flush = dma_fence_get(*fence);
+        } else
               amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);

The Linux kernel coding style uses braces for all branches of a conditional statement, if one branch uses braces. [1]

By the way scripts/checkpatch.pl --strict … would also have found this:

```
$ scripts/checkpatch.pl --strict 0001-drm-amdgpu-Fix-one-use-after-free-of-VM.patch
CHECK: braces {} should be used on all arms of this statement
#53: FILE: drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:935:
+               if (fence && *fence &&
[...]
+               } else
[...]

CHECK: Unbalanced braces around else statement
#58: FILE: drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:940:
+               } else

total: 0 errors, 0 warnings, 2 checks, 54 lines checked

NOTE: For some of the reported defects, checkpatch may be able to
      mechanically convert to the typical style using --fix or 
--fix-inplace.

0001-drm-amdgpu-Fix-one-use-after-free-of-VM.patch has style problems, please review.
```

Kind regards,

Paul

Quoted 50 lines
Quoted 42 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +1 -0
           tlb_cb = NULL;
       }
@@ -2094,6 +2097,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, 
 struct amdgpu_vm *vm)
           vm->update_funcs = &amdgpu_vm_sdma_funcs;
       vm->last_update = NULL;
       vm->last_unlocked = dma_fence_get_stub();
+    vm->last_tlb_flush = dma_fence_get_stub();
       mutex_init(&vm->eviction_lock);
       vm->evicting = false;
@@ -2132,6 +2136,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, 
 struct amdgpu_vm *vm)
       vm->root.bo = NULL;

  error_free_delayed:
+    dma_fence_put(vm->last_tlb_flush);
      dma_fence_put(vm->last_unlocked);
      drm_sched_entity_destroy(&vm->delayed);

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
@@ -2248,6 +2253,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, 
 struct amdgpu_vm *vm)
       struct amdgpu_bo_va_mapping *mapping, *tmp;
       bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
       struct amdgpu_bo *root;
+    unsigned long flags;
       int i;
       amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
@@ -2257,6 +2263,11 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, 
 struct amdgpu_vm *vm)
       amdgpu_vm_set_pasid(adev, vm, 0);
       dma_fence_wait(vm->last_unlocked, false);
       dma_fence_put(vm->last_unlocked);
+    dma_fence_wait(vm->last_tlb_flush, false);
+    /* Make sure that all fence callbacks have completed */
+    spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
+    spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
+    dma_fence_put(vm->last_tlb_flush);
       list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
           if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +0 -0
index 1a814fbffff8..6b06a214f05f 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h struct amdgpu_vm { · line 290

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
@@ -286,6 +286,7 @@ struct amdgpu_vm {
       /* Last finished delayed update */
       atomic64_t        tlb_seq;
+    struct dma_fence    *last_tlb_flush;
       /* Last unlocked submission to the scheduler entities */
       struct dma_fence    *last_unlocked;

[1]: https://www.kernel.org/doc/html/latest/process/coding-style.html#placing-braces-and-spaces

maintainer
#3 lore raw
In reply to xinhui pan

Am 14.04.22 um 07:03 schrieb xinhui pan:

Quoted 41 lines

VM might already be freed when amdgpu_vm_tlb_seq_cb() is called. We see the calltrace below.

Fix it by keeping the last flush fence around and wait for it to signal

BUG kmalloc-4k (Not tainted): Poison overwritten

0xffff9c88630414e8-0xffff9c88630414e8 @offset=5352. First byte 0x6c instead of 0x6b Allocated in amdgpu_driver_open_kms+0x9d/0x360 [amdgpu] age=44 cpu=0 pid=2343

  __slab_alloc.isra.0+0x4f/0x90
  kmem_cache_alloc_trace+0x6b8/0x7a0
  amdgpu_driver_open_kms+0x9d/0x360 [amdgpu]
  drm_file_alloc+0x222/0x3e0 [drm]
  drm_open+0x11d/0x410 [drm]
Freed in amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu] age=22 cpu=1
pid=2485
  kfree+0x4a2/0x580
  amdgpu_driver_postclose_kms+0x3e9/0x550 [amdgpu]
  drm_file_free+0x24e/0x3c0 [drm]
  drm_close_helper.isra.0+0x90/0xb0 [drm]
  drm_release+0x97/0x1a0 [drm]
  __fput+0xb6/0x280
  ____fput+0xe/0x10
  task_work_run+0x64/0xb0
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 17 ++++++++++++++---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  1 +
  2 files changed, 15 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 645ce28277c2..cd5aa7edd451 100644
@@ -932,9 +932,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 
  	if (flush_tlb || params.table_freed) {
  		tlb_cb->vm = vm;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 941

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -3
-		if (!fence || !*fence ||
-		    dma_fence_add_callback(*fence, &tlb_cb->cb,
-					   amdgpu_vm_tlb_seq_cb))
+		if (fence && *fence &&
+		    !dma_fence_add_callback(*fence, &tlb_cb->cb,
+					   amdgpu_vm_tlb_seq_cb)) {
+			dma_fence_put(vm->last_tlb_flush);
+			vm->last_tlb_flush = dma_fence_get(*fence);
+		} else
  			amdgpu_vm_tlb_seq_cb(NULL, &tlb_cb->cb);

One style nit pick here: When one side of an if/else has {}, the other side should have {} as well.

Apart from that the patch is Reviewed-by: Christian König <[email protected]>

Thanks,
Christian.

Quoted 50 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -0
  		tlb_cb = NULL;
  	}
@@ -2094,6 +2097,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  		vm->update_funcs = &amdgpu_vm_sdma_funcs;
  	vm->last_update = NULL;
  	vm->last_unlocked = dma_fence_get_stub();
+	vm->last_tlb_flush = dma_fence_get_stub();
 
  	mutex_init(&vm->eviction_lock);
  	vm->evicting = false;
@@ -2132,6 +2136,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  	vm->root.bo = NULL;
 
  error_free_delayed:
+	dma_fence_put(vm->last_tlb_flush);
  	dma_fence_put(vm->last_unlocked);
  	drm_sched_entity_destroy(&vm->delayed);
 
@@ -2248,6 +2253,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  	struct amdgpu_bo_va_mapping *mapping, *tmp;
  	bool prt_fini_needed = !!adev->gmc.gmc_funcs->set_prt;
  	struct amdgpu_bo *root;
+	unsigned long flags;
  	int i;
 
  	amdgpu_amdkfd_gpuvm_destroy_cb(adev, vm);
@@ -2257,6 +2263,11 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
  	amdgpu_vm_set_pasid(adev, vm, 0);
  	dma_fence_wait(vm->last_unlocked, false);
  	dma_fence_put(vm->last_unlocked);
+	dma_fence_wait(vm->last_tlb_flush, false);
+	/* Make sure that all fence callbacks have completed */
+	spin_lock_irqsave(vm->last_tlb_flush->lock, flags);
+	spin_unlock_irqrestore(vm->last_tlb_flush->lock, flags);
+	dma_fence_put(vm->last_tlb_flush);
 
  	list_for_each_entry_safe(mapping, tmp, &vm->freed, list) {
  		if (mapping->flags & AMDGPU_PTE_PRT && prt_fini_needed) {
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
index 1a814fbffff8..6b06a214f05f 100644
@@ -286,6 +286,7 @@ struct amdgpu_vm {
 
  	/* Last finished delayed update */
  	atomic64_t		tlb_seq;
+	struct dma_fence	*last_tlb_flush;
 
  	/* Last unlocked submission to the scheduler entities */
  	struct dma_fence	*last_unlocked;