Skip to contentSkip to conversation

drm/ttm, amdgpu: fix use-after-free in recent deadlock fix

PATCH In review Nicolai Hähnle 5 participants 8 messages Feb 14, 2017 3 messages not in the archive

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover drm/ttm, amdgpu: fix use-after-free in recent deadlock fix

Hi all,

based on my current theory on how a deadlock could happen in the buffer allocation code, these two patches should fix the deadlock without having a use-after-free.

I'm still working on a way to clean up the ttm_bo_init sequence overall, but I'm separating these two out for a hopefully quick fix, since in my book use-after-frees are worse than deadlocks.

Samuel, I'd very much appreciate if you could check that the deadlocks in Hitman remain fixed with these two patches.

Thanks,
Nicolai

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): 20170214091845.1225-1-nhaehnle-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#2 lore raw

PATCH 1/2 drm/ttm: never add BO that failed to validate to the LRU list

From: Nicolai Hähnle <[email protected]>

Fixes a potential race condition in amdgpu that looks as follows:

Task 1: attempt ttm_bo_init, but ttm_bo_validate fails Task 1: add BO to global list anyway
Task 2: grabs hold of the BO, waits on its reservation lock Task 1: releases its reference of the BO; never gives up the
reservation lock

The patch "drm/amdgpu: fix a potential deadlock in amdgpu_bo_create_restricted()" attempts to fix that by releasing the reservation lock in amdgpu code; unfortunately, it introduces a use-after-free when this race _doesn't_ happen.

This patch should fix the race properly by never adding the BO to the global list in the first place.

Cc Samuel Pitoiset <[email protected]>
Cc zhoucm1 <[email protected]>
---
 drivers/gpu/drm/ttm/ttm_bo.c | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +7 -5
index 239a957..76bee42 100644
@@ -1215,18 +1215,20 @@ int ttm_bo_init(struct ttm_bo_device *bdev,
 	if (likely(!ret))
 		ret = ttm_bo_validate(bo, placement, interruptible, false);
 
-	if (!resv) {
+	if (!resv)
 		ttm_bo_unreserve(bo);
 
-	} else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
+	if (unlikely(ret)) {
+		ttm_bo_unref(&bo);
+		return ret;
+	}
+
+	if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
 		spin_lock(&bo->glob->lru_lock);
 		ttm_bo_add_to_lru(bo);
 		spin_unlock(&bo->glob->lru_lock);
 	}
 
-	if (unlikely(ret))
-		ttm_bo_unref(&bo);
-
 	return ret;
 }
 EXPORT_SYMBOL(ttm_bo_init);
Signature · 6 lines
2.9.3

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
#3 lore raw

Re: [PATCH 1/2] drm/ttm: never add BO that failed to validate to the LRU list

Am 14.02.2017 um 10:18 schrieb Nicolai Hähnle:

Quoted 15 lines

From: Nicolai Hähnle <[email protected]>

Fixes a potential race condition in amdgpu that looks as follows:

Task 1: attempt ttm_bo_init, but ttm_bo_validate fails Task 1: add BO to global list anyway
Task 2: grabs hold of the BO, waits on its reservation lock Task 1: releases its reference of the BO; never gives up the
reservation lock

The patch "drm/amdgpu: fix a potential deadlock in amdgpu_bo_create_restricted()" attempts to fix that by releasing the reservation lock in amdgpu code; unfortunately, it introduces a use-after-free when this race _doesn't_ happen.

Quoted 6 lines

This patch should fix the race properly by never adding the BO to the global list in the first place.

Cc Samuel Pitoiset <[email protected]>
Cc zhoucm1 <[email protected]>

NAK, that is actually not correct either.

The previous implementation doesn't have an use after free, but actually a memory leak.

I completely agree that adding the BO to the LRU in case of an error is incorrect, but ttm_bo_add_to_lru() will add some references to the BO.

So the following ttm_bo_unref() will just drop the initial reference.

Let's clean up this mess by moving the freeing of the BO in case of an error to the caller.

Regards,
Christian.

Quoted 24 lines
---
  drivers/gpu/drm/ttm/ttm_bo.c | 12 +++++++-----
  1 file changed, 7 insertions(+), 5 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +7 -2
index 239a957..76bee42 100644
@@ -1215,18 +1215,20 @@ int ttm_bo_init(struct ttm_bo_device *bdev,
  	if (likely(!ret))
  		ret = ttm_bo_validate(bo, placement, interruptible, false);
 
-	if (!resv) {
+	if (!resv)
  		ttm_bo_unreserve(bo);
 
-	} else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
+	if (unlikely(ret)) {
+		ttm_bo_unref(&bo);
+		return ret;
+	}
+
+	if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
  		spin_lock(&bo->glob->lru_lock);

drivers/gpu/drm/ttm/ttm_bo.c int ttm_bo_init(struct ttm_bo_device *bdev, · line 1234

Quoted 10 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -3
  		ttm_bo_add_to_lru(bo);
  		spin_unlock(&bo->glob->lru_lock);
  	}
 
-	if (unlikely(ret))
-		ttm_bo_unref(&bo);
-
  	return ret;
  }
  EXPORT_SYMBOL(ttm_bo_init);

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

A message that is not in the archive (sent off-list or not yet received): ef1d2448-4296-5eaf-a852-5000bd16a947-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#5 lore raw

Re: [PATCH 1/2] drm/ttm: never add BO that failed to validate to the LRU list

Quoted 23 lines · Christian König

Am 14.02.2017 um 10:18 schrieb Nicolai Hähnle:

Quoted 15 lines

From: Nicolai Hähnle <[email protected]>

Fixes a potential race condition in amdgpu that looks as follows:

Task 1: attempt ttm_bo_init, but ttm_bo_validate fails Task 1: add BO to global list anyway
Task 2: grabs hold of the BO, waits on its reservation lock Task 1: releases its reference of the BO; never gives up the
reservation lock

The patch "drm/amdgpu: fix a potential deadlock in amdgpu_bo_create_restricted()" attempts to fix that by releasing the reservation lock in amdgpu code; unfortunately, it introduces a use-after-free when this race _doesn't_ happen.

Quoted 6 lines

This patch should fix the race properly by never adding the BO to the global list in the first place.

Cc Samuel Pitoiset <[email protected]>
Cc zhoucm1 <[email protected]>
Quoted 7 lines

NAK, that is actually not correct either.

The previous implementation doesn't have an use after free, but actually a memory leak.

I completely agree that adding the BO to the LRU in case of an error is incorrect, but ttm_bo_add_to_lru() will add some references to the BO.

Yes, but those are bo->list_krefs, not bo->krefs. ttm_bo_unref will put the bo->kref, which should lead to ttm_bo_release. This will then add the buffer to the ddestroy list and eventually put all the list_krefs. (Or possibly, if the ttm_bo_wait in ttm_bo_cleanup_refs_or_queue fails, it will explicitly delete the buffer from the LRU, which should amount to the same, I think).

Or perhaps I'm still missing something, I'm not too familiar with the whole BO reference counting.

Quoted 4 lines

So the following ttm_bo_unref() will just drop the initial reference.

Let's clean up this mess by moving the freeing of the BO in case of an error to the caller.

Yes, see my other patches.

Thanks,
Nicolai

Quoted 27 lines

Regards,
Christian.

Quoted 23 lines
---
  drivers/gpu/drm/ttm/ttm_bo.c | 12 +++++++-----
  1 file changed, 7 insertions(+), 5 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +7 -0
index 239a957..76bee42 100644
@@ -1215,18 +1215,20 @@ int ttm_bo_init(struct ttm_bo_device *bdev,
      if (likely(!ret))
          ret = ttm_bo_validate(bo, placement, interruptible, false);
  -    if (!resv) {
+    if (!resv)
          ttm_bo_unreserve(bo);
  -    } else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
+    if (unlikely(ret)) {
+        ttm_bo_unref(&bo);
+        return ret;
+    }
+
+    if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
          spin_lock(&bo->glob->lru_lock);
          ttm_bo_add_to_lru(bo);
Quoted 10 lines

drivers/gpu/drm/ttm/ttm_bo.c int ttm_bo_init(struct ttm_bo_device *bdev, · line 1234

Quoted 8 lines
drivers/gpu/drm/ttm/ttm_bo.c +0 -2
          spin_unlock(&bo->glob->lru_lock);
      }
  -    if (unlikely(ret))
-        ttm_bo_unref(&bo);
-
      return ret;
  }
  EXPORT_SYMBOL(ttm_bo_init);

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 3ac4923c-61b9-8ee9-e9cb-d7a60fc51dbf-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#7 lore raw

Re: [PATCH 1/2] drm/ttm: never add BO that failed to validate to the LRU list

Am 14.02.2017 um 12:37 schrieb Nicolai Hähnle:

Quoted 29 lines · Christian König
Quoted 28 lines

Am 14.02.2017 um 10:18 schrieb Nicolai Hähnle:

Quoted 15 lines

From: Nicolai Hähnle <[email protected]>

Fixes a potential race condition in amdgpu that looks as follows:

Task 1: attempt ttm_bo_init, but ttm_bo_validate fails Task 1: add BO to global list anyway
Task 2: grabs hold of the BO, waits on its reservation lock Task 1: releases its reference of the BO; never gives up the
reservation lock

The patch "drm/amdgpu: fix a potential deadlock in amdgpu_bo_create_restricted()" attempts to fix that by releasing the reservation lock in amdgpu code; unfortunately, it introduces a use-after-free when this race _doesn't_ happen.

Quoted 6 lines

This patch should fix the race properly by never adding the BO to the global list in the first place.

Cc Samuel Pitoiset <[email protected]>
Cc zhoucm1 <[email protected]>

NAK, that is actually not correct either.

The previous implementation doesn't have an use after free, but actually a memory leak.

Quoted 6 lines
Quoted 2 lines

I completely agree that adding the BO to the LRU in case of an error is incorrect, but ttm_bo_add_to_lru() will add some references to the BO.

Yes, but those are bo->list_krefs, not bo->krefs. ttm_bo_unref will put the bo->kref, which should lead to ttm_bo_release. This will then add the buffer to the ddestroy list and eventually put all the list_krefs. (Or possibly, if the ttm_bo_wait in ttm_bo_cleanup_refs_or_queue fails, it will explicitly delete the buffer from the LRU, which should amount to the same, I think).

Or perhaps I'm still missing something, I'm not too familiar with the whole BO reference counting.

Ah, yes of course, never mind.

Quoted 8 lines
Quoted 6 lines

So the following ttm_bo_unref() will just drop the initial reference.

Let's clean up this mess by moving the freeing of the BO in case of an error to the caller.

Yes, see my other patches.

In this case the set is Reviewed-by: Christian König <[email protected]>.

Regards,
Christian.

Quoted 43 lines

Thanks,
Nicolai

Quoted 37 lines

Regards,
Christian.

Quoted 31 lines
---
  drivers/gpu/drm/ttm/ttm_bo.c | 12 +++++++-----
  1 file changed, 7 insertions(+), 5 deletions(-)
drivers/gpu/drm/ttm/ttm_bo.c +7 -2
index 239a957..76bee42 100644
@@ -1215,18 +1215,20 @@ int ttm_bo_init(struct ttm_bo_device *bdev,
      if (likely(!ret))
          ret = ttm_bo_validate(bo, placement, interruptible, false);
  -    if (!resv) {
+    if (!resv)
          ttm_bo_unreserve(bo);
  -    } else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
+    if (unlikely(ret)) {
+        ttm_bo_unref(&bo);
+        return ret;
+    }
+
+    if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
          spin_lock(&bo->glob->lru_lock);
          ttm_bo_add_to_lru(bo);
          spin_unlock(&bo->glob->lru_lock);
      }
  -    if (unlikely(ret))
-        ttm_bo_unref(&bo);
-
      return ret;
  }
  EXPORT_SYMBOL(ttm_bo_init);

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#8 lore raw
In reply to a message not in the archive
Quoted 6 lines · Nicolai Hähnle

Hi all,

based on my current theory on how a deadlock could happen in the buffer allocation code, these two patches should fix the deadlock without having a use-after-free.

Quoted 6 lines

I'm still working on a way to clean up the ttm_bo_init sequence overall, but I'm separating these two out for a hopefully quick fix, since in my book use-after-frees are worse than deadlocks.

Samuel, I'd very much appreciate if you could check that the deadlocks in Hitman remain fixed with these two patches.

Sure, I will check.

Quoted 9 lines

Thanks,
Nicolai

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#9 lore raw
In reply to a message not in the archive
Quoted 6 lines · Nicolai Hähnle

Hi all,

based on my current theory on how a deadlock could happen in the buffer allocation code, these two patches should fix the deadlock without having a use-after-free.

Quoted 6 lines

I'm still working on a way to clean up the ttm_bo_init sequence overall, but I'm separating these two out for a hopefully quick fix, since in my book use-after-frees are worse than deadlocks.

Samuel, I'd very much appreciate if you could check that the deadlocks in Hitman remain fixed with these two patches.

Looks good. Thanks!

Quoted 9 lines

Thanks,
Nicolai

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#10 lore raw

PATCH 2/2 Revert "drm/amdgpu: fix a potential deadlock in amdgpu_bo_create_restricted()"

In reply to Nicolai Hähnle

From: Nicolai Hähnle <[email protected]>

This reverts commit 38fc4856ad98f230bc91da0421dec69e4aee40f8, which introduces a use-after-free.

The underlying bug should be properly fixed with "drm/ttm: never add BO that failed to validate to the LRU list".

Cc Samuel Pitoiset <[email protected]>
Cc zhoucm1 <[email protected]>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c +1 -4
index 556236a..d1ef1d0 100644
@@ -403,11 +403,8 @@ int amdgpu_bo_create_restricted(struct amdgpu_device *adev,
 			&bo->placement, page_align, !kernel, NULL,
 			acc_size, sg, resv ? resv : &bo->tbo.ttm_resv,
 			&amdgpu_ttm_bo_destroy);
-	if (unlikely(r != 0)) {
-		if (!resv)
-			ww_mutex_unlock(&bo->tbo.resv->lock);
+	if (unlikely(r != 0))
 		return r;
-	}
 
 	bo->tbo.priority = ilog2(bo->tbo.num_pages);
 	if (kernel)
Signature · 6 lines
2.9.3

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel