[1/3] drm/vmwgfx: Don't use memory accounting for kernel-side fence objects

Commit Message

Thomas Hellstrom Dec. 2, 2014, 11:59 a.m. UTC

Kernel side fence objects are used when unbinding resources and may thus be
created as part of a memory reclaim operation. This might trigger recursive
memory reclaims and result in the kernel running out of stack space.

So a simple way out is to avoid accounting of these fence objects.
In principle this is OK since while user-space can trigger the creation of
such objects, it can't really hold on to them. However, their lifetime is
quite long, so some form of accounting should perhaps be implemented in the
future.

Fixes kernel crashes when running, for example viewperf11 ensight-04 test 3
with low system memory settings.

Cc: <[e-mail]>
Signed-off-by: Thomas Hellstrom <[e-mail]>
Reviewed-by: Jakob Bornecrantz <[e-mail]>
Reviewed-by: Sinclair Yeh <[e-mail]>
---
 drivers/gpu/drm/vmwgfx/vmwgfx_fence.c | 22 ++--------------------
 1 file changed, 2 insertions(+), 20 deletions(-)

Comments

Dave Airlie Dec. 3, 2014, 1:06 a.m. UTC | #1

On 2 December 2014 at 21:59, Thomas Hellstrom <[e-mail]> wrote:
> Kernel side fence objects are used when unbinding resources and may thus be
> created as part of a memory reclaim operation. This might trigger recursive
> memory reclaims and result in the kernel running out of stack space.
>
> So a simple way out is to avoid accounting of these fence objects.
> In principle this is OK since while user-space can trigger the creation of
> such objects, it can't really hold on to them. However, their lifetime is
> quite long, so some form of accounting should perhaps be implemented in the
> future.
>
> Fixes kernel crashes when running, for example viewperf11 ensight-04 test 3
> with low system memory settings.

are these 3 intended for fixes? can you send me a git pull for them
soon, so they don't miss Linus release.

Dave.

Thomas Hellstrom Dec. 3, 2014, 6:30 a.m. UTC | #2

On 12/03/2014 02:06 AM, Dave Airlie wrote:
> On 2 December 2014 at 21:59, Thomas Hellstrom <[e-mail]> wrote:
>> Kernel side fence objects are used when unbinding resources and may thus be
>> created as part of a memory reclaim operation. This might trigger recursive
>> memory reclaims and result in the kernel running out of stack space.
>>
>> So a simple way out is to avoid accounting of these fence objects.
>> In principle this is OK since while user-space can trigger the creation of
>> such objects, it can't really hold on to them. However, their lifetime is
>> quite long, so some form of accounting should perhaps be implemented in the
>> future.
>>
>> Fixes kernel crashes when running, for example viewperf11 ensight-04 test 3
>> with low system memory settings.
> are these 3 intended for fixes? can you send me a git pull for them
> soon, so they don't miss Linus release.
>
> Dave.
Hi.

Actually, these bugs have been quite long-standing and I feel a bit
uncomfortable including the fixes this late in the release cycle. I'd
rather wait to 3.19 unless you think otherwise.

Thanks,
Thomas

Patch

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
index 197164f..6773938 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c
@@ -545,35 +545,19 @@  void vmw_fence_obj_flush(struct vmw_fence_obj *fence)
 
 static void vmw_fence_destroy(struct vmw_fence_obj *fence)
 {
-	struct vmw_fence_manager *fman = fman_from_fence(fence);
-
 	fence_free(&fence->base);
-
-	/*
-	 * Free kernel space accounting.
-	 */
-	ttm_mem_global_free(vmw_mem_glob(fman->dev_priv),
-			    fman->fence_size);
 }
 
 int vmw_fence_create(struct vmw_fence_manager *fman,
 		     uint32_t seqno,
 		     struct vmw_fence_obj **p_fence)
 {
-	struct ttm_mem_global *mem_glob = vmw_mem_glob(fman->dev_priv);
 	struct vmw_fence_obj *fence;
 	int ret;
 
-	ret = ttm_mem_global_alloc(mem_glob, fman->fence_size,
-				   false, false);
-	if (unlikely(ret != 0))
-		return ret;
-
 	fence = kzalloc(sizeof(*fence), GFP_KERNEL);
-	if (unlikely(fence == NULL)) {
-		ret = -ENOMEM;
-		goto out_no_object;
-	}
+	if (unlikely(fence == NULL))
+		return -ENOMEM;
 
 	ret = vmw_fence_obj_init(fman, fence, seqno,
 				 vmw_fence_destroy);
@@ -585,8 +569,6 @@  int vmw_fence_create(struct vmw_fence_manager *fman,
 
 out_err_init:
 	kfree(fence);
-out_no_object:
-	ttm_mem_global_free(mem_glob, fman->fence_size);
 	return ret;
 }