From: Zack Rusin zack.rusin@broadcom.com
commit e58337100721f3cc0c7424a18730e4f39844934f upstream.
Introduce a version of the fence ops that on release doesn't remove the fence from the pending list, and thus doesn't require a lock to fix poll->fence wait->fence unref deadlocks.
vmwgfx overwrites the wait callback to iterate over the list of all fences and update their status, to do that it holds a lock to prevent the list modifcations from other threads. The fence destroy callback both deletes the fence and removes it from the list of pending fences, for which it holds a lock.
dma buf polling cb unrefs a fence after it's been signaled: so the poll calls the wait, which signals the fences, which are being destroyed. The destruction tries to acquire the lock on the pending fences list which it can never get because it's held by the wait from which it was called.
Old bug, but not a lot of userspace apps were using dma-buf polling interfaces. Fix those, in particular this fixes KDE stalls/deadlock.
Signed-off-by: Zack Rusin zack.rusin@broadcom.com Fixes: 2298e804e96e ("drm/vmwgfx: rework to new fence interface, v2") Cc: Broadcom internal kernel review list bcm-kernel-feedback-list@broadcom.com Cc: dri-devel@lists.freedesktop.org Cc: stable@vger.kernel.org # v6.2+ Reviewed-by: Maaz Mombasawala maaz.mombasawala@broadcom.com Reviewed-by: Martin Krastev martin.krastev@broadcom.com Link: https://patchwork.freedesktop.org/patch/msgid/20240722184313.181318-2-zack.r... [Minor context change fixed] Signed-off-by: Zhi Yang Zhi.Yang@windriver.com Signed-off-by: He Zhe zhe.he@windriver.com --- Build test passed. --- drivers/gpu/drm/vmwgfx/vmwgfx_fence.c | 17 +++++++---------- 1 file changed, 7 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c index 6bacdb7583df..0505f87d13c0 100644 --- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c @@ -32,7 +32,6 @@ #define VMW_FENCE_WRAP (1 << 31)
struct vmw_fence_manager { - int num_fence_objects; struct vmw_private *dev_priv; spinlock_t lock; struct list_head fence_list; @@ -113,13 +112,13 @@ static void vmw_fence_obj_destroy(struct dma_fence *f) { struct vmw_fence_obj *fence = container_of(f, struct vmw_fence_obj, base); - struct vmw_fence_manager *fman = fman_from_fence(fence);
- spin_lock(&fman->lock); - list_del_init(&fence->head); - --fman->num_fence_objects; - spin_unlock(&fman->lock); + if (!list_empty(&fence->head)) { + spin_lock(&fman->lock); + list_del_init(&fence->head); + spin_unlock(&fman->lock); + } fence->destroy(fence); }
@@ -250,7 +249,6 @@ static const struct dma_fence_ops vmw_fence_ops = { .release = vmw_fence_obj_destroy, };
- /** * Execute signal actions on fences recently signaled. * This is done from a workqueue so we don't have to execute @@ -353,7 +351,6 @@ static int vmw_fence_obj_init(struct vmw_fence_manager *fman, goto out_unlock; } list_add_tail(&fence->head, &fman->fence_list); - ++fman->num_fence_objects;
out_unlock: spin_unlock(&fman->lock); @@ -402,7 +399,7 @@ static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman, { u32 goal_seqno; u32 *fifo_mem; - struct vmw_fence_obj *fence; + struct vmw_fence_obj *fence, *next_fence;
if (likely(!fman->seqno_valid)) return false; @@ -413,7 +410,7 @@ static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman, return false;
fman->seqno_valid = false; - list_for_each_entry(fence, &fman->fence_list, head) { + list_for_each_entry_safe(fence, next_fence, &fman->fence_list, head) { if (!list_empty(&fence->seq_passed_actions)) { fman->seqno_valid = true; vmw_mmio_write(fence->base.seqno,
[ Sasha's backport helper bot ]
Hi,
✅ All tests passed successfully. No issues detected. No action required from the submitter.
The upstream commit SHA1 provided is correct: e58337100721f3cc0c7424a18730e4f39844934f
WARNING: Author mismatch between patch and upstream commit: Backport author: Zhi YangZhi.Yang@eng.windriver.com Commit author: Zack Rusinzack.rusin@broadcom.com
Status in newer kernel trees: 6.14.y | Present (exact SHA1) 6.12.y | Present (exact SHA1) 6.6.y | Present (different SHA1: a8943969f9ea) 6.1.y | Present (different SHA1: 3b933b16c996) 5.15.y | Present (different SHA1: 9e20d028d8d1)
Note: The patch differs from the upstream commit: --- 1: e58337100721f ! 1: cdcabc8bb962b drm/vmwgfx: Fix a deadlock in dma buf fence polling @@ Metadata ## Commit message ## drm/vmwgfx: Fix a deadlock in dma buf fence polling
+ commit e58337100721f3cc0c7424a18730e4f39844934f upstream. + Introduce a version of the fence ops that on release doesn't remove the fence from the pending list, and thus doesn't require a lock to fix poll->fence wait->fence unref deadlocks. @@ Commit message Reviewed-by: Maaz Mombasawala maaz.mombasawala@broadcom.com Reviewed-by: Martin Krastev martin.krastev@broadcom.com Link: https://patchwork.freedesktop.org/patch/msgid/20240722184313.181318-2-zack.r... + [Minor context change fixed] + Signed-off-by: Zhi Yang Zhi.Yang@windriver.com + Signed-off-by: He Zhe zhe.he@windriver.com
## drivers/gpu/drm/vmwgfx/vmwgfx_fence.c ## @@ @@ drivers/gpu/drm/vmwgfx/vmwgfx_fence.c: static const struct dma_fence_ops vmw_fen };
- - /* + /** * Execute signal actions on fences recently signaled. * This is done from a workqueue so we don't have to execute @@ drivers/gpu/drm/vmwgfx/vmwgfx_fence.c: static int vmw_fence_obj_init(struct vmw_fence_manager *fman, @@ drivers/gpu/drm/vmwgfx/vmwgfx_fence.c: static int vmw_fence_obj_init(struct vmw_ out_unlock: spin_unlock(&fman->lock); @@ drivers/gpu/drm/vmwgfx/vmwgfx_fence.c: static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman, - u32 passed_seqno) { u32 goal_seqno; + u32 *fifo_mem; - struct vmw_fence_obj *fence; + struct vmw_fence_obj *fence, *next_fence;
@@ drivers/gpu/drm/vmwgfx/vmwgfx_fence.c: static bool vmw_fence_goal_new_locked(str + list_for_each_entry_safe(fence, next_fence, &fman->fence_list, head) { if (!list_empty(&fence->seq_passed_actions)) { fman->seqno_valid = true; - vmw_fence_goal_write(fman->dev_priv, + vmw_mmio_write(fence->base.seqno, ---
Results of testing on various branches:
| Branch | Patch Apply | Build Test | |---------------------------|-------------|------------| | stable/linux-5.10.y | Success | Success |
linux-stable-mirror@lists.linaro.org