On Tue, Jul 2, 2024 at 5:12 AM Zack Rusin zack.rusin@broadcom.com wrote:

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+

---

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 5efc6a766f64..588d50ababf6 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;
  

@@ -124,13 +123,13 @@ 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);
  

}

@@ -257,7 +256,6 @@ 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

@@ -355,7 +353,6 @@ 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); @@ -403,7 +400,7 @@ static bool vmw_fence_goal_new_locked(struct vmw_fence_manager *fman, u32 passed_seqno) { u32 goal_seqno;

•   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_fence_goal_write(fman->dev_priv,
  

-- 2.43.0

LGTM

Reviewed-by: Martin Krastev martin.krastev@broadcom.com

Regards, Martin