diff --git a/src/freedreno/drm/freedreno_bo.c b/src/freedreno/drm/freedreno_bo.c
index 5229eaf17aa..d45f30dfa3c 100644
--- a/src/freedreno/drm/freedreno_bo.c
+++ b/src/freedreno/drm/freedreno_bo.c
@@ -372,7 +372,7 @@ fd_bo_fini_common(struct fd_bo *bo)
    VG_BO_FREE(bo);
 
    for (int i = 0; i < bo->nr_fences; i++)
-      fd_pipe_del(bo->fences[i].pipe);
+      fd_fence_del(bo->fences[i]);
 
    if (bo->fences != &bo->_inline_fence)
       free(bo->fences);
@@ -420,9 +420,16 @@ close_handles(struct fd_device *dev, uint32_t *handles, unsigned cnt)
 static void
 bo_flush(struct fd_bo *bo)
 {
-   for (int i = 0; i < bo->nr_fences; i++) {
-      struct fd_bo_fence *f = &bo->fences[i];
-      fd_pipe_flush(f->pipe, f->fence);
+   simple_mtx_lock(&fence_lock);
+   unsigned nr = bo->nr_fences;
+   struct fd_fence *fences[nr];
+   for (unsigned i = 0; i < nr; i++)
+      fences[i] = fd_fence_ref_locked(bo->fences[i]);
+   simple_mtx_unlock(&fence_lock);
+
+   for (unsigned i = 0; i < nr; i++) {
+      fd_fence_flush(bo->fences[i]);
+      fd_fence_del(fences[i]);
    }
 }
 
@@ -598,13 +605,11 @@ cleanup_fences(struct fd_bo *bo)
    simple_mtx_assert_locked(&fence_lock);
 
    for (int i = 0; i < bo->nr_fences; i++) {
-      struct fd_bo_fence *f = &bo->fences[i];
+      struct fd_fence *f = bo->fences[i];
 
-      if (fd_fence_before(f->pipe->control->fence, f->fence))
+      if (fd_fence_before(f->pipe->control->fence, f->ufence))
          continue;
 
-      struct fd_pipe *pipe = f->pipe;
-
       bo->nr_fences--;
 
       if (bo->nr_fences > 0) {
@@ -613,12 +618,12 @@ cleanup_fences(struct fd_bo *bo)
          i--;
       }
 
-      fd_pipe_del_locked(pipe);
+      fd_fence_del_locked(f);
    }
 }
 
 void
-fd_bo_add_fence(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t fence)
+fd_bo_add_fence(struct fd_bo *bo, struct fd_fence *fence)
 {
    simple_mtx_assert_locked(&fence_lock);
 
@@ -626,13 +631,16 @@ fd_bo_add_fence(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t fence)
       return;
 
    /* The common case is bo re-used on the same pipe it had previously
-    * been used on:
+    * been used on, so just replace the previous fence.
     */
    for (int i = 0; i < bo->nr_fences; i++) {
-      struct fd_bo_fence *f = &bo->fences[i];
-      if (f->pipe == pipe) {
-         assert(fd_fence_before(f->fence, fence));
-         f->fence = fence;
+      struct fd_fence *f = bo->fences[i];
+      if (f == fence)
+         return;
+      if (f->pipe == fence->pipe) {
+         assert(fd_fence_before(f->ufence, fence->ufence));
+         fd_fence_del_locked(f);
+         bo->fences[i] = fd_fence_ref_locked(fence);
          return;
       }
    }
@@ -650,10 +658,7 @@ fd_bo_add_fence(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t fence)
       APPEND(bo, fences, bo->_inline_fence);
    }
 
-   APPEND(bo, fences, (struct fd_bo_fence){
-      .pipe = fd_pipe_ref_locked(pipe),
-      .fence = fence,
-   });
+   APPEND(bo, fences, fd_fence_ref_locked(fence));
 }
 
 enum fd_bo_state
diff --git a/src/freedreno/drm/freedreno_drmif.h b/src/freedreno/drm/freedreno_drmif.h
index 7c14d93ab61..d88fd510318 100644
--- a/src/freedreno/drm/freedreno_drmif.h
+++ b/src/freedreno/drm/freedreno_drmif.h
@@ -208,15 +208,6 @@ int fd_pipe_wait_timeout(struct fd_pipe *pipe, const struct fd_fence *fence,
 /* buffer-object functions:
  */
 
-struct fd_bo_fence {
-   /* For non-shared buffers, track the last pipe the buffer was active
-    * on, and the per-pipe fence value that indicates when the buffer is
-    * idle:
-    */
-   uint32_t fence;
-   struct fd_pipe *pipe;
-};
-
 struct fd_bo {
    struct fd_device *dev;
    uint32_t size;
@@ -244,13 +235,13 @@ struct fd_bo {
    time_t free_time;      /* time when added to bucket-list */
 
    unsigned short nr_fences, max_fences;
-   struct fd_bo_fence *fences;
+   struct fd_fence **fences;
 
    /* In the common case, there is no more than one fence attached.
     * This provides storage for the fences table until it grows to
     * be larger than a single element.
     */
-   struct fd_bo_fence _inline_fence;
+   struct fd_fence *_inline_fence;
 };
 
 struct fd_bo *_fd_bo_new(struct fd_device *dev, uint32_t size, uint32_t flags);
diff --git a/src/freedreno/drm/freedreno_priv.h b/src/freedreno/drm/freedreno_priv.h
index 4772743064e..b166b0b8735 100644
--- a/src/freedreno/drm/freedreno_priv.h
+++ b/src/freedreno/drm/freedreno_priv.h
@@ -342,7 +342,7 @@ struct fd_bo_funcs {
    bool (*prefer_upload)(struct fd_bo *bo, unsigned len);
 };
 
-void fd_bo_add_fence(struct fd_bo *bo, struct fd_pipe *pipe, uint32_t fence);
+void fd_bo_add_fence(struct fd_bo *bo, struct fd_fence *fence);
 
 enum fd_bo_state {
    FD_BO_STATE_IDLE,
diff --git a/src/freedreno/drm/freedreno_ringbuffer_sp.c b/src/freedreno/drm/freedreno_ringbuffer_sp.c
index f78ff44fb44..a6dd3ed03cb 100644
--- a/src/freedreno/drm/freedreno_ringbuffer_sp.c
+++ b/src/freedreno/drm/freedreno_ringbuffer_sp.c
@@ -179,9 +179,11 @@ fd_submit_sp_flush_prep(struct fd_submit *submit, int in_fence_fd,
    for (unsigned i = 0; i < primary->u.nr_cmds; i++)
       fd_submit_append_bo(fd_submit, primary->u.cmds[i].ring_bo);
 
+   out_fence->ufence = submit->fence;
+
    simple_mtx_lock(&fence_lock);
    for (unsigned i = 0; i < fd_submit->nr_bos; i++) {
-      fd_bo_add_fence(fd_submit->bos[i], submit->pipe, submit->fence);
+      fd_bo_add_fence(fd_submit->bos[i], out_fence);
       has_shared |= fd_submit->bos[i]->alloc_flags & FD_BO_SHARED;
    }
    simple_mtx_unlock(&fence_lock);
diff --git a/src/freedreno/drm/msm/msm_ringbuffer.c b/src/freedreno/drm/msm/msm_ringbuffer.c
index dd9cb0c0f79..3bfcce55dd7 100644
--- a/src/freedreno/drm/msm/msm_ringbuffer.c
+++ b/src/freedreno/drm/msm/msm_ringbuffer.c
@@ -345,7 +345,7 @@ msm_submit_flush(struct fd_submit *submit, int in_fence_fd, bool use_fence_fd)
 
    simple_mtx_lock(&fence_lock);
    for (unsigned j = 0; j < msm_submit->nr_bos; j++) {
-      fd_bo_add_fence(msm_submit->bos[j], submit->pipe, submit->fence);
+      fd_bo_add_fence(msm_submit->bos[j], out_fence);
    }
    simple_mtx_unlock(&fence_lock);
 
diff --git a/src/freedreno/drm/msm/msm_ringbuffer_sp.c b/src/freedreno/drm/msm/msm_ringbuffer_sp.c
index 2403837b5bb..55639d80661 100644
--- a/src/freedreno/drm/msm/msm_ringbuffer_sp.c
+++ b/src/freedreno/drm/msm/msm_ringbuffer_sp.c
@@ -152,7 +152,6 @@ flush_submit_list(struct list_head *submit_list)
       msm_dump_submit(&req);
    } else if (!ret) {
       fd_submit->out_fence->kfence = req.fence;
-      fd_submit->out_fence->ufence = fd_submit->base.fence;
       fd_submit->out_fence->fence_fd = req.fence_fd;
    }
 
diff --git a/src/freedreno/drm/virtio/virtio_ringbuffer.c b/src/freedreno/drm/virtio/virtio_ringbuffer.c
index 42cee1df270..c4e1b4d7a67 100644
--- a/src/freedreno/drm/virtio/virtio_ringbuffer.c
+++ b/src/freedreno/drm/virtio/virtio_ringbuffer.c
@@ -176,7 +176,6 @@ flush_submit_list(struct list_head *submit_list)
    struct fd_fence *out_fence = fd_submit->out_fence;
 
    out_fence->kfence = kfence;
-   out_fence->ufence = fd_submit->base.fence;
 
    /* Even if gallium driver hasn't requested a fence-fd, request one.
     * This way, if we have to block waiting for the fence, we can do