diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c
index cae9ac6379a5..aba811c6aa0d 100644
--- a/drivers/gpu/drm/i915/i915_active.c
+++ b/drivers/gpu/drm/i915/i915_active.c
@@ -457,8 +457,11 @@ int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
 		}
 	} while (unlikely(is_barrier(active)));
 
-	if (!__i915_active_fence_set(active, fence))
+	fence = __i915_active_fence_set(active, fence);
+	if (!fence)
 		__i915_active_acquire(ref);
+	else
+		dma_fence_put(fence);
 
 out:
 	i915_active_release(ref);
@@ -477,13 +480,9 @@ __i915_active_set_fence(struct i915_active *ref,
 		return NULL;
 	}
 
-	rcu_read_lock();
 	prev = __i915_active_fence_set(active, fence);
-	if (prev)
-		prev = dma_fence_get_rcu(prev);
-	else
+	if (!prev)
 		__i915_active_acquire(ref);
-	rcu_read_unlock();
 
 	return prev;
 }
@@ -1050,10 +1049,11 @@ void i915_request_add_active_barriers(struct i915_request *rq)
  *
  * Records the new @fence as the last active fence along its timeline in
  * this active tracker, moving the tracking callbacks from the previous
- * fence onto this one. Returns the previous fence (if not already completed),
- * which the caller must ensure is executed before the new fence. To ensure
- * that the order of fences within the timeline of the i915_active_fence is
- * understood, it should be locked by the caller.
+ * fence onto this one. Gets and returns a reference to the previous fence
+ * (if not already completed), which the caller must put after making sure
+ * that it is executed before the new fence. To ensure that the order of
+ * fences within the timeline of the i915_active_fence is understood, it
+ * should be locked by the caller.
  */
 struct dma_fence *
 __i915_active_fence_set(struct i915_active_fence *active,
@@ -1062,7 +1062,23 @@ __i915_active_fence_set(struct i915_active_fence *active,
 	struct dma_fence *prev;
 	unsigned long flags;
 
-	if (fence == rcu_access_pointer(active->fence))
+	/*
+	 * In case of fences embedded in i915_requests, their memory is
+	 * SLAB_FAILSAFE_BY_RCU, then it can be reused right after release
+	 * by new requests.  Then, there is a risk of passing back a pointer
+	 * to a new, completely unrelated fence that reuses the same memory
+	 * while tracked under a different active tracker.  Combined with i915
+	 * perf open/close operations that build await dependencies between
+	 * engine kernel context requests and user requests from different
+	 * timelines, this can lead to dependency loops and infinite waits.
+	 *
+	 * As a countermeasure, we try to get a reference to the active->fence
+	 * first, so if we succeed and pass it back to our user then it is not
+	 * released and potentially reused by an unrelated request before the
+	 * user has a chance to set up an await dependency on it.
+	 */
+	prev = i915_active_fence_get(active);
+	if (fence == prev)
 		return fence;
 
 	GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
@@ -1071,27 +1087,56 @@ __i915_active_fence_set(struct i915_active_fence *active,
 	 * Consider that we have two threads arriving (A and B), with
 	 * C already resident as the active->fence.
 	 *
-	 * A does the xchg first, and so it sees C or NULL depending
-	 * on the timing of the interrupt handler. If it is NULL, the
-	 * previous fence must have been signaled and we know that
-	 * we are first on the timeline. If it is still present,
-	 * we acquire the lock on that fence and serialise with the interrupt
-	 * handler, in the process removing it from any future interrupt
-	 * callback. A will then wait on C before executing (if present).
-	 *
-	 * As B is second, it sees A as the previous fence and so waits for
-	 * it to complete its transition and takes over the occupancy for
-	 * itself -- remembering that it needs to wait on A before executing.
+	 * Both A and B have got a reference to C or NULL, depending on the
+	 * timing of the interrupt handler.  Let's assume that if A has got C
+	 * then it has locked C first (before B).
 	 *
 	 * Note the strong ordering of the timeline also provides consistent
 	 * nesting rules for the fence->lock; the inner lock is always the
 	 * older lock.
 	 */
 	spin_lock_irqsave(fence->lock, flags);
-	prev = xchg(__active_fence_slot(active), fence);
-	if (prev) {
-		GEM_BUG_ON(prev == fence);
+	if (prev)
 		spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
+
+	/*
+	 * A does the cmpxchg first, and so it sees C or NULL, as before, or
+	 * something else, depending on the timing of other threads and/or
+	 * interrupt handler.  If not the same as before then A unlocks C if
+	 * applicable and retries, starting from an attempt to get a new
+	 * active->fence.  Meanwhile, B follows the same path as A.
+	 * Once A succeeds with cmpxch, B fails again, retires, gets A from
+	 * active->fence, locks it as soon as A completes, and possibly
+	 * succeeds with cmpxchg.
+	 */
+	while (cmpxchg(__active_fence_slot(active), prev, fence) != prev) {
+		if (prev) {
+			spin_unlock(prev->lock);
+			dma_fence_put(prev);
+		}
+		spin_unlock_irqrestore(fence->lock, flags);
+
+		prev = i915_active_fence_get(active);
+		GEM_BUG_ON(prev == fence);
+
+		spin_lock_irqsave(fence->lock, flags);
+		if (prev)
+			spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
+	}
+
+	/*
+	 * If prev is NULL then the previous fence must have been signaled
+	 * and we know that we are first on the timeline.  If it is still
+	 * present then, having the lock on that fence already acquired, we
+	 * serialise with the interrupt handler, in the process of removing it
+	 * from any future interrupt callback.  A will then wait on C before
+	 * executing (if present).
+	 *
+	 * As B is second, it sees A as the previous fence and so waits for
+	 * it to complete its transition and takes over the occupancy for
+	 * itself -- remembering that it needs to wait on A before executing.
+	 */
+	if (prev) {
 		__list_del_entry(&active->cb.node);
 		spin_unlock(prev->lock); /* serialise with prev->cb_list */
 	}
@@ -1108,11 +1153,7 @@ int i915_active_fence_set(struct i915_active_fence *active,
 	int err = 0;
 
 	/* Must maintain timeline ordering wrt previous active requests */
-	rcu_read_lock();
 	fence = __i915_active_fence_set(active, &rq->fence);
-	if (fence) /* but the previous fence may not belong to that timeline! */
-		fence = dma_fence_get_rcu(fence);
-	rcu_read_unlock();
 	if (fence) {
 		err = i915_request_await_dma_fence(rq, fence);
 		dma_fence_put(fence);
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c
index 896389f93029..eda56b2fdc68 100644
--- a/drivers/gpu/drm/i915/i915_request.c
+++ b/drivers/gpu/drm/i915/i915_request.c
@@ -1525,6 +1525,8 @@ __i915_request_add_to_timeline(struct i915_request *rq)
 							 &rq->dep,
 							 0);
 	}
+	if (prev)
+		i915_request_put(prev);
 
 	/*
 	 * Make sure that no request gazumped us - if it was allocated after