On 12/11/25 16:04, Tvrtko Ursulin wrote: ...

...

@@ -90,6 +73,11 @@ static int test_signaling(void *arg)           goto err_free;       }   +    if (rcu_dereference_protected(f->ops, true)) { +        pr_err("Fence ops not cleared on signal\n"); +        goto err_free; +    }

Bump to after the signaled check just below? Otherwise the signaled state hasn't been ascertained yet.

Done. I've put it to the end of the test.

...

if (!dma_fence_is_signaled(f)) {           pr_err("Fence not reporting signaled\n");           goto err_free; @@ -540,19 +528,7 @@ int dma_fence(void)           SUBTEST(test_stub),           SUBTEST(race_signal_callback),       }; -    int ret;         pr_info("sizeof(dma_fence)=%zu\n", sizeof(struct dma_fence));

-    slab_fences = KMEM_CACHE(mock_fence, -                 SLAB_TYPESAFE_BY_RCU |

Hm.. race_signal_callback looks like it could be depending on SLAB_TYPESAFE_BY_RCU. To you not?

Hui? As far as I can see it doesn't.

The race_signal_callback test just depends on the general RCU functionality of fences.

Regards, Christian.

Regards,

Tvrtko

...

-                 SLAB_HWCACHE_ALIGN); -    if (!slab_fences) -        return -ENOMEM;

-    ret = subtests(tests, NULL);

-    kmem_cache_destroy(slab_fences);

-    return ret; +    return subtests(tests, NULL);   }