Am 31.10.2017 um 18:58 schrieb Michel Dänzer: > On 25/10/17 05:43 PM, Michel Dänzer wrote: >> KASAN caught another use-after-free on my development machine today, see >> the attached dmesg excerpt. There haven't been any related changes in >> amd-staging-drm-next since yesterday, so maybe userspace is just >> tickling the kernel differently (e.g. piglit runs some more tests in >> parallel now). It's not reproducible every time, but it just happened a >> second time (with an amd-staging-drm-next commit from about a week ago). > I took a closer look, and I think I see what's happening. The > use-after-free happens at: > > reservation_object_wait_timeout_rcu+0xe02/0xe90 > ttm_bo_cleanup_refs_and_unlock+0x271/0x990 [ttm] (ttm_bo.c:530) > ttm_mem_evict_first+0x263/0x4a0 [ttm] > > The memory was freed at: > > [reservation_object_fini] > ttm_bo_cleanup_refs_and_unlock+0x517/0x990 [ttm] (ttm_bo.c:564) > ttm_mem_evict_first+0x263/0x4a0 [ttm] > > So it's two processes handling the same BO in ttm_mem_evict_first -> > ttm_bo_cleanup_refs_and_unlock. The first one unreserved the BO before > calling reservation_object_wait_timeout_rcu. Meanwhile, the other one > manages to reserve the BO and get all the way to the end of > ttm_bo_cleanup_refs_and_unlock, destroying bo->ttm_resv. Then > reservation_object_wait_timeout_rcu in the first process still accesses > memory which bo->ttm_resv pointed to => boom. Good catch. But this means that just grabbing another reference before calling reservation_object_wait_timeout_rcu() and we should be on the safe side, shouldn't we ? Going to take a closer look tomorrow, today is a holiday here and I'm actually ill again once more :( Christian.
