CVE-2026-68237

In the Linux kernel, the following vulnerability has been resolved:<br /> <br /> drm/amdgpu/userq: fix indefinite fence wait during GPU reset<br /> <br /> pre_reset only force-completes fences of MAPPED queues. A queue in any<br /> other state (e.g. mid-eviction) keeps its last_fence pending; after a<br /> GPU reset that fence never signals, so the eviction/suspend worker and<br /> process teardown (amdgpu_evf_mgr_flush_suspend) wait on it forever and<br /> wedge the machine:<br /> <br /> INFO: task kworker/6:28 blocked for more than 120 seconds.<br /> Workqueue: events amdgpu_eviction_fence_suspend_worker [amdgpu]<br /> Call Trace:<br /> dma_fence_wait_timeout+0x7e/0x130<br /> amdgpu_userq_evict+0x67/0x140 [amdgpu]<br /> amdgpu_eviction_fence_suspend_worker+0xd8/0x160 [amdgpu]<br /> process_scheduled_works+0xa6/0x420<br /> <br /> Force-complete every queue&amp;#39;s fence regardless of state. The unmap and<br /> mark-hung step stays gated on MAPPED, since unmapping a queue that is<br /> not mapped is invalid.<br /> <br /> (cherry picked from commit 9102b39fa924dcc3dc75a3137bfa9633c40b88c0)