]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commitdiff
drm/amdgpu/fence: Fix oops due to non-matching drm_sched init/fini
authorGuilherme G. Piccoli <gpiccoli@igalia.com>
Thu, 2 Feb 2023 13:48:56 +0000 (10:48 -0300)
committerAlex Deucher <alexander.deucher@amd.com>
Thu, 9 Feb 2023 03:32:50 +0000 (22:32 -0500)
Currently amdgpu calls drm_sched_fini() from the fence driver sw fini
routine - such function is expected to be called only after the
respective init function - drm_sched_init() - was executed successfully.

Happens that we faced a driver probe failure in the Steam Deck
recently, and the function drm_sched_fini() was called even without
its counter-part had been previously called, causing the following oops:

amdgpu: probe of 0000:04:00.0 failed with error -110
BUG: kernel NULL pointer dereference, address: 0000000000000090
PGD 0 P4D 0
Oops: 0002 [#1] PREEMPT SMP NOPTI
CPU: 0 PID: 609 Comm: systemd-udevd Not tainted 6.2.0-rc3-gpiccoli #338
Hardware name: Valve Jupiter/Jupiter, BIOS F7A0113 11/04/2022
RIP: 0010:drm_sched_fini+0x84/0xa0 [gpu_sched]
[...]
Call Trace:
 <TASK>
 amdgpu_fence_driver_sw_fini+0xc8/0xd0 [amdgpu]
 amdgpu_device_fini_sw+0x2b/0x3b0 [amdgpu]
 amdgpu_driver_release_kms+0x16/0x30 [amdgpu]
 devm_drm_dev_init_release+0x49/0x70
 [...]

To prevent that, check if the drm_sched was properly initialized for a
given ring before calling its fini counter-part.

Notice ideally we'd use sched.ready for that; such field is set as the latest
thing on drm_sched_init(). But amdgpu seems to "override" the meaning of such
field - in the above oops for example, it was a GFX ring causing the crash, and
the sched.ready field was set to true in the ring init routine, regardless of
the state of the DRM scheduler. Hence, we ended-up using sched.ops as per
Christian's suggestion [0], and also removed the no_scheduler check [1].

[0] https://lore.kernel.org/amd-gfx/984ee981-2906-0eaf-ccec-9f80975cb136@amd.com/
[1] https://lore.kernel.org/amd-gfx/cd0e2994-f85f-d837-609f-7056d5fb7231@amd.com/

Fixes: 067f44c8b459 ("drm/amdgpu: avoid over-handle of fence driver fini in s3 test (v2)")
Suggested-by: Christian König <christian.koenig@amd.com>
Cc: Guchun Chen <guchun.chen@amd.com>
Cc: Luben Tuikov <luben.tuikov@amd.com>
Cc: Mario Limonciello <mario.limonciello@amd.com>
Reviewed-by: Luben Tuikov <luben.tuikov@amd.com>
Signed-off-by: Guilherme G. Piccoli <gpiccoli@igalia.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c

index 00444203220d2a00acdd3d981dd86e1b22615d56..faff4a3f96e6e8911e573daf46b4b894103a76e3 100644 (file)
@@ -618,7 +618,13 @@ void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev)
                if (!ring || !ring->fence_drv.initialized)
                        continue;
 
-               if (!ring->no_scheduler)
+               /*
+                * Notice we check for sched.ops since there's some
+                * override on the meaning of sched.ready by amdgpu.
+                * The natural check would be sched.ready, which is
+                * set as drm_sched_init() finishes...
+                */
+               if (ring->sched.ops)
                        drm_sched_fini(&ring->sched);
 
                for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)