[PATCH v2] drm/amdgpu/fence: Fix oops due to non-matching drm_sched init/fini

[Christian König]

Am 01.02.23 um 17:48 schrieb Guilherme G. Piccoli:
> 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].
>
> [0] https://lore.kernel.org/amd-gfx/984ee981-2906-0eaf-ccec-9f80975cb136@amd.com/
>
> Fixes: 067f44c8b459 ("drm/amdgpu: avoid over-handle of fence driver fini in s3 test (v2)")
> Suggested-by: Christian König <christian.koenig at amd.com>
> Cc: Guchun Chen <guchun.chen at amd.com>
> Cc: Luben Tuikov <luben.tuikov at amd.com>
> Cc: Mario Limonciello <mario.limonciello at amd.com>
> Signed-off-by: Guilherme G. Piccoli <gpiccoli at igalia.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 8 +++++++-
>   1 file changed, 7 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> index 00444203220d..3b962cb680a6 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c
> @@ -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->no_scheduler && ring->sched.ops)
>   			drm_sched_fini(&ring->sched);

I think we should drop the check for no_scheduler here and just call 
drm_sched_fini() when the scheduler instance was initialized before.

Background is that I've found a couple of places where we potentially 
set no_scheduler to false after the scheduler was already initialized.

This is then probably leading to a memory leak or worth.

Regards,
Christian.

>   
>   		for (j = 0; j <= ring->fence_drv.num_fences_mask; ++j)