amdgpu vaapi failure due to dma_resv_locking_ctx returning 0

[valpackett]

Finally got around to debugging the video acceleration on Vega again. (HEVC encoding was working, H264 encoding and ALL decoding broke at some point)

Turns out this "double check" is failing:

drm-kmod/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c

Lines 1730 to 1732 in 2b7442a

	/* Double check that the BO is reserved by this CS */
	if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
	return -EINVAL;

because the dma_resv_locking_ctx(..) result is zero. Everything is working fine when the return is replaced with a printf.

---

how to reproduce: easiest command is avcstreamoutdemo (from libva utils) but it won't really be successful (still some assertion failure) when it's working, just the amdgpu error would disappear.

mpv -v --hwdec=vaapi some-thing.mp4 (mp4 either h264 or hevc if vainfo lists hevc) should start working when fixed

[valpackett]

Looking a bit deeper..

dma_resv_locking_ctx == READ_ONCE(obj->lock.ctx)
struct dma_resv *obj -> struct ww_mutex lock

struct ww_acquire_ctx {
};
struct ww_mutex {
	struct mutex base;
	struct cv condvar;
	struct ww_acquire_ctx *ctx;
};

there is no 'ctx =' in ww_mutex.h, seems like nothing ever sets that field. ww_mutex_lock(_interruptible) just ignores the ctx argument! I'll try setting it there.

[valpackett]

https://reviews.freebsd.org/D26657

[evadot]

So this can be closed now that it's commited ?

[valpackett]

Yep