Re: [PATCH] drm/amdgpu: improve debug VRAM access performance using sdma

[Christian König <christian.koenig@xxxxxxx>]

Am 04.01.22 um 20:12 schrieb Jonathan Kim:
> For better performance during VRAM access for debugged processes, do
> read/write copies over SDMA.
>
> In order to fulfill post mortem debugging on a broken device, fallback to
> stable MMIO access when gpu recovery is disabled or when job submission
> time outs are set to max.  Failed SDMA access should automatically fall
> back to MMIO access.
>
> Use a pre-allocated GTT bounce buffer pre-mapped into GART to avoid
> page-table updates and TLB flushes on access.
>
> Signed-off-by: Jonathan Kim <jonathan.kim@xxxxxxx>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 78 +++++++++++++++++++++++++
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h |  5 +-
>   2 files changed, 82 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index 367abed1d6e6..512df4c09772 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -48,6 +48,7 @@
>   #include <drm/ttm/ttm_range_manager.h>
>   
>   #include <drm/amdgpu_drm.h>
> +#include <drm/drm_drv.h>
>   
>   #include "amdgpu.h"
>   #include "amdgpu_object.h"
> @@ -1429,6 +1430,70 @@ static void amdgpu_ttm_vram_mm_access(struct amdgpu_device *adev, loff_t pos,
>   	}
>   }
>   
> +static int amdgpu_ttm_access_memory_sdma(struct ttm_buffer_object *bo,
> +					unsigned long offset, void *buf, int len, int write)
> +{
> +	struct amdgpu_bo *abo = ttm_to_amdgpu_bo(bo);
> +	struct amdgpu_device *adev = amdgpu_ttm_adev(abo->tbo.bdev);
> +	struct amdgpu_job *job;
> +	struct dma_fence *fence;
> +	uint64_t src_addr, dst_addr;
> +	unsigned int num_dw;
> +	int r, idx;
> +
> +	if (len != PAGE_SIZE)
> +		return -EINVAL;
> +
> +	if (!adev->mman.sdma_access_ptr)
> +		return -EACCES;
> +
> +	r = drm_dev_enter(adev_to_drm(adev), &idx);
> +	if (r)
> +		return r;
> +
> +	if (write)
> +		memcpy(adev->mman.sdma_access_ptr, buf, len);
> +
> +	num_dw = ALIGN(adev->mman.buffer_funcs->copy_num_dw, 8);
> +	r = amdgpu_job_alloc_with_ib(adev, num_dw * 4, AMDGPU_IB_POOL_DELAYED, &job);
> +	if (r)
> +		goto out;
> +
> +	src_addr = write ? amdgpu_bo_gpu_offset(adev->mman.sdma_access_bo) :
> +			amdgpu_bo_gpu_offset(abo);
> +	dst_addr = write ? amdgpu_bo_gpu_offset(abo) :
> +			amdgpu_bo_gpu_offset(adev->mman.sdma_access_bo);

I suggest to write this as

src_addr = a;
dst_addr = b;
if (write)
    swap(src_addr, dst_addr);

This way we are not duplicating getting the different offsets.

> +	amdgpu_emit_copy_buffer(adev, &job->ibs[0], src_addr, dst_addr, PAGE_SIZE, false);
> +
> +	amdgpu_ring_pad_ib(adev->mman.buffer_funcs_ring, &job->ibs[0]);
> +	WARN_ON(job->ibs[0].length_dw > num_dw);
> +
> +	r = amdgpu_job_submit(job, &adev->mman.entity, AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
> +	if (r) {
> +		amdgpu_job_free(job);
> +		goto out;
> +	}
> +
> +	if (!dma_fence_wait_timeout(fence, false, adev->sdma_timeout))
> +		r = -ETIMEDOUT;
> +	dma_fence_put(fence);
> +
> +	if (!(r || write))
> +		memcpy(buf, adev->mman.sdma_access_ptr, len);
> +out:
> +	drm_dev_exit(idx);
> +	return r;
> +}
> +
> +static inline bool amdgpu_ttm_allow_post_mortem_debug(struct amdgpu_device *adev)
> +{
> +	return amdgpu_gpu_recovery == 0 ||
> +		adev->gfx_timeout == MAX_SCHEDULE_TIMEOUT ||
> +		adev->compute_timeout == MAX_SCHEDULE_TIMEOUT ||
> +		adev->sdma_timeout == MAX_SCHEDULE_TIMEOUT ||
> +		adev->video_timeout == MAX_SCHEDULE_TIMEOUT;
> +}

This should probably be inside amdgpu_device.c

> +
>   /**
>    * amdgpu_ttm_access_memory - Read or Write memory that backs a buffer object.
>    *
> @@ -1453,6 +1518,10 @@ static int amdgpu_ttm_access_memory(struct ttm_buffer_object *bo,
>   	if (bo->resource->mem_type != TTM_PL_VRAM)
>   		return -EIO;
>   
> +	if (!amdgpu_ttm_allow_post_mortem_debug(adev) &&
> +			!amdgpu_ttm_access_memory_sdma(bo, offset, buf, len, write))
> +		return len;
> +
>   	amdgpu_res_first(bo->resource, offset, len, &cursor);
>   	while (cursor.remaining) {
>   		size_t count, size = cursor.size;
> @@ -1793,6 +1862,12 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>   		return r;
>   	}
>   
> +	if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
> +				AMDGPU_GEM_DOMAIN_GTT,
> +				&adev->mman.sdma_access_bo, NULL,
> +				adev->mman.sdma_access_ptr))
> +		DRM_WARN("Debug VRAM access will use slowpath MM access\n");
> +
>   	return 0;
>   }
>   
> @@ -1823,6 +1898,9 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev)
>   	ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_OA);
>   	ttm_device_fini(&adev->mman.bdev);
>   	adev->mman.initialized = false;
> +	if (adev->mman.sdma_access_ptr)

You can drop that if. Free functions can usually take a NULL pointer.

Apart from those nit picks looks good to me as well.

Regards,
Christian.

> +		amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
> +					&adev->mman.sdma_access_ptr);
>   	DRM_INFO("amdgpu: ttm finalized\n");
>   }
>   
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> index 91a087f9dc7c..b0116c4a768f 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> @@ -98,6 +98,10 @@ struct amdgpu_mman {
>   	u64		fw_vram_usage_size;
>   	struct amdgpu_bo	*fw_vram_usage_reserved_bo;
>   	void		*fw_vram_usage_va;
> +
> +	/* PAGE_SIZE'd BO for process memory r/w over SDMA. */
> +	struct amdgpu_bo	*sdma_access_bo;
> +	void			*sdma_access_ptr;
>   };
>   
>   struct amdgpu_copy_mem {
> @@ -193,5 +197,4 @@ uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
>   int amdgpu_ttm_evict_resources(struct amdgpu_device *adev, int mem_type);
>   
>   void amdgpu_ttm_debugfs_init(struct amdgpu_device *adev);
> -
>   #endif