-----Original Message-----
From: amd-gfx [mailto:amd-gfx-boun...@lists.freedesktop.org] On Behalf Of
Christian König
Sent: Monday, October 09, 2017 5:08 AM
To: Kasiviswanathan, Harish <harish.kasiviswanat...@amd.com>;
amd-gfx@lists.freedesktop.org
Subject: Re: [PATCH v2 1/2] drm/amdgpu: Refactor amdgpu_move_blit
Am 07.10.2017 um 00:08 schrieb Harish Kasiviswanathan:
> Add more generic function amdgpu_copy_ttm_mem_to_mem() that supports
> arbitrary copy size, offsets and two BOs (source & dest.).
>
> This is useful for KFD Cross Memory Attach feature where data needs to
> be copied from BOs from different processes
>
> v2: Add struct amdgpu_copy_mem and changed amdgpu_copy_ttm_mem_to_mem()
> function parameters to use the struct
>
> Change-Id: I848d541a84a1c2d12827d9dcf6d9054d854b4159
> Signed-off-by: Harish Kasiviswanathan <harish.kasiviswanat...@amd.com>
Looks like a nice cleanup to me, with the notes below fixed the patch is
Reviewed-by: Christian König <christian.koe...@amd.com>.
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 169
> +++++++++++++++++++++++---------
> drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h | 12 +++
> 2 files changed, 132 insertions(+), 49 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index 1086f03..dda4f08 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -289,97 +289,168 @@ static uint64_t amdgpu_mm_node_addr(struct
> ttm_buffer_object *bo,
> return addr;
> }
>
> -static int amdgpu_move_blit(struct ttm_buffer_object *bo,
> - bool evict, bool no_wait_gpu,
> - struct ttm_mem_reg *new_mem,
> - struct ttm_mem_reg *old_mem)
> +/**
> + * amdgpu_copy_ttm_mem_to_mem - Helper function for copy
> + *
> + * The function copies @size bytes from {src->mem + src->offset} to
> + * {dst->mem + dst->offset}. src->bo and dst->bo could be same BO for a
> + * move and different for a BO to BO copy.
> + *
> + * @f: Returns the last fence if multiple jobs are submitted.
> + */
> +int amdgpu_copy_ttm_mem_to_mem(struct amdgpu_device *adev,
Please name that amdgpu_ttm_copy_.... to note that the function is part
of the TTM subsystem in amdgpu.
[HK] - Will fix the name.
> + struct amdgpu_copy_mem *src,
> + struct amdgpu_copy_mem *dst,
> + uint64_t size,
> + struct reservation_object *resv,
> + struct dma_fence **f)
> {
> - struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
> struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
> -
> - struct drm_mm_node *old_mm, *new_mm;
> - uint64_t old_start, old_size, new_start, new_size;
> - unsigned long num_pages;
> + struct drm_mm_node *src_mm, *dst_mm;
> + uint64_t src_node_start, dst_node_start, src_node_size,
> + dst_node_size, src_page_offset, dst_page_offset;
> struct dma_fence *fence = NULL;
> - int r;
> -
> - BUILD_BUG_ON((PAGE_SIZE % AMDGPU_GPU_PAGE_SIZE) != 0);
> + int r = 0;
> + const uint64_t GTT_MAX_BYTES = (AMDGPU_GTT_MAX_TRANSFER_SIZE *
> + AMDGPU_GPU_PAGE_SIZE);
>
> if (!ring->ready) {
> DRM_ERROR("Trying to move memory with ring turned off.\n");
> return -EINVAL;
> }
>
> - old_mm = old_mem->mm_node;
> - old_size = old_mm->size;
> - old_start = amdgpu_mm_node_addr(bo, old_mm, old_mem);
> + src_mm = src->mem->mm_node;
> + while (src->offset >= (src_mm->size << PAGE_SHIFT)) {
> + src->offset -= (src_mm->size << PAGE_SHIFT);
> + ++src_mm;
> + }
> + src_node_start = amdgpu_mm_node_addr(src->bo, src_mm, src->mem) +
> + src->offset;
> + src_node_size = (src_mm->size << PAGE_SHIFT) - src->offset;
> + src_page_offset = src_node_start & (PAGE_SIZE - 1);
>
> - new_mm = new_mem->mm_node;
> - new_size = new_mm->size;
> - new_start = amdgpu_mm_node_addr(bo, new_mm, new_mem);
> + dst_mm = dst->mem->mm_node;
> + while (dst->offset >= (dst_mm->size << PAGE_SHIFT)) {
> + dst->offset -= (dst_mm->size << PAGE_SHIFT);
> + ++dst_mm;
> + }
> + dst_node_start = amdgpu_mm_node_addr(dst->bo, dst_mm, dst->mem) +
> + dst->offset;
> + dst_node_size = (dst_mm->size << PAGE_SHIFT) - dst->offset;
> + dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
>
> - num_pages = new_mem->num_pages;
> mutex_lock(&adev->mman.gtt_window_lock);
> - while (num_pages) {
> - unsigned long cur_pages = min(min(old_size, new_size),
> -
> (u64)AMDGPU_GTT_MAX_TRANSFER_SIZE);
> - uint64_t from = old_start, to = new_start;
> +
> + while (size) {
> + unsigned long cur_size;
> + uint64_t from = src_node_start, to = dst_node_start;
> struct dma_fence *next;
>
> - if (old_mem->mem_type == TTM_PL_TT &&
> - !amdgpu_gtt_mgr_is_allocated(old_mem)) {
> - r = amdgpu_map_buffer(bo, old_mem, cur_pages,
> - old_start, 0, ring, &from);
> + /* Copy size cannot exceed GTT_MAX_BYTES. So if src or dst
> + * begins at an offset, then adjust the size accordingly
> + */
> + cur_size = min3(min(src_node_size, dst_node_size), size,
> + GTT_MAX_BYTES);
> + if (cur_size + src_page_offset > GTT_MAX_BYTES ||
> + cur_size + dst_page_offset > GTT_MAX_BYTES)
> + cur_size -= max(src_page_offset, dst_page_offset);
> +
> + /* Map only what needs to be accessed. Map src to window 0 and
> + * dst to window 1
> + */
> + if (src->mem->mem_type == TTM_PL_TT &&
> + !amdgpu_gtt_mgr_is_allocated(src->mem)) {
> + r = amdgpu_map_buffer(src->bo, src->mem,
> + PFN_UP(cur_size + src_page_offset),
> + src_node_start, 0, ring,
> + &from);
> if (r)
> goto error;
> + /* Adjust the offset because amdgpu_map_buffer returns
> + * start of mapped page
> + */
> + from += src_page_offset;
> }
>
> - if (new_mem->mem_type == TTM_PL_TT &&
> - !amdgpu_gtt_mgr_is_allocated(new_mem)) {
> - r = amdgpu_map_buffer(bo, new_mem, cur_pages,
> - new_start, 1, ring, &to);
> + if (dst->mem->mem_type == TTM_PL_TT &&
> + !amdgpu_gtt_mgr_is_allocated(dst->mem)) {
> + r = amdgpu_map_buffer(dst->bo, dst->mem,
> + PFN_UP(cur_size + dst_page_offset),
> + dst_node_start, 1, ring,
> + &to);
> if (r)
> goto error;
> + to += dst_page_offset;
> }
>
> - r = amdgpu_copy_buffer(ring, from, to,
> - cur_pages * PAGE_SIZE,
> - bo->resv, &next, false, true);
> + r = amdgpu_copy_buffer(ring, from, to, cur_size,
> + resv, &next, false, true);
> if (r)
> goto error;
>
> dma_fence_put(fence);
> fence = next;
>
> - num_pages -= cur_pages;
> - if (!num_pages)
> + size -= cur_size;
> + if (!size)
> break;
>
> - old_size -= cur_pages;
> - if (!old_size) {
> - old_start = amdgpu_mm_node_addr(bo, ++old_mm, old_mem);
> - old_size = old_mm->size;
> + src_node_size -= cur_size;
> + if (!src_node_size) {
> + src_node_start = amdgpu_mm_node_addr(src->bo, ++src_mm,
> + src->mem);
> + src_node_size = (src_mm->size << PAGE_SHIFT);
> } else {
> - old_start += cur_pages * PAGE_SIZE;
> + src_node_start += cur_size;
> + src_page_offset = src_node_start & (PAGE_SIZE - 1);
> }
> -
> - new_size -= cur_pages;
> - if (!new_size) {
> - new_start = amdgpu_mm_node_addr(bo, ++new_mm, new_mem);
> - new_size = new_mm->size;
> + dst_node_size -= cur_size;
> + if (!dst_node_size) {
> + dst_node_start = amdgpu_mm_node_addr(dst->bo, ++dst_mm,
> + dst->mem);
> + dst_node_size = (dst_mm->size << PAGE_SHIFT);
> } else {
> - new_start += cur_pages * PAGE_SIZE;
> + dst_node_start += cur_size;
> + dst_page_offset = dst_node_start & (PAGE_SIZE - 1);
> }
> }
> +error:
> mutex_unlock(&adev->mman.gtt_window_lock);
> + if (f)
> + *f = dma_fence_get(fence);
> + dma_fence_put(fence);
> + return r;
> +}
> +
> +
> +static int amdgpu_move_blit(struct ttm_buffer_object *bo,
> + bool evict, bool no_wait_gpu,
> + struct ttm_mem_reg *new_mem,
> + struct ttm_mem_reg *old_mem)
> +{
I think I would rather prefer to change the callers of that function,
but not 100% sure.
How much change would that be? But not a blocking issue.
[HK]: The function is called 3 times in amdgpu_ttm.c. I don't see an elegant
way to avoid this function. So for now I am leaving it as it is.
> + struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
> + struct amdgpu_copy_mem src, dst;
> + struct dma_fence *fence = NULL;
> + int r;
> +
> + src.bo = bo;
> + dst.bo = bo;
> + src.mem = old_mem;
> + dst.mem = new_mem;
> + src.offset = 0;
> + dst.offset = 0;
> +
> + r = amdgpu_copy_ttm_mem_to_mem(adev, &src, &dst,
> + new_mem->num_pages << PAGE_SHIFT,
> + bo->resv, &fence);
> + if (r)
> + goto error;
>
> r = ttm_bo_pipeline_move(bo, fence, evict, new_mem);
> dma_fence_put(fence);
> return r;
>
> error:
> - mutex_unlock(&adev->mman.gtt_window_lock);
> -
> if (fence)
> dma_fence_wait(fence, false);
> dma_fence_put(fence);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> index 7abae68..2188034 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> @@ -58,6 +58,12 @@ struct amdgpu_mman {
> struct amd_sched_entity entity;
> };
>
> +struct amdgpu_copy_mem {
> + struct ttm_buffer_object *bo;
> + struct ttm_mem_reg *mem;
> + unsigned long offset;
> +};
> +
> extern const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func;
> extern const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func;
>
> @@ -72,6 +78,12 @@ int amdgpu_copy_buffer(struct amdgpu_ring *ring, uint64_t
> src_offset,
> struct reservation_object *resv,
> struct dma_fence **fence, bool direct_submit,
> bool vm_needs_flush);
> +int amdgpu_copy_ttm_mem_to_mem(struct amdgpu_device *adev,
> + struct amdgpu_copy_mem *src,
> + struct amdgpu_copy_mem *dst,
> + uint64_t size,
> + struct reservation_object *resv,
> + struct dma_fence **f);
> int amdgpu_fill_buffer(struct amdgpu_bo *bo,
> uint64_t src_data,
> struct reservation_object *resv,
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
