2 * Copyright 2012 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
22 * based on nouveau_prime.c
24 * Authors: Alex Deucher
28 * DOC: PRIME Buffer Sharing
30 * The following callback implementations are used for :ref:`sharing GEM buffer
31 * objects between different devices via PRIME <prime_buffer_sharing>`.
35 #include "amdgpu_display.h"
36 #include "amdgpu_gem.h"
37 #include <drm/amdgpu_drm.h>
38 #include <linux/dma-buf.h>
39 #include <linux/dma-fence-array.h>
42 * amdgpu_gem_prime_get_sg_table - &drm_driver.gem_prime_get_sg_table
44 * @obj: GEM buffer object (BO)
47 * A scatter/gather table for the pinned pages of the BO's memory.
49 struct sg_table *amdgpu_gem_prime_get_sg_table(struct drm_gem_object *obj)
51 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
52 int npages = bo->tbo.num_pages;
54 return drm_prime_pages_to_sg(bo->tbo.ttm->pages, npages);
58 * amdgpu_gem_prime_vmap - &dma_buf_ops.vmap implementation
61 * Sets up an in-kernel virtual mapping of the BO's memory.
64 * The virtual address of the mapping or an error pointer.
66 void *amdgpu_gem_prime_vmap(struct drm_gem_object *obj)
68 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
71 ret = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages,
76 return bo->dma_buf_vmap.virtual;
80 * amdgpu_gem_prime_vunmap - &dma_buf_ops.vunmap implementation
82 * @vaddr: Virtual address (unused)
84 * Tears down the in-kernel virtual mapping of the BO's memory.
86 void amdgpu_gem_prime_vunmap(struct drm_gem_object *obj, void *vaddr)
88 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
90 ttm_bo_kunmap(&bo->dma_buf_vmap);
94 * amdgpu_gem_prime_mmap - &drm_driver.gem_prime_mmap implementation
96 * @vma: Virtual memory area
98 * Sets up a userspace mapping of the BO's memory in the given
99 * virtual memory area.
102 * 0 on success or a negative error code on failure.
104 int amdgpu_gem_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
106 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
107 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
108 unsigned asize = amdgpu_bo_size(bo);
117 /* Check for valid size. */
118 if (asize < vma->vm_end - vma->vm_start)
121 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
122 (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) {
125 vma->vm_pgoff += amdgpu_bo_mmap_offset(bo) >> PAGE_SHIFT;
127 /* prime mmap does not need to check access, so allow here */
128 ret = drm_vma_node_allow(&obj->vma_node, vma->vm_file->private_data);
132 ret = ttm_bo_mmap(vma->vm_file, vma, &adev->mman.bdev);
133 drm_vma_node_revoke(&obj->vma_node, vma->vm_file->private_data);
139 * amdgpu_gem_prime_import_sg_table - &drm_driver.gem_prime_import_sg_table
142 * @attach: DMA-buf attachment
143 * @sg: Scatter/gather table
145 * Imports shared DMA buffer memory exported by another device.
148 * A new GEM BO of the given DRM device, representing the memory
149 * described by the given DMA-buf attachment and scatter/gather table.
151 struct drm_gem_object *
152 amdgpu_gem_prime_import_sg_table(struct drm_device *dev,
153 struct dma_buf_attachment *attach,
156 struct reservation_object *resv = attach->dmabuf->resv;
157 struct amdgpu_device *adev = dev->dev_private;
158 struct amdgpu_bo *bo;
159 struct amdgpu_bo_param bp;
162 memset(&bp, 0, sizeof(bp));
163 bp.size = attach->dmabuf->size;
164 bp.byte_align = PAGE_SIZE;
165 bp.domain = AMDGPU_GEM_DOMAIN_CPU;
167 bp.type = ttm_bo_type_sg;
169 ww_mutex_lock(&resv->lock, NULL);
170 ret = amdgpu_bo_create(adev, &bp, &bo);
175 bo->tbo.ttm->sg = sg;
176 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
177 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
178 if (attach->dmabuf->ops != &amdgpu_dmabuf_ops)
179 bo->prime_shared_count = 1;
181 ww_mutex_unlock(&resv->lock);
182 return &bo->gem_base;
185 ww_mutex_unlock(&resv->lock);
190 __reservation_object_make_exclusive(struct reservation_object *obj)
192 struct dma_fence **fences;
196 if (!reservation_object_get_list(obj)) /* no shared fences to convert */
199 r = reservation_object_get_fences_rcu(obj, NULL, &count, &fences);
204 /* Now that was unexpected. */
205 } else if (count == 1) {
206 reservation_object_add_excl_fence(obj, fences[0]);
207 dma_fence_put(fences[0]);
210 struct dma_fence_array *array;
212 array = dma_fence_array_create(count, fences,
213 dma_fence_context_alloc(1), 0,
218 reservation_object_add_excl_fence(obj, &array->base);
219 dma_fence_put(&array->base);
226 dma_fence_put(fences[count]);
232 * amdgpu_gem_map_attach - &dma_buf_ops.attach implementation
233 * @dma_buf: Shared DMA buffer
234 * @attach: DMA-buf attachment
236 * Makes sure that the shared DMA buffer can be accessed by the target device.
237 * For now, simply pins it to the GTT domain, where it should be accessible by
241 * 0 on success or a negative error code on failure.
243 static int amdgpu_gem_map_attach(struct dma_buf *dma_buf,
244 struct dma_buf_attachment *attach)
246 struct drm_gem_object *obj = dma_buf->priv;
247 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
248 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
251 r = drm_gem_map_attach(dma_buf, attach);
255 r = amdgpu_bo_reserve(bo, false);
256 if (unlikely(r != 0))
260 if (attach->dev->driver != adev->dev->driver) {
262 * We only create shared fences for internal use, but importers
263 * of the dmabuf rely on exclusive fences for implicitly
264 * tracking write hazards. As any of the current fences may
265 * correspond to a write, we need to convert all existing
266 * fences on the reservation object into a single exclusive
269 r = __reservation_object_make_exclusive(bo->tbo.resv);
271 goto error_unreserve;
274 /* pin buffer into GTT */
275 r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
277 goto error_unreserve;
279 if (attach->dev->driver != adev->dev->driver)
280 bo->prime_shared_count++;
283 amdgpu_bo_unreserve(bo);
287 drm_gem_map_detach(dma_buf, attach);
292 * amdgpu_gem_map_detach - &dma_buf_ops.detach implementation
293 * @dma_buf: Shared DMA buffer
294 * @attach: DMA-buf attachment
296 * This is called when a shared DMA buffer no longer needs to be accessible by
297 * another device. For now, simply unpins the buffer from GTT.
299 static void amdgpu_gem_map_detach(struct dma_buf *dma_buf,
300 struct dma_buf_attachment *attach)
302 struct drm_gem_object *obj = dma_buf->priv;
303 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
304 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
307 ret = amdgpu_bo_reserve(bo, true);
308 if (unlikely(ret != 0))
312 if (attach->dev->driver != adev->dev->driver && bo->prime_shared_count)
313 bo->prime_shared_count--;
314 amdgpu_bo_unreserve(bo);
317 drm_gem_map_detach(dma_buf, attach);
321 * amdgpu_gem_prime_res_obj - &drm_driver.gem_prime_res_obj implementation
325 * The BO's reservation object.
327 struct reservation_object *amdgpu_gem_prime_res_obj(struct drm_gem_object *obj)
329 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
335 * amdgpu_gem_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
336 * @dma_buf: Shared DMA buffer
337 * @direction: Direction of DMA transfer
339 * This is called before CPU access to the shared DMA buffer's memory. If it's
340 * a read access, the buffer is moved to the GTT domain if possible, for optimal
341 * CPU read performance.
344 * 0 on success or a negative error code on failure.
346 static int amdgpu_gem_begin_cpu_access(struct dma_buf *dma_buf,
347 enum dma_data_direction direction)
349 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
350 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
351 struct ttm_operation_ctx ctx = { true, false };
352 u32 domain = amdgpu_display_supported_domains(adev);
354 bool reads = (direction == DMA_BIDIRECTIONAL ||
355 direction == DMA_FROM_DEVICE);
357 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
361 ret = amdgpu_bo_reserve(bo, false);
362 if (unlikely(ret != 0))
365 if (!bo->pin_count && (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
366 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
367 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
370 amdgpu_bo_unreserve(bo);
374 const struct dma_buf_ops amdgpu_dmabuf_ops = {
375 .attach = amdgpu_gem_map_attach,
376 .detach = amdgpu_gem_map_detach,
377 .map_dma_buf = drm_gem_map_dma_buf,
378 .unmap_dma_buf = drm_gem_unmap_dma_buf,
379 .release = drm_gem_dmabuf_release,
380 .begin_cpu_access = amdgpu_gem_begin_cpu_access,
381 .mmap = drm_gem_dmabuf_mmap,
382 .vmap = drm_gem_dmabuf_vmap,
383 .vunmap = drm_gem_dmabuf_vunmap,
387 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
390 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
392 * The main work is done by the &drm_gem_prime_export helper, which in turn
393 * uses &amdgpu_gem_prime_res_obj.
396 * Shared DMA buffer representing the GEM BO from the given device.
398 struct dma_buf *amdgpu_gem_prime_export(struct drm_device *dev,
399 struct drm_gem_object *gobj,
402 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
405 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
406 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
407 return ERR_PTR(-EPERM);
409 buf = drm_gem_prime_export(dev, gobj, flags);
411 buf->file->f_mapping = dev->anon_inode->i_mapping;
412 buf->ops = &amdgpu_dmabuf_ops;
419 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
421 * @dma_buf: Shared DMA buffer
423 * The main work is done by the &drm_gem_prime_import helper, which in turn
424 * uses &amdgpu_gem_prime_import_sg_table.
427 * GEM BO representing the shared DMA buffer for the given device.
429 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
430 struct dma_buf *dma_buf)
432 struct drm_gem_object *obj;
434 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
436 if (obj->dev == dev) {
438 * Importing dmabuf exported from out own gem increases
439 * refcount on gem itself instead of f_count of dmabuf.
441 drm_gem_object_get(obj);
446 return drm_gem_prime_import(dev, dma_buf);