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>`.
37 #include "amdgpu_display.h"
38 #include <drm/amdgpu_drm.h>
39 #include <linux/dma-buf.h>
41 static const struct dma_buf_ops amdgpu_dmabuf_ops;
44 * amdgpu_gem_prime_get_sg_table - &drm_driver.gem_prime_get_sg_table
46 * @obj: GEM buffer object
49 * A scatter/gather table for the pinned pages of the buffer object's memory.
51 struct sg_table *amdgpu_gem_prime_get_sg_table(struct drm_gem_object *obj)
53 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
54 int npages = bo->tbo.num_pages;
56 return drm_prime_pages_to_sg(bo->tbo.ttm->pages, npages);
60 * amdgpu_gem_prime_vmap - &dma_buf_ops.vmap implementation
61 * @obj: GEM buffer object
63 * Sets up an in-kernel virtual mapping of the buffer object's memory.
66 * The virtual address of the mapping or an error pointer.
68 void *amdgpu_gem_prime_vmap(struct drm_gem_object *obj)
70 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
73 ret = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages,
78 return bo->dma_buf_vmap.virtual;
82 * amdgpu_gem_prime_vunmap - &dma_buf_ops.vunmap implementation
83 * @obj: GEM buffer object
84 * @vaddr: virtual address (unused)
86 * Tears down the in-kernel virtual mapping of the buffer object's memory.
88 void amdgpu_gem_prime_vunmap(struct drm_gem_object *obj, void *vaddr)
90 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
92 ttm_bo_kunmap(&bo->dma_buf_vmap);
96 * amdgpu_gem_prime_mmap - &drm_driver.gem_prime_mmap implementation
97 * @obj: GEM buffer object
98 * @vma: virtual memory area
100 * Sets up a userspace mapping of the buffer object's memory in the given
101 * virtual memory area.
104 * 0 on success or negative error code.
106 int amdgpu_gem_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)
108 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
109 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
110 unsigned asize = amdgpu_bo_size(bo);
119 /* Check for valid size. */
120 if (asize < vma->vm_end - vma->vm_start)
123 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
124 (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) {
127 vma->vm_pgoff += amdgpu_bo_mmap_offset(bo) >> PAGE_SHIFT;
129 /* prime mmap does not need to check access, so allow here */
130 ret = drm_vma_node_allow(&obj->vma_node, vma->vm_file->private_data);
134 ret = ttm_bo_mmap(vma->vm_file, vma, &adev->mman.bdev);
135 drm_vma_node_revoke(&obj->vma_node, vma->vm_file->private_data);
141 * amdgpu_gem_prime_import_sg_table - &drm_driver.gem_prime_import_sg_table
144 * @attach: DMA-buf attachment
145 * @sg: Scatter/gather table
147 * Import shared DMA buffer memory exported by another device.
150 * A new GEM buffer object of the given DRM device, representing the memory
151 * described by the given DMA-buf attachment and scatter/gather table.
153 struct drm_gem_object *
154 amdgpu_gem_prime_import_sg_table(struct drm_device *dev,
155 struct dma_buf_attachment *attach,
158 struct reservation_object *resv = attach->dmabuf->resv;
159 struct amdgpu_device *adev = dev->dev_private;
160 struct amdgpu_bo *bo;
161 struct amdgpu_bo_param bp;
164 memset(&bp, 0, sizeof(bp));
165 bp.size = attach->dmabuf->size;
166 bp.byte_align = PAGE_SIZE;
167 bp.domain = AMDGPU_GEM_DOMAIN_CPU;
169 bp.type = ttm_bo_type_sg;
171 ww_mutex_lock(&resv->lock, NULL);
172 ret = amdgpu_bo_create(adev, &bp, &bo);
177 bo->tbo.ttm->sg = sg;
178 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
179 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
180 if (attach->dmabuf->ops != &amdgpu_dmabuf_ops)
181 bo->prime_shared_count = 1;
183 ww_mutex_unlock(&resv->lock);
184 return &bo->gem_base;
187 ww_mutex_unlock(&resv->lock);
192 * amdgpu_gem_map_attach - &dma_buf_ops.attach implementation
193 * @dma_buf: shared DMA buffer
194 * @target_dev: target device
195 * @attach: DMA-buf attachment
197 * Makes sure that the shared DMA buffer can be accessed by the target device.
198 * For now, simply pins it to the GTT domain, where it should be accessible by
202 * 0 on success or negative error code.
204 static int amdgpu_gem_map_attach(struct dma_buf *dma_buf,
205 struct device *target_dev,
206 struct dma_buf_attachment *attach)
208 struct drm_gem_object *obj = dma_buf->priv;
209 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
210 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
213 r = drm_gem_map_attach(dma_buf, target_dev, attach);
217 r = amdgpu_bo_reserve(bo, false);
218 if (unlikely(r != 0))
222 if (attach->dev->driver != adev->dev->driver) {
224 * Wait for all shared fences to complete before we switch to future
225 * use of exclusive fence on this prime shared bo.
227 r = reservation_object_wait_timeout_rcu(bo->tbo.resv,
229 MAX_SCHEDULE_TIMEOUT);
230 if (unlikely(r < 0)) {
231 DRM_DEBUG_PRIME("Fence wait failed: %li\n", r);
232 goto error_unreserve;
236 /* pin buffer into GTT */
237 r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT, NULL);
239 goto error_unreserve;
241 if (attach->dev->driver != adev->dev->driver)
242 bo->prime_shared_count++;
245 amdgpu_bo_unreserve(bo);
249 drm_gem_map_detach(dma_buf, attach);
254 * amdgpu_gem_map_detach - &dma_buf_ops.detach implementation
255 * @dma_buf: shared DMA buffer
256 * @attach: DMA-buf attachment
258 * This is called when a shared DMA buffer no longer needs to be accessible by
259 * the other device. For now, simply unpins the buffer from GTT.
261 static void amdgpu_gem_map_detach(struct dma_buf *dma_buf,
262 struct dma_buf_attachment *attach)
264 struct drm_gem_object *obj = dma_buf->priv;
265 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
266 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
269 ret = amdgpu_bo_reserve(bo, true);
270 if (unlikely(ret != 0))
274 if (attach->dev->driver != adev->dev->driver && bo->prime_shared_count)
275 bo->prime_shared_count--;
276 amdgpu_bo_unreserve(bo);
279 drm_gem_map_detach(dma_buf, attach);
283 * amdgpu_gem_prime_res_obj - &drm_driver.gem_prime_res_obj implementation
284 * @obj: GEM buffer object
287 * The buffer object's reservation object.
289 struct reservation_object *amdgpu_gem_prime_res_obj(struct drm_gem_object *obj)
291 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
297 * amdgpu_gem_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
298 * @dma_buf: shared DMA buffer
299 * @direction: direction of DMA transfer
301 * This is called before CPU access to the shared DMA buffer's memory. If it's
302 * a read access, the buffer is moved to the GTT domain if possible, for optimal
303 * CPU read performance.
306 * 0 on success or negative error code.
308 static int amdgpu_gem_begin_cpu_access(struct dma_buf *dma_buf,
309 enum dma_data_direction direction)
311 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
312 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
313 struct ttm_operation_ctx ctx = { true, false };
314 u32 domain = amdgpu_display_supported_domains(adev);
316 bool reads = (direction == DMA_BIDIRECTIONAL ||
317 direction == DMA_FROM_DEVICE);
319 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
323 ret = amdgpu_bo_reserve(bo, false);
324 if (unlikely(ret != 0))
327 if (!bo->pin_count && (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
328 amdgpu_ttm_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
329 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
332 amdgpu_bo_unreserve(bo);
336 static const struct dma_buf_ops amdgpu_dmabuf_ops = {
337 .attach = amdgpu_gem_map_attach,
338 .detach = amdgpu_gem_map_detach,
339 .map_dma_buf = drm_gem_map_dma_buf,
340 .unmap_dma_buf = drm_gem_unmap_dma_buf,
341 .release = drm_gem_dmabuf_release,
342 .begin_cpu_access = amdgpu_gem_begin_cpu_access,
343 .map = drm_gem_dmabuf_kmap,
344 .map_atomic = drm_gem_dmabuf_kmap_atomic,
345 .unmap = drm_gem_dmabuf_kunmap,
346 .unmap_atomic = drm_gem_dmabuf_kunmap_atomic,
347 .mmap = drm_gem_dmabuf_mmap,
348 .vmap = drm_gem_dmabuf_vmap,
349 .vunmap = drm_gem_dmabuf_vunmap,
353 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
355 * @gobj: GEM buffer object
356 * @flags: flags like DRM_CLOEXEC and DRM_RDWR
358 * The main work is done by the &drm_gem_prime_export helper, which in turn
359 * uses &amdgpu_gem_prime_res_obj.
362 * Shared DMA buffer representing the GEM buffer object from the given device.
364 struct dma_buf *amdgpu_gem_prime_export(struct drm_device *dev,
365 struct drm_gem_object *gobj,
368 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
371 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
372 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
373 return ERR_PTR(-EPERM);
375 buf = drm_gem_prime_export(dev, gobj, flags);
377 buf->file->f_mapping = dev->anon_inode->i_mapping;
378 buf->ops = &amdgpu_dmabuf_ops;
385 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
387 * @dma_buf: Shared DMA buffer
389 * The main work is done by the &drm_gem_prime_import helper, which in turn
390 * uses &amdgpu_gem_prime_import_sg_table.
393 * GEM buffer object representing the shared DMA buffer for the given device.
395 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
396 struct dma_buf *dma_buf)
398 struct drm_gem_object *obj;
400 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
402 if (obj->dev == dev) {
404 * Importing dmabuf exported from out own gem increases
405 * refcount on gem itself instead of f_count of dmabuf.
407 drm_gem_object_get(obj);
412 return drm_gem_prime_import(dev, dma_buf);