2 * Copyright 2019 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 "amdgpu_dma_buf.h"
38 #include <drm/amdgpu_drm.h>
39 #include <linux/dma-buf.h>
40 #include <linux/dma-fence-array.h>
43 * amdgpu_gem_prime_vmap - &dma_buf_ops.vmap implementation
46 * Sets up an in-kernel virtual mapping of the BO's memory.
49 * The virtual address of the mapping or an error pointer.
51 void *amdgpu_gem_prime_vmap(struct drm_gem_object
*obj
)
53 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(obj
);
56 ret
= ttm_bo_kmap(&bo
->tbo
, 0, bo
->tbo
.num_pages
,
61 return bo
->dma_buf_vmap
.virtual;
65 * amdgpu_gem_prime_vunmap - &dma_buf_ops.vunmap implementation
67 * @vaddr: Virtual address (unused)
69 * Tears down the in-kernel virtual mapping of the BO's memory.
71 void amdgpu_gem_prime_vunmap(struct drm_gem_object
*obj
, void *vaddr
)
73 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(obj
);
75 ttm_bo_kunmap(&bo
->dma_buf_vmap
);
79 * amdgpu_gem_prime_mmap - &drm_driver.gem_prime_mmap implementation
81 * @vma: Virtual memory area
83 * Sets up a userspace mapping of the BO's memory in the given
84 * virtual memory area.
87 * 0 on success or a negative error code on failure.
89 int amdgpu_gem_prime_mmap(struct drm_gem_object
*obj
,
90 struct vm_area_struct
*vma
)
92 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(obj
);
93 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
94 unsigned asize
= amdgpu_bo_size(bo
);
103 /* Check for valid size. */
104 if (asize
< vma
->vm_end
- vma
->vm_start
)
107 if (amdgpu_ttm_tt_get_usermm(bo
->tbo
.ttm
) ||
108 (bo
->flags
& AMDGPU_GEM_CREATE_NO_CPU_ACCESS
)) {
111 vma
->vm_pgoff
+= amdgpu_bo_mmap_offset(bo
) >> PAGE_SHIFT
;
113 /* prime mmap does not need to check access, so allow here */
114 ret
= drm_vma_node_allow(&obj
->vma_node
, vma
->vm_file
->private_data
);
118 ret
= ttm_bo_mmap(vma
->vm_file
, vma
, &adev
->mman
.bdev
);
119 drm_vma_node_revoke(&obj
->vma_node
, vma
->vm_file
->private_data
);
125 __dma_resv_make_exclusive(struct dma_resv
*obj
)
127 struct dma_fence
**fences
;
131 if (!dma_resv_get_list(obj
)) /* no shared fences to convert */
134 r
= dma_resv_get_fences_rcu(obj
, NULL
, &count
, &fences
);
139 /* Now that was unexpected. */
140 } else if (count
== 1) {
141 dma_resv_add_excl_fence(obj
, fences
[0]);
142 dma_fence_put(fences
[0]);
145 struct dma_fence_array
*array
;
147 array
= dma_fence_array_create(count
, fences
,
148 dma_fence_context_alloc(1), 0,
153 dma_resv_add_excl_fence(obj
, &array
->base
);
154 dma_fence_put(&array
->base
);
161 dma_fence_put(fences
[count
]);
167 * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
169 * @dmabuf: DMA-buf where we attach to
170 * @attach: attachment to add
172 * Add the attachment as user to the exported DMA-buf.
174 static int amdgpu_dma_buf_attach(struct dma_buf
*dmabuf
,
175 struct dma_buf_attachment
*attach
)
177 struct drm_gem_object
*obj
= dmabuf
->priv
;
178 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(obj
);
179 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
182 if (attach
->dev
->driver
== adev
->dev
->driver
)
185 r
= amdgpu_bo_reserve(bo
, false);
186 if (unlikely(r
!= 0))
190 * We only create shared fences for internal use, but importers
191 * of the dmabuf rely on exclusive fences for implicitly
192 * tracking write hazards. As any of the current fences may
193 * correspond to a write, we need to convert all existing
194 * fences on the reservation object into a single exclusive
197 r
= __dma_resv_make_exclusive(bo
->tbo
.base
.resv
);
201 bo
->prime_shared_count
++;
202 amdgpu_bo_unreserve(bo
);
207 * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
209 * @dmabuf: DMA-buf where we remove the attachment from
210 * @attach: the attachment to remove
212 * Called when an attachment is removed from the DMA-buf.
214 static void amdgpu_dma_buf_detach(struct dma_buf
*dmabuf
,
215 struct dma_buf_attachment
*attach
)
217 struct drm_gem_object
*obj
= dmabuf
->priv
;
218 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(obj
);
219 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
221 if (attach
->dev
->driver
!= adev
->dev
->driver
&& bo
->prime_shared_count
)
222 bo
->prime_shared_count
--;
226 * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
227 * @attach: DMA-buf attachment
228 * @dir: DMA direction
230 * Makes sure that the shared DMA buffer can be accessed by the target device.
231 * For now, simply pins it to the GTT domain, where it should be accessible by
235 * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
238 static struct sg_table
*amdgpu_dma_buf_map(struct dma_buf_attachment
*attach
,
239 enum dma_data_direction dir
)
241 struct dma_buf
*dma_buf
= attach
->dmabuf
;
242 struct drm_gem_object
*obj
= dma_buf
->priv
;
243 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(obj
);
244 struct sg_table
*sgt
;
247 r
= amdgpu_bo_pin(bo
, AMDGPU_GEM_DOMAIN_GTT
);
251 sgt
= drm_prime_pages_to_sg(bo
->tbo
.ttm
->pages
, bo
->tbo
.num_pages
);
255 if (!dma_map_sg_attrs(attach
->dev
, sgt
->sgl
, sgt
->nents
, dir
,
256 DMA_ATTR_SKIP_CPU_SYNC
))
264 return ERR_PTR(-ENOMEM
);
268 * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
269 * @attach: DMA-buf attachment
270 * @sgt: sg_table to unmap
271 * @dir: DMA direction
273 * This is called when a shared DMA buffer no longer needs to be accessible by
274 * another device. For now, simply unpins the buffer from GTT.
276 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment
*attach
,
277 struct sg_table
*sgt
,
278 enum dma_data_direction dir
)
280 struct drm_gem_object
*obj
= attach
->dmabuf
->priv
;
281 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(obj
);
283 dma_unmap_sg(attach
->dev
, sgt
->sgl
, sgt
->nents
, dir
);
290 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
291 * @dma_buf: Shared DMA buffer
292 * @direction: Direction of DMA transfer
294 * This is called before CPU access to the shared DMA buffer's memory. If it's
295 * a read access, the buffer is moved to the GTT domain if possible, for optimal
296 * CPU read performance.
299 * 0 on success or a negative error code on failure.
301 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf
*dma_buf
,
302 enum dma_data_direction direction
)
304 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(dma_buf
->priv
);
305 struct amdgpu_device
*adev
= amdgpu_ttm_adev(bo
->tbo
.bdev
);
306 struct ttm_operation_ctx ctx
= { true, false };
307 u32 domain
= amdgpu_display_supported_domains(adev
, bo
->flags
);
309 bool reads
= (direction
== DMA_BIDIRECTIONAL
||
310 direction
== DMA_FROM_DEVICE
);
312 if (!reads
|| !(domain
& AMDGPU_GEM_DOMAIN_GTT
))
316 ret
= amdgpu_bo_reserve(bo
, false);
317 if (unlikely(ret
!= 0))
320 if (!bo
->pin_count
&& (bo
->allowed_domains
& AMDGPU_GEM_DOMAIN_GTT
)) {
321 amdgpu_bo_placement_from_domain(bo
, AMDGPU_GEM_DOMAIN_GTT
);
322 ret
= ttm_bo_validate(&bo
->tbo
, &bo
->placement
, &ctx
);
325 amdgpu_bo_unreserve(bo
);
329 const struct dma_buf_ops amdgpu_dmabuf_ops
= {
330 .dynamic_mapping
= true,
331 .attach
= amdgpu_dma_buf_attach
,
332 .detach
= amdgpu_dma_buf_detach
,
333 .map_dma_buf
= amdgpu_dma_buf_map
,
334 .unmap_dma_buf
= amdgpu_dma_buf_unmap
,
335 .release
= drm_gem_dmabuf_release
,
336 .begin_cpu_access
= amdgpu_dma_buf_begin_cpu_access
,
337 .mmap
= drm_gem_dmabuf_mmap
,
338 .vmap
= drm_gem_dmabuf_vmap
,
339 .vunmap
= drm_gem_dmabuf_vunmap
,
343 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
345 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
347 * The main work is done by the &drm_gem_prime_export helper.
350 * Shared DMA buffer representing the GEM BO from the given device.
352 struct dma_buf
*amdgpu_gem_prime_export(struct drm_gem_object
*gobj
,
355 struct amdgpu_bo
*bo
= gem_to_amdgpu_bo(gobj
);
358 if (amdgpu_ttm_tt_get_usermm(bo
->tbo
.ttm
) ||
359 bo
->flags
& AMDGPU_GEM_CREATE_VM_ALWAYS_VALID
)
360 return ERR_PTR(-EPERM
);
362 buf
= drm_gem_prime_export(gobj
, flags
);
364 buf
->ops
= &amdgpu_dmabuf_ops
;
370 * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
375 * Creates an empty SG BO for DMA-buf import.
378 * A new GEM BO of the given DRM device, representing the memory
379 * described by the given DMA-buf attachment and scatter/gather table.
381 static struct drm_gem_object
*
382 amdgpu_dma_buf_create_obj(struct drm_device
*dev
, struct dma_buf
*dma_buf
)
384 struct dma_resv
*resv
= dma_buf
->resv
;
385 struct amdgpu_device
*adev
= dev
->dev_private
;
386 struct amdgpu_bo
*bo
;
387 struct amdgpu_bo_param bp
;
390 memset(&bp
, 0, sizeof(bp
));
391 bp
.size
= dma_buf
->size
;
392 bp
.byte_align
= PAGE_SIZE
;
393 bp
.domain
= AMDGPU_GEM_DOMAIN_CPU
;
395 bp
.type
= ttm_bo_type_sg
;
397 dma_resv_lock(resv
, NULL
);
398 ret
= amdgpu_bo_create(adev
, &bp
, &bo
);
402 bo
->allowed_domains
= AMDGPU_GEM_DOMAIN_GTT
;
403 bo
->preferred_domains
= AMDGPU_GEM_DOMAIN_GTT
;
404 if (dma_buf
->ops
!= &amdgpu_dmabuf_ops
)
405 bo
->prime_shared_count
= 1;
407 dma_resv_unlock(resv
);
408 return &bo
->tbo
.base
;
411 dma_resv_unlock(resv
);
416 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
418 * @dma_buf: Shared DMA buffer
420 * Import a dma_buf into a the driver and potentially create a new GEM object.
423 * GEM BO representing the shared DMA buffer for the given device.
425 struct drm_gem_object
*amdgpu_gem_prime_import(struct drm_device
*dev
,
426 struct dma_buf
*dma_buf
)
428 struct dma_buf_attachment
*attach
;
429 struct drm_gem_object
*obj
;
431 if (dma_buf
->ops
== &amdgpu_dmabuf_ops
) {
433 if (obj
->dev
== dev
) {
435 * Importing dmabuf exported from out own gem increases
436 * refcount on gem itself instead of f_count of dmabuf.
438 drm_gem_object_get(obj
);
443 obj
= amdgpu_dma_buf_create_obj(dev
, dma_buf
);
447 attach
= dma_buf_dynamic_attach(dma_buf
, dev
->dev
, true);
448 if (IS_ERR(attach
)) {
449 drm_gem_object_put(obj
);
450 return ERR_CAST(attach
);
453 get_dma_buf(dma_buf
);
454 obj
->import_attach
= attach
;