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>
41 #include <linux/pci-p2pdma.h>
44 * amdgpu_gem_prime_vmap - &dma_buf_ops.vmap implementation
47 * Sets up an in-kernel virtual mapping of the BO's memory.
50 * The virtual address of the mapping or an error pointer.
52 void *amdgpu_gem_prime_vmap(struct drm_gem_object *obj)
54 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
57 ret = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages,
62 return bo->dma_buf_vmap.virtual;
66 * amdgpu_gem_prime_vunmap - &dma_buf_ops.vunmap implementation
68 * @vaddr: Virtual address (unused)
70 * Tears down the in-kernel virtual mapping of the BO's memory.
72 void amdgpu_gem_prime_vunmap(struct drm_gem_object *obj, void *vaddr)
74 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
76 ttm_bo_kunmap(&bo->dma_buf_vmap);
80 * amdgpu_gem_prime_mmap - &drm_driver.gem_prime_mmap implementation
82 * @vma: Virtual memory area
84 * Sets up a userspace mapping of the BO's memory in the given
85 * virtual memory area.
88 * 0 on success or a negative error code on failure.
90 int amdgpu_gem_prime_mmap(struct drm_gem_object *obj,
91 struct vm_area_struct *vma)
93 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
94 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
95 unsigned asize = amdgpu_bo_size(bo);
104 /* Check for valid size. */
105 if (asize < vma->vm_end - vma->vm_start)
108 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
109 (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) {
112 vma->vm_pgoff += amdgpu_bo_mmap_offset(bo) >> PAGE_SHIFT;
114 /* prime mmap does not need to check access, so allow here */
115 ret = drm_vma_node_allow(&obj->vma_node, vma->vm_file->private_data);
119 ret = ttm_bo_mmap(vma->vm_file, vma, &adev->mman.bdev);
120 drm_vma_node_revoke(&obj->vma_node, vma->vm_file->private_data);
126 __dma_resv_make_exclusive(struct dma_resv *obj)
128 struct dma_fence **fences;
132 if (!dma_resv_get_list(obj)) /* no shared fences to convert */
135 r = dma_resv_get_fences_rcu(obj, NULL, &count, &fences);
140 /* Now that was unexpected. */
141 } else if (count == 1) {
142 dma_resv_add_excl_fence(obj, fences[0]);
143 dma_fence_put(fences[0]);
146 struct dma_fence_array *array;
148 array = dma_fence_array_create(count, fences,
149 dma_fence_context_alloc(1), 0,
154 dma_resv_add_excl_fence(obj, &array->base);
155 dma_fence_put(&array->base);
162 dma_fence_put(fences[count]);
168 * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
170 * @dmabuf: DMA-buf where we attach to
171 * @attach: attachment to add
173 * Add the attachment as user to the exported DMA-buf.
175 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
176 struct dma_buf_attachment *attach)
178 struct drm_gem_object *obj = dmabuf->priv;
179 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
180 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
183 if (pci_p2pdma_distance_many(adev->pdev, &attach->dev, 1, true) < 0)
184 attach->peer2peer = false;
186 if (attach->dev->driver == adev->dev->driver)
189 r = amdgpu_bo_reserve(bo, false);
190 if (unlikely(r != 0))
194 * We only create shared fences for internal use, but importers
195 * of the dmabuf rely on exclusive fences for implicitly
196 * tracking write hazards. As any of the current fences may
197 * correspond to a write, we need to convert all existing
198 * fences on the reservation object into a single exclusive
201 r = __dma_resv_make_exclusive(bo->tbo.base.resv);
205 bo->prime_shared_count++;
206 amdgpu_bo_unreserve(bo);
211 * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
213 * @dmabuf: DMA-buf where we remove the attachment from
214 * @attach: the attachment to remove
216 * Called when an attachment is removed from the DMA-buf.
218 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
219 struct dma_buf_attachment *attach)
221 struct drm_gem_object *obj = dmabuf->priv;
222 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
223 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
225 if (attach->dev->driver != adev->dev->driver && bo->prime_shared_count)
226 bo->prime_shared_count--;
230 * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
232 * @attach: attachment to pin down
234 * Pin the BO which is backing the DMA-buf so that it can't move any more.
236 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
238 struct drm_gem_object *obj = attach->dmabuf->priv;
239 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
241 /* pin buffer into GTT */
242 return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
246 * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
248 * @attach: attachment to unpin
250 * Unpin a previously pinned BO to make it movable again.
252 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
254 struct drm_gem_object *obj = attach->dmabuf->priv;
255 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
261 * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
262 * @attach: DMA-buf attachment
263 * @dir: DMA direction
265 * Makes sure that the shared DMA buffer can be accessed by the target device.
266 * For now, simply pins it to the GTT domain, where it should be accessible by
270 * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
273 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
274 enum dma_data_direction dir)
276 struct dma_buf *dma_buf = attach->dmabuf;
277 struct drm_gem_object *obj = dma_buf->priv;
278 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
279 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
280 struct sg_table *sgt;
283 if (!bo->pin_count) {
284 /* move buffer into GTT or VRAM */
285 struct ttm_operation_ctx ctx = { false, false };
286 unsigned domains = AMDGPU_GEM_DOMAIN_GTT;
288 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
290 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
291 domains |= AMDGPU_GEM_DOMAIN_VRAM;
293 amdgpu_bo_placement_from_domain(bo, domains);
294 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
298 } else if (!(amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type) &
299 AMDGPU_GEM_DOMAIN_GTT)) {
300 return ERR_PTR(-EBUSY);
303 switch (bo->tbo.mem.mem_type) {
305 sgt = drm_prime_pages_to_sg(obj->dev,
311 if (dma_map_sgtable(attach->dev, sgt, dir,
312 DMA_ATTR_SKIP_CPU_SYNC))
317 r = amdgpu_vram_mgr_alloc_sgt(adev, &bo->tbo.mem, attach->dev,
323 return ERR_PTR(-EINVAL);
331 return ERR_PTR(-EBUSY);
335 * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
336 * @attach: DMA-buf attachment
337 * @sgt: sg_table to unmap
338 * @dir: DMA direction
340 * This is called when a shared DMA buffer no longer needs to be accessible by
341 * another device. For now, simply unpins the buffer from GTT.
343 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
344 struct sg_table *sgt,
345 enum dma_data_direction dir)
347 struct dma_buf *dma_buf = attach->dmabuf;
348 struct drm_gem_object *obj = dma_buf->priv;
349 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
350 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
352 if (sgt->sgl->page_link) {
353 dma_unmap_sgtable(attach->dev, sgt, dir, 0);
357 amdgpu_vram_mgr_free_sgt(adev, attach->dev, dir, sgt);
362 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
363 * @dma_buf: Shared DMA buffer
364 * @direction: Direction of DMA transfer
366 * This is called before CPU access to the shared DMA buffer's memory. If it's
367 * a read access, the buffer is moved to the GTT domain if possible, for optimal
368 * CPU read performance.
371 * 0 on success or a negative error code on failure.
373 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
374 enum dma_data_direction direction)
376 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
377 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
378 struct ttm_operation_ctx ctx = { true, false };
379 u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
381 bool reads = (direction == DMA_BIDIRECTIONAL ||
382 direction == DMA_FROM_DEVICE);
384 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
388 ret = amdgpu_bo_reserve(bo, false);
389 if (unlikely(ret != 0))
392 if (!bo->pin_count && (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
393 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
394 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
397 amdgpu_bo_unreserve(bo);
401 const struct dma_buf_ops amdgpu_dmabuf_ops = {
402 .attach = amdgpu_dma_buf_attach,
403 .detach = amdgpu_dma_buf_detach,
404 .pin = amdgpu_dma_buf_pin,
405 .unpin = amdgpu_dma_buf_unpin,
406 .map_dma_buf = amdgpu_dma_buf_map,
407 .unmap_dma_buf = amdgpu_dma_buf_unmap,
408 .release = drm_gem_dmabuf_release,
409 .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
410 .mmap = drm_gem_dmabuf_mmap,
411 .vmap = drm_gem_dmabuf_vmap,
412 .vunmap = drm_gem_dmabuf_vunmap,
416 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
418 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
420 * The main work is done by the &drm_gem_prime_export helper.
423 * Shared DMA buffer representing the GEM BO from the given device.
425 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
428 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
431 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
432 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
433 return ERR_PTR(-EPERM);
435 buf = drm_gem_prime_export(gobj, flags);
437 buf->ops = &amdgpu_dmabuf_ops;
443 * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
448 * Creates an empty SG BO for DMA-buf import.
451 * A new GEM BO of the given DRM device, representing the memory
452 * described by the given DMA-buf attachment and scatter/gather table.
454 static struct drm_gem_object *
455 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
457 struct dma_resv *resv = dma_buf->resv;
458 struct amdgpu_device *adev = dev->dev_private;
459 struct amdgpu_bo *bo;
460 struct amdgpu_bo_param bp;
463 memset(&bp, 0, sizeof(bp));
464 bp.size = dma_buf->size;
465 bp.byte_align = PAGE_SIZE;
466 bp.domain = AMDGPU_GEM_DOMAIN_CPU;
468 bp.type = ttm_bo_type_sg;
470 dma_resv_lock(resv, NULL);
471 ret = amdgpu_bo_create(adev, &bp, &bo);
475 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
476 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
477 if (dma_buf->ops != &amdgpu_dmabuf_ops)
478 bo->prime_shared_count = 1;
480 dma_resv_unlock(resv);
481 return &bo->tbo.base;
484 dma_resv_unlock(resv);
489 * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
491 * @attach: the DMA-buf attachment
493 * Invalidate the DMA-buf attachment, making sure that the we re-create the
494 * mapping before the next use.
497 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
499 struct drm_gem_object *obj = attach->importer_priv;
500 struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
501 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
502 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
503 struct ttm_operation_ctx ctx = { false, false };
504 struct ttm_placement placement = {};
505 struct amdgpu_vm_bo_base *bo_base;
508 if (bo->tbo.mem.mem_type == TTM_PL_SYSTEM)
511 r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
513 DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
517 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
518 struct amdgpu_vm *vm = bo_base->vm;
519 struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
522 /* When we get an error here it means that somebody
523 * else is holding the VM lock and updating page tables
524 * So we can just continue here.
526 r = dma_resv_lock(resv, ticket);
531 /* TODO: This is more problematic and we actually need
532 * to allow page tables updates without holding the
535 if (!dma_resv_trylock(resv))
539 r = amdgpu_vm_clear_freed(adev, vm, NULL);
541 r = amdgpu_vm_handle_moved(adev, vm);
543 if (r && r != -EBUSY)
544 DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
547 dma_resv_unlock(resv);
551 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
552 .allow_peer2peer = true,
553 .move_notify = amdgpu_dma_buf_move_notify
557 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
559 * @dma_buf: Shared DMA buffer
561 * Import a dma_buf into a the driver and potentially create a new GEM object.
564 * GEM BO representing the shared DMA buffer for the given device.
566 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
567 struct dma_buf *dma_buf)
569 struct dma_buf_attachment *attach;
570 struct drm_gem_object *obj;
572 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
574 if (obj->dev == dev) {
576 * Importing dmabuf exported from out own gem increases
577 * refcount on gem itself instead of f_count of dmabuf.
579 drm_gem_object_get(obj);
584 obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
588 attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
589 &amdgpu_dma_buf_attach_ops, obj);
590 if (IS_ERR(attach)) {
591 drm_gem_object_put(obj);
592 return ERR_CAST(attach);
595 get_dma_buf(dma_buf);
596 obj->import_attach = attach;