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 "amdgpu_xgmi.h"
39 #include <drm/amdgpu_drm.h>
40 #include <linux/dma-buf.h>
41 #include <linux/dma-fence-array.h>
42 #include <linux/pci-p2pdma.h>
43 #include <linux/pm_runtime.h>
46 * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
48 * @dmabuf: DMA-buf where we attach to
49 * @attach: attachment to add
51 * Add the attachment as user to the exported DMA-buf.
53 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
54 struct dma_buf_attachment *attach)
56 struct drm_gem_object *obj = dmabuf->priv;
57 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
58 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
61 if (pci_p2pdma_distance_many(adev->pdev, &attach->dev, 1, true) < 0)
62 attach->peer2peer = false;
64 r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
71 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
76 * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
78 * @dmabuf: DMA-buf where we remove the attachment from
79 * @attach: the attachment to remove
81 * Called when an attachment is removed from the DMA-buf.
83 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
84 struct dma_buf_attachment *attach)
86 struct drm_gem_object *obj = dmabuf->priv;
87 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
88 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
90 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
91 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
95 * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
97 * @attach: attachment to pin down
99 * Pin the BO which is backing the DMA-buf so that it can't move any more.
101 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
103 struct drm_gem_object *obj = attach->dmabuf->priv;
104 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
107 /* pin buffer into GTT */
108 r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
112 if (bo->tbo.moving) {
113 r = dma_fence_wait(bo->tbo.moving, true);
123 * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
125 * @attach: attachment to unpin
127 * Unpin a previously pinned BO to make it movable again.
129 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
131 struct drm_gem_object *obj = attach->dmabuf->priv;
132 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
138 * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
139 * @attach: DMA-buf attachment
140 * @dir: DMA direction
142 * Makes sure that the shared DMA buffer can be accessed by the target device.
143 * For now, simply pins it to the GTT domain, where it should be accessible by
147 * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
150 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
151 enum dma_data_direction dir)
153 struct dma_buf *dma_buf = attach->dmabuf;
154 struct drm_gem_object *obj = dma_buf->priv;
155 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
156 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
157 struct sg_table *sgt;
160 if (!bo->tbo.pin_count) {
161 /* move buffer into GTT or VRAM */
162 struct ttm_operation_ctx ctx = { false, false };
163 unsigned domains = AMDGPU_GEM_DOMAIN_GTT;
165 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
167 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
168 domains |= AMDGPU_GEM_DOMAIN_VRAM;
170 amdgpu_bo_placement_from_domain(bo, domains);
171 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
175 } else if (!(amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type) &
176 AMDGPU_GEM_DOMAIN_GTT)) {
177 return ERR_PTR(-EBUSY);
180 switch (bo->tbo.resource->mem_type) {
182 sgt = drm_prime_pages_to_sg(obj->dev,
184 bo->tbo.ttm->num_pages);
188 if (dma_map_sgtable(attach->dev, sgt, dir,
189 DMA_ATTR_SKIP_CPU_SYNC))
194 r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
195 bo->tbo.base.size, attach->dev,
201 return ERR_PTR(-EINVAL);
209 return ERR_PTR(-EBUSY);
213 * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
214 * @attach: DMA-buf attachment
215 * @sgt: sg_table to unmap
216 * @dir: DMA direction
218 * This is called when a shared DMA buffer no longer needs to be accessible by
219 * another device. For now, simply unpins the buffer from GTT.
221 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
222 struct sg_table *sgt,
223 enum dma_data_direction dir)
225 if (sgt->sgl->page_link) {
226 dma_unmap_sgtable(attach->dev, sgt, dir, 0);
230 amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
235 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
236 * @dma_buf: Shared DMA buffer
237 * @direction: Direction of DMA transfer
239 * This is called before CPU access to the shared DMA buffer's memory. If it's
240 * a read access, the buffer is moved to the GTT domain if possible, for optimal
241 * CPU read performance.
244 * 0 on success or a negative error code on failure.
246 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
247 enum dma_data_direction direction)
249 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
250 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
251 struct ttm_operation_ctx ctx = { true, false };
252 u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
254 bool reads = (direction == DMA_BIDIRECTIONAL ||
255 direction == DMA_FROM_DEVICE);
257 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
261 ret = amdgpu_bo_reserve(bo, false);
262 if (unlikely(ret != 0))
265 if (!bo->tbo.pin_count &&
266 (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
267 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
268 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
271 amdgpu_bo_unreserve(bo);
275 const struct dma_buf_ops amdgpu_dmabuf_ops = {
276 .attach = amdgpu_dma_buf_attach,
277 .detach = amdgpu_dma_buf_detach,
278 .pin = amdgpu_dma_buf_pin,
279 .unpin = amdgpu_dma_buf_unpin,
280 .map_dma_buf = amdgpu_dma_buf_map,
281 .unmap_dma_buf = amdgpu_dma_buf_unmap,
282 .release = drm_gem_dmabuf_release,
283 .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
284 .mmap = drm_gem_dmabuf_mmap,
285 .vmap = drm_gem_dmabuf_vmap,
286 .vunmap = drm_gem_dmabuf_vunmap,
290 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
292 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
294 * The main work is done by the &drm_gem_prime_export helper.
297 * Shared DMA buffer representing the GEM BO from the given device.
299 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
302 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
305 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
306 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
307 return ERR_PTR(-EPERM);
309 buf = drm_gem_prime_export(gobj, flags);
311 buf->ops = &amdgpu_dmabuf_ops;
317 * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
322 * Creates an empty SG BO for DMA-buf import.
325 * A new GEM BO of the given DRM device, representing the memory
326 * described by the given DMA-buf attachment and scatter/gather table.
328 static struct drm_gem_object *
329 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
331 struct dma_resv *resv = dma_buf->resv;
332 struct amdgpu_device *adev = drm_to_adev(dev);
333 struct drm_gem_object *gobj;
334 struct amdgpu_bo *bo;
338 dma_resv_lock(resv, NULL);
340 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
341 struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
343 flags |= other->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC;
346 ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
347 AMDGPU_GEM_DOMAIN_CPU, flags,
348 ttm_bo_type_sg, resv, &gobj);
352 bo = gem_to_amdgpu_bo(gobj);
353 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
354 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
356 dma_resv_unlock(resv);
360 dma_resv_unlock(resv);
365 * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
367 * @attach: the DMA-buf attachment
369 * Invalidate the DMA-buf attachment, making sure that the we re-create the
370 * mapping before the next use.
373 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
375 struct drm_gem_object *obj = attach->importer_priv;
376 struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
377 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
378 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
379 struct ttm_operation_ctx ctx = { false, false };
380 struct ttm_placement placement = {};
381 struct amdgpu_vm_bo_base *bo_base;
384 if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
387 r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
389 DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
393 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
394 struct amdgpu_vm *vm = bo_base->vm;
395 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
398 /* When we get an error here it means that somebody
399 * else is holding the VM lock and updating page tables
400 * So we can just continue here.
402 r = dma_resv_lock(resv, ticket);
407 /* TODO: This is more problematic and we actually need
408 * to allow page tables updates without holding the
411 if (!dma_resv_trylock(resv))
415 r = amdgpu_vm_clear_freed(adev, vm, NULL);
417 r = amdgpu_vm_handle_moved(adev, vm);
419 if (r && r != -EBUSY)
420 DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
423 dma_resv_unlock(resv);
427 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
428 .allow_peer2peer = true,
429 .move_notify = amdgpu_dma_buf_move_notify
433 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
435 * @dma_buf: Shared DMA buffer
437 * Import a dma_buf into a the driver and potentially create a new GEM object.
440 * GEM BO representing the shared DMA buffer for the given device.
442 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
443 struct dma_buf *dma_buf)
445 struct dma_buf_attachment *attach;
446 struct drm_gem_object *obj;
448 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
450 if (obj->dev == dev) {
452 * Importing dmabuf exported from out own gem increases
453 * refcount on gem itself instead of f_count of dmabuf.
455 drm_gem_object_get(obj);
460 obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
464 attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
465 &amdgpu_dma_buf_attach_ops, obj);
466 if (IS_ERR(attach)) {
467 drm_gem_object_put(obj);
468 return ERR_CAST(attach);
471 get_dma_buf(dma_buf);
472 obj->import_attach = attach;
477 * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
479 * @adev: amdgpu_device pointer of the importer
480 * @bo: amdgpu buffer object
483 * True if dmabuf accessible over xgmi, false otherwise.
485 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
486 struct amdgpu_bo *bo)
488 struct drm_gem_object *obj = &bo->tbo.base;
489 struct drm_gem_object *gobj;
491 if (obj->import_attach) {
492 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
494 if (dma_buf->ops != &amdgpu_dmabuf_ops)
495 /* No XGMI with non AMD GPUs */
498 gobj = dma_buf->priv;
499 bo = gem_to_amdgpu_bo(gobj);
502 if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
503 (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))