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 <drm/ttm/ttm_tt.h>
41 #include <linux/dma-buf.h>
42 #include <linux/dma-fence-array.h>
43 #include <linux/pci-p2pdma.h>
44 #include <linux/pm_runtime.h>
47 * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
49 * @dmabuf: DMA-buf where we attach to
50 * @attach: attachment to add
52 * Add the attachment as user to the exported DMA-buf.
54 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
55 struct dma_buf_attachment *attach)
57 struct drm_gem_object *obj = dmabuf->priv;
58 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
59 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
62 if (pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0)
63 attach->peer2peer = false;
65 r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
72 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
77 * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
79 * @dmabuf: DMA-buf where we remove the attachment from
80 * @attach: the attachment to remove
82 * Called when an attachment is removed from the DMA-buf.
84 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
85 struct dma_buf_attachment *attach)
87 struct drm_gem_object *obj = dmabuf->priv;
88 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
89 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
91 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
92 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
96 * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
98 * @attach: attachment to pin down
100 * Pin the BO which is backing the DMA-buf so that it can't move any more.
102 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
104 struct drm_gem_object *obj = attach->dmabuf->priv;
105 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
107 /* pin buffer into GTT */
108 return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
112 * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
114 * @attach: attachment to unpin
116 * Unpin a previously pinned BO to make it movable again.
118 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
120 struct drm_gem_object *obj = attach->dmabuf->priv;
121 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
127 * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
128 * @attach: DMA-buf attachment
129 * @dir: DMA direction
131 * Makes sure that the shared DMA buffer can be accessed by the target device.
132 * For now, simply pins it to the GTT domain, where it should be accessible by
136 * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
139 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
140 enum dma_data_direction dir)
142 struct dma_buf *dma_buf = attach->dmabuf;
143 struct drm_gem_object *obj = dma_buf->priv;
144 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
145 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
146 struct sg_table *sgt;
149 if (!bo->tbo.pin_count) {
150 /* move buffer into GTT or VRAM */
151 struct ttm_operation_ctx ctx = { false, false };
152 unsigned int domains = AMDGPU_GEM_DOMAIN_GTT;
154 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
156 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
157 domains |= AMDGPU_GEM_DOMAIN_VRAM;
159 amdgpu_bo_placement_from_domain(bo, domains);
160 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
164 } else if (!(amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type) &
165 AMDGPU_GEM_DOMAIN_GTT)) {
166 return ERR_PTR(-EBUSY);
169 switch (bo->tbo.resource->mem_type) {
171 sgt = drm_prime_pages_to_sg(obj->dev,
173 bo->tbo.ttm->num_pages);
177 if (dma_map_sgtable(attach->dev, sgt, dir,
178 DMA_ATTR_SKIP_CPU_SYNC))
183 r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
184 bo->tbo.base.size, attach->dev,
190 return ERR_PTR(-EINVAL);
198 return ERR_PTR(-EBUSY);
202 * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
203 * @attach: DMA-buf attachment
204 * @sgt: sg_table to unmap
205 * @dir: DMA direction
207 * This is called when a shared DMA buffer no longer needs to be accessible by
208 * another device. For now, simply unpins the buffer from GTT.
210 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
211 struct sg_table *sgt,
212 enum dma_data_direction dir)
214 if (sgt->sgl->page_link) {
215 dma_unmap_sgtable(attach->dev, sgt, dir, 0);
219 amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
224 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
225 * @dma_buf: Shared DMA buffer
226 * @direction: Direction of DMA transfer
228 * This is called before CPU access to the shared DMA buffer's memory. If it's
229 * a read access, the buffer is moved to the GTT domain if possible, for optimal
230 * CPU read performance.
233 * 0 on success or a negative error code on failure.
235 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
236 enum dma_data_direction direction)
238 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
239 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
240 struct ttm_operation_ctx ctx = { true, false };
241 u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
243 bool reads = (direction == DMA_BIDIRECTIONAL ||
244 direction == DMA_FROM_DEVICE);
246 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
250 ret = amdgpu_bo_reserve(bo, false);
251 if (unlikely(ret != 0))
254 if (!bo->tbo.pin_count &&
255 (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
256 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
257 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
260 amdgpu_bo_unreserve(bo);
264 const struct dma_buf_ops amdgpu_dmabuf_ops = {
265 .attach = amdgpu_dma_buf_attach,
266 .detach = amdgpu_dma_buf_detach,
267 .pin = amdgpu_dma_buf_pin,
268 .unpin = amdgpu_dma_buf_unpin,
269 .map_dma_buf = amdgpu_dma_buf_map,
270 .unmap_dma_buf = amdgpu_dma_buf_unmap,
271 .release = drm_gem_dmabuf_release,
272 .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
273 .mmap = drm_gem_dmabuf_mmap,
274 .vmap = drm_gem_dmabuf_vmap,
275 .vunmap = drm_gem_dmabuf_vunmap,
279 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
281 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
283 * The main work is done by the &drm_gem_prime_export helper.
286 * Shared DMA buffer representing the GEM BO from the given device.
288 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
291 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
294 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
295 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
296 return ERR_PTR(-EPERM);
298 buf = drm_gem_prime_export(gobj, flags);
300 buf->ops = &amdgpu_dmabuf_ops;
306 * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
311 * Creates an empty SG BO for DMA-buf import.
314 * A new GEM BO of the given DRM device, representing the memory
315 * described by the given DMA-buf attachment and scatter/gather table.
317 static struct drm_gem_object *
318 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
320 struct dma_resv *resv = dma_buf->resv;
321 struct amdgpu_device *adev = drm_to_adev(dev);
322 struct drm_gem_object *gobj;
323 struct amdgpu_bo *bo;
327 dma_resv_lock(resv, NULL);
329 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
330 struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
332 flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
333 AMDGPU_GEM_CREATE_COHERENT |
334 AMDGPU_GEM_CREATE_UNCACHED);
337 ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
338 AMDGPU_GEM_DOMAIN_CPU, flags,
339 ttm_bo_type_sg, resv, &gobj, 0);
343 bo = gem_to_amdgpu_bo(gobj);
344 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
345 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
347 dma_resv_unlock(resv);
351 dma_resv_unlock(resv);
356 * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
358 * @attach: the DMA-buf attachment
360 * Invalidate the DMA-buf attachment, making sure that the we re-create the
361 * mapping before the next use.
364 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
366 struct drm_gem_object *obj = attach->importer_priv;
367 struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
368 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
369 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
370 struct ttm_operation_ctx ctx = { false, false };
371 struct ttm_placement placement = {};
372 struct amdgpu_vm_bo_base *bo_base;
375 if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
378 r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
380 DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
384 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
385 struct amdgpu_vm *vm = bo_base->vm;
386 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
389 /* When we get an error here it means that somebody
390 * else is holding the VM lock and updating page tables
391 * So we can just continue here.
393 r = dma_resv_lock(resv, ticket);
398 /* TODO: This is more problematic and we actually need
399 * to allow page tables updates without holding the
402 if (!dma_resv_trylock(resv))
406 r = amdgpu_vm_clear_freed(adev, vm, NULL);
408 r = amdgpu_vm_handle_moved(adev, vm);
410 if (r && r != -EBUSY)
411 DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
414 dma_resv_unlock(resv);
418 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
419 .allow_peer2peer = true,
420 .move_notify = amdgpu_dma_buf_move_notify
424 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
426 * @dma_buf: Shared DMA buffer
428 * Import a dma_buf into a the driver and potentially create a new GEM object.
431 * GEM BO representing the shared DMA buffer for the given device.
433 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
434 struct dma_buf *dma_buf)
436 struct dma_buf_attachment *attach;
437 struct drm_gem_object *obj;
439 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
441 if (obj->dev == dev) {
443 * Importing dmabuf exported from out own gem increases
444 * refcount on gem itself instead of f_count of dmabuf.
446 drm_gem_object_get(obj);
451 obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
455 attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
456 &amdgpu_dma_buf_attach_ops, obj);
457 if (IS_ERR(attach)) {
458 drm_gem_object_put(obj);
459 return ERR_CAST(attach);
462 get_dma_buf(dma_buf);
463 obj->import_attach = attach;
468 * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
470 * @adev: amdgpu_device pointer of the importer
471 * @bo: amdgpu buffer object
474 * True if dmabuf accessible over xgmi, false otherwise.
476 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
477 struct amdgpu_bo *bo)
479 struct drm_gem_object *obj = &bo->tbo.base;
480 struct drm_gem_object *gobj;
482 if (obj->import_attach) {
483 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
485 if (dma_buf->ops != &amdgpu_dmabuf_ops)
486 /* No XGMI with non AMD GPUs */
489 gobj = dma_buf->priv;
490 bo = gem_to_amdgpu_bo(gobj);
493 if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
494 (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))