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 __dma_resv_make_exclusive(struct dma_resv *obj)
48 struct dma_fence **fences;
52 if (!dma_resv_get_list(obj)) /* no shared fences to convert */
55 r = dma_resv_get_fences_rcu(obj, NULL, &count, &fences);
60 /* Now that was unexpected. */
61 } else if (count == 1) {
62 dma_resv_add_excl_fence(obj, fences[0]);
63 dma_fence_put(fences[0]);
66 struct dma_fence_array *array;
68 array = dma_fence_array_create(count, fences,
69 dma_fence_context_alloc(1), 0,
74 dma_resv_add_excl_fence(obj, &array->base);
75 dma_fence_put(&array->base);
82 dma_fence_put(fences[count]);
88 * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
90 * @dmabuf: DMA-buf where we attach to
91 * @attach: attachment to add
93 * Add the attachment as user to the exported DMA-buf.
95 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
96 struct dma_buf_attachment *attach)
98 struct drm_gem_object *obj = dmabuf->priv;
99 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
100 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
103 if (pci_p2pdma_distance_many(adev->pdev, &attach->dev, 1, true) < 0)
104 attach->peer2peer = false;
106 if (attach->dev->driver == adev->dev->driver)
109 r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
113 r = amdgpu_bo_reserve(bo, false);
114 if (unlikely(r != 0))
118 * We only create shared fences for internal use, but importers
119 * of the dmabuf rely on exclusive fences for implicitly
120 * tracking write hazards. As any of the current fences may
121 * correspond to a write, we need to convert all existing
122 * fences on the reservation object into a single exclusive
125 r = __dma_resv_make_exclusive(bo->tbo.base.resv);
129 bo->prime_shared_count++;
130 amdgpu_bo_unreserve(bo);
134 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
139 * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
141 * @dmabuf: DMA-buf where we remove the attachment from
142 * @attach: the attachment to remove
144 * Called when an attachment is removed from the DMA-buf.
146 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
147 struct dma_buf_attachment *attach)
149 struct drm_gem_object *obj = dmabuf->priv;
150 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
151 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
153 if (attach->dev->driver != adev->dev->driver && bo->prime_shared_count)
154 bo->prime_shared_count--;
156 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
157 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
161 * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
163 * @attach: attachment to pin down
165 * Pin the BO which is backing the DMA-buf so that it can't move any more.
167 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
169 struct drm_gem_object *obj = attach->dmabuf->priv;
170 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
172 /* pin buffer into GTT */
173 return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
177 * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
179 * @attach: attachment to unpin
181 * Unpin a previously pinned BO to make it movable again.
183 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
185 struct drm_gem_object *obj = attach->dmabuf->priv;
186 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
192 * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
193 * @attach: DMA-buf attachment
194 * @dir: DMA direction
196 * Makes sure that the shared DMA buffer can be accessed by the target device.
197 * For now, simply pins it to the GTT domain, where it should be accessible by
201 * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
204 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
205 enum dma_data_direction dir)
207 struct dma_buf *dma_buf = attach->dmabuf;
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);
211 struct sg_table *sgt;
214 if (!bo->tbo.pin_count) {
215 /* move buffer into GTT or VRAM */
216 struct ttm_operation_ctx ctx = { false, false };
217 unsigned domains = AMDGPU_GEM_DOMAIN_GTT;
219 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
221 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
222 domains |= AMDGPU_GEM_DOMAIN_VRAM;
224 amdgpu_bo_placement_from_domain(bo, domains);
225 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
229 } else if (!(amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type) &
230 AMDGPU_GEM_DOMAIN_GTT)) {
231 return ERR_PTR(-EBUSY);
234 switch (bo->tbo.mem.mem_type) {
236 sgt = drm_prime_pages_to_sg(obj->dev,
238 bo->tbo.ttm->num_pages);
242 if (dma_map_sgtable(attach->dev, sgt, dir,
243 DMA_ATTR_SKIP_CPU_SYNC))
248 r = amdgpu_vram_mgr_alloc_sgt(adev, &bo->tbo.mem, 0,
249 bo->tbo.base.size, attach->dev, dir, &sgt);
254 return ERR_PTR(-EINVAL);
262 return ERR_PTR(-EBUSY);
266 * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
267 * @attach: DMA-buf attachment
268 * @sgt: sg_table to unmap
269 * @dir: DMA direction
271 * This is called when a shared DMA buffer no longer needs to be accessible by
272 * another device. For now, simply unpins the buffer from GTT.
274 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
275 struct sg_table *sgt,
276 enum dma_data_direction dir)
278 if (sgt->sgl->page_link) {
279 dma_unmap_sgtable(attach->dev, sgt, dir, 0);
283 amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
288 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
289 * @dma_buf: Shared DMA buffer
290 * @direction: Direction of DMA transfer
292 * This is called before CPU access to the shared DMA buffer's memory. If it's
293 * a read access, the buffer is moved to the GTT domain if possible, for optimal
294 * CPU read performance.
297 * 0 on success or a negative error code on failure.
299 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
300 enum dma_data_direction direction)
302 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
303 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
304 struct ttm_operation_ctx ctx = { true, false };
305 u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
307 bool reads = (direction == DMA_BIDIRECTIONAL ||
308 direction == DMA_FROM_DEVICE);
310 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
314 ret = amdgpu_bo_reserve(bo, false);
315 if (unlikely(ret != 0))
318 if (!bo->tbo.pin_count &&
319 (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
320 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
321 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
324 amdgpu_bo_unreserve(bo);
328 const struct dma_buf_ops amdgpu_dmabuf_ops = {
329 .attach = amdgpu_dma_buf_attach,
330 .detach = amdgpu_dma_buf_detach,
331 .pin = amdgpu_dma_buf_pin,
332 .unpin = amdgpu_dma_buf_unpin,
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 = drm_to_adev(dev);
386 struct drm_gem_object *gobj;
387 struct amdgpu_bo *bo;
391 dma_resv_lock(resv, NULL);
393 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
394 struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
396 flags |= other->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC;
399 ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
400 AMDGPU_GEM_DOMAIN_CPU, flags,
401 ttm_bo_type_sg, resv, &gobj);
405 bo = gem_to_amdgpu_bo(gobj);
406 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
407 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
408 if (dma_buf->ops != &amdgpu_dmabuf_ops)
409 bo->prime_shared_count = 1;
411 dma_resv_unlock(resv);
415 dma_resv_unlock(resv);
420 * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
422 * @attach: the DMA-buf attachment
424 * Invalidate the DMA-buf attachment, making sure that the we re-create the
425 * mapping before the next use.
428 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
430 struct drm_gem_object *obj = attach->importer_priv;
431 struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
432 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
433 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
434 struct ttm_operation_ctx ctx = { false, false };
435 struct ttm_placement placement = {};
436 struct amdgpu_vm_bo_base *bo_base;
439 if (bo->tbo.mem.mem_type == TTM_PL_SYSTEM)
442 r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
444 DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
448 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
449 struct amdgpu_vm *vm = bo_base->vm;
450 struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
453 /* When we get an error here it means that somebody
454 * else is holding the VM lock and updating page tables
455 * So we can just continue here.
457 r = dma_resv_lock(resv, ticket);
462 /* TODO: This is more problematic and we actually need
463 * to allow page tables updates without holding the
466 if (!dma_resv_trylock(resv))
470 r = amdgpu_vm_clear_freed(adev, vm, NULL);
472 r = amdgpu_vm_handle_moved(adev, vm);
474 if (r && r != -EBUSY)
475 DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
478 dma_resv_unlock(resv);
482 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
483 .allow_peer2peer = true,
484 .move_notify = amdgpu_dma_buf_move_notify
488 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
490 * @dma_buf: Shared DMA buffer
492 * Import a dma_buf into a the driver and potentially create a new GEM object.
495 * GEM BO representing the shared DMA buffer for the given device.
497 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
498 struct dma_buf *dma_buf)
500 struct dma_buf_attachment *attach;
501 struct drm_gem_object *obj;
503 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
505 if (obj->dev == dev) {
507 * Importing dmabuf exported from out own gem increases
508 * refcount on gem itself instead of f_count of dmabuf.
510 drm_gem_object_get(obj);
515 obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
519 attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
520 &amdgpu_dma_buf_attach_ops, obj);
521 if (IS_ERR(attach)) {
522 drm_gem_object_put(obj);
523 return ERR_CAST(attach);
526 get_dma_buf(dma_buf);
527 obj->import_attach = attach;
532 * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
534 * @adev: amdgpu_device pointer of the importer
535 * @bo: amdgpu buffer object
538 * True if dmabuf accessible over xgmi, false otherwise.
540 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
541 struct amdgpu_bo *bo)
543 struct drm_gem_object *obj = &bo->tbo.base;
544 struct drm_gem_object *gobj;
546 if (obj->import_attach) {
547 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
549 if (dma_buf->ops != &amdgpu_dmabuf_ops)
550 /* No XGMI with non AMD GPUs */
553 gobj = dma_buf->priv;
554 bo = gem_to_amdgpu_bo(gobj);
557 if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
558 (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))