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>
45 #include "amdgpu_trace.h"
48 * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
50 * @dmabuf: DMA-buf where we attach to
51 * @attach: attachment to add
53 * Add the attachment as user to the exported DMA-buf.
55 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
56 struct dma_buf_attachment *attach)
58 struct drm_gem_object *obj = dmabuf->priv;
59 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
60 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
63 if (pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0)
64 attach->peer2peer = false;
66 r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
67 trace_amdgpu_runpm_reference_dumps(1, __func__);
74 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
75 trace_amdgpu_runpm_reference_dumps(0, __func__);
80 * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
82 * @dmabuf: DMA-buf where we remove the attachment from
83 * @attach: the attachment to remove
85 * Called when an attachment is removed from the DMA-buf.
87 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
88 struct dma_buf_attachment *attach)
90 struct drm_gem_object *obj = dmabuf->priv;
91 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
92 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
94 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
95 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
96 trace_amdgpu_runpm_reference_dumps(0, __func__);
100 * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
102 * @attach: attachment to pin down
104 * Pin the BO which is backing the DMA-buf so that it can't move any more.
106 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
108 struct drm_gem_object *obj = attach->dmabuf->priv;
109 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
111 /* pin buffer into GTT */
112 return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
116 * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
118 * @attach: attachment to unpin
120 * Unpin a previously pinned BO to make it movable again.
122 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
124 struct drm_gem_object *obj = attach->dmabuf->priv;
125 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
131 * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
132 * @attach: DMA-buf attachment
133 * @dir: DMA direction
135 * Makes sure that the shared DMA buffer can be accessed by the target device.
136 * For now, simply pins it to the GTT domain, where it should be accessible by
140 * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
143 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
144 enum dma_data_direction dir)
146 struct dma_buf *dma_buf = attach->dmabuf;
147 struct drm_gem_object *obj = dma_buf->priv;
148 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
149 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
150 struct sg_table *sgt;
153 if (!bo->tbo.pin_count) {
154 /* move buffer into GTT or VRAM */
155 struct ttm_operation_ctx ctx = { false, false };
156 unsigned int domains = AMDGPU_GEM_DOMAIN_GTT;
158 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
160 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
161 domains |= AMDGPU_GEM_DOMAIN_VRAM;
163 amdgpu_bo_placement_from_domain(bo, domains);
164 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
168 } else if (bo->tbo.resource->mem_type != TTM_PL_TT) {
169 return ERR_PTR(-EBUSY);
172 switch (bo->tbo.resource->mem_type) {
174 sgt = drm_prime_pages_to_sg(obj->dev,
176 bo->tbo.ttm->num_pages);
180 if (dma_map_sgtable(attach->dev, sgt, dir,
181 DMA_ATTR_SKIP_CPU_SYNC))
186 r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
187 bo->tbo.base.size, attach->dev,
193 return ERR_PTR(-EINVAL);
201 return ERR_PTR(-EBUSY);
205 * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
206 * @attach: DMA-buf attachment
207 * @sgt: sg_table to unmap
208 * @dir: DMA direction
210 * This is called when a shared DMA buffer no longer needs to be accessible by
211 * another device. For now, simply unpins the buffer from GTT.
213 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
214 struct sg_table *sgt,
215 enum dma_data_direction dir)
217 if (sgt->sgl->page_link) {
218 dma_unmap_sgtable(attach->dev, sgt, dir, 0);
222 amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
227 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
228 * @dma_buf: Shared DMA buffer
229 * @direction: Direction of DMA transfer
231 * This is called before CPU access to the shared DMA buffer's memory. If it's
232 * a read access, the buffer is moved to the GTT domain if possible, for optimal
233 * CPU read performance.
236 * 0 on success or a negative error code on failure.
238 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
239 enum dma_data_direction direction)
241 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
242 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
243 struct ttm_operation_ctx ctx = { true, false };
244 u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
246 bool reads = (direction == DMA_BIDIRECTIONAL ||
247 direction == DMA_FROM_DEVICE);
249 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
253 ret = amdgpu_bo_reserve(bo, false);
254 if (unlikely(ret != 0))
257 if (!bo->tbo.pin_count &&
258 (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
259 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
260 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
263 amdgpu_bo_unreserve(bo);
267 const struct dma_buf_ops amdgpu_dmabuf_ops = {
268 .attach = amdgpu_dma_buf_attach,
269 .detach = amdgpu_dma_buf_detach,
270 .pin = amdgpu_dma_buf_pin,
271 .unpin = amdgpu_dma_buf_unpin,
272 .map_dma_buf = amdgpu_dma_buf_map,
273 .unmap_dma_buf = amdgpu_dma_buf_unmap,
274 .release = drm_gem_dmabuf_release,
275 .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
276 .mmap = drm_gem_dmabuf_mmap,
277 .vmap = drm_gem_dmabuf_vmap,
278 .vunmap = drm_gem_dmabuf_vunmap,
282 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
284 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
286 * The main work is done by the &drm_gem_prime_export helper.
289 * Shared DMA buffer representing the GEM BO from the given device.
291 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
294 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
297 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
298 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
299 return ERR_PTR(-EPERM);
301 buf = drm_gem_prime_export(gobj, flags);
303 buf->ops = &amdgpu_dmabuf_ops;
309 * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
314 * Creates an empty SG BO for DMA-buf import.
317 * A new GEM BO of the given DRM device, representing the memory
318 * described by the given DMA-buf attachment and scatter/gather table.
320 static struct drm_gem_object *
321 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
323 struct dma_resv *resv = dma_buf->resv;
324 struct amdgpu_device *adev = drm_to_adev(dev);
325 struct drm_gem_object *gobj;
326 struct amdgpu_bo *bo;
330 dma_resv_lock(resv, NULL);
332 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
333 struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
335 flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
336 AMDGPU_GEM_CREATE_COHERENT |
337 AMDGPU_GEM_CREATE_EXT_COHERENT |
338 AMDGPU_GEM_CREATE_UNCACHED);
341 ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
342 AMDGPU_GEM_DOMAIN_CPU, flags,
343 ttm_bo_type_sg, resv, &gobj, 0);
347 bo = gem_to_amdgpu_bo(gobj);
348 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
349 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
351 dma_resv_unlock(resv);
355 dma_resv_unlock(resv);
360 * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
362 * @attach: the DMA-buf attachment
364 * Invalidate the DMA-buf attachment, making sure that the we re-create the
365 * mapping before the next use.
368 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
370 struct drm_gem_object *obj = attach->importer_priv;
371 struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
372 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
373 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
374 struct ttm_operation_ctx ctx = { false, false };
375 struct ttm_placement placement = {};
376 struct amdgpu_vm_bo_base *bo_base;
379 /* FIXME: This should be after the "if", but needs a fix to make sure
380 * DMABuf imports are initialized in the right VM list.
382 amdgpu_vm_bo_invalidate(adev, bo, false);
383 if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
386 r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
388 DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
392 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
393 struct amdgpu_vm *vm = bo_base->vm;
394 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
397 /* When we get an error here it means that somebody
398 * else is holding the VM lock and updating page tables
399 * So we can just continue here.
401 r = dma_resv_lock(resv, ticket);
406 /* TODO: This is more problematic and we actually need
407 * to allow page tables updates without holding the
410 if (!dma_resv_trylock(resv))
414 /* Reserve fences for two SDMA page table updates */
415 r = dma_resv_reserve_fences(resv, 2);
417 r = amdgpu_vm_clear_freed(adev, vm, NULL);
419 r = amdgpu_vm_handle_moved(adev, vm, ticket);
421 if (r && r != -EBUSY)
422 DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
425 dma_resv_unlock(resv);
429 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
430 .allow_peer2peer = true,
431 .move_notify = amdgpu_dma_buf_move_notify
435 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
437 * @dma_buf: Shared DMA buffer
439 * Import a dma_buf into a the driver and potentially create a new GEM object.
442 * GEM BO representing the shared DMA buffer for the given device.
444 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
445 struct dma_buf *dma_buf)
447 struct dma_buf_attachment *attach;
448 struct drm_gem_object *obj;
450 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
452 if (obj->dev == dev) {
454 * Importing dmabuf exported from out own gem increases
455 * refcount on gem itself instead of f_count of dmabuf.
457 drm_gem_object_get(obj);
462 obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
466 attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
467 &amdgpu_dma_buf_attach_ops, obj);
468 if (IS_ERR(attach)) {
469 drm_gem_object_put(obj);
470 return ERR_CAST(attach);
473 get_dma_buf(dma_buf);
474 obj->import_attach = attach;
479 * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
481 * @adev: amdgpu_device pointer of the importer
482 * @bo: amdgpu buffer object
485 * True if dmabuf accessible over xgmi, false otherwise.
487 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
488 struct amdgpu_bo *bo)
490 struct drm_gem_object *obj = &bo->tbo.base;
491 struct drm_gem_object *gobj;
493 if (obj->import_attach) {
494 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
496 if (dma_buf->ops != &amdgpu_dmabuf_ops)
497 /* No XGMI with non AMD GPUs */
500 gobj = dma_buf->priv;
501 bo = gem_to_amdgpu_bo(gobj);
504 if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
505 (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))