xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c (revision 3f1c07fc21c68bd3bd2df9d2c9441f6485e934d9)
1 /*
2  * Copyright 2019 Advanced Micro Devices, Inc.
3  *
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:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
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.
21  *
22  * based on nouveau_prime.c
23  *
24  * Authors: Alex Deucher
25  */
26 
27 /**
28  * DOC: PRIME Buffer Sharing
29  *
30  * The following callback implementations are used for :ref:`sharing GEM buffer
31  * objects between different devices via PRIME <prime_buffer_sharing>`.
32  */
33 
34 #include "amdgpu.h"
35 #include "amdgpu_display.h"
36 #include "amdgpu_gem.h"
37 #include "amdgpu_dma_buf.h"
38 #include "amdgpu_xgmi.h"
39 #include "amdgpu_vm.h"
40 #include <drm/amdgpu_drm.h>
41 #include <drm/ttm/ttm_tt.h>
42 #include <linux/dma-buf.h>
43 #include <linux/dma-fence-array.h>
44 #include <linux/pci-p2pdma.h>
45 
46 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops;
47 
48 /**
49  * dma_buf_attach_adev - Helper to get adev of an attachment
50  *
51  * @attach: attachment
52  *
53  * Returns:
54  * A struct amdgpu_device * if the attaching device is an amdgpu device or
55  * partition, NULL otherwise.
56  */
dma_buf_attach_adev(struct dma_buf_attachment * attach)57 static struct amdgpu_device *dma_buf_attach_adev(struct dma_buf_attachment *attach)
58 {
59 	if (attach->importer_ops == &amdgpu_dma_buf_attach_ops) {
60 		struct drm_gem_object *obj = attach->importer_priv;
61 		struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
62 
63 		return amdgpu_ttm_adev(bo->tbo.bdev);
64 	}
65 
66 	return NULL;
67 }
68 
69 /**
70  * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
71  *
72  * @dmabuf: DMA-buf where we attach to
73  * @attach: attachment to add
74  *
75  * Add the attachment as user to the exported DMA-buf.
76  */
amdgpu_dma_buf_attach(struct dma_buf * dmabuf,struct dma_buf_attachment * attach)77 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
78 				 struct dma_buf_attachment *attach)
79 {
80 	struct amdgpu_device *attach_adev = dma_buf_attach_adev(attach);
81 	struct drm_gem_object *obj = dmabuf->priv;
82 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
83 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
84 	int r;
85 
86 	/*
87 	 * Disable peer-to-peer access for DCC-enabled VRAM surfaces on GFX12+.
88 	 * Such buffers cannot be safely accessed over P2P due to device-local
89 	 * compression metadata. Fallback to system-memory path instead.
90 	 * Device supports GFX12 (GC 12.x or newer)
91 	 * BO was created with the AMDGPU_GEM_CREATE_GFX12_DCC flag
92 	 *
93 	 */
94 	if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0) &&
95 	    bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC)
96 		attach->peer2peer = false;
97 
98 	/*
99 	 * Disable peer-to-peer access for DCC-enabled VRAM surfaces on GFX12+.
100 	 * Such buffers cannot be safely accessed over P2P due to device-local
101 	 * compression metadata. Fallback to system-memory path instead.
102 	 * Device supports GFX12 (GC 12.x or newer)
103 	 * BO was created with the AMDGPU_GEM_CREATE_GFX12_DCC flag
104 	 *
105 	 */
106 	if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0) &&
107 	    bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC)
108 		attach->peer2peer = false;
109 
110 	if (!amdgpu_dmabuf_is_xgmi_accessible(attach_adev, bo) &&
111 	    pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0)
112 		attach->peer2peer = false;
113 
114 	r = dma_resv_lock(bo->tbo.base.resv, NULL);
115 	if (r)
116 		return r;
117 
118 	amdgpu_vm_bo_update_shared(bo);
119 
120 	dma_resv_unlock(bo->tbo.base.resv);
121 
122 	return 0;
123 }
124 
125 /**
126  * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
127  *
128  * @attach: attachment to pin down
129  *
130  * Pin the BO which is backing the DMA-buf so that it can't move any more.
131  */
amdgpu_dma_buf_pin(struct dma_buf_attachment * attach)132 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
133 {
134 	struct dma_buf *dmabuf = attach->dmabuf;
135 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(dmabuf->priv);
136 	u32 domains = bo->allowed_domains;
137 
138 	dma_resv_assert_held(dmabuf->resv);
139 
140 	/* Try pinning into VRAM to allow P2P with RDMA NICs without ODP
141 	 * support if all attachments can do P2P. If any attachment can't do
142 	 * P2P just pin into GTT instead.
143 	 *
144 	 * To avoid with conflicting pinnings between GPUs and RDMA when move
145 	 * notifiers are disabled, only allow pinning in VRAM when move
146 	 * notiers are enabled.
147 	 */
148 	if (!IS_ENABLED(CONFIG_DMABUF_MOVE_NOTIFY)) {
149 		domains &= ~AMDGPU_GEM_DOMAIN_VRAM;
150 	} else {
151 		list_for_each_entry(attach, &dmabuf->attachments, node)
152 			if (!attach->peer2peer)
153 				domains &= ~AMDGPU_GEM_DOMAIN_VRAM;
154 	}
155 
156 	if (domains & AMDGPU_GEM_DOMAIN_VRAM)
157 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
158 
159 	if (WARN_ON(!domains))
160 		return -EINVAL;
161 
162 	return amdgpu_bo_pin(bo, domains);
163 }
164 
165 /**
166  * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
167  *
168  * @attach: attachment to unpin
169  *
170  * Unpin a previously pinned BO to make it movable again.
171  */
amdgpu_dma_buf_unpin(struct dma_buf_attachment * attach)172 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
173 {
174 	struct drm_gem_object *obj = attach->dmabuf->priv;
175 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
176 
177 	amdgpu_bo_unpin(bo);
178 }
179 
180 /**
181  * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
182  * @attach: DMA-buf attachment
183  * @dir: DMA direction
184  *
185  * Makes sure that the shared DMA buffer can be accessed by the target device.
186  * For now, simply pins it to the GTT domain, where it should be accessible by
187  * all DMA devices.
188  *
189  * Returns:
190  * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
191  * code.
192  */
amdgpu_dma_buf_map(struct dma_buf_attachment * attach,enum dma_data_direction dir)193 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
194 					   enum dma_data_direction dir)
195 {
196 	struct dma_buf *dma_buf = attach->dmabuf;
197 	struct drm_gem_object *obj = dma_buf->priv;
198 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
199 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
200 	struct sg_table *sgt;
201 	long r;
202 
203 	if (!bo->tbo.pin_count) {
204 		/* move buffer into GTT or VRAM */
205 		struct ttm_operation_ctx ctx = { false, false };
206 		unsigned int domains = AMDGPU_GEM_DOMAIN_GTT;
207 
208 		if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
209 		    attach->peer2peer) {
210 			bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
211 			domains |= AMDGPU_GEM_DOMAIN_VRAM;
212 		}
213 		amdgpu_bo_placement_from_domain(bo, domains);
214 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
215 		if (r)
216 			return ERR_PTR(r);
217 	}
218 
219 	switch (bo->tbo.resource->mem_type) {
220 	case TTM_PL_TT:
221 		sgt = drm_prime_pages_to_sg(obj->dev,
222 					    bo->tbo.ttm->pages,
223 					    bo->tbo.ttm->num_pages);
224 		if (IS_ERR(sgt))
225 			return sgt;
226 
227 		if (dma_map_sgtable(attach->dev, sgt, dir,
228 				    DMA_ATTR_SKIP_CPU_SYNC))
229 			goto error_free;
230 		break;
231 
232 	case TTM_PL_VRAM:
233 		/* XGMI-accessible memory should never be DMA-mapped */
234 		if (WARN_ON(amdgpu_dmabuf_is_xgmi_accessible(
235 				dma_buf_attach_adev(attach), bo)))
236 			return ERR_PTR(-EINVAL);
237 
238 		r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
239 					      bo->tbo.base.size, attach->dev,
240 					      dir, &sgt);
241 		if (r)
242 			return ERR_PTR(r);
243 		break;
244 	default:
245 		return ERR_PTR(-EINVAL);
246 	}
247 
248 	return sgt;
249 
250 error_free:
251 	sg_free_table(sgt);
252 	kfree(sgt);
253 	return ERR_PTR(-EBUSY);
254 }
255 
256 /**
257  * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
258  * @attach: DMA-buf attachment
259  * @sgt: sg_table to unmap
260  * @dir: DMA direction
261  *
262  * This is called when a shared DMA buffer no longer needs to be accessible by
263  * another device. For now, simply unpins the buffer from GTT.
264  */
amdgpu_dma_buf_unmap(struct dma_buf_attachment * attach,struct sg_table * sgt,enum dma_data_direction dir)265 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
266 				 struct sg_table *sgt,
267 				 enum dma_data_direction dir)
268 {
269 	if (sg_page(sgt->sgl)) {
270 		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
271 		sg_free_table(sgt);
272 		kfree(sgt);
273 	} else {
274 		amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
275 	}
276 }
277 
278 /**
279  * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
280  * @dma_buf: Shared DMA buffer
281  * @direction: Direction of DMA transfer
282  *
283  * This is called before CPU access to the shared DMA buffer's memory. If it's
284  * a read access, the buffer is moved to the GTT domain if possible, for optimal
285  * CPU read performance.
286  *
287  * Returns:
288  * 0 on success or a negative error code on failure.
289  */
amdgpu_dma_buf_begin_cpu_access(struct dma_buf * dma_buf,enum dma_data_direction direction)290 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
291 					   enum dma_data_direction direction)
292 {
293 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
294 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
295 	struct ttm_operation_ctx ctx = { true, false };
296 	u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
297 	int ret;
298 	bool reads = (direction == DMA_BIDIRECTIONAL ||
299 		      direction == DMA_FROM_DEVICE);
300 
301 	if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
302 		return 0;
303 
304 	/* move to gtt */
305 	ret = amdgpu_bo_reserve(bo, false);
306 	if (unlikely(ret != 0))
307 		return ret;
308 
309 	if (!bo->tbo.pin_count &&
310 	    (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
311 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
312 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
313 	}
314 
315 	amdgpu_bo_unreserve(bo);
316 	return ret;
317 }
318 
amdgpu_dma_buf_vmap(struct dma_buf * dma_buf,struct iosys_map * map)319 static int amdgpu_dma_buf_vmap(struct dma_buf *dma_buf, struct iosys_map *map)
320 {
321 	struct drm_gem_object *obj = dma_buf->priv;
322 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
323 	int ret;
324 
325 	/*
326 	 * Pin to keep buffer in place while it's vmap'ed. The actual
327 	 * domain is not that important as long as it's mapable. Using
328 	 * GTT and VRAM should be compatible with most use cases.
329 	 */
330 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT | AMDGPU_GEM_DOMAIN_VRAM);
331 	if (ret)
332 		return ret;
333 	ret = drm_gem_dmabuf_vmap(dma_buf, map);
334 	if (ret)
335 		amdgpu_bo_unpin(bo);
336 
337 	return ret;
338 }
339 
amdgpu_dma_buf_vunmap(struct dma_buf * dma_buf,struct iosys_map * map)340 static void amdgpu_dma_buf_vunmap(struct dma_buf *dma_buf, struct iosys_map *map)
341 {
342 	struct drm_gem_object *obj = dma_buf->priv;
343 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
344 
345 	drm_gem_dmabuf_vunmap(dma_buf, map);
346 	amdgpu_bo_unpin(bo);
347 }
348 
349 const struct dma_buf_ops amdgpu_dmabuf_ops = {
350 	.attach = amdgpu_dma_buf_attach,
351 	.pin = amdgpu_dma_buf_pin,
352 	.unpin = amdgpu_dma_buf_unpin,
353 	.map_dma_buf = amdgpu_dma_buf_map,
354 	.unmap_dma_buf = amdgpu_dma_buf_unmap,
355 	.release = drm_gem_dmabuf_release,
356 	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
357 	.mmap = drm_gem_dmabuf_mmap,
358 	.vmap = amdgpu_dma_buf_vmap,
359 	.vunmap = amdgpu_dma_buf_vunmap,
360 };
361 
362 /**
363  * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
364  * @gobj: GEM BO
365  * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
366  *
367  * The main work is done by the &drm_gem_prime_export helper.
368  *
369  * Returns:
370  * Shared DMA buffer representing the GEM BO from the given device.
371  */
amdgpu_gem_prime_export(struct drm_gem_object * gobj,int flags)372 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
373 					int flags)
374 {
375 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
376 	struct dma_buf *buf;
377 	struct ttm_operation_ctx ctx = {
378 		.interruptible = true,
379 		.no_wait_gpu = true,
380 		/* We opt to avoid OOM on system pages allocations */
381 		.gfp_retry_mayfail = true,
382 		.allow_res_evict = false,
383 	};
384 	int ret;
385 
386 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
387 	    bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
388 		return ERR_PTR(-EPERM);
389 
390 	ret = ttm_bo_setup_export(&bo->tbo, &ctx);
391 	if (ret)
392 		return ERR_PTR(ret);
393 
394 	buf = drm_gem_prime_export(gobj, flags);
395 	if (!IS_ERR(buf))
396 		buf->ops = &amdgpu_dmabuf_ops;
397 
398 	return buf;
399 }
400 
401 /**
402  * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
403  *
404  * @dev: DRM device
405  * @dma_buf: DMA-buf
406  *
407  * Creates an empty SG BO for DMA-buf import.
408  *
409  * Returns:
410  * A new GEM BO of the given DRM device, representing the memory
411  * described by the given DMA-buf attachment and scatter/gather table.
412  */
413 static struct drm_gem_object *
amdgpu_dma_buf_create_obj(struct drm_device * dev,struct dma_buf * dma_buf)414 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
415 {
416 	struct dma_resv *resv = dma_buf->resv;
417 	struct amdgpu_device *adev = drm_to_adev(dev);
418 	struct drm_gem_object *gobj;
419 	struct amdgpu_bo *bo;
420 	uint64_t flags = 0;
421 	int ret;
422 
423 	dma_resv_lock(resv, NULL);
424 
425 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
426 		struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
427 
428 		flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
429 					 AMDGPU_GEM_CREATE_COHERENT |
430 					 AMDGPU_GEM_CREATE_EXT_COHERENT |
431 					 AMDGPU_GEM_CREATE_UNCACHED);
432 	}
433 
434 	ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
435 				       AMDGPU_GEM_DOMAIN_CPU, flags,
436 				       ttm_bo_type_sg, resv, &gobj, 0);
437 	if (ret)
438 		goto error;
439 
440 	bo = gem_to_amdgpu_bo(gobj);
441 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
442 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
443 
444 	dma_resv_unlock(resv);
445 	return gobj;
446 
447 error:
448 	dma_resv_unlock(resv);
449 	return ERR_PTR(ret);
450 }
451 
452 /**
453  * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
454  *
455  * @attach: the DMA-buf attachment
456  *
457  * Invalidate the DMA-buf attachment, making sure that the we re-create the
458  * mapping before the next use.
459  */
460 static void
amdgpu_dma_buf_move_notify(struct dma_buf_attachment * attach)461 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
462 {
463 	struct drm_gem_object *obj = attach->importer_priv;
464 	struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
465 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
466 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
467 	struct ttm_operation_ctx ctx = { false, false };
468 	struct ttm_placement placement = {};
469 	struct amdgpu_vm_bo_base *bo_base;
470 	int r;
471 
472 	/* FIXME: This should be after the "if", but needs a fix to make sure
473 	 * DMABuf imports are initialized in the right VM list.
474 	 */
475 	amdgpu_vm_bo_invalidate(bo, false);
476 	if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
477 		return;
478 
479 	r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
480 	if (r) {
481 		DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
482 		return;
483 	}
484 
485 	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
486 		struct amdgpu_vm *vm = bo_base->vm;
487 		struct dma_resv *resv = vm->root.bo->tbo.base.resv;
488 
489 		if (ticket) {
490 			/* When we get an error here it means that somebody
491 			 * else is holding the VM lock and updating page tables
492 			 * So we can just continue here.
493 			 */
494 			r = dma_resv_lock(resv, ticket);
495 			if (r)
496 				continue;
497 
498 		} else {
499 			/* TODO: This is more problematic and we actually need
500 			 * to allow page tables updates without holding the
501 			 * lock.
502 			 */
503 			if (!dma_resv_trylock(resv))
504 				continue;
505 		}
506 
507 		/* Reserve fences for two SDMA page table updates */
508 		r = dma_resv_reserve_fences(resv, 2);
509 		if (!r)
510 			r = amdgpu_vm_clear_freed(adev, vm, NULL);
511 		if (!r)
512 			r = amdgpu_vm_handle_moved(adev, vm, ticket);
513 
514 		if (r && r != -EBUSY)
515 			DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
516 				  r);
517 
518 		dma_resv_unlock(resv);
519 	}
520 }
521 
522 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
523 	.allow_peer2peer = true,
524 	.move_notify = amdgpu_dma_buf_move_notify
525 };
526 
527 /**
528  * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
529  * @dev: DRM device
530  * @dma_buf: Shared DMA buffer
531  *
532  * Import a dma_buf into a the driver and potentially create a new GEM object.
533  *
534  * Returns:
535  * GEM BO representing the shared DMA buffer for the given device.
536  */
amdgpu_gem_prime_import(struct drm_device * dev,struct dma_buf * dma_buf)537 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
538 					       struct dma_buf *dma_buf)
539 {
540 	struct dma_buf_attachment *attach;
541 	struct drm_gem_object *obj;
542 
543 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
544 		obj = dma_buf->priv;
545 		if (obj->dev == dev) {
546 			/*
547 			 * Importing dmabuf exported from out own gem increases
548 			 * refcount on gem itself instead of f_count of dmabuf.
549 			 */
550 			drm_gem_object_get(obj);
551 			return obj;
552 		}
553 	}
554 
555 	obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
556 	if (IS_ERR(obj))
557 		return obj;
558 
559 	attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
560 					&amdgpu_dma_buf_attach_ops, obj);
561 	if (IS_ERR(attach)) {
562 		drm_gem_object_put(obj);
563 		return ERR_CAST(attach);
564 	}
565 
566 	get_dma_buf(dma_buf);
567 	obj->import_attach = attach;
568 	return obj;
569 }
570 
571 /**
572  * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
573  *
574  * @adev: amdgpu_device pointer of the importer
575  * @bo: amdgpu buffer object
576  *
577  * Returns:
578  * True if dmabuf accessible over xgmi, false otherwise.
579  */
amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device * adev,struct amdgpu_bo * bo)580 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
581 				      struct amdgpu_bo *bo)
582 {
583 	struct drm_gem_object *obj = &bo->tbo.base;
584 	struct drm_gem_object *gobj;
585 
586 	if (!adev)
587 		return false;
588 
589 	if (drm_gem_is_imported(obj)) {
590 		struct dma_buf *dma_buf = obj->import_attach->dmabuf;
591 
592 		if (dma_buf->ops != &amdgpu_dmabuf_ops)
593 			/* No XGMI with non AMD GPUs */
594 			return false;
595 
596 		gobj = dma_buf->priv;
597 		bo = gem_to_amdgpu_bo(gobj);
598 	}
599 
600 	if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
601 			(bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))
602 		return true;
603 
604 	return false;
605 }
606