xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c (revision 40288c9206c17eb66a603262e06a58d300d0f279)
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 "amdgpu_ttm.h"
41 #include <drm/amdgpu_drm.h>
42 #include <drm/ttm/ttm_tt.h>
43 #include <linux/dma-buf.h>
44 #include <linux/dma-fence-array.h>
45 #include <linux/pci-p2pdma.h>
46 #include <linux/pm_runtime.h>
47 
48 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops;
49 
50 /**
51  * dma_buf_attach_adev - Helper to get adev of an attachment
52  *
53  * @attach: attachment
54  *
55  * Returns:
56  * A struct amdgpu_device * if the attaching device is an amdgpu device or
57  * partition, NULL otherwise.
58  */
dma_buf_attach_adev(struct dma_buf_attachment * attach)59 static struct amdgpu_device *dma_buf_attach_adev(struct dma_buf_attachment *attach)
60 {
61 	if (attach->importer_ops == &amdgpu_dma_buf_attach_ops) {
62 		struct drm_gem_object *obj = attach->importer_priv;
63 		struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
64 
65 		return amdgpu_ttm_adev(bo->tbo.bdev);
66 	}
67 
68 	return NULL;
69 }
70 
71 /**
72  * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
73  *
74  * @dmabuf: DMA-buf where we attach to
75  * @attach: attachment to add
76  *
77  * Add the attachment as user to the exported DMA-buf.
78  */
amdgpu_dma_buf_attach(struct dma_buf * dmabuf,struct dma_buf_attachment * attach)79 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
80 				 struct dma_buf_attachment *attach)
81 {
82 	struct amdgpu_device *attach_adev = dma_buf_attach_adev(attach);
83 	struct drm_gem_object *obj = dmabuf->priv;
84 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
85 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
86 	int r;
87 
88 	/*
89 	 * Disable peer-to-peer access for DCC-enabled VRAM surfaces on GFX12+.
90 	 * Such buffers cannot be safely accessed over P2P due to device-local
91 	 * compression metadata. Fallback to system-memory path instead.
92 	 * Device supports GFX12 (GC 12.x or newer)
93 	 * BO was created with the AMDGPU_GEM_CREATE_GFX12_DCC flag
94 	 *
95 	 */
96 	if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0) &&
97 	    bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC)
98 		attach->peer2peer = false;
99 
100 	if (!amdgpu_dmabuf_is_xgmi_accessible(attach_adev, bo) &&
101 	    pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0)
102 		attach->peer2peer = false;
103 
104 	/*
105 	 * Only allow P2P while the exporter is active, and keep it active
106 	 * until detach.  With runtime PM disabled take a plain reference so
107 	 * the put in detach stays balanced.
108 	 */
109 	if (attach->peer2peer) {
110 		struct device *dev = adev_to_drm(adev)->dev;
111 		int ret = pm_runtime_get_if_active(dev);
112 
113 		if (!ret)
114 			attach->peer2peer = false;
115 		else if (ret < 0)
116 			pm_runtime_get_noresume(dev);
117 	}
118 
119 	r = dma_resv_lock(bo->tbo.base.resv, NULL);
120 	if (r)
121 		goto err_pm_put;
122 
123 	amdgpu_vm_bo_update_shared(bo);
124 
125 	dma_resv_unlock(bo->tbo.base.resv);
126 
127 	return 0;
128 
129 err_pm_put:
130 	if (attach->peer2peer)
131 		pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
132 	return r;
133 }
134 
135 /**
136  * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
137  *
138  * @dmabuf: DMA-buf where we remove the attachment from
139  * @attach: the attachment to remove
140  *
141  * Drop the runtime PM reference taken in amdgpu_dma_buf_attach().
142  */
amdgpu_dma_buf_detach(struct dma_buf * dmabuf,struct dma_buf_attachment * attach)143 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
144 				  struct dma_buf_attachment *attach)
145 {
146 	struct drm_gem_object *obj = dmabuf->priv;
147 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
148 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
149 
150 	if (attach->peer2peer)
151 		pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
152 }
153 
154 /**
155  * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
156  *
157  * @attach: attachment to pin down
158  *
159  * Pin the BO which is backing the DMA-buf so that it can't move any more.
160  */
amdgpu_dma_buf_pin(struct dma_buf_attachment * attach)161 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
162 {
163 	struct dma_buf *dmabuf = attach->dmabuf;
164 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(dmabuf->priv);
165 	u32 domains = bo->allowed_domains;
166 
167 	dma_resv_assert_held(dmabuf->resv);
168 
169 	/* Try pinning into VRAM to allow P2P with RDMA NICs without ODP
170 	 * support if all attachments can do P2P. If any attachment can't do
171 	 * P2P just pin into GTT instead.
172 	 *
173 	 * To avoid with conflicting pinnings between GPUs and RDMA when move
174 	 * notifiers are disabled, only allow pinning in VRAM when move
175 	 * notiers are enabled.
176 	 */
177 	list_for_each_entry(attach, &dmabuf->attachments, node)
178 		if (!attach->peer2peer)
179 			domains &= ~AMDGPU_GEM_DOMAIN_VRAM;
180 
181 	if (domains & AMDGPU_GEM_DOMAIN_VRAM)
182 		bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
183 
184 	if (WARN_ON(!domains))
185 		return -EINVAL;
186 
187 	return amdgpu_bo_pin(bo, domains);
188 }
189 
190 /**
191  * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
192  *
193  * @attach: attachment to unpin
194  *
195  * Unpin a previously pinned BO to make it movable again.
196  */
amdgpu_dma_buf_unpin(struct dma_buf_attachment * attach)197 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
198 {
199 	struct drm_gem_object *obj = attach->dmabuf->priv;
200 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
201 
202 	amdgpu_bo_unpin(bo);
203 }
204 
205 /**
206  * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
207  * @attach: DMA-buf attachment
208  * @dir: DMA direction
209  *
210  * Makes sure that the shared DMA buffer can be accessed by the target device.
211  * For now, simply pins it to the GTT domain, where it should be accessible by
212  * all DMA devices.
213  *
214  * Returns:
215  * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
216  * code.
217  */
amdgpu_dma_buf_map(struct dma_buf_attachment * attach,enum dma_data_direction dir)218 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
219 					   enum dma_data_direction dir)
220 {
221 	struct dma_buf *dma_buf = attach->dmabuf;
222 	struct drm_gem_object *obj = dma_buf->priv;
223 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
224 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
225 	struct sg_table *sgt;
226 	long r;
227 
228 	if (!bo->tbo.pin_count) {
229 		/* move buffer into GTT or VRAM */
230 		struct ttm_operation_ctx ctx = { false, false };
231 		unsigned int domains = AMDGPU_GEM_DOMAIN_GTT;
232 
233 		if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
234 		    attach->peer2peer) {
235 			bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
236 			domains |= AMDGPU_GEM_DOMAIN_VRAM;
237 		}
238 		amdgpu_bo_placement_from_domain(bo, domains);
239 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
240 		if (r)
241 			return ERR_PTR(r);
242 	}
243 
244 	switch (bo->tbo.resource->mem_type) {
245 	case TTM_PL_TT:
246 		sgt = drm_prime_pages_to_sg(obj->dev,
247 					    bo->tbo.ttm->pages,
248 					    bo->tbo.ttm->num_pages);
249 		if (IS_ERR(sgt))
250 			return sgt;
251 
252 		if (dma_map_sgtable(attach->dev, sgt, dir,
253 				    DMA_ATTR_SKIP_CPU_SYNC))
254 			goto error_free;
255 		break;
256 
257 	case TTM_PL_VRAM:
258 		/* XGMI-accessible memory should never be DMA-mapped */
259 		if (WARN_ON(amdgpu_dmabuf_is_xgmi_accessible(
260 				dma_buf_attach_adev(attach), bo)))
261 			return ERR_PTR(-EINVAL);
262 
263 		r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
264 					      bo->tbo.base.size, attach->dev,
265 					      dir, &sgt);
266 		if (r)
267 			return ERR_PTR(r);
268 		break;
269 
270 	case AMDGPU_PL_MMIO_REMAP:
271 		r = amdgpu_ttm_mmio_remap_alloc_sgt(adev, bo->tbo.resource,
272 						    attach->dev, dir, &sgt);
273 		if (r)
274 			return ERR_PTR(r);
275 		break;
276 
277 	default:
278 		return ERR_PTR(-EINVAL);
279 	}
280 
281 	return sgt;
282 
283 error_free:
284 	sg_free_table(sgt);
285 	kfree(sgt);
286 	return ERR_PTR(-EBUSY);
287 }
288 
289 /**
290  * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
291  * @attach: DMA-buf attachment
292  * @sgt: sg_table to unmap
293  * @dir: DMA direction
294  *
295  * This is called when a shared DMA buffer no longer needs to be accessible by
296  * another device. For now, simply unpins the buffer from GTT.
297  */
amdgpu_dma_buf_unmap(struct dma_buf_attachment * attach,struct sg_table * sgt,enum dma_data_direction dir)298 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
299 				 struct sg_table *sgt,
300 				 enum dma_data_direction dir)
301 {
302 	struct drm_gem_object *obj = attach->dmabuf->priv;
303 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
304 
305 	if (bo->tbo.resource &&
306 	    bo->tbo.resource->mem_type == AMDGPU_PL_MMIO_REMAP) {
307 		amdgpu_ttm_mmio_remap_free_sgt(attach->dev, dir, sgt);
308 		return;
309 	}
310 
311 	if (sg_page(sgt->sgl)) {
312 		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
313 		sg_free_table(sgt);
314 		kfree(sgt);
315 	} else {
316 		amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
317 	}
318 }
319 
320 /**
321  * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
322  * @dma_buf: Shared DMA buffer
323  * @direction: Direction of DMA transfer
324  *
325  * This is called before CPU access to the shared DMA buffer's memory. If it's
326  * a read access, the buffer is moved to the GTT domain if possible, for optimal
327  * CPU read performance.
328  *
329  * Returns:
330  * 0 on success or a negative error code on failure.
331  */
amdgpu_dma_buf_begin_cpu_access(struct dma_buf * dma_buf,enum dma_data_direction direction)332 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
333 					   enum dma_data_direction direction)
334 {
335 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
336 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
337 	struct ttm_operation_ctx ctx = { true, false };
338 	u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
339 	int ret;
340 	bool reads = (direction == DMA_BIDIRECTIONAL ||
341 		      direction == DMA_FROM_DEVICE);
342 
343 	if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
344 		return 0;
345 
346 	/* move to gtt */
347 	ret = amdgpu_bo_reserve(bo, false);
348 	if (unlikely(ret != 0))
349 		return ret;
350 
351 	if (!bo->tbo.pin_count &&
352 	    (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
353 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
354 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
355 	}
356 
357 	amdgpu_bo_unreserve(bo);
358 	return ret;
359 }
360 
amdgpu_dma_buf_vmap(struct dma_buf * dma_buf,struct iosys_map * map)361 static int amdgpu_dma_buf_vmap(struct dma_buf *dma_buf, struct iosys_map *map)
362 {
363 	struct drm_gem_object *obj = dma_buf->priv;
364 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
365 	int ret;
366 
367 	/*
368 	 * Pin to keep buffer in place while it's vmap'ed. The actual
369 	 * domain is not that important as long as it's mapable. Using
370 	 * GTT and VRAM should be compatible with most use cases.
371 	 */
372 	ret = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT | AMDGPU_GEM_DOMAIN_VRAM);
373 	if (ret)
374 		return ret;
375 	ret = drm_gem_dmabuf_vmap(dma_buf, map);
376 	if (ret)
377 		amdgpu_bo_unpin(bo);
378 
379 	return ret;
380 }
381 
amdgpu_dma_buf_vunmap(struct dma_buf * dma_buf,struct iosys_map * map)382 static void amdgpu_dma_buf_vunmap(struct dma_buf *dma_buf, struct iosys_map *map)
383 {
384 	struct drm_gem_object *obj = dma_buf->priv;
385 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
386 
387 	drm_gem_dmabuf_vunmap(dma_buf, map);
388 	amdgpu_bo_unpin(bo);
389 }
390 
391 const struct dma_buf_ops amdgpu_dmabuf_ops = {
392 	.attach = amdgpu_dma_buf_attach,
393 	.detach = amdgpu_dma_buf_detach,
394 	.pin = amdgpu_dma_buf_pin,
395 	.unpin = amdgpu_dma_buf_unpin,
396 	.map_dma_buf = amdgpu_dma_buf_map,
397 	.unmap_dma_buf = amdgpu_dma_buf_unmap,
398 	.release = drm_gem_dmabuf_release,
399 	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
400 	.mmap = drm_gem_dmabuf_mmap,
401 	.vmap = amdgpu_dma_buf_vmap,
402 	.vunmap = amdgpu_dma_buf_vunmap,
403 };
404 
405 /**
406  * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
407  * @gobj: GEM BO
408  * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
409  *
410  * The main work is done by the &drm_gem_prime_export helper.
411  *
412  * Returns:
413  * Shared DMA buffer representing the GEM BO from the given device.
414  */
amdgpu_gem_prime_export(struct drm_gem_object * gobj,int flags)415 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
416 					int flags)
417 {
418 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
419 	struct dma_buf *buf;
420 	struct ttm_operation_ctx ctx = {
421 		.interruptible = true,
422 		.no_wait_gpu = true,
423 		/* We opt to avoid OOM on system pages allocations */
424 		.gfp_retry_mayfail = true,
425 		.allow_res_evict = false,
426 	};
427 	int ret;
428 
429 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
430 	    bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
431 		return ERR_PTR(-EPERM);
432 
433 	ret = ttm_bo_setup_export(&bo->tbo, &ctx);
434 	if (ret)
435 		return ERR_PTR(ret);
436 
437 	buf = drm_gem_prime_export(gobj, flags);
438 	if (!IS_ERR(buf))
439 		buf->ops = &amdgpu_dmabuf_ops;
440 
441 	return buf;
442 }
443 
444 /**
445  * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
446  *
447  * @dev: DRM device
448  * @dma_buf: DMA-buf
449  *
450  * Creates an empty SG BO for DMA-buf import.
451  *
452  * Returns:
453  * A new GEM BO of the given DRM device, representing the memory
454  * described by the given DMA-buf attachment and scatter/gather table.
455  */
456 static struct drm_gem_object *
amdgpu_dma_buf_create_obj(struct drm_device * dev,struct dma_buf * dma_buf)457 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
458 {
459 	struct dma_resv *resv = dma_buf->resv;
460 	struct amdgpu_device *adev = drm_to_adev(dev);
461 	struct drm_gem_object *gobj;
462 	struct amdgpu_bo *bo;
463 	uint64_t flags = 0;
464 	int ret;
465 
466 	dma_resv_lock(resv, NULL);
467 
468 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
469 		struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
470 
471 		flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
472 					 AMDGPU_GEM_CREATE_COHERENT |
473 					 AMDGPU_GEM_CREATE_EXT_COHERENT |
474 					 AMDGPU_GEM_CREATE_UNCACHED);
475 	}
476 
477 	ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
478 				       AMDGPU_GEM_DOMAIN_CPU, flags,
479 				       ttm_bo_type_sg, resv, &gobj, 0);
480 	if (ret)
481 		goto error;
482 
483 	bo = gem_to_amdgpu_bo(gobj);
484 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
485 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
486 
487 	dma_resv_unlock(resv);
488 	return gobj;
489 
490 error:
491 	dma_resv_unlock(resv);
492 	return ERR_PTR(ret);
493 }
494 
495 /**
496  * amdgpu_dma_buf_move_notify - &attach.invalidate_mappings implementation
497  *
498  * @attach: the DMA-buf attachment
499  *
500  * Invalidate the DMA-buf attachment, making sure that the we re-create the
501  * mapping before the next use.
502  */
503 static void
amdgpu_dma_buf_move_notify(struct dma_buf_attachment * attach)504 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
505 {
506 	struct drm_gem_object *obj = attach->importer_priv;
507 	struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
508 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
509 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
510 	struct ttm_operation_ctx ctx = { false, false };
511 	struct ttm_placement placement = {};
512 	struct amdgpu_vm_bo_base *bo_base;
513 	int r;
514 
515 	/* FIXME: This should be after the "if", but needs a fix to make sure
516 	 * DMABuf imports are initialized in the right VM list.
517 	 */
518 	amdgpu_vm_bo_invalidate(bo, false);
519 	if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
520 		return;
521 
522 	r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
523 	if (r) {
524 		DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
525 		return;
526 	}
527 
528 	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
529 		struct amdgpu_vm *vm = bo_base->vm;
530 		struct dma_resv *resv = vm->root.bo->tbo.base.resv;
531 
532 		if (ticket) {
533 			/* When we get an error here it means that somebody
534 			 * else is holding the VM lock and updating page tables
535 			 * So we can just continue here.
536 			 */
537 			r = dma_resv_lock(resv, ticket);
538 			if (r)
539 				continue;
540 
541 		} else {
542 			/* TODO: This is more problematic and we actually need
543 			 * to allow page tables updates without holding the
544 			 * lock.
545 			 */
546 			if (!dma_resv_trylock(resv))
547 				continue;
548 		}
549 
550 		/* Reserve fences for two SDMA page table updates */
551 		r = dma_resv_reserve_fences(resv, 2);
552 		if (!r)
553 			r = amdgpu_vm_clear_freed(adev, vm, NULL);
554 
555 		/* Don't pass 'ticket' to amdgpu_vm_handle_moved: we want the clear=true
556 		 * path to be used otherwise we might update the PT of another process
557 		 * while it's using the BO.
558 		 * With clear=true, amdgpu_vm_bo_update will sync to command submission
559 		 * from the same VM.
560 		 */
561 		if (!r)
562 			r = amdgpu_vm_handle_moved(adev, vm, NULL);
563 
564 		if (r && r != -EBUSY)
565 			DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
566 				  r);
567 
568 		dma_resv_unlock(resv);
569 	}
570 }
571 
572 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
573 	.allow_peer2peer = true,
574 	.invalidate_mappings = amdgpu_dma_buf_move_notify
575 };
576 
577 /**
578  * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
579  * @dev: DRM device
580  * @dma_buf: Shared DMA buffer
581  *
582  * Import a dma_buf into a the driver and potentially create a new GEM object.
583  *
584  * Returns:
585  * GEM BO representing the shared DMA buffer for the given device.
586  */
amdgpu_gem_prime_import(struct drm_device * dev,struct dma_buf * dma_buf)587 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
588 					       struct dma_buf *dma_buf)
589 {
590 	struct dma_buf_attachment *attach;
591 	struct drm_gem_object *obj;
592 
593 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
594 		obj = dma_buf->priv;
595 		if (obj->dev == dev) {
596 			/*
597 			 * Importing dmabuf exported from out own gem increases
598 			 * refcount on gem itself instead of f_count of dmabuf.
599 			 */
600 			drm_gem_object_get(obj);
601 			return obj;
602 		}
603 	}
604 
605 	obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
606 	if (IS_ERR(obj))
607 		return obj;
608 
609 	attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
610 					&amdgpu_dma_buf_attach_ops, obj);
611 	if (IS_ERR(attach)) {
612 		drm_gem_object_put(obj);
613 		return ERR_CAST(attach);
614 	}
615 
616 	get_dma_buf(dma_buf);
617 	obj->import_attach = attach;
618 	return obj;
619 }
620 
621 /**
622  * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
623  *
624  * @adev: amdgpu_device pointer of the importer
625  * @bo: amdgpu buffer object
626  *
627  * Returns:
628  * True if dmabuf accessible over xgmi, false otherwise.
629  */
amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device * adev,struct amdgpu_bo * bo)630 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
631 				      struct amdgpu_bo *bo)
632 {
633 	struct drm_gem_object *obj = &bo->tbo.base;
634 	struct drm_gem_object *gobj;
635 
636 	if (!adev)
637 		return false;
638 
639 	if (drm_gem_is_imported(obj)) {
640 		struct dma_buf *dma_buf = obj->import_attach->dmabuf;
641 
642 		if (dma_buf->ops != &amdgpu_dmabuf_ops)
643 			/* No XGMI with non AMD GPUs */
644 			return false;
645 
646 		gobj = dma_buf->priv;
647 		bo = gem_to_amdgpu_bo(gobj);
648 	}
649 
650 	if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
651 			(bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))
652 		return true;
653 
654 	return false;
655 }
656