xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c (revision 2c1ed907520c50326b8f604907a8478b27881a2e)
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 /**
47  * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
48  *
49  * @dmabuf: DMA-buf where we attach to
50  * @attach: attachment to add
51  *
52  * Add the attachment as user to the exported DMA-buf.
53  */
amdgpu_dma_buf_attach(struct dma_buf * dmabuf,struct dma_buf_attachment * attach)54 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
55 				 struct dma_buf_attachment *attach)
56 {
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);
60 
61 	if (pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0)
62 		attach->peer2peer = false;
63 
64 	amdgpu_vm_bo_update_shared(bo);
65 
66 	return 0;
67 }
68 
69 /**
70  * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
71  *
72  * @attach: attachment to pin down
73  *
74  * Pin the BO which is backing the DMA-buf so that it can't move any more.
75  */
amdgpu_dma_buf_pin(struct dma_buf_attachment * attach)76 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
77 {
78 	struct drm_gem_object *obj = attach->dmabuf->priv;
79 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
80 
81 	/* pin buffer into GTT */
82 	return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
83 }
84 
85 /**
86  * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
87  *
88  * @attach: attachment to unpin
89  *
90  * Unpin a previously pinned BO to make it movable again.
91  */
amdgpu_dma_buf_unpin(struct dma_buf_attachment * attach)92 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
93 {
94 	struct drm_gem_object *obj = attach->dmabuf->priv;
95 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
96 
97 	amdgpu_bo_unpin(bo);
98 }
99 
100 /**
101  * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
102  * @attach: DMA-buf attachment
103  * @dir: DMA direction
104  *
105  * Makes sure that the shared DMA buffer can be accessed by the target device.
106  * For now, simply pins it to the GTT domain, where it should be accessible by
107  * all DMA devices.
108  *
109  * Returns:
110  * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
111  * code.
112  */
amdgpu_dma_buf_map(struct dma_buf_attachment * attach,enum dma_data_direction dir)113 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
114 					   enum dma_data_direction dir)
115 {
116 	struct dma_buf *dma_buf = attach->dmabuf;
117 	struct drm_gem_object *obj = dma_buf->priv;
118 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
119 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
120 	struct sg_table *sgt;
121 	long r;
122 
123 	if (!bo->tbo.pin_count) {
124 		/* move buffer into GTT or VRAM */
125 		struct ttm_operation_ctx ctx = { false, false };
126 		unsigned int domains = AMDGPU_GEM_DOMAIN_GTT;
127 
128 		if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
129 		    attach->peer2peer) {
130 			bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
131 			domains |= AMDGPU_GEM_DOMAIN_VRAM;
132 		}
133 		amdgpu_bo_placement_from_domain(bo, domains);
134 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
135 		if (r)
136 			return ERR_PTR(r);
137 
138 	} else if (bo->tbo.resource->mem_type != TTM_PL_TT) {
139 		return ERR_PTR(-EBUSY);
140 	}
141 
142 	switch (bo->tbo.resource->mem_type) {
143 	case TTM_PL_TT:
144 		sgt = drm_prime_pages_to_sg(obj->dev,
145 					    bo->tbo.ttm->pages,
146 					    bo->tbo.ttm->num_pages);
147 		if (IS_ERR(sgt))
148 			return sgt;
149 
150 		if (dma_map_sgtable(attach->dev, sgt, dir,
151 				    DMA_ATTR_SKIP_CPU_SYNC))
152 			goto error_free;
153 		break;
154 
155 	case TTM_PL_VRAM:
156 		r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
157 					      bo->tbo.base.size, attach->dev,
158 					      dir, &sgt);
159 		if (r)
160 			return ERR_PTR(r);
161 		break;
162 	default:
163 		return ERR_PTR(-EINVAL);
164 	}
165 
166 	return sgt;
167 
168 error_free:
169 	sg_free_table(sgt);
170 	kfree(sgt);
171 	return ERR_PTR(-EBUSY);
172 }
173 
174 /**
175  * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
176  * @attach: DMA-buf attachment
177  * @sgt: sg_table to unmap
178  * @dir: DMA direction
179  *
180  * This is called when a shared DMA buffer no longer needs to be accessible by
181  * another device. For now, simply unpins the buffer from GTT.
182  */
amdgpu_dma_buf_unmap(struct dma_buf_attachment * attach,struct sg_table * sgt,enum dma_data_direction dir)183 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
184 				 struct sg_table *sgt,
185 				 enum dma_data_direction dir)
186 {
187 	if (sgt->sgl->page_link) {
188 		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
189 		sg_free_table(sgt);
190 		kfree(sgt);
191 	} else {
192 		amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
193 	}
194 }
195 
196 /**
197  * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
198  * @dma_buf: Shared DMA buffer
199  * @direction: Direction of DMA transfer
200  *
201  * This is called before CPU access to the shared DMA buffer's memory. If it's
202  * a read access, the buffer is moved to the GTT domain if possible, for optimal
203  * CPU read performance.
204  *
205  * Returns:
206  * 0 on success or a negative error code on failure.
207  */
amdgpu_dma_buf_begin_cpu_access(struct dma_buf * dma_buf,enum dma_data_direction direction)208 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
209 					   enum dma_data_direction direction)
210 {
211 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
212 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
213 	struct ttm_operation_ctx ctx = { true, false };
214 	u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
215 	int ret;
216 	bool reads = (direction == DMA_BIDIRECTIONAL ||
217 		      direction == DMA_FROM_DEVICE);
218 
219 	if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
220 		return 0;
221 
222 	/* move to gtt */
223 	ret = amdgpu_bo_reserve(bo, false);
224 	if (unlikely(ret != 0))
225 		return ret;
226 
227 	if (!bo->tbo.pin_count &&
228 	    (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
229 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
230 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
231 	}
232 
233 	amdgpu_bo_unreserve(bo);
234 	return ret;
235 }
236 
237 const struct dma_buf_ops amdgpu_dmabuf_ops = {
238 	.attach = amdgpu_dma_buf_attach,
239 	.pin = amdgpu_dma_buf_pin,
240 	.unpin = amdgpu_dma_buf_unpin,
241 	.map_dma_buf = amdgpu_dma_buf_map,
242 	.unmap_dma_buf = amdgpu_dma_buf_unmap,
243 	.release = drm_gem_dmabuf_release,
244 	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
245 	.mmap = drm_gem_dmabuf_mmap,
246 	.vmap = drm_gem_dmabuf_vmap,
247 	.vunmap = drm_gem_dmabuf_vunmap,
248 };
249 
250 /**
251  * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
252  * @gobj: GEM BO
253  * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
254  *
255  * The main work is done by the &drm_gem_prime_export helper.
256  *
257  * Returns:
258  * Shared DMA buffer representing the GEM BO from the given device.
259  */
amdgpu_gem_prime_export(struct drm_gem_object * gobj,int flags)260 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
261 					int flags)
262 {
263 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
264 	struct dma_buf *buf;
265 
266 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
267 	    bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
268 		return ERR_PTR(-EPERM);
269 
270 	buf = drm_gem_prime_export(gobj, flags);
271 	if (!IS_ERR(buf))
272 		buf->ops = &amdgpu_dmabuf_ops;
273 
274 	return buf;
275 }
276 
277 /**
278  * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
279  *
280  * @dev: DRM device
281  * @dma_buf: DMA-buf
282  *
283  * Creates an empty SG BO for DMA-buf import.
284  *
285  * Returns:
286  * A new GEM BO of the given DRM device, representing the memory
287  * described by the given DMA-buf attachment and scatter/gather table.
288  */
289 static struct drm_gem_object *
amdgpu_dma_buf_create_obj(struct drm_device * dev,struct dma_buf * dma_buf)290 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
291 {
292 	struct dma_resv *resv = dma_buf->resv;
293 	struct amdgpu_device *adev = drm_to_adev(dev);
294 	struct drm_gem_object *gobj;
295 	struct amdgpu_bo *bo;
296 	uint64_t flags = 0;
297 	int ret;
298 
299 	dma_resv_lock(resv, NULL);
300 
301 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
302 		struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
303 
304 		flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
305 					 AMDGPU_GEM_CREATE_COHERENT |
306 					 AMDGPU_GEM_CREATE_EXT_COHERENT |
307 					 AMDGPU_GEM_CREATE_UNCACHED);
308 	}
309 
310 	ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
311 				       AMDGPU_GEM_DOMAIN_CPU, flags,
312 				       ttm_bo_type_sg, resv, &gobj, 0);
313 	if (ret)
314 		goto error;
315 
316 	bo = gem_to_amdgpu_bo(gobj);
317 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
318 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
319 
320 	dma_resv_unlock(resv);
321 	return gobj;
322 
323 error:
324 	dma_resv_unlock(resv);
325 	return ERR_PTR(ret);
326 }
327 
328 /**
329  * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
330  *
331  * @attach: the DMA-buf attachment
332  *
333  * Invalidate the DMA-buf attachment, making sure that the we re-create the
334  * mapping before the next use.
335  */
336 static void
amdgpu_dma_buf_move_notify(struct dma_buf_attachment * attach)337 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
338 {
339 	struct drm_gem_object *obj = attach->importer_priv;
340 	struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
341 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
342 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
343 	struct ttm_operation_ctx ctx = { false, false };
344 	struct ttm_placement placement = {};
345 	struct amdgpu_vm_bo_base *bo_base;
346 	int r;
347 
348 	/* FIXME: This should be after the "if", but needs a fix to make sure
349 	 * DMABuf imports are initialized in the right VM list.
350 	 */
351 	amdgpu_vm_bo_invalidate(bo, false);
352 	if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
353 		return;
354 
355 	r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
356 	if (r) {
357 		DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
358 		return;
359 	}
360 
361 	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
362 		struct amdgpu_vm *vm = bo_base->vm;
363 		struct dma_resv *resv = vm->root.bo->tbo.base.resv;
364 
365 		if (ticket) {
366 			/* When we get an error here it means that somebody
367 			 * else is holding the VM lock and updating page tables
368 			 * So we can just continue here.
369 			 */
370 			r = dma_resv_lock(resv, ticket);
371 			if (r)
372 				continue;
373 
374 		} else {
375 			/* TODO: This is more problematic and we actually need
376 			 * to allow page tables updates without holding the
377 			 * lock.
378 			 */
379 			if (!dma_resv_trylock(resv))
380 				continue;
381 		}
382 
383 		/* Reserve fences for two SDMA page table updates */
384 		r = dma_resv_reserve_fences(resv, 2);
385 		if (!r)
386 			r = amdgpu_vm_clear_freed(adev, vm, NULL);
387 		if (!r)
388 			r = amdgpu_vm_handle_moved(adev, vm, ticket);
389 
390 		if (r && r != -EBUSY)
391 			DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
392 				  r);
393 
394 		dma_resv_unlock(resv);
395 	}
396 }
397 
398 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
399 	.allow_peer2peer = true,
400 	.move_notify = amdgpu_dma_buf_move_notify
401 };
402 
403 /**
404  * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
405  * @dev: DRM device
406  * @dma_buf: Shared DMA buffer
407  *
408  * Import a dma_buf into a the driver and potentially create a new GEM object.
409  *
410  * Returns:
411  * GEM BO representing the shared DMA buffer for the given device.
412  */
amdgpu_gem_prime_import(struct drm_device * dev,struct dma_buf * dma_buf)413 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
414 					       struct dma_buf *dma_buf)
415 {
416 	struct dma_buf_attachment *attach;
417 	struct drm_gem_object *obj;
418 
419 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
420 		obj = dma_buf->priv;
421 		if (obj->dev == dev) {
422 			/*
423 			 * Importing dmabuf exported from out own gem increases
424 			 * refcount on gem itself instead of f_count of dmabuf.
425 			 */
426 			drm_gem_object_get(obj);
427 			return obj;
428 		}
429 	}
430 
431 	obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
432 	if (IS_ERR(obj))
433 		return obj;
434 
435 	attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
436 					&amdgpu_dma_buf_attach_ops, obj);
437 	if (IS_ERR(attach)) {
438 		drm_gem_object_put(obj);
439 		return ERR_CAST(attach);
440 	}
441 
442 	get_dma_buf(dma_buf);
443 	obj->import_attach = attach;
444 	return obj;
445 }
446 
447 /**
448  * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
449  *
450  * @adev: amdgpu_device pointer of the importer
451  * @bo: amdgpu buffer object
452  *
453  * Returns:
454  * True if dmabuf accessible over xgmi, false otherwise.
455  */
amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device * adev,struct amdgpu_bo * bo)456 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
457 				      struct amdgpu_bo *bo)
458 {
459 	struct drm_gem_object *obj = &bo->tbo.base;
460 	struct drm_gem_object *gobj;
461 
462 	if (obj->import_attach) {
463 		struct dma_buf *dma_buf = obj->import_attach->dmabuf;
464 
465 		if (dma_buf->ops != &amdgpu_dmabuf_ops)
466 			/* No XGMI with non AMD GPUs */
467 			return false;
468 
469 		gobj = dma_buf->priv;
470 		bo = gem_to_amdgpu_bo(gobj);
471 	}
472 
473 	if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
474 			(bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))
475 		return true;
476 
477 	return false;
478 }
479