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 <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 
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 	int r;
61 
62 	if (pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0)
63 		attach->peer2peer = false;
64 
65 	r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
66 	if (r < 0)
67 		goto out;
68 
69 	return 0;
70 
71 out:
72 	pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
73 	return r;
74 }
75 
76 /**
77  * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
78  *
79  * @dmabuf: DMA-buf where we remove the attachment from
80  * @attach: the attachment to remove
81  *
82  * Called when an attachment is removed from the DMA-buf.
83  */
amdgpu_dma_buf_detach(struct dma_buf * dmabuf,struct dma_buf_attachment * attach)84 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
85 				  struct dma_buf_attachment *attach)
86 {
87 	struct drm_gem_object *obj = dmabuf->priv;
88 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
89 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
90 
91 	pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
92 	pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
93 }
94 
95 /**
96  * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
97  *
98  * @attach: attachment to pin down
99  *
100  * Pin the BO which is backing the DMA-buf so that it can't move any more.
101  */
amdgpu_dma_buf_pin(struct dma_buf_attachment * attach)102 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
103 {
104 	struct drm_gem_object *obj = attach->dmabuf->priv;
105 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
106 
107 	/* pin buffer into GTT */
108 	return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
109 }
110 
111 /**
112  * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
113  *
114  * @attach: attachment to unpin
115  *
116  * Unpin a previously pinned BO to make it movable again.
117  */
amdgpu_dma_buf_unpin(struct dma_buf_attachment * attach)118 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
119 {
120 	struct drm_gem_object *obj = attach->dmabuf->priv;
121 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
122 
123 	amdgpu_bo_unpin(bo);
124 }
125 
126 /**
127  * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
128  * @attach: DMA-buf attachment
129  * @dir: DMA direction
130  *
131  * Makes sure that the shared DMA buffer can be accessed by the target device.
132  * For now, simply pins it to the GTT domain, where it should be accessible by
133  * all DMA devices.
134  *
135  * Returns:
136  * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
137  * code.
138  */
amdgpu_dma_buf_map(struct dma_buf_attachment * attach,enum dma_data_direction dir)139 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
140 					   enum dma_data_direction dir)
141 {
142 	struct dma_buf *dma_buf = attach->dmabuf;
143 	struct drm_gem_object *obj = dma_buf->priv;
144 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
145 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
146 	struct sg_table *sgt;
147 	long r;
148 
149 	if (!bo->tbo.pin_count) {
150 		/* move buffer into GTT or VRAM */
151 		struct ttm_operation_ctx ctx = { false, false };
152 		unsigned domains = AMDGPU_GEM_DOMAIN_GTT;
153 
154 		if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
155 		    attach->peer2peer) {
156 			bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
157 			domains |= AMDGPU_GEM_DOMAIN_VRAM;
158 		}
159 		amdgpu_bo_placement_from_domain(bo, domains);
160 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
161 		if (r)
162 			return ERR_PTR(r);
163 
164 	} else if (!(amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type) &
165 		     AMDGPU_GEM_DOMAIN_GTT)) {
166 		return ERR_PTR(-EBUSY);
167 	}
168 
169 	switch (bo->tbo.resource->mem_type) {
170 	case TTM_PL_TT:
171 		sgt = drm_prime_pages_to_sg(obj->dev,
172 					    bo->tbo.ttm->pages,
173 					    bo->tbo.ttm->num_pages);
174 		if (IS_ERR(sgt))
175 			return sgt;
176 
177 		if (dma_map_sgtable(attach->dev, sgt, dir,
178 				    DMA_ATTR_SKIP_CPU_SYNC))
179 			goto error_free;
180 		break;
181 
182 	case TTM_PL_VRAM:
183 		r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
184 					      bo->tbo.base.size, attach->dev,
185 					      dir, &sgt);
186 		if (r)
187 			return ERR_PTR(r);
188 		break;
189 	default:
190 		return ERR_PTR(-EINVAL);
191 	}
192 
193 	return sgt;
194 
195 error_free:
196 	sg_free_table(sgt);
197 	kfree(sgt);
198 	return ERR_PTR(-EBUSY);
199 }
200 
201 /**
202  * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
203  * @attach: DMA-buf attachment
204  * @sgt: sg_table to unmap
205  * @dir: DMA direction
206  *
207  * This is called when a shared DMA buffer no longer needs to be accessible by
208  * another device. For now, simply unpins the buffer from GTT.
209  */
amdgpu_dma_buf_unmap(struct dma_buf_attachment * attach,struct sg_table * sgt,enum dma_data_direction dir)210 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
211 				 struct sg_table *sgt,
212 				 enum dma_data_direction dir)
213 {
214 	if (sgt->sgl->page_link) {
215 		dma_unmap_sgtable(attach->dev, sgt, dir, 0);
216 		sg_free_table(sgt);
217 		kfree(sgt);
218 	} else {
219 		amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
220 	}
221 }
222 
223 /**
224  * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
225  * @dma_buf: Shared DMA buffer
226  * @direction: Direction of DMA transfer
227  *
228  * This is called before CPU access to the shared DMA buffer's memory. If it's
229  * a read access, the buffer is moved to the GTT domain if possible, for optimal
230  * CPU read performance.
231  *
232  * Returns:
233  * 0 on success or a negative error code on failure.
234  */
amdgpu_dma_buf_begin_cpu_access(struct dma_buf * dma_buf,enum dma_data_direction direction)235 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
236 					   enum dma_data_direction direction)
237 {
238 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
239 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
240 	struct ttm_operation_ctx ctx = { true, false };
241 	u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
242 	int ret;
243 	bool reads = (direction == DMA_BIDIRECTIONAL ||
244 		      direction == DMA_FROM_DEVICE);
245 
246 	if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
247 		return 0;
248 
249 	/* move to gtt */
250 	ret = amdgpu_bo_reserve(bo, false);
251 	if (unlikely(ret != 0))
252 		return ret;
253 
254 	if (!bo->tbo.pin_count &&
255 	    (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
256 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
257 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
258 	}
259 
260 	amdgpu_bo_unreserve(bo);
261 	return ret;
262 }
263 
264 const struct dma_buf_ops amdgpu_dmabuf_ops = {
265 	.attach = amdgpu_dma_buf_attach,
266 	.detach = amdgpu_dma_buf_detach,
267 	.pin = amdgpu_dma_buf_pin,
268 	.unpin = amdgpu_dma_buf_unpin,
269 	.map_dma_buf = amdgpu_dma_buf_map,
270 	.unmap_dma_buf = amdgpu_dma_buf_unmap,
271 	.release = drm_gem_dmabuf_release,
272 	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
273 	.mmap = drm_gem_dmabuf_mmap,
274 	.vmap = drm_gem_dmabuf_vmap,
275 	.vunmap = drm_gem_dmabuf_vunmap,
276 };
277 
278 /**
279  * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
280  * @gobj: GEM BO
281  * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
282  *
283  * The main work is done by the &drm_gem_prime_export helper.
284  *
285  * Returns:
286  * Shared DMA buffer representing the GEM BO from the given device.
287  */
amdgpu_gem_prime_export(struct drm_gem_object * gobj,int flags)288 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
289 					int flags)
290 {
291 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
292 	struct dma_buf *buf;
293 
294 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
295 	    bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
296 		return ERR_PTR(-EPERM);
297 
298 	buf = drm_gem_prime_export(gobj, flags);
299 	if (!IS_ERR(buf))
300 		buf->ops = &amdgpu_dmabuf_ops;
301 
302 	return buf;
303 }
304 
305 /**
306  * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
307  *
308  * @dev: DRM device
309  * @dma_buf: DMA-buf
310  *
311  * Creates an empty SG BO for DMA-buf import.
312  *
313  * Returns:
314  * A new GEM BO of the given DRM device, representing the memory
315  * described by the given DMA-buf attachment and scatter/gather table.
316  */
317 static struct drm_gem_object *
amdgpu_dma_buf_create_obj(struct drm_device * dev,struct dma_buf * dma_buf)318 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
319 {
320 	struct dma_resv *resv = dma_buf->resv;
321 	struct amdgpu_device *adev = drm_to_adev(dev);
322 	struct drm_gem_object *gobj;
323 	struct amdgpu_bo *bo;
324 	uint64_t flags = 0;
325 	int ret;
326 
327 	dma_resv_lock(resv, NULL);
328 
329 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
330 		struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
331 
332 		flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
333 					 AMDGPU_GEM_CREATE_COHERENT |
334 					 AMDGPU_GEM_CREATE_UNCACHED);
335 	}
336 
337 	ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
338 				       AMDGPU_GEM_DOMAIN_CPU, flags,
339 				       ttm_bo_type_sg, resv, &gobj);
340 	if (ret)
341 		goto error;
342 
343 	bo = gem_to_amdgpu_bo(gobj);
344 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
345 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
346 
347 	dma_resv_unlock(resv);
348 	return gobj;
349 
350 error:
351 	dma_resv_unlock(resv);
352 	return ERR_PTR(ret);
353 }
354 
355 /**
356  * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
357  *
358  * @attach: the DMA-buf attachment
359  *
360  * Invalidate the DMA-buf attachment, making sure that the we re-create the
361  * mapping before the next use.
362  */
363 static void
amdgpu_dma_buf_move_notify(struct dma_buf_attachment * attach)364 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
365 {
366 	struct drm_gem_object *obj = attach->importer_priv;
367 	struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
368 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
369 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
370 	struct ttm_operation_ctx ctx = { false, false };
371 	struct ttm_placement placement = {};
372 	struct amdgpu_vm_bo_base *bo_base;
373 	int r;
374 
375 	if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
376 		return;
377 
378 	r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
379 	if (r) {
380 		DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
381 		return;
382 	}
383 
384 	for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
385 		struct amdgpu_vm *vm = bo_base->vm;
386 		struct dma_resv *resv = vm->root.bo->tbo.base.resv;
387 
388 		if (ticket) {
389 			/* When we get an error here it means that somebody
390 			 * else is holding the VM lock and updating page tables
391 			 * So we can just continue here.
392 			 */
393 			r = dma_resv_lock(resv, ticket);
394 			if (r)
395 				continue;
396 
397 		} else {
398 			/* TODO: This is more problematic and we actually need
399 			 * to allow page tables updates without holding the
400 			 * lock.
401 			 */
402 			if (!dma_resv_trylock(resv))
403 				continue;
404 		}
405 
406 		r = amdgpu_vm_clear_freed(adev, vm, NULL);
407 		if (!r)
408 			r = amdgpu_vm_handle_moved(adev, vm);
409 
410 		if (r && r != -EBUSY)
411 			DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
412 				  r);
413 
414 		dma_resv_unlock(resv);
415 	}
416 }
417 
418 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
419 	.allow_peer2peer = true,
420 	.move_notify = amdgpu_dma_buf_move_notify
421 };
422 
423 /**
424  * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
425  * @dev: DRM device
426  * @dma_buf: Shared DMA buffer
427  *
428  * Import a dma_buf into a the driver and potentially create a new GEM object.
429  *
430  * Returns:
431  * GEM BO representing the shared DMA buffer for the given device.
432  */
amdgpu_gem_prime_import(struct drm_device * dev,struct dma_buf * dma_buf)433 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
434 					       struct dma_buf *dma_buf)
435 {
436 	struct dma_buf_attachment *attach;
437 	struct drm_gem_object *obj;
438 
439 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
440 		obj = dma_buf->priv;
441 		if (obj->dev == dev) {
442 			/*
443 			 * Importing dmabuf exported from out own gem increases
444 			 * refcount on gem itself instead of f_count of dmabuf.
445 			 */
446 			drm_gem_object_get(obj);
447 			return obj;
448 		}
449 	}
450 
451 	obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
452 	if (IS_ERR(obj))
453 		return obj;
454 
455 	attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
456 					&amdgpu_dma_buf_attach_ops, obj);
457 	if (IS_ERR(attach)) {
458 		drm_gem_object_put(obj);
459 		return ERR_CAST(attach);
460 	}
461 
462 	get_dma_buf(dma_buf);
463 	obj->import_attach = attach;
464 	return obj;
465 }
466 
467 /**
468  * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
469  *
470  * @adev: amdgpu_device pointer of the importer
471  * @bo: amdgpu buffer object
472  *
473  * Returns:
474  * True if dmabuf accessible over xgmi, false otherwise.
475  */
amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device * adev,struct amdgpu_bo * bo)476 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
477 				      struct amdgpu_bo *bo)
478 {
479 	struct drm_gem_object *obj = &bo->tbo.base;
480 	struct drm_gem_object *gobj;
481 
482 	if (obj->import_attach) {
483 		struct dma_buf *dma_buf = obj->import_attach->dmabuf;
484 
485 		if (dma_buf->ops != &amdgpu_dmabuf_ops)
486 			/* No XGMI with non AMD GPUs */
487 			return false;
488 
489 		gobj = dma_buf->priv;
490 		bo = gem_to_amdgpu_bo(gobj);
491 	}
492 
493 	if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
494 			(bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))
495 		return true;
496 
497 	return false;
498 }
499