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 <drm/amdgpu_drm.h>
38 #include <linux/dma-buf.h>
39 #include <linux/dma-fence-array.h>
40 
41 /**
42  * amdgpu_gem_prime_get_sg_table - &drm_driver.gem_prime_get_sg_table
43  * implementation
44  * @obj: GEM buffer object (BO)
45  *
46  * Returns:
47  * A scatter/gather table for the pinned pages of the BO's memory.
48  */
49 struct sg_table *amdgpu_gem_prime_get_sg_table(struct drm_gem_object *obj)
50 {
51 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
52 	int npages = bo->tbo.num_pages;
53 
54 	return drm_prime_pages_to_sg(bo->tbo.ttm->pages, npages);
55 }
56 
57 /**
58  * amdgpu_gem_prime_vmap - &dma_buf_ops.vmap implementation
59  * @obj: GEM BO
60  *
61  * Sets up an in-kernel virtual mapping of the BO's memory.
62  *
63  * Returns:
64  * The virtual address of the mapping or an error pointer.
65  */
66 void *amdgpu_gem_prime_vmap(struct drm_gem_object *obj)
67 {
68 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
69 	int ret;
70 
71 	ret = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages,
72 			  &bo->dma_buf_vmap);
73 	if (ret)
74 		return ERR_PTR(ret);
75 
76 	return bo->dma_buf_vmap.virtual;
77 }
78 
79 /**
80  * amdgpu_gem_prime_vunmap - &dma_buf_ops.vunmap implementation
81  * @obj: GEM BO
82  * @vaddr: Virtual address (unused)
83  *
84  * Tears down the in-kernel virtual mapping of the BO's memory.
85  */
86 void amdgpu_gem_prime_vunmap(struct drm_gem_object *obj, void *vaddr)
87 {
88 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
89 
90 	ttm_bo_kunmap(&bo->dma_buf_vmap);
91 }
92 
93 /**
94  * amdgpu_gem_prime_mmap - &drm_driver.gem_prime_mmap implementation
95  * @obj: GEM BO
96  * @vma: Virtual memory area
97  *
98  * Sets up a userspace mapping of the BO's memory in the given
99  * virtual memory area.
100  *
101  * Returns:
102  * 0 on success or a negative error code on failure.
103  */
104 int amdgpu_gem_prime_mmap(struct drm_gem_object *obj,
105 			  struct vm_area_struct *vma)
106 {
107 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
108 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
109 	unsigned asize = amdgpu_bo_size(bo);
110 	int ret;
111 
112 	if (!vma->vm_file)
113 		return -ENODEV;
114 
115 	if (adev == NULL)
116 		return -ENODEV;
117 
118 	/* Check for valid size. */
119 	if (asize < vma->vm_end - vma->vm_start)
120 		return -EINVAL;
121 
122 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
123 	    (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)) {
124 		return -EPERM;
125 	}
126 	vma->vm_pgoff += amdgpu_bo_mmap_offset(bo) >> PAGE_SHIFT;
127 
128 	/* prime mmap does not need to check access, so allow here */
129 	ret = drm_vma_node_allow(&obj->vma_node, vma->vm_file->private_data);
130 	if (ret)
131 		return ret;
132 
133 	ret = ttm_bo_mmap(vma->vm_file, vma, &adev->mman.bdev);
134 	drm_vma_node_revoke(&obj->vma_node, vma->vm_file->private_data);
135 
136 	return ret;
137 }
138 
139 static int
140 __reservation_object_make_exclusive(struct reservation_object *obj)
141 {
142 	struct dma_fence **fences;
143 	unsigned int count;
144 	int r;
145 
146 	if (!reservation_object_get_list(obj)) /* no shared fences to convert */
147 		return 0;
148 
149 	r = reservation_object_get_fences_rcu(obj, NULL, &count, &fences);
150 	if (r)
151 		return r;
152 
153 	if (count == 0) {
154 		/* Now that was unexpected. */
155 	} else if (count == 1) {
156 		reservation_object_add_excl_fence(obj, fences[0]);
157 		dma_fence_put(fences[0]);
158 		kfree(fences);
159 	} else {
160 		struct dma_fence_array *array;
161 
162 		array = dma_fence_array_create(count, fences,
163 					       dma_fence_context_alloc(1), 0,
164 					       false);
165 		if (!array)
166 			goto err_fences_put;
167 
168 		reservation_object_add_excl_fence(obj, &array->base);
169 		dma_fence_put(&array->base);
170 	}
171 
172 	return 0;
173 
174 err_fences_put:
175 	while (count--)
176 		dma_fence_put(fences[count]);
177 	kfree(fences);
178 	return -ENOMEM;
179 }
180 
181 /**
182  * amdgpu_dma_buf_map_attach - &dma_buf_ops.attach implementation
183  * @dma_buf: Shared DMA buffer
184  * @attach: DMA-buf attachment
185  *
186  * Makes sure that the shared DMA buffer can be accessed by the target device.
187  * For now, simply pins it to the GTT domain, where it should be accessible by
188  * all DMA devices.
189  *
190  * Returns:
191  * 0 on success or a negative error code on failure.
192  */
193 static int amdgpu_dma_buf_map_attach(struct dma_buf *dma_buf,
194 				     struct dma_buf_attachment *attach)
195 {
196 	struct drm_gem_object *obj = dma_buf->priv;
197 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
198 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
199 	long r;
200 
201 	r = drm_gem_map_attach(dma_buf, attach);
202 	if (r)
203 		return r;
204 
205 	r = amdgpu_bo_reserve(bo, false);
206 	if (unlikely(r != 0))
207 		goto error_detach;
208 
209 
210 	if (attach->dev->driver != adev->dev->driver) {
211 		/*
212 		 * We only create shared fences for internal use, but importers
213 		 * of the dmabuf rely on exclusive fences for implicitly
214 		 * tracking write hazards. As any of the current fences may
215 		 * correspond to a write, we need to convert all existing
216 		 * fences on the reservation object into a single exclusive
217 		 * fence.
218 		 */
219 		r = __reservation_object_make_exclusive(bo->tbo.resv);
220 		if (r)
221 			goto error_unreserve;
222 	}
223 
224 	/* pin buffer into GTT */
225 	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
226 	if (r)
227 		goto error_unreserve;
228 
229 	if (attach->dev->driver != adev->dev->driver)
230 		bo->prime_shared_count++;
231 
232 error_unreserve:
233 	amdgpu_bo_unreserve(bo);
234 
235 error_detach:
236 	if (r)
237 		drm_gem_map_detach(dma_buf, attach);
238 	return r;
239 }
240 
241 /**
242  * amdgpu_dma_buf_map_detach - &dma_buf_ops.detach implementation
243  * @dma_buf: Shared DMA buffer
244  * @attach: DMA-buf attachment
245  *
246  * This is called when a shared DMA buffer no longer needs to be accessible by
247  * another device. For now, simply unpins the buffer from GTT.
248  */
249 static void amdgpu_dma_buf_map_detach(struct dma_buf *dma_buf,
250 				      struct dma_buf_attachment *attach)
251 {
252 	struct drm_gem_object *obj = dma_buf->priv;
253 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
254 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
255 	int ret = 0;
256 
257 	ret = amdgpu_bo_reserve(bo, true);
258 	if (unlikely(ret != 0))
259 		goto error;
260 
261 	amdgpu_bo_unpin(bo);
262 	if (attach->dev->driver != adev->dev->driver && bo->prime_shared_count)
263 		bo->prime_shared_count--;
264 	amdgpu_bo_unreserve(bo);
265 
266 error:
267 	drm_gem_map_detach(dma_buf, attach);
268 }
269 
270 /**
271  * amdgpu_gem_prime_res_obj - &drm_driver.gem_prime_res_obj implementation
272  * @obj: GEM BO
273  *
274  * Returns:
275  * The BO's reservation object.
276  */
277 struct reservation_object *amdgpu_gem_prime_res_obj(struct drm_gem_object *obj)
278 {
279 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
280 
281 	return bo->tbo.resv;
282 }
283 
284 /**
285  * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
286  * @dma_buf: Shared DMA buffer
287  * @direction: Direction of DMA transfer
288  *
289  * This is called before CPU access to the shared DMA buffer's memory. If it's
290  * a read access, the buffer is moved to the GTT domain if possible, for optimal
291  * CPU read performance.
292  *
293  * Returns:
294  * 0 on success or a negative error code on failure.
295  */
296 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
297 					   enum dma_data_direction direction)
298 {
299 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
300 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
301 	struct ttm_operation_ctx ctx = { true, false };
302 	u32 domain = amdgpu_display_supported_domains(adev);
303 	int ret;
304 	bool reads = (direction == DMA_BIDIRECTIONAL ||
305 		      direction == DMA_FROM_DEVICE);
306 
307 	if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
308 		return 0;
309 
310 	/* move to gtt */
311 	ret = amdgpu_bo_reserve(bo, false);
312 	if (unlikely(ret != 0))
313 		return ret;
314 
315 	if (!bo->pin_count && (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
316 		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
317 		ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
318 	}
319 
320 	amdgpu_bo_unreserve(bo);
321 	return ret;
322 }
323 
324 const struct dma_buf_ops amdgpu_dmabuf_ops = {
325 	.attach = amdgpu_dma_buf_map_attach,
326 	.detach = amdgpu_dma_buf_map_detach,
327 	.map_dma_buf = drm_gem_map_dma_buf,
328 	.unmap_dma_buf = drm_gem_unmap_dma_buf,
329 	.release = drm_gem_dmabuf_release,
330 	.begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
331 	.mmap = drm_gem_dmabuf_mmap,
332 	.vmap = drm_gem_dmabuf_vmap,
333 	.vunmap = drm_gem_dmabuf_vunmap,
334 };
335 
336 /**
337  * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
338  * @dev: DRM device
339  * @gobj: GEM BO
340  * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
341  *
342  * The main work is done by the &drm_gem_prime_export helper, which in turn
343  * uses &amdgpu_gem_prime_res_obj.
344  *
345  * Returns:
346  * Shared DMA buffer representing the GEM BO from the given device.
347  */
348 struct dma_buf *amdgpu_gem_prime_export(struct drm_device *dev,
349 					struct drm_gem_object *gobj,
350 					int flags)
351 {
352 	struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
353 	struct dma_buf *buf;
354 
355 	if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
356 	    bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
357 		return ERR_PTR(-EPERM);
358 
359 	buf = drm_gem_prime_export(dev, gobj, flags);
360 	if (!IS_ERR(buf)) {
361 		buf->file->f_mapping = dev->anon_inode->i_mapping;
362 		buf->ops = &amdgpu_dmabuf_ops;
363 	}
364 
365 	return buf;
366 }
367 
368 /**
369  * amdgpu_gem_prime_import_sg_table - &drm_driver.gem_prime_import_sg_table
370  * implementation
371  * @dev: DRM device
372  * @attach: DMA-buf attachment
373  * @sg: Scatter/gather table
374  *
375  * Imports shared DMA buffer memory exported by another device.
376  *
377  * Returns:
378  * A new GEM BO of the given DRM device, representing the memory
379  * described by the given DMA-buf attachment and scatter/gather table.
380  */
381 struct drm_gem_object *
382 amdgpu_gem_prime_import_sg_table(struct drm_device *dev,
383 				 struct dma_buf_attachment *attach,
384 				 struct sg_table *sg)
385 {
386 	struct reservation_object *resv = attach->dmabuf->resv;
387 	struct amdgpu_device *adev = dev->dev_private;
388 	struct amdgpu_bo *bo;
389 	struct amdgpu_bo_param bp;
390 	int ret;
391 
392 	memset(&bp, 0, sizeof(bp));
393 	bp.size = attach->dmabuf->size;
394 	bp.byte_align = PAGE_SIZE;
395 	bp.domain = AMDGPU_GEM_DOMAIN_CPU;
396 	bp.flags = 0;
397 	bp.type = ttm_bo_type_sg;
398 	bp.resv = resv;
399 	ww_mutex_lock(&resv->lock, NULL);
400 	ret = amdgpu_bo_create(adev, &bp, &bo);
401 	if (ret)
402 		goto error;
403 
404 	bo->tbo.sg = sg;
405 	bo->tbo.ttm->sg = sg;
406 	bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
407 	bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
408 	if (attach->dmabuf->ops != &amdgpu_dmabuf_ops)
409 		bo->prime_shared_count = 1;
410 
411 	ww_mutex_unlock(&resv->lock);
412 	return &bo->gem_base;
413 
414 error:
415 	ww_mutex_unlock(&resv->lock);
416 	return ERR_PTR(ret);
417 }
418 
419 /**
420  * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
421  * @dev: DRM device
422  * @dma_buf: Shared DMA buffer
423  *
424  * The main work is done by the &drm_gem_prime_import helper, which in turn
425  * uses &amdgpu_gem_prime_import_sg_table.
426  *
427  * Returns:
428  * GEM BO representing the shared DMA buffer for the given device.
429  */
430 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
431 					    struct dma_buf *dma_buf)
432 {
433 	struct drm_gem_object *obj;
434 
435 	if (dma_buf->ops == &amdgpu_dmabuf_ops) {
436 		obj = dma_buf->priv;
437 		if (obj->dev == dev) {
438 			/*
439 			 * Importing dmabuf exported from out own gem increases
440 			 * refcount on gem itself instead of f_count of dmabuf.
441 			 */
442 			drm_gem_object_get(obj);
443 			return obj;
444 		}
445 	}
446 
447 	return drm_gem_prime_import(dev, dma_buf);
448 }
449