Loading...
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#include "amdgpu_trace.h"
46
47/**
48 * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
49 *
50 * @dmabuf: DMA-buf where we attach to
51 * @attach: attachment to add
52 *
53 * Add the attachment as user to the exported DMA-buf.
54 */
55static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
56 struct dma_buf_attachment *attach)
57{
58 struct drm_gem_object *obj = dmabuf->priv;
59 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
60 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
61 int r;
62
63 if (pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0)
64 attach->peer2peer = false;
65
66 r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
67 trace_amdgpu_runpm_reference_dumps(1, __func__);
68 if (r < 0)
69 goto out;
70
71 return 0;
72
73out:
74 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
75 trace_amdgpu_runpm_reference_dumps(0, __func__);
76 return r;
77}
78
79/**
80 * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
81 *
82 * @dmabuf: DMA-buf where we remove the attachment from
83 * @attach: the attachment to remove
84 *
85 * Called when an attachment is removed from the DMA-buf.
86 */
87static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
88 struct dma_buf_attachment *attach)
89{
90 struct drm_gem_object *obj = dmabuf->priv;
91 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
92 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
93
94 pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
95 pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
96 trace_amdgpu_runpm_reference_dumps(0, __func__);
97}
98
99/**
100 * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
101 *
102 * @attach: attachment to pin down
103 *
104 * Pin the BO which is backing the DMA-buf so that it can't move any more.
105 */
106static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
107{
108 struct drm_gem_object *obj = attach->dmabuf->priv;
109 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
110
111 /* pin buffer into GTT */
112 return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
113}
114
115/**
116 * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
117 *
118 * @attach: attachment to unpin
119 *
120 * Unpin a previously pinned BO to make it movable again.
121 */
122static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
123{
124 struct drm_gem_object *obj = attach->dmabuf->priv;
125 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
126
127 amdgpu_bo_unpin(bo);
128}
129
130/**
131 * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
132 * @attach: DMA-buf attachment
133 * @dir: DMA direction
134 *
135 * Makes sure that the shared DMA buffer can be accessed by the target device.
136 * For now, simply pins it to the GTT domain, where it should be accessible by
137 * all DMA devices.
138 *
139 * Returns:
140 * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
141 * code.
142 */
143static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
144 enum dma_data_direction dir)
145{
146 struct dma_buf *dma_buf = attach->dmabuf;
147 struct drm_gem_object *obj = dma_buf->priv;
148 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
149 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
150 struct sg_table *sgt;
151 long r;
152
153 if (!bo->tbo.pin_count) {
154 /* move buffer into GTT or VRAM */
155 struct ttm_operation_ctx ctx = { false, false };
156 unsigned int domains = AMDGPU_GEM_DOMAIN_GTT;
157
158 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
159 attach->peer2peer) {
160 bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
161 domains |= AMDGPU_GEM_DOMAIN_VRAM;
162 }
163 amdgpu_bo_placement_from_domain(bo, domains);
164 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
165 if (r)
166 return ERR_PTR(r);
167
168 } else if (!(amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type) &
169 AMDGPU_GEM_DOMAIN_GTT)) {
170 return ERR_PTR(-EBUSY);
171 }
172
173 switch (bo->tbo.resource->mem_type) {
174 case TTM_PL_TT:
175 sgt = drm_prime_pages_to_sg(obj->dev,
176 bo->tbo.ttm->pages,
177 bo->tbo.ttm->num_pages);
178 if (IS_ERR(sgt))
179 return sgt;
180
181 if (dma_map_sgtable(attach->dev, sgt, dir,
182 DMA_ATTR_SKIP_CPU_SYNC))
183 goto error_free;
184 break;
185
186 case TTM_PL_VRAM:
187 r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
188 bo->tbo.base.size, attach->dev,
189 dir, &sgt);
190 if (r)
191 return ERR_PTR(r);
192 break;
193 default:
194 return ERR_PTR(-EINVAL);
195 }
196
197 return sgt;
198
199error_free:
200 sg_free_table(sgt);
201 kfree(sgt);
202 return ERR_PTR(-EBUSY);
203}
204
205/**
206 * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
207 * @attach: DMA-buf attachment
208 * @sgt: sg_table to unmap
209 * @dir: DMA direction
210 *
211 * This is called when a shared DMA buffer no longer needs to be accessible by
212 * another device. For now, simply unpins the buffer from GTT.
213 */
214static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
215 struct sg_table *sgt,
216 enum dma_data_direction dir)
217{
218 if (sgt->sgl->page_link) {
219 dma_unmap_sgtable(attach->dev, sgt, dir, 0);
220 sg_free_table(sgt);
221 kfree(sgt);
222 } else {
223 amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
224 }
225}
226
227/**
228 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
229 * @dma_buf: Shared DMA buffer
230 * @direction: Direction of DMA transfer
231 *
232 * This is called before CPU access to the shared DMA buffer's memory. If it's
233 * a read access, the buffer is moved to the GTT domain if possible, for optimal
234 * CPU read performance.
235 *
236 * Returns:
237 * 0 on success or a negative error code on failure.
238 */
239static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
240 enum dma_data_direction direction)
241{
242 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
243 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
244 struct ttm_operation_ctx ctx = { true, false };
245 u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
246 int ret;
247 bool reads = (direction == DMA_BIDIRECTIONAL ||
248 direction == DMA_FROM_DEVICE);
249
250 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
251 return 0;
252
253 /* move to gtt */
254 ret = amdgpu_bo_reserve(bo, false);
255 if (unlikely(ret != 0))
256 return ret;
257
258 if (!bo->tbo.pin_count &&
259 (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
260 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
261 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
262 }
263
264 amdgpu_bo_unreserve(bo);
265 return ret;
266}
267
268const struct dma_buf_ops amdgpu_dmabuf_ops = {
269 .attach = amdgpu_dma_buf_attach,
270 .detach = amdgpu_dma_buf_detach,
271 .pin = amdgpu_dma_buf_pin,
272 .unpin = amdgpu_dma_buf_unpin,
273 .map_dma_buf = amdgpu_dma_buf_map,
274 .unmap_dma_buf = amdgpu_dma_buf_unmap,
275 .release = drm_gem_dmabuf_release,
276 .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
277 .mmap = drm_gem_dmabuf_mmap,
278 .vmap = drm_gem_dmabuf_vmap,
279 .vunmap = drm_gem_dmabuf_vunmap,
280};
281
282/**
283 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
284 * @gobj: GEM BO
285 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
286 *
287 * The main work is done by the &drm_gem_prime_export helper.
288 *
289 * Returns:
290 * Shared DMA buffer representing the GEM BO from the given device.
291 */
292struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
293 int flags)
294{
295 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
296 struct dma_buf *buf;
297
298 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
299 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
300 return ERR_PTR(-EPERM);
301
302 buf = drm_gem_prime_export(gobj, flags);
303 if (!IS_ERR(buf))
304 buf->ops = &amdgpu_dmabuf_ops;
305
306 return buf;
307}
308
309/**
310 * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
311 *
312 * @dev: DRM device
313 * @dma_buf: DMA-buf
314 *
315 * Creates an empty SG BO for DMA-buf import.
316 *
317 * Returns:
318 * A new GEM BO of the given DRM device, representing the memory
319 * described by the given DMA-buf attachment and scatter/gather table.
320 */
321static struct drm_gem_object *
322amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
323{
324 struct dma_resv *resv = dma_buf->resv;
325 struct amdgpu_device *adev = drm_to_adev(dev);
326 struct drm_gem_object *gobj;
327 struct amdgpu_bo *bo;
328 uint64_t flags = 0;
329 int ret;
330
331 dma_resv_lock(resv, NULL);
332
333 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
334 struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
335
336 flags |= other->flags & (AMDGPU_GEM_CREATE_CPU_GTT_USWC |
337 AMDGPU_GEM_CREATE_COHERENT |
338 AMDGPU_GEM_CREATE_EXT_COHERENT |
339 AMDGPU_GEM_CREATE_UNCACHED);
340 }
341
342 ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
343 AMDGPU_GEM_DOMAIN_CPU, flags,
344 ttm_bo_type_sg, resv, &gobj, 0);
345 if (ret)
346 goto error;
347
348 bo = gem_to_amdgpu_bo(gobj);
349 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
350 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
351
352 dma_resv_unlock(resv);
353 return gobj;
354
355error:
356 dma_resv_unlock(resv);
357 return ERR_PTR(ret);
358}
359
360/**
361 * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
362 *
363 * @attach: the DMA-buf attachment
364 *
365 * Invalidate the DMA-buf attachment, making sure that the we re-create the
366 * mapping before the next use.
367 */
368static void
369amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
370{
371 struct drm_gem_object *obj = attach->importer_priv;
372 struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
373 struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
374 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
375 struct ttm_operation_ctx ctx = { false, false };
376 struct ttm_placement placement = {};
377 struct amdgpu_vm_bo_base *bo_base;
378 int r;
379
380 /* FIXME: This should be after the "if", but needs a fix to make sure
381 * DMABuf imports are initialized in the right VM list.
382 */
383 amdgpu_vm_bo_invalidate(adev, bo, false);
384 if (!bo->tbo.resource || bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
385 return;
386
387 r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
388 if (r) {
389 DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
390 return;
391 }
392
393 for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
394 struct amdgpu_vm *vm = bo_base->vm;
395 struct dma_resv *resv = vm->root.bo->tbo.base.resv;
396
397 if (ticket) {
398 /* When we get an error here it means that somebody
399 * else is holding the VM lock and updating page tables
400 * So we can just continue here.
401 */
402 r = dma_resv_lock(resv, ticket);
403 if (r)
404 continue;
405
406 } else {
407 /* TODO: This is more problematic and we actually need
408 * to allow page tables updates without holding the
409 * lock.
410 */
411 if (!dma_resv_trylock(resv))
412 continue;
413 }
414
415 /* Reserve fences for two SDMA page table updates */
416 r = dma_resv_reserve_fences(resv, 2);
417 if (!r)
418 r = amdgpu_vm_clear_freed(adev, vm, NULL);
419 if (!r)
420 r = amdgpu_vm_handle_moved(adev, vm, ticket);
421
422 if (r && r != -EBUSY)
423 DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
424 r);
425
426 dma_resv_unlock(resv);
427 }
428}
429
430static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
431 .allow_peer2peer = true,
432 .move_notify = amdgpu_dma_buf_move_notify
433};
434
435/**
436 * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
437 * @dev: DRM device
438 * @dma_buf: Shared DMA buffer
439 *
440 * Import a dma_buf into a the driver and potentially create a new GEM object.
441 *
442 * Returns:
443 * GEM BO representing the shared DMA buffer for the given device.
444 */
445struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
446 struct dma_buf *dma_buf)
447{
448 struct dma_buf_attachment *attach;
449 struct drm_gem_object *obj;
450
451 if (dma_buf->ops == &amdgpu_dmabuf_ops) {
452 obj = dma_buf->priv;
453 if (obj->dev == dev) {
454 /*
455 * Importing dmabuf exported from out own gem increases
456 * refcount on gem itself instead of f_count of dmabuf.
457 */
458 drm_gem_object_get(obj);
459 return obj;
460 }
461 }
462
463 obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
464 if (IS_ERR(obj))
465 return obj;
466
467 attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
468 &amdgpu_dma_buf_attach_ops, obj);
469 if (IS_ERR(attach)) {
470 drm_gem_object_put(obj);
471 return ERR_CAST(attach);
472 }
473
474 get_dma_buf(dma_buf);
475 obj->import_attach = attach;
476 return obj;
477}
478
479/**
480 * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
481 *
482 * @adev: amdgpu_device pointer of the importer
483 * @bo: amdgpu buffer object
484 *
485 * Returns:
486 * True if dmabuf accessible over xgmi, false otherwise.
487 */
488bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
489 struct amdgpu_bo *bo)
490{
491 struct drm_gem_object *obj = &bo->tbo.base;
492 struct drm_gem_object *gobj;
493
494 if (obj->import_attach) {
495 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
496
497 if (dma_buf->ops != &amdgpu_dmabuf_ops)
498 /* No XGMI with non AMD GPUs */
499 return false;
500
501 gobj = dma_buf->priv;
502 bo = gem_to_amdgpu_bo(gobj);
503 }
504
505 if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
506 (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))
507 return true;
508
509 return false;
510}
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 */
49struct 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 */
66void *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 */
86void 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 */
104int 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
139static int
140__dma_resv_make_exclusive(struct dma_resv *obj)
141{
142 struct dma_fence **fences;
143 unsigned int count;
144 int r;
145
146 if (!dma_resv_get_list(obj)) /* no shared fences to convert */
147 return 0;
148
149 r = dma_resv_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 dma_resv_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 dma_resv_add_excl_fence(obj, &array->base);
169 dma_fence_put(&array->base);
170 }
171
172 return 0;
173
174err_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 */
193static 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 = __dma_resv_make_exclusive(bo->tbo.base.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
232error_unreserve:
233 amdgpu_bo_unreserve(bo);
234
235error_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 */
249static 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
266error:
267 drm_gem_map_detach(dma_buf, attach);
268}
269
270/**
271 * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
272 * @dma_buf: Shared DMA buffer
273 * @direction: Direction of DMA transfer
274 *
275 * This is called before CPU access to the shared DMA buffer's memory. If it's
276 * a read access, the buffer is moved to the GTT domain if possible, for optimal
277 * CPU read performance.
278 *
279 * Returns:
280 * 0 on success or a negative error code on failure.
281 */
282static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
283 enum dma_data_direction direction)
284{
285 struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
286 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
287 struct ttm_operation_ctx ctx = { true, false };
288 u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
289 int ret;
290 bool reads = (direction == DMA_BIDIRECTIONAL ||
291 direction == DMA_FROM_DEVICE);
292
293 if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
294 return 0;
295
296 /* move to gtt */
297 ret = amdgpu_bo_reserve(bo, false);
298 if (unlikely(ret != 0))
299 return ret;
300
301 if (!bo->pin_count && (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
302 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
303 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
304 }
305
306 amdgpu_bo_unreserve(bo);
307 return ret;
308}
309
310const struct dma_buf_ops amdgpu_dmabuf_ops = {
311 .attach = amdgpu_dma_buf_map_attach,
312 .detach = amdgpu_dma_buf_map_detach,
313 .map_dma_buf = drm_gem_map_dma_buf,
314 .unmap_dma_buf = drm_gem_unmap_dma_buf,
315 .release = drm_gem_dmabuf_release,
316 .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
317 .mmap = drm_gem_dmabuf_mmap,
318 .vmap = drm_gem_dmabuf_vmap,
319 .vunmap = drm_gem_dmabuf_vunmap,
320};
321
322/**
323 * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
324 * @dev: DRM device
325 * @gobj: GEM BO
326 * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
327 *
328 * The main work is done by the &drm_gem_prime_export helper.
329 *
330 * Returns:
331 * Shared DMA buffer representing the GEM BO from the given device.
332 */
333struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
334 int flags)
335{
336 struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
337 struct dma_buf *buf;
338
339 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
340 bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
341 return ERR_PTR(-EPERM);
342
343 buf = drm_gem_prime_export(gobj, flags);
344 if (!IS_ERR(buf)) {
345 buf->file->f_mapping = gobj->dev->anon_inode->i_mapping;
346 buf->ops = &amdgpu_dmabuf_ops;
347 }
348
349 return buf;
350}
351
352/**
353 * amdgpu_gem_prime_import_sg_table - &drm_driver.gem_prime_import_sg_table
354 * implementation
355 * @dev: DRM device
356 * @attach: DMA-buf attachment
357 * @sg: Scatter/gather table
358 *
359 * Imports shared DMA buffer memory exported by another device.
360 *
361 * Returns:
362 * A new GEM BO of the given DRM device, representing the memory
363 * described by the given DMA-buf attachment and scatter/gather table.
364 */
365struct drm_gem_object *
366amdgpu_gem_prime_import_sg_table(struct drm_device *dev,
367 struct dma_buf_attachment *attach,
368 struct sg_table *sg)
369{
370 struct dma_resv *resv = attach->dmabuf->resv;
371 struct amdgpu_device *adev = dev->dev_private;
372 struct amdgpu_bo *bo;
373 struct amdgpu_bo_param bp;
374 int ret;
375
376 memset(&bp, 0, sizeof(bp));
377 bp.size = attach->dmabuf->size;
378 bp.byte_align = PAGE_SIZE;
379 bp.domain = AMDGPU_GEM_DOMAIN_CPU;
380 bp.flags = 0;
381 bp.type = ttm_bo_type_sg;
382 bp.resv = resv;
383 dma_resv_lock(resv, NULL);
384 ret = amdgpu_bo_create(adev, &bp, &bo);
385 if (ret)
386 goto error;
387
388 bo->tbo.sg = sg;
389 bo->tbo.ttm->sg = sg;
390 bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
391 bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
392 if (attach->dmabuf->ops != &amdgpu_dmabuf_ops)
393 bo->prime_shared_count = 1;
394
395 dma_resv_unlock(resv);
396 return &bo->tbo.base;
397
398error:
399 dma_resv_unlock(resv);
400 return ERR_PTR(ret);
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 * The main work is done by the &drm_gem_prime_import helper, which in turn
409 * uses &amdgpu_gem_prime_import_sg_table.
410 *
411 * Returns:
412 * GEM BO representing the shared DMA buffer for the given device.
413 */
414struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
415 struct dma_buf *dma_buf)
416{
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 return drm_gem_prime_import(dev, dma_buf);
432}