]> Git Repo - linux.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c
Merge tag 'drm-intel-next-2021-06-09' of git://anongit.freedesktop.org/drm/drm-intel...
[linux.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_dma_buf.c
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 <linux/dma-buf.h>
41 #include <linux/dma-fence-array.h>
42 #include <linux/pci-p2pdma.h>
43 #include <linux/pm_runtime.h>
44
45 static int
46 __dma_resv_make_exclusive(struct dma_resv *obj)
47 {
48         struct dma_fence **fences;
49         unsigned int count;
50         int r;
51
52         if (!dma_resv_shared_list(obj)) /* no shared fences to convert */
53                 return 0;
54
55         r = dma_resv_get_fences(obj, NULL, &count, &fences);
56         if (r)
57                 return r;
58
59         if (count == 0) {
60                 /* Now that was unexpected. */
61         } else if (count == 1) {
62                 dma_resv_add_excl_fence(obj, fences[0]);
63                 dma_fence_put(fences[0]);
64                 kfree(fences);
65         } else {
66                 struct dma_fence_array *array;
67
68                 array = dma_fence_array_create(count, fences,
69                                                dma_fence_context_alloc(1), 0,
70                                                false);
71                 if (!array)
72                         goto err_fences_put;
73
74                 dma_resv_add_excl_fence(obj, &array->base);
75                 dma_fence_put(&array->base);
76         }
77
78         return 0;
79
80 err_fences_put:
81         while (count--)
82                 dma_fence_put(fences[count]);
83         kfree(fences);
84         return -ENOMEM;
85 }
86
87 /**
88  * amdgpu_dma_buf_attach - &dma_buf_ops.attach implementation
89  *
90  * @dmabuf: DMA-buf where we attach to
91  * @attach: attachment to add
92  *
93  * Add the attachment as user to the exported DMA-buf.
94  */
95 static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf,
96                                  struct dma_buf_attachment *attach)
97 {
98         struct drm_gem_object *obj = dmabuf->priv;
99         struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
100         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
101         int r;
102
103         if (pci_p2pdma_distance_many(adev->pdev, &attach->dev, 1, true) < 0)
104                 attach->peer2peer = false;
105
106         if (attach->dev->driver == adev->dev->driver)
107                 return 0;
108
109         r = pm_runtime_get_sync(adev_to_drm(adev)->dev);
110         if (r < 0)
111                 goto out;
112
113         r = amdgpu_bo_reserve(bo, false);
114         if (unlikely(r != 0))
115                 goto out;
116
117         /*
118          * We only create shared fences for internal use, but importers
119          * of the dmabuf rely on exclusive fences for implicitly
120          * tracking write hazards. As any of the current fences may
121          * correspond to a write, we need to convert all existing
122          * fences on the reservation object into a single exclusive
123          * fence.
124          */
125         r = __dma_resv_make_exclusive(bo->tbo.base.resv);
126         if (r)
127                 goto out;
128
129         bo->prime_shared_count++;
130         amdgpu_bo_unreserve(bo);
131         return 0;
132
133 out:
134         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
135         return r;
136 }
137
138 /**
139  * amdgpu_dma_buf_detach - &dma_buf_ops.detach implementation
140  *
141  * @dmabuf: DMA-buf where we remove the attachment from
142  * @attach: the attachment to remove
143  *
144  * Called when an attachment is removed from the DMA-buf.
145  */
146 static void amdgpu_dma_buf_detach(struct dma_buf *dmabuf,
147                                   struct dma_buf_attachment *attach)
148 {
149         struct drm_gem_object *obj = dmabuf->priv;
150         struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
151         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
152
153         if (attach->dev->driver != adev->dev->driver && bo->prime_shared_count)
154                 bo->prime_shared_count--;
155
156         pm_runtime_mark_last_busy(adev_to_drm(adev)->dev);
157         pm_runtime_put_autosuspend(adev_to_drm(adev)->dev);
158 }
159
160 /**
161  * amdgpu_dma_buf_pin - &dma_buf_ops.pin implementation
162  *
163  * @attach: attachment to pin down
164  *
165  * Pin the BO which is backing the DMA-buf so that it can't move any more.
166  */
167 static int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)
168 {
169         struct drm_gem_object *obj = attach->dmabuf->priv;
170         struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
171
172         /* pin buffer into GTT */
173         return amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_GTT);
174 }
175
176 /**
177  * amdgpu_dma_buf_unpin - &dma_buf_ops.unpin implementation
178  *
179  * @attach: attachment to unpin
180  *
181  * Unpin a previously pinned BO to make it movable again.
182  */
183 static void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)
184 {
185         struct drm_gem_object *obj = attach->dmabuf->priv;
186         struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
187
188         amdgpu_bo_unpin(bo);
189 }
190
191 /**
192  * amdgpu_dma_buf_map - &dma_buf_ops.map_dma_buf implementation
193  * @attach: DMA-buf attachment
194  * @dir: DMA direction
195  *
196  * Makes sure that the shared DMA buffer can be accessed by the target device.
197  * For now, simply pins it to the GTT domain, where it should be accessible by
198  * all DMA devices.
199  *
200  * Returns:
201  * sg_table filled with the DMA addresses to use or ERR_PRT with negative error
202  * code.
203  */
204 static struct sg_table *amdgpu_dma_buf_map(struct dma_buf_attachment *attach,
205                                            enum dma_data_direction dir)
206 {
207         struct dma_buf *dma_buf = attach->dmabuf;
208         struct drm_gem_object *obj = dma_buf->priv;
209         struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
210         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
211         struct sg_table *sgt;
212         long r;
213
214         if (!bo->tbo.pin_count) {
215                 /* move buffer into GTT or VRAM */
216                 struct ttm_operation_ctx ctx = { false, false };
217                 unsigned domains = AMDGPU_GEM_DOMAIN_GTT;
218
219                 if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM &&
220                     attach->peer2peer) {
221                         bo->flags |= AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED;
222                         domains |= AMDGPU_GEM_DOMAIN_VRAM;
223                 }
224                 amdgpu_bo_placement_from_domain(bo, domains);
225                 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
226                 if (r)
227                         return ERR_PTR(r);
228
229         } else if (!(amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type) &
230                      AMDGPU_GEM_DOMAIN_GTT)) {
231                 return ERR_PTR(-EBUSY);
232         }
233
234         switch (bo->tbo.resource->mem_type) {
235         case TTM_PL_TT:
236                 sgt = drm_prime_pages_to_sg(obj->dev,
237                                             bo->tbo.ttm->pages,
238                                             bo->tbo.ttm->num_pages);
239                 if (IS_ERR(sgt))
240                         return sgt;
241
242                 if (dma_map_sgtable(attach->dev, sgt, dir,
243                                     DMA_ATTR_SKIP_CPU_SYNC))
244                         goto error_free;
245                 break;
246
247         case TTM_PL_VRAM:
248                 r = amdgpu_vram_mgr_alloc_sgt(adev, bo->tbo.resource, 0,
249                                               bo->tbo.base.size, attach->dev,
250                                               dir, &sgt);
251                 if (r)
252                         return ERR_PTR(r);
253                 break;
254         default:
255                 return ERR_PTR(-EINVAL);
256         }
257
258         return sgt;
259
260 error_free:
261         sg_free_table(sgt);
262         kfree(sgt);
263         return ERR_PTR(-EBUSY);
264 }
265
266 /**
267  * amdgpu_dma_buf_unmap - &dma_buf_ops.unmap_dma_buf implementation
268  * @attach: DMA-buf attachment
269  * @sgt: sg_table to unmap
270  * @dir: DMA direction
271  *
272  * This is called when a shared DMA buffer no longer needs to be accessible by
273  * another device. For now, simply unpins the buffer from GTT.
274  */
275 static void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach,
276                                  struct sg_table *sgt,
277                                  enum dma_data_direction dir)
278 {
279         if (sgt->sgl->page_link) {
280                 dma_unmap_sgtable(attach->dev, sgt, dir, 0);
281                 sg_free_table(sgt);
282                 kfree(sgt);
283         } else {
284                 amdgpu_vram_mgr_free_sgt(attach->dev, dir, sgt);
285         }
286 }
287
288 /**
289  * amdgpu_dma_buf_begin_cpu_access - &dma_buf_ops.begin_cpu_access implementation
290  * @dma_buf: Shared DMA buffer
291  * @direction: Direction of DMA transfer
292  *
293  * This is called before CPU access to the shared DMA buffer's memory. If it's
294  * a read access, the buffer is moved to the GTT domain if possible, for optimal
295  * CPU read performance.
296  *
297  * Returns:
298  * 0 on success or a negative error code on failure.
299  */
300 static int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf,
301                                            enum dma_data_direction direction)
302 {
303         struct amdgpu_bo *bo = gem_to_amdgpu_bo(dma_buf->priv);
304         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
305         struct ttm_operation_ctx ctx = { true, false };
306         u32 domain = amdgpu_display_supported_domains(adev, bo->flags);
307         int ret;
308         bool reads = (direction == DMA_BIDIRECTIONAL ||
309                       direction == DMA_FROM_DEVICE);
310
311         if (!reads || !(domain & AMDGPU_GEM_DOMAIN_GTT))
312                 return 0;
313
314         /* move to gtt */
315         ret = amdgpu_bo_reserve(bo, false);
316         if (unlikely(ret != 0))
317                 return ret;
318
319         if (!bo->tbo.pin_count &&
320             (bo->allowed_domains & AMDGPU_GEM_DOMAIN_GTT)) {
321                 amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_GTT);
322                 ret = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
323         }
324
325         amdgpu_bo_unreserve(bo);
326         return ret;
327 }
328
329 const struct dma_buf_ops amdgpu_dmabuf_ops = {
330         .attach = amdgpu_dma_buf_attach,
331         .detach = amdgpu_dma_buf_detach,
332         .pin = amdgpu_dma_buf_pin,
333         .unpin = amdgpu_dma_buf_unpin,
334         .map_dma_buf = amdgpu_dma_buf_map,
335         .unmap_dma_buf = amdgpu_dma_buf_unmap,
336         .release = drm_gem_dmabuf_release,
337         .begin_cpu_access = amdgpu_dma_buf_begin_cpu_access,
338         .mmap = drm_gem_dmabuf_mmap,
339         .vmap = drm_gem_dmabuf_vmap,
340         .vunmap = drm_gem_dmabuf_vunmap,
341 };
342
343 /**
344  * amdgpu_gem_prime_export - &drm_driver.gem_prime_export implementation
345  * @gobj: GEM BO
346  * @flags: Flags such as DRM_CLOEXEC and DRM_RDWR.
347  *
348  * The main work is done by the &drm_gem_prime_export helper.
349  *
350  * Returns:
351  * Shared DMA buffer representing the GEM BO from the given device.
352  */
353 struct dma_buf *amdgpu_gem_prime_export(struct drm_gem_object *gobj,
354                                         int flags)
355 {
356         struct amdgpu_bo *bo = gem_to_amdgpu_bo(gobj);
357         struct dma_buf *buf;
358
359         if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm) ||
360             bo->flags & AMDGPU_GEM_CREATE_VM_ALWAYS_VALID)
361                 return ERR_PTR(-EPERM);
362
363         buf = drm_gem_prime_export(gobj, flags);
364         if (!IS_ERR(buf))
365                 buf->ops = &amdgpu_dmabuf_ops;
366
367         return buf;
368 }
369
370 /**
371  * amdgpu_dma_buf_create_obj - create BO for DMA-buf import
372  *
373  * @dev: DRM device
374  * @dma_buf: DMA-buf
375  *
376  * Creates an empty SG BO for DMA-buf import.
377  *
378  * Returns:
379  * A new GEM BO of the given DRM device, representing the memory
380  * described by the given DMA-buf attachment and scatter/gather table.
381  */
382 static struct drm_gem_object *
383 amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)
384 {
385         struct dma_resv *resv = dma_buf->resv;
386         struct amdgpu_device *adev = drm_to_adev(dev);
387         struct drm_gem_object *gobj;
388         struct amdgpu_bo *bo;
389         uint64_t flags = 0;
390         int ret;
391
392         dma_resv_lock(resv, NULL);
393
394         if (dma_buf->ops == &amdgpu_dmabuf_ops) {
395                 struct amdgpu_bo *other = gem_to_amdgpu_bo(dma_buf->priv);
396
397                 flags |= other->flags & AMDGPU_GEM_CREATE_CPU_GTT_USWC;
398         }
399
400         ret = amdgpu_gem_object_create(adev, dma_buf->size, PAGE_SIZE,
401                                        AMDGPU_GEM_DOMAIN_CPU, flags,
402                                        ttm_bo_type_sg, resv, &gobj);
403         if (ret)
404                 goto error;
405
406         bo = gem_to_amdgpu_bo(gobj);
407         bo->allowed_domains = AMDGPU_GEM_DOMAIN_GTT;
408         bo->preferred_domains = AMDGPU_GEM_DOMAIN_GTT;
409         if (dma_buf->ops != &amdgpu_dmabuf_ops)
410                 bo->prime_shared_count = 1;
411
412         dma_resv_unlock(resv);
413         return gobj;
414
415 error:
416         dma_resv_unlock(resv);
417         return ERR_PTR(ret);
418 }
419
420 /**
421  * amdgpu_dma_buf_move_notify - &attach.move_notify implementation
422  *
423  * @attach: the DMA-buf attachment
424  *
425  * Invalidate the DMA-buf attachment, making sure that the we re-create the
426  * mapping before the next use.
427  */
428 static void
429 amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)
430 {
431         struct drm_gem_object *obj = attach->importer_priv;
432         struct ww_acquire_ctx *ticket = dma_resv_locking_ctx(obj->resv);
433         struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj);
434         struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
435         struct ttm_operation_ctx ctx = { false, false };
436         struct ttm_placement placement = {};
437         struct amdgpu_vm_bo_base *bo_base;
438         int r;
439
440         if (bo->tbo.resource->mem_type == TTM_PL_SYSTEM)
441                 return;
442
443         r = ttm_bo_validate(&bo->tbo, &placement, &ctx);
444         if (r) {
445                 DRM_ERROR("Failed to invalidate DMA-buf import (%d))\n", r);
446                 return;
447         }
448
449         for (bo_base = bo->vm_bo; bo_base; bo_base = bo_base->next) {
450                 struct amdgpu_vm *vm = bo_base->vm;
451                 struct dma_resv *resv = vm->root.base.bo->tbo.base.resv;
452
453                 if (ticket) {
454                         /* When we get an error here it means that somebody
455                          * else is holding the VM lock and updating page tables
456                          * So we can just continue here.
457                          */
458                         r = dma_resv_lock(resv, ticket);
459                         if (r)
460                                 continue;
461
462                 } else {
463                         /* TODO: This is more problematic and we actually need
464                          * to allow page tables updates without holding the
465                          * lock.
466                          */
467                         if (!dma_resv_trylock(resv))
468                                 continue;
469                 }
470
471                 r = amdgpu_vm_clear_freed(adev, vm, NULL);
472                 if (!r)
473                         r = amdgpu_vm_handle_moved(adev, vm);
474
475                 if (r && r != -EBUSY)
476                         DRM_ERROR("Failed to invalidate VM page tables (%d))\n",
477                                   r);
478
479                 dma_resv_unlock(resv);
480         }
481 }
482
483 static const struct dma_buf_attach_ops amdgpu_dma_buf_attach_ops = {
484         .allow_peer2peer = true,
485         .move_notify = amdgpu_dma_buf_move_notify
486 };
487
488 /**
489  * amdgpu_gem_prime_import - &drm_driver.gem_prime_import implementation
490  * @dev: DRM device
491  * @dma_buf: Shared DMA buffer
492  *
493  * Import a dma_buf into a the driver and potentially create a new GEM object.
494  *
495  * Returns:
496  * GEM BO representing the shared DMA buffer for the given device.
497  */
498 struct drm_gem_object *amdgpu_gem_prime_import(struct drm_device *dev,
499                                                struct dma_buf *dma_buf)
500 {
501         struct dma_buf_attachment *attach;
502         struct drm_gem_object *obj;
503
504         if (dma_buf->ops == &amdgpu_dmabuf_ops) {
505                 obj = dma_buf->priv;
506                 if (obj->dev == dev) {
507                         /*
508                          * Importing dmabuf exported from out own gem increases
509                          * refcount on gem itself instead of f_count of dmabuf.
510                          */
511                         drm_gem_object_get(obj);
512                         return obj;
513                 }
514         }
515
516         obj = amdgpu_dma_buf_create_obj(dev, dma_buf);
517         if (IS_ERR(obj))
518                 return obj;
519
520         attach = dma_buf_dynamic_attach(dma_buf, dev->dev,
521                                         &amdgpu_dma_buf_attach_ops, obj);
522         if (IS_ERR(attach)) {
523                 drm_gem_object_put(obj);
524                 return ERR_CAST(attach);
525         }
526
527         get_dma_buf(dma_buf);
528         obj->import_attach = attach;
529         return obj;
530 }
531
532 /**
533  * amdgpu_dmabuf_is_xgmi_accessible - Check if xgmi available for P2P transfer
534  *
535  * @adev: amdgpu_device pointer of the importer
536  * @bo: amdgpu buffer object
537  *
538  * Returns:
539  * True if dmabuf accessible over xgmi, false otherwise.
540  */
541 bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev,
542                                       struct amdgpu_bo *bo)
543 {
544         struct drm_gem_object *obj = &bo->tbo.base;
545         struct drm_gem_object *gobj;
546
547         if (obj->import_attach) {
548                 struct dma_buf *dma_buf = obj->import_attach->dmabuf;
549
550                 if (dma_buf->ops != &amdgpu_dmabuf_ops)
551                         /* No XGMI with non AMD GPUs */
552                         return false;
553
554                 gobj = dma_buf->priv;
555                 bo = gem_to_amdgpu_bo(gobj);
556         }
557
558         if (amdgpu_xgmi_same_hive(adev, amdgpu_ttm_adev(bo->tbo.bdev)) &&
559                         (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM))
560                 return true;
561
562         return false;
563 }
This page took 0.070936 seconds and 4 git commands to generate.