Loading...
1/*
2 * Copyright 2008 Jerome Glisse.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
23 *
24 * Authors:
25 * Jerome Glisse <glisse@freedesktop.org>
26 */
27
28#include <linux/file.h>
29#include <linux/pagemap.h>
30#include <linux/sync_file.h>
31#include <linux/dma-buf.h>
32
33#include <drm/amdgpu_drm.h>
34#include <drm/drm_syncobj.h>
35#include "amdgpu.h"
36#include "amdgpu_trace.h"
37#include "amdgpu_gmc.h"
38#include "amdgpu_gem.h"
39#include "amdgpu_ras.h"
40
41static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
42 struct drm_amdgpu_cs_chunk_fence *data,
43 uint32_t *offset)
44{
45 struct drm_gem_object *gobj;
46 struct amdgpu_bo *bo;
47 unsigned long size;
48 int r;
49
50 gobj = drm_gem_object_lookup(p->filp, data->handle);
51 if (gobj == NULL)
52 return -EINVAL;
53
54 bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
55 p->uf_entry.priority = 0;
56 p->uf_entry.tv.bo = &bo->tbo;
57 /* One for TTM and one for the CS job */
58 p->uf_entry.tv.num_shared = 2;
59
60 drm_gem_object_put(gobj);
61
62 size = amdgpu_bo_size(bo);
63 if (size != PAGE_SIZE || (data->offset + 8) > size) {
64 r = -EINVAL;
65 goto error_unref;
66 }
67
68 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
69 r = -EINVAL;
70 goto error_unref;
71 }
72
73 *offset = data->offset;
74
75 return 0;
76
77error_unref:
78 amdgpu_bo_unref(&bo);
79 return r;
80}
81
82static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
83 struct drm_amdgpu_bo_list_in *data)
84{
85 int r;
86 struct drm_amdgpu_bo_list_entry *info = NULL;
87
88 r = amdgpu_bo_create_list_entry_array(data, &info);
89 if (r)
90 return r;
91
92 r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
93 &p->bo_list);
94 if (r)
95 goto error_free;
96
97 kvfree(info);
98 return 0;
99
100error_free:
101 if (info)
102 kvfree(info);
103
104 return r;
105}
106
107static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
108{
109 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
110 struct amdgpu_vm *vm = &fpriv->vm;
111 uint64_t *chunk_array_user;
112 uint64_t *chunk_array;
113 unsigned size, num_ibs = 0;
114 uint32_t uf_offset = 0;
115 int i;
116 int ret;
117
118 if (cs->in.num_chunks == 0)
119 return 0;
120
121 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
122 if (!chunk_array)
123 return -ENOMEM;
124
125 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
126 if (!p->ctx) {
127 ret = -EINVAL;
128 goto free_chunk;
129 }
130
131 mutex_lock(&p->ctx->lock);
132
133 /* skip guilty context job */
134 if (atomic_read(&p->ctx->guilty) == 1) {
135 ret = -ECANCELED;
136 goto free_chunk;
137 }
138
139 /* get chunks */
140 chunk_array_user = u64_to_user_ptr(cs->in.chunks);
141 if (copy_from_user(chunk_array, chunk_array_user,
142 sizeof(uint64_t)*cs->in.num_chunks)) {
143 ret = -EFAULT;
144 goto free_chunk;
145 }
146
147 p->nchunks = cs->in.num_chunks;
148 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
149 GFP_KERNEL);
150 if (!p->chunks) {
151 ret = -ENOMEM;
152 goto free_chunk;
153 }
154
155 for (i = 0; i < p->nchunks; i++) {
156 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
157 struct drm_amdgpu_cs_chunk user_chunk;
158 uint32_t __user *cdata;
159
160 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
161 if (copy_from_user(&user_chunk, chunk_ptr,
162 sizeof(struct drm_amdgpu_cs_chunk))) {
163 ret = -EFAULT;
164 i--;
165 goto free_partial_kdata;
166 }
167 p->chunks[i].chunk_id = user_chunk.chunk_id;
168 p->chunks[i].length_dw = user_chunk.length_dw;
169
170 size = p->chunks[i].length_dw;
171 cdata = u64_to_user_ptr(user_chunk.chunk_data);
172
173 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
174 if (p->chunks[i].kdata == NULL) {
175 ret = -ENOMEM;
176 i--;
177 goto free_partial_kdata;
178 }
179 size *= sizeof(uint32_t);
180 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
181 ret = -EFAULT;
182 goto free_partial_kdata;
183 }
184
185 switch (p->chunks[i].chunk_id) {
186 case AMDGPU_CHUNK_ID_IB:
187 ++num_ibs;
188 break;
189
190 case AMDGPU_CHUNK_ID_FENCE:
191 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
192 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
193 ret = -EINVAL;
194 goto free_partial_kdata;
195 }
196
197 ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
198 &uf_offset);
199 if (ret)
200 goto free_partial_kdata;
201
202 break;
203
204 case AMDGPU_CHUNK_ID_BO_HANDLES:
205 size = sizeof(struct drm_amdgpu_bo_list_in);
206 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
207 ret = -EINVAL;
208 goto free_partial_kdata;
209 }
210
211 ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
212 if (ret)
213 goto free_partial_kdata;
214
215 break;
216
217 case AMDGPU_CHUNK_ID_DEPENDENCIES:
218 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
219 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
220 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
221 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
222 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
223 break;
224
225 default:
226 ret = -EINVAL;
227 goto free_partial_kdata;
228 }
229 }
230
231 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
232 if (ret)
233 goto free_all_kdata;
234
235 if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
236 ret = -ECANCELED;
237 goto free_all_kdata;
238 }
239
240 if (p->uf_entry.tv.bo)
241 p->job->uf_addr = uf_offset;
242 kfree(chunk_array);
243
244 /* Use this opportunity to fill in task info for the vm */
245 amdgpu_vm_set_task_info(vm);
246
247 return 0;
248
249free_all_kdata:
250 i = p->nchunks - 1;
251free_partial_kdata:
252 for (; i >= 0; i--)
253 kvfree(p->chunks[i].kdata);
254 kfree(p->chunks);
255 p->chunks = NULL;
256 p->nchunks = 0;
257free_chunk:
258 kfree(chunk_array);
259
260 return ret;
261}
262
263/* Convert microseconds to bytes. */
264static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
265{
266 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
267 return 0;
268
269 /* Since accum_us is incremented by a million per second, just
270 * multiply it by the number of MB/s to get the number of bytes.
271 */
272 return us << adev->mm_stats.log2_max_MBps;
273}
274
275static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
276{
277 if (!adev->mm_stats.log2_max_MBps)
278 return 0;
279
280 return bytes >> adev->mm_stats.log2_max_MBps;
281}
282
283/* Returns how many bytes TTM can move right now. If no bytes can be moved,
284 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
285 * which means it can go over the threshold once. If that happens, the driver
286 * will be in debt and no other buffer migrations can be done until that debt
287 * is repaid.
288 *
289 * This approach allows moving a buffer of any size (it's important to allow
290 * that).
291 *
292 * The currency is simply time in microseconds and it increases as the clock
293 * ticks. The accumulated microseconds (us) are converted to bytes and
294 * returned.
295 */
296static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
297 u64 *max_bytes,
298 u64 *max_vis_bytes)
299{
300 s64 time_us, increment_us;
301 u64 free_vram, total_vram, used_vram;
302
303 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
304 * throttling.
305 *
306 * It means that in order to get full max MBps, at least 5 IBs per
307 * second must be submitted and not more than 200ms apart from each
308 * other.
309 */
310 const s64 us_upper_bound = 200000;
311
312 if (!adev->mm_stats.log2_max_MBps) {
313 *max_bytes = 0;
314 *max_vis_bytes = 0;
315 return;
316 }
317
318 total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
319 used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
320 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
321
322 spin_lock(&adev->mm_stats.lock);
323
324 /* Increase the amount of accumulated us. */
325 time_us = ktime_to_us(ktime_get());
326 increment_us = time_us - adev->mm_stats.last_update_us;
327 adev->mm_stats.last_update_us = time_us;
328 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
329 us_upper_bound);
330
331 /* This prevents the short period of low performance when the VRAM
332 * usage is low and the driver is in debt or doesn't have enough
333 * accumulated us to fill VRAM quickly.
334 *
335 * The situation can occur in these cases:
336 * - a lot of VRAM is freed by userspace
337 * - the presence of a big buffer causes a lot of evictions
338 * (solution: split buffers into smaller ones)
339 *
340 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
341 * accum_us to a positive number.
342 */
343 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
344 s64 min_us;
345
346 /* Be more aggresive on dGPUs. Try to fill a portion of free
347 * VRAM now.
348 */
349 if (!(adev->flags & AMD_IS_APU))
350 min_us = bytes_to_us(adev, free_vram / 4);
351 else
352 min_us = 0; /* Reset accum_us on APUs. */
353
354 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
355 }
356
357 /* This is set to 0 if the driver is in debt to disallow (optional)
358 * buffer moves.
359 */
360 *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
361
362 /* Do the same for visible VRAM if half of it is free */
363 if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
364 u64 total_vis_vram = adev->gmc.visible_vram_size;
365 u64 used_vis_vram =
366 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
367
368 if (used_vis_vram < total_vis_vram) {
369 u64 free_vis_vram = total_vis_vram - used_vis_vram;
370 adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
371 increment_us, us_upper_bound);
372
373 if (free_vis_vram >= total_vis_vram / 2)
374 adev->mm_stats.accum_us_vis =
375 max(bytes_to_us(adev, free_vis_vram / 2),
376 adev->mm_stats.accum_us_vis);
377 }
378
379 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
380 } else {
381 *max_vis_bytes = 0;
382 }
383
384 spin_unlock(&adev->mm_stats.lock);
385}
386
387/* Report how many bytes have really been moved for the last command
388 * submission. This can result in a debt that can stop buffer migrations
389 * temporarily.
390 */
391void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
392 u64 num_vis_bytes)
393{
394 spin_lock(&adev->mm_stats.lock);
395 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
396 adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
397 spin_unlock(&adev->mm_stats.lock);
398}
399
400static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
401 struct amdgpu_bo *bo)
402{
403 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
404 struct ttm_operation_ctx ctx = {
405 .interruptible = true,
406 .no_wait_gpu = false,
407 .resv = bo->tbo.base.resv,
408 .flags = 0
409 };
410 uint32_t domain;
411 int r;
412
413 if (bo->pin_count)
414 return 0;
415
416 /* Don't move this buffer if we have depleted our allowance
417 * to move it. Don't move anything if the threshold is zero.
418 */
419 if (p->bytes_moved < p->bytes_moved_threshold &&
420 (!bo->tbo.base.dma_buf ||
421 list_empty(&bo->tbo.base.dma_buf->attachments))) {
422 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
423 (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
424 /* And don't move a CPU_ACCESS_REQUIRED BO to limited
425 * visible VRAM if we've depleted our allowance to do
426 * that.
427 */
428 if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
429 domain = bo->preferred_domains;
430 else
431 domain = bo->allowed_domains;
432 } else {
433 domain = bo->preferred_domains;
434 }
435 } else {
436 domain = bo->allowed_domains;
437 }
438
439retry:
440 amdgpu_bo_placement_from_domain(bo, domain);
441 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
442
443 p->bytes_moved += ctx.bytes_moved;
444 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
445 amdgpu_bo_in_cpu_visible_vram(bo))
446 p->bytes_moved_vis += ctx.bytes_moved;
447
448 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
449 domain = bo->allowed_domains;
450 goto retry;
451 }
452
453 return r;
454}
455
456static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
457{
458 struct amdgpu_cs_parser *p = param;
459 int r;
460
461 r = amdgpu_cs_bo_validate(p, bo);
462 if (r)
463 return r;
464
465 if (bo->shadow)
466 r = amdgpu_cs_bo_validate(p, bo->shadow);
467
468 return r;
469}
470
471static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
472 struct list_head *validated)
473{
474 struct ttm_operation_ctx ctx = { true, false };
475 struct amdgpu_bo_list_entry *lobj;
476 int r;
477
478 list_for_each_entry(lobj, validated, tv.head) {
479 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
480 struct mm_struct *usermm;
481
482 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
483 if (usermm && usermm != current->mm)
484 return -EPERM;
485
486 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
487 lobj->user_invalidated && lobj->user_pages) {
488 amdgpu_bo_placement_from_domain(bo,
489 AMDGPU_GEM_DOMAIN_CPU);
490 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
491 if (r)
492 return r;
493
494 amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
495 lobj->user_pages);
496 }
497
498 r = amdgpu_cs_validate(p, bo);
499 if (r)
500 return r;
501
502 kvfree(lobj->user_pages);
503 lobj->user_pages = NULL;
504 }
505 return 0;
506}
507
508static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
509 union drm_amdgpu_cs *cs)
510{
511 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
512 struct amdgpu_vm *vm = &fpriv->vm;
513 struct amdgpu_bo_list_entry *e;
514 struct list_head duplicates;
515 struct amdgpu_bo *gds;
516 struct amdgpu_bo *gws;
517 struct amdgpu_bo *oa;
518 int r;
519
520 INIT_LIST_HEAD(&p->validated);
521
522 /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
523 if (cs->in.bo_list_handle) {
524 if (p->bo_list)
525 return -EINVAL;
526
527 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
528 &p->bo_list);
529 if (r)
530 return r;
531 } else if (!p->bo_list) {
532 /* Create a empty bo_list when no handle is provided */
533 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
534 &p->bo_list);
535 if (r)
536 return r;
537 }
538
539 /* One for TTM and one for the CS job */
540 amdgpu_bo_list_for_each_entry(e, p->bo_list)
541 e->tv.num_shared = 2;
542
543 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
544
545 INIT_LIST_HEAD(&duplicates);
546 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
547
548 if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
549 list_add(&p->uf_entry.tv.head, &p->validated);
550
551 /* Get userptr backing pages. If pages are updated after registered
552 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
553 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
554 */
555 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
556 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
557 bool userpage_invalidated = false;
558 int i;
559
560 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
561 sizeof(struct page *),
562 GFP_KERNEL | __GFP_ZERO);
563 if (!e->user_pages) {
564 DRM_ERROR("calloc failure\n");
565 return -ENOMEM;
566 }
567
568 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
569 if (r) {
570 kvfree(e->user_pages);
571 e->user_pages = NULL;
572 return r;
573 }
574
575 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
576 if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
577 userpage_invalidated = true;
578 break;
579 }
580 }
581 e->user_invalidated = userpage_invalidated;
582 }
583
584 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
585 &duplicates);
586 if (unlikely(r != 0)) {
587 if (r != -ERESTARTSYS)
588 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
589 goto out;
590 }
591
592 amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
593 &p->bytes_moved_vis_threshold);
594 p->bytes_moved = 0;
595 p->bytes_moved_vis = 0;
596
597 r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
598 amdgpu_cs_validate, p);
599 if (r) {
600 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
601 goto error_validate;
602 }
603
604 r = amdgpu_cs_list_validate(p, &duplicates);
605 if (r)
606 goto error_validate;
607
608 r = amdgpu_cs_list_validate(p, &p->validated);
609 if (r)
610 goto error_validate;
611
612 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
613 p->bytes_moved_vis);
614
615 gds = p->bo_list->gds_obj;
616 gws = p->bo_list->gws_obj;
617 oa = p->bo_list->oa_obj;
618
619 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
620 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
621
622 /* Make sure we use the exclusive slot for shared BOs */
623 if (bo->prime_shared_count)
624 e->tv.num_shared = 0;
625 e->bo_va = amdgpu_vm_bo_find(vm, bo);
626 }
627
628 if (gds) {
629 p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
630 p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
631 }
632 if (gws) {
633 p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
634 p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
635 }
636 if (oa) {
637 p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
638 p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
639 }
640
641 if (!r && p->uf_entry.tv.bo) {
642 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
643
644 r = amdgpu_ttm_alloc_gart(&uf->tbo);
645 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
646 }
647
648error_validate:
649 if (r)
650 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
651out:
652 return r;
653}
654
655static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
656{
657 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
658 struct amdgpu_bo_list_entry *e;
659 int r;
660
661 list_for_each_entry(e, &p->validated, tv.head) {
662 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
663 struct dma_resv *resv = bo->tbo.base.resv;
664 enum amdgpu_sync_mode sync_mode;
665
666 sync_mode = amdgpu_bo_explicit_sync(bo) ?
667 AMDGPU_SYNC_EXPLICIT : AMDGPU_SYNC_NE_OWNER;
668 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, sync_mode,
669 &fpriv->vm);
670 if (r)
671 return r;
672 }
673 return 0;
674}
675
676/**
677 * cs_parser_fini() - clean parser states
678 * @parser: parser structure holding parsing context.
679 * @error: error number
680 *
681 * If error is set than unvalidate buffer, otherwise just free memory
682 * used by parsing context.
683 **/
684static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
685 bool backoff)
686{
687 unsigned i;
688
689 if (error && backoff)
690 ttm_eu_backoff_reservation(&parser->ticket,
691 &parser->validated);
692
693 for (i = 0; i < parser->num_post_deps; i++) {
694 drm_syncobj_put(parser->post_deps[i].syncobj);
695 kfree(parser->post_deps[i].chain);
696 }
697 kfree(parser->post_deps);
698
699 dma_fence_put(parser->fence);
700
701 if (parser->ctx) {
702 mutex_unlock(&parser->ctx->lock);
703 amdgpu_ctx_put(parser->ctx);
704 }
705 if (parser->bo_list)
706 amdgpu_bo_list_put(parser->bo_list);
707
708 for (i = 0; i < parser->nchunks; i++)
709 kvfree(parser->chunks[i].kdata);
710 kfree(parser->chunks);
711 if (parser->job)
712 amdgpu_job_free(parser->job);
713 if (parser->uf_entry.tv.bo) {
714 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
715
716 amdgpu_bo_unref(&uf);
717 }
718}
719
720static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
721{
722 struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
723 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
724 struct amdgpu_device *adev = p->adev;
725 struct amdgpu_vm *vm = &fpriv->vm;
726 struct amdgpu_bo_list_entry *e;
727 struct amdgpu_bo_va *bo_va;
728 struct amdgpu_bo *bo;
729 int r;
730
731 /* Only for UVD/VCE VM emulation */
732 if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
733 unsigned i, j;
734
735 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
736 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
737 struct amdgpu_bo_va_mapping *m;
738 struct amdgpu_bo *aobj = NULL;
739 struct amdgpu_cs_chunk *chunk;
740 uint64_t offset, va_start;
741 struct amdgpu_ib *ib;
742 uint8_t *kptr;
743
744 chunk = &p->chunks[i];
745 ib = &p->job->ibs[j];
746 chunk_ib = chunk->kdata;
747
748 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
749 continue;
750
751 va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
752 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
753 if (r) {
754 DRM_ERROR("IB va_start is invalid\n");
755 return r;
756 }
757
758 if ((va_start + chunk_ib->ib_bytes) >
759 (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
760 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
761 return -EINVAL;
762 }
763
764 /* the IB should be reserved at this point */
765 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
766 if (r) {
767 return r;
768 }
769
770 offset = m->start * AMDGPU_GPU_PAGE_SIZE;
771 kptr += va_start - offset;
772
773 if (ring->funcs->parse_cs) {
774 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
775 amdgpu_bo_kunmap(aobj);
776
777 r = amdgpu_ring_parse_cs(ring, p, j);
778 if (r)
779 return r;
780 } else {
781 ib->ptr = (uint32_t *)kptr;
782 r = amdgpu_ring_patch_cs_in_place(ring, p, j);
783 amdgpu_bo_kunmap(aobj);
784 if (r)
785 return r;
786 }
787
788 j++;
789 }
790 }
791
792 if (!p->job->vm)
793 return amdgpu_cs_sync_rings(p);
794
795
796 r = amdgpu_vm_clear_freed(adev, vm, NULL);
797 if (r)
798 return r;
799
800 r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
801 if (r)
802 return r;
803
804 r = amdgpu_sync_vm_fence(&p->job->sync, fpriv->prt_va->last_pt_update);
805 if (r)
806 return r;
807
808 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
809 bo_va = fpriv->csa_va;
810 BUG_ON(!bo_va);
811 r = amdgpu_vm_bo_update(adev, bo_va, false);
812 if (r)
813 return r;
814
815 r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
816 if (r)
817 return r;
818 }
819
820 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
821 /* ignore duplicates */
822 bo = ttm_to_amdgpu_bo(e->tv.bo);
823 if (!bo)
824 continue;
825
826 bo_va = e->bo_va;
827 if (bo_va == NULL)
828 continue;
829
830 r = amdgpu_vm_bo_update(adev, bo_va, false);
831 if (r)
832 return r;
833
834 r = amdgpu_sync_vm_fence(&p->job->sync, bo_va->last_pt_update);
835 if (r)
836 return r;
837 }
838
839 r = amdgpu_vm_handle_moved(adev, vm);
840 if (r)
841 return r;
842
843 r = amdgpu_vm_update_pdes(adev, vm, false);
844 if (r)
845 return r;
846
847 r = amdgpu_sync_vm_fence(&p->job->sync, vm->last_update);
848 if (r)
849 return r;
850
851 p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.base.bo);
852
853 if (amdgpu_vm_debug) {
854 /* Invalidate all BOs to test for userspace bugs */
855 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
856 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
857
858 /* ignore duplicates */
859 if (!bo)
860 continue;
861
862 amdgpu_vm_bo_invalidate(adev, bo, false);
863 }
864 }
865
866 return amdgpu_cs_sync_rings(p);
867}
868
869static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
870 struct amdgpu_cs_parser *parser)
871{
872 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
873 struct amdgpu_vm *vm = &fpriv->vm;
874 int r, ce_preempt = 0, de_preempt = 0;
875 struct amdgpu_ring *ring;
876 int i, j;
877
878 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
879 struct amdgpu_cs_chunk *chunk;
880 struct amdgpu_ib *ib;
881 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
882 struct drm_sched_entity *entity;
883
884 chunk = &parser->chunks[i];
885 ib = &parser->job->ibs[j];
886 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
887
888 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
889 continue;
890
891 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
892 (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
893 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
894 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
895 ce_preempt++;
896 else
897 de_preempt++;
898 }
899
900 /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
901 if (ce_preempt > 1 || de_preempt > 1)
902 return -EINVAL;
903 }
904
905 r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
906 chunk_ib->ip_instance, chunk_ib->ring,
907 &entity);
908 if (r)
909 return r;
910
911 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
912 parser->job->preamble_status |=
913 AMDGPU_PREAMBLE_IB_PRESENT;
914
915 if (parser->entity && parser->entity != entity)
916 return -EINVAL;
917
918 /* Return if there is no run queue associated with this entity.
919 * Possibly because of disabled HW IP*/
920 if (entity->rq == NULL)
921 return -EINVAL;
922
923 parser->entity = entity;
924
925 ring = to_amdgpu_ring(entity->rq->sched);
926 r = amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
927 chunk_ib->ib_bytes : 0,
928 AMDGPU_IB_POOL_DELAYED, ib);
929 if (r) {
930 DRM_ERROR("Failed to get ib !\n");
931 return r;
932 }
933
934 ib->gpu_addr = chunk_ib->va_start;
935 ib->length_dw = chunk_ib->ib_bytes / 4;
936 ib->flags = chunk_ib->flags;
937
938 j++;
939 }
940
941 /* MM engine doesn't support user fences */
942 ring = to_amdgpu_ring(parser->entity->rq->sched);
943 if (parser->job->uf_addr && ring->funcs->no_user_fence)
944 return -EINVAL;
945
946 return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->entity);
947}
948
949static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
950 struct amdgpu_cs_chunk *chunk)
951{
952 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
953 unsigned num_deps;
954 int i, r;
955 struct drm_amdgpu_cs_chunk_dep *deps;
956
957 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
958 num_deps = chunk->length_dw * 4 /
959 sizeof(struct drm_amdgpu_cs_chunk_dep);
960
961 for (i = 0; i < num_deps; ++i) {
962 struct amdgpu_ctx *ctx;
963 struct drm_sched_entity *entity;
964 struct dma_fence *fence;
965
966 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
967 if (ctx == NULL)
968 return -EINVAL;
969
970 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
971 deps[i].ip_instance,
972 deps[i].ring, &entity);
973 if (r) {
974 amdgpu_ctx_put(ctx);
975 return r;
976 }
977
978 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
979 amdgpu_ctx_put(ctx);
980
981 if (IS_ERR(fence))
982 return PTR_ERR(fence);
983 else if (!fence)
984 continue;
985
986 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
987 struct drm_sched_fence *s_fence;
988 struct dma_fence *old = fence;
989
990 s_fence = to_drm_sched_fence(fence);
991 fence = dma_fence_get(&s_fence->scheduled);
992 dma_fence_put(old);
993 }
994
995 r = amdgpu_sync_fence(&p->job->sync, fence);
996 dma_fence_put(fence);
997 if (r)
998 return r;
999 }
1000 return 0;
1001}
1002
1003static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1004 uint32_t handle, u64 point,
1005 u64 flags)
1006{
1007 struct dma_fence *fence;
1008 int r;
1009
1010 r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1011 if (r) {
1012 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1013 handle, point, r);
1014 return r;
1015 }
1016
1017 r = amdgpu_sync_fence(&p->job->sync, fence);
1018 dma_fence_put(fence);
1019
1020 return r;
1021}
1022
1023static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1024 struct amdgpu_cs_chunk *chunk)
1025{
1026 struct drm_amdgpu_cs_chunk_sem *deps;
1027 unsigned num_deps;
1028 int i, r;
1029
1030 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1031 num_deps = chunk->length_dw * 4 /
1032 sizeof(struct drm_amdgpu_cs_chunk_sem);
1033 for (i = 0; i < num_deps; ++i) {
1034 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1035 0, 0);
1036 if (r)
1037 return r;
1038 }
1039
1040 return 0;
1041}
1042
1043
1044static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1045 struct amdgpu_cs_chunk *chunk)
1046{
1047 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1048 unsigned num_deps;
1049 int i, r;
1050
1051 syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1052 num_deps = chunk->length_dw * 4 /
1053 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1054 for (i = 0; i < num_deps; ++i) {
1055 r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1056 syncobj_deps[i].handle,
1057 syncobj_deps[i].point,
1058 syncobj_deps[i].flags);
1059 if (r)
1060 return r;
1061 }
1062
1063 return 0;
1064}
1065
1066static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1067 struct amdgpu_cs_chunk *chunk)
1068{
1069 struct drm_amdgpu_cs_chunk_sem *deps;
1070 unsigned num_deps;
1071 int i;
1072
1073 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1074 num_deps = chunk->length_dw * 4 /
1075 sizeof(struct drm_amdgpu_cs_chunk_sem);
1076
1077 if (p->post_deps)
1078 return -EINVAL;
1079
1080 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1081 GFP_KERNEL);
1082 p->num_post_deps = 0;
1083
1084 if (!p->post_deps)
1085 return -ENOMEM;
1086
1087
1088 for (i = 0; i < num_deps; ++i) {
1089 p->post_deps[i].syncobj =
1090 drm_syncobj_find(p->filp, deps[i].handle);
1091 if (!p->post_deps[i].syncobj)
1092 return -EINVAL;
1093 p->post_deps[i].chain = NULL;
1094 p->post_deps[i].point = 0;
1095 p->num_post_deps++;
1096 }
1097
1098 return 0;
1099}
1100
1101
1102static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1103 struct amdgpu_cs_chunk *chunk)
1104{
1105 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1106 unsigned num_deps;
1107 int i;
1108
1109 syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1110 num_deps = chunk->length_dw * 4 /
1111 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1112
1113 if (p->post_deps)
1114 return -EINVAL;
1115
1116 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1117 GFP_KERNEL);
1118 p->num_post_deps = 0;
1119
1120 if (!p->post_deps)
1121 return -ENOMEM;
1122
1123 for (i = 0; i < num_deps; ++i) {
1124 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1125
1126 dep->chain = NULL;
1127 if (syncobj_deps[i].point) {
1128 dep->chain = kmalloc(sizeof(*dep->chain), GFP_KERNEL);
1129 if (!dep->chain)
1130 return -ENOMEM;
1131 }
1132
1133 dep->syncobj = drm_syncobj_find(p->filp,
1134 syncobj_deps[i].handle);
1135 if (!dep->syncobj) {
1136 kfree(dep->chain);
1137 return -EINVAL;
1138 }
1139 dep->point = syncobj_deps[i].point;
1140 p->num_post_deps++;
1141 }
1142
1143 return 0;
1144}
1145
1146static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1147 struct amdgpu_cs_parser *p)
1148{
1149 int i, r;
1150
1151 for (i = 0; i < p->nchunks; ++i) {
1152 struct amdgpu_cs_chunk *chunk;
1153
1154 chunk = &p->chunks[i];
1155
1156 switch (chunk->chunk_id) {
1157 case AMDGPU_CHUNK_ID_DEPENDENCIES:
1158 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1159 r = amdgpu_cs_process_fence_dep(p, chunk);
1160 if (r)
1161 return r;
1162 break;
1163 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1164 r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1165 if (r)
1166 return r;
1167 break;
1168 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1169 r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1170 if (r)
1171 return r;
1172 break;
1173 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1174 r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1175 if (r)
1176 return r;
1177 break;
1178 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1179 r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1180 if (r)
1181 return r;
1182 break;
1183 }
1184 }
1185
1186 return 0;
1187}
1188
1189static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1190{
1191 int i;
1192
1193 for (i = 0; i < p->num_post_deps; ++i) {
1194 if (p->post_deps[i].chain && p->post_deps[i].point) {
1195 drm_syncobj_add_point(p->post_deps[i].syncobj,
1196 p->post_deps[i].chain,
1197 p->fence, p->post_deps[i].point);
1198 p->post_deps[i].chain = NULL;
1199 } else {
1200 drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1201 p->fence);
1202 }
1203 }
1204}
1205
1206static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1207 union drm_amdgpu_cs *cs)
1208{
1209 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1210 struct drm_sched_entity *entity = p->entity;
1211 struct amdgpu_bo_list_entry *e;
1212 struct amdgpu_job *job;
1213 uint64_t seq;
1214 int r;
1215
1216 job = p->job;
1217 p->job = NULL;
1218
1219 r = drm_sched_job_init(&job->base, entity, &fpriv->vm);
1220 if (r)
1221 goto error_unlock;
1222
1223 /* No memory allocation is allowed while holding the notifier lock.
1224 * The lock is held until amdgpu_cs_submit is finished and fence is
1225 * added to BOs.
1226 */
1227 mutex_lock(&p->adev->notifier_lock);
1228
1229 /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1230 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1231 */
1232 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1233 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1234
1235 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1236 }
1237 if (r) {
1238 r = -EAGAIN;
1239 goto error_abort;
1240 }
1241
1242 p->fence = dma_fence_get(&job->base.s_fence->finished);
1243
1244 amdgpu_ctx_add_fence(p->ctx, entity, p->fence, &seq);
1245 amdgpu_cs_post_dependencies(p);
1246
1247 if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1248 !p->ctx->preamble_presented) {
1249 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1250 p->ctx->preamble_presented = true;
1251 }
1252
1253 cs->out.handle = seq;
1254 job->uf_sequence = seq;
1255
1256 amdgpu_job_free_resources(job);
1257
1258 trace_amdgpu_cs_ioctl(job);
1259 amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1260 drm_sched_entity_push_job(&job->base, entity);
1261
1262 amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1263
1264 ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1265 mutex_unlock(&p->adev->notifier_lock);
1266
1267 return 0;
1268
1269error_abort:
1270 drm_sched_job_cleanup(&job->base);
1271 mutex_unlock(&p->adev->notifier_lock);
1272
1273error_unlock:
1274 amdgpu_job_free(job);
1275 return r;
1276}
1277
1278int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1279{
1280 struct amdgpu_device *adev = dev->dev_private;
1281 union drm_amdgpu_cs *cs = data;
1282 struct amdgpu_cs_parser parser = {};
1283 bool reserved_buffers = false;
1284 int i, r;
1285
1286 if (amdgpu_ras_intr_triggered())
1287 return -EHWPOISON;
1288
1289 if (!adev->accel_working)
1290 return -EBUSY;
1291
1292 parser.adev = adev;
1293 parser.filp = filp;
1294
1295 r = amdgpu_cs_parser_init(&parser, data);
1296 if (r) {
1297 DRM_ERROR("Failed to initialize parser %d!\n", r);
1298 goto out;
1299 }
1300
1301 r = amdgpu_cs_ib_fill(adev, &parser);
1302 if (r)
1303 goto out;
1304
1305 r = amdgpu_cs_dependencies(adev, &parser);
1306 if (r) {
1307 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1308 goto out;
1309 }
1310
1311 r = amdgpu_cs_parser_bos(&parser, data);
1312 if (r) {
1313 if (r == -ENOMEM)
1314 DRM_ERROR("Not enough memory for command submission!\n");
1315 else if (r != -ERESTARTSYS && r != -EAGAIN)
1316 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1317 goto out;
1318 }
1319
1320 reserved_buffers = true;
1321
1322 for (i = 0; i < parser.job->num_ibs; i++)
1323 trace_amdgpu_cs(&parser, i);
1324
1325 r = amdgpu_cs_vm_handling(&parser);
1326 if (r)
1327 goto out;
1328
1329 r = amdgpu_cs_submit(&parser, cs);
1330
1331out:
1332 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1333
1334 return r;
1335}
1336
1337/**
1338 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1339 *
1340 * @dev: drm device
1341 * @data: data from userspace
1342 * @filp: file private
1343 *
1344 * Wait for the command submission identified by handle to finish.
1345 */
1346int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1347 struct drm_file *filp)
1348{
1349 union drm_amdgpu_wait_cs *wait = data;
1350 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1351 struct drm_sched_entity *entity;
1352 struct amdgpu_ctx *ctx;
1353 struct dma_fence *fence;
1354 long r;
1355
1356 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1357 if (ctx == NULL)
1358 return -EINVAL;
1359
1360 r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1361 wait->in.ring, &entity);
1362 if (r) {
1363 amdgpu_ctx_put(ctx);
1364 return r;
1365 }
1366
1367 fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1368 if (IS_ERR(fence))
1369 r = PTR_ERR(fence);
1370 else if (fence) {
1371 r = dma_fence_wait_timeout(fence, true, timeout);
1372 if (r > 0 && fence->error)
1373 r = fence->error;
1374 dma_fence_put(fence);
1375 } else
1376 r = 1;
1377
1378 amdgpu_ctx_put(ctx);
1379 if (r < 0)
1380 return r;
1381
1382 memset(wait, 0, sizeof(*wait));
1383 wait->out.status = (r == 0);
1384
1385 return 0;
1386}
1387
1388/**
1389 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1390 *
1391 * @adev: amdgpu device
1392 * @filp: file private
1393 * @user: drm_amdgpu_fence copied from user space
1394 */
1395static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1396 struct drm_file *filp,
1397 struct drm_amdgpu_fence *user)
1398{
1399 struct drm_sched_entity *entity;
1400 struct amdgpu_ctx *ctx;
1401 struct dma_fence *fence;
1402 int r;
1403
1404 ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1405 if (ctx == NULL)
1406 return ERR_PTR(-EINVAL);
1407
1408 r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1409 user->ring, &entity);
1410 if (r) {
1411 amdgpu_ctx_put(ctx);
1412 return ERR_PTR(r);
1413 }
1414
1415 fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1416 amdgpu_ctx_put(ctx);
1417
1418 return fence;
1419}
1420
1421int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1422 struct drm_file *filp)
1423{
1424 struct amdgpu_device *adev = dev->dev_private;
1425 union drm_amdgpu_fence_to_handle *info = data;
1426 struct dma_fence *fence;
1427 struct drm_syncobj *syncobj;
1428 struct sync_file *sync_file;
1429 int fd, r;
1430
1431 fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1432 if (IS_ERR(fence))
1433 return PTR_ERR(fence);
1434
1435 if (!fence)
1436 fence = dma_fence_get_stub();
1437
1438 switch (info->in.what) {
1439 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1440 r = drm_syncobj_create(&syncobj, 0, fence);
1441 dma_fence_put(fence);
1442 if (r)
1443 return r;
1444 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1445 drm_syncobj_put(syncobj);
1446 return r;
1447
1448 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1449 r = drm_syncobj_create(&syncobj, 0, fence);
1450 dma_fence_put(fence);
1451 if (r)
1452 return r;
1453 r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1454 drm_syncobj_put(syncobj);
1455 return r;
1456
1457 case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1458 fd = get_unused_fd_flags(O_CLOEXEC);
1459 if (fd < 0) {
1460 dma_fence_put(fence);
1461 return fd;
1462 }
1463
1464 sync_file = sync_file_create(fence);
1465 dma_fence_put(fence);
1466 if (!sync_file) {
1467 put_unused_fd(fd);
1468 return -ENOMEM;
1469 }
1470
1471 fd_install(fd, sync_file->file);
1472 info->out.handle = fd;
1473 return 0;
1474
1475 default:
1476 return -EINVAL;
1477 }
1478}
1479
1480/**
1481 * amdgpu_cs_wait_all_fence - wait on all fences to signal
1482 *
1483 * @adev: amdgpu device
1484 * @filp: file private
1485 * @wait: wait parameters
1486 * @fences: array of drm_amdgpu_fence
1487 */
1488static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1489 struct drm_file *filp,
1490 union drm_amdgpu_wait_fences *wait,
1491 struct drm_amdgpu_fence *fences)
1492{
1493 uint32_t fence_count = wait->in.fence_count;
1494 unsigned int i;
1495 long r = 1;
1496
1497 for (i = 0; i < fence_count; i++) {
1498 struct dma_fence *fence;
1499 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1500
1501 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1502 if (IS_ERR(fence))
1503 return PTR_ERR(fence);
1504 else if (!fence)
1505 continue;
1506
1507 r = dma_fence_wait_timeout(fence, true, timeout);
1508 dma_fence_put(fence);
1509 if (r < 0)
1510 return r;
1511
1512 if (r == 0)
1513 break;
1514
1515 if (fence->error)
1516 return fence->error;
1517 }
1518
1519 memset(wait, 0, sizeof(*wait));
1520 wait->out.status = (r > 0);
1521
1522 return 0;
1523}
1524
1525/**
1526 * amdgpu_cs_wait_any_fence - wait on any fence to signal
1527 *
1528 * @adev: amdgpu device
1529 * @filp: file private
1530 * @wait: wait parameters
1531 * @fences: array of drm_amdgpu_fence
1532 */
1533static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1534 struct drm_file *filp,
1535 union drm_amdgpu_wait_fences *wait,
1536 struct drm_amdgpu_fence *fences)
1537{
1538 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1539 uint32_t fence_count = wait->in.fence_count;
1540 uint32_t first = ~0;
1541 struct dma_fence **array;
1542 unsigned int i;
1543 long r;
1544
1545 /* Prepare the fence array */
1546 array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1547
1548 if (array == NULL)
1549 return -ENOMEM;
1550
1551 for (i = 0; i < fence_count; i++) {
1552 struct dma_fence *fence;
1553
1554 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1555 if (IS_ERR(fence)) {
1556 r = PTR_ERR(fence);
1557 goto err_free_fence_array;
1558 } else if (fence) {
1559 array[i] = fence;
1560 } else { /* NULL, the fence has been already signaled */
1561 r = 1;
1562 first = i;
1563 goto out;
1564 }
1565 }
1566
1567 r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1568 &first);
1569 if (r < 0)
1570 goto err_free_fence_array;
1571
1572out:
1573 memset(wait, 0, sizeof(*wait));
1574 wait->out.status = (r > 0);
1575 wait->out.first_signaled = first;
1576
1577 if (first < fence_count && array[first])
1578 r = array[first]->error;
1579 else
1580 r = 0;
1581
1582err_free_fence_array:
1583 for (i = 0; i < fence_count; i++)
1584 dma_fence_put(array[i]);
1585 kfree(array);
1586
1587 return r;
1588}
1589
1590/**
1591 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1592 *
1593 * @dev: drm device
1594 * @data: data from userspace
1595 * @filp: file private
1596 */
1597int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1598 struct drm_file *filp)
1599{
1600 struct amdgpu_device *adev = dev->dev_private;
1601 union drm_amdgpu_wait_fences *wait = data;
1602 uint32_t fence_count = wait->in.fence_count;
1603 struct drm_amdgpu_fence *fences_user;
1604 struct drm_amdgpu_fence *fences;
1605 int r;
1606
1607 /* Get the fences from userspace */
1608 fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1609 GFP_KERNEL);
1610 if (fences == NULL)
1611 return -ENOMEM;
1612
1613 fences_user = u64_to_user_ptr(wait->in.fences);
1614 if (copy_from_user(fences, fences_user,
1615 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1616 r = -EFAULT;
1617 goto err_free_fences;
1618 }
1619
1620 if (wait->in.wait_all)
1621 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1622 else
1623 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1624
1625err_free_fences:
1626 kfree(fences);
1627
1628 return r;
1629}
1630
1631/**
1632 * amdgpu_cs_find_bo_va - find bo_va for VM address
1633 *
1634 * @parser: command submission parser context
1635 * @addr: VM address
1636 * @bo: resulting BO of the mapping found
1637 *
1638 * Search the buffer objects in the command submission context for a certain
1639 * virtual memory address. Returns allocation structure when found, NULL
1640 * otherwise.
1641 */
1642int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1643 uint64_t addr, struct amdgpu_bo **bo,
1644 struct amdgpu_bo_va_mapping **map)
1645{
1646 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1647 struct ttm_operation_ctx ctx = { false, false };
1648 struct amdgpu_vm *vm = &fpriv->vm;
1649 struct amdgpu_bo_va_mapping *mapping;
1650 int r;
1651
1652 addr /= AMDGPU_GPU_PAGE_SIZE;
1653
1654 mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1655 if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1656 return -EINVAL;
1657
1658 *bo = mapping->bo_va->base.bo;
1659 *map = mapping;
1660
1661 /* Double check that the BO is reserved by this CS */
1662 if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1663 return -EINVAL;
1664
1665 if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1666 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1667 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1668 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1669 if (r)
1670 return r;
1671 }
1672
1673 return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1674}
1/*
2 * Copyright 2008 Jerome Glisse.
3 * All Rights Reserved.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
23 *
24 * Authors:
25 * Jerome Glisse <glisse@freedesktop.org>
26 */
27#include <linux/list_sort.h>
28#include <linux/pagemap.h>
29#include <drm/drmP.h>
30#include <drm/amdgpu_drm.h>
31#include "amdgpu.h"
32#include "amdgpu_trace.h"
33
34int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
35 u32 ip_instance, u32 ring,
36 struct amdgpu_ring **out_ring)
37{
38 /* Right now all IPs have only one instance - multiple rings. */
39 if (ip_instance != 0) {
40 DRM_ERROR("invalid ip instance: %d\n", ip_instance);
41 return -EINVAL;
42 }
43
44 switch (ip_type) {
45 default:
46 DRM_ERROR("unknown ip type: %d\n", ip_type);
47 return -EINVAL;
48 case AMDGPU_HW_IP_GFX:
49 if (ring < adev->gfx.num_gfx_rings) {
50 *out_ring = &adev->gfx.gfx_ring[ring];
51 } else {
52 DRM_ERROR("only %d gfx rings are supported now\n",
53 adev->gfx.num_gfx_rings);
54 return -EINVAL;
55 }
56 break;
57 case AMDGPU_HW_IP_COMPUTE:
58 if (ring < adev->gfx.num_compute_rings) {
59 *out_ring = &adev->gfx.compute_ring[ring];
60 } else {
61 DRM_ERROR("only %d compute rings are supported now\n",
62 adev->gfx.num_compute_rings);
63 return -EINVAL;
64 }
65 break;
66 case AMDGPU_HW_IP_DMA:
67 if (ring < adev->sdma.num_instances) {
68 *out_ring = &adev->sdma.instance[ring].ring;
69 } else {
70 DRM_ERROR("only %d SDMA rings are supported\n",
71 adev->sdma.num_instances);
72 return -EINVAL;
73 }
74 break;
75 case AMDGPU_HW_IP_UVD:
76 *out_ring = &adev->uvd.ring;
77 break;
78 case AMDGPU_HW_IP_VCE:
79 if (ring < 2){
80 *out_ring = &adev->vce.ring[ring];
81 } else {
82 DRM_ERROR("only two VCE rings are supported\n");
83 return -EINVAL;
84 }
85 break;
86 }
87 return 0;
88}
89
90static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
91 struct amdgpu_user_fence *uf,
92 struct drm_amdgpu_cs_chunk_fence *fence_data)
93{
94 struct drm_gem_object *gobj;
95 uint32_t handle;
96
97 handle = fence_data->handle;
98 gobj = drm_gem_object_lookup(p->adev->ddev, p->filp,
99 fence_data->handle);
100 if (gobj == NULL)
101 return -EINVAL;
102
103 uf->bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
104 uf->offset = fence_data->offset;
105
106 if (amdgpu_ttm_tt_get_usermm(uf->bo->tbo.ttm)) {
107 drm_gem_object_unreference_unlocked(gobj);
108 return -EINVAL;
109 }
110
111 p->uf_entry.robj = amdgpu_bo_ref(uf->bo);
112 p->uf_entry.priority = 0;
113 p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
114 p->uf_entry.tv.shared = true;
115 p->uf_entry.user_pages = NULL;
116
117 drm_gem_object_unreference_unlocked(gobj);
118 return 0;
119}
120
121int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
122{
123 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
124 union drm_amdgpu_cs *cs = data;
125 uint64_t *chunk_array_user;
126 uint64_t *chunk_array;
127 struct amdgpu_user_fence uf = {};
128 unsigned size, num_ibs = 0;
129 int i;
130 int ret;
131
132 if (cs->in.num_chunks == 0)
133 return 0;
134
135 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
136 if (!chunk_array)
137 return -ENOMEM;
138
139 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
140 if (!p->ctx) {
141 ret = -EINVAL;
142 goto free_chunk;
143 }
144
145 /* get chunks */
146 chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
147 if (copy_from_user(chunk_array, chunk_array_user,
148 sizeof(uint64_t)*cs->in.num_chunks)) {
149 ret = -EFAULT;
150 goto put_ctx;
151 }
152
153 p->nchunks = cs->in.num_chunks;
154 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
155 GFP_KERNEL);
156 if (!p->chunks) {
157 ret = -ENOMEM;
158 goto put_ctx;
159 }
160
161 for (i = 0; i < p->nchunks; i++) {
162 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
163 struct drm_amdgpu_cs_chunk user_chunk;
164 uint32_t __user *cdata;
165
166 chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
167 if (copy_from_user(&user_chunk, chunk_ptr,
168 sizeof(struct drm_amdgpu_cs_chunk))) {
169 ret = -EFAULT;
170 i--;
171 goto free_partial_kdata;
172 }
173 p->chunks[i].chunk_id = user_chunk.chunk_id;
174 p->chunks[i].length_dw = user_chunk.length_dw;
175
176 size = p->chunks[i].length_dw;
177 cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
178
179 p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
180 if (p->chunks[i].kdata == NULL) {
181 ret = -ENOMEM;
182 i--;
183 goto free_partial_kdata;
184 }
185 size *= sizeof(uint32_t);
186 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
187 ret = -EFAULT;
188 goto free_partial_kdata;
189 }
190
191 switch (p->chunks[i].chunk_id) {
192 case AMDGPU_CHUNK_ID_IB:
193 ++num_ibs;
194 break;
195
196 case AMDGPU_CHUNK_ID_FENCE:
197 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
198 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
199 ret = -EINVAL;
200 goto free_partial_kdata;
201 }
202
203 ret = amdgpu_cs_user_fence_chunk(p, &uf, (void *)p->chunks[i].kdata);
204 if (ret)
205 goto free_partial_kdata;
206
207 break;
208
209 case AMDGPU_CHUNK_ID_DEPENDENCIES:
210 break;
211
212 default:
213 ret = -EINVAL;
214 goto free_partial_kdata;
215 }
216 }
217
218 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job);
219 if (ret)
220 goto free_all_kdata;
221
222 p->job->uf = uf;
223
224 kfree(chunk_array);
225 return 0;
226
227free_all_kdata:
228 i = p->nchunks - 1;
229free_partial_kdata:
230 for (; i >= 0; i--)
231 drm_free_large(p->chunks[i].kdata);
232 kfree(p->chunks);
233put_ctx:
234 amdgpu_ctx_put(p->ctx);
235free_chunk:
236 kfree(chunk_array);
237
238 return ret;
239}
240
241/* Returns how many bytes TTM can move per IB.
242 */
243static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
244{
245 u64 real_vram_size = adev->mc.real_vram_size;
246 u64 vram_usage = atomic64_read(&adev->vram_usage);
247
248 /* This function is based on the current VRAM usage.
249 *
250 * - If all of VRAM is free, allow relocating the number of bytes that
251 * is equal to 1/4 of the size of VRAM for this IB.
252
253 * - If more than one half of VRAM is occupied, only allow relocating
254 * 1 MB of data for this IB.
255 *
256 * - From 0 to one half of used VRAM, the threshold decreases
257 * linearly.
258 * __________________
259 * 1/4 of -|\ |
260 * VRAM | \ |
261 * | \ |
262 * | \ |
263 * | \ |
264 * | \ |
265 * | \ |
266 * | \________|1 MB
267 * |----------------|
268 * VRAM 0 % 100 %
269 * used used
270 *
271 * Note: It's a threshold, not a limit. The threshold must be crossed
272 * for buffer relocations to stop, so any buffer of an arbitrary size
273 * can be moved as long as the threshold isn't crossed before
274 * the relocation takes place. We don't want to disable buffer
275 * relocations completely.
276 *
277 * The idea is that buffers should be placed in VRAM at creation time
278 * and TTM should only do a minimum number of relocations during
279 * command submission. In practice, you need to submit at least
280 * a dozen IBs to move all buffers to VRAM if they are in GTT.
281 *
282 * Also, things can get pretty crazy under memory pressure and actual
283 * VRAM usage can change a lot, so playing safe even at 50% does
284 * consistently increase performance.
285 */
286
287 u64 half_vram = real_vram_size >> 1;
288 u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
289 u64 bytes_moved_threshold = half_free_vram >> 1;
290 return max(bytes_moved_threshold, 1024*1024ull);
291}
292
293int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
294 struct list_head *validated)
295{
296 struct amdgpu_bo_list_entry *lobj;
297 u64 initial_bytes_moved;
298 int r;
299
300 list_for_each_entry(lobj, validated, tv.head) {
301 struct amdgpu_bo *bo = lobj->robj;
302 bool binding_userptr = false;
303 struct mm_struct *usermm;
304 uint32_t domain;
305
306 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
307 if (usermm && usermm != current->mm)
308 return -EPERM;
309
310 /* Check if we have user pages and nobody bound the BO already */
311 if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
312 size_t size = sizeof(struct page *);
313
314 size *= bo->tbo.ttm->num_pages;
315 memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
316 binding_userptr = true;
317 }
318
319 if (bo->pin_count)
320 continue;
321
322 /* Avoid moving this one if we have moved too many buffers
323 * for this IB already.
324 *
325 * Note that this allows moving at least one buffer of
326 * any size, because it doesn't take the current "bo"
327 * into account. We don't want to disallow buffer moves
328 * completely.
329 */
330 if (p->bytes_moved <= p->bytes_moved_threshold)
331 domain = bo->prefered_domains;
332 else
333 domain = bo->allowed_domains;
334
335 retry:
336 amdgpu_ttm_placement_from_domain(bo, domain);
337 initial_bytes_moved = atomic64_read(&bo->adev->num_bytes_moved);
338 r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
339 p->bytes_moved += atomic64_read(&bo->adev->num_bytes_moved) -
340 initial_bytes_moved;
341
342 if (unlikely(r)) {
343 if (r != -ERESTARTSYS && domain != bo->allowed_domains) {
344 domain = bo->allowed_domains;
345 goto retry;
346 }
347 return r;
348 }
349
350 if (binding_userptr) {
351 drm_free_large(lobj->user_pages);
352 lobj->user_pages = NULL;
353 }
354 }
355 return 0;
356}
357
358static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
359 union drm_amdgpu_cs *cs)
360{
361 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
362 struct amdgpu_bo_list_entry *e;
363 struct list_head duplicates;
364 bool need_mmap_lock = false;
365 unsigned i, tries = 10;
366 int r;
367
368 INIT_LIST_HEAD(&p->validated);
369
370 p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
371 if (p->bo_list) {
372 need_mmap_lock = p->bo_list->first_userptr !=
373 p->bo_list->num_entries;
374 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
375 }
376
377 INIT_LIST_HEAD(&duplicates);
378 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
379
380 if (p->job->uf.bo)
381 list_add(&p->uf_entry.tv.head, &p->validated);
382
383 if (need_mmap_lock)
384 down_read(¤t->mm->mmap_sem);
385
386 while (1) {
387 struct list_head need_pages;
388 unsigned i;
389
390 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
391 &duplicates);
392 if (unlikely(r != 0))
393 goto error_free_pages;
394
395 /* Without a BO list we don't have userptr BOs */
396 if (!p->bo_list)
397 break;
398
399 INIT_LIST_HEAD(&need_pages);
400 for (i = p->bo_list->first_userptr;
401 i < p->bo_list->num_entries; ++i) {
402
403 e = &p->bo_list->array[i];
404
405 if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
406 &e->user_invalidated) && e->user_pages) {
407
408 /* We acquired a page array, but somebody
409 * invalidated it. Free it an try again
410 */
411 release_pages(e->user_pages,
412 e->robj->tbo.ttm->num_pages,
413 false);
414 drm_free_large(e->user_pages);
415 e->user_pages = NULL;
416 }
417
418 if (e->robj->tbo.ttm->state != tt_bound &&
419 !e->user_pages) {
420 list_del(&e->tv.head);
421 list_add(&e->tv.head, &need_pages);
422
423 amdgpu_bo_unreserve(e->robj);
424 }
425 }
426
427 if (list_empty(&need_pages))
428 break;
429
430 /* Unreserve everything again. */
431 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
432
433 /* We tried to often, just abort */
434 if (!--tries) {
435 r = -EDEADLK;
436 goto error_free_pages;
437 }
438
439 /* Fill the page arrays for all useptrs. */
440 list_for_each_entry(e, &need_pages, tv.head) {
441 struct ttm_tt *ttm = e->robj->tbo.ttm;
442
443 e->user_pages = drm_calloc_large(ttm->num_pages,
444 sizeof(struct page*));
445 if (!e->user_pages) {
446 r = -ENOMEM;
447 goto error_free_pages;
448 }
449
450 r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
451 if (r) {
452 drm_free_large(e->user_pages);
453 e->user_pages = NULL;
454 goto error_free_pages;
455 }
456 }
457
458 /* And try again. */
459 list_splice(&need_pages, &p->validated);
460 }
461
462 amdgpu_vm_get_pt_bos(&fpriv->vm, &duplicates);
463
464 p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
465 p->bytes_moved = 0;
466
467 r = amdgpu_cs_list_validate(p, &duplicates);
468 if (r)
469 goto error_validate;
470
471 r = amdgpu_cs_list_validate(p, &p->validated);
472 if (r)
473 goto error_validate;
474
475 if (p->bo_list) {
476 struct amdgpu_vm *vm = &fpriv->vm;
477 unsigned i;
478
479 for (i = 0; i < p->bo_list->num_entries; i++) {
480 struct amdgpu_bo *bo = p->bo_list->array[i].robj;
481
482 p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
483 }
484 }
485
486error_validate:
487 if (r) {
488 amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
489 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
490 }
491
492error_free_pages:
493
494 if (need_mmap_lock)
495 up_read(¤t->mm->mmap_sem);
496
497 if (p->bo_list) {
498 for (i = p->bo_list->first_userptr;
499 i < p->bo_list->num_entries; ++i) {
500 e = &p->bo_list->array[i];
501
502 if (!e->user_pages)
503 continue;
504
505 release_pages(e->user_pages,
506 e->robj->tbo.ttm->num_pages,
507 false);
508 drm_free_large(e->user_pages);
509 }
510 }
511
512 return r;
513}
514
515static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
516{
517 struct amdgpu_bo_list_entry *e;
518 int r;
519
520 list_for_each_entry(e, &p->validated, tv.head) {
521 struct reservation_object *resv = e->robj->tbo.resv;
522 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
523
524 if (r)
525 return r;
526 }
527 return 0;
528}
529
530static int cmp_size_smaller_first(void *priv, struct list_head *a,
531 struct list_head *b)
532{
533 struct amdgpu_bo_list_entry *la = list_entry(a, struct amdgpu_bo_list_entry, tv.head);
534 struct amdgpu_bo_list_entry *lb = list_entry(b, struct amdgpu_bo_list_entry, tv.head);
535
536 /* Sort A before B if A is smaller. */
537 return (int)la->robj->tbo.num_pages - (int)lb->robj->tbo.num_pages;
538}
539
540/**
541 * cs_parser_fini() - clean parser states
542 * @parser: parser structure holding parsing context.
543 * @error: error number
544 *
545 * If error is set than unvalidate buffer, otherwise just free memory
546 * used by parsing context.
547 **/
548static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
549{
550 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
551 unsigned i;
552
553 if (!error) {
554 amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);
555
556 /* Sort the buffer list from the smallest to largest buffer,
557 * which affects the order of buffers in the LRU list.
558 * This assures that the smallest buffers are added first
559 * to the LRU list, so they are likely to be later evicted
560 * first, instead of large buffers whose eviction is more
561 * expensive.
562 *
563 * This slightly lowers the number of bytes moved by TTM
564 * per frame under memory pressure.
565 */
566 list_sort(NULL, &parser->validated, cmp_size_smaller_first);
567
568 ttm_eu_fence_buffer_objects(&parser->ticket,
569 &parser->validated,
570 parser->fence);
571 } else if (backoff) {
572 ttm_eu_backoff_reservation(&parser->ticket,
573 &parser->validated);
574 }
575 fence_put(parser->fence);
576
577 if (parser->ctx)
578 amdgpu_ctx_put(parser->ctx);
579 if (parser->bo_list)
580 amdgpu_bo_list_put(parser->bo_list);
581
582 for (i = 0; i < parser->nchunks; i++)
583 drm_free_large(parser->chunks[i].kdata);
584 kfree(parser->chunks);
585 if (parser->job)
586 amdgpu_job_free(parser->job);
587 amdgpu_bo_unref(&parser->uf_entry.robj);
588}
589
590static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
591 struct amdgpu_vm *vm)
592{
593 struct amdgpu_device *adev = p->adev;
594 struct amdgpu_bo_va *bo_va;
595 struct amdgpu_bo *bo;
596 int i, r;
597
598 r = amdgpu_vm_update_page_directory(adev, vm);
599 if (r)
600 return r;
601
602 r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
603 if (r)
604 return r;
605
606 r = amdgpu_vm_clear_freed(adev, vm);
607 if (r)
608 return r;
609
610 if (p->bo_list) {
611 for (i = 0; i < p->bo_list->num_entries; i++) {
612 struct fence *f;
613
614 /* ignore duplicates */
615 bo = p->bo_list->array[i].robj;
616 if (!bo)
617 continue;
618
619 bo_va = p->bo_list->array[i].bo_va;
620 if (bo_va == NULL)
621 continue;
622
623 r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
624 if (r)
625 return r;
626
627 f = bo_va->last_pt_update;
628 r = amdgpu_sync_fence(adev, &p->job->sync, f);
629 if (r)
630 return r;
631 }
632
633 }
634
635 r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
636
637 if (amdgpu_vm_debug && p->bo_list) {
638 /* Invalidate all BOs to test for userspace bugs */
639 for (i = 0; i < p->bo_list->num_entries; i++) {
640 /* ignore duplicates */
641 bo = p->bo_list->array[i].robj;
642 if (!bo)
643 continue;
644
645 amdgpu_vm_bo_invalidate(adev, bo);
646 }
647 }
648
649 return r;
650}
651
652static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
653 struct amdgpu_cs_parser *p)
654{
655 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
656 struct amdgpu_vm *vm = &fpriv->vm;
657 struct amdgpu_ring *ring = p->job->ring;
658 int i, r;
659
660 /* Only for UVD/VCE VM emulation */
661 if (ring->funcs->parse_cs) {
662 for (i = 0; i < p->job->num_ibs; i++) {
663 r = amdgpu_ring_parse_cs(ring, p, i);
664 if (r)
665 return r;
666 }
667 }
668
669 r = amdgpu_bo_vm_update_pte(p, vm);
670 if (!r)
671 amdgpu_cs_sync_rings(p);
672
673 return r;
674}
675
676static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
677{
678 if (r == -EDEADLK) {
679 r = amdgpu_gpu_reset(adev);
680 if (!r)
681 r = -EAGAIN;
682 }
683 return r;
684}
685
686static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
687 struct amdgpu_cs_parser *parser)
688{
689 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
690 struct amdgpu_vm *vm = &fpriv->vm;
691 int i, j;
692 int r;
693
694 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
695 struct amdgpu_cs_chunk *chunk;
696 struct amdgpu_ib *ib;
697 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
698 struct amdgpu_ring *ring;
699
700 chunk = &parser->chunks[i];
701 ib = &parser->job->ibs[j];
702 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
703
704 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
705 continue;
706
707 r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
708 chunk_ib->ip_instance, chunk_ib->ring,
709 &ring);
710 if (r)
711 return r;
712
713 if (parser->job->ring && parser->job->ring != ring)
714 return -EINVAL;
715
716 parser->job->ring = ring;
717
718 if (ring->funcs->parse_cs) {
719 struct amdgpu_bo_va_mapping *m;
720 struct amdgpu_bo *aobj = NULL;
721 uint64_t offset;
722 uint8_t *kptr;
723
724 m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
725 &aobj);
726 if (!aobj) {
727 DRM_ERROR("IB va_start is invalid\n");
728 return -EINVAL;
729 }
730
731 if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
732 (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
733 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
734 return -EINVAL;
735 }
736
737 /* the IB should be reserved at this point */
738 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
739 if (r) {
740 return r;
741 }
742
743 offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
744 kptr += chunk_ib->va_start - offset;
745
746 r = amdgpu_ib_get(adev, NULL, chunk_ib->ib_bytes, ib);
747 if (r) {
748 DRM_ERROR("Failed to get ib !\n");
749 return r;
750 }
751
752 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
753 amdgpu_bo_kunmap(aobj);
754 } else {
755 r = amdgpu_ib_get(adev, vm, 0, ib);
756 if (r) {
757 DRM_ERROR("Failed to get ib !\n");
758 return r;
759 }
760
761 ib->gpu_addr = chunk_ib->va_start;
762 }
763
764 ib->length_dw = chunk_ib->ib_bytes / 4;
765 ib->flags = chunk_ib->flags;
766 ib->ctx = parser->ctx;
767 j++;
768 }
769
770 /* add GDS resources to first IB */
771 if (parser->bo_list) {
772 struct amdgpu_bo *gds = parser->bo_list->gds_obj;
773 struct amdgpu_bo *gws = parser->bo_list->gws_obj;
774 struct amdgpu_bo *oa = parser->bo_list->oa_obj;
775 struct amdgpu_ib *ib = &parser->job->ibs[0];
776
777 if (gds) {
778 ib->gds_base = amdgpu_bo_gpu_offset(gds);
779 ib->gds_size = amdgpu_bo_size(gds);
780 }
781 if (gws) {
782 ib->gws_base = amdgpu_bo_gpu_offset(gws);
783 ib->gws_size = amdgpu_bo_size(gws);
784 }
785 if (oa) {
786 ib->oa_base = amdgpu_bo_gpu_offset(oa);
787 ib->oa_size = amdgpu_bo_size(oa);
788 }
789 }
790 /* wrap the last IB with user fence */
791 if (parser->job->uf.bo) {
792 struct amdgpu_ib *ib = &parser->job->ibs[parser->job->num_ibs - 1];
793
794 /* UVD & VCE fw doesn't support user fences */
795 if (parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
796 parser->job->ring->type == AMDGPU_RING_TYPE_VCE)
797 return -EINVAL;
798
799 ib->user = &parser->job->uf;
800 }
801
802 return 0;
803}
804
805static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
806 struct amdgpu_cs_parser *p)
807{
808 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
809 int i, j, r;
810
811 for (i = 0; i < p->nchunks; ++i) {
812 struct drm_amdgpu_cs_chunk_dep *deps;
813 struct amdgpu_cs_chunk *chunk;
814 unsigned num_deps;
815
816 chunk = &p->chunks[i];
817
818 if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
819 continue;
820
821 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
822 num_deps = chunk->length_dw * 4 /
823 sizeof(struct drm_amdgpu_cs_chunk_dep);
824
825 for (j = 0; j < num_deps; ++j) {
826 struct amdgpu_ring *ring;
827 struct amdgpu_ctx *ctx;
828 struct fence *fence;
829
830 r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
831 deps[j].ip_instance,
832 deps[j].ring, &ring);
833 if (r)
834 return r;
835
836 ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
837 if (ctx == NULL)
838 return -EINVAL;
839
840 fence = amdgpu_ctx_get_fence(ctx, ring,
841 deps[j].handle);
842 if (IS_ERR(fence)) {
843 r = PTR_ERR(fence);
844 amdgpu_ctx_put(ctx);
845 return r;
846
847 } else if (fence) {
848 r = amdgpu_sync_fence(adev, &p->job->sync,
849 fence);
850 fence_put(fence);
851 amdgpu_ctx_put(ctx);
852 if (r)
853 return r;
854 }
855 }
856 }
857
858 return 0;
859}
860
861static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
862 union drm_amdgpu_cs *cs)
863{
864 struct amdgpu_ring *ring = p->job->ring;
865 struct amd_sched_fence *fence;
866 struct amdgpu_job *job;
867
868 job = p->job;
869 p->job = NULL;
870
871 job->base.sched = &ring->sched;
872 job->base.s_entity = &p->ctx->rings[ring->idx].entity;
873 job->owner = p->filp;
874
875 fence = amd_sched_fence_create(job->base.s_entity, p->filp);
876 if (!fence) {
877 amdgpu_job_free(job);
878 return -ENOMEM;
879 }
880
881 job->base.s_fence = fence;
882 p->fence = fence_get(&fence->base);
883
884 cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring,
885 &fence->base);
886 job->ibs[job->num_ibs - 1].sequence = cs->out.handle;
887
888 trace_amdgpu_cs_ioctl(job);
889 amd_sched_entity_push_job(&job->base);
890
891 return 0;
892}
893
894int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
895{
896 struct amdgpu_device *adev = dev->dev_private;
897 union drm_amdgpu_cs *cs = data;
898 struct amdgpu_cs_parser parser = {};
899 bool reserved_buffers = false;
900 int i, r;
901
902 if (!adev->accel_working)
903 return -EBUSY;
904
905 parser.adev = adev;
906 parser.filp = filp;
907
908 r = amdgpu_cs_parser_init(&parser, data);
909 if (r) {
910 DRM_ERROR("Failed to initialize parser !\n");
911 amdgpu_cs_parser_fini(&parser, r, false);
912 r = amdgpu_cs_handle_lockup(adev, r);
913 return r;
914 }
915 r = amdgpu_cs_parser_bos(&parser, data);
916 if (r == -ENOMEM)
917 DRM_ERROR("Not enough memory for command submission!\n");
918 else if (r && r != -ERESTARTSYS)
919 DRM_ERROR("Failed to process the buffer list %d!\n", r);
920 else if (!r) {
921 reserved_buffers = true;
922 r = amdgpu_cs_ib_fill(adev, &parser);
923 }
924
925 if (!r) {
926 r = amdgpu_cs_dependencies(adev, &parser);
927 if (r)
928 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
929 }
930
931 if (r)
932 goto out;
933
934 for (i = 0; i < parser.job->num_ibs; i++)
935 trace_amdgpu_cs(&parser, i);
936
937 r = amdgpu_cs_ib_vm_chunk(adev, &parser);
938 if (r)
939 goto out;
940
941 r = amdgpu_cs_submit(&parser, cs);
942
943out:
944 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
945 r = amdgpu_cs_handle_lockup(adev, r);
946 return r;
947}
948
949/**
950 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
951 *
952 * @dev: drm device
953 * @data: data from userspace
954 * @filp: file private
955 *
956 * Wait for the command submission identified by handle to finish.
957 */
958int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
959 struct drm_file *filp)
960{
961 union drm_amdgpu_wait_cs *wait = data;
962 struct amdgpu_device *adev = dev->dev_private;
963 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
964 struct amdgpu_ring *ring = NULL;
965 struct amdgpu_ctx *ctx;
966 struct fence *fence;
967 long r;
968
969 r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
970 wait->in.ring, &ring);
971 if (r)
972 return r;
973
974 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
975 if (ctx == NULL)
976 return -EINVAL;
977
978 fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
979 if (IS_ERR(fence))
980 r = PTR_ERR(fence);
981 else if (fence) {
982 r = fence_wait_timeout(fence, true, timeout);
983 fence_put(fence);
984 } else
985 r = 1;
986
987 amdgpu_ctx_put(ctx);
988 if (r < 0)
989 return r;
990
991 memset(wait, 0, sizeof(*wait));
992 wait->out.status = (r == 0);
993
994 return 0;
995}
996
997/**
998 * amdgpu_cs_find_bo_va - find bo_va for VM address
999 *
1000 * @parser: command submission parser context
1001 * @addr: VM address
1002 * @bo: resulting BO of the mapping found
1003 *
1004 * Search the buffer objects in the command submission context for a certain
1005 * virtual memory address. Returns allocation structure when found, NULL
1006 * otherwise.
1007 */
1008struct amdgpu_bo_va_mapping *
1009amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1010 uint64_t addr, struct amdgpu_bo **bo)
1011{
1012 struct amdgpu_bo_va_mapping *mapping;
1013 unsigned i;
1014
1015 if (!parser->bo_list)
1016 return NULL;
1017
1018 addr /= AMDGPU_GPU_PAGE_SIZE;
1019
1020 for (i = 0; i < parser->bo_list->num_entries; i++) {
1021 struct amdgpu_bo_list_entry *lobj;
1022
1023 lobj = &parser->bo_list->array[i];
1024 if (!lobj->bo_va)
1025 continue;
1026
1027 list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1028 if (mapping->it.start > addr ||
1029 addr > mapping->it.last)
1030 continue;
1031
1032 *bo = lobj->bo_va->bo;
1033 return mapping;
1034 }
1035
1036 list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
1037 if (mapping->it.start > addr ||
1038 addr > mapping->it.last)
1039 continue;
1040
1041 *bo = lobj->bo_va->bo;
1042 return mapping;
1043 }
1044 }
1045
1046 return NULL;
1047}