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#include <linux/pagemap.h>
28#include <linux/sync_file.h>
29#include <drm/drmP.h>
30#include <drm/amdgpu_drm.h>
31#include <drm/drm_syncobj.h>
32#include "amdgpu.h"
33#include "amdgpu_trace.h"
34
35static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
36 struct drm_amdgpu_cs_chunk_fence *data,
37 uint32_t *offset)
38{
39 struct drm_gem_object *gobj;
40 unsigned long size;
41
42 gobj = drm_gem_object_lookup(p->filp, data->handle);
43 if (gobj == NULL)
44 return -EINVAL;
45
46 p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
47 p->uf_entry.priority = 0;
48 p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
49 p->uf_entry.tv.shared = true;
50 p->uf_entry.user_pages = NULL;
51
52 size = amdgpu_bo_size(p->uf_entry.robj);
53 if (size != PAGE_SIZE || (data->offset + 8) > size)
54 return -EINVAL;
55
56 *offset = data->offset;
57
58 drm_gem_object_put_unlocked(gobj);
59
60 if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
61 amdgpu_bo_unref(&p->uf_entry.robj);
62 return -EINVAL;
63 }
64
65 return 0;
66}
67
68static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
69{
70 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
71 struct amdgpu_vm *vm = &fpriv->vm;
72 union drm_amdgpu_cs *cs = data;
73 uint64_t *chunk_array_user;
74 uint64_t *chunk_array;
75 unsigned size, num_ibs = 0;
76 uint32_t uf_offset = 0;
77 int i;
78 int ret;
79
80 if (cs->in.num_chunks == 0)
81 return 0;
82
83 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
84 if (!chunk_array)
85 return -ENOMEM;
86
87 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
88 if (!p->ctx) {
89 ret = -EINVAL;
90 goto free_chunk;
91 }
92
93 /* skip guilty context job */
94 if (atomic_read(&p->ctx->guilty) == 1) {
95 ret = -ECANCELED;
96 goto free_chunk;
97 }
98
99 mutex_lock(&p->ctx->lock);
100
101 /* get chunks */
102 chunk_array_user = u64_to_user_ptr(cs->in.chunks);
103 if (copy_from_user(chunk_array, chunk_array_user,
104 sizeof(uint64_t)*cs->in.num_chunks)) {
105 ret = -EFAULT;
106 goto free_chunk;
107 }
108
109 p->nchunks = cs->in.num_chunks;
110 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
111 GFP_KERNEL);
112 if (!p->chunks) {
113 ret = -ENOMEM;
114 goto free_chunk;
115 }
116
117 for (i = 0; i < p->nchunks; i++) {
118 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
119 struct drm_amdgpu_cs_chunk user_chunk;
120 uint32_t __user *cdata;
121
122 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
123 if (copy_from_user(&user_chunk, chunk_ptr,
124 sizeof(struct drm_amdgpu_cs_chunk))) {
125 ret = -EFAULT;
126 i--;
127 goto free_partial_kdata;
128 }
129 p->chunks[i].chunk_id = user_chunk.chunk_id;
130 p->chunks[i].length_dw = user_chunk.length_dw;
131
132 size = p->chunks[i].length_dw;
133 cdata = u64_to_user_ptr(user_chunk.chunk_data);
134
135 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
136 if (p->chunks[i].kdata == NULL) {
137 ret = -ENOMEM;
138 i--;
139 goto free_partial_kdata;
140 }
141 size *= sizeof(uint32_t);
142 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
143 ret = -EFAULT;
144 goto free_partial_kdata;
145 }
146
147 switch (p->chunks[i].chunk_id) {
148 case AMDGPU_CHUNK_ID_IB:
149 ++num_ibs;
150 break;
151
152 case AMDGPU_CHUNK_ID_FENCE:
153 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
154 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
155 ret = -EINVAL;
156 goto free_partial_kdata;
157 }
158
159 ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
160 &uf_offset);
161 if (ret)
162 goto free_partial_kdata;
163
164 break;
165
166 case AMDGPU_CHUNK_ID_DEPENDENCIES:
167 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
168 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
169 break;
170
171 default:
172 ret = -EINVAL;
173 goto free_partial_kdata;
174 }
175 }
176
177 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
178 if (ret)
179 goto free_all_kdata;
180
181 if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
182 ret = -ECANCELED;
183 goto free_all_kdata;
184 }
185
186 if (p->uf_entry.robj)
187 p->job->uf_addr = uf_offset;
188 kfree(chunk_array);
189 return 0;
190
191free_all_kdata:
192 i = p->nchunks - 1;
193free_partial_kdata:
194 for (; i >= 0; i--)
195 kvfree(p->chunks[i].kdata);
196 kfree(p->chunks);
197 p->chunks = NULL;
198 p->nchunks = 0;
199free_chunk:
200 kfree(chunk_array);
201
202 return ret;
203}
204
205/* Convert microseconds to bytes. */
206static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
207{
208 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
209 return 0;
210
211 /* Since accum_us is incremented by a million per second, just
212 * multiply it by the number of MB/s to get the number of bytes.
213 */
214 return us << adev->mm_stats.log2_max_MBps;
215}
216
217static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
218{
219 if (!adev->mm_stats.log2_max_MBps)
220 return 0;
221
222 return bytes >> adev->mm_stats.log2_max_MBps;
223}
224
225/* Returns how many bytes TTM can move right now. If no bytes can be moved,
226 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
227 * which means it can go over the threshold once. If that happens, the driver
228 * will be in debt and no other buffer migrations can be done until that debt
229 * is repaid.
230 *
231 * This approach allows moving a buffer of any size (it's important to allow
232 * that).
233 *
234 * The currency is simply time in microseconds and it increases as the clock
235 * ticks. The accumulated microseconds (us) are converted to bytes and
236 * returned.
237 */
238static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
239 u64 *max_bytes,
240 u64 *max_vis_bytes)
241{
242 s64 time_us, increment_us;
243 u64 free_vram, total_vram, used_vram;
244
245 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
246 * throttling.
247 *
248 * It means that in order to get full max MBps, at least 5 IBs per
249 * second must be submitted and not more than 200ms apart from each
250 * other.
251 */
252 const s64 us_upper_bound = 200000;
253
254 if (!adev->mm_stats.log2_max_MBps) {
255 *max_bytes = 0;
256 *max_vis_bytes = 0;
257 return;
258 }
259
260 total_vram = adev->gmc.real_vram_size - adev->vram_pin_size;
261 used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
262 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
263
264 spin_lock(&adev->mm_stats.lock);
265
266 /* Increase the amount of accumulated us. */
267 time_us = ktime_to_us(ktime_get());
268 increment_us = time_us - adev->mm_stats.last_update_us;
269 adev->mm_stats.last_update_us = time_us;
270 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
271 us_upper_bound);
272
273 /* This prevents the short period of low performance when the VRAM
274 * usage is low and the driver is in debt or doesn't have enough
275 * accumulated us to fill VRAM quickly.
276 *
277 * The situation can occur in these cases:
278 * - a lot of VRAM is freed by userspace
279 * - the presence of a big buffer causes a lot of evictions
280 * (solution: split buffers into smaller ones)
281 *
282 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
283 * accum_us to a positive number.
284 */
285 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
286 s64 min_us;
287
288 /* Be more aggresive on dGPUs. Try to fill a portion of free
289 * VRAM now.
290 */
291 if (!(adev->flags & AMD_IS_APU))
292 min_us = bytes_to_us(adev, free_vram / 4);
293 else
294 min_us = 0; /* Reset accum_us on APUs. */
295
296 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
297 }
298
299 /* This is set to 0 if the driver is in debt to disallow (optional)
300 * buffer moves.
301 */
302 *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
303
304 /* Do the same for visible VRAM if half of it is free */
305 if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size) {
306 u64 total_vis_vram = adev->gmc.visible_vram_size;
307 u64 used_vis_vram =
308 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
309
310 if (used_vis_vram < total_vis_vram) {
311 u64 free_vis_vram = total_vis_vram - used_vis_vram;
312 adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
313 increment_us, us_upper_bound);
314
315 if (free_vis_vram >= total_vis_vram / 2)
316 adev->mm_stats.accum_us_vis =
317 max(bytes_to_us(adev, free_vis_vram / 2),
318 adev->mm_stats.accum_us_vis);
319 }
320
321 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
322 } else {
323 *max_vis_bytes = 0;
324 }
325
326 spin_unlock(&adev->mm_stats.lock);
327}
328
329/* Report how many bytes have really been moved for the last command
330 * submission. This can result in a debt that can stop buffer migrations
331 * temporarily.
332 */
333void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
334 u64 num_vis_bytes)
335{
336 spin_lock(&adev->mm_stats.lock);
337 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
338 adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
339 spin_unlock(&adev->mm_stats.lock);
340}
341
342static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
343 struct amdgpu_bo *bo)
344{
345 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
346 struct ttm_operation_ctx ctx = {
347 .interruptible = true,
348 .no_wait_gpu = false,
349 .resv = bo->tbo.resv,
350 .flags = 0
351 };
352 uint32_t domain;
353 int r;
354
355 if (bo->pin_count)
356 return 0;
357
358 /* Don't move this buffer if we have depleted our allowance
359 * to move it. Don't move anything if the threshold is zero.
360 */
361 if (p->bytes_moved < p->bytes_moved_threshold) {
362 if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
363 (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
364 /* And don't move a CPU_ACCESS_REQUIRED BO to limited
365 * visible VRAM if we've depleted our allowance to do
366 * that.
367 */
368 if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
369 domain = bo->preferred_domains;
370 else
371 domain = bo->allowed_domains;
372 } else {
373 domain = bo->preferred_domains;
374 }
375 } else {
376 domain = bo->allowed_domains;
377 }
378
379retry:
380 amdgpu_ttm_placement_from_domain(bo, domain);
381 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
382
383 p->bytes_moved += ctx.bytes_moved;
384 if (adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
385 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
386 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT)
387 p->bytes_moved_vis += ctx.bytes_moved;
388
389 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
390 domain = bo->allowed_domains;
391 goto retry;
392 }
393
394 return r;
395}
396
397/* Last resort, try to evict something from the current working set */
398static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
399 struct amdgpu_bo *validated)
400{
401 uint32_t domain = validated->allowed_domains;
402 struct ttm_operation_ctx ctx = { true, false };
403 int r;
404
405 if (!p->evictable)
406 return false;
407
408 for (;&p->evictable->tv.head != &p->validated;
409 p->evictable = list_prev_entry(p->evictable, tv.head)) {
410
411 struct amdgpu_bo_list_entry *candidate = p->evictable;
412 struct amdgpu_bo *bo = candidate->robj;
413 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
414 u64 initial_bytes_moved, bytes_moved;
415 bool update_bytes_moved_vis;
416 uint32_t other;
417
418 /* If we reached our current BO we can forget it */
419 if (candidate->robj == validated)
420 break;
421
422 /* We can't move pinned BOs here */
423 if (bo->pin_count)
424 continue;
425
426 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
427
428 /* Check if this BO is in one of the domains we need space for */
429 if (!(other & domain))
430 continue;
431
432 /* Check if we can move this BO somewhere else */
433 other = bo->allowed_domains & ~domain;
434 if (!other)
435 continue;
436
437 /* Good we can try to move this BO somewhere else */
438 amdgpu_ttm_placement_from_domain(bo, other);
439 update_bytes_moved_vis =
440 adev->gmc.visible_vram_size < adev->gmc.real_vram_size &&
441 bo->tbo.mem.mem_type == TTM_PL_VRAM &&
442 bo->tbo.mem.start < adev->gmc.visible_vram_size >> PAGE_SHIFT;
443 initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
444 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
445 bytes_moved = atomic64_read(&adev->num_bytes_moved) -
446 initial_bytes_moved;
447 p->bytes_moved += bytes_moved;
448 if (update_bytes_moved_vis)
449 p->bytes_moved_vis += bytes_moved;
450
451 if (unlikely(r))
452 break;
453
454 p->evictable = list_prev_entry(p->evictable, tv.head);
455 list_move(&candidate->tv.head, &p->validated);
456
457 return true;
458 }
459
460 return false;
461}
462
463static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
464{
465 struct amdgpu_cs_parser *p = param;
466 int r;
467
468 do {
469 r = amdgpu_cs_bo_validate(p, bo);
470 } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
471 if (r)
472 return r;
473
474 if (bo->shadow)
475 r = amdgpu_cs_bo_validate(p, bo->shadow);
476
477 return r;
478}
479
480static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
481 struct list_head *validated)
482{
483 struct ttm_operation_ctx ctx = { true, false };
484 struct amdgpu_bo_list_entry *lobj;
485 int r;
486
487 list_for_each_entry(lobj, validated, tv.head) {
488 struct amdgpu_bo *bo = lobj->robj;
489 bool binding_userptr = false;
490 struct mm_struct *usermm;
491
492 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
493 if (usermm && usermm != current->mm)
494 return -EPERM;
495
496 /* Check if we have user pages and nobody bound the BO already */
497 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
498 lobj->user_pages) {
499 amdgpu_ttm_placement_from_domain(bo,
500 AMDGPU_GEM_DOMAIN_CPU);
501 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
502 if (r)
503 return r;
504 amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
505 lobj->user_pages);
506 binding_userptr = true;
507 }
508
509 if (p->evictable == lobj)
510 p->evictable = NULL;
511
512 r = amdgpu_cs_validate(p, bo);
513 if (r)
514 return r;
515
516 if (binding_userptr) {
517 kvfree(lobj->user_pages);
518 lobj->user_pages = NULL;
519 }
520 }
521 return 0;
522}
523
524static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
525 union drm_amdgpu_cs *cs)
526{
527 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
528 struct amdgpu_bo_list_entry *e;
529 struct list_head duplicates;
530 unsigned i, tries = 10;
531 int r;
532
533 INIT_LIST_HEAD(&p->validated);
534
535 p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
536 if (p->bo_list) {
537 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
538 if (p->bo_list->first_userptr != p->bo_list->num_entries)
539 p->mn = amdgpu_mn_get(p->adev);
540 }
541
542 INIT_LIST_HEAD(&duplicates);
543 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
544
545 if (p->uf_entry.robj && !p->uf_entry.robj->parent)
546 list_add(&p->uf_entry.tv.head, &p->validated);
547
548 while (1) {
549 struct list_head need_pages;
550 unsigned i;
551
552 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
553 &duplicates);
554 if (unlikely(r != 0)) {
555 if (r != -ERESTARTSYS)
556 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
557 goto error_free_pages;
558 }
559
560 /* Without a BO list we don't have userptr BOs */
561 if (!p->bo_list)
562 break;
563
564 INIT_LIST_HEAD(&need_pages);
565 for (i = p->bo_list->first_userptr;
566 i < p->bo_list->num_entries; ++i) {
567 struct amdgpu_bo *bo;
568
569 e = &p->bo_list->array[i];
570 bo = e->robj;
571
572 if (amdgpu_ttm_tt_userptr_invalidated(bo->tbo.ttm,
573 &e->user_invalidated) && e->user_pages) {
574
575 /* We acquired a page array, but somebody
576 * invalidated it. Free it and try again
577 */
578 release_pages(e->user_pages,
579 bo->tbo.ttm->num_pages);
580 kvfree(e->user_pages);
581 e->user_pages = NULL;
582 }
583
584 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm) &&
585 !e->user_pages) {
586 list_del(&e->tv.head);
587 list_add(&e->tv.head, &need_pages);
588
589 amdgpu_bo_unreserve(e->robj);
590 }
591 }
592
593 if (list_empty(&need_pages))
594 break;
595
596 /* Unreserve everything again. */
597 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
598
599 /* We tried too many times, just abort */
600 if (!--tries) {
601 r = -EDEADLK;
602 DRM_ERROR("deadlock in %s\n", __func__);
603 goto error_free_pages;
604 }
605
606 /* Fill the page arrays for all userptrs. */
607 list_for_each_entry(e, &need_pages, tv.head) {
608 struct ttm_tt *ttm = e->robj->tbo.ttm;
609
610 e->user_pages = kvmalloc_array(ttm->num_pages,
611 sizeof(struct page*),
612 GFP_KERNEL | __GFP_ZERO);
613 if (!e->user_pages) {
614 r = -ENOMEM;
615 DRM_ERROR("calloc failure in %s\n", __func__);
616 goto error_free_pages;
617 }
618
619 r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
620 if (r) {
621 DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
622 kvfree(e->user_pages);
623 e->user_pages = NULL;
624 goto error_free_pages;
625 }
626 }
627
628 /* And try again. */
629 list_splice(&need_pages, &p->validated);
630 }
631
632 amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
633 &p->bytes_moved_vis_threshold);
634 p->bytes_moved = 0;
635 p->bytes_moved_vis = 0;
636 p->evictable = list_last_entry(&p->validated,
637 struct amdgpu_bo_list_entry,
638 tv.head);
639
640 r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
641 amdgpu_cs_validate, p);
642 if (r) {
643 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
644 goto error_validate;
645 }
646
647 r = amdgpu_cs_list_validate(p, &duplicates);
648 if (r) {
649 DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
650 goto error_validate;
651 }
652
653 r = amdgpu_cs_list_validate(p, &p->validated);
654 if (r) {
655 DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
656 goto error_validate;
657 }
658
659 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
660 p->bytes_moved_vis);
661 if (p->bo_list) {
662 struct amdgpu_bo *gds = p->bo_list->gds_obj;
663 struct amdgpu_bo *gws = p->bo_list->gws_obj;
664 struct amdgpu_bo *oa = p->bo_list->oa_obj;
665 struct amdgpu_vm *vm = &fpriv->vm;
666 unsigned i;
667
668 for (i = 0; i < p->bo_list->num_entries; i++) {
669 struct amdgpu_bo *bo = p->bo_list->array[i].robj;
670
671 p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
672 }
673
674 if (gds) {
675 p->job->gds_base = amdgpu_bo_gpu_offset(gds);
676 p->job->gds_size = amdgpu_bo_size(gds);
677 }
678 if (gws) {
679 p->job->gws_base = amdgpu_bo_gpu_offset(gws);
680 p->job->gws_size = amdgpu_bo_size(gws);
681 }
682 if (oa) {
683 p->job->oa_base = amdgpu_bo_gpu_offset(oa);
684 p->job->oa_size = amdgpu_bo_size(oa);
685 }
686 }
687
688 if (!r && p->uf_entry.robj) {
689 struct amdgpu_bo *uf = p->uf_entry.robj;
690
691 r = amdgpu_ttm_alloc_gart(&uf->tbo);
692 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
693 }
694
695error_validate:
696 if (r)
697 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
698
699error_free_pages:
700
701 if (p->bo_list) {
702 for (i = p->bo_list->first_userptr;
703 i < p->bo_list->num_entries; ++i) {
704 e = &p->bo_list->array[i];
705
706 if (!e->user_pages)
707 continue;
708
709 release_pages(e->user_pages,
710 e->robj->tbo.ttm->num_pages);
711 kvfree(e->user_pages);
712 }
713 }
714
715 return r;
716}
717
718static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
719{
720 struct amdgpu_bo_list_entry *e;
721 int r;
722
723 list_for_each_entry(e, &p->validated, tv.head) {
724 struct reservation_object *resv = e->robj->tbo.resv;
725 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp,
726 amdgpu_bo_explicit_sync(e->robj));
727
728 if (r)
729 return r;
730 }
731 return 0;
732}
733
734/**
735 * cs_parser_fini() - clean parser states
736 * @parser: parser structure holding parsing context.
737 * @error: error number
738 *
739 * If error is set than unvalidate buffer, otherwise just free memory
740 * used by parsing context.
741 **/
742static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
743 bool backoff)
744{
745 unsigned i;
746
747 if (error && backoff)
748 ttm_eu_backoff_reservation(&parser->ticket,
749 &parser->validated);
750
751 for (i = 0; i < parser->num_post_dep_syncobjs; i++)
752 drm_syncobj_put(parser->post_dep_syncobjs[i]);
753 kfree(parser->post_dep_syncobjs);
754
755 dma_fence_put(parser->fence);
756
757 if (parser->ctx) {
758 mutex_unlock(&parser->ctx->lock);
759 amdgpu_ctx_put(parser->ctx);
760 }
761 if (parser->bo_list)
762 amdgpu_bo_list_put(parser->bo_list);
763
764 for (i = 0; i < parser->nchunks; i++)
765 kvfree(parser->chunks[i].kdata);
766 kfree(parser->chunks);
767 if (parser->job)
768 amdgpu_job_free(parser->job);
769 amdgpu_bo_unref(&parser->uf_entry.robj);
770}
771
772static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p)
773{
774 struct amdgpu_device *adev = p->adev;
775 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
776 struct amdgpu_vm *vm = &fpriv->vm;
777 struct amdgpu_bo_va *bo_va;
778 struct amdgpu_bo *bo;
779 int i, r;
780
781 r = amdgpu_vm_clear_freed(adev, vm, NULL);
782 if (r)
783 return r;
784
785 r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
786 if (r)
787 return r;
788
789 r = amdgpu_sync_fence(adev, &p->job->sync,
790 fpriv->prt_va->last_pt_update, false);
791 if (r)
792 return r;
793
794 if (amdgpu_sriov_vf(adev)) {
795 struct dma_fence *f;
796
797 bo_va = fpriv->csa_va;
798 BUG_ON(!bo_va);
799 r = amdgpu_vm_bo_update(adev, bo_va, false);
800 if (r)
801 return r;
802
803 f = bo_va->last_pt_update;
804 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
805 if (r)
806 return r;
807 }
808
809 if (p->bo_list) {
810 for (i = 0; i < p->bo_list->num_entries; i++) {
811 struct dma_fence *f;
812
813 /* ignore duplicates */
814 bo = p->bo_list->array[i].robj;
815 if (!bo)
816 continue;
817
818 bo_va = p->bo_list->array[i].bo_va;
819 if (bo_va == NULL)
820 continue;
821
822 r = amdgpu_vm_bo_update(adev, bo_va, false);
823 if (r)
824 return r;
825
826 f = bo_va->last_pt_update;
827 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
828 if (r)
829 return r;
830 }
831
832 }
833
834 r = amdgpu_vm_handle_moved(adev, vm);
835 if (r)
836 return r;
837
838 r = amdgpu_vm_update_directories(adev, vm);
839 if (r)
840 return r;
841
842 r = amdgpu_sync_fence(adev, &p->job->sync, vm->last_update, false);
843 if (r)
844 return r;
845
846 if (amdgpu_vm_debug && p->bo_list) {
847 /* Invalidate all BOs to test for userspace bugs */
848 for (i = 0; i < p->bo_list->num_entries; i++) {
849 /* ignore duplicates */
850 bo = p->bo_list->array[i].robj;
851 if (!bo)
852 continue;
853
854 amdgpu_vm_bo_invalidate(adev, bo, false);
855 }
856 }
857
858 return r;
859}
860
861static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
862 struct amdgpu_cs_parser *p)
863{
864 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
865 struct amdgpu_vm *vm = &fpriv->vm;
866 struct amdgpu_ring *ring = p->job->ring;
867 int r;
868
869 /* Only for UVD/VCE VM emulation */
870 if (p->job->ring->funcs->parse_cs) {
871 unsigned i, j;
872
873 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
874 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
875 struct amdgpu_bo_va_mapping *m;
876 struct amdgpu_bo *aobj = NULL;
877 struct amdgpu_cs_chunk *chunk;
878 uint64_t offset, va_start;
879 struct amdgpu_ib *ib;
880 uint8_t *kptr;
881
882 chunk = &p->chunks[i];
883 ib = &p->job->ibs[j];
884 chunk_ib = chunk->kdata;
885
886 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
887 continue;
888
889 va_start = chunk_ib->va_start & AMDGPU_VA_HOLE_MASK;
890 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
891 if (r) {
892 DRM_ERROR("IB va_start is invalid\n");
893 return r;
894 }
895
896 if ((va_start + chunk_ib->ib_bytes) >
897 (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
898 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
899 return -EINVAL;
900 }
901
902 /* the IB should be reserved at this point */
903 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
904 if (r) {
905 return r;
906 }
907
908 offset = m->start * AMDGPU_GPU_PAGE_SIZE;
909 kptr += va_start - offset;
910
911 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
912 amdgpu_bo_kunmap(aobj);
913
914 r = amdgpu_ring_parse_cs(ring, p, j);
915 if (r)
916 return r;
917
918 j++;
919 }
920 }
921
922 if (p->job->vm) {
923 p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->root.base.bo);
924
925 r = amdgpu_bo_vm_update_pte(p);
926 if (r)
927 return r;
928 }
929
930 return amdgpu_cs_sync_rings(p);
931}
932
933static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
934 struct amdgpu_cs_parser *parser)
935{
936 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
937 struct amdgpu_vm *vm = &fpriv->vm;
938 int i, j;
939 int r, ce_preempt = 0, de_preempt = 0;
940
941 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
942 struct amdgpu_cs_chunk *chunk;
943 struct amdgpu_ib *ib;
944 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
945 struct amdgpu_ring *ring;
946
947 chunk = &parser->chunks[i];
948 ib = &parser->job->ibs[j];
949 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
950
951 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
952 continue;
953
954 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX && amdgpu_sriov_vf(adev)) {
955 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
956 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
957 ce_preempt++;
958 else
959 de_preempt++;
960 }
961
962 /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
963 if (ce_preempt > 1 || de_preempt > 1)
964 return -EINVAL;
965 }
966
967 r = amdgpu_queue_mgr_map(adev, &parser->ctx->queue_mgr, chunk_ib->ip_type,
968 chunk_ib->ip_instance, chunk_ib->ring, &ring);
969 if (r)
970 return r;
971
972 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
973 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
974 if (!parser->ctx->preamble_presented) {
975 parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
976 parser->ctx->preamble_presented = true;
977 }
978 }
979
980 if (parser->job->ring && parser->job->ring != ring)
981 return -EINVAL;
982
983 parser->job->ring = ring;
984
985 r = amdgpu_ib_get(adev, vm,
986 ring->funcs->parse_cs ? chunk_ib->ib_bytes : 0,
987 ib);
988 if (r) {
989 DRM_ERROR("Failed to get ib !\n");
990 return r;
991 }
992
993 ib->gpu_addr = chunk_ib->va_start;
994 ib->length_dw = chunk_ib->ib_bytes / 4;
995 ib->flags = chunk_ib->flags;
996
997 j++;
998 }
999
1000 /* UVD & VCE fw doesn't support user fences */
1001 if (parser->job->uf_addr && (
1002 parser->job->ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
1003 parser->job->ring->funcs->type == AMDGPU_RING_TYPE_VCE))
1004 return -EINVAL;
1005
1006 return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->job->ring->idx);
1007}
1008
1009static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
1010 struct amdgpu_cs_chunk *chunk)
1011{
1012 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1013 unsigned num_deps;
1014 int i, r;
1015 struct drm_amdgpu_cs_chunk_dep *deps;
1016
1017 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
1018 num_deps = chunk->length_dw * 4 /
1019 sizeof(struct drm_amdgpu_cs_chunk_dep);
1020
1021 for (i = 0; i < num_deps; ++i) {
1022 struct amdgpu_ring *ring;
1023 struct amdgpu_ctx *ctx;
1024 struct dma_fence *fence;
1025
1026 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
1027 if (ctx == NULL)
1028 return -EINVAL;
1029
1030 r = amdgpu_queue_mgr_map(p->adev, &ctx->queue_mgr,
1031 deps[i].ip_type,
1032 deps[i].ip_instance,
1033 deps[i].ring, &ring);
1034 if (r) {
1035 amdgpu_ctx_put(ctx);
1036 return r;
1037 }
1038
1039 fence = amdgpu_ctx_get_fence(ctx, ring,
1040 deps[i].handle);
1041 if (IS_ERR(fence)) {
1042 r = PTR_ERR(fence);
1043 amdgpu_ctx_put(ctx);
1044 return r;
1045 } else if (fence) {
1046 r = amdgpu_sync_fence(p->adev, &p->job->sync, fence,
1047 true);
1048 dma_fence_put(fence);
1049 amdgpu_ctx_put(ctx);
1050 if (r)
1051 return r;
1052 }
1053 }
1054 return 0;
1055}
1056
1057static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1058 uint32_t handle)
1059{
1060 int r;
1061 struct dma_fence *fence;
1062 r = drm_syncobj_find_fence(p->filp, handle, &fence);
1063 if (r)
1064 return r;
1065
1066 r = amdgpu_sync_fence(p->adev, &p->job->sync, fence, true);
1067 dma_fence_put(fence);
1068
1069 return r;
1070}
1071
1072static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1073 struct amdgpu_cs_chunk *chunk)
1074{
1075 unsigned num_deps;
1076 int i, r;
1077 struct drm_amdgpu_cs_chunk_sem *deps;
1078
1079 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1080 num_deps = chunk->length_dw * 4 /
1081 sizeof(struct drm_amdgpu_cs_chunk_sem);
1082
1083 for (i = 0; i < num_deps; ++i) {
1084 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle);
1085 if (r)
1086 return r;
1087 }
1088 return 0;
1089}
1090
1091static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1092 struct amdgpu_cs_chunk *chunk)
1093{
1094 unsigned num_deps;
1095 int i;
1096 struct drm_amdgpu_cs_chunk_sem *deps;
1097 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1098 num_deps = chunk->length_dw * 4 /
1099 sizeof(struct drm_amdgpu_cs_chunk_sem);
1100
1101 p->post_dep_syncobjs = kmalloc_array(num_deps,
1102 sizeof(struct drm_syncobj *),
1103 GFP_KERNEL);
1104 p->num_post_dep_syncobjs = 0;
1105
1106 if (!p->post_dep_syncobjs)
1107 return -ENOMEM;
1108
1109 for (i = 0; i < num_deps; ++i) {
1110 p->post_dep_syncobjs[i] = drm_syncobj_find(p->filp, deps[i].handle);
1111 if (!p->post_dep_syncobjs[i])
1112 return -EINVAL;
1113 p->num_post_dep_syncobjs++;
1114 }
1115 return 0;
1116}
1117
1118static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1119 struct amdgpu_cs_parser *p)
1120{
1121 int i, r;
1122
1123 for (i = 0; i < p->nchunks; ++i) {
1124 struct amdgpu_cs_chunk *chunk;
1125
1126 chunk = &p->chunks[i];
1127
1128 if (chunk->chunk_id == AMDGPU_CHUNK_ID_DEPENDENCIES) {
1129 r = amdgpu_cs_process_fence_dep(p, chunk);
1130 if (r)
1131 return r;
1132 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_IN) {
1133 r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1134 if (r)
1135 return r;
1136 } else if (chunk->chunk_id == AMDGPU_CHUNK_ID_SYNCOBJ_OUT) {
1137 r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1138 if (r)
1139 return r;
1140 }
1141 }
1142
1143 return 0;
1144}
1145
1146static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1147{
1148 int i;
1149
1150 for (i = 0; i < p->num_post_dep_syncobjs; ++i)
1151 drm_syncobj_replace_fence(p->post_dep_syncobjs[i], p->fence);
1152}
1153
1154static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1155 union drm_amdgpu_cs *cs)
1156{
1157 struct amdgpu_ring *ring = p->job->ring;
1158 struct drm_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1159 struct amdgpu_job *job;
1160 unsigned i;
1161 uint64_t seq;
1162
1163 int r;
1164
1165 amdgpu_mn_lock(p->mn);
1166 if (p->bo_list) {
1167 for (i = p->bo_list->first_userptr;
1168 i < p->bo_list->num_entries; ++i) {
1169 struct amdgpu_bo *bo = p->bo_list->array[i].robj;
1170
1171 if (amdgpu_ttm_tt_userptr_needs_pages(bo->tbo.ttm)) {
1172 amdgpu_mn_unlock(p->mn);
1173 return -ERESTARTSYS;
1174 }
1175 }
1176 }
1177
1178 job = p->job;
1179 p->job = NULL;
1180
1181 r = drm_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1182 if (r) {
1183 amdgpu_job_free(job);
1184 amdgpu_mn_unlock(p->mn);
1185 return r;
1186 }
1187
1188 job->owner = p->filp;
1189 job->fence_ctx = entity->fence_context;
1190 p->fence = dma_fence_get(&job->base.s_fence->finished);
1191
1192 r = amdgpu_ctx_add_fence(p->ctx, ring, p->fence, &seq);
1193 if (r) {
1194 dma_fence_put(p->fence);
1195 dma_fence_put(&job->base.s_fence->finished);
1196 amdgpu_job_free(job);
1197 amdgpu_mn_unlock(p->mn);
1198 return r;
1199 }
1200
1201 amdgpu_cs_post_dependencies(p);
1202
1203 cs->out.handle = seq;
1204 job->uf_sequence = seq;
1205
1206 amdgpu_job_free_resources(job);
1207 amdgpu_ring_priority_get(job->ring, job->base.s_priority);
1208
1209 trace_amdgpu_cs_ioctl(job);
1210 drm_sched_entity_push_job(&job->base, entity);
1211
1212 ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1213 amdgpu_mn_unlock(p->mn);
1214
1215 return 0;
1216}
1217
1218int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1219{
1220 struct amdgpu_device *adev = dev->dev_private;
1221 union drm_amdgpu_cs *cs = data;
1222 struct amdgpu_cs_parser parser = {};
1223 bool reserved_buffers = false;
1224 int i, r;
1225
1226 if (!adev->accel_working)
1227 return -EBUSY;
1228
1229 parser.adev = adev;
1230 parser.filp = filp;
1231
1232 r = amdgpu_cs_parser_init(&parser, data);
1233 if (r) {
1234 DRM_ERROR("Failed to initialize parser !\n");
1235 goto out;
1236 }
1237
1238 r = amdgpu_cs_ib_fill(adev, &parser);
1239 if (r)
1240 goto out;
1241
1242 r = amdgpu_cs_parser_bos(&parser, data);
1243 if (r) {
1244 if (r == -ENOMEM)
1245 DRM_ERROR("Not enough memory for command submission!\n");
1246 else if (r != -ERESTARTSYS)
1247 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1248 goto out;
1249 }
1250
1251 reserved_buffers = true;
1252
1253 r = amdgpu_cs_dependencies(adev, &parser);
1254 if (r) {
1255 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1256 goto out;
1257 }
1258
1259 for (i = 0; i < parser.job->num_ibs; i++)
1260 trace_amdgpu_cs(&parser, i);
1261
1262 r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1263 if (r)
1264 goto out;
1265
1266 r = amdgpu_cs_submit(&parser, cs);
1267
1268out:
1269 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1270 return r;
1271}
1272
1273/**
1274 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1275 *
1276 * @dev: drm device
1277 * @data: data from userspace
1278 * @filp: file private
1279 *
1280 * Wait for the command submission identified by handle to finish.
1281 */
1282int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1283 struct drm_file *filp)
1284{
1285 union drm_amdgpu_wait_cs *wait = data;
1286 struct amdgpu_device *adev = dev->dev_private;
1287 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1288 struct amdgpu_ring *ring = NULL;
1289 struct amdgpu_ctx *ctx;
1290 struct dma_fence *fence;
1291 long r;
1292
1293 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1294 if (ctx == NULL)
1295 return -EINVAL;
1296
1297 r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr,
1298 wait->in.ip_type, wait->in.ip_instance,
1299 wait->in.ring, &ring);
1300 if (r) {
1301 amdgpu_ctx_put(ctx);
1302 return r;
1303 }
1304
1305 fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
1306 if (IS_ERR(fence))
1307 r = PTR_ERR(fence);
1308 else if (fence) {
1309 r = dma_fence_wait_timeout(fence, true, timeout);
1310 if (r > 0 && fence->error)
1311 r = fence->error;
1312 dma_fence_put(fence);
1313 } else
1314 r = 1;
1315
1316 amdgpu_ctx_put(ctx);
1317 if (r < 0)
1318 return r;
1319
1320 memset(wait, 0, sizeof(*wait));
1321 wait->out.status = (r == 0);
1322
1323 return 0;
1324}
1325
1326/**
1327 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1328 *
1329 * @adev: amdgpu device
1330 * @filp: file private
1331 * @user: drm_amdgpu_fence copied from user space
1332 */
1333static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1334 struct drm_file *filp,
1335 struct drm_amdgpu_fence *user)
1336{
1337 struct amdgpu_ring *ring;
1338 struct amdgpu_ctx *ctx;
1339 struct dma_fence *fence;
1340 int r;
1341
1342 ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1343 if (ctx == NULL)
1344 return ERR_PTR(-EINVAL);
1345
1346 r = amdgpu_queue_mgr_map(adev, &ctx->queue_mgr, user->ip_type,
1347 user->ip_instance, user->ring, &ring);
1348 if (r) {
1349 amdgpu_ctx_put(ctx);
1350 return ERR_PTR(r);
1351 }
1352
1353 fence = amdgpu_ctx_get_fence(ctx, ring, user->seq_no);
1354 amdgpu_ctx_put(ctx);
1355
1356 return fence;
1357}
1358
1359int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1360 struct drm_file *filp)
1361{
1362 struct amdgpu_device *adev = dev->dev_private;
1363 union drm_amdgpu_fence_to_handle *info = data;
1364 struct dma_fence *fence;
1365 struct drm_syncobj *syncobj;
1366 struct sync_file *sync_file;
1367 int fd, r;
1368
1369 fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1370 if (IS_ERR(fence))
1371 return PTR_ERR(fence);
1372
1373 switch (info->in.what) {
1374 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1375 r = drm_syncobj_create(&syncobj, 0, fence);
1376 dma_fence_put(fence);
1377 if (r)
1378 return r;
1379 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1380 drm_syncobj_put(syncobj);
1381 return r;
1382
1383 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1384 r = drm_syncobj_create(&syncobj, 0, fence);
1385 dma_fence_put(fence);
1386 if (r)
1387 return r;
1388 r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1389 drm_syncobj_put(syncobj);
1390 return r;
1391
1392 case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1393 fd = get_unused_fd_flags(O_CLOEXEC);
1394 if (fd < 0) {
1395 dma_fence_put(fence);
1396 return fd;
1397 }
1398
1399 sync_file = sync_file_create(fence);
1400 dma_fence_put(fence);
1401 if (!sync_file) {
1402 put_unused_fd(fd);
1403 return -ENOMEM;
1404 }
1405
1406 fd_install(fd, sync_file->file);
1407 info->out.handle = fd;
1408 return 0;
1409
1410 default:
1411 return -EINVAL;
1412 }
1413}
1414
1415/**
1416 * amdgpu_cs_wait_all_fence - wait on all fences to signal
1417 *
1418 * @adev: amdgpu device
1419 * @filp: file private
1420 * @wait: wait parameters
1421 * @fences: array of drm_amdgpu_fence
1422 */
1423static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1424 struct drm_file *filp,
1425 union drm_amdgpu_wait_fences *wait,
1426 struct drm_amdgpu_fence *fences)
1427{
1428 uint32_t fence_count = wait->in.fence_count;
1429 unsigned int i;
1430 long r = 1;
1431
1432 for (i = 0; i < fence_count; i++) {
1433 struct dma_fence *fence;
1434 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1435
1436 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1437 if (IS_ERR(fence))
1438 return PTR_ERR(fence);
1439 else if (!fence)
1440 continue;
1441
1442 r = dma_fence_wait_timeout(fence, true, timeout);
1443 dma_fence_put(fence);
1444 if (r < 0)
1445 return r;
1446
1447 if (r == 0)
1448 break;
1449
1450 if (fence->error)
1451 return fence->error;
1452 }
1453
1454 memset(wait, 0, sizeof(*wait));
1455 wait->out.status = (r > 0);
1456
1457 return 0;
1458}
1459
1460/**
1461 * amdgpu_cs_wait_any_fence - wait on any fence to signal
1462 *
1463 * @adev: amdgpu device
1464 * @filp: file private
1465 * @wait: wait parameters
1466 * @fences: array of drm_amdgpu_fence
1467 */
1468static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1469 struct drm_file *filp,
1470 union drm_amdgpu_wait_fences *wait,
1471 struct drm_amdgpu_fence *fences)
1472{
1473 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1474 uint32_t fence_count = wait->in.fence_count;
1475 uint32_t first = ~0;
1476 struct dma_fence **array;
1477 unsigned int i;
1478 long r;
1479
1480 /* Prepare the fence array */
1481 array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1482
1483 if (array == NULL)
1484 return -ENOMEM;
1485
1486 for (i = 0; i < fence_count; i++) {
1487 struct dma_fence *fence;
1488
1489 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1490 if (IS_ERR(fence)) {
1491 r = PTR_ERR(fence);
1492 goto err_free_fence_array;
1493 } else if (fence) {
1494 array[i] = fence;
1495 } else { /* NULL, the fence has been already signaled */
1496 r = 1;
1497 first = i;
1498 goto out;
1499 }
1500 }
1501
1502 r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1503 &first);
1504 if (r < 0)
1505 goto err_free_fence_array;
1506
1507out:
1508 memset(wait, 0, sizeof(*wait));
1509 wait->out.status = (r > 0);
1510 wait->out.first_signaled = first;
1511
1512 if (first < fence_count && array[first])
1513 r = array[first]->error;
1514 else
1515 r = 0;
1516
1517err_free_fence_array:
1518 for (i = 0; i < fence_count; i++)
1519 dma_fence_put(array[i]);
1520 kfree(array);
1521
1522 return r;
1523}
1524
1525/**
1526 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1527 *
1528 * @dev: drm device
1529 * @data: data from userspace
1530 * @filp: file private
1531 */
1532int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1533 struct drm_file *filp)
1534{
1535 struct amdgpu_device *adev = dev->dev_private;
1536 union drm_amdgpu_wait_fences *wait = data;
1537 uint32_t fence_count = wait->in.fence_count;
1538 struct drm_amdgpu_fence *fences_user;
1539 struct drm_amdgpu_fence *fences;
1540 int r;
1541
1542 /* Get the fences from userspace */
1543 fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1544 GFP_KERNEL);
1545 if (fences == NULL)
1546 return -ENOMEM;
1547
1548 fences_user = u64_to_user_ptr(wait->in.fences);
1549 if (copy_from_user(fences, fences_user,
1550 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1551 r = -EFAULT;
1552 goto err_free_fences;
1553 }
1554
1555 if (wait->in.wait_all)
1556 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1557 else
1558 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1559
1560err_free_fences:
1561 kfree(fences);
1562
1563 return r;
1564}
1565
1566/**
1567 * amdgpu_cs_find_bo_va - find bo_va for VM address
1568 *
1569 * @parser: command submission parser context
1570 * @addr: VM address
1571 * @bo: resulting BO of the mapping found
1572 *
1573 * Search the buffer objects in the command submission context for a certain
1574 * virtual memory address. Returns allocation structure when found, NULL
1575 * otherwise.
1576 */
1577int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1578 uint64_t addr, struct amdgpu_bo **bo,
1579 struct amdgpu_bo_va_mapping **map)
1580{
1581 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1582 struct ttm_operation_ctx ctx = { false, false };
1583 struct amdgpu_vm *vm = &fpriv->vm;
1584 struct amdgpu_bo_va_mapping *mapping;
1585 int r;
1586
1587 addr /= AMDGPU_GPU_PAGE_SIZE;
1588
1589 mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1590 if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1591 return -EINVAL;
1592
1593 *bo = mapping->bo_va->base.bo;
1594 *map = mapping;
1595
1596 /* Double check that the BO is reserved by this CS */
1597 if (READ_ONCE((*bo)->tbo.resv->lock.ctx) != &parser->ticket)
1598 return -EINVAL;
1599
1600 if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1601 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1602 amdgpu_ttm_placement_from_domain(*bo, (*bo)->allowed_domains);
1603 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1604 if (r)
1605 return r;
1606 }
1607
1608 return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1609}
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
32#include <drm/amdgpu_drm.h>
33#include <drm/drm_syncobj.h>
34#include "amdgpu.h"
35#include "amdgpu_trace.h"
36#include "amdgpu_gmc.h"
37#include "amdgpu_gem.h"
38
39static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
40 struct drm_amdgpu_cs_chunk_fence *data,
41 uint32_t *offset)
42{
43 struct drm_gem_object *gobj;
44 struct amdgpu_bo *bo;
45 unsigned long size;
46 int r;
47
48 gobj = drm_gem_object_lookup(p->filp, data->handle);
49 if (gobj == NULL)
50 return -EINVAL;
51
52 bo = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
53 p->uf_entry.priority = 0;
54 p->uf_entry.tv.bo = &bo->tbo;
55 /* One for TTM and one for the CS job */
56 p->uf_entry.tv.num_shared = 2;
57
58 drm_gem_object_put_unlocked(gobj);
59
60 size = amdgpu_bo_size(bo);
61 if (size != PAGE_SIZE || (data->offset + 8) > size) {
62 r = -EINVAL;
63 goto error_unref;
64 }
65
66 if (amdgpu_ttm_tt_get_usermm(bo->tbo.ttm)) {
67 r = -EINVAL;
68 goto error_unref;
69 }
70
71 *offset = data->offset;
72
73 return 0;
74
75error_unref:
76 amdgpu_bo_unref(&bo);
77 return r;
78}
79
80static int amdgpu_cs_bo_handles_chunk(struct amdgpu_cs_parser *p,
81 struct drm_amdgpu_bo_list_in *data)
82{
83 int r;
84 struct drm_amdgpu_bo_list_entry *info = NULL;
85
86 r = amdgpu_bo_create_list_entry_array(data, &info);
87 if (r)
88 return r;
89
90 r = amdgpu_bo_list_create(p->adev, p->filp, info, data->bo_number,
91 &p->bo_list);
92 if (r)
93 goto error_free;
94
95 kvfree(info);
96 return 0;
97
98error_free:
99 if (info)
100 kvfree(info);
101
102 return r;
103}
104
105static int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, union drm_amdgpu_cs *cs)
106{
107 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
108 struct amdgpu_vm *vm = &fpriv->vm;
109 uint64_t *chunk_array_user;
110 uint64_t *chunk_array;
111 unsigned size, num_ibs = 0;
112 uint32_t uf_offset = 0;
113 int i;
114 int ret;
115
116 if (cs->in.num_chunks == 0)
117 return 0;
118
119 chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
120 if (!chunk_array)
121 return -ENOMEM;
122
123 p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
124 if (!p->ctx) {
125 ret = -EINVAL;
126 goto free_chunk;
127 }
128
129 mutex_lock(&p->ctx->lock);
130
131 /* skip guilty context job */
132 if (atomic_read(&p->ctx->guilty) == 1) {
133 ret = -ECANCELED;
134 goto free_chunk;
135 }
136
137 /* get chunks */
138 chunk_array_user = u64_to_user_ptr(cs->in.chunks);
139 if (copy_from_user(chunk_array, chunk_array_user,
140 sizeof(uint64_t)*cs->in.num_chunks)) {
141 ret = -EFAULT;
142 goto free_chunk;
143 }
144
145 p->nchunks = cs->in.num_chunks;
146 p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
147 GFP_KERNEL);
148 if (!p->chunks) {
149 ret = -ENOMEM;
150 goto free_chunk;
151 }
152
153 for (i = 0; i < p->nchunks; i++) {
154 struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
155 struct drm_amdgpu_cs_chunk user_chunk;
156 uint32_t __user *cdata;
157
158 chunk_ptr = u64_to_user_ptr(chunk_array[i]);
159 if (copy_from_user(&user_chunk, chunk_ptr,
160 sizeof(struct drm_amdgpu_cs_chunk))) {
161 ret = -EFAULT;
162 i--;
163 goto free_partial_kdata;
164 }
165 p->chunks[i].chunk_id = user_chunk.chunk_id;
166 p->chunks[i].length_dw = user_chunk.length_dw;
167
168 size = p->chunks[i].length_dw;
169 cdata = u64_to_user_ptr(user_chunk.chunk_data);
170
171 p->chunks[i].kdata = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
172 if (p->chunks[i].kdata == NULL) {
173 ret = -ENOMEM;
174 i--;
175 goto free_partial_kdata;
176 }
177 size *= sizeof(uint32_t);
178 if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
179 ret = -EFAULT;
180 goto free_partial_kdata;
181 }
182
183 switch (p->chunks[i].chunk_id) {
184 case AMDGPU_CHUNK_ID_IB:
185 ++num_ibs;
186 break;
187
188 case AMDGPU_CHUNK_ID_FENCE:
189 size = sizeof(struct drm_amdgpu_cs_chunk_fence);
190 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
191 ret = -EINVAL;
192 goto free_partial_kdata;
193 }
194
195 ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
196 &uf_offset);
197 if (ret)
198 goto free_partial_kdata;
199
200 break;
201
202 case AMDGPU_CHUNK_ID_BO_HANDLES:
203 size = sizeof(struct drm_amdgpu_bo_list_in);
204 if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
205 ret = -EINVAL;
206 goto free_partial_kdata;
207 }
208
209 ret = amdgpu_cs_bo_handles_chunk(p, p->chunks[i].kdata);
210 if (ret)
211 goto free_partial_kdata;
212
213 break;
214
215 case AMDGPU_CHUNK_ID_DEPENDENCIES:
216 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
217 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
218 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
219 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
220 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
221 break;
222
223 default:
224 ret = -EINVAL;
225 goto free_partial_kdata;
226 }
227 }
228
229 ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
230 if (ret)
231 goto free_all_kdata;
232
233 if (p->ctx->vram_lost_counter != p->job->vram_lost_counter) {
234 ret = -ECANCELED;
235 goto free_all_kdata;
236 }
237
238 if (p->uf_entry.tv.bo)
239 p->job->uf_addr = uf_offset;
240 kfree(chunk_array);
241
242 /* Use this opportunity to fill in task info for the vm */
243 amdgpu_vm_set_task_info(vm);
244
245 return 0;
246
247free_all_kdata:
248 i = p->nchunks - 1;
249free_partial_kdata:
250 for (; i >= 0; i--)
251 kvfree(p->chunks[i].kdata);
252 kfree(p->chunks);
253 p->chunks = NULL;
254 p->nchunks = 0;
255free_chunk:
256 kfree(chunk_array);
257
258 return ret;
259}
260
261/* Convert microseconds to bytes. */
262static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
263{
264 if (us <= 0 || !adev->mm_stats.log2_max_MBps)
265 return 0;
266
267 /* Since accum_us is incremented by a million per second, just
268 * multiply it by the number of MB/s to get the number of bytes.
269 */
270 return us << adev->mm_stats.log2_max_MBps;
271}
272
273static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
274{
275 if (!adev->mm_stats.log2_max_MBps)
276 return 0;
277
278 return bytes >> adev->mm_stats.log2_max_MBps;
279}
280
281/* Returns how many bytes TTM can move right now. If no bytes can be moved,
282 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
283 * which means it can go over the threshold once. If that happens, the driver
284 * will be in debt and no other buffer migrations can be done until that debt
285 * is repaid.
286 *
287 * This approach allows moving a buffer of any size (it's important to allow
288 * that).
289 *
290 * The currency is simply time in microseconds and it increases as the clock
291 * ticks. The accumulated microseconds (us) are converted to bytes and
292 * returned.
293 */
294static void amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev,
295 u64 *max_bytes,
296 u64 *max_vis_bytes)
297{
298 s64 time_us, increment_us;
299 u64 free_vram, total_vram, used_vram;
300
301 /* Allow a maximum of 200 accumulated ms. This is basically per-IB
302 * throttling.
303 *
304 * It means that in order to get full max MBps, at least 5 IBs per
305 * second must be submitted and not more than 200ms apart from each
306 * other.
307 */
308 const s64 us_upper_bound = 200000;
309
310 if (!adev->mm_stats.log2_max_MBps) {
311 *max_bytes = 0;
312 *max_vis_bytes = 0;
313 return;
314 }
315
316 total_vram = adev->gmc.real_vram_size - atomic64_read(&adev->vram_pin_size);
317 used_vram = amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
318 free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;
319
320 spin_lock(&adev->mm_stats.lock);
321
322 /* Increase the amount of accumulated us. */
323 time_us = ktime_to_us(ktime_get());
324 increment_us = time_us - adev->mm_stats.last_update_us;
325 adev->mm_stats.last_update_us = time_us;
326 adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
327 us_upper_bound);
328
329 /* This prevents the short period of low performance when the VRAM
330 * usage is low and the driver is in debt or doesn't have enough
331 * accumulated us to fill VRAM quickly.
332 *
333 * The situation can occur in these cases:
334 * - a lot of VRAM is freed by userspace
335 * - the presence of a big buffer causes a lot of evictions
336 * (solution: split buffers into smaller ones)
337 *
338 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
339 * accum_us to a positive number.
340 */
341 if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
342 s64 min_us;
343
344 /* Be more aggresive on dGPUs. Try to fill a portion of free
345 * VRAM now.
346 */
347 if (!(adev->flags & AMD_IS_APU))
348 min_us = bytes_to_us(adev, free_vram / 4);
349 else
350 min_us = 0; /* Reset accum_us on APUs. */
351
352 adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
353 }
354
355 /* This is set to 0 if the driver is in debt to disallow (optional)
356 * buffer moves.
357 */
358 *max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);
359
360 /* Do the same for visible VRAM if half of it is free */
361 if (!amdgpu_gmc_vram_full_visible(&adev->gmc)) {
362 u64 total_vis_vram = adev->gmc.visible_vram_size;
363 u64 used_vis_vram =
364 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM]);
365
366 if (used_vis_vram < total_vis_vram) {
367 u64 free_vis_vram = total_vis_vram - used_vis_vram;
368 adev->mm_stats.accum_us_vis = min(adev->mm_stats.accum_us_vis +
369 increment_us, us_upper_bound);
370
371 if (free_vis_vram >= total_vis_vram / 2)
372 adev->mm_stats.accum_us_vis =
373 max(bytes_to_us(adev, free_vis_vram / 2),
374 adev->mm_stats.accum_us_vis);
375 }
376
377 *max_vis_bytes = us_to_bytes(adev, adev->mm_stats.accum_us_vis);
378 } else {
379 *max_vis_bytes = 0;
380 }
381
382 spin_unlock(&adev->mm_stats.lock);
383}
384
385/* Report how many bytes have really been moved for the last command
386 * submission. This can result in a debt that can stop buffer migrations
387 * temporarily.
388 */
389void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes,
390 u64 num_vis_bytes)
391{
392 spin_lock(&adev->mm_stats.lock);
393 adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
394 adev->mm_stats.accum_us_vis -= bytes_to_us(adev, num_vis_bytes);
395 spin_unlock(&adev->mm_stats.lock);
396}
397
398static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
399 struct amdgpu_bo *bo)
400{
401 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
402 struct ttm_operation_ctx ctx = {
403 .interruptible = true,
404 .no_wait_gpu = false,
405 .resv = bo->tbo.base.resv,
406 .flags = 0
407 };
408 uint32_t domain;
409 int r;
410
411 if (bo->pin_count)
412 return 0;
413
414 /* Don't move this buffer if we have depleted our allowance
415 * to move it. Don't move anything if the threshold is zero.
416 */
417 if (p->bytes_moved < p->bytes_moved_threshold) {
418 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
419 (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)) {
420 /* And don't move a CPU_ACCESS_REQUIRED BO to limited
421 * visible VRAM if we've depleted our allowance to do
422 * that.
423 */
424 if (p->bytes_moved_vis < p->bytes_moved_vis_threshold)
425 domain = bo->preferred_domains;
426 else
427 domain = bo->allowed_domains;
428 } else {
429 domain = bo->preferred_domains;
430 }
431 } else {
432 domain = bo->allowed_domains;
433 }
434
435retry:
436 amdgpu_bo_placement_from_domain(bo, domain);
437 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
438
439 p->bytes_moved += ctx.bytes_moved;
440 if (!amdgpu_gmc_vram_full_visible(&adev->gmc) &&
441 amdgpu_bo_in_cpu_visible_vram(bo))
442 p->bytes_moved_vis += ctx.bytes_moved;
443
444 if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
445 domain = bo->allowed_domains;
446 goto retry;
447 }
448
449 return r;
450}
451
452/* Last resort, try to evict something from the current working set */
453static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
454 struct amdgpu_bo *validated)
455{
456 uint32_t domain = validated->allowed_domains;
457 struct ttm_operation_ctx ctx = { true, false };
458 int r;
459
460 if (!p->evictable)
461 return false;
462
463 for (;&p->evictable->tv.head != &p->validated;
464 p->evictable = list_prev_entry(p->evictable, tv.head)) {
465
466 struct amdgpu_bo_list_entry *candidate = p->evictable;
467 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(candidate->tv.bo);
468 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
469 bool update_bytes_moved_vis;
470 uint32_t other;
471
472 /* If we reached our current BO we can forget it */
473 if (bo == validated)
474 break;
475
476 /* We can't move pinned BOs here */
477 if (bo->pin_count)
478 continue;
479
480 other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
481
482 /* Check if this BO is in one of the domains we need space for */
483 if (!(other & domain))
484 continue;
485
486 /* Check if we can move this BO somewhere else */
487 other = bo->allowed_domains & ~domain;
488 if (!other)
489 continue;
490
491 /* Good we can try to move this BO somewhere else */
492 update_bytes_moved_vis =
493 !amdgpu_gmc_vram_full_visible(&adev->gmc) &&
494 amdgpu_bo_in_cpu_visible_vram(bo);
495 amdgpu_bo_placement_from_domain(bo, other);
496 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
497 p->bytes_moved += ctx.bytes_moved;
498 if (update_bytes_moved_vis)
499 p->bytes_moved_vis += ctx.bytes_moved;
500
501 if (unlikely(r))
502 break;
503
504 p->evictable = list_prev_entry(p->evictable, tv.head);
505 list_move(&candidate->tv.head, &p->validated);
506
507 return true;
508 }
509
510 return false;
511}
512
513static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
514{
515 struct amdgpu_cs_parser *p = param;
516 int r;
517
518 do {
519 r = amdgpu_cs_bo_validate(p, bo);
520 } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
521 if (r)
522 return r;
523
524 if (bo->shadow)
525 r = amdgpu_cs_bo_validate(p, bo->shadow);
526
527 return r;
528}
529
530static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
531 struct list_head *validated)
532{
533 struct ttm_operation_ctx ctx = { true, false };
534 struct amdgpu_bo_list_entry *lobj;
535 int r;
536
537 list_for_each_entry(lobj, validated, tv.head) {
538 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(lobj->tv.bo);
539 struct mm_struct *usermm;
540
541 usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
542 if (usermm && usermm != current->mm)
543 return -EPERM;
544
545 if (amdgpu_ttm_tt_is_userptr(bo->tbo.ttm) &&
546 lobj->user_invalidated && lobj->user_pages) {
547 amdgpu_bo_placement_from_domain(bo,
548 AMDGPU_GEM_DOMAIN_CPU);
549 r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
550 if (r)
551 return r;
552
553 amdgpu_ttm_tt_set_user_pages(bo->tbo.ttm,
554 lobj->user_pages);
555 }
556
557 if (p->evictable == lobj)
558 p->evictable = NULL;
559
560 r = amdgpu_cs_validate(p, bo);
561 if (r)
562 return r;
563
564 kvfree(lobj->user_pages);
565 lobj->user_pages = NULL;
566 }
567 return 0;
568}
569
570static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
571 union drm_amdgpu_cs *cs)
572{
573 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
574 struct amdgpu_vm *vm = &fpriv->vm;
575 struct amdgpu_bo_list_entry *e;
576 struct list_head duplicates;
577 struct amdgpu_bo *gds;
578 struct amdgpu_bo *gws;
579 struct amdgpu_bo *oa;
580 int r;
581
582 INIT_LIST_HEAD(&p->validated);
583
584 /* p->bo_list could already be assigned if AMDGPU_CHUNK_ID_BO_HANDLES is present */
585 if (cs->in.bo_list_handle) {
586 if (p->bo_list)
587 return -EINVAL;
588
589 r = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle,
590 &p->bo_list);
591 if (r)
592 return r;
593 } else if (!p->bo_list) {
594 /* Create a empty bo_list when no handle is provided */
595 r = amdgpu_bo_list_create(p->adev, p->filp, NULL, 0,
596 &p->bo_list);
597 if (r)
598 return r;
599 }
600
601 /* One for TTM and one for the CS job */
602 amdgpu_bo_list_for_each_entry(e, p->bo_list)
603 e->tv.num_shared = 2;
604
605 amdgpu_bo_list_get_list(p->bo_list, &p->validated);
606 if (p->bo_list->first_userptr != p->bo_list->num_entries)
607 p->mn = amdgpu_mn_get(p->adev, AMDGPU_MN_TYPE_GFX);
608
609 INIT_LIST_HEAD(&duplicates);
610 amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
611
612 if (p->uf_entry.tv.bo && !ttm_to_amdgpu_bo(p->uf_entry.tv.bo)->parent)
613 list_add(&p->uf_entry.tv.head, &p->validated);
614
615 /* Get userptr backing pages. If pages are updated after registered
616 * in amdgpu_gem_userptr_ioctl(), amdgpu_cs_list_validate() will do
617 * amdgpu_ttm_backend_bind() to flush and invalidate new pages
618 */
619 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
620 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
621 bool userpage_invalidated = false;
622 int i;
623
624 e->user_pages = kvmalloc_array(bo->tbo.ttm->num_pages,
625 sizeof(struct page *),
626 GFP_KERNEL | __GFP_ZERO);
627 if (!e->user_pages) {
628 DRM_ERROR("calloc failure\n");
629 return -ENOMEM;
630 }
631
632 r = amdgpu_ttm_tt_get_user_pages(bo, e->user_pages);
633 if (r) {
634 kvfree(e->user_pages);
635 e->user_pages = NULL;
636 return r;
637 }
638
639 for (i = 0; i < bo->tbo.ttm->num_pages; i++) {
640 if (bo->tbo.ttm->pages[i] != e->user_pages[i]) {
641 userpage_invalidated = true;
642 break;
643 }
644 }
645 e->user_invalidated = userpage_invalidated;
646 }
647
648 r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
649 &duplicates, false);
650 if (unlikely(r != 0)) {
651 if (r != -ERESTARTSYS)
652 DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
653 goto out;
654 }
655
656 amdgpu_cs_get_threshold_for_moves(p->adev, &p->bytes_moved_threshold,
657 &p->bytes_moved_vis_threshold);
658 p->bytes_moved = 0;
659 p->bytes_moved_vis = 0;
660 p->evictable = list_last_entry(&p->validated,
661 struct amdgpu_bo_list_entry,
662 tv.head);
663
664 r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
665 amdgpu_cs_validate, p);
666 if (r) {
667 DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
668 goto error_validate;
669 }
670
671 r = amdgpu_cs_list_validate(p, &duplicates);
672 if (r)
673 goto error_validate;
674
675 r = amdgpu_cs_list_validate(p, &p->validated);
676 if (r)
677 goto error_validate;
678
679 amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved,
680 p->bytes_moved_vis);
681
682 gds = p->bo_list->gds_obj;
683 gws = p->bo_list->gws_obj;
684 oa = p->bo_list->oa_obj;
685
686 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
687 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
688
689 /* Make sure we use the exclusive slot for shared BOs */
690 if (bo->prime_shared_count)
691 e->tv.num_shared = 0;
692 e->bo_va = amdgpu_vm_bo_find(vm, bo);
693 }
694
695 if (gds) {
696 p->job->gds_base = amdgpu_bo_gpu_offset(gds) >> PAGE_SHIFT;
697 p->job->gds_size = amdgpu_bo_size(gds) >> PAGE_SHIFT;
698 }
699 if (gws) {
700 p->job->gws_base = amdgpu_bo_gpu_offset(gws) >> PAGE_SHIFT;
701 p->job->gws_size = amdgpu_bo_size(gws) >> PAGE_SHIFT;
702 }
703 if (oa) {
704 p->job->oa_base = amdgpu_bo_gpu_offset(oa) >> PAGE_SHIFT;
705 p->job->oa_size = amdgpu_bo_size(oa) >> PAGE_SHIFT;
706 }
707
708 if (!r && p->uf_entry.tv.bo) {
709 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(p->uf_entry.tv.bo);
710
711 r = amdgpu_ttm_alloc_gart(&uf->tbo);
712 p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
713 }
714
715error_validate:
716 if (r)
717 ttm_eu_backoff_reservation(&p->ticket, &p->validated);
718out:
719 return r;
720}
721
722static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
723{
724 struct amdgpu_bo_list_entry *e;
725 int r;
726
727 list_for_each_entry(e, &p->validated, tv.head) {
728 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
729 struct dma_resv *resv = bo->tbo.base.resv;
730
731 r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp,
732 amdgpu_bo_explicit_sync(bo));
733
734 if (r)
735 return r;
736 }
737 return 0;
738}
739
740/**
741 * cs_parser_fini() - clean parser states
742 * @parser: parser structure holding parsing context.
743 * @error: error number
744 *
745 * If error is set than unvalidate buffer, otherwise just free memory
746 * used by parsing context.
747 **/
748static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error,
749 bool backoff)
750{
751 unsigned i;
752
753 if (error && backoff)
754 ttm_eu_backoff_reservation(&parser->ticket,
755 &parser->validated);
756
757 for (i = 0; i < parser->num_post_deps; i++) {
758 drm_syncobj_put(parser->post_deps[i].syncobj);
759 kfree(parser->post_deps[i].chain);
760 }
761 kfree(parser->post_deps);
762
763 dma_fence_put(parser->fence);
764
765 if (parser->ctx) {
766 mutex_unlock(&parser->ctx->lock);
767 amdgpu_ctx_put(parser->ctx);
768 }
769 if (parser->bo_list)
770 amdgpu_bo_list_put(parser->bo_list);
771
772 for (i = 0; i < parser->nchunks; i++)
773 kvfree(parser->chunks[i].kdata);
774 kfree(parser->chunks);
775 if (parser->job)
776 amdgpu_job_free(parser->job);
777 if (parser->uf_entry.tv.bo) {
778 struct amdgpu_bo *uf = ttm_to_amdgpu_bo(parser->uf_entry.tv.bo);
779
780 amdgpu_bo_unref(&uf);
781 }
782}
783
784static int amdgpu_cs_vm_handling(struct amdgpu_cs_parser *p)
785{
786 struct amdgpu_ring *ring = to_amdgpu_ring(p->entity->rq->sched);
787 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
788 struct amdgpu_device *adev = p->adev;
789 struct amdgpu_vm *vm = &fpriv->vm;
790 struct amdgpu_bo_list_entry *e;
791 struct amdgpu_bo_va *bo_va;
792 struct amdgpu_bo *bo;
793 int r;
794
795 /* Only for UVD/VCE VM emulation */
796 if (ring->funcs->parse_cs || ring->funcs->patch_cs_in_place) {
797 unsigned i, j;
798
799 for (i = 0, j = 0; i < p->nchunks && j < p->job->num_ibs; i++) {
800 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
801 struct amdgpu_bo_va_mapping *m;
802 struct amdgpu_bo *aobj = NULL;
803 struct amdgpu_cs_chunk *chunk;
804 uint64_t offset, va_start;
805 struct amdgpu_ib *ib;
806 uint8_t *kptr;
807
808 chunk = &p->chunks[i];
809 ib = &p->job->ibs[j];
810 chunk_ib = chunk->kdata;
811
812 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
813 continue;
814
815 va_start = chunk_ib->va_start & AMDGPU_GMC_HOLE_MASK;
816 r = amdgpu_cs_find_mapping(p, va_start, &aobj, &m);
817 if (r) {
818 DRM_ERROR("IB va_start is invalid\n");
819 return r;
820 }
821
822 if ((va_start + chunk_ib->ib_bytes) >
823 (m->last + 1) * AMDGPU_GPU_PAGE_SIZE) {
824 DRM_ERROR("IB va_start+ib_bytes is invalid\n");
825 return -EINVAL;
826 }
827
828 /* the IB should be reserved at this point */
829 r = amdgpu_bo_kmap(aobj, (void **)&kptr);
830 if (r) {
831 return r;
832 }
833
834 offset = m->start * AMDGPU_GPU_PAGE_SIZE;
835 kptr += va_start - offset;
836
837 if (ring->funcs->parse_cs) {
838 memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
839 amdgpu_bo_kunmap(aobj);
840
841 r = amdgpu_ring_parse_cs(ring, p, j);
842 if (r)
843 return r;
844 } else {
845 ib->ptr = (uint32_t *)kptr;
846 r = amdgpu_ring_patch_cs_in_place(ring, p, j);
847 amdgpu_bo_kunmap(aobj);
848 if (r)
849 return r;
850 }
851
852 j++;
853 }
854 }
855
856 if (!p->job->vm)
857 return amdgpu_cs_sync_rings(p);
858
859
860 r = amdgpu_vm_clear_freed(adev, vm, NULL);
861 if (r)
862 return r;
863
864 r = amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
865 if (r)
866 return r;
867
868 r = amdgpu_sync_fence(adev, &p->job->sync,
869 fpriv->prt_va->last_pt_update, false);
870 if (r)
871 return r;
872
873 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
874 struct dma_fence *f;
875
876 bo_va = fpriv->csa_va;
877 BUG_ON(!bo_va);
878 r = amdgpu_vm_bo_update(adev, bo_va, false);
879 if (r)
880 return r;
881
882 f = bo_va->last_pt_update;
883 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
884 if (r)
885 return r;
886 }
887
888 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
889 struct dma_fence *f;
890
891 /* ignore duplicates */
892 bo = ttm_to_amdgpu_bo(e->tv.bo);
893 if (!bo)
894 continue;
895
896 bo_va = e->bo_va;
897 if (bo_va == NULL)
898 continue;
899
900 r = amdgpu_vm_bo_update(adev, bo_va, false);
901 if (r)
902 return r;
903
904 f = bo_va->last_pt_update;
905 r = amdgpu_sync_fence(adev, &p->job->sync, f, false);
906 if (r)
907 return r;
908 }
909
910 r = amdgpu_vm_handle_moved(adev, vm);
911 if (r)
912 return r;
913
914 r = amdgpu_vm_update_directories(adev, vm);
915 if (r)
916 return r;
917
918 r = amdgpu_sync_fence(adev, &p->job->sync, vm->last_update, false);
919 if (r)
920 return r;
921
922 p->job->vm_pd_addr = amdgpu_gmc_pd_addr(vm->root.base.bo);
923
924 if (amdgpu_vm_debug) {
925 /* Invalidate all BOs to test for userspace bugs */
926 amdgpu_bo_list_for_each_entry(e, p->bo_list) {
927 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
928
929 /* ignore duplicates */
930 if (!bo)
931 continue;
932
933 amdgpu_vm_bo_invalidate(adev, bo, false);
934 }
935 }
936
937 return amdgpu_cs_sync_rings(p);
938}
939
940static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
941 struct amdgpu_cs_parser *parser)
942{
943 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
944 struct amdgpu_vm *vm = &fpriv->vm;
945 int r, ce_preempt = 0, de_preempt = 0;
946 struct amdgpu_ring *ring;
947 int i, j;
948
949 for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
950 struct amdgpu_cs_chunk *chunk;
951 struct amdgpu_ib *ib;
952 struct drm_amdgpu_cs_chunk_ib *chunk_ib;
953 struct drm_sched_entity *entity;
954
955 chunk = &parser->chunks[i];
956 ib = &parser->job->ibs[j];
957 chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;
958
959 if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
960 continue;
961
962 if (chunk_ib->ip_type == AMDGPU_HW_IP_GFX &&
963 (amdgpu_mcbp || amdgpu_sriov_vf(adev))) {
964 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREEMPT) {
965 if (chunk_ib->flags & AMDGPU_IB_FLAG_CE)
966 ce_preempt++;
967 else
968 de_preempt++;
969 }
970
971 /* each GFX command submit allows 0 or 1 IB preemptible for CE & DE */
972 if (ce_preempt > 1 || de_preempt > 1)
973 return -EINVAL;
974 }
975
976 r = amdgpu_ctx_get_entity(parser->ctx, chunk_ib->ip_type,
977 chunk_ib->ip_instance, chunk_ib->ring,
978 &entity);
979 if (r)
980 return r;
981
982 if (chunk_ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
983 parser->job->preamble_status |=
984 AMDGPU_PREAMBLE_IB_PRESENT;
985
986 if (parser->entity && parser->entity != entity)
987 return -EINVAL;
988
989 parser->entity = entity;
990
991 ring = to_amdgpu_ring(entity->rq->sched);
992 r = amdgpu_ib_get(adev, vm, ring->funcs->parse_cs ?
993 chunk_ib->ib_bytes : 0, ib);
994 if (r) {
995 DRM_ERROR("Failed to get ib !\n");
996 return r;
997 }
998
999 ib->gpu_addr = chunk_ib->va_start;
1000 ib->length_dw = chunk_ib->ib_bytes / 4;
1001 ib->flags = chunk_ib->flags;
1002
1003 j++;
1004 }
1005
1006 /* MM engine doesn't support user fences */
1007 ring = to_amdgpu_ring(parser->entity->rq->sched);
1008 if (parser->job->uf_addr && ring->funcs->no_user_fence)
1009 return -EINVAL;
1010
1011 return amdgpu_ctx_wait_prev_fence(parser->ctx, parser->entity);
1012}
1013
1014static int amdgpu_cs_process_fence_dep(struct amdgpu_cs_parser *p,
1015 struct amdgpu_cs_chunk *chunk)
1016{
1017 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1018 unsigned num_deps;
1019 int i, r;
1020 struct drm_amdgpu_cs_chunk_dep *deps;
1021
1022 deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
1023 num_deps = chunk->length_dw * 4 /
1024 sizeof(struct drm_amdgpu_cs_chunk_dep);
1025
1026 for (i = 0; i < num_deps; ++i) {
1027 struct amdgpu_ctx *ctx;
1028 struct drm_sched_entity *entity;
1029 struct dma_fence *fence;
1030
1031 ctx = amdgpu_ctx_get(fpriv, deps[i].ctx_id);
1032 if (ctx == NULL)
1033 return -EINVAL;
1034
1035 r = amdgpu_ctx_get_entity(ctx, deps[i].ip_type,
1036 deps[i].ip_instance,
1037 deps[i].ring, &entity);
1038 if (r) {
1039 amdgpu_ctx_put(ctx);
1040 return r;
1041 }
1042
1043 fence = amdgpu_ctx_get_fence(ctx, entity, deps[i].handle);
1044 amdgpu_ctx_put(ctx);
1045
1046 if (IS_ERR(fence))
1047 return PTR_ERR(fence);
1048 else if (!fence)
1049 continue;
1050
1051 if (chunk->chunk_id == AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES) {
1052 struct drm_sched_fence *s_fence;
1053 struct dma_fence *old = fence;
1054
1055 s_fence = to_drm_sched_fence(fence);
1056 fence = dma_fence_get(&s_fence->scheduled);
1057 dma_fence_put(old);
1058 }
1059
1060 r = amdgpu_sync_fence(p->adev, &p->job->sync, fence, true);
1061 dma_fence_put(fence);
1062 if (r)
1063 return r;
1064 }
1065 return 0;
1066}
1067
1068static int amdgpu_syncobj_lookup_and_add_to_sync(struct amdgpu_cs_parser *p,
1069 uint32_t handle, u64 point,
1070 u64 flags)
1071{
1072 struct dma_fence *fence;
1073 int r;
1074
1075 r = drm_syncobj_find_fence(p->filp, handle, point, flags, &fence);
1076 if (r) {
1077 DRM_ERROR("syncobj %u failed to find fence @ %llu (%d)!\n",
1078 handle, point, r);
1079 return r;
1080 }
1081
1082 r = amdgpu_sync_fence(p->adev, &p->job->sync, fence, true);
1083 dma_fence_put(fence);
1084
1085 return r;
1086}
1087
1088static int amdgpu_cs_process_syncobj_in_dep(struct amdgpu_cs_parser *p,
1089 struct amdgpu_cs_chunk *chunk)
1090{
1091 struct drm_amdgpu_cs_chunk_sem *deps;
1092 unsigned num_deps;
1093 int i, r;
1094
1095 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1096 num_deps = chunk->length_dw * 4 /
1097 sizeof(struct drm_amdgpu_cs_chunk_sem);
1098 for (i = 0; i < num_deps; ++i) {
1099 r = amdgpu_syncobj_lookup_and_add_to_sync(p, deps[i].handle,
1100 0, 0);
1101 if (r)
1102 return r;
1103 }
1104
1105 return 0;
1106}
1107
1108
1109static int amdgpu_cs_process_syncobj_timeline_in_dep(struct amdgpu_cs_parser *p,
1110 struct amdgpu_cs_chunk *chunk)
1111{
1112 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1113 unsigned num_deps;
1114 int i, r;
1115
1116 syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1117 num_deps = chunk->length_dw * 4 /
1118 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1119 for (i = 0; i < num_deps; ++i) {
1120 r = amdgpu_syncobj_lookup_and_add_to_sync(p,
1121 syncobj_deps[i].handle,
1122 syncobj_deps[i].point,
1123 syncobj_deps[i].flags);
1124 if (r)
1125 return r;
1126 }
1127
1128 return 0;
1129}
1130
1131static int amdgpu_cs_process_syncobj_out_dep(struct amdgpu_cs_parser *p,
1132 struct amdgpu_cs_chunk *chunk)
1133{
1134 struct drm_amdgpu_cs_chunk_sem *deps;
1135 unsigned num_deps;
1136 int i;
1137
1138 deps = (struct drm_amdgpu_cs_chunk_sem *)chunk->kdata;
1139 num_deps = chunk->length_dw * 4 /
1140 sizeof(struct drm_amdgpu_cs_chunk_sem);
1141
1142 if (p->post_deps)
1143 return -EINVAL;
1144
1145 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1146 GFP_KERNEL);
1147 p->num_post_deps = 0;
1148
1149 if (!p->post_deps)
1150 return -ENOMEM;
1151
1152
1153 for (i = 0; i < num_deps; ++i) {
1154 p->post_deps[i].syncobj =
1155 drm_syncobj_find(p->filp, deps[i].handle);
1156 if (!p->post_deps[i].syncobj)
1157 return -EINVAL;
1158 p->post_deps[i].chain = NULL;
1159 p->post_deps[i].point = 0;
1160 p->num_post_deps++;
1161 }
1162
1163 return 0;
1164}
1165
1166
1167static int amdgpu_cs_process_syncobj_timeline_out_dep(struct amdgpu_cs_parser *p,
1168 struct amdgpu_cs_chunk *chunk)
1169{
1170 struct drm_amdgpu_cs_chunk_syncobj *syncobj_deps;
1171 unsigned num_deps;
1172 int i;
1173
1174 syncobj_deps = (struct drm_amdgpu_cs_chunk_syncobj *)chunk->kdata;
1175 num_deps = chunk->length_dw * 4 /
1176 sizeof(struct drm_amdgpu_cs_chunk_syncobj);
1177
1178 if (p->post_deps)
1179 return -EINVAL;
1180
1181 p->post_deps = kmalloc_array(num_deps, sizeof(*p->post_deps),
1182 GFP_KERNEL);
1183 p->num_post_deps = 0;
1184
1185 if (!p->post_deps)
1186 return -ENOMEM;
1187
1188 for (i = 0; i < num_deps; ++i) {
1189 struct amdgpu_cs_post_dep *dep = &p->post_deps[i];
1190
1191 dep->chain = NULL;
1192 if (syncobj_deps[i].point) {
1193 dep->chain = kmalloc(sizeof(*dep->chain), GFP_KERNEL);
1194 if (!dep->chain)
1195 return -ENOMEM;
1196 }
1197
1198 dep->syncobj = drm_syncobj_find(p->filp,
1199 syncobj_deps[i].handle);
1200 if (!dep->syncobj) {
1201 kfree(dep->chain);
1202 return -EINVAL;
1203 }
1204 dep->point = syncobj_deps[i].point;
1205 p->num_post_deps++;
1206 }
1207
1208 return 0;
1209}
1210
1211static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
1212 struct amdgpu_cs_parser *p)
1213{
1214 int i, r;
1215
1216 for (i = 0; i < p->nchunks; ++i) {
1217 struct amdgpu_cs_chunk *chunk;
1218
1219 chunk = &p->chunks[i];
1220
1221 switch (chunk->chunk_id) {
1222 case AMDGPU_CHUNK_ID_DEPENDENCIES:
1223 case AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES:
1224 r = amdgpu_cs_process_fence_dep(p, chunk);
1225 if (r)
1226 return r;
1227 break;
1228 case AMDGPU_CHUNK_ID_SYNCOBJ_IN:
1229 r = amdgpu_cs_process_syncobj_in_dep(p, chunk);
1230 if (r)
1231 return r;
1232 break;
1233 case AMDGPU_CHUNK_ID_SYNCOBJ_OUT:
1234 r = amdgpu_cs_process_syncobj_out_dep(p, chunk);
1235 if (r)
1236 return r;
1237 break;
1238 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_WAIT:
1239 r = amdgpu_cs_process_syncobj_timeline_in_dep(p, chunk);
1240 if (r)
1241 return r;
1242 break;
1243 case AMDGPU_CHUNK_ID_SYNCOBJ_TIMELINE_SIGNAL:
1244 r = amdgpu_cs_process_syncobj_timeline_out_dep(p, chunk);
1245 if (r)
1246 return r;
1247 break;
1248 }
1249 }
1250
1251 return 0;
1252}
1253
1254static void amdgpu_cs_post_dependencies(struct amdgpu_cs_parser *p)
1255{
1256 int i;
1257
1258 for (i = 0; i < p->num_post_deps; ++i) {
1259 if (p->post_deps[i].chain && p->post_deps[i].point) {
1260 drm_syncobj_add_point(p->post_deps[i].syncobj,
1261 p->post_deps[i].chain,
1262 p->fence, p->post_deps[i].point);
1263 p->post_deps[i].chain = NULL;
1264 } else {
1265 drm_syncobj_replace_fence(p->post_deps[i].syncobj,
1266 p->fence);
1267 }
1268 }
1269}
1270
1271static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
1272 union drm_amdgpu_cs *cs)
1273{
1274 struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
1275 struct drm_sched_entity *entity = p->entity;
1276 enum drm_sched_priority priority;
1277 struct amdgpu_ring *ring;
1278 struct amdgpu_bo_list_entry *e;
1279 struct amdgpu_job *job;
1280 uint64_t seq;
1281 int r;
1282
1283 job = p->job;
1284 p->job = NULL;
1285
1286 r = drm_sched_job_init(&job->base, entity, p->filp);
1287 if (r)
1288 goto error_unlock;
1289
1290 /* No memory allocation is allowed while holding the mn lock.
1291 * p->mn is hold until amdgpu_cs_submit is finished and fence is added
1292 * to BOs.
1293 */
1294 amdgpu_mn_lock(p->mn);
1295
1296 /* If userptr are invalidated after amdgpu_cs_parser_bos(), return
1297 * -EAGAIN, drmIoctl in libdrm will restart the amdgpu_cs_ioctl.
1298 */
1299 amdgpu_bo_list_for_each_userptr_entry(e, p->bo_list) {
1300 struct amdgpu_bo *bo = ttm_to_amdgpu_bo(e->tv.bo);
1301
1302 r |= !amdgpu_ttm_tt_get_user_pages_done(bo->tbo.ttm);
1303 }
1304 if (r) {
1305 r = -EAGAIN;
1306 goto error_abort;
1307 }
1308
1309 job->owner = p->filp;
1310 p->fence = dma_fence_get(&job->base.s_fence->finished);
1311
1312 amdgpu_ctx_add_fence(p->ctx, entity, p->fence, &seq);
1313 amdgpu_cs_post_dependencies(p);
1314
1315 if ((job->preamble_status & AMDGPU_PREAMBLE_IB_PRESENT) &&
1316 !p->ctx->preamble_presented) {
1317 job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
1318 p->ctx->preamble_presented = true;
1319 }
1320
1321 cs->out.handle = seq;
1322 job->uf_sequence = seq;
1323
1324 amdgpu_job_free_resources(job);
1325
1326 trace_amdgpu_cs_ioctl(job);
1327 amdgpu_vm_bo_trace_cs(&fpriv->vm, &p->ticket);
1328 priority = job->base.s_priority;
1329 drm_sched_entity_push_job(&job->base, entity);
1330
1331 ring = to_amdgpu_ring(entity->rq->sched);
1332 amdgpu_ring_priority_get(ring, priority);
1333
1334 amdgpu_vm_move_to_lru_tail(p->adev, &fpriv->vm);
1335
1336 ttm_eu_fence_buffer_objects(&p->ticket, &p->validated, p->fence);
1337 amdgpu_mn_unlock(p->mn);
1338
1339 return 0;
1340
1341error_abort:
1342 drm_sched_job_cleanup(&job->base);
1343 amdgpu_mn_unlock(p->mn);
1344
1345error_unlock:
1346 amdgpu_job_free(job);
1347 return r;
1348}
1349
1350int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
1351{
1352 struct amdgpu_device *adev = dev->dev_private;
1353 union drm_amdgpu_cs *cs = data;
1354 struct amdgpu_cs_parser parser = {};
1355 bool reserved_buffers = false;
1356 int i, r;
1357
1358 if (!adev->accel_working)
1359 return -EBUSY;
1360
1361 parser.adev = adev;
1362 parser.filp = filp;
1363
1364 r = amdgpu_cs_parser_init(&parser, data);
1365 if (r) {
1366 DRM_ERROR("Failed to initialize parser %d!\n", r);
1367 goto out;
1368 }
1369
1370 r = amdgpu_cs_ib_fill(adev, &parser);
1371 if (r)
1372 goto out;
1373
1374 r = amdgpu_cs_dependencies(adev, &parser);
1375 if (r) {
1376 DRM_ERROR("Failed in the dependencies handling %d!\n", r);
1377 goto out;
1378 }
1379
1380 r = amdgpu_cs_parser_bos(&parser, data);
1381 if (r) {
1382 if (r == -ENOMEM)
1383 DRM_ERROR("Not enough memory for command submission!\n");
1384 else if (r != -ERESTARTSYS && r != -EAGAIN)
1385 DRM_ERROR("Failed to process the buffer list %d!\n", r);
1386 goto out;
1387 }
1388
1389 reserved_buffers = true;
1390
1391 for (i = 0; i < parser.job->num_ibs; i++)
1392 trace_amdgpu_cs(&parser, i);
1393
1394 r = amdgpu_cs_vm_handling(&parser);
1395 if (r)
1396 goto out;
1397
1398 r = amdgpu_cs_submit(&parser, cs);
1399
1400out:
1401 amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1402
1403 return r;
1404}
1405
1406/**
1407 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
1408 *
1409 * @dev: drm device
1410 * @data: data from userspace
1411 * @filp: file private
1412 *
1413 * Wait for the command submission identified by handle to finish.
1414 */
1415int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
1416 struct drm_file *filp)
1417{
1418 union drm_amdgpu_wait_cs *wait = data;
1419 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1420 struct drm_sched_entity *entity;
1421 struct amdgpu_ctx *ctx;
1422 struct dma_fence *fence;
1423 long r;
1424
1425 ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
1426 if (ctx == NULL)
1427 return -EINVAL;
1428
1429 r = amdgpu_ctx_get_entity(ctx, wait->in.ip_type, wait->in.ip_instance,
1430 wait->in.ring, &entity);
1431 if (r) {
1432 amdgpu_ctx_put(ctx);
1433 return r;
1434 }
1435
1436 fence = amdgpu_ctx_get_fence(ctx, entity, wait->in.handle);
1437 if (IS_ERR(fence))
1438 r = PTR_ERR(fence);
1439 else if (fence) {
1440 r = dma_fence_wait_timeout(fence, true, timeout);
1441 if (r > 0 && fence->error)
1442 r = fence->error;
1443 dma_fence_put(fence);
1444 } else
1445 r = 1;
1446
1447 amdgpu_ctx_put(ctx);
1448 if (r < 0)
1449 return r;
1450
1451 memset(wait, 0, sizeof(*wait));
1452 wait->out.status = (r == 0);
1453
1454 return 0;
1455}
1456
1457/**
1458 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
1459 *
1460 * @adev: amdgpu device
1461 * @filp: file private
1462 * @user: drm_amdgpu_fence copied from user space
1463 */
1464static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
1465 struct drm_file *filp,
1466 struct drm_amdgpu_fence *user)
1467{
1468 struct drm_sched_entity *entity;
1469 struct amdgpu_ctx *ctx;
1470 struct dma_fence *fence;
1471 int r;
1472
1473 ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
1474 if (ctx == NULL)
1475 return ERR_PTR(-EINVAL);
1476
1477 r = amdgpu_ctx_get_entity(ctx, user->ip_type, user->ip_instance,
1478 user->ring, &entity);
1479 if (r) {
1480 amdgpu_ctx_put(ctx);
1481 return ERR_PTR(r);
1482 }
1483
1484 fence = amdgpu_ctx_get_fence(ctx, entity, user->seq_no);
1485 amdgpu_ctx_put(ctx);
1486
1487 return fence;
1488}
1489
1490int amdgpu_cs_fence_to_handle_ioctl(struct drm_device *dev, void *data,
1491 struct drm_file *filp)
1492{
1493 struct amdgpu_device *adev = dev->dev_private;
1494 union drm_amdgpu_fence_to_handle *info = data;
1495 struct dma_fence *fence;
1496 struct drm_syncobj *syncobj;
1497 struct sync_file *sync_file;
1498 int fd, r;
1499
1500 fence = amdgpu_cs_get_fence(adev, filp, &info->in.fence);
1501 if (IS_ERR(fence))
1502 return PTR_ERR(fence);
1503
1504 if (!fence)
1505 fence = dma_fence_get_stub();
1506
1507 switch (info->in.what) {
1508 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ:
1509 r = drm_syncobj_create(&syncobj, 0, fence);
1510 dma_fence_put(fence);
1511 if (r)
1512 return r;
1513 r = drm_syncobj_get_handle(filp, syncobj, &info->out.handle);
1514 drm_syncobj_put(syncobj);
1515 return r;
1516
1517 case AMDGPU_FENCE_TO_HANDLE_GET_SYNCOBJ_FD:
1518 r = drm_syncobj_create(&syncobj, 0, fence);
1519 dma_fence_put(fence);
1520 if (r)
1521 return r;
1522 r = drm_syncobj_get_fd(syncobj, (int*)&info->out.handle);
1523 drm_syncobj_put(syncobj);
1524 return r;
1525
1526 case AMDGPU_FENCE_TO_HANDLE_GET_SYNC_FILE_FD:
1527 fd = get_unused_fd_flags(O_CLOEXEC);
1528 if (fd < 0) {
1529 dma_fence_put(fence);
1530 return fd;
1531 }
1532
1533 sync_file = sync_file_create(fence);
1534 dma_fence_put(fence);
1535 if (!sync_file) {
1536 put_unused_fd(fd);
1537 return -ENOMEM;
1538 }
1539
1540 fd_install(fd, sync_file->file);
1541 info->out.handle = fd;
1542 return 0;
1543
1544 default:
1545 return -EINVAL;
1546 }
1547}
1548
1549/**
1550 * amdgpu_cs_wait_all_fence - wait on all fences to signal
1551 *
1552 * @adev: amdgpu device
1553 * @filp: file private
1554 * @wait: wait parameters
1555 * @fences: array of drm_amdgpu_fence
1556 */
1557static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
1558 struct drm_file *filp,
1559 union drm_amdgpu_wait_fences *wait,
1560 struct drm_amdgpu_fence *fences)
1561{
1562 uint32_t fence_count = wait->in.fence_count;
1563 unsigned int i;
1564 long r = 1;
1565
1566 for (i = 0; i < fence_count; i++) {
1567 struct dma_fence *fence;
1568 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1569
1570 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1571 if (IS_ERR(fence))
1572 return PTR_ERR(fence);
1573 else if (!fence)
1574 continue;
1575
1576 r = dma_fence_wait_timeout(fence, true, timeout);
1577 dma_fence_put(fence);
1578 if (r < 0)
1579 return r;
1580
1581 if (r == 0)
1582 break;
1583
1584 if (fence->error)
1585 return fence->error;
1586 }
1587
1588 memset(wait, 0, sizeof(*wait));
1589 wait->out.status = (r > 0);
1590
1591 return 0;
1592}
1593
1594/**
1595 * amdgpu_cs_wait_any_fence - wait on any fence to signal
1596 *
1597 * @adev: amdgpu device
1598 * @filp: file private
1599 * @wait: wait parameters
1600 * @fences: array of drm_amdgpu_fence
1601 */
1602static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
1603 struct drm_file *filp,
1604 union drm_amdgpu_wait_fences *wait,
1605 struct drm_amdgpu_fence *fences)
1606{
1607 unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
1608 uint32_t fence_count = wait->in.fence_count;
1609 uint32_t first = ~0;
1610 struct dma_fence **array;
1611 unsigned int i;
1612 long r;
1613
1614 /* Prepare the fence array */
1615 array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);
1616
1617 if (array == NULL)
1618 return -ENOMEM;
1619
1620 for (i = 0; i < fence_count; i++) {
1621 struct dma_fence *fence;
1622
1623 fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
1624 if (IS_ERR(fence)) {
1625 r = PTR_ERR(fence);
1626 goto err_free_fence_array;
1627 } else if (fence) {
1628 array[i] = fence;
1629 } else { /* NULL, the fence has been already signaled */
1630 r = 1;
1631 first = i;
1632 goto out;
1633 }
1634 }
1635
1636 r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
1637 &first);
1638 if (r < 0)
1639 goto err_free_fence_array;
1640
1641out:
1642 memset(wait, 0, sizeof(*wait));
1643 wait->out.status = (r > 0);
1644 wait->out.first_signaled = first;
1645
1646 if (first < fence_count && array[first])
1647 r = array[first]->error;
1648 else
1649 r = 0;
1650
1651err_free_fence_array:
1652 for (i = 0; i < fence_count; i++)
1653 dma_fence_put(array[i]);
1654 kfree(array);
1655
1656 return r;
1657}
1658
1659/**
1660 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
1661 *
1662 * @dev: drm device
1663 * @data: data from userspace
1664 * @filp: file private
1665 */
1666int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
1667 struct drm_file *filp)
1668{
1669 struct amdgpu_device *adev = dev->dev_private;
1670 union drm_amdgpu_wait_fences *wait = data;
1671 uint32_t fence_count = wait->in.fence_count;
1672 struct drm_amdgpu_fence *fences_user;
1673 struct drm_amdgpu_fence *fences;
1674 int r;
1675
1676 /* Get the fences from userspace */
1677 fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
1678 GFP_KERNEL);
1679 if (fences == NULL)
1680 return -ENOMEM;
1681
1682 fences_user = u64_to_user_ptr(wait->in.fences);
1683 if (copy_from_user(fences, fences_user,
1684 sizeof(struct drm_amdgpu_fence) * fence_count)) {
1685 r = -EFAULT;
1686 goto err_free_fences;
1687 }
1688
1689 if (wait->in.wait_all)
1690 r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
1691 else
1692 r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);
1693
1694err_free_fences:
1695 kfree(fences);
1696
1697 return r;
1698}
1699
1700/**
1701 * amdgpu_cs_find_bo_va - find bo_va for VM address
1702 *
1703 * @parser: command submission parser context
1704 * @addr: VM address
1705 * @bo: resulting BO of the mapping found
1706 *
1707 * Search the buffer objects in the command submission context for a certain
1708 * virtual memory address. Returns allocation structure when found, NULL
1709 * otherwise.
1710 */
1711int amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
1712 uint64_t addr, struct amdgpu_bo **bo,
1713 struct amdgpu_bo_va_mapping **map)
1714{
1715 struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
1716 struct ttm_operation_ctx ctx = { false, false };
1717 struct amdgpu_vm *vm = &fpriv->vm;
1718 struct amdgpu_bo_va_mapping *mapping;
1719 int r;
1720
1721 addr /= AMDGPU_GPU_PAGE_SIZE;
1722
1723 mapping = amdgpu_vm_bo_lookup_mapping(vm, addr);
1724 if (!mapping || !mapping->bo_va || !mapping->bo_va->base.bo)
1725 return -EINVAL;
1726
1727 *bo = mapping->bo_va->base.bo;
1728 *map = mapping;
1729
1730 /* Double check that the BO is reserved by this CS */
1731 if (dma_resv_locking_ctx((*bo)->tbo.base.resv) != &parser->ticket)
1732 return -EINVAL;
1733
1734 if (!((*bo)->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)) {
1735 (*bo)->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
1736 amdgpu_bo_placement_from_domain(*bo, (*bo)->allowed_domains);
1737 r = ttm_bo_validate(&(*bo)->tbo, &(*bo)->placement, &ctx);
1738 if (r)
1739 return r;
1740 }
1741
1742 return amdgpu_ttm_alloc_gart(&(*bo)->tbo);
1743}