amdgpu_cs.c 33.6 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
/*
 * Copyright 2008 Jerome Glisse.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * PRECISION INSIGHT AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
 * DEALINGS IN THE SOFTWARE.
 *
 * Authors:
 *    Jerome Glisse <glisse@freedesktop.org>
 */
27
#include <linux/pagemap.h>
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
#include <drm/drmP.h>
#include <drm/amdgpu_drm.h>
#include "amdgpu.h"
#include "amdgpu_trace.h"

int amdgpu_cs_get_ring(struct amdgpu_device *adev, u32 ip_type,
		       u32 ip_instance, u32 ring,
		       struct amdgpu_ring **out_ring)
{
	/* Right now all IPs have only one instance - multiple rings. */
	if (ip_instance != 0) {
		DRM_ERROR("invalid ip instance: %d\n", ip_instance);
		return -EINVAL;
	}

	switch (ip_type) {
	default:
		DRM_ERROR("unknown ip type: %d\n", ip_type);
		return -EINVAL;
	case AMDGPU_HW_IP_GFX:
		if (ring < adev->gfx.num_gfx_rings) {
			*out_ring = &adev->gfx.gfx_ring[ring];
		} else {
			DRM_ERROR("only %d gfx rings are supported now\n",
				  adev->gfx.num_gfx_rings);
			return -EINVAL;
		}
		break;
	case AMDGPU_HW_IP_COMPUTE:
		if (ring < adev->gfx.num_compute_rings) {
			*out_ring = &adev->gfx.compute_ring[ring];
		} else {
			DRM_ERROR("only %d compute rings are supported now\n",
				  adev->gfx.num_compute_rings);
			return -EINVAL;
		}
		break;
	case AMDGPU_HW_IP_DMA:
66
67
		if (ring < adev->sdma.num_instances) {
			*out_ring = &adev->sdma.instance[ring].ring;
68
		} else {
69
70
			DRM_ERROR("only %d SDMA rings are supported\n",
				  adev->sdma.num_instances);
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
			return -EINVAL;
		}
		break;
	case AMDGPU_HW_IP_UVD:
		*out_ring = &adev->uvd.ring;
		break;
	case AMDGPU_HW_IP_VCE:
		if (ring < 2){
			*out_ring = &adev->vce.ring[ring];
		} else {
			DRM_ERROR("only two VCE rings are supported\n");
			return -EINVAL;
		}
		break;
	}
86
87
88
89
90
91
92

	if (!(*out_ring && (*out_ring)->adev)) {
		DRM_ERROR("Ring %d is not initialized on IP %d\n",
			  ring, ip_type);
		return -EINVAL;
	}

93
94
95
	return 0;
}

96
static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
97
98
				      struct drm_amdgpu_cs_chunk_fence *data,
				      uint32_t *offset)
99
100
{
	struct drm_gem_object *gobj;
101
	unsigned long size;
102

103
	gobj = drm_gem_object_lookup(p->filp, data->handle);
104
105
106
	if (gobj == NULL)
		return -EINVAL;

107
	p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
108
109
110
	p->uf_entry.priority = 0;
	p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
	p->uf_entry.tv.shared = true;
111
	p->uf_entry.user_pages = NULL;
112
113
114
115
116

	size = amdgpu_bo_size(p->uf_entry.robj);
	if (size != PAGE_SIZE || (data->offset + 8) > size)
		return -EINVAL;

117
	*offset = data->offset;
118
119

	drm_gem_object_unreference_unlocked(gobj);
120
121
122
123
124
125

	if (amdgpu_ttm_tt_get_usermm(p->uf_entry.robj->tbo.ttm)) {
		amdgpu_bo_unref(&p->uf_entry.robj);
		return -EINVAL;
	}

126
127
128
	return 0;
}

129
130
int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
{
131
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
132
	struct amdgpu_vm *vm = &fpriv->vm;
133
134
	union drm_amdgpu_cs *cs = data;
	uint64_t *chunk_array_user;
135
	uint64_t *chunk_array;
136
	unsigned size, num_ibs = 0;
137
	uint32_t uf_offset = 0;
138
	int i;
139
	int ret;
140

141
142
143
144
145
146
	if (cs->in.num_chunks == 0)
		return 0;

	chunk_array = kmalloc_array(cs->in.num_chunks, sizeof(uint64_t), GFP_KERNEL);
	if (!chunk_array)
		return -ENOMEM;
147

148
149
	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
	if (!p->ctx) {
150
151
		ret = -EINVAL;
		goto free_chunk;
152
	}
153

154
	/* get chunks */
155
	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
156
157
	if (copy_from_user(chunk_array, chunk_array_user,
			   sizeof(uint64_t)*cs->in.num_chunks)) {
158
		ret = -EFAULT;
159
		goto put_ctx;
160
161
162
	}

	p->nchunks = cs->in.num_chunks;
163
	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
164
			    GFP_KERNEL);
165
166
	if (!p->chunks) {
		ret = -ENOMEM;
167
		goto put_ctx;
168
169
170
171
172
173
174
	}

	for (i = 0; i < p->nchunks; i++) {
		struct drm_amdgpu_cs_chunk __user **chunk_ptr = NULL;
		struct drm_amdgpu_cs_chunk user_chunk;
		uint32_t __user *cdata;

175
		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
176
177
		if (copy_from_user(&user_chunk, chunk_ptr,
				       sizeof(struct drm_amdgpu_cs_chunk))) {
178
179
180
			ret = -EFAULT;
			i--;
			goto free_partial_kdata;
181
182
183
184
185
		}
		p->chunks[i].chunk_id = user_chunk.chunk_id;
		p->chunks[i].length_dw = user_chunk.length_dw;

		size = p->chunks[i].length_dw;
186
		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
187
188
189

		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
		if (p->chunks[i].kdata == NULL) {
190
191
192
			ret = -ENOMEM;
			i--;
			goto free_partial_kdata;
193
194
195
		}
		size *= sizeof(uint32_t);
		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
196
197
			ret = -EFAULT;
			goto free_partial_kdata;
198
199
		}

200
201
		switch (p->chunks[i].chunk_id) {
		case AMDGPU_CHUNK_ID_IB:
202
			++num_ibs;
203
204
205
			break;

		case AMDGPU_CHUNK_ID_FENCE:
206
			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
207
			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
208
209
				ret = -EINVAL;
				goto free_partial_kdata;
210
			}
211

212
213
			ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
							 &uf_offset);
214
215
216
			if (ret)
				goto free_partial_kdata;

217
218
			break;

219
220
221
		case AMDGPU_CHUNK_ID_DEPENDENCIES:
			break;

222
		default:
223
224
			ret = -EINVAL;
			goto free_partial_kdata;
225
226
227
		}
	}

228
	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
229
	if (ret)
230
		goto free_all_kdata;
231

232
233
	if (p->uf_entry.robj)
		p->job->uf_addr = uf_offset;
234
	kfree(chunk_array);
235
236
237
238
239
240
241
242
	return 0;

free_all_kdata:
	i = p->nchunks - 1;
free_partial_kdata:
	for (; i >= 0; i--)
		drm_free_large(p->chunks[i].kdata);
	kfree(p->chunks);
243
put_ctx:
244
245
246
247
248
	amdgpu_ctx_put(p->ctx);
free_chunk:
	kfree(chunk_array);

	return ret;
249
250
}

251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
/* Convert microseconds to bytes. */
static u64 us_to_bytes(struct amdgpu_device *adev, s64 us)
{
	if (us <= 0 || !adev->mm_stats.log2_max_MBps)
		return 0;

	/* Since accum_us is incremented by a million per second, just
	 * multiply it by the number of MB/s to get the number of bytes.
	 */
	return us << adev->mm_stats.log2_max_MBps;
}

static s64 bytes_to_us(struct amdgpu_device *adev, u64 bytes)
{
	if (!adev->mm_stats.log2_max_MBps)
		return 0;

	return bytes >> adev->mm_stats.log2_max_MBps;
}

/* Returns how many bytes TTM can move right now. If no bytes can be moved,
 * it returns 0. If it returns non-zero, it's OK to move at least one buffer,
 * which means it can go over the threshold once. If that happens, the driver
 * will be in debt and no other buffer migrations can be done until that debt
 * is repaid.
 *
 * This approach allows moving a buffer of any size (it's important to allow
 * that).
 *
 * The currency is simply time in microseconds and it increases as the clock
 * ticks. The accumulated microseconds (us) are converted to bytes and
 * returned.
283
284
285
 */
static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
{
286
287
288
	s64 time_us, increment_us;
	u64 max_bytes;
	u64 free_vram, total_vram, used_vram;
289

290
291
	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
	 * throttling.
292
	 *
293
294
295
296
297
	 * It means that in order to get full max MBps, at least 5 IBs per
	 * second must be submitted and not more than 200ms apart from each
	 * other.
	 */
	const s64 us_upper_bound = 200000;
298

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
	if (!adev->mm_stats.log2_max_MBps)
		return 0;

	total_vram = adev->mc.real_vram_size - adev->vram_pin_size;
	used_vram = atomic64_read(&adev->vram_usage);
	free_vram = used_vram >= total_vram ? 0 : total_vram - used_vram;

	spin_lock(&adev->mm_stats.lock);

	/* Increase the amount of accumulated us. */
	time_us = ktime_to_us(ktime_get());
	increment_us = time_us - adev->mm_stats.last_update_us;
	adev->mm_stats.last_update_us = time_us;
	adev->mm_stats.accum_us = min(adev->mm_stats.accum_us + increment_us,
                                      us_upper_bound);

	/* This prevents the short period of low performance when the VRAM
	 * usage is low and the driver is in debt or doesn't have enough
	 * accumulated us to fill VRAM quickly.
318
	 *
319
320
321
322
	 * The situation can occur in these cases:
	 * - a lot of VRAM is freed by userspace
	 * - the presence of a big buffer causes a lot of evictions
	 *   (solution: split buffers into smaller ones)
323
	 *
324
325
	 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
	 * accum_us to a positive number.
326
	 */
327
328
329
330
331
332
333
334
335
336
337
338
339
	if (free_vram >= 128 * 1024 * 1024 || free_vram >= total_vram / 8) {
		s64 min_us;

		/* Be more aggresive on dGPUs. Try to fill a portion of free
		 * VRAM now.
		 */
		if (!(adev->flags & AMD_IS_APU))
			min_us = bytes_to_us(adev, free_vram / 4);
		else
			min_us = 0; /* Reset accum_us on APUs. */

		adev->mm_stats.accum_us = max(min_us, adev->mm_stats.accum_us);
	}
340

341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
	/* This returns 0 if the driver is in debt to disallow (optional)
	 * buffer moves.
	 */
	max_bytes = us_to_bytes(adev, adev->mm_stats.accum_us);

	spin_unlock(&adev->mm_stats.lock);
	return max_bytes;
}

/* Report how many bytes have really been moved for the last command
 * submission. This can result in a debt that can stop buffer migrations
 * temporarily.
 */
static void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev,
					 u64 num_bytes)
{
	spin_lock(&adev->mm_stats.lock);
	adev->mm_stats.accum_us -= bytes_to_us(adev, num_bytes);
	spin_unlock(&adev->mm_stats.lock);
360
361
}

362
363
364
static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
				 struct amdgpu_bo *bo)
{
365
	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
366
367
368
369
370
371
372
	u64 initial_bytes_moved;
	uint32_t domain;
	int r;

	if (bo->pin_count)
		return 0;

373
374
	/* Don't move this buffer if we have depleted our allowance
	 * to move it. Don't move anything if the threshold is zero.
375
	 */
376
	if (p->bytes_moved < p->bytes_moved_threshold)
377
378
379
380
381
382
		domain = bo->prefered_domains;
	else
		domain = bo->allowed_domains;

retry:
	amdgpu_ttm_placement_from_domain(bo, domain);
383
	initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
384
	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
385
	p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
386
387
		initial_bytes_moved;

388
389
390
	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
		domain = bo->allowed_domains;
		goto retry;
391
392
393
394
395
	}

	return r;
}

396
397
/* Last resort, try to evict something from the current working set */
static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
398
				struct amdgpu_bo *validated)
399
{
400
	uint32_t domain = validated->allowed_domains;
401
402
403
404
405
406
407
408
409
410
	int r;

	if (!p->evictable)
		return false;

	for (;&p->evictable->tv.head != &p->validated;
	     p->evictable = list_prev_entry(p->evictable, tv.head)) {

		struct amdgpu_bo_list_entry *candidate = p->evictable;
		struct amdgpu_bo *bo = candidate->robj;
411
		struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
412
413
414
415
		u64 initial_bytes_moved;
		uint32_t other;

		/* If we reached our current BO we can forget it */
416
		if (candidate->robj == validated)
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
			break;

		other = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);

		/* Check if this BO is in one of the domains we need space for */
		if (!(other & domain))
			continue;

		/* Check if we can move this BO somewhere else */
		other = bo->allowed_domains & ~domain;
		if (!other)
			continue;

		/* Good we can try to move this BO somewhere else */
		amdgpu_ttm_placement_from_domain(bo, other);
432
		initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
433
		r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
434
		p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
435
436
437
438
439
440
441
442
443
444
445
446
447
448
			initial_bytes_moved;

		if (unlikely(r))
			break;

		p->evictable = list_prev_entry(p->evictable, tv.head);
		list_move(&candidate->tv.head, &p->validated);

		return true;
	}

	return false;
}

449
450
451
452
453
454
455
456
457
458
459
460
static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo)
{
	struct amdgpu_cs_parser *p = param;
	int r;

	do {
		r = amdgpu_cs_bo_validate(p, bo);
	} while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo));
	if (r)
		return r;

	if (bo->shadow)
461
		r = amdgpu_cs_bo_validate(p, bo->shadow);
462
463
464
465

	return r;
}

466
static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
467
			    struct list_head *validated)
468
469
470
471
{
	struct amdgpu_bo_list_entry *lobj;
	int r;

472
	list_for_each_entry(lobj, validated, tv.head) {
473
		struct amdgpu_bo *bo = lobj->robj;
474
		bool binding_userptr = false;
475
		struct mm_struct *usermm;
476

477
478
479
480
		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
		if (usermm && usermm != current->mm)
			return -EPERM;

481
482
483
484
485
486
487
488
489
		/* Check if we have user pages and nobody bound the BO already */
		if (lobj->user_pages && bo->tbo.ttm->state != tt_bound) {
			size_t size = sizeof(struct page *);

			size *= bo->tbo.ttm->num_pages;
			memcpy(bo->tbo.ttm->pages, lobj->user_pages, size);
			binding_userptr = true;
		}

490
491
492
		if (p->evictable == lobj)
			p->evictable = NULL;

493
		r = amdgpu_cs_validate(p, bo);
494
		if (r)
495
			return r;
496

497
498
499
500
		if (binding_userptr) {
			drm_free_large(lobj->user_pages);
			lobj->user_pages = NULL;
		}
501
502
503
504
	}
	return 0;
}

505
506
static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
				union drm_amdgpu_cs *cs)
507
508
{
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
509
	struct amdgpu_bo_list_entry *e;
510
	struct list_head duplicates;
511
	bool need_mmap_lock = false;
512
	unsigned i, tries = 10;
513
	int r;
514

515
516
517
	INIT_LIST_HEAD(&p->validated);

	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
518
	if (p->bo_list) {
519
520
		need_mmap_lock = p->bo_list->first_userptr !=
			p->bo_list->num_entries;
521
		amdgpu_bo_list_get_list(p->bo_list, &p->validated);
522
	}
523

524
	INIT_LIST_HEAD(&duplicates);
525
	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
526

527
	if (p->uf_entry.robj)
528
529
		list_add(&p->uf_entry.tv.head, &p->validated);

530
531
532
	if (need_mmap_lock)
		down_read(&current->mm->mmap_sem);

533
534
535
536
537
538
	while (1) {
		struct list_head need_pages;
		unsigned i;

		r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
					   &duplicates);
539
		if (unlikely(r != 0)) {
540
541
			if (r != -ERESTARTSYS)
				DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
542
			goto error_free_pages;
543
		}
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582

		/* Without a BO list we don't have userptr BOs */
		if (!p->bo_list)
			break;

		INIT_LIST_HEAD(&need_pages);
		for (i = p->bo_list->first_userptr;
		     i < p->bo_list->num_entries; ++i) {

			e = &p->bo_list->array[i];

			if (amdgpu_ttm_tt_userptr_invalidated(e->robj->tbo.ttm,
				 &e->user_invalidated) && e->user_pages) {

				/* We acquired a page array, but somebody
				 * invalidated it. Free it an try again
				 */
				release_pages(e->user_pages,
					      e->robj->tbo.ttm->num_pages,
					      false);
				drm_free_large(e->user_pages);
				e->user_pages = NULL;
			}

			if (e->robj->tbo.ttm->state != tt_bound &&
			    !e->user_pages) {
				list_del(&e->tv.head);
				list_add(&e->tv.head, &need_pages);

				amdgpu_bo_unreserve(e->robj);
			}
		}

		if (list_empty(&need_pages))
			break;

		/* Unreserve everything again. */
		ttm_eu_backoff_reservation(&p->ticket, &p->validated);

583
		/* We tried too many times, just abort */
584
585
		if (!--tries) {
			r = -EDEADLK;
586
			DRM_ERROR("deadlock in %s\n", __func__);
587
588
589
590
591
592
593
594
595
596
597
			goto error_free_pages;
		}

		/* Fill the page arrays for all useptrs. */
		list_for_each_entry(e, &need_pages, tv.head) {
			struct ttm_tt *ttm = e->robj->tbo.ttm;

			e->user_pages = drm_calloc_large(ttm->num_pages,
							 sizeof(struct page*));
			if (!e->user_pages) {
				r = -ENOMEM;
598
				DRM_ERROR("calloc failure in %s\n", __func__);
599
600
601
602
603
				goto error_free_pages;
			}

			r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
			if (r) {
604
				DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
605
606
607
608
609
610
611
612
613
				drm_free_large(e->user_pages);
				e->user_pages = NULL;
				goto error_free_pages;
			}
		}

		/* And try again. */
		list_splice(&need_pages, &p->validated);
	}
614

615
616
	p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
	p->bytes_moved = 0;
617
618
619
	p->evictable = list_last_entry(&p->validated,
				       struct amdgpu_bo_list_entry,
				       tv.head);
620

621
622
623
624
625
626
627
	r = amdgpu_vm_validate_pt_bos(p->adev, &fpriv->vm,
				      amdgpu_cs_validate, p);
	if (r) {
		DRM_ERROR("amdgpu_vm_validate_pt_bos() failed.\n");
		goto error_validate;
	}

628
	r = amdgpu_cs_list_validate(p, &duplicates);
629
630
	if (r) {
		DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
631
		goto error_validate;
632
	}
633

634
	r = amdgpu_cs_list_validate(p, &p->validated);
635
636
	if (r) {
		DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
637
		goto error_validate;
638
	}
639

640
641
	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved);

642
643
644
	fpriv->vm.last_eviction_counter =
		atomic64_read(&p->adev->num_evictions);

645
	if (p->bo_list) {
646
647
648
		struct amdgpu_bo *gds = p->bo_list->gds_obj;
		struct amdgpu_bo *gws = p->bo_list->gws_obj;
		struct amdgpu_bo *oa = p->bo_list->oa_obj;
649
650
651
652
653
654
655
656
		struct amdgpu_vm *vm = &fpriv->vm;
		unsigned i;

		for (i = 0; i < p->bo_list->num_entries; i++) {
			struct amdgpu_bo *bo = p->bo_list->array[i].robj;

			p->bo_list->array[i].bo_va = amdgpu_vm_bo_find(vm, bo);
		}
657
658
659
660
661
662
663
664
665
666
667
668
669

		if (gds) {
			p->job->gds_base = amdgpu_bo_gpu_offset(gds);
			p->job->gds_size = amdgpu_bo_size(gds);
		}
		if (gws) {
			p->job->gws_base = amdgpu_bo_gpu_offset(gws);
			p->job->gws_size = amdgpu_bo_size(gws);
		}
		if (oa) {
			p->job->oa_base = amdgpu_bo_gpu_offset(oa);
			p->job->oa_size = amdgpu_bo_size(oa);
		}
670
	}
671

672
673
674
	if (!r && p->uf_entry.robj) {
		struct amdgpu_bo *uf = p->uf_entry.robj;

675
		r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
676
677
		p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
	}
678

679
error_validate:
680
681
	if (r) {
		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
682
		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
683
	}
684

685
686
error_free_pages:

687
688
689
	if (need_mmap_lock)
		up_read(&current->mm->mmap_sem);

690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
	if (p->bo_list) {
		for (i = p->bo_list->first_userptr;
		     i < p->bo_list->num_entries; ++i) {
			e = &p->bo_list->array[i];

			if (!e->user_pages)
				continue;

			release_pages(e->user_pages,
				      e->robj->tbo.ttm->num_pages,
				      false);
			drm_free_large(e->user_pages);
		}
	}

705
706
707
708
709
710
711
712
713
714
	return r;
}

static int amdgpu_cs_sync_rings(struct amdgpu_cs_parser *p)
{
	struct amdgpu_bo_list_entry *e;
	int r;

	list_for_each_entry(e, &p->validated, tv.head) {
		struct reservation_object *resv = e->robj->tbo.resv;
715
		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
716
717
718
719
720
721
722

		if (r)
			return r;
	}
	return 0;
}

723
724
725
726
727
728
729
730
731
/**
 * cs_parser_fini() - clean parser states
 * @parser:	parser structure holding parsing context.
 * @error:	error number
 *
 * If error is set than unvalidate buffer, otherwise just free memory
 * used by parsing context.
 **/
static void amdgpu_cs_parser_fini(struct amdgpu_cs_parser *parser, int error, bool backoff)
732
{
733
	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
734
735
	unsigned i;

736
	if (!error) {
737
738
		amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);

739
		ttm_eu_fence_buffer_objects(&parser->ticket,
740
741
					    &parser->validated,
					    parser->fence);
742
743
744
745
	} else if (backoff) {
		ttm_eu_backoff_reservation(&parser->ticket,
					   &parser->validated);
	}
746
	dma_fence_put(parser->fence);
747

748
749
	if (parser->ctx)
		amdgpu_ctx_put(parser->ctx);
750
751
752
	if (parser->bo_list)
		amdgpu_bo_list_put(parser->bo_list);

753
754
755
	for (i = 0; i < parser->nchunks; i++)
		drm_free_large(parser->chunks[i].kdata);
	kfree(parser->chunks);
756
757
	if (parser->job)
		amdgpu_job_free(parser->job);
758
	amdgpu_bo_unref(&parser->uf_entry.robj);
759
760
761
762
763
764
765
766
767
768
769
770
771
772
}

static int amdgpu_bo_vm_update_pte(struct amdgpu_cs_parser *p,
				   struct amdgpu_vm *vm)
{
	struct amdgpu_device *adev = p->adev;
	struct amdgpu_bo_va *bo_va;
	struct amdgpu_bo *bo;
	int i, r;

	r = amdgpu_vm_update_page_directory(adev, vm);
	if (r)
		return r;

773
	r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
774
775
776
	if (r)
		return r;

777
778
779
780
781
782
	r = amdgpu_vm_clear_freed(adev, vm);
	if (r)
		return r;

	if (p->bo_list) {
		for (i = 0; i < p->bo_list->num_entries; i++) {
783
			struct dma_fence *f;
784

785
786
787
788
789
790
791
792
793
			/* ignore duplicates */
			bo = p->bo_list->array[i].robj;
			if (!bo)
				continue;

			bo_va = p->bo_list->array[i].bo_va;
			if (bo_va == NULL)
				continue;

794
			r = amdgpu_vm_bo_update(adev, bo_va, false);
795
796
797
			if (r)
				return r;

798
			f = bo_va->last_pt_update;
799
			r = amdgpu_sync_fence(adev, &p->job->sync, f);
800
801
			if (r)
				return r;
802
		}
803
804
805

	}

806
	r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
807
808
809
810
811
812
813
814
815
816
817

	if (amdgpu_vm_debug && p->bo_list) {
		/* Invalidate all BOs to test for userspace bugs */
		for (i = 0; i < p->bo_list->num_entries; i++) {
			/* ignore duplicates */
			bo = p->bo_list->array[i].robj;
			if (!bo)
				continue;

			amdgpu_vm_bo_invalidate(adev, bo);
		}
818
819
	}

820
	return r;
821
822
823
}

static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
824
				 struct amdgpu_cs_parser *p)
825
{
826
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
827
	struct amdgpu_vm *vm = &fpriv->vm;
828
	struct amdgpu_ring *ring = p->job->ring;
829
830
831
	int i, r;

	/* Only for UVD/VCE VM emulation */
832
833
834
	if (ring->funcs->parse_cs) {
		for (i = 0; i < p->job->num_ibs; i++) {
			r = amdgpu_ring_parse_cs(ring, p, i);
835
836
837
			if (r)
				return r;
		}
838
839
840
	}

	if (p->job->vm) {
841
		p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
842

843
844
845
846
		r = amdgpu_bo_vm_update_pte(p, vm);
		if (r)
			return r;
	}
847

848
	return amdgpu_cs_sync_rings(p);
849
850
851
852
853
854
855
856
857
858
}

static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
			     struct amdgpu_cs_parser *parser)
{
	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
	struct amdgpu_vm *vm = &fpriv->vm;
	int i, j;
	int r;

859
	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
860
861
862
863
864
865
		struct amdgpu_cs_chunk *chunk;
		struct amdgpu_ib *ib;
		struct drm_amdgpu_cs_chunk_ib *chunk_ib;
		struct amdgpu_ring *ring;

		chunk = &parser->chunks[i];
866
		ib = &parser->job->ibs[j];
867
868
869
870
871
872
873
874
		chunk_ib = (struct drm_amdgpu_cs_chunk_ib *)chunk->kdata;

		if (chunk->chunk_id != AMDGPU_CHUNK_ID_IB)
			continue;

		r = amdgpu_cs_get_ring(adev, chunk_ib->ip_type,
				       chunk_ib->ip_instance, chunk_ib->ring,
				       &ring);
875
		if (r)
876
877
			return r;

878
879
880
881
882
883
884
885
		if (ib->flags & AMDGPU_IB_FLAG_PREAMBLE) {
			parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT;
			if (!parser->ctx->preamble_presented) {
				parser->job->preamble_status |= AMDGPU_PREAMBLE_IB_PRESENT_FIRST;
				parser->ctx->preamble_presented = true;
			}
		}

886
887
888
889
890
		if (parser->job->ring && parser->job->ring != ring)
			return -EINVAL;

		parser->job->ring = ring;

891
		if (ring->funcs->parse_cs) {
892
			struct amdgpu_bo_va_mapping *m;
893
			struct amdgpu_bo *aobj = NULL;
894
895
			uint64_t offset;
			uint8_t *kptr;
896

897
898
			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
						   &aobj);
899
900
901
			if (!aobj) {
				DRM_ERROR("IB va_start is invalid\n");
				return -EINVAL;
902
903
			}

904
905
906
907
908
909
			if ((chunk_ib->va_start + chunk_ib->ib_bytes) >
			    (m->it.last + 1) * AMDGPU_GPU_PAGE_SIZE) {
				DRM_ERROR("IB va_start+ib_bytes is invalid\n");
				return -EINVAL;
			}

910
			/* the IB should be reserved at this point */
911
			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
912
913
914
915
			if (r) {
				return r;
			}

916
917
918
			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
			kptr += chunk_ib->va_start - offset;

919
			r =  amdgpu_ib_get(adev, vm, chunk_ib->ib_bytes, ib);
920
921
922
923
924
925
926
927
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
			amdgpu_bo_kunmap(aobj);
		} else {
928
			r =  amdgpu_ib_get(adev, vm, 0, ib);
929
930
931
932
933
934
935
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

		}

936
		ib->gpu_addr = chunk_ib->va_start;
937
		ib->length_dw = chunk_ib->ib_bytes / 4;
938
		ib->flags = chunk_ib->flags;
939
940
941
		j++;
	}

942
	/* UVD & VCE fw doesn't support user fences */
943
	if (parser->job->uf_addr && (
944
945
	    parser->job->ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
	    parser->job->ring->funcs->type == AMDGPU_RING_TYPE_VCE))
946
		return -EINVAL;
947
948
949
950

	return 0;
}

951
952
953
static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
				  struct amdgpu_cs_parser *p)
{
954
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
	int i, j, r;

	for (i = 0; i < p->nchunks; ++i) {
		struct drm_amdgpu_cs_chunk_dep *deps;
		struct amdgpu_cs_chunk *chunk;
		unsigned num_deps;

		chunk = &p->chunks[i];

		if (chunk->chunk_id != AMDGPU_CHUNK_ID_DEPENDENCIES)
			continue;

		deps = (struct drm_amdgpu_cs_chunk_dep *)chunk->kdata;
		num_deps = chunk->length_dw * 4 /
			sizeof(struct drm_amdgpu_cs_chunk_dep);

		for (j = 0; j < num_deps; ++j) {
			struct amdgpu_ring *ring;
973
			struct amdgpu_ctx *ctx;
974
			struct dma_fence *fence;
975
976
977
978
979
980
981

			r = amdgpu_cs_get_ring(adev, deps[j].ip_type,
					       deps[j].ip_instance,
					       deps[j].ring, &ring);
			if (r)
				return r;

982
983
984
985
			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
			if (ctx == NULL)
				return -EINVAL;

986
987
988
989
			fence = amdgpu_ctx_get_fence(ctx, ring,
						     deps[j].handle);
			if (IS_ERR(fence)) {
				r = PTR_ERR(fence);
990
				amdgpu_ctx_put(ctx);
991
				return r;
992

993
			} else if (fence) {
994
995
				r = amdgpu_sync_fence(adev, &p->job->sync,
						      fence);
996
				dma_fence_put(fence);
997
998
999
1000
				amdgpu_ctx_put(ctx);
				if (r)
					return r;
			}
1001
1002
1003
1004
1005
1006
		}
	}

	return 0;
}

1007
1008
1009
static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
			    union drm_amdgpu_cs *cs)
{
1010
	struct amdgpu_ring *ring = p->job->ring;
1011
	struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1012
	struct amdgpu_job *job;
1013
	int r;
1014

1015
1016
	job = p->job;
	p->job = NULL;
1017

1018
	r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1019
	if (r) {
1020
		amdgpu_job_free(job);
1021
		return r;
1022
1023
	}

1024
	job->owner = p->filp;
1025
	job->fence_ctx = entity->fence_context;
1026
	p->fence = dma_fence_get(&job->base.s_fence->finished);
1027
	cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
1028
	job->uf_sequence = cs->out.handle;
1029
	amdgpu_job_free_resources(job);
1030
1031
1032
1033
1034
1035
1036

	trace_amdgpu_cs_ioctl(job);
	amd_sched_entity_push_job(&job->base);

	return 0;
}

1037
1038
1039
1040
int amdgpu_cs_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)
{
	struct amdgpu_device *adev = dev->dev_private;
	union drm_amdgpu_cs *cs = data;
1041
	struct amdgpu_cs_parser parser = {};
1042
1043
	bool reserved_buffers = false;
	int i, r;
1044

1045
	if (!adev->accel_working)
1046
		return -EBUSY;
1047

1048
1049
1050
1051
	parser.adev = adev;
	parser.filp = filp;

	r = amdgpu_cs_parser_init(&parser, data);
1052
	if (r) {
1053
		DRM_ERROR("Failed to initialize parser !\n");
1054
		goto out;
1055
1056
	}

1057
1058
1059
1060
1061
1062
1063
	r = amdgpu_cs_parser_bos(&parser, data);
	if (r) {
		if (r == -ENOMEM)
			DRM_ERROR("Not enough memory for command submission!\n");
		else if (r != -ERESTARTSYS)
			DRM_ERROR("Failed to process the buffer list %d!\n", r);
		goto out;
1064
1065
	}

1066
1067
	reserved_buffers = true;
	r = amdgpu_cs_ib_fill(adev, &parser);
1068
1069
1070
	if (r)
		goto out;

1071
1072
1073
1074
1075
1076
	r = amdgpu_cs_dependencies(adev, &parser);
	if (r) {
		DRM_ERROR("Failed in the dependencies handling %d!\n", r);
		goto out;
	}

1077
	for (i = 0; i < parser.job->num_ibs; i++)
1078
		trace_amdgpu_cs(&parser, i);
1079

1080
	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1081
1082
1083
	if (r)
		goto out;

1084
	r = amdgpu_cs_submit(&parser, cs);
1085
1086

out:
1087
	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
	return r;
}

/**
 * amdgpu_cs_wait_ioctl - wait for a command submission to finish
 *
 * @dev: drm device
 * @data: data from userspace
 * @filp: file private
 *
 * Wait for the command submission identified by handle to finish.
 */
int amdgpu_cs_wait_ioctl(struct drm_device *dev, void *data,
			 struct drm_file *filp)
{
	union drm_amdgpu_wait_cs *wait = data;
	struct amdgpu_device *adev = dev->dev_private;
	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout);
1106
	struct amdgpu_ring *ring = NULL;
1107
	struct amdgpu_ctx *ctx;
1108
	struct dma_fence *fence;
1109
1110
	long r;

1111
1112
1113
1114
1115
	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
			       wait->in.ring, &ring);
	if (r)
		return r;

1116
1117
1118
	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
	if (ctx == NULL)
		return -EINVAL;
1119

1120
1121
1122
1123
	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
	if (IS_ERR(fence))
		r = PTR_ERR(fence);
	else if (fence) {
1124
1125
		r = dma_fence_wait_timeout(fence, true, timeout);
		dma_fence_put(fence);
1126
1127
	} else
		r = 1;
1128

1129
	amdgpu_ctx_put(ctx);
1130
1131
1132
1133
1134
1135
1136
1137
1138
	if (r < 0)
		return r;

	memset(wait, 0, sizeof(*wait));
	wait->out.status = (r == 0);

	return 0;
}

1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
/**
 * amdgpu_cs_get_fence - helper to get fence from drm_amdgpu_fence
 *
 * @adev: amdgpu device
 * @filp: file private
 * @user: drm_amdgpu_fence copied from user space
 */
static struct dma_fence *amdgpu_cs_get_fence(struct amdgpu_device *adev,
					     struct drm_file *filp,
					     struct drm_amdgpu_fence *user)
{
	struct amdgpu_ring *ring;
	struct amdgpu_ctx *ctx;
	struct dma_fence *fence;
	int r;

	r = amdgpu_cs_get_ring(adev, user->ip_type, user->ip_instance,
			       user->ring, &ring);
	if (r)
		return ERR_PTR(r);

	ctx = amdgpu_ctx_get(filp->driver_priv, user->ctx_id);
	if (ctx == NULL)
		return ERR_PTR(-EINVAL);

	fence = amdgpu_ctx_get_fence(ctx, ring, user->seq_no);
	amdgpu_ctx_put(ctx);

	return fence;
}

/**
 * amdgpu_cs_wait_all_fence - wait on all fences to signal
 *
 * @adev: amdgpu device
 * @filp: file private
 * @wait: wait parameters
 * @fences: array of drm_amdgpu_fence
 */
static int amdgpu_cs_wait_all_fences(struct amdgpu_device *adev,
				     struct drm_file *filp,
				     union drm_amdgpu_wait_fences *wait,
				     struct drm_amdgpu_fence *fences)
{
	uint32_t fence_count = wait->in.fence_count;
	unsigned int i;
	long r = 1;

	for (i = 0; i < fence_count; i++) {
		struct dma_fence *fence;
		unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);

		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
		if (IS_ERR(fence))
			return PTR_ERR(fence);
		else if (!fence)
			continue;

		r = dma_fence_wait_timeout(fence, true, timeout);
		if (r < 0)
			return r;

		if (r == 0)
			break;
	}

	memset(wait, 0, sizeof(*wait));
	wait->out.status = (r > 0);

	return 0;
}

/**
 * amdgpu_cs_wait_any_fence - wait on any fence to signal
 *
 * @adev: amdgpu device
 * @filp: file private
 * @wait: wait parameters
 *