amdgpu_cs.c 24.4 KB
Newer Older
A
Alex Deucher 已提交
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>
A
Alex Deucher 已提交
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:
A
Alex Deucher 已提交
66 67
		if (ring < adev->sdma.num_instances) {
			*out_ring = &adev->sdma.instance[ring].ring;
A
Alex Deucher 已提交
68
		} else {
A
Alex Deucher 已提交
69 70
			DRM_ERROR("only %d SDMA rings are supported\n",
				  adev->sdma.num_instances);
A
Alex Deucher 已提交
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88
			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;
	}
	return 0;
}

89
static int amdgpu_cs_user_fence_chunk(struct amdgpu_cs_parser *p,
90 91
				      struct drm_amdgpu_cs_chunk_fence *data,
				      uint32_t *offset)
92 93 94
{
	struct drm_gem_object *gobj;

95
	gobj = drm_gem_object_lookup(p->filp, data->handle);
96 97 98
	if (gobj == NULL)
		return -EINVAL;

99
	p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
100 101 102
	p->uf_entry.priority = 0;
	p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
	p->uf_entry.tv.shared = true;
103
	p->uf_entry.user_pages = NULL;
104
	*offset = data->offset;
105 106

	drm_gem_object_unreference_unlocked(gobj);
107 108 109 110 111 112

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

113 114 115
	return 0;
}

A
Alex Deucher 已提交
116 117
int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
{
118
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
119
	struct amdgpu_vm *vm = &fpriv->vm;
A
Alex Deucher 已提交
120 121
	union drm_amdgpu_cs *cs = data;
	uint64_t *chunk_array_user;
122
	uint64_t *chunk_array;
123
	unsigned size, num_ibs = 0;
124
	uint32_t uf_offset = 0;
125
	int i;
126
	int ret;
A
Alex Deucher 已提交
127

128 129 130 131 132 133
	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;
A
Alex Deucher 已提交
134

135 136
	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
	if (!p->ctx) {
137 138
		ret = -EINVAL;
		goto free_chunk;
139
	}
140

A
Alex Deucher 已提交
141
	/* get chunks */
142
	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
A
Alex Deucher 已提交
143 144
	if (copy_from_user(chunk_array, chunk_array_user,
			   sizeof(uint64_t)*cs->in.num_chunks)) {
145
		ret = -EFAULT;
146
		goto put_ctx;
A
Alex Deucher 已提交
147 148 149
	}

	p->nchunks = cs->in.num_chunks;
150
	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
A
Alex Deucher 已提交
151
			    GFP_KERNEL);
152 153
	if (!p->chunks) {
		ret = -ENOMEM;
154
		goto put_ctx;
A
Alex Deucher 已提交
155 156 157 158 159 160 161
	}

	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;

162
		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
A
Alex Deucher 已提交
163 164
		if (copy_from_user(&user_chunk, chunk_ptr,
				       sizeof(struct drm_amdgpu_cs_chunk))) {
165 166 167
			ret = -EFAULT;
			i--;
			goto free_partial_kdata;
A
Alex Deucher 已提交
168 169 170 171 172
		}
		p->chunks[i].chunk_id = user_chunk.chunk_id;
		p->chunks[i].length_dw = user_chunk.length_dw;

		size = p->chunks[i].length_dw;
173
		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
A
Alex Deucher 已提交
174 175 176

		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
		if (p->chunks[i].kdata == NULL) {
177 178 179
			ret = -ENOMEM;
			i--;
			goto free_partial_kdata;
A
Alex Deucher 已提交
180 181 182
		}
		size *= sizeof(uint32_t);
		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
183 184
			ret = -EFAULT;
			goto free_partial_kdata;
A
Alex Deucher 已提交
185 186
		}

187 188
		switch (p->chunks[i].chunk_id) {
		case AMDGPU_CHUNK_ID_IB:
189
			++num_ibs;
190 191 192
			break;

		case AMDGPU_CHUNK_ID_FENCE:
A
Alex Deucher 已提交
193
			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
194
			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
195 196
				ret = -EINVAL;
				goto free_partial_kdata;
A
Alex Deucher 已提交
197
			}
198

199 200
			ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
							 &uf_offset);
201 202 203
			if (ret)
				goto free_partial_kdata;

204 205
			break;

206 207 208
		case AMDGPU_CHUNK_ID_DEPENDENCIES:
			break;

209
		default:
210 211
			ret = -EINVAL;
			goto free_partial_kdata;
A
Alex Deucher 已提交
212 213 214
		}
	}

215
	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
216
	if (ret)
C
Christian König 已提交
217
		goto free_all_kdata;
A
Alex Deucher 已提交
218

219 220 221 222
	if (p->uf_entry.robj) {
		p->job->uf_bo = amdgpu_bo_ref(p->uf_entry.robj);
		p->job->uf_offset = uf_offset;
	}
223

A
Alex Deucher 已提交
224
	kfree(chunk_array);
225 226 227 228 229 230 231 232
	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);
233
put_ctx:
234 235 236 237 238
	amdgpu_ctx_put(p->ctx);
free_chunk:
	kfree(chunk_array);

	return ret;
A
Alex Deucher 已提交
239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
}

/* Returns how many bytes TTM can move per IB.
 */
static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
{
	u64 real_vram_size = adev->mc.real_vram_size;
	u64 vram_usage = atomic64_read(&adev->vram_usage);

	/* This function is based on the current VRAM usage.
	 *
	 * - If all of VRAM is free, allow relocating the number of bytes that
	 *   is equal to 1/4 of the size of VRAM for this IB.

	 * - If more than one half of VRAM is occupied, only allow relocating
	 *   1 MB of data for this IB.
	 *
	 * - From 0 to one half of used VRAM, the threshold decreases
	 *   linearly.
	 *         __________________
	 * 1/4 of -|\               |
	 * VRAM    | \              |
	 *         |  \             |
	 *         |   \            |
	 *         |    \           |
	 *         |     \          |
	 *         |      \         |
	 *         |       \________|1 MB
	 *         |----------------|
	 *    VRAM 0 %             100 %
	 *         used            used
	 *
	 * Note: It's a threshold, not a limit. The threshold must be crossed
	 * for buffer relocations to stop, so any buffer of an arbitrary size
	 * can be moved as long as the threshold isn't crossed before
	 * the relocation takes place. We don't want to disable buffer
	 * relocations completely.
	 *
	 * The idea is that buffers should be placed in VRAM at creation time
	 * and TTM should only do a minimum number of relocations during
	 * command submission. In practice, you need to submit at least
	 * a dozen IBs to move all buffers to VRAM if they are in GTT.
	 *
	 * Also, things can get pretty crazy under memory pressure and actual
	 * VRAM usage can change a lot, so playing safe even at 50% does
	 * consistently increase performance.
	 */

	u64 half_vram = real_vram_size >> 1;
	u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage;
	u64 bytes_moved_threshold = half_free_vram >> 1;
	return max(bytes_moved_threshold, 1024*1024ull);
}

293
int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
294
			    struct list_head *validated)
A
Alex Deucher 已提交
295 296
{
	struct amdgpu_bo_list_entry *lobj;
297
	u64 initial_bytes_moved;
A
Alex Deucher 已提交
298 299
	int r;

300
	list_for_each_entry(lobj, validated, tv.head) {
301
		struct amdgpu_bo *bo = lobj->robj;
302
		bool binding_userptr = false;
303
		struct mm_struct *usermm;
304
		uint32_t domain;
A
Alex Deucher 已提交
305

306 307 308 309
		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
		if (usermm && usermm != current->mm)
			return -EPERM;

310 311 312 313 314 315 316 317 318
		/* 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;
		}

319 320 321 322 323 324 325 326 327 328 329 330
		if (bo->pin_count)
			continue;

		/* Avoid moving this one if we have moved too many buffers
		 * for this IB already.
		 *
		 * Note that this allows moving at least one buffer of
		 * any size, because it doesn't take the current "bo"
		 * into account. We don't want to disallow buffer moves
		 * completely.
		 */
		if (p->bytes_moved <= p->bytes_moved_threshold)
331
			domain = bo->prefered_domains;
332
		else
333
			domain = bo->allowed_domains;
334 335 336 337 338 339 340 341 342

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

		if (unlikely(r)) {
343 344
			if (r != -ERESTARTSYS && domain != bo->allowed_domains) {
				domain = bo->allowed_domains;
345
				goto retry;
A
Alex Deucher 已提交
346
			}
347
			return r;
A
Alex Deucher 已提交
348
		}
349 350 351 352 353

		if (binding_userptr) {
			drm_free_large(lobj->user_pages);
			lobj->user_pages = NULL;
		}
A
Alex Deucher 已提交
354 355 356 357
	}
	return 0;
}

358 359
static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
				union drm_amdgpu_cs *cs)
A
Alex Deucher 已提交
360 361
{
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
362
	struct amdgpu_bo_list_entry *e;
363
	struct list_head duplicates;
364
	bool need_mmap_lock = false;
365
	unsigned i, tries = 10;
366
	int r;
A
Alex Deucher 已提交
367

368 369 370
	INIT_LIST_HEAD(&p->validated);

	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
371
	if (p->bo_list) {
372 373
		need_mmap_lock = p->bo_list->first_userptr !=
			p->bo_list->num_entries;
374
		amdgpu_bo_list_get_list(p->bo_list, &p->validated);
375
	}
A
Alex Deucher 已提交
376

377
	INIT_LIST_HEAD(&duplicates);
378
	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
A
Alex Deucher 已提交
379

380
	if (p->uf_entry.robj)
381 382
		list_add(&p->uf_entry.tv.head, &p->validated);

A
Alex Deucher 已提交
383 384 385
	if (need_mmap_lock)
		down_read(&current->mm->mmap_sem);

386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
	while (1) {
		struct list_head need_pages;
		unsigned i;

		r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
					   &duplicates);
		if (unlikely(r != 0))
			goto error_free_pages;

		/* 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);

		/* We tried to often, just abort */
		if (!--tries) {
			r = -EDEADLK;
			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;
				goto error_free_pages;
			}

			r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
			if (r) {
				drm_free_large(e->user_pages);
				e->user_pages = NULL;
				goto error_free_pages;
			}
		}

		/* And try again. */
		list_splice(&need_pages, &p->validated);
	}
461

462
	amdgpu_vm_get_pt_bos(p->adev, &fpriv->vm, &duplicates);
463

464 465 466 467
	p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
	p->bytes_moved = 0;

	r = amdgpu_cs_list_validate(p, &duplicates);
468 469 470
	if (r)
		goto error_validate;

471
	r = amdgpu_cs_list_validate(p, &p->validated);
472 473 474
	if (r)
		goto error_validate;

475 476 477
	fpriv->vm.last_eviction_counter =
		atomic64_read(&p->adev->num_evictions);

478
	if (p->bo_list) {
479 480 481
		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;
482 483 484 485 486 487 488 489
		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);
		}
490 491 492 493 494 495 496 497 498 499 500 501 502

		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);
		}
503
	}
504 505

error_validate:
506 507
	if (r) {
		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
508
		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
509
	}
A
Alex Deucher 已提交
510

511 512
error_free_pages:

A
Alex Deucher 已提交
513 514 515
	if (need_mmap_lock)
		up_read(&current->mm->mmap_sem);

516 517 518 519 520 521 522 523 524 525 526 527 528 529 530
	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);
		}
	}

A
Alex Deucher 已提交
531 532 533 534 535 536 537 538 539 540
	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;
541
		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
A
Alex Deucher 已提交
542 543 544 545 546 547 548

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

549 550 551 552 553 554 555 556 557
/**
 * 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)
C
Chunming Zhou 已提交
558
{
559
	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
560 561
	unsigned i;

A
Alex Deucher 已提交
562
	if (!error) {
563 564
		amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);

A
Alex Deucher 已提交
565
		ttm_eu_fence_buffer_objects(&parser->ticket,
566 567
					    &parser->validated,
					    parser->fence);
A
Alex Deucher 已提交
568 569 570 571
	} else if (backoff) {
		ttm_eu_backoff_reservation(&parser->ticket,
					   &parser->validated);
	}
572
	fence_put(parser->fence);
573

574 575
	if (parser->ctx)
		amdgpu_ctx_put(parser->ctx);
576 577 578
	if (parser->bo_list)
		amdgpu_bo_list_put(parser->bo_list);

A
Alex Deucher 已提交
579 580 581
	for (i = 0; i < parser->nchunks; i++)
		drm_free_large(parser->chunks[i].kdata);
	kfree(parser->chunks);
582 583
	if (parser->job)
		amdgpu_job_free(parser->job);
584
	amdgpu_bo_unref(&parser->uf_entry.robj);
A
Alex Deucher 已提交
585 586 587 588 589 590 591 592 593 594 595 596 597 598
}

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;

599
	r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
600 601 602
	if (r)
		return r;

A
Alex Deucher 已提交
603 604 605 606 607 608
	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++) {
609 610
			struct fence *f;

A
Alex Deucher 已提交
611 612 613 614 615 616 617 618 619 620 621 622 623
			/* 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;

			r = amdgpu_vm_bo_update(adev, bo_va, &bo->tbo.mem);
			if (r)
				return r;

624
			f = bo_va->last_pt_update;
625
			r = amdgpu_sync_fence(adev, &p->job->sync, f);
626 627
			if (r)
				return r;
A
Alex Deucher 已提交
628
		}
629 630 631

	}

632
	r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
633 634 635 636 637 638 639 640 641 642 643

	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);
		}
A
Alex Deucher 已提交
644 645
	}

646
	return r;
A
Alex Deucher 已提交
647 648 649
}

static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
650
				 struct amdgpu_cs_parser *p)
A
Alex Deucher 已提交
651
{
652
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
A
Alex Deucher 已提交
653
	struct amdgpu_vm *vm = &fpriv->vm;
654
	struct amdgpu_ring *ring = p->job->ring;
A
Alex Deucher 已提交
655 656 657
	int i, r;

	/* Only for UVD/VCE VM emulation */
658 659 660
	if (ring->funcs->parse_cs) {
		for (i = 0; i < p->job->num_ibs; i++) {
			r = amdgpu_ring_parse_cs(ring, p, i);
A
Alex Deucher 已提交
661 662 663 664 665
			if (r)
				return r;
		}
	}

666 667
	p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);

668
	r = amdgpu_bo_vm_update_pte(p, vm);
669
	if (!r)
670
		amdgpu_cs_sync_rings(p);
A
Alex Deucher 已提交
671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692

	return r;
}

static int amdgpu_cs_handle_lockup(struct amdgpu_device *adev, int r)
{
	if (r == -EDEADLK) {
		r = amdgpu_gpu_reset(adev);
		if (!r)
			r = -EAGAIN;
	}
	return r;
}

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;

693
	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
A
Alex Deucher 已提交
694 695 696 697 698 699
		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];
700
		ib = &parser->job->ibs[j];
A
Alex Deucher 已提交
701 702 703 704 705 706 707 708
		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);
709
		if (r)
A
Alex Deucher 已提交
710 711
			return r;

712 713 714 715 716
		if (parser->job->ring && parser->job->ring != ring)
			return -EINVAL;

		parser->job->ring = ring;

A
Alex Deucher 已提交
717
		if (ring->funcs->parse_cs) {
718
			struct amdgpu_bo_va_mapping *m;
719
			struct amdgpu_bo *aobj = NULL;
720 721
			uint64_t offset;
			uint8_t *kptr;
722

723 724
			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
						   &aobj);
725 726 727
			if (!aobj) {
				DRM_ERROR("IB va_start is invalid\n");
				return -EINVAL;
A
Alex Deucher 已提交
728 729
			}

730 731 732 733 734 735
			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;
			}

736
			/* the IB should be reserved at this point */
737
			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
A
Alex Deucher 已提交
738 739 740 741
			if (r) {
				return r;
			}

742 743 744
			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
			kptr += chunk_ib->va_start - offset;

745
			r =  amdgpu_ib_get(adev, NULL, chunk_ib->ib_bytes, ib);
A
Alex Deucher 已提交
746 747 748 749 750 751 752 753
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
			amdgpu_bo_kunmap(aobj);
		} else {
754
			r =  amdgpu_ib_get(adev, vm, 0, ib);
A
Alex Deucher 已提交
755 756 757 758 759 760 761 762
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

			ib->gpu_addr = chunk_ib->va_start;
		}

763
		ib->length_dw = chunk_ib->ib_bytes / 4;
764
		ib->flags = chunk_ib->flags;
A
Alex Deucher 已提交
765 766 767
		j++;
	}

768 769 770 771 772
	/* UVD & VCE fw doesn't support user fences */
	if (parser->job->uf_bo && (
	    parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
	    parser->job->ring->type == AMDGPU_RING_TYPE_VCE))
		return -EINVAL;
A
Alex Deucher 已提交
773 774 775 776

	return 0;
}

777 778 779
static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
				  struct amdgpu_cs_parser *p)
{
780
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798
	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;
799
			struct amdgpu_ctx *ctx;
800
			struct fence *fence;
801 802 803 804 805 806 807

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

808 809 810 811
			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
			if (ctx == NULL)
				return -EINVAL;

812 813 814 815
			fence = amdgpu_ctx_get_fence(ctx, ring,
						     deps[j].handle);
			if (IS_ERR(fence)) {
				r = PTR_ERR(fence);
816
				amdgpu_ctx_put(ctx);
817
				return r;
818

819
			} else if (fence) {
820 821
				r = amdgpu_sync_fence(adev, &p->job->sync,
						      fence);
822 823 824 825 826
				fence_put(fence);
				amdgpu_ctx_put(ctx);
				if (r)
					return r;
			}
827 828 829 830 831 832
		}
	}

	return 0;
}

833 834 835
static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
			    union drm_amdgpu_cs *cs)
{
836
	struct amdgpu_ring *ring = p->job->ring;
837
	struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
838
	struct fence *fence;
839
	struct amdgpu_job *job;
840
	int r;
841

842 843
	job = p->job;
	p->job = NULL;
844

845
	r = amd_sched_job_init(&job->base, &ring->sched,
846
			       entity, p->filp, &fence);
847
	if (r) {
848
		amdgpu_job_free(job);
849
		return r;
850 851
	}

852
	job->owner = p->filp;
853
	job->ctx = entity->fence_context;
854 855
	p->fence = fence_get(fence);
	cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, fence);
856
	job->uf_sequence = cs->out.handle;
857 858 859 860 861 862 863

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

	return 0;
}

C
Chunming Zhou 已提交
864 865 866 867
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;
868
	struct amdgpu_cs_parser parser = {};
869 870
	bool reserved_buffers = false;
	int i, r;
C
Chunming Zhou 已提交
871

872
	if (!adev->accel_working)
C
Chunming Zhou 已提交
873
		return -EBUSY;
874

875 876 877 878
	parser.adev = adev;
	parser.filp = filp;

	r = amdgpu_cs_parser_init(&parser, data);
A
Alex Deucher 已提交
879
	if (r) {
C
Chunming Zhou 已提交
880
		DRM_ERROR("Failed to initialize parser !\n");
881
		amdgpu_cs_parser_fini(&parser, r, false);
A
Alex Deucher 已提交
882 883 884
		r = amdgpu_cs_handle_lockup(adev, r);
		return r;
	}
885
	r = amdgpu_cs_parser_bos(&parser, data);
886 887 888 889 890 891
	if (r == -ENOMEM)
		DRM_ERROR("Not enough memory for command submission!\n");
	else if (r && r != -ERESTARTSYS)
		DRM_ERROR("Failed to process the buffer list %d!\n", r);
	else if (!r) {
		reserved_buffers = true;
892
		r = amdgpu_cs_ib_fill(adev, &parser);
893 894 895
	}

	if (!r) {
896
		r = amdgpu_cs_dependencies(adev, &parser);
897 898 899 900 901 902 903
		if (r)
			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
	}

	if (r)
		goto out;

904
	for (i = 0; i < parser.job->num_ibs; i++)
905
		trace_amdgpu_cs(&parser, i);
906

907
	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
908 909 910
	if (r)
		goto out;

C
Christian König 已提交
911
	r = amdgpu_cs_submit(&parser, cs);
A
Alex Deucher 已提交
912 913

out:
914
	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
A
Alex Deucher 已提交
915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933
	r = amdgpu_cs_handle_lockup(adev, r);
	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);
934
	struct amdgpu_ring *ring = NULL;
935
	struct amdgpu_ctx *ctx;
936
	struct fence *fence;
A
Alex Deucher 已提交
937 938
	long r;

939 940 941 942 943
	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
			       wait->in.ring, &ring);
	if (r)
		return r;

944 945 946
	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
	if (ctx == NULL)
		return -EINVAL;
A
Alex Deucher 已提交
947

948 949 950 951 952 953 954 955
	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
	if (IS_ERR(fence))
		r = PTR_ERR(fence);
	else if (fence) {
		r = fence_wait_timeout(fence, true, timeout);
		fence_put(fence);
	} else
		r = 1;
C
Chunming Zhou 已提交
956

957
	amdgpu_ctx_put(ctx);
A
Alex Deucher 已提交
958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982
	if (r < 0)
		return r;

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

	return 0;
}

/**
 * amdgpu_cs_find_bo_va - find bo_va for VM address
 *
 * @parser: command submission parser context
 * @addr: VM address
 * @bo: resulting BO of the mapping found
 *
 * Search the buffer objects in the command submission context for a certain
 * virtual memory address. Returns allocation structure when found, NULL
 * otherwise.
 */
struct amdgpu_bo_va_mapping *
amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
		       uint64_t addr, struct amdgpu_bo **bo)
{
	struct amdgpu_bo_va_mapping *mapping;
983 984 985 986
	unsigned i;

	if (!parser->bo_list)
		return NULL;
A
Alex Deucher 已提交
987 988 989

	addr /= AMDGPU_GPU_PAGE_SIZE;

990 991 992 993 994
	for (i = 0; i < parser->bo_list->num_entries; i++) {
		struct amdgpu_bo_list_entry *lobj;

		lobj = &parser->bo_list->array[i];
		if (!lobj->bo_va)
A
Alex Deucher 已提交
995 996
			continue;

997
		list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
998 999 1000 1001
			if (mapping->it.start > addr ||
			    addr > mapping->it.last)
				continue;

1002
			*bo = lobj->bo_va->bo;
1003 1004 1005
			return mapping;
		}

1006
		list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
A
Alex Deucher 已提交
1007 1008 1009 1010
			if (mapping->it.start > addr ||
			    addr > mapping->it.last)
				continue;

1011
			*bo = lobj->bo_va->bo;
A
Alex Deucher 已提交
1012 1013 1014 1015 1016 1017
			return mapping;
		}
	}

	return NULL;
}