amdgpu_cs.c 29.5 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
{
	struct drm_gem_object *gobj;
94
	unsigned long size;
95

96
	gobj = drm_gem_object_lookup(p->filp, data->handle);
97 98 99
	if (gobj == NULL)
		return -EINVAL;

100
	p->uf_entry.robj = amdgpu_bo_ref(gem_to_amdgpu_bo(gobj));
101 102 103
	p->uf_entry.priority = 0;
	p->uf_entry.tv.bo = &p->uf_entry.robj->tbo;
	p->uf_entry.tv.shared = true;
104
	p->uf_entry.user_pages = NULL;
105 106 107 108 109

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

110
	*offset = data->offset;
111 112

	drm_gem_object_unreference_unlocked(gobj);
113 114 115 116 117 118

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

119 120 121
	return 0;
}

A
Alex Deucher 已提交
122 123
int amdgpu_cs_parser_init(struct amdgpu_cs_parser *p, void *data)
{
124
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
125
	struct amdgpu_vm *vm = &fpriv->vm;
A
Alex Deucher 已提交
126 127
	union drm_amdgpu_cs *cs = data;
	uint64_t *chunk_array_user;
128
	uint64_t *chunk_array;
129
	unsigned size, num_ibs = 0;
130
	uint32_t uf_offset = 0;
131
	int i;
132
	int ret;
A
Alex Deucher 已提交
133

134 135 136 137 138 139
	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
140

141 142
	p->ctx = amdgpu_ctx_get(fpriv, cs->in.ctx_id);
	if (!p->ctx) {
143 144
		ret = -EINVAL;
		goto free_chunk;
145
	}
146

A
Alex Deucher 已提交
147
	/* get chunks */
148
	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
A
Alex Deucher 已提交
149 150
	if (copy_from_user(chunk_array, chunk_array_user,
			   sizeof(uint64_t)*cs->in.num_chunks)) {
151
		ret = -EFAULT;
152
		goto put_ctx;
A
Alex Deucher 已提交
153 154 155
	}

	p->nchunks = cs->in.num_chunks;
156
	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
A
Alex Deucher 已提交
157
			    GFP_KERNEL);
158 159
	if (!p->chunks) {
		ret = -ENOMEM;
160
		goto put_ctx;
A
Alex Deucher 已提交
161 162 163 164 165 166 167
	}

	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;

168
		chunk_ptr = (void __user *)(unsigned long)chunk_array[i];
A
Alex Deucher 已提交
169 170
		if (copy_from_user(&user_chunk, chunk_ptr,
				       sizeof(struct drm_amdgpu_cs_chunk))) {
171 172 173
			ret = -EFAULT;
			i--;
			goto free_partial_kdata;
A
Alex Deucher 已提交
174 175 176 177 178
		}
		p->chunks[i].chunk_id = user_chunk.chunk_id;
		p->chunks[i].length_dw = user_chunk.length_dw;

		size = p->chunks[i].length_dw;
179
		cdata = (void __user *)(unsigned long)user_chunk.chunk_data;
A
Alex Deucher 已提交
180 181 182

		p->chunks[i].kdata = drm_malloc_ab(size, sizeof(uint32_t));
		if (p->chunks[i].kdata == NULL) {
183 184 185
			ret = -ENOMEM;
			i--;
			goto free_partial_kdata;
A
Alex Deucher 已提交
186 187 188
		}
		size *= sizeof(uint32_t);
		if (copy_from_user(p->chunks[i].kdata, cdata, size)) {
189 190
			ret = -EFAULT;
			goto free_partial_kdata;
A
Alex Deucher 已提交
191 192
		}

193 194
		switch (p->chunks[i].chunk_id) {
		case AMDGPU_CHUNK_ID_IB:
195
			++num_ibs;
196 197 198
			break;

		case AMDGPU_CHUNK_ID_FENCE:
A
Alex Deucher 已提交
199
			size = sizeof(struct drm_amdgpu_cs_chunk_fence);
200
			if (p->chunks[i].length_dw * sizeof(uint32_t) < size) {
201 202
				ret = -EINVAL;
				goto free_partial_kdata;
A
Alex Deucher 已提交
203
			}
204

205 206
			ret = amdgpu_cs_user_fence_chunk(p, p->chunks[i].kdata,
							 &uf_offset);
207 208 209
			if (ret)
				goto free_partial_kdata;

210 211
			break;

212 213 214
		case AMDGPU_CHUNK_ID_DEPENDENCIES:
			break;

215
		default:
216 217
			ret = -EINVAL;
			goto free_partial_kdata;
A
Alex Deucher 已提交
218 219 220
		}
	}

221
	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
222
	if (ret)
C
Christian König 已提交
223
		goto free_all_kdata;
A
Alex Deucher 已提交
224

225 226
	if (p->uf_entry.robj)
		p->job->uf_addr = uf_offset;
A
Alex Deucher 已提交
227
	kfree(chunk_array);
228 229 230 231 232 233 234 235
	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);
236
put_ctx:
237 238 239 240 241
	amdgpu_ctx_put(p->ctx);
free_chunk:
	kfree(chunk_array);

	return ret;
A
Alex Deucher 已提交
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
/* 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.
A
Alex Deucher 已提交
276 277 278
 */
static u64 amdgpu_cs_get_threshold_for_moves(struct amdgpu_device *adev)
{
279 280 281
	s64 time_us, increment_us;
	u64 max_bytes;
	u64 free_vram, total_vram, used_vram;
A
Alex Deucher 已提交
282

283 284
	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
	 * throttling.
A
Alex Deucher 已提交
285
	 *
286 287 288 289 290
	 * 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;
A
Alex Deucher 已提交
291

292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310
	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.
A
Alex Deucher 已提交
311
	 *
312 313 314 315
	 * 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)
A
Alex Deucher 已提交
316
	 *
317 318
	 * If 128 MB or 1/8th of VRAM is free, start filling it now by setting
	 * accum_us to a positive number.
A
Alex Deucher 已提交
319
	 */
320 321 322 323 324 325 326 327 328 329 330 331 332
	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);
	}
A
Alex Deucher 已提交
333

334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352
	/* 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);
A
Alex Deucher 已提交
353 354
}

355 356 357
static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
				 struct amdgpu_bo *bo)
{
358
	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
359 360 361 362 363 364 365
	u64 initial_bytes_moved;
	uint32_t domain;
	int r;

	if (bo->pin_count)
		return 0;

366 367
	/* Don't move this buffer if we have depleted our allowance
	 * to move it. Don't move anything if the threshold is zero.
368
	 */
369
	if (p->bytes_moved < p->bytes_moved_threshold)
370 371 372 373 374 375
		domain = bo->prefered_domains;
	else
		domain = bo->allowed_domains;

retry:
	amdgpu_ttm_placement_from_domain(bo, domain);
376
	initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
377
	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
378
	p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
379 380
		initial_bytes_moved;

381 382 383
	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
		domain = bo->allowed_domains;
		goto retry;
384 385 386 387 388
	}

	return r;
}

389 390 391 392 393 394 395 396 397 398 399 400 401 402 403
/* Last resort, try to evict something from the current working set */
static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
				struct amdgpu_bo_list_entry *lobj)
{
	uint32_t domain = lobj->robj->allowed_domains;
	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;
404
		struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
		u64 initial_bytes_moved;
		uint32_t other;

		/* If we reached our current BO we can forget it */
		if (candidate == lobj)
			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);
425
		initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
426
		r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
427
		p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
428 429 430 431 432 433 434 435 436 437 438 439 440 441
			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;
}

442
static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
443
			    struct list_head *validated)
A
Alex Deucher 已提交
444 445 446 447
{
	struct amdgpu_bo_list_entry *lobj;
	int r;

448
	list_for_each_entry(lobj, validated, tv.head) {
449
		struct amdgpu_bo *bo = lobj->robj;
450
		bool binding_userptr = false;
451
		struct mm_struct *usermm;
A
Alex Deucher 已提交
452

453 454 455 456
		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
		if (usermm && usermm != current->mm)
			return -EPERM;

457 458 459 460 461 462 463 464 465
		/* 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;
		}

466 467 468 469 470 471
		if (p->evictable == lobj)
			p->evictable = NULL;

		do {
			r = amdgpu_cs_bo_validate(p, bo);
		} while (r == -ENOMEM && amdgpu_cs_try_evict(p, lobj));
472
		if (r)
473
			return r;
474

475 476 477 478
		if (bo->shadow) {
			r = amdgpu_cs_bo_validate(p, bo);
			if (r)
				return r;
A
Alex Deucher 已提交
479
		}
480 481 482 483 484

		if (binding_userptr) {
			drm_free_large(lobj->user_pages);
			lobj->user_pages = NULL;
		}
A
Alex Deucher 已提交
485 486 487 488
	}
	return 0;
}

489 490
static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
				union drm_amdgpu_cs *cs)
A
Alex Deucher 已提交
491 492
{
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
493
	struct amdgpu_bo_list_entry *e;
494
	struct list_head duplicates;
495
	bool need_mmap_lock = false;
496
	unsigned i, tries = 10;
497
	int r;
A
Alex Deucher 已提交
498

499 500 501
	INIT_LIST_HEAD(&p->validated);

	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
502
	if (p->bo_list) {
503 504
		need_mmap_lock = p->bo_list->first_userptr !=
			p->bo_list->num_entries;
505
		amdgpu_bo_list_get_list(p->bo_list, &p->validated);
506
	}
A
Alex Deucher 已提交
507

508
	INIT_LIST_HEAD(&duplicates);
509
	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
A
Alex Deucher 已提交
510

511
	if (p->uf_entry.robj)
512 513
		list_add(&p->uf_entry.tv.head, &p->validated);

A
Alex Deucher 已提交
514 515 516
	if (need_mmap_lock)
		down_read(&current->mm->mmap_sem);

517 518 519 520 521 522
	while (1) {
		struct list_head need_pages;
		unsigned i;

		r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
					   &duplicates);
523 524
		if (unlikely(r != 0)) {
			DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
525
			goto error_free_pages;
526
		}
527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565

		/* 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);

566
		/* We tried too many times, just abort */
567 568
		if (!--tries) {
			r = -EDEADLK;
569
			DRM_ERROR("deadlock in %s\n", __func__);
570 571 572 573 574 575 576 577 578 579 580
			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;
581
				DRM_ERROR("calloc failure in %s\n", __func__);
582 583 584 585 586
				goto error_free_pages;
			}

			r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
			if (r) {
587
				DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
588 589 590 591 592 593 594 595 596
				drm_free_large(e->user_pages);
				e->user_pages = NULL;
				goto error_free_pages;
			}
		}

		/* And try again. */
		list_splice(&need_pages, &p->validated);
	}
597

598
	amdgpu_vm_get_pt_bos(p->adev, &fpriv->vm, &duplicates);
599

600 601
	p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
	p->bytes_moved = 0;
602 603 604
	p->evictable = list_last_entry(&p->validated,
				       struct amdgpu_bo_list_entry,
				       tv.head);
605 606

	r = amdgpu_cs_list_validate(p, &duplicates);
607 608
	if (r) {
		DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
609
		goto error_validate;
610
	}
611

612
	r = amdgpu_cs_list_validate(p, &p->validated);
613 614
	if (r) {
		DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
615
		goto error_validate;
616
	}
617

618 619
	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved);

620 621 622
	fpriv->vm.last_eviction_counter =
		atomic64_read(&p->adev->num_evictions);

623
	if (p->bo_list) {
624 625 626
		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;
627 628 629 630 631 632 633 634
		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);
		}
635 636 637 638 639 640 641 642 643 644 645 646 647

		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);
		}
648
	}
649

650 651 652
	if (!r && p->uf_entry.robj) {
		struct amdgpu_bo *uf = p->uf_entry.robj;

653
		r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
654 655
		p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
	}
656

657
error_validate:
658 659
	if (r) {
		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
660
		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
661
	}
A
Alex Deucher 已提交
662

663 664
error_free_pages:

A
Alex Deucher 已提交
665 666 667
	if (need_mmap_lock)
		up_read(&current->mm->mmap_sem);

668 669 670 671 672 673 674 675 676 677 678 679 680 681 682
	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
683 684 685 686 687 688 689 690 691 692
	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;
693
		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
A
Alex Deucher 已提交
694 695 696 697 698 699 700

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

701 702 703 704 705 706 707 708 709
/**
 * 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
710
{
711
	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
712 713
	unsigned i;

A
Alex Deucher 已提交
714
	if (!error) {
715 716
		amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);

A
Alex Deucher 已提交
717
		ttm_eu_fence_buffer_objects(&parser->ticket,
718 719
					    &parser->validated,
					    parser->fence);
A
Alex Deucher 已提交
720 721 722 723
	} else if (backoff) {
		ttm_eu_backoff_reservation(&parser->ticket,
					   &parser->validated);
	}
724
	fence_put(parser->fence);
725

726 727
	if (parser->ctx)
		amdgpu_ctx_put(parser->ctx);
728 729 730
	if (parser->bo_list)
		amdgpu_bo_list_put(parser->bo_list);

A
Alex Deucher 已提交
731 732 733
	for (i = 0; i < parser->nchunks; i++)
		drm_free_large(parser->chunks[i].kdata);
	kfree(parser->chunks);
734 735
	if (parser->job)
		amdgpu_job_free(parser->job);
736
	amdgpu_bo_unref(&parser->uf_entry.robj);
A
Alex Deucher 已提交
737 738 739 740 741 742 743 744 745 746 747 748 749 750
}

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;

751
	r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
752 753 754
	if (r)
		return r;

A
Alex Deucher 已提交
755 756 757 758 759 760
	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++) {
761 762
			struct fence *f;

A
Alex Deucher 已提交
763 764 765 766 767 768 769 770 771
			/* 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;

772
			r = amdgpu_vm_bo_update(adev, bo_va, false);
A
Alex Deucher 已提交
773 774 775
			if (r)
				return r;

776
			f = bo_va->last_pt_update;
777
			r = amdgpu_sync_fence(adev, &p->job->sync, f);
778 779
			if (r)
				return r;
A
Alex Deucher 已提交
780
		}
781 782 783

	}

784
	r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
785 786 787 788 789 790 791 792 793 794 795

	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
796 797
	}

798
	return r;
A
Alex Deucher 已提交
799 800 801
}

static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
802
				 struct amdgpu_cs_parser *p)
A
Alex Deucher 已提交
803
{
804
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
A
Alex Deucher 已提交
805
	struct amdgpu_vm *vm = &fpriv->vm;
806
	struct amdgpu_ring *ring = p->job->ring;
A
Alex Deucher 已提交
807 808 809
	int i, r;

	/* Only for UVD/VCE VM emulation */
810
	if (ring->funcs->parse_cs) {
811
		p->job->vm = NULL;
812 813
		for (i = 0; i < p->job->num_ibs; i++) {
			r = amdgpu_ring_parse_cs(ring, p, i);
A
Alex Deucher 已提交
814 815 816
			if (r)
				return r;
		}
817 818
	} else {
		p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
819

820 821 822 823
		r = amdgpu_bo_vm_update_pte(p, vm);
		if (r)
			return r;
	}
A
Alex Deucher 已提交
824

825
	return amdgpu_cs_sync_rings(p);
A
Alex Deucher 已提交
826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845
}

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;

846
	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
A
Alex Deucher 已提交
847 848 849 850 851 852
		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];
853
		ib = &parser->job->ibs[j];
A
Alex Deucher 已提交
854 855 856 857 858 859 860 861
		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);
862
		if (r)
A
Alex Deucher 已提交
863 864
			return r;

865 866 867 868 869 870 871 872
		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;
			}
		}

873 874 875 876 877
		if (parser->job->ring && parser->job->ring != ring)
			return -EINVAL;

		parser->job->ring = ring;

A
Alex Deucher 已提交
878
		if (ring->funcs->parse_cs) {
879
			struct amdgpu_bo_va_mapping *m;
880
			struct amdgpu_bo *aobj = NULL;
881 882
			uint64_t offset;
			uint8_t *kptr;
883

884 885
			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
						   &aobj);
886 887 888
			if (!aobj) {
				DRM_ERROR("IB va_start is invalid\n");
				return -EINVAL;
A
Alex Deucher 已提交
889 890
			}

891 892 893 894 895 896
			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;
			}

897
			/* the IB should be reserved at this point */
898
			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
A
Alex Deucher 已提交
899 900 901 902
			if (r) {
				return r;
			}

903 904 905
			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
			kptr += chunk_ib->va_start - offset;

906
			r =  amdgpu_ib_get(adev, NULL, chunk_ib->ib_bytes, ib);
A
Alex Deucher 已提交
907 908 909 910 911 912 913 914
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
			amdgpu_bo_kunmap(aobj);
		} else {
915
			r =  amdgpu_ib_get(adev, vm, 0, ib);
A
Alex Deucher 已提交
916 917 918 919 920 921 922 923
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

			ib->gpu_addr = chunk_ib->va_start;
		}

924
		ib->length_dw = chunk_ib->ib_bytes / 4;
925
		ib->flags = chunk_ib->flags;
A
Alex Deucher 已提交
926 927 928
		j++;
	}

929
	/* UVD & VCE fw doesn't support user fences */
930
	if (parser->job->uf_addr && (
931 932 933
	    parser->job->ring->type == AMDGPU_RING_TYPE_UVD ||
	    parser->job->ring->type == AMDGPU_RING_TYPE_VCE))
		return -EINVAL;
A
Alex Deucher 已提交
934 935 936 937

	return 0;
}

938 939 940
static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
				  struct amdgpu_cs_parser *p)
{
941
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959
	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;
960
			struct amdgpu_ctx *ctx;
961
			struct fence *fence;
962 963 964 965 966 967 968

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

969 970 971 972
			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
			if (ctx == NULL)
				return -EINVAL;

973 974 975 976
			fence = amdgpu_ctx_get_fence(ctx, ring,
						     deps[j].handle);
			if (IS_ERR(fence)) {
				r = PTR_ERR(fence);
977
				amdgpu_ctx_put(ctx);
978
				return r;
979

980
			} else if (fence) {
981 982
				r = amdgpu_sync_fence(adev, &p->job->sync,
						      fence);
983 984 985 986 987
				fence_put(fence);
				amdgpu_ctx_put(ctx);
				if (r)
					return r;
			}
988 989 990 991 992 993
		}
	}

	return 0;
}

994 995 996
static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
			    union drm_amdgpu_cs *cs)
{
997
	struct amdgpu_ring *ring = p->job->ring;
998
	struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
999
	struct amdgpu_job *job;
1000
	int r;
1001

1002 1003
	job = p->job;
	p->job = NULL;
1004

1005
	r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1006
	if (r) {
1007
		amdgpu_job_free(job);
1008
		return r;
1009 1010
	}

1011
	job->owner = p->filp;
1012
	job->fence_ctx = entity->fence_context;
1013 1014
	p->fence = fence_get(&job->base.s_fence->finished);
	cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
1015
	job->uf_sequence = cs->out.handle;
1016
	amdgpu_job_free_resources(job);
1017 1018 1019 1020 1021 1022 1023

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

	return 0;
}

C
Chunming Zhou 已提交
1024 1025 1026 1027
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;
1028
	struct amdgpu_cs_parser parser = {};
1029 1030
	bool reserved_buffers = false;
	int i, r;
C
Chunming Zhou 已提交
1031

1032
	if (!adev->accel_working)
C
Chunming Zhou 已提交
1033
		return -EBUSY;
1034

1035 1036 1037 1038
	parser.adev = adev;
	parser.filp = filp;

	r = amdgpu_cs_parser_init(&parser, data);
A
Alex Deucher 已提交
1039
	if (r) {
C
Chunming Zhou 已提交
1040
		DRM_ERROR("Failed to initialize parser !\n");
1041
		amdgpu_cs_parser_fini(&parser, r, false);
A
Alex Deucher 已提交
1042 1043 1044
		r = amdgpu_cs_handle_lockup(adev, r);
		return r;
	}
1045
	r = amdgpu_cs_parser_bos(&parser, data);
1046 1047 1048 1049 1050 1051
	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;
1052
		r = amdgpu_cs_ib_fill(adev, &parser);
1053 1054 1055
	}

	if (!r) {
1056
		r = amdgpu_cs_dependencies(adev, &parser);
1057 1058 1059 1060 1061 1062 1063
		if (r)
			DRM_ERROR("Failed in the dependencies handling %d!\n", r);
	}

	if (r)
		goto out;

1064
	for (i = 0; i < parser.job->num_ibs; i++)
1065
		trace_amdgpu_cs(&parser, i);
1066

1067
	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1068 1069 1070
	if (r)
		goto out;

C
Christian König 已提交
1071
	r = amdgpu_cs_submit(&parser, cs);
A
Alex Deucher 已提交
1072 1073

out:
1074
	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
A
Alex Deucher 已提交
1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093
	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);
1094
	struct amdgpu_ring *ring = NULL;
1095
	struct amdgpu_ctx *ctx;
1096
	struct fence *fence;
A
Alex Deucher 已提交
1097 1098
	long r;

1099 1100 1101 1102 1103
	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
			       wait->in.ring, &ring);
	if (r)
		return r;

1104 1105 1106
	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
	if (ctx == NULL)
		return -EINVAL;
A
Alex Deucher 已提交
1107

1108 1109 1110 1111 1112 1113 1114 1115
	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
1116

1117
	amdgpu_ctx_put(ctx);
A
Alex Deucher 已提交
1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142
	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;
1143 1144 1145 1146
	unsigned i;

	if (!parser->bo_list)
		return NULL;
A
Alex Deucher 已提交
1147 1148 1149

	addr /= AMDGPU_GPU_PAGE_SIZE;

1150 1151 1152 1153 1154
	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
1155 1156
			continue;

1157
		list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1158 1159 1160 1161
			if (mapping->it.start > addr ||
			    addr > mapping->it.last)
				continue;

1162
			*bo = lobj->bo_va->bo;
1163 1164 1165
			return mapping;
		}

1166
		list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
A
Alex Deucher 已提交
1167 1168 1169 1170
			if (mapping->it.start > addr ||
			    addr > mapping->it.last)
				continue;

1171
			*bo = lobj->bo_va->bo;
A
Alex Deucher 已提交
1172 1173 1174 1175 1176 1177
			return mapping;
		}
	}

	return NULL;
}
1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196

/**
 * amdgpu_cs_sysvm_access_required - make BOs accessible by the system VM
 *
 * @parser: command submission parser context
 *
 * Helper for UVD/VCE VM emulation, make sure BOs are accessible by the system VM.
 */
int amdgpu_cs_sysvm_access_required(struct amdgpu_cs_parser *parser)
{
	unsigned i;
	int r;

	if (!parser->bo_list)
		return 0;

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

1197
		r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
1198 1199
		if (unlikely(r))
			return r;
1200 1201 1202 1203 1204 1205 1206 1207 1208

		if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS)
			continue;

		bo->flags |= AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
		amdgpu_ttm_placement_from_domain(bo, bo->allowed_domains);
		r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false);
		if (unlikely(r))
			return r;
1209 1210 1211 1212
	}

	return 0;
}