amdgpu_cs.c 33.9 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
			return -EINVAL;
		}
		break;
	case AMDGPU_HW_IP_UVD:
		*out_ring = &adev->uvd.ring;
		break;
	case AMDGPU_HW_IP_VCE:
78
		if (ring < adev->vce.num_rings){
A
Alex Deucher 已提交
79 80
			*out_ring = &adev->vce.ring[ring];
		} else {
81
			DRM_ERROR("only %d VCE rings are supported\n", adev->vce.num_rings);
A
Alex Deucher 已提交
82 83 84 85
			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;
	}

A
Alex Deucher 已提交
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;
}

A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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

A
Alex Deucher 已提交
154
	/* get chunks */
155
	chunk_array_user = (uint64_t __user *)(unsigned long)(cs->in.chunks);
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
160 161 162
	}

	p->nchunks = cs->in.num_chunks;
163
	p->chunks = kmalloc_array(p->nchunks, sizeof(struct amdgpu_cs_chunk),
A
Alex Deucher 已提交
164
			    GFP_KERNEL);
165 166
	if (!p->chunks) {
		ret = -ENOMEM;
167
		goto put_ctx;
A
Alex Deucher 已提交
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];
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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:
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
225 226 227
		}
	}

228
	ret = amdgpu_job_alloc(p->adev, num_ibs, &p->job, vm);
229
	if (ret)
C
Christian König 已提交
230
		goto free_all_kdata;
A
Alex Deucher 已提交
231

232 233
	if (p->uf_entry.robj)
		p->job->uf_addr = uf_offset;
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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.
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
289

290 291
	/* Allow a maximum of 200 accumulated ms. This is basically per-IB
	 * throttling.
A
Alex Deucher 已提交
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;
A
Alex Deucher 已提交
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.
A
Alex Deucher 已提交
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)
A
Alex Deucher 已提交
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.
A
Alex Deucher 已提交
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);
	}
A
Alex Deucher 已提交
340

341 342 343 344 345 346 347 348 349 350 351 352 353
	/* 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.
 */
354
void amdgpu_cs_report_moved_bytes(struct amdgpu_device *adev, u64 num_bytes)
355 356 357 358
{
	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
359 360
}

361 362 363
static int amdgpu_cs_bo_validate(struct amdgpu_cs_parser *p,
				 struct amdgpu_bo *bo)
{
364
	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
365 366 367 368 369 370 371
	u64 initial_bytes_moved;
	uint32_t domain;
	int r;

	if (bo->pin_count)
		return 0;

372 373
	/* Don't move this buffer if we have depleted our allowance
	 * to move it. Don't move anything if the threshold is zero.
374
	 */
375
	if (p->bytes_moved < p->bytes_moved_threshold)
376 377 378 379 380 381
		domain = bo->prefered_domains;
	else
		domain = bo->allowed_domains;

retry:
	amdgpu_ttm_placement_from_domain(bo, domain);
382
	initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
383
	r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
384
	p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
385 386
		initial_bytes_moved;

387 388 389
	if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
		domain = bo->allowed_domains;
		goto retry;
390 391 392 393 394
	}

	return r;
}

395 396
/* Last resort, try to evict something from the current working set */
static bool amdgpu_cs_try_evict(struct amdgpu_cs_parser *p,
397
				struct amdgpu_bo *validated)
398
{
399
	uint32_t domain = validated->allowed_domains;
400 401 402 403 404 405 406 407 408 409
	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;
410
		struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
411 412 413 414
		u64 initial_bytes_moved;
		uint32_t other;

		/* If we reached our current BO we can forget it */
415
		if (candidate->robj == validated)
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430
			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);
431
		initial_bytes_moved = atomic64_read(&adev->num_bytes_moved);
432
		r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false);
433
		p->bytes_moved += atomic64_read(&adev->num_bytes_moved) -
434 435 436 437 438 439 440 441 442 443 444 445 446 447
			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;
}

448 449 450 451 452 453 454 455 456 457 458 459
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)
A
Alex Xie 已提交
460
		r = amdgpu_cs_bo_validate(p, bo->shadow);
461 462 463 464

	return r;
}

465
static int amdgpu_cs_list_validate(struct amdgpu_cs_parser *p,
466
			    struct list_head *validated)
A
Alex Deucher 已提交
467 468 469 470
{
	struct amdgpu_bo_list_entry *lobj;
	int r;

471
	list_for_each_entry(lobj, validated, tv.head) {
472
		struct amdgpu_bo *bo = lobj->robj;
473
		bool binding_userptr = false;
474
		struct mm_struct *usermm;
A
Alex Deucher 已提交
475

476 477 478 479
		usermm = amdgpu_ttm_tt_get_usermm(bo->tbo.ttm);
		if (usermm && usermm != current->mm)
			return -EPERM;

480 481 482 483 484 485 486 487 488
		/* 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;
		}

489 490 491
		if (p->evictable == lobj)
			p->evictable = NULL;

492
		r = amdgpu_cs_validate(p, bo);
493
		if (r)
494
			return r;
495

496 497 498 499
		if (binding_userptr) {
			drm_free_large(lobj->user_pages);
			lobj->user_pages = NULL;
		}
A
Alex Deucher 已提交
500 501 502 503
	}
	return 0;
}

504 505
static int amdgpu_cs_parser_bos(struct amdgpu_cs_parser *p,
				union drm_amdgpu_cs *cs)
A
Alex Deucher 已提交
506 507
{
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
508
	struct amdgpu_bo_list_entry *e;
509
	struct list_head duplicates;
510
	bool need_mmap_lock = false;
511
	unsigned i, tries = 10;
512
	int r;
A
Alex Deucher 已提交
513

514 515 516
	INIT_LIST_HEAD(&p->validated);

	p->bo_list = amdgpu_bo_list_get(fpriv, cs->in.bo_list_handle);
517
	if (p->bo_list) {
518 519
		need_mmap_lock = p->bo_list->first_userptr !=
			p->bo_list->num_entries;
520
		amdgpu_bo_list_get_list(p->bo_list, &p->validated);
521
	}
A
Alex Deucher 已提交
522

523
	INIT_LIST_HEAD(&duplicates);
524
	amdgpu_vm_get_pd_bo(&fpriv->vm, &p->validated, &p->vm_pd);
A
Alex Deucher 已提交
525

526
	if (p->uf_entry.robj)
527 528
		list_add(&p->uf_entry.tv.head, &p->validated);

A
Alex Deucher 已提交
529 530 531
	if (need_mmap_lock)
		down_read(&current->mm->mmap_sem);

532 533 534 535 536 537
	while (1) {
		struct list_head need_pages;
		unsigned i;

		r = ttm_eu_reserve_buffers(&p->ticket, &p->validated, true,
					   &duplicates);
538
		if (unlikely(r != 0)) {
539 540
			if (r != -ERESTARTSYS)
				DRM_ERROR("ttm_eu_reserve_buffers failed.\n");
541
			goto error_free_pages;
542
		}
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

		/* 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);

582
		/* We tried too many times, just abort */
583 584
		if (!--tries) {
			r = -EDEADLK;
585
			DRM_ERROR("deadlock in %s\n", __func__);
586 587 588 589 590 591 592 593 594 595 596
			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;
597
				DRM_ERROR("calloc failure in %s\n", __func__);
598 599 600 601 602
				goto error_free_pages;
			}

			r = amdgpu_ttm_tt_get_user_pages(ttm, e->user_pages);
			if (r) {
603
				DRM_ERROR("amdgpu_ttm_tt_get_user_pages failed.\n");
604 605 606 607 608 609 610 611 612
				drm_free_large(e->user_pages);
				e->user_pages = NULL;
				goto error_free_pages;
			}
		}

		/* And try again. */
		list_splice(&need_pages, &p->validated);
	}
613

614 615
	p->bytes_moved_threshold = amdgpu_cs_get_threshold_for_moves(p->adev);
	p->bytes_moved = 0;
616 617 618
	p->evictable = list_last_entry(&p->validated,
				       struct amdgpu_bo_list_entry,
				       tv.head);
619

620 621 622 623 624 625 626
	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;
	}

627
	r = amdgpu_cs_list_validate(p, &duplicates);
628 629
	if (r) {
		DRM_ERROR("amdgpu_cs_list_validate(duplicates) failed.\n");
630
		goto error_validate;
631
	}
632

633
	r = amdgpu_cs_list_validate(p, &p->validated);
634 635
	if (r) {
		DRM_ERROR("amdgpu_cs_list_validate(validated) failed.\n");
636
		goto error_validate;
637
	}
638

639 640
	amdgpu_cs_report_moved_bytes(p->adev, p->bytes_moved);

641 642 643
	fpriv->vm.last_eviction_counter =
		atomic64_read(&p->adev->num_evictions);

644
	if (p->bo_list) {
645 646 647
		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;
648 649 650 651 652 653 654 655
		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);
		}
656 657 658 659 660 661 662 663 664 665 666 667 668

		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);
		}
669
	}
670

671 672 673
	if (!r && p->uf_entry.robj) {
		struct amdgpu_bo *uf = p->uf_entry.robj;

674
		r = amdgpu_ttm_bind(&uf->tbo, &uf->tbo.mem);
675 676
		p->job->uf_addr += amdgpu_bo_gpu_offset(uf);
	}
677

678
error_validate:
679 680
	if (r) {
		amdgpu_vm_move_pt_bos_in_lru(p->adev, &fpriv->vm);
681
		ttm_eu_backoff_reservation(&p->ticket, &p->validated);
682
	}
A
Alex Deucher 已提交
683

684 685
error_free_pages:

A
Alex Deucher 已提交
686 687 688
	if (need_mmap_lock)
		up_read(&current->mm->mmap_sem);

689 690 691 692 693 694 695 696 697 698 699 700 701 702 703
	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
704 705 706 707 708 709 710 711 712 713
	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;
714
		r = amdgpu_sync_resv(p->adev, &p->job->sync, resv, p->filp);
A
Alex Deucher 已提交
715 716 717 718 719 720 721

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

722 723 724 725 726 727 728 729 730
/**
 * 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
731
{
732
	struct amdgpu_fpriv *fpriv = parser->filp->driver_priv;
733 734
	unsigned i;

A
Alex Deucher 已提交
735
	if (!error) {
736 737
		amdgpu_vm_move_pt_bos_in_lru(parser->adev, &fpriv->vm);

A
Alex Deucher 已提交
738
		ttm_eu_fence_buffer_objects(&parser->ticket,
739 740
					    &parser->validated,
					    parser->fence);
A
Alex Deucher 已提交
741 742 743 744
	} else if (backoff) {
		ttm_eu_backoff_reservation(&parser->ticket,
					   &parser->validated);
	}
745
	dma_fence_put(parser->fence);
746

747 748
	if (parser->ctx)
		amdgpu_ctx_put(parser->ctx);
749 750 751
	if (parser->bo_list)
		amdgpu_bo_list_put(parser->bo_list);

A
Alex Deucher 已提交
752 753 754
	for (i = 0; i < parser->nchunks; i++)
		drm_free_large(parser->chunks[i].kdata);
	kfree(parser->chunks);
755 756
	if (parser->job)
		amdgpu_job_free(parser->job);
757
	amdgpu_bo_unref(&parser->uf_entry.robj);
A
Alex Deucher 已提交
758 759 760 761 762 763 764 765 766 767 768 769 770 771
}

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;

772
	r = amdgpu_sync_fence(adev, &p->job->sync, vm->page_directory_fence);
773 774 775
	if (r)
		return r;

A
Alex Deucher 已提交
776 777 778 779
	r = amdgpu_vm_clear_freed(adev, vm);
	if (r)
		return r;

M
Monk Liu 已提交
780 781 782 783 784 785 786 787 788 789 790 791 792 793
	if (amdgpu_sriov_vf(adev)) {
		struct dma_fence *f;
		bo_va = vm->csa_bo_va;
		BUG_ON(!bo_va);
		r = amdgpu_vm_bo_update(adev, bo_va, false);
		if (r)
			return r;

		f = bo_va->last_pt_update;
		r = amdgpu_sync_fence(adev, &p->job->sync, f);
		if (r)
			return r;
	}

A
Alex Deucher 已提交
794 795
	if (p->bo_list) {
		for (i = 0; i < p->bo_list->num_entries; i++) {
796
			struct dma_fence *f;
797

A
Alex Deucher 已提交
798 799 800 801 802 803 804 805 806
			/* 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;

807
			r = amdgpu_vm_bo_update(adev, bo_va, false);
A
Alex Deucher 已提交
808 809 810
			if (r)
				return r;

811
			f = bo_va->last_pt_update;
812
			r = amdgpu_sync_fence(adev, &p->job->sync, f);
813 814
			if (r)
				return r;
A
Alex Deucher 已提交
815
		}
816 817 818

	}

819
	r = amdgpu_vm_clear_invalids(adev, vm, &p->job->sync);
820 821 822 823 824 825 826 827 828 829 830

	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
831 832
	}

833
	return r;
A
Alex Deucher 已提交
834 835 836
}

static int amdgpu_cs_ib_vm_chunk(struct amdgpu_device *adev,
837
				 struct amdgpu_cs_parser *p)
A
Alex Deucher 已提交
838
{
839
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
A
Alex Deucher 已提交
840
	struct amdgpu_vm *vm = &fpriv->vm;
841
	struct amdgpu_ring *ring = p->job->ring;
A
Alex Deucher 已提交
842 843 844
	int i, r;

	/* Only for UVD/VCE VM emulation */
845 846 847
	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
848 849 850
			if (r)
				return r;
		}
851 852 853
	}

	if (p->job->vm) {
854
		p->job->vm_pd_addr = amdgpu_bo_gpu_offset(vm->page_directory);
855

856 857 858 859
		r = amdgpu_bo_vm_update_pte(p, vm);
		if (r)
			return r;
	}
A
Alex Deucher 已提交
860

861
	return amdgpu_cs_sync_rings(p);
A
Alex Deucher 已提交
862 863 864 865 866 867 868 869 870 871
}

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;

872
	for (i = 0, j = 0; i < parser->nchunks && j < parser->job->num_ibs; i++) {
A
Alex Deucher 已提交
873 874 875 876 877 878
		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];
879
		ib = &parser->job->ibs[j];
A
Alex Deucher 已提交
880 881 882 883 884 885 886 887
		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);
888
		if (r)
A
Alex Deucher 已提交
889 890
			return r;

891 892 893 894 895 896 897 898
		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;
			}
		}

899 900 901 902 903
		if (parser->job->ring && parser->job->ring != ring)
			return -EINVAL;

		parser->job->ring = ring;

A
Alex Deucher 已提交
904
		if (ring->funcs->parse_cs) {
905
			struct amdgpu_bo_va_mapping *m;
906
			struct amdgpu_bo *aobj = NULL;
907 908
			uint64_t offset;
			uint8_t *kptr;
909

910 911
			m = amdgpu_cs_find_mapping(parser, chunk_ib->va_start,
						   &aobj);
912 913 914
			if (!aobj) {
				DRM_ERROR("IB va_start is invalid\n");
				return -EINVAL;
A
Alex Deucher 已提交
915 916
			}

917 918 919 920 921 922
			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;
			}

923
			/* the IB should be reserved at this point */
924
			r = amdgpu_bo_kmap(aobj, (void **)&kptr);
A
Alex Deucher 已提交
925 926 927 928
			if (r) {
				return r;
			}

929 930 931
			offset = ((uint64_t)m->it.start) * AMDGPU_GPU_PAGE_SIZE;
			kptr += chunk_ib->va_start - offset;

932
			r =  amdgpu_ib_get(adev, vm, chunk_ib->ib_bytes, ib);
A
Alex Deucher 已提交
933 934 935 936 937 938 939 940
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

			memcpy(ib->ptr, kptr, chunk_ib->ib_bytes);
			amdgpu_bo_kunmap(aobj);
		} else {
941
			r =  amdgpu_ib_get(adev, vm, 0, ib);
A
Alex Deucher 已提交
942 943 944 945 946 947 948
			if (r) {
				DRM_ERROR("Failed to get ib !\n");
				return r;
			}

		}

949
		ib->gpu_addr = chunk_ib->va_start;
950
		ib->length_dw = chunk_ib->ib_bytes / 4;
951
		ib->flags = chunk_ib->flags;
A
Alex Deucher 已提交
952 953 954
		j++;
	}

955
	/* UVD & VCE fw doesn't support user fences */
956
	if (parser->job->uf_addr && (
957 958
	    parser->job->ring->funcs->type == AMDGPU_RING_TYPE_UVD ||
	    parser->job->ring->funcs->type == AMDGPU_RING_TYPE_VCE))
959
		return -EINVAL;
A
Alex Deucher 已提交
960 961 962 963

	return 0;
}

964 965 966
static int amdgpu_cs_dependencies(struct amdgpu_device *adev,
				  struct amdgpu_cs_parser *p)
{
967
	struct amdgpu_fpriv *fpriv = p->filp->driver_priv;
968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985
	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;
986
			struct amdgpu_ctx *ctx;
987
			struct dma_fence *fence;
988 989 990 991 992 993 994

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

995 996 997 998
			ctx = amdgpu_ctx_get(fpriv, deps[j].ctx_id);
			if (ctx == NULL)
				return -EINVAL;

999 1000 1001 1002
			fence = amdgpu_ctx_get_fence(ctx, ring,
						     deps[j].handle);
			if (IS_ERR(fence)) {
				r = PTR_ERR(fence);
1003
				amdgpu_ctx_put(ctx);
1004
				return r;
1005

1006
			} else if (fence) {
1007 1008
				r = amdgpu_sync_fence(adev, &p->job->sync,
						      fence);
1009
				dma_fence_put(fence);
1010 1011 1012 1013
				amdgpu_ctx_put(ctx);
				if (r)
					return r;
			}
1014 1015 1016 1017 1018 1019
		}
	}

	return 0;
}

1020 1021 1022
static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
			    union drm_amdgpu_cs *cs)
{
1023
	struct amdgpu_ring *ring = p->job->ring;
1024
	struct amd_sched_entity *entity = &p->ctx->rings[ring->idx].entity;
1025
	struct amdgpu_job *job;
1026
	int r;
1027

1028 1029
	job = p->job;
	p->job = NULL;
1030

1031
	r = amd_sched_job_init(&job->base, &ring->sched, entity, p->filp);
1032
	if (r) {
1033
		amdgpu_job_free(job);
1034
		return r;
1035 1036
	}

1037
	job->owner = p->filp;
1038
	job->fence_ctx = entity->fence_context;
1039
	p->fence = dma_fence_get(&job->base.s_fence->finished);
1040
	cs->out.handle = amdgpu_ctx_add_fence(p->ctx, ring, p->fence);
1041
	job->uf_sequence = cs->out.handle;
1042
	amdgpu_job_free_resources(job);
1043 1044 1045 1046 1047 1048 1049

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

	return 0;
}

C
Chunming Zhou 已提交
1050 1051 1052 1053
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;
1054
	struct amdgpu_cs_parser parser = {};
1055 1056
	bool reserved_buffers = false;
	int i, r;
C
Chunming Zhou 已提交
1057

1058
	if (!adev->accel_working)
C
Chunming Zhou 已提交
1059
		return -EBUSY;
1060

1061 1062 1063 1064
	parser.adev = adev;
	parser.filp = filp;

	r = amdgpu_cs_parser_init(&parser, data);
A
Alex Deucher 已提交
1065
	if (r) {
C
Chunming Zhou 已提交
1066
		DRM_ERROR("Failed to initialize parser !\n");
1067
		goto out;
1068 1069
	}

1070 1071 1072 1073 1074 1075 1076
	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;
1077 1078
	}

1079 1080
	reserved_buffers = true;
	r = amdgpu_cs_ib_fill(adev, &parser);
1081 1082 1083
	if (r)
		goto out;

1084 1085 1086 1087 1088 1089
	r = amdgpu_cs_dependencies(adev, &parser);
	if (r) {
		DRM_ERROR("Failed in the dependencies handling %d!\n", r);
		goto out;
	}

1090
	for (i = 0; i < parser.job->num_ibs; i++)
1091
		trace_amdgpu_cs(&parser, i);
1092

1093
	r = amdgpu_cs_ib_vm_chunk(adev, &parser);
1094 1095 1096
	if (r)
		goto out;

C
Christian König 已提交
1097
	r = amdgpu_cs_submit(&parser, cs);
A
Alex Deucher 已提交
1098 1099

out:
1100
	amdgpu_cs_parser_fini(&parser, r, reserved_buffers);
A
Alex Deucher 已提交
1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118
	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);
1119
	struct amdgpu_ring *ring = NULL;
1120
	struct amdgpu_ctx *ctx;
1121
	struct dma_fence *fence;
A
Alex Deucher 已提交
1122 1123
	long r;

1124 1125 1126 1127 1128
	r = amdgpu_cs_get_ring(adev, wait->in.ip_type, wait->in.ip_instance,
			       wait->in.ring, &ring);
	if (r)
		return r;

1129 1130 1131
	ctx = amdgpu_ctx_get(filp->driver_priv, wait->in.ctx_id);
	if (ctx == NULL)
		return -EINVAL;
A
Alex Deucher 已提交
1132

1133 1134 1135 1136
	fence = amdgpu_ctx_get_fence(ctx, ring, wait->in.handle);
	if (IS_ERR(fence))
		r = PTR_ERR(fence);
	else if (fence) {
1137 1138
		r = dma_fence_wait_timeout(fence, true, timeout);
		dma_fence_put(fence);
1139 1140
	} else
		r = 1;
C
Chunming Zhou 已提交
1141

1142
	amdgpu_ctx_put(ctx);
A
Alex Deucher 已提交
1143 1144 1145 1146 1147 1148 1149 1150 1151
	if (r < 0)
		return r;

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

	return 0;
}

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 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325
/**
 * 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
 * @fences: array of drm_amdgpu_fence
 */
static int amdgpu_cs_wait_any_fence(struct amdgpu_device *adev,
				    struct drm_file *filp,
				    union drm_amdgpu_wait_fences *wait,
				    struct drm_amdgpu_fence *fences)
{
	unsigned long timeout = amdgpu_gem_timeout(wait->in.timeout_ns);
	uint32_t fence_count = wait->in.fence_count;
	uint32_t first = ~0;
	struct dma_fence **array;
	unsigned int i;
	long r;

	/* Prepare the fence array */
	array = kcalloc(fence_count, sizeof(struct dma_fence *), GFP_KERNEL);

	if (array == NULL)
		return -ENOMEM;

	for (i = 0; i < fence_count; i++) {
		struct dma_fence *fence;

		fence = amdgpu_cs_get_fence(adev, filp, &fences[i]);
		if (IS_ERR(fence)) {
			r = PTR_ERR(fence);
			goto err_free_fence_array;
		} else if (fence) {
			array[i] = fence;
		} else { /* NULL, the fence has been already signaled */
			r = 1;
			goto out;
		}
	}

	r = dma_fence_wait_any_timeout(array, fence_count, true, timeout,
				       &first);
	if (r < 0)
		goto err_free_fence_array;

out:
	memset(wait, 0, sizeof(*wait));
	wait->out.status = (r > 0);
	wait->out.first_signaled = first;
	/* set return value 0 to indicate success */
	r = 0;

err_free_fence_array:
	for (i = 0; i < fence_count; i++)
		dma_fence_put(array[i]);
	kfree(array);

	return r;
}

/**
 * amdgpu_cs_wait_fences_ioctl - wait for multiple command submissions to finish
 *
 * @dev: drm device
 * @data: data from userspace
 * @filp: file private
 */
int amdgpu_cs_wait_fences_ioctl(struct drm_device *dev, void *data,
				struct drm_file *filp)
{
	struct amdgpu_device *adev = dev->dev_private;
	union drm_amdgpu_wait_fences *wait = data;
	uint32_t fence_count = wait->in.fence_count;
	struct drm_amdgpu_fence *fences_user;
	struct drm_amdgpu_fence *fences;
	int r;

	/* Get the fences from userspace */
	fences = kmalloc_array(fence_count, sizeof(struct drm_amdgpu_fence),
			GFP_KERNEL);
	if (fences == NULL)
		return -ENOMEM;

	fences_user = (void __user *)(unsigned long)(wait->in.fences);
	if (copy_from_user(fences, fences_user,
		sizeof(struct drm_amdgpu_fence) * fence_count)) {
		r = -EFAULT;
		goto err_free_fences;
	}

	if (wait->in.wait_all)
		r = amdgpu_cs_wait_all_fences(adev, filp, wait, fences);
	else
		r = amdgpu_cs_wait_any_fence(adev, filp, wait, fences);

err_free_fences:
	kfree(fences);

	return r;
}

A
Alex Deucher 已提交
1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341
/**
 * 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;
1342 1343 1344 1345
	unsigned i;

	if (!parser->bo_list)
		return NULL;
A
Alex Deucher 已提交
1346 1347 1348

	addr /= AMDGPU_GPU_PAGE_SIZE;

1349 1350 1351 1352 1353
	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
1354 1355
			continue;

1356
		list_for_each_entry(mapping, &lobj->bo_va->valids, list) {
1357 1358 1359 1360
			if (mapping->it.start > addr ||
			    addr > mapping->it.last)
				continue;

1361
			*bo = lobj->bo_va->bo;
1362 1363 1364
			return mapping;
		}

1365
		list_for_each_entry(mapping, &lobj->bo_va->invalids, list) {
A
Alex Deucher 已提交
1366 1367 1368 1369
			if (mapping->it.start > addr ||
			    addr > mapping->it.last)
				continue;

1370
			*bo = lobj->bo_va->bo;
A
Alex Deucher 已提交
1371 1372 1373 1374 1375 1376
			return mapping;
		}
	}

	return NULL;
}
1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395

/**
 * 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;

1396
		r = amdgpu_ttm_bind(&bo->tbo, &bo->tbo.mem);
1397 1398
		if (unlikely(r))
			return r;
1399 1400 1401 1402 1403 1404 1405 1406 1407

		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;
1408 1409 1410 1411
	}

	return 0;
}