ttm_bo.c 44.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
/**************************************************************************
 *
 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
 * 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, sub license, 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 NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE COPYRIGHT HOLDERS, AUTHORS 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: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
 */

J
Joe Perches 已提交
31 32
#define pr_fmt(fmt) "[TTM] " fmt

33 34 35
#include <drm/ttm/ttm_module.h>
#include <drm/ttm/ttm_bo_driver.h>
#include <drm/ttm/ttm_placement.h>
36 37 38 39 40 41
#include <linux/jiffies.h>
#include <linux/slab.h>
#include <linux/sched.h>
#include <linux/mm.h>
#include <linux/file.h>
#include <linux/module.h>
A
Arun Sharma 已提交
42
#include <linux/atomic.h>
43
#include <linux/reservation.h>
44

45 46 47 48 49 50 51
static void ttm_bo_global_kobj_release(struct kobject *kobj);

static struct attribute ttm_bo_count = {
	.name = "bo_count",
	.mode = S_IRUGO
};

52 53
static inline int ttm_mem_type_from_place(const struct ttm_place *place,
					  uint32_t *mem_type)
54
{
55
	int pos;
56

57 58 59 60 61 62
	pos = ffs(place->flags & TTM_PL_MASK_MEM);
	if (unlikely(!pos))
		return -EINVAL;

	*mem_type = pos - 1;
	return 0;
63 64
}

65
static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
66
{
67
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
68
	struct drm_printer p = drm_debug_printer(TTM_PFX);
69

J
Joe Perches 已提交
70 71 72
	pr_err("    has_type: %d\n", man->has_type);
	pr_err("    use_type: %d\n", man->use_type);
	pr_err("    flags: 0x%08X\n", man->flags);
73
	pr_err("    gpu_offset: 0x%08llX\n", man->gpu_offset);
J
Joe Perches 已提交
74 75 76
	pr_err("    size: %llu\n", man->size);
	pr_err("    available_caching: 0x%08X\n", man->available_caching);
	pr_err("    default_caching: 0x%08X\n", man->default_caching);
77
	if (mem_type != TTM_PL_SYSTEM)
78
		(*man->func->debug)(man, &p);
79 80 81 82 83 84 85
}

static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
					struct ttm_placement *placement)
{
	int i, ret, mem_type;

J
Joe Perches 已提交
86 87 88
	pr_err("No space for %p (%lu pages, %luK, %luM)\n",
	       bo, bo->mem.num_pages, bo->mem.size >> 10,
	       bo->mem.size >> 20);
89
	for (i = 0; i < placement->num_placement; i++) {
90
		ret = ttm_mem_type_from_place(&placement->placement[i],
91 92 93
						&mem_type);
		if (ret)
			return;
J
Joe Perches 已提交
94
		pr_err("  placement[%d]=0x%08X (%d)\n",
95
		       i, placement->placement[i].flags, mem_type);
96
		ttm_mem_type_debug(bo->bdev, mem_type);
97 98 99
	}
}

100 101 102 103 104 105 106
static ssize_t ttm_bo_global_show(struct kobject *kobj,
				  struct attribute *attr,
				  char *buffer)
{
	struct ttm_bo_global *glob =
		container_of(kobj, struct ttm_bo_global, kobj);

107 108
	return snprintf(buffer, PAGE_SIZE, "%d\n",
				atomic_read(&glob->bo_count));
109 110 111 112 113 114 115
}

static struct attribute *ttm_bo_global_attrs[] = {
	&ttm_bo_count,
	NULL
};

116
static const struct sysfs_ops ttm_bo_global_ops = {
117 118 119 120 121 122 123 124 125
	.show = &ttm_bo_global_show
};

static struct kobj_type ttm_bo_glob_kobj_type  = {
	.release = &ttm_bo_global_kobj_release,
	.sysfs_ops = &ttm_bo_global_ops,
	.default_attrs = ttm_bo_global_attrs
};

126 127 128 129 130 131 132 133 134 135 136

static inline uint32_t ttm_bo_type_flags(unsigned type)
{
	return 1 << (type);
}

static void ttm_bo_release_list(struct kref *list_kref)
{
	struct ttm_buffer_object *bo =
	    container_of(list_kref, struct ttm_buffer_object, list_kref);
	struct ttm_bo_device *bdev = bo->bdev;
137
	size_t acc_size = bo->acc_size;
138

139 140
	BUG_ON(kref_read(&bo->list_kref));
	BUG_ON(kref_read(&bo->kref));
141 142 143 144
	BUG_ON(atomic_read(&bo->cpu_writers));
	BUG_ON(bo->mem.mm_node != NULL);
	BUG_ON(!list_empty(&bo->lru));
	BUG_ON(!list_empty(&bo->ddestroy));
145
	ttm_tt_destroy(bo->ttm);
146
	atomic_dec(&bo->glob->bo_count);
147
	dma_fence_put(bo->moving);
148
	reservation_object_fini(&bo->ttm_resv);
149
	mutex_destroy(&bo->wu_mutex);
150 151 152 153 154
	if (bo->destroy)
		bo->destroy(bo);
	else {
		kfree(bo);
	}
155
	ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
156 157
}

158
void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
159 160
{
	struct ttm_bo_device *bdev = bo->bdev;
161
	struct ttm_mem_type_manager *man;
162

163
	reservation_object_assert_held(bo->resv);
164 165 166 167 168

	if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {

		BUG_ON(!list_empty(&bo->lru));

169 170
		man = &bdev->man[bo->mem.mem_type];
		list_add_tail(&bo->lru, &man->lru[bo->priority]);
171 172
		kref_get(&bo->list_kref);

173 174
		if (bo->ttm && !(bo->ttm->page_flags &
				 (TTM_PAGE_FLAG_SG | TTM_PAGE_FLAG_SWAPPED))) {
175 176
			list_add_tail(&bo->swap,
				      &bo->glob->swap_lru[bo->priority]);
177 178 179 180
			kref_get(&bo->list_kref);
		}
	}
}
181
EXPORT_SYMBOL(ttm_bo_add_to_lru);
182

183 184 185 186 187 188
static void ttm_bo_ref_bug(struct kref *list_kref)
{
	BUG();
}

void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
189 190 191
{
	if (!list_empty(&bo->swap)) {
		list_del_init(&bo->swap);
192
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
193 194 195
	}
	if (!list_empty(&bo->lru)) {
		list_del_init(&bo->lru);
196
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
197 198
	}

199 200 201 202
	/*
	 * TODO: Add a driver hook to delete from
	 * driver-specific LRU's here.
	 */
203 204
}

205
void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
206
{
207
	spin_lock(&bo->glob->lru_lock);
208
	ttm_bo_del_from_lru(bo);
209
	spin_unlock(&bo->glob->lru_lock);
210
}
211
EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
212

213 214
void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
{
215
	reservation_object_assert_held(bo->resv);
216

217
	ttm_bo_del_from_lru(bo);
218
	ttm_bo_add_to_lru(bo);
219 220 221
}
EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);

222 223 224 225 226 227
/*
 * Call bo->mutex locked.
 */
static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
{
	struct ttm_bo_device *bdev = bo->bdev;
228
	struct ttm_bo_global *glob = bo->glob;
229 230 231
	int ret = 0;
	uint32_t page_flags = 0;

232
	reservation_object_assert_held(bo->resv);
233 234
	bo->ttm = NULL;

D
Dave Airlie 已提交
235 236 237
	if (bdev->need_dma32)
		page_flags |= TTM_PAGE_FLAG_DMA32;

238 239 240 241 242
	switch (bo->type) {
	case ttm_bo_type_device:
		if (zero_alloc)
			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
	case ttm_bo_type_kernel:
243 244
		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
						      page_flags, glob->dummy_read_page);
245 246 247
		if (unlikely(bo->ttm == NULL))
			ret = -ENOMEM;
		break;
248 249 250 251 252 253 254 255 256 257
	case ttm_bo_type_sg:
		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
						      page_flags | TTM_PAGE_FLAG_SG,
						      glob->dummy_read_page);
		if (unlikely(bo->ttm == NULL)) {
			ret = -ENOMEM;
			break;
		}
		bo->ttm->sg = bo->sg;
		break;
258
	default:
J
Joe Perches 已提交
259
		pr_err("Illegal buffer object type\n");
260 261 262 263 264 265 266 267
		ret = -EINVAL;
		break;
	}

	return ret;
}

static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
268 269
				  struct ttm_mem_reg *mem, bool evict,
				  struct ttm_operation_ctx *ctx)
270 271 272 273 274 275 276 277 278
{
	struct ttm_bo_device *bdev = bo->bdev;
	bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
	bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
	struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
	struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
	int ret = 0;

	if (old_is_pci || new_is_pci ||
279 280 281 282 283 284 285
	    ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
		ret = ttm_mem_io_lock(old_man, true);
		if (unlikely(ret != 0))
			goto out_err;
		ttm_bo_unmap_virtual_locked(bo);
		ttm_mem_io_unlock(old_man);
	}
286 287 288 289 290

	/*
	 * Create and bind a ttm if required.
	 */

291 292
	if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
		if (bo->ttm == NULL) {
293 294
			bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
			ret = ttm_bo_add_ttm(bo, zero);
295 296 297
			if (ret)
				goto out_err;
		}
298 299 300

		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
		if (ret)
301
			goto out_err;
302 303

		if (mem->mem_type != TTM_PL_SYSTEM) {
304
			ret = ttm_tt_bind(bo->ttm, mem, ctx);
305 306 307 308 309
			if (ret)
				goto out_err;
		}

		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
310
			if (bdev->driver->move_notify)
311
				bdev->driver->move_notify(bo, evict, mem);
312
			bo->mem = *mem;
313 314 315 316 317
			mem->mm_node = NULL;
			goto moved;
		}
	}

318
	if (bdev->driver->move_notify)
319
		bdev->driver->move_notify(bo, evict, mem);
320

321 322
	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
323
		ret = ttm_bo_move_ttm(bo, ctx, mem);
324
	else if (bdev->driver->move)
325
		ret = bdev->driver->move(bo, evict, ctx, mem);
326
	else
327
		ret = ttm_bo_move_memcpy(bo, ctx, mem);
328

329 330 331 332 333
	if (ret) {
		if (bdev->driver->move_notify) {
			struct ttm_mem_reg tmp_mem = *mem;
			*mem = bo->mem;
			bo->mem = tmp_mem;
334
			bdev->driver->move_notify(bo, false, mem);
335
			bo->mem = *mem;
336
			*mem = tmp_mem;
337
		}
338

339 340
		goto out_err;
	}
341

342 343
moved:
	if (bo->evicted) {
344 345 346 347 348
		if (bdev->driver->invalidate_caches) {
			ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
			if (ret)
				pr_err("Can not flush read caches\n");
		}
349 350 351
		bo->evicted = false;
	}

C
Christian König 已提交
352
	if (bo->mem.mm_node)
353
		bo->offset = (bo->mem.start << PAGE_SHIFT) +
354
		    bdev->man[bo->mem.mem_type].gpu_offset;
C
Christian König 已提交
355
	else
356
		bo->offset = 0;
357

358
	ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
359 360 361 362
	return 0;

out_err:
	new_man = &bdev->man[bo->mem.mem_type];
363
	if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
364 365 366 367 368 369 370
		ttm_tt_destroy(bo->ttm);
		bo->ttm = NULL;
	}

	return ret;
}

371
/**
372
 * Call bo::reserved.
373
 * Will release GPU memory type usage on destruction.
374 375 376
 * This is the place to put in driver specific hooks to release
 * driver private resources.
 * Will release the bo::reserved lock.
377 378 379 380
 */

static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
{
381
	if (bo->bdev->driver->move_notify)
382
		bo->bdev->driver->move_notify(bo, false, NULL);
383

384 385
	ttm_tt_destroy(bo->ttm);
	bo->ttm = NULL;
386
	ttm_bo_mem_put(bo, &bo->mem);
387 388
}

389 390 391 392 393 394 395
static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
{
	int r;

	if (bo->resv == &bo->ttm_resv)
		return 0;

396
	BUG_ON(!reservation_object_trylock(&bo->ttm_resv));
397 398

	r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv);
399
	if (r)
400 401 402 403 404
		reservation_object_unlock(&bo->ttm_resv);

	return r;
}

405 406 407
static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
{
	struct reservation_object_list *fobj;
408
	struct dma_fence *fence;
409 410
	int i;

411 412
	fobj = reservation_object_get_list(&bo->ttm_resv);
	fence = reservation_object_get_excl(&bo->ttm_resv);
413
	if (fence && !fence->ops->signaled)
414
		dma_fence_enable_sw_signaling(fence);
415 416 417 418 419 420

	for (i = 0; fobj && i < fobj->shared_count; ++i) {
		fence = rcu_dereference_protected(fobj->shared[i],
					reservation_object_held(bo->resv));

		if (!fence->ops->signaled)
421
			dma_fence_enable_sw_signaling(fence);
422 423 424
	}
}

425
static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
426 427
{
	struct ttm_bo_device *bdev = bo->bdev;
428
	struct ttm_bo_global *glob = bo->glob;
429 430
	int ret;

431 432 433 434 435 436 437 438 439 440 441
	ret = ttm_bo_individualize_resv(bo);
	if (ret) {
		/* Last resort, if we fail to allocate memory for the
		 * fences block for the BO to become idle
		 */
		reservation_object_wait_timeout_rcu(bo->resv, true, false,
						    30 * HZ);
		spin_lock(&glob->lru_lock);
		goto error;
	}

442
	spin_lock(&glob->lru_lock);
443
	ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY;
M
Maarten Lankhorst 已提交
444
	if (!ret) {
445
		if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) {
446
			ttm_bo_del_from_lru(bo);
M
Maarten Lankhorst 已提交
447
			spin_unlock(&glob->lru_lock);
448
			if (bo->resv != &bo->ttm_resv)
449
				reservation_object_unlock(&bo->ttm_resv);
450

M
Maarten Lankhorst 已提交
451
			ttm_bo_cleanup_memtype_use(bo);
452
			reservation_object_unlock(bo->resv);
M
Maarten Lankhorst 已提交
453
			return;
454 455 456
		}

		ttm_bo_flush_all_fences(bo);
457 458 459 460 461 462 463 464 465 466 467

		/*
		 * Make NO_EVICT bos immediately available to
		 * shrinkers, now that they are queued for
		 * destruction.
		 */
		if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
			bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
			ttm_bo_add_to_lru(bo);
		}

468
		reservation_object_unlock(bo->resv);
469
	}
470 471
	if (bo->resv != &bo->ttm_resv)
		reservation_object_unlock(&bo->ttm_resv);
472

473
error:
474 475 476 477 478 479 480 481 482
	kref_get(&bo->list_kref);
	list_add_tail(&bo->ddestroy, &bdev->ddestroy);
	spin_unlock(&glob->lru_lock);

	schedule_delayed_work(&bdev->wq,
			      ((HZ / 100) < 1) ? 1 : HZ / 100);
}

/**
483
 * function ttm_bo_cleanup_refs
484 485 486
 * If bo idle, remove from delayed- and lru lists, and unref.
 * If not idle, do nothing.
 *
487
 * Must be called with lru_lock and reservation held, this function
488
 * will drop the lru lock and optionally the reservation lock before returning.
489
 *
490 491
 * @interruptible         Any sleeps should occur interruptibly.
 * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
492
 * @unlock_resv           Unlock the reservation lock as well.
493 494
 */

495 496 497
static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
			       bool interruptible, bool no_wait_gpu,
			       bool unlock_resv)
498 499
{
	struct ttm_bo_global *glob = bo->glob;
500
	struct reservation_object *resv;
501
	int ret;
502

503 504 505 506 507 508 509 510 511
	if (unlikely(list_empty(&bo->ddestroy)))
		resv = bo->resv;
	else
		resv = &bo->ttm_resv;

	if (reservation_object_test_signaled_rcu(resv, true))
		ret = 0;
	else
		ret = -EBUSY;
512

513
	if (ret && !no_wait_gpu) {
M
Maarten Lankhorst 已提交
514
		long lret;
515

516 517
		if (unlock_resv)
			reservation_object_unlock(bo->resv);
M
Maarten Lankhorst 已提交
518 519
		spin_unlock(&glob->lru_lock);

520
		lret = reservation_object_wait_timeout_rcu(resv, true,
M
Maarten Lankhorst 已提交
521 522 523 524 525 526 527
							   interruptible,
							   30 * HZ);

		if (lret < 0)
			return lret;
		else if (lret == 0)
			return -EBUSY;
528

529
		spin_lock(&glob->lru_lock);
530 531 532 533 534 535 536 537 538
		if (unlock_resv && !reservation_object_trylock(bo->resv)) {
			/*
			 * We raced, and lost, someone else holds the reservation now,
			 * and is probably busy in ttm_bo_cleanup_memtype_use.
			 *
			 * Even if it's not the case, because we finished waiting any
			 * delayed destruction would succeed, so just return success
			 * here.
			 */
539 540 541
			spin_unlock(&glob->lru_lock);
			return 0;
		}
542
		ret = 0;
543
	}
544

545
	if (ret || unlikely(list_empty(&bo->ddestroy))) {
546 547
		if (unlock_resv)
			reservation_object_unlock(bo->resv);
548
		spin_unlock(&glob->lru_lock);
549
		return ret;
550 551
	}

552
	ttm_bo_del_from_lru(bo);
553
	list_del_init(&bo->ddestroy);
554
	kref_put(&bo->list_kref, ttm_bo_ref_bug);
555 556 557

	spin_unlock(&glob->lru_lock);
	ttm_bo_cleanup_memtype_use(bo);
558 559 560

	if (unlock_resv)
		reservation_object_unlock(bo->resv);
561 562

	return 0;
563 564 565 566 567 568
}

/**
 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 * encountered buffers.
 */
569
static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
570
{
571
	struct ttm_bo_global *glob = bdev->glob;
572 573
	struct list_head removed;
	bool empty;
574

575
	INIT_LIST_HEAD(&removed);
576

577 578 579
	spin_lock(&glob->lru_lock);
	while (!list_empty(&bdev->ddestroy)) {
		struct ttm_buffer_object *bo;
580

581 582 583 584
		bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
				      ddestroy);
		kref_get(&bo->list_kref);
		list_move_tail(&bo->ddestroy, &removed);
585

586 587 588
		if (remove_all || bo->resv != &bo->ttm_resv) {
			spin_unlock(&glob->lru_lock);
			reservation_object_lock(bo->resv, NULL);
589

590 591 592 593 594
			spin_lock(&glob->lru_lock);
			ttm_bo_cleanup_refs(bo, false, !remove_all, true);

		} else if (reservation_object_trylock(bo->resv)) {
			ttm_bo_cleanup_refs(bo, false, !remove_all, true);
595 596
		} else {
			spin_unlock(&glob->lru_lock);
597
		}
598

599
		kref_put(&bo->list_kref, ttm_bo_release_list);
600
		spin_lock(&glob->lru_lock);
601
	}
602 603
	list_splice_tail(&removed, &bdev->ddestroy);
	empty = list_empty(&bdev->ddestroy);
604
	spin_unlock(&glob->lru_lock);
605 606

	return empty;
607 608 609 610 611 612 613
}

static void ttm_bo_delayed_workqueue(struct work_struct *work)
{
	struct ttm_bo_device *bdev =
	    container_of(work, struct ttm_bo_device, wq.work);

614
	if (!ttm_bo_delayed_delete(bdev, false)) {
615 616 617 618 619 620 621 622 623 624
		schedule_delayed_work(&bdev->wq,
				      ((HZ / 100) < 1) ? 1 : HZ / 100);
	}
}

static void ttm_bo_release(struct kref *kref)
{
	struct ttm_buffer_object *bo =
	    container_of(kref, struct ttm_buffer_object, kref);
	struct ttm_bo_device *bdev = bo->bdev;
625
	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
626

627
	drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
628 629 630
	ttm_mem_io_lock(man, false);
	ttm_mem_io_free_vm(bo);
	ttm_mem_io_unlock(man);
631
	ttm_bo_cleanup_refs_or_queue(bo);
632 633 634 635 636 637 638 639 640 641 642 643
	kref_put(&bo->list_kref, ttm_bo_release_list);
}

void ttm_bo_unref(struct ttm_buffer_object **p_bo)
{
	struct ttm_buffer_object *bo = *p_bo;

	*p_bo = NULL;
	kref_put(&bo->kref, ttm_bo_release);
}
EXPORT_SYMBOL(ttm_bo_unref);

644 645 646 647 648 649 650 651 652 653 654 655 656 657
int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
{
	return cancel_delayed_work_sync(&bdev->wq);
}
EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);

void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
{
	if (resched)
		schedule_delayed_work(&bdev->wq,
				      ((HZ / 100) < 1) ? 1 : HZ / 100);
}
EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);

658 659
static int ttm_bo_evict(struct ttm_buffer_object *bo,
			struct ttm_operation_ctx *ctx)
660 661 662
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_reg evict_mem;
663 664
	struct ttm_placement placement;
	int ret = 0;
665

666
	reservation_object_assert_held(bo->resv);
667 668 669

	evict_mem = bo->mem;
	evict_mem.mm_node = NULL;
670 671
	evict_mem.bus.io_reserved_vm = false;
	evict_mem.bus.io_reserved_count = 0;
672

673 674
	placement.num_placement = 0;
	placement.num_busy_placement = 0;
675
	bdev->driver->evict_flags(bo, &placement);
676
	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
677
	if (ret) {
678
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
679 680
			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
			       bo);
681 682
			ttm_bo_mem_space_debug(bo, &placement);
		}
683 684 685
		goto out;
	}

686
	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
687
	if (unlikely(ret)) {
688
		if (ret != -ERESTARTSYS)
J
Joe Perches 已提交
689
			pr_err("Buffer eviction failed\n");
690
		ttm_bo_mem_put(bo, &evict_mem);
691 692
		goto out;
	}
693 694 695 696 697
	bo->evicted = true;
out:
	return ret;
}

698 699 700 701 702 703 704 705 706 707 708 709 710 711
bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
			      const struct ttm_place *place)
{
	/* Don't evict this BO if it's outside of the
	 * requested placement range
	 */
	if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
	    (place->lpfn && place->lpfn <= bo->mem.start))
		return false;

	return true;
}
EXPORT_SYMBOL(ttm_bo_eviction_valuable);

712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739
/**
 * Check the target bo is allowable to be evicted or swapout, including cases:
 *
 * a. if share same reservation object with ctx->resv, have assumption
 * reservation objects should already be locked, so not lock again and
 * return true directly when either the opreation allow_reserved_eviction
 * or the target bo already is in delayed free list;
 *
 * b. Otherwise, trylock it.
 */
static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
			struct ttm_operation_ctx *ctx, bool *locked)
{
	bool ret = false;

	*locked = false;
	if (bo->resv == ctx->resv) {
		reservation_object_assert_held(bo->resv);
		if (ctx->allow_reserved_eviction || !list_empty(&bo->ddestroy))
			ret = true;
	} else {
		*locked = reservation_object_trylock(bo->resv);
		ret = *locked;
	}

	return ret;
}

740
static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
741 742
			       uint32_t mem_type,
			       const struct ttm_place *place,
743
			       struct ttm_operation_ctx *ctx)
744 745 746
{
	struct ttm_bo_global *glob = bdev->glob;
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
747 748
	struct ttm_buffer_object *bo = NULL;
	bool locked = false;
749
	unsigned i;
750
	int ret;
751

752
	spin_lock(&glob->lru_lock);
753 754
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &man->lru[i], lru) {
755 756
			if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked))
				continue;
757

758 759
			if (place && !bdev->driver->eviction_valuable(bo,
								      place)) {
760 761
				if (locked)
					reservation_object_unlock(bo->resv);
762 763 764
				continue;
			}
			break;
765
		}
766

767 768
		/* If the inner loop terminated early, we have our candidate */
		if (&bo->lru != &man->lru[i])
769
			break;
770 771

		bo = NULL;
772 773
	}

774
	if (!bo) {
775
		spin_unlock(&glob->lru_lock);
776
		return -EBUSY;
777 778
	}

779
	kref_get(&bo->list_kref);
780

781
	if (!list_empty(&bo->ddestroy)) {
782 783
		ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
					  ctx->no_wait_gpu, locked);
784
		kref_put(&bo->list_kref, ttm_bo_release_list);
785
		return ret;
786 787
	}

788
	ttm_bo_del_from_lru(bo);
789
	spin_unlock(&glob->lru_lock);
790

791
	ret = ttm_bo_evict(bo, ctx);
792
	if (locked) {
793
		ttm_bo_unreserve(bo);
794 795
	} else {
		spin_lock(&glob->lru_lock);
796
		ttm_bo_add_to_lru(bo);
797 798
		spin_unlock(&glob->lru_lock);
	}
799

800
	kref_put(&bo->list_kref, ttm_bo_release_list);
801 802 803
	return ret;
}

804 805
void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
{
806
	struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
807

808 809
	if (mem->mm_node)
		(*man->func->put_node)(man, mem);
810 811 812
}
EXPORT_SYMBOL(ttm_bo_mem_put);

813 814 815 816 817 818 819
/**
 * Add the last move fence to the BO and reserve a new shared slot.
 */
static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
				 struct ttm_mem_type_manager *man,
				 struct ttm_mem_reg *mem)
{
820
	struct dma_fence *fence;
821 822 823
	int ret;

	spin_lock(&man->move_lock);
824
	fence = dma_fence_get(man->move);
825 826 827 828 829 830 831 832 833
	spin_unlock(&man->move_lock);

	if (fence) {
		reservation_object_add_shared_fence(bo->resv, fence);

		ret = reservation_object_reserve_shared(bo->resv);
		if (unlikely(ret))
			return ret;

834
		dma_fence_put(bo->moving);
835 836 837 838 839 840
		bo->moving = fence;
	}

	return 0;
}

841 842 843 844
/**
 * Repeatedly evict memory from the LRU for @mem_type until we create enough
 * space, or we've evicted everything and there isn't enough space.
 */
845 846
static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
					uint32_t mem_type,
847
					const struct ttm_place *place,
848
					struct ttm_mem_reg *mem,
849
					struct ttm_operation_ctx *ctx)
850
{
851
	struct ttm_bo_device *bdev = bo->bdev;
852 853 854 855
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	int ret;

	do {
856
		ret = (*man->func->get_node)(man, bo, place, mem);
857 858
		if (unlikely(ret != 0))
			return ret;
859
		if (mem->mm_node)
860
			break;
R
Roger He 已提交
861
		ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
862 863 864 865
		if (unlikely(ret != 0))
			return ret;
	} while (1);
	mem->mem_type = mem_type;
866
	return ttm_bo_add_move_fence(bo, man, mem);
867 868
}

869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893
static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
				      uint32_t cur_placement,
				      uint32_t proposed_placement)
{
	uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
	uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;

	/**
	 * Keep current caching if possible.
	 */

	if ((cur_placement & caching) != 0)
		result |= (cur_placement & caching);
	else if ((man->default_caching & caching) != 0)
		result |= man->default_caching;
	else if ((TTM_PL_FLAG_CACHED & caching) != 0)
		result |= TTM_PL_FLAG_CACHED;
	else if ((TTM_PL_FLAG_WC & caching) != 0)
		result |= TTM_PL_FLAG_WC;
	else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
		result |= TTM_PL_FLAG_UNCACHED;

	return result;
}

894 895
static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
				 uint32_t mem_type,
896
				 const struct ttm_place *place,
897
				 uint32_t *masked_placement)
898 899 900
{
	uint32_t cur_flags = ttm_bo_type_flags(mem_type);

901
	if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
902 903
		return false;

904
	if ((place->flags & man->available_caching) == 0)
905 906
		return false;

907
	cur_flags |= (place->flags & man->available_caching);
908 909

	*masked_placement = cur_flags;
910 911 912 913 914 915 916 917 918 919 920 921
	return true;
}

/**
 * Creates space for memory region @mem according to its type.
 *
 * This function first searches for free space in compatible memory types in
 * the priority order defined by the driver.  If free space isn't found, then
 * ttm_bo_mem_force_space is attempted in priority order to evict and find
 * space.
 */
int ttm_bo_mem_space(struct ttm_buffer_object *bo,
922 923
			struct ttm_placement *placement,
			struct ttm_mem_reg *mem,
924
			struct ttm_operation_ctx *ctx)
925 926 927 928 929 930 931
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_type_manager *man;
	uint32_t mem_type = TTM_PL_SYSTEM;
	uint32_t cur_flags = 0;
	bool type_found = false;
	bool type_ok = false;
932
	bool has_erestartsys = false;
933
	int i, ret;
934

935 936 937 938
	ret = reservation_object_reserve_shared(bo->resv);
	if (unlikely(ret))
		return ret;

939
	mem->mm_node = NULL;
940
	for (i = 0; i < placement->num_placement; ++i) {
941 942 943
		const struct ttm_place *place = &placement->placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
944 945
		if (ret)
			return ret;
946
		man = &bdev->man[mem_type];
947 948
		if (!man->has_type || !man->use_type)
			continue;
949

950
		type_ok = ttm_bo_mt_compatible(man, mem_type, place,
951
						&cur_flags);
952 953 954 955

		if (!type_ok)
			continue;

956
		type_found = true;
957 958
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
959 960 961 962
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
963
		ttm_flag_masked(&cur_flags, place->flags,
964
				~TTM_PL_MASK_MEMTYPE);
965

966 967 968
		if (mem_type == TTM_PL_SYSTEM)
			break;

969 970 971
		ret = (*man->func->get_node)(man, bo, place, mem);
		if (unlikely(ret))
			return ret;
972 973 974 975 976 977 978

		if (mem->mm_node) {
			ret = ttm_bo_add_move_fence(bo, man, mem);
			if (unlikely(ret)) {
				(*man->func->put_node)(man, mem);
				return ret;
			}
979
			break;
980
		}
981 982
	}

983
	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
984 985 986 987 988
		mem->mem_type = mem_type;
		mem->placement = cur_flags;
		return 0;
	}

989
	for (i = 0; i < placement->num_busy_placement; ++i) {
990 991 992
		const struct ttm_place *place = &placement->busy_placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
993 994
		if (ret)
			return ret;
995
		man = &bdev->man[mem_type];
996
		if (!man->has_type || !man->use_type)
997
			continue;
998
		if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
999 1000
			continue;

1001
		type_found = true;
1002 1003
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
1004 1005 1006 1007
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
1008
		ttm_flag_masked(&cur_flags, place->flags,
1009
				~TTM_PL_MASK_MEMTYPE);
1010

1011 1012 1013 1014 1015 1016 1017
		if (mem_type == TTM_PL_SYSTEM) {
			mem->mem_type = mem_type;
			mem->placement = cur_flags;
			mem->mm_node = NULL;
			return 0;
		}

1018
		ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx);
1019 1020 1021 1022
		if (ret == 0 && mem->mm_node) {
			mem->placement = cur_flags;
			return 0;
		}
1023 1024
		if (ret == -ERESTARTSYS)
			has_erestartsys = true;
1025
	}
1026 1027

	if (!type_found) {
1028
		pr_err(TTM_PFX "No compatible memory type found\n");
1029 1030 1031 1032
		return -EINVAL;
	}

	return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1033 1034 1035
}
EXPORT_SYMBOL(ttm_bo_mem_space);

1036
static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1037 1038
			      struct ttm_placement *placement,
			      struct ttm_operation_ctx *ctx)
1039 1040 1041 1042
{
	int ret = 0;
	struct ttm_mem_reg mem;

1043
	reservation_object_assert_held(bo->resv);
1044 1045 1046 1047

	mem.num_pages = bo->num_pages;
	mem.size = mem.num_pages << PAGE_SHIFT;
	mem.page_alignment = bo->mem.page_alignment;
1048 1049
	mem.bus.io_reserved_vm = false;
	mem.bus.io_reserved_count = 0;
1050 1051 1052
	/*
	 * Determine where to move the buffer.
	 */
1053
	ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1054 1055
	if (ret)
		goto out_unlock;
1056
	ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1057
out_unlock:
1058 1059
	if (ret && mem.mm_node)
		ttm_bo_mem_put(bo, &mem);
1060 1061 1062
	return ret;
}

1063 1064 1065 1066
static bool ttm_bo_places_compat(const struct ttm_place *places,
				 unsigned num_placement,
				 struct ttm_mem_reg *mem,
				 uint32_t *new_flags)
1067
{
1068
	unsigned i;
1069

1070 1071
	for (i = 0; i < num_placement; i++) {
		const struct ttm_place *heap = &places[i];
1072

1073
		if (mem->mm_node && (mem->start < heap->fpfn ||
1074
		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1075 1076 1077
			continue;

		*new_flags = heap->flags;
1078
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1079 1080 1081
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
		    (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
		     (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1082
			return true;
1083
	}
1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100
	return false;
}

bool ttm_bo_mem_compat(struct ttm_placement *placement,
		       struct ttm_mem_reg *mem,
		       uint32_t *new_flags)
{
	if (ttm_bo_places_compat(placement->placement, placement->num_placement,
				 mem, new_flags))
		return true;

	if ((placement->busy_placement != placement->placement ||
	     placement->num_busy_placement > placement->num_placement) &&
	    ttm_bo_places_compat(placement->busy_placement,
				 placement->num_busy_placement,
				 mem, new_flags))
		return true;
1101 1102

	return false;
1103
}
1104
EXPORT_SYMBOL(ttm_bo_mem_compat);
1105

1106
int ttm_bo_validate(struct ttm_buffer_object *bo,
1107 1108
		    struct ttm_placement *placement,
		    struct ttm_operation_ctx *ctx)
1109 1110
{
	int ret;
1111
	uint32_t new_flags;
1112

1113
	reservation_object_assert_held(bo->resv);
1114 1115 1116
	/*
	 * Check whether we need to move buffer.
	 */
1117
	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1118
		ret = ttm_bo_move_buffer(bo, placement, ctx);
1119
		if (ret)
1120
			return ret;
1121 1122 1123 1124 1125
	} else {
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the compatible memory placement flags to the active flags
		 */
1126
		ttm_flag_masked(&bo->mem.placement, new_flags,
1127
				~TTM_PL_MASK_MEMTYPE);
1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138
	}
	/*
	 * We might need to add a TTM.
	 */
	if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
		ret = ttm_bo_add_ttm(bo, true);
		if (ret)
			return ret;
	}
	return 0;
}
1139
EXPORT_SYMBOL(ttm_bo_validate);
1140

1141 1142 1143 1144 1145 1146
int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
			 struct ttm_buffer_object *bo,
			 unsigned long size,
			 enum ttm_bo_type type,
			 struct ttm_placement *placement,
			 uint32_t page_alignment,
1147
			 struct ttm_operation_ctx *ctx,
1148 1149 1150 1151 1152
			 struct file *persistent_swap_storage,
			 size_t acc_size,
			 struct sg_table *sg,
			 struct reservation_object *resv,
			 void (*destroy) (struct ttm_buffer_object *))
1153
{
1154
	int ret = 0;
1155
	unsigned long num_pages;
1156
	struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1157
	bool locked;
1158

1159
	ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1160
	if (ret) {
J
Joe Perches 已提交
1161
		pr_err("Out of kernel memory\n");
1162 1163 1164 1165 1166 1167
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
		return -ENOMEM;
	}
1168 1169 1170

	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
	if (num_pages == 0) {
J
Joe Perches 已提交
1171
		pr_err("Illegal buffer object size\n");
1172 1173 1174 1175
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
1176
		ttm_mem_global_free(mem_glob, acc_size);
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186
		return -EINVAL;
	}
	bo->destroy = destroy;

	kref_init(&bo->kref);
	kref_init(&bo->list_kref);
	atomic_set(&bo->cpu_writers, 0);
	INIT_LIST_HEAD(&bo->lru);
	INIT_LIST_HEAD(&bo->ddestroy);
	INIT_LIST_HEAD(&bo->swap);
1187
	INIT_LIST_HEAD(&bo->io_reserve_lru);
1188
	mutex_init(&bo->wu_mutex);
1189
	bo->bdev = bdev;
1190
	bo->glob = bdev->glob;
1191 1192
	bo->type = type;
	bo->num_pages = num_pages;
1193
	bo->mem.size = num_pages << PAGE_SHIFT;
1194 1195 1196 1197
	bo->mem.mem_type = TTM_PL_SYSTEM;
	bo->mem.num_pages = bo->num_pages;
	bo->mem.mm_node = NULL;
	bo->mem.page_alignment = page_alignment;
1198 1199
	bo->mem.bus.io_reserved_vm = false;
	bo->mem.bus.io_reserved_count = 0;
1200
	bo->moving = NULL;
1201
	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
J
Jan Engelhardt 已提交
1202
	bo->persistent_swap_storage = persistent_swap_storage;
1203
	bo->acc_size = acc_size;
1204
	bo->sg = sg;
1205 1206
	if (resv) {
		bo->resv = resv;
1207
		reservation_object_assert_held(bo->resv);
1208 1209 1210
	} else {
		bo->resv = &bo->ttm_resv;
	}
1211
	reservation_object_init(&bo->ttm_resv);
1212
	atomic_inc(&bo->glob->bo_count);
1213
	drm_vma_node_reset(&bo->vma_node);
1214
	bo->priority = 0;
1215 1216 1217 1218 1219

	/*
	 * For ttm_bo_type_device buffers, allocate
	 * address space from the device.
	 */
1220 1221
	if (bo->type == ttm_bo_type_device ||
	    bo->type == ttm_bo_type_sg)
1222 1223
		ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
					 bo->mem.num_pages);
1224

1225 1226 1227 1228
	/* passed reservation objects should already be locked,
	 * since otherwise lockdep will be angered in radeon.
	 */
	if (!resv) {
1229
		locked = reservation_object_trylock(bo->resv);
1230 1231
		WARN_ON(!locked);
	}
1232

1233 1234
	if (likely(!ret))
		ret = ttm_bo_validate(bo, placement, ctx);
1235

1236
	if (unlikely(ret)) {
1237 1238 1239
		if (!resv)
			ttm_bo_unreserve(bo);

1240 1241 1242 1243 1244
		ttm_bo_unref(&bo);
		return ret;
	}

	if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1245 1246 1247 1248 1249
		spin_lock(&bo->glob->lru_lock);
		ttm_bo_add_to_lru(bo);
		spin_unlock(&bo->glob->lru_lock);
	}

1250 1251
	return ret;
}
1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266
EXPORT_SYMBOL(ttm_bo_init_reserved);

int ttm_bo_init(struct ttm_bo_device *bdev,
		struct ttm_buffer_object *bo,
		unsigned long size,
		enum ttm_bo_type type,
		struct ttm_placement *placement,
		uint32_t page_alignment,
		bool interruptible,
		struct file *persistent_swap_storage,
		size_t acc_size,
		struct sg_table *sg,
		struct reservation_object *resv,
		void (*destroy) (struct ttm_buffer_object *))
{
1267
	struct ttm_operation_ctx ctx = { interruptible, false };
1268 1269 1270
	int ret;

	ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1271
				   page_alignment, &ctx,
1272 1273 1274 1275 1276 1277 1278 1279 1280 1281
				   persistent_swap_storage, acc_size,
				   sg, resv, destroy);
	if (ret)
		return ret;

	if (!resv)
		ttm_bo_unreserve(bo);

	return 0;
}
1282
EXPORT_SYMBOL(ttm_bo_init);
1283

1284 1285 1286
size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
		       unsigned long bo_size,
		       unsigned struct_size)
1287
{
1288 1289
	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
	size_t size = 0;
1290

1291
	size += ttm_round_pot(struct_size);
F
Felix Kuehling 已提交
1292
	size += ttm_round_pot(npages * sizeof(void *));
1293 1294
	size += ttm_round_pot(sizeof(struct ttm_tt));
	return size;
1295
}
1296 1297 1298 1299 1300 1301 1302 1303 1304 1305
EXPORT_SYMBOL(ttm_bo_acc_size);

size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
			   unsigned long bo_size,
			   unsigned struct_size)
{
	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
	size_t size = 0;

	size += ttm_round_pot(struct_size);
F
Felix Kuehling 已提交
1306
	size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1307 1308 1309 1310
	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
	return size;
}
EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1311

1312 1313 1314 1315 1316 1317
int ttm_bo_create(struct ttm_bo_device *bdev,
			unsigned long size,
			enum ttm_bo_type type,
			struct ttm_placement *placement,
			uint32_t page_alignment,
			bool interruptible,
J
Jan Engelhardt 已提交
1318
			struct file *persistent_swap_storage,
1319
			struct ttm_buffer_object **p_bo)
1320 1321
{
	struct ttm_buffer_object *bo;
1322
	size_t acc_size;
1323
	int ret;
1324 1325

	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1326
	if (unlikely(bo == NULL))
1327 1328
		return -ENOMEM;

1329
	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1330
	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1331
			  interruptible, persistent_swap_storage, acc_size,
1332
			  NULL, NULL, NULL);
1333 1334 1335 1336 1337
	if (likely(ret == 0))
		*p_bo = bo;

	return ret;
}
T
Thomas Hellstrom 已提交
1338
EXPORT_SYMBOL(ttm_bo_create);
1339 1340

static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1341
				   unsigned mem_type)
1342
{
1343
	struct ttm_operation_ctx ctx = { false, false };
1344
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1345
	struct ttm_bo_global *glob = bdev->glob;
1346
	struct dma_fence *fence;
1347
	int ret;
1348
	unsigned i;
1349 1350 1351 1352 1353

	/*
	 * Can't use standard list traversal since we're unlocking.
	 */

1354
	spin_lock(&glob->lru_lock);
1355 1356 1357
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		while (!list_empty(&man->lru[i])) {
			spin_unlock(&glob->lru_lock);
R
Roger He 已提交
1358
			ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx);
1359
			if (ret)
1360
				return ret;
1361
			spin_lock(&glob->lru_lock);
1362
		}
1363
	}
1364
	spin_unlock(&glob->lru_lock);
1365 1366

	spin_lock(&man->move_lock);
1367
	fence = dma_fence_get(man->move);
1368 1369 1370
	spin_unlock(&man->move_lock);

	if (fence) {
1371 1372
		ret = dma_fence_wait(fence, false);
		dma_fence_put(fence);
1373 1374
		if (ret)
			return ret;
1375 1376
	}

1377 1378 1379 1380 1381
	return 0;
}

int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
{
R
Roel Kluin 已提交
1382
	struct ttm_mem_type_manager *man;
1383 1384 1385
	int ret = -EINVAL;

	if (mem_type >= TTM_NUM_MEM_TYPES) {
J
Joe Perches 已提交
1386
		pr_err("Illegal memory type %d\n", mem_type);
1387 1388
		return ret;
	}
R
Roel Kluin 已提交
1389
	man = &bdev->man[mem_type];
1390 1391

	if (!man->has_type) {
J
Joe Perches 已提交
1392 1393
		pr_err("Trying to take down uninitialized memory manager type %u\n",
		       mem_type);
1394 1395 1396 1397 1398 1399 1400 1401
		return ret;
	}

	man->use_type = false;
	man->has_type = false;

	ret = 0;
	if (mem_type > 0) {
1402 1403 1404 1405 1406
		ret = ttm_bo_force_list_clean(bdev, mem_type);
		if (ret) {
			pr_err("Cleanup eviction failed\n");
			return ret;
		}
1407

1408
		ret = (*man->func->takedown)(man);
1409 1410
	}

1411 1412 1413
	dma_fence_put(man->move);
	man->move = NULL;

1414 1415 1416 1417 1418 1419 1420 1421 1422
	return ret;
}
EXPORT_SYMBOL(ttm_bo_clean_mm);

int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
{
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];

	if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
J
Joe Perches 已提交
1423
		pr_err("Illegal memory manager memory type %u\n", mem_type);
1424 1425 1426 1427
		return -EINVAL;
	}

	if (!man->has_type) {
J
Joe Perches 已提交
1428
		pr_err("Memory type %u has not been initialized\n", mem_type);
1429 1430 1431
		return 0;
	}

1432
	return ttm_bo_force_list_clean(bdev, mem_type);
1433 1434 1435 1436
}
EXPORT_SYMBOL(ttm_bo_evict_mm);

int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1437
			unsigned long p_size)
1438
{
H
Huang Rui 已提交
1439
	int ret;
1440
	struct ttm_mem_type_manager *man;
1441
	unsigned i;
1442

1443
	BUG_ON(type >= TTM_NUM_MEM_TYPES);
1444
	man = &bdev->man[type];
1445
	BUG_ON(man->has_type);
1446 1447 1448
	man->io_reserve_fastpath = true;
	man->use_io_reserve_lru = false;
	mutex_init(&man->io_reserve_mutex);
1449
	spin_lock_init(&man->move_lock);
1450
	INIT_LIST_HEAD(&man->io_reserve_lru);
1451 1452 1453 1454

	ret = bdev->driver->init_mem_type(bdev, type, man);
	if (ret)
		return ret;
1455
	man->bdev = bdev;
1456 1457

	if (type != TTM_PL_SYSTEM) {
1458
		ret = (*man->func->init)(man, p_size);
1459 1460 1461 1462 1463 1464 1465
		if (ret)
			return ret;
	}
	man->has_type = true;
	man->use_type = true;
	man->size = p_size;

1466 1467
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&man->lru[i]);
1468
	man->move = NULL;
1469 1470 1471 1472 1473

	return 0;
}
EXPORT_SYMBOL(ttm_bo_init_mm);

1474 1475 1476 1477 1478 1479 1480 1481 1482
static void ttm_bo_global_kobj_release(struct kobject *kobj)
{
	struct ttm_bo_global *glob =
		container_of(kobj, struct ttm_bo_global, kobj);

	__free_page(glob->dummy_read_page);
	kfree(glob);
}

1483
void ttm_bo_global_release(struct drm_global_reference *ref)
1484 1485 1486 1487 1488 1489 1490 1491
{
	struct ttm_bo_global *glob = ref->object;

	kobject_del(&glob->kobj);
	kobject_put(&glob->kobj);
}
EXPORT_SYMBOL(ttm_bo_global_release);

1492
int ttm_bo_global_init(struct drm_global_reference *ref)
1493 1494 1495 1496 1497
{
	struct ttm_bo_global_ref *bo_ref =
		container_of(ref, struct ttm_bo_global_ref, ref);
	struct ttm_bo_global *glob = ref->object;
	int ret;
1498
	unsigned i;
1499 1500 1501 1502

	mutex_init(&glob->device_list_mutex);
	spin_lock_init(&glob->lru_lock);
	glob->mem_glob = bo_ref->mem_glob;
1503
	glob->mem_glob->bo_glob = glob;
1504 1505 1506 1507 1508 1509 1510
	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);

	if (unlikely(glob->dummy_read_page == NULL)) {
		ret = -ENOMEM;
		goto out_no_drp;
	}

1511 1512
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&glob->swap_lru[i]);
1513 1514 1515
	INIT_LIST_HEAD(&glob->device_list);
	atomic_set(&glob->bo_count, 0);

1516 1517
	ret = kobject_init_and_add(
		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1518 1519 1520 1521 1522 1523 1524 1525 1526 1527
	if (unlikely(ret != 0))
		kobject_put(&glob->kobj);
	return ret;
out_no_drp:
	kfree(glob);
	return ret;
}
EXPORT_SYMBOL(ttm_bo_global_init);


1528 1529 1530 1531 1532
int ttm_bo_device_release(struct ttm_bo_device *bdev)
{
	int ret = 0;
	unsigned i = TTM_NUM_MEM_TYPES;
	struct ttm_mem_type_manager *man;
1533
	struct ttm_bo_global *glob = bdev->glob;
1534 1535 1536 1537 1538 1539 1540

	while (i--) {
		man = &bdev->man[i];
		if (man->has_type) {
			man->use_type = false;
			if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
				ret = -EBUSY;
J
Joe Perches 已提交
1541 1542
				pr_err("DRM memory manager type %d is not clean\n",
				       i);
1543 1544 1545 1546 1547
			}
			man->has_type = false;
		}
	}

1548 1549 1550 1551
	mutex_lock(&glob->device_list_mutex);
	list_del(&bdev->device_list);
	mutex_unlock(&glob->device_list_mutex);

1552
	cancel_delayed_work_sync(&bdev->wq);
1553

1554
	if (ttm_bo_delayed_delete(bdev, true))
1555
		pr_debug("Delayed destroy list was clean\n");
1556

1557
	spin_lock(&glob->lru_lock);
1558 1559
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		if (list_empty(&bdev->man[0].lru[0]))
1560
			pr_debug("Swap list %d was clean\n", i);
1561
	spin_unlock(&glob->lru_lock);
1562

1563
	drm_vma_offset_manager_destroy(&bdev->vma_manager);
1564 1565 1566 1567 1568 1569

	return ret;
}
EXPORT_SYMBOL(ttm_bo_device_release);

int ttm_bo_device_init(struct ttm_bo_device *bdev,
1570 1571
		       struct ttm_bo_global *glob,
		       struct ttm_bo_driver *driver,
1572
		       struct address_space *mapping,
D
Dave Airlie 已提交
1573
		       uint64_t file_page_offset,
D
Dave Airlie 已提交
1574
		       bool need_dma32)
1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585
{
	int ret = -EINVAL;

	bdev->driver = driver;

	memset(bdev->man, 0, sizeof(bdev->man));

	/*
	 * Initialize the system memory buffer type.
	 * Other types need to be driver / IOCTL initialized.
	 */
1586
	ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1587
	if (unlikely(ret != 0))
1588
		goto out_no_sys;
1589

1590 1591
	drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
				    0x10000000);
1592 1593
	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
	INIT_LIST_HEAD(&bdev->ddestroy);
1594
	bdev->dev_mapping = mapping;
1595
	bdev->glob = glob;
D
Dave Airlie 已提交
1596
	bdev->need_dma32 = need_dma32;
1597 1598 1599
	mutex_lock(&glob->device_list_mutex);
	list_add_tail(&bdev->device_list, &glob->device_list);
	mutex_unlock(&glob->device_list_mutex);
1600 1601

	return 0;
1602
out_no_sys:
1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627
	return ret;
}
EXPORT_SYMBOL(ttm_bo_device_init);

/*
 * buffer object vm functions.
 */

bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
{
	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];

	if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
		if (mem->mem_type == TTM_PL_SYSTEM)
			return false;

		if (man->flags & TTM_MEMTYPE_FLAG_CMA)
			return false;

		if (mem->placement & TTM_PL_FLAG_CACHED)
			return false;
	}
	return true;
}

1628
void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1629 1630 1631
{
	struct ttm_bo_device *bdev = bo->bdev;

1632
	drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1633
	ttm_mem_io_free_vm(bo);
1634
}
1635 1636 1637 1638 1639 1640 1641 1642 1643

void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];

	ttm_mem_io_lock(man, false);
	ttm_bo_unmap_virtual_locked(bo);
	ttm_mem_io_unlock(man);
1644
}
1645 1646


1647
EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1648 1649

int ttm_bo_wait(struct ttm_buffer_object *bo,
1650
		bool interruptible, bool no_wait)
1651
{
C
Christian König 已提交
1652 1653 1654 1655 1656 1657 1658 1659
	long timeout = 15 * HZ;

	if (no_wait) {
		if (reservation_object_test_signaled_rcu(bo->resv, true))
			return 0;
		else
			return -EBUSY;
	}
1660

C
Christian König 已提交
1661 1662
	timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
						      interruptible, timeout);
1663 1664 1665 1666 1667 1668
	if (timeout < 0)
		return timeout;

	if (timeout == 0)
		return -EBUSY;

C
Christian König 已提交
1669
	reservation_object_add_excl_fence(bo->resv, NULL);
1670
	return 0;
1671 1672 1673 1674 1675 1676 1677 1678
}
EXPORT_SYMBOL(ttm_bo_wait);

int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
{
	int ret = 0;

	/*
1679
	 * Using ttm_bo_reserve makes sure the lru lists are updated.
1680 1681
	 */

1682
	ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1683 1684
	if (unlikely(ret != 0))
		return ret;
1685
	ret = ttm_bo_wait(bo, true, no_wait);
1686 1687 1688 1689 1690
	if (likely(ret == 0))
		atomic_inc(&bo->cpu_writers);
	ttm_bo_unreserve(bo);
	return ret;
}
1691
EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1692 1693 1694

void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
{
1695
	atomic_dec(&bo->cpu_writers);
1696
}
1697
EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1698 1699 1700 1701 1702

/**
 * A buffer object shrink method that tries to swap out the first
 * buffer object on the bo_global::swap_lru list.
 */
1703
int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1704 1705 1706
{
	struct ttm_buffer_object *bo;
	int ret = -EBUSY;
1707
	bool locked;
1708
	unsigned i;
1709

1710
	spin_lock(&glob->lru_lock);
1711 1712
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1713 1714
			if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) {
				ret = 0;
1715
				break;
1716
			}
1717
		}
1718 1719 1720
		if (!ret)
			break;
	}
1721

1722 1723 1724 1725
	if (ret) {
		spin_unlock(&glob->lru_lock);
		return ret;
	}
1726

1727
	kref_get(&bo->list_kref);
1728

1729
	if (!list_empty(&bo->ddestroy)) {
1730
		ret = ttm_bo_cleanup_refs(bo, false, false, true);
1731 1732
		kref_put(&bo->list_kref, ttm_bo_release_list);
		return ret;
1733 1734
	}

1735
	ttm_bo_del_from_lru(bo);
1736
	spin_unlock(&glob->lru_lock);
1737 1738

	/**
1739
	 * Move to system cached
1740 1741
	 */

1742 1743
	if (bo->mem.mem_type != TTM_PL_SYSTEM ||
	    bo->ttm->caching_state != tt_cached) {
1744
		struct ttm_operation_ctx ctx = { false, false };
1745 1746 1747 1748 1749 1750 1751
		struct ttm_mem_reg evict_mem;

		evict_mem = bo->mem;
		evict_mem.mm_node = NULL;
		evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
		evict_mem.mem_type = TTM_PL_SYSTEM;

1752
		ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1753 1754 1755 1756
		if (unlikely(ret != 0))
			goto out;
	}

1757 1758 1759 1760 1761 1762 1763 1764
	/**
	 * Make sure BO is idle.
	 */

	ret = ttm_bo_wait(bo, false, false);
	if (unlikely(ret != 0))
		goto out;

1765 1766 1767 1768 1769 1770 1771
	ttm_bo_unmap_virtual(bo);

	/**
	 * Swap out. Buffer will be swapped in again as soon as
	 * anyone tries to access a ttm page.
	 */

1772 1773 1774
	if (bo->bdev->driver->swap_notify)
		bo->bdev->driver->swap_notify(bo);

J
Jan Engelhardt 已提交
1775
	ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1776 1777 1778 1779 1780 1781 1782
out:

	/**
	 *
	 * Unreserve without putting on LRU to avoid swapping out an
	 * already swapped buffer.
	 */
1783 1784
	if (locked)
		reservation_object_unlock(bo->resv);
1785 1786 1787
	kref_put(&bo->list_kref, ttm_bo_release_list);
	return ret;
}
1788
EXPORT_SYMBOL(ttm_bo_swapout);
1789 1790 1791

void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
{
1792 1793 1794 1795 1796 1797
	struct ttm_operation_ctx ctx = {
		.interruptible = false,
		.no_wait_gpu = false
	};

	while (ttm_bo_swapout(bdev->glob, &ctx) == 0)
1798 1799
		;
}
1800
EXPORT_SYMBOL(ttm_bo_swapout_all);
1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823

/**
 * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
 * unreserved
 *
 * @bo: Pointer to buffer
 */
int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
{
	int ret;

	/*
	 * In the absense of a wait_unlocked API,
	 * Use the bo::wu_mutex to avoid triggering livelocks due to
	 * concurrent use of this function. Note that this use of
	 * bo::wu_mutex can go away if we change locking order to
	 * mmap_sem -> bo::reserve.
	 */
	ret = mutex_lock_interruptible(&bo->wu_mutex);
	if (unlikely(ret != 0))
		return -ERESTARTSYS;
	if (!ww_mutex_is_locked(&bo->resv->lock))
		goto out_unlock;
1824 1825 1826
	ret = reservation_object_lock_interruptible(bo->resv, NULL);
	if (ret == -EINTR)
		ret = -ERESTARTSYS;
1827 1828
	if (unlikely(ret != 0))
		goto out_unlock;
1829
	reservation_object_unlock(bo->resv);
1830 1831 1832 1833 1834

out_unlock:
	mutex_unlock(&bo->wu_mutex);
	return ret;
}