ttm_bo.c 42.5 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

#define TTM_ASSERT_LOCKED(param)
#define TTM_DEBUG(fmt, arg...)
#define TTM_BO_HASH_ORDER 13

static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50 51 52 53 54 55 56
static void ttm_bo_global_kobj_release(struct kobject *kobj);

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

57 58
static inline int ttm_mem_type_from_place(const struct ttm_place *place,
					  uint32_t *mem_type)
59 60 61 62
{
	int i;

	for (i = 0; i <= TTM_PL_PRIV5; i++)
63
		if (place->flags & (1 << i)) {
64 65 66 67 68 69
			*mem_type = i;
			return 0;
		}
	return -EINVAL;
}

70
static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
71
{
72 73
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];

J
Joe Perches 已提交
74 75 76
	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);
77
	pr_err("    gpu_offset: 0x%08llX\n", man->gpu_offset);
J
Joe Perches 已提交
78 79 80
	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);
81 82
	if (mem_type != TTM_PL_SYSTEM)
		(*man->func->debug)(man, TTM_PFX);
83 84 85 86 87 88 89
}

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
90 91 92
	pr_err("No space for %p (%lu pages, %luK, %luM)\n",
	       bo, bo->mem.num_pages, bo->mem.size >> 10,
	       bo->mem.size >> 20);
93
	for (i = 0; i < placement->num_placement; i++) {
94
		ret = ttm_mem_type_from_place(&placement->placement[i],
95 96 97
						&mem_type);
		if (ret)
			return;
J
Joe Perches 已提交
98
		pr_err("  placement[%d]=0x%08X (%d)\n",
99
		       i, placement->placement[i].flags, mem_type);
100
		ttm_mem_type_debug(bo->bdev, mem_type);
101 102 103
	}
}

104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
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);

	return snprintf(buffer, PAGE_SIZE, "%lu\n",
			(unsigned long) atomic_read(&glob->bo_count));
}

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

120
static const struct sysfs_ops ttm_bo_global_ops = {
121 122 123 124 125 126 127 128 129
	.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
};

130 131 132 133 134 135 136 137 138 139 140

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;
141
	size_t acc_size = bo->acc_size;
142 143 144 145 146 147 148 149 150 151

	BUG_ON(atomic_read(&bo->list_kref.refcount));
	BUG_ON(atomic_read(&bo->kref.refcount));
	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));

	if (bo->ttm)
		ttm_tt_destroy(bo->ttm);
152
	atomic_dec(&bo->glob->bo_count);
153 154
	if (bo->resv == &bo->ttm_resv)
		reservation_object_fini(&bo->ttm_resv);
155
	mutex_destroy(&bo->wu_mutex);
156 157 158 159 160
	if (bo->destroy)
		bo->destroy(bo);
	else {
		kfree(bo);
	}
161
	ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
162 163
}

164
void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
165 166 167 168
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_type_manager *man;

169
	lockdep_assert_held(&bo->resv->lock.base);
170 171 172 173 174 175 176 177 178

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

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

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

179
		if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
180
			list_add_tail(&bo->swap, &bo->glob->swap_lru);
181 182 183 184
			kref_get(&bo->list_kref);
		}
	}
}
185
EXPORT_SYMBOL(ttm_bo_add_to_lru);
186

187
int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
188
{
189
	struct ttm_bo_device *bdev = bo->bdev;
190 191
	int put_count = 0;

192 193 194
	if (bdev->driver->lru_removal)
		bdev->driver->lru_removal(bo);

195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211
	if (!list_empty(&bo->swap)) {
		list_del_init(&bo->swap);
		++put_count;
	}
	if (!list_empty(&bo->lru)) {
		list_del_init(&bo->lru);
		++put_count;
	}

	return put_count;
}

static void ttm_bo_ref_bug(struct kref *list_kref)
{
	BUG();
}

212 213 214
void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
			 bool never_free)
{
215 216
	kref_sub(&bo->list_kref, count,
		 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
217 218
}

219
void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
220
{
221
	int put_count;
222

223 224 225 226
	spin_lock(&bo->glob->lru_lock);
	put_count = ttm_bo_del_from_lru(bo);
	spin_unlock(&bo->glob->lru_lock);
	ttm_bo_list_ref_sub(bo, put_count, true);
227
}
228
EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
229

230 231 232 233 234 235 236
void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_type_manager *man;

	lockdep_assert_held(&bo->resv->lock.base);

237 238 239
	if (bdev->driver->lru_removal)
		bdev->driver->lru_removal(bo);

240 241 242 243 244 245 246 247 248 249 250 251 252 253
	if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
		list_del_init(&bo->swap);
		list_del_init(&bo->lru);

	} else {
		if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG))
			list_move_tail(&bo->swap, &bo->glob->swap_lru);

		man = &bdev->man[bo->mem.mem_type];
		list_move_tail(&bo->lru, &man->lru);
	}
}
EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);

254 255 256 257 258 259
/*
 * 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;
260
	struct ttm_bo_global *glob = bo->glob;
261 262 263 264 265 266
	int ret = 0;
	uint32_t page_flags = 0;

	TTM_ASSERT_LOCKED(&bo->mutex);
	bo->ttm = NULL;

D
Dave Airlie 已提交
267 268 269
	if (bdev->need_dma32)
		page_flags |= TTM_PAGE_FLAG_DMA32;

270 271 272 273 274
	switch (bo->type) {
	case ttm_bo_type_device:
		if (zero_alloc)
			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
	case ttm_bo_type_kernel:
275 276
		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
						      page_flags, glob->dummy_read_page);
277 278 279
		if (unlikely(bo->ttm == NULL))
			ret = -ENOMEM;
		break;
280 281 282 283 284 285 286 287 288 289
	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;
290
	default:
J
Joe Perches 已提交
291
		pr_err("Illegal buffer object type\n");
292 293 294 295 296 297 298 299 300
		ret = -EINVAL;
		break;
	}

	return ret;
}

static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
				  struct ttm_mem_reg *mem,
301
				  bool evict, bool interruptible,
302
				  bool no_wait_gpu)
303 304 305 306 307 308 309 310 311
{
	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 ||
312 313 314 315 316 317 318
	    ((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);
	}
319 320 321 322 323

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

324 325
	if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
		if (bo->ttm == NULL) {
326 327
			bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
			ret = ttm_bo_add_ttm(bo, zero);
328 329 330
			if (ret)
				goto out_err;
		}
331 332 333

		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
		if (ret)
334
			goto out_err;
335 336 337 338 339 340 341 342

		if (mem->mem_type != TTM_PL_SYSTEM) {
			ret = ttm_tt_bind(bo->ttm, mem);
			if (ret)
				goto out_err;
		}

		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
343 344
			if (bdev->driver->move_notify)
				bdev->driver->move_notify(bo, mem);
345
			bo->mem = *mem;
346 347 348 349 350
			mem->mm_node = NULL;
			goto moved;
		}
	}

351 352 353
	if (bdev->driver->move_notify)
		bdev->driver->move_notify(bo, mem);

354 355
	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
356
		ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
357 358
	else if (bdev->driver->move)
		ret = bdev->driver->move(bo, evict, interruptible,
359
					 no_wait_gpu, mem);
360
	else
361
		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
362

363 364 365 366 367 368 369
	if (ret) {
		if (bdev->driver->move_notify) {
			struct ttm_mem_reg tmp_mem = *mem;
			*mem = bo->mem;
			bo->mem = tmp_mem;
			bdev->driver->move_notify(bo, mem);
			bo->mem = *mem;
370
			*mem = tmp_mem;
371
		}
372

373 374
		goto out_err;
	}
375

376 377
moved:
	if (bo->evicted) {
378 379 380 381 382
		if (bdev->driver->invalidate_caches) {
			ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
			if (ret)
				pr_err("Can not flush read caches\n");
		}
383 384 385 386
		bo->evicted = false;
	}

	if (bo->mem.mm_node) {
387
		bo->offset = (bo->mem.start << PAGE_SHIFT) +
388 389
		    bdev->man[bo->mem.mem_type].gpu_offset;
		bo->cur_placement = bo->mem.placement;
390 391
	} else
		bo->offset = 0;
392 393 394 395 396 397 398 399 400 401 402 403 404 405

	return 0;

out_err:
	new_man = &bdev->man[bo->mem.mem_type];
	if ((new_man->flags & TTM_MEMTYPE_FLAG_FIXED) && bo->ttm) {
		ttm_tt_unbind(bo->ttm);
		ttm_tt_destroy(bo->ttm);
		bo->ttm = NULL;
	}

	return ret;
}

406
/**
407
 * Call bo::reserved.
408
 * Will release GPU memory type usage on destruction.
409 410 411
 * This is the place to put in driver specific hooks to release
 * driver private resources.
 * Will release the bo::reserved lock.
412 413 414 415
 */

static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
{
416 417 418
	if (bo->bdev->driver->move_notify)
		bo->bdev->driver->move_notify(bo, NULL);

419 420 421 422 423
	if (bo->ttm) {
		ttm_tt_unbind(bo->ttm);
		ttm_tt_destroy(bo->ttm);
		bo->ttm = NULL;
	}
424
	ttm_bo_mem_put(bo, &bo->mem);
425

426
	ww_mutex_unlock (&bo->resv->lock);
427 428
}

429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448
static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
{
	struct reservation_object_list *fobj;
	struct fence *fence;
	int i;

	fobj = reservation_object_get_list(bo->resv);
	fence = reservation_object_get_excl(bo->resv);
	if (fence && !fence->ops->signaled)
		fence_enable_sw_signaling(fence);

	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)
			fence_enable_sw_signaling(fence);
	}
}

449
static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
450 451
{
	struct ttm_bo_device *bdev = bo->bdev;
452
	struct ttm_bo_global *glob = bo->glob;
453
	int put_count;
454 455
	int ret;

456
	spin_lock(&glob->lru_lock);
457
	ret = __ttm_bo_reserve(bo, false, true, NULL);
458

M
Maarten Lankhorst 已提交
459
	if (!ret) {
460
		if (!ttm_bo_wait(bo, false, true)) {
M
Maarten Lankhorst 已提交
461
			put_count = ttm_bo_del_from_lru(bo);
462

M
Maarten Lankhorst 已提交
463 464
			spin_unlock(&glob->lru_lock);
			ttm_bo_cleanup_memtype_use(bo);
465

M
Maarten Lankhorst 已提交
466
			ttm_bo_list_ref_sub(bo, put_count, true);
467

M
Maarten Lankhorst 已提交
468
			return;
469 470
		} else
			ttm_bo_flush_all_fences(bo);
471 472 473 474 475 476 477 478 479 480 481

		/*
		 * 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);
		}

482
		__ttm_bo_unreserve(bo);
483
	}
484 485 486 487 488 489 490 491 492 493

	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);
}

/**
494
 * function ttm_bo_cleanup_refs_and_unlock
495 496 497
 * If bo idle, remove from delayed- and lru lists, and unref.
 * If not idle, do nothing.
 *
498 499 500
 * Must be called with lru_lock and reservation held, this function
 * will drop both before returning.
 *
501 502 503 504
 * @interruptible         Any sleeps should occur interruptibly.
 * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
 */

505 506 507
static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
					  bool interruptible,
					  bool no_wait_gpu)
508 509 510
{
	struct ttm_bo_global *glob = bo->glob;
	int put_count;
511
	int ret;
512

513
	ret = ttm_bo_wait(bo, false, true);
514

515
	if (ret && !no_wait_gpu) {
M
Maarten Lankhorst 已提交
516 517 518 519 520 521 522 523 524 525 526 527 528
		long lret;
		ww_mutex_unlock(&bo->resv->lock);
		spin_unlock(&glob->lru_lock);

		lret = reservation_object_wait_timeout_rcu(bo->resv,
							   true,
							   interruptible,
							   30 * HZ);

		if (lret < 0)
			return lret;
		else if (lret == 0)
			return -EBUSY;
529

530
		spin_lock(&glob->lru_lock);
531
		ret = __ttm_bo_reserve(bo, false, true, NULL);
532

533 534 535 536 537 538 539 540 541 542 543 544
		/*
		 * 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.
		 */
		if (ret) {
			spin_unlock(&glob->lru_lock);
			return 0;
		}
545 546 547 548 549

		/*
		 * remove sync_obj with ttm_bo_wait, the wait should be
		 * finished, and no new wait object should have been added.
		 */
550
		ret = ttm_bo_wait(bo, false, true);
551 552
		WARN_ON(ret);
	}
553

554
	if (ret || unlikely(list_empty(&bo->ddestroy))) {
555
		__ttm_bo_unreserve(bo);
556
		spin_unlock(&glob->lru_lock);
557
		return ret;
558 559
	}

560 561 562 563 564 565 566
	put_count = ttm_bo_del_from_lru(bo);
	list_del_init(&bo->ddestroy);
	++put_count;

	spin_unlock(&glob->lru_lock);
	ttm_bo_cleanup_memtype_use(bo);

567
	ttm_bo_list_ref_sub(bo, put_count, true);
568 569

	return 0;
570 571 572 573 574 575 576 577 578
}

/**
 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 * encountered buffers.
 */

static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
{
579
	struct ttm_bo_global *glob = bdev->glob;
580 581
	struct ttm_buffer_object *entry = NULL;
	int ret = 0;
582

583
	spin_lock(&glob->lru_lock);
584 585 586 587 588 589 590 591 592 593 594 595 596
	if (list_empty(&bdev->ddestroy))
		goto out_unlock;

	entry = list_first_entry(&bdev->ddestroy,
		struct ttm_buffer_object, ddestroy);
	kref_get(&entry->list_kref);

	for (;;) {
		struct ttm_buffer_object *nentry = NULL;

		if (entry->ddestroy.next != &bdev->ddestroy) {
			nentry = list_first_entry(&entry->ddestroy,
				struct ttm_buffer_object, ddestroy);
597 598 599
			kref_get(&nentry->list_kref);
		}

600
		ret = __ttm_bo_reserve(entry, false, true, NULL);
601 602
		if (remove_all && ret) {
			spin_unlock(&glob->lru_lock);
603
			ret = __ttm_bo_reserve(entry, false, false, NULL);
604 605 606
			spin_lock(&glob->lru_lock);
		}

607 608 609 610 611 612
		if (!ret)
			ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
							     !remove_all);
		else
			spin_unlock(&glob->lru_lock);

613
		kref_put(&entry->list_kref, ttm_bo_release_list);
614 615 616 617
		entry = nentry;

		if (ret || !entry)
			goto out;
618

619
		spin_lock(&glob->lru_lock);
620
		if (list_empty(&entry->ddestroy))
621 622 623
			break;
	}

624 625 626 627 628
out_unlock:
	spin_unlock(&glob->lru_lock);
out:
	if (entry)
		kref_put(&entry->list_kref, ttm_bo_release_list);
629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
	return ret;
}

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

	if (ttm_bo_delayed_delete(bdev, false)) {
		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;
648
	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
649

650
	drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
651 652 653
	ttm_mem_io_lock(man, false);
	ttm_mem_io_free_vm(bo);
	ttm_mem_io_unlock(man);
654
	ttm_bo_cleanup_refs_or_queue(bo);
655 656 657 658 659 660 661 662 663 664 665 666
	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);

667 668 669 670 671 672 673 674 675 676 677 678 679 680
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);

681
static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
682
			bool no_wait_gpu)
683 684 685
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_reg evict_mem;
686 687
	struct ttm_placement placement;
	int ret = 0;
688

689
	ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
690

691
	if (unlikely(ret != 0)) {
692
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
693
			pr_err("Failed to expire sync object before buffer eviction\n");
694
		}
695 696 697
		goto out;
	}

698
	lockdep_assert_held(&bo->resv->lock.base);
699 700 701

	evict_mem = bo->mem;
	evict_mem.mm_node = NULL;
702 703
	evict_mem.bus.io_reserved_vm = false;
	evict_mem.bus.io_reserved_count = 0;
704

705 706
	placement.num_placement = 0;
	placement.num_busy_placement = 0;
707 708
	bdev->driver->evict_flags(bo, &placement);
	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
709
				no_wait_gpu);
710
	if (ret) {
711
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
712 713
			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
			       bo);
714 715
			ttm_bo_mem_space_debug(bo, &placement);
		}
716 717 718 719
		goto out;
	}

	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
720
				     no_wait_gpu);
721
	if (ret) {
722
		if (ret != -ERESTARTSYS)
J
Joe Perches 已提交
723
			pr_err("Buffer eviction failed\n");
724
		ttm_bo_mem_put(bo, &evict_mem);
725 726
		goto out;
	}
727 728 729 730 731 732 733
	bo->evicted = true;
out:
	return ret;
}

static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
				uint32_t mem_type,
734
				const struct ttm_place *place,
735
				bool interruptible,
736
				bool no_wait_gpu)
737 738 739 740
{
	struct ttm_bo_global *glob = bdev->glob;
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	struct ttm_buffer_object *bo;
741
	int ret = -EBUSY, put_count;
742

743
	spin_lock(&glob->lru_lock);
744
	list_for_each_entry(bo, &man->lru, lru) {
745
		ret = __ttm_bo_reserve(bo, false, true, NULL);
746 747 748 749 750 751 752 753 754 755 756 757 758
		if (!ret) {
			if (place && (place->fpfn || place->lpfn)) {
				/* 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)) {
					__ttm_bo_unreserve(bo);
					ret = -EBUSY;
					continue;
				}
			}

759
			break;
760
		}
761 762 763
	}

	if (ret) {
764
		spin_unlock(&glob->lru_lock);
765
		return ret;
766 767
	}

768
	kref_get(&bo->list_kref);
769

770
	if (!list_empty(&bo->ddestroy)) {
771 772
		ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
						     no_wait_gpu);
773
		kref_put(&bo->list_kref, ttm_bo_release_list);
774
		return ret;
775 776
	}

777
	put_count = ttm_bo_del_from_lru(bo);
778
	spin_unlock(&glob->lru_lock);
779 780 781

	BUG_ON(ret != 0);

782
	ttm_bo_list_ref_sub(bo, put_count, true);
783

784
	ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
785
	ttm_bo_unreserve(bo);
786

787
	kref_put(&bo->list_kref, ttm_bo_release_list);
788 789 790
	return ret;
}

791 792
void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
{
793
	struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
794

795 796
	if (mem->mm_node)
		(*man->func->put_node)(man, mem);
797 798 799
}
EXPORT_SYMBOL(ttm_bo_mem_put);

800 801 802 803
/**
 * 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.
 */
804 805
static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
					uint32_t mem_type,
806
					const struct ttm_place *place,
807
					struct ttm_mem_reg *mem,
808 809
					bool interruptible,
					bool no_wait_gpu)
810
{
811
	struct ttm_bo_device *bdev = bo->bdev;
812 813 814 815
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	int ret;

	do {
816
		ret = (*man->func->get_node)(man, bo, place, mem);
817 818
		if (unlikely(ret != 0))
			return ret;
819
		if (mem->mm_node)
820
			break;
821
		ret = ttm_mem_evict_first(bdev, mem_type, place,
822
					  interruptible, no_wait_gpu);
823 824 825
		if (unlikely(ret != 0))
			return ret;
	} while (1);
826
	if (mem->mm_node == NULL)
827 828 829 830 831
		return -ENOMEM;
	mem->mem_type = mem_type;
	return 0;
}

832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856
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;
}

857 858
static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
				 uint32_t mem_type,
859
				 const struct ttm_place *place,
860
				 uint32_t *masked_placement)
861 862 863
{
	uint32_t cur_flags = ttm_bo_type_flags(mem_type);

864
	if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
865 866
		return false;

867
	if ((place->flags & man->available_caching) == 0)
868 869
		return false;

870
	cur_flags |= (place->flags & man->available_caching);
871 872

	*masked_placement = cur_flags;
873 874 875 876 877 878 879 880 881 882 883 884
	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,
885 886
			struct ttm_placement *placement,
			struct ttm_mem_reg *mem,
887
			bool interruptible,
888
			bool no_wait_gpu)
889 890 891 892 893 894 895
{
	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;
896
	bool has_erestartsys = false;
897
	int i, ret;
898 899

	mem->mm_node = NULL;
900
	for (i = 0; i < placement->num_placement; ++i) {
901 902 903
		const struct ttm_place *place = &placement->placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
904 905
		if (ret)
			return ret;
906
		man = &bdev->man[mem_type];
907 908
		if (!man->has_type || !man->use_type)
			continue;
909

910
		type_ok = ttm_bo_mt_compatible(man, mem_type, place,
911
						&cur_flags);
912 913 914 915

		if (!type_ok)
			continue;

916
		type_found = true;
917 918
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
919 920 921 922
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
923
		ttm_flag_masked(&cur_flags, place->flags,
924
				~TTM_PL_MASK_MEMTYPE);
925

926 927 928
		if (mem_type == TTM_PL_SYSTEM)
			break;

929 930 931 932
		ret = (*man->func->get_node)(man, bo, place, mem);
		if (unlikely(ret))
			return ret;
		
933
		if (mem->mm_node)
934 935 936
			break;
	}

937
	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
938 939 940 941 942
		mem->mem_type = mem_type;
		mem->placement = cur_flags;
		return 0;
	}

943
	for (i = 0; i < placement->num_busy_placement; ++i) {
944 945 946
		const struct ttm_place *place = &placement->busy_placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
947 948
		if (ret)
			return ret;
949
		man = &bdev->man[mem_type];
950
		if (!man->has_type || !man->use_type)
951
			continue;
952
		if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
953 954
			continue;

955
		type_found = true;
956 957
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
958 959 960 961
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
962
		ttm_flag_masked(&cur_flags, place->flags,
963
				~TTM_PL_MASK_MEMTYPE);
964

965 966 967 968 969 970 971
		if (mem_type == TTM_PL_SYSTEM) {
			mem->mem_type = mem_type;
			mem->placement = cur_flags;
			mem->mm_node = NULL;
			return 0;
		}

972
		ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
973
						interruptible, no_wait_gpu);
974 975 976 977
		if (ret == 0 && mem->mm_node) {
			mem->placement = cur_flags;
			return 0;
		}
978 979
		if (ret == -ERESTARTSYS)
			has_erestartsys = true;
980
	}
981 982 983 984 985 986 987

	if (!type_found) {
		printk(KERN_ERR TTM_PFX "No compatible memory type found.\n");
		return -EINVAL;
	}

	return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
988 989 990
}
EXPORT_SYMBOL(ttm_bo_mem_space);

991
static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
992
			struct ttm_placement *placement,
993
			bool interruptible,
994
			bool no_wait_gpu)
995 996 997 998
{
	int ret = 0;
	struct ttm_mem_reg mem;

999
	lockdep_assert_held(&bo->resv->lock.base);
1000 1001

	/*
1002 1003
	 * Don't wait for the BO on initial allocation. This is important when
	 * the BO has an imported reservation object.
1004
	 */
1005 1006 1007 1008 1009 1010
	if (bo->mem.mem_type != TTM_PL_SYSTEM || bo->ttm != NULL) {
		/*
		 * FIXME: It's possible to pipeline buffer moves.
		 * Have the driver move function wait for idle when necessary,
		 * instead of doing it here.
		 */
1011
		ret = ttm_bo_wait(bo, interruptible, no_wait_gpu);
1012 1013 1014
		if (ret)
			return ret;
	}
1015 1016 1017
	mem.num_pages = bo->num_pages;
	mem.size = mem.num_pages << PAGE_SHIFT;
	mem.page_alignment = bo->mem.page_alignment;
1018 1019
	mem.bus.io_reserved_vm = false;
	mem.bus.io_reserved_count = 0;
1020 1021 1022
	/*
	 * Determine where to move the buffer.
	 */
1023 1024
	ret = ttm_bo_mem_space(bo, placement, &mem,
			       interruptible, no_wait_gpu);
1025 1026
	if (ret)
		goto out_unlock;
1027 1028
	ret = ttm_bo_handle_move_mem(bo, &mem, false,
				     interruptible, no_wait_gpu);
1029
out_unlock:
1030 1031
	if (ret && mem.mm_node)
		ttm_bo_mem_put(bo, &mem);
1032 1033 1034
	return ret;
}

1035 1036 1037
static bool ttm_bo_mem_compat(struct ttm_placement *placement,
			      struct ttm_mem_reg *mem,
			      uint32_t *new_flags)
1038
{
1039
	int i;
1040

1041
	for (i = 0; i < placement->num_placement; i++) {
1042
		const struct ttm_place *heap = &placement->placement[i];
1043
		if (mem->mm_node &&
1044
		    (mem->start < heap->fpfn ||
1045
		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1046 1047 1048
			continue;

		*new_flags = heap->flags;
1049 1050 1051 1052 1053 1054
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM))
			return true;
	}

	for (i = 0; i < placement->num_busy_placement; i++) {
1055
		const struct ttm_place *heap = &placement->busy_placement[i];
1056
		if (mem->mm_node &&
1057
		    (mem->start < heap->fpfn ||
1058
		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1059 1060 1061
			continue;

		*new_flags = heap->flags;
1062 1063 1064
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM))
			return true;
1065
	}
1066 1067

	return false;
1068 1069
}

1070 1071
int ttm_bo_validate(struct ttm_buffer_object *bo,
			struct ttm_placement *placement,
1072
			bool interruptible,
1073
			bool no_wait_gpu)
1074 1075
{
	int ret;
1076
	uint32_t new_flags;
1077

1078
	lockdep_assert_held(&bo->resv->lock.base);
1079 1080 1081
	/*
	 * Check whether we need to move buffer.
	 */
1082
	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1083 1084
		ret = ttm_bo_move_buffer(bo, placement, interruptible,
					 no_wait_gpu);
1085
		if (ret)
1086
			return ret;
1087 1088 1089 1090 1091
	} else {
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the compatible memory placement flags to the active flags
		 */
1092
		ttm_flag_masked(&bo->mem.placement, new_flags,
1093
				~TTM_PL_MASK_MEMTYPE);
1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104
	}
	/*
	 * 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;
}
1105
EXPORT_SYMBOL(ttm_bo_validate);
1106

1107 1108 1109 1110 1111 1112 1113
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,
J
Jan Engelhardt 已提交
1114
		struct file *persistent_swap_storage,
1115
		size_t acc_size,
1116
		struct sg_table *sg,
1117
		struct reservation_object *resv,
1118
		void (*destroy) (struct ttm_buffer_object *))
1119
{
1120
	int ret = 0;
1121
	unsigned long num_pages;
1122
	struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1123
	bool locked;
1124 1125 1126

	ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
	if (ret) {
J
Joe Perches 已提交
1127
		pr_err("Out of kernel memory\n");
1128 1129 1130 1131 1132 1133
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
		return -ENOMEM;
	}
1134 1135 1136

	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
	if (num_pages == 0) {
J
Joe Perches 已提交
1137
		pr_err("Illegal buffer object size\n");
1138 1139 1140 1141
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
1142
		ttm_mem_global_free(mem_glob, acc_size);
1143 1144 1145 1146 1147 1148 1149 1150 1151 1152
		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);
1153
	INIT_LIST_HEAD(&bo->io_reserve_lru);
1154
	mutex_init(&bo->wu_mutex);
1155
	bo->bdev = bdev;
1156
	bo->glob = bdev->glob;
1157 1158
	bo->type = type;
	bo->num_pages = num_pages;
1159
	bo->mem.size = num_pages << PAGE_SHIFT;
1160 1161 1162 1163
	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;
1164 1165
	bo->mem.bus.io_reserved_vm = false;
	bo->mem.bus.io_reserved_count = 0;
1166 1167
	bo->priv_flags = 0;
	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
J
Jan Engelhardt 已提交
1168
	bo->persistent_swap_storage = persistent_swap_storage;
1169
	bo->acc_size = acc_size;
1170
	bo->sg = sg;
1171 1172 1173 1174 1175 1176 1177
	if (resv) {
		bo->resv = resv;
		lockdep_assert_held(&bo->resv->lock.base);
	} else {
		bo->resv = &bo->ttm_resv;
		reservation_object_init(&bo->ttm_resv);
	}
1178
	atomic_inc(&bo->glob->bo_count);
1179
	drm_vma_node_reset(&bo->vma_node);
1180 1181 1182 1183 1184

	/*
	 * For ttm_bo_type_device buffers, allocate
	 * address space from the device.
	 */
1185 1186
	if (bo->type == ttm_bo_type_device ||
	    bo->type == ttm_bo_type_sg)
1187 1188
		ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
					 bo->mem.num_pages);
1189

1190 1191 1192 1193 1194 1195 1196
	/* passed reservation objects should already be locked,
	 * since otherwise lockdep will be angered in radeon.
	 */
	if (!resv) {
		locked = ww_mutex_trylock(&bo->resv->lock);
		WARN_ON(!locked);
	}
1197

1198 1199
	if (likely(!ret))
		ret = ttm_bo_validate(bo, placement, interruptible, false);
1200

1201
	if (!resv) {
1202
		ttm_bo_unreserve(bo);
1203

1204 1205 1206 1207 1208 1209
	} else if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
		spin_lock(&bo->glob->lru_lock);
		ttm_bo_add_to_lru(bo);
		spin_unlock(&bo->glob->lru_lock);
	}

1210 1211
	if (unlikely(ret))
		ttm_bo_unref(&bo);
1212 1213 1214

	return ret;
}
1215
EXPORT_SYMBOL(ttm_bo_init);
1216

1217 1218 1219
size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
		       unsigned long bo_size,
		       unsigned struct_size)
1220
{
1221 1222
	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
	size_t size = 0;
1223

1224
	size += ttm_round_pot(struct_size);
F
Felix Kuehling 已提交
1225
	size += ttm_round_pot(npages * sizeof(void *));
1226 1227
	size += ttm_round_pot(sizeof(struct ttm_tt));
	return size;
1228
}
1229 1230 1231 1232 1233 1234 1235 1236 1237 1238
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
1239
	size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1240 1241 1242 1243
	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
	return size;
}
EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1244

1245 1246 1247 1248 1249 1250
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
1251
			struct file *persistent_swap_storage,
1252
			struct ttm_buffer_object **p_bo)
1253 1254
{
	struct ttm_buffer_object *bo;
1255
	size_t acc_size;
1256
	int ret;
1257 1258

	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1259
	if (unlikely(bo == NULL))
1260 1261
		return -ENOMEM;

1262
	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1263
	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1264
			  interruptible, persistent_swap_storage, acc_size,
1265
			  NULL, NULL, NULL);
1266 1267 1268 1269 1270
	if (likely(ret == 0))
		*p_bo = bo;

	return ret;
}
T
Thomas Hellstrom 已提交
1271
EXPORT_SYMBOL(ttm_bo_create);
1272 1273

static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1274
					unsigned mem_type, bool allow_errors)
1275
{
1276
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1277
	struct ttm_bo_global *glob = bdev->glob;
1278 1279 1280 1281 1282 1283
	int ret;

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

1284
	spin_lock(&glob->lru_lock);
1285
	while (!list_empty(&man->lru)) {
1286
		spin_unlock(&glob->lru_lock);
1287
		ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false);
1288 1289 1290 1291
		if (ret) {
			if (allow_errors) {
				return ret;
			} else {
J
Joe Perches 已提交
1292
				pr_err("Cleanup eviction failed\n");
1293 1294
			}
		}
1295
		spin_lock(&glob->lru_lock);
1296
	}
1297
	spin_unlock(&glob->lru_lock);
1298 1299 1300 1301 1302
	return 0;
}

int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
{
R
Roel Kluin 已提交
1303
	struct ttm_mem_type_manager *man;
1304 1305 1306
	int ret = -EINVAL;

	if (mem_type >= TTM_NUM_MEM_TYPES) {
J
Joe Perches 已提交
1307
		pr_err("Illegal memory type %d\n", mem_type);
1308 1309
		return ret;
	}
R
Roel Kluin 已提交
1310
	man = &bdev->man[mem_type];
1311 1312

	if (!man->has_type) {
J
Joe Perches 已提交
1313 1314
		pr_err("Trying to take down uninitialized memory manager type %u\n",
		       mem_type);
1315 1316 1317 1318 1319 1320 1321 1322
		return ret;
	}

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

	ret = 0;
	if (mem_type > 0) {
1323
		ttm_bo_force_list_clean(bdev, mem_type, false);
1324

1325
		ret = (*man->func->takedown)(man);
1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336
	}

	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
1337
		pr_err("Illegal memory manager memory type %u\n", mem_type);
1338 1339 1340 1341
		return -EINVAL;
	}

	if (!man->has_type) {
J
Joe Perches 已提交
1342
		pr_err("Memory type %u has not been initialized\n", mem_type);
1343 1344 1345
		return 0;
	}

1346
	return ttm_bo_force_list_clean(bdev, mem_type, true);
1347 1348 1349 1350
}
EXPORT_SYMBOL(ttm_bo_evict_mm);

int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1351
			unsigned long p_size)
1352 1353 1354 1355
{
	int ret = -EINVAL;
	struct ttm_mem_type_manager *man;

1356
	BUG_ON(type >= TTM_NUM_MEM_TYPES);
1357
	man = &bdev->man[type];
1358
	BUG_ON(man->has_type);
1359 1360 1361 1362
	man->io_reserve_fastpath = true;
	man->use_io_reserve_lru = false;
	mutex_init(&man->io_reserve_mutex);
	INIT_LIST_HEAD(&man->io_reserve_lru);
1363 1364 1365 1366

	ret = bdev->driver->init_mem_type(bdev, type, man);
	if (ret)
		return ret;
1367
	man->bdev = bdev;
1368 1369 1370

	ret = 0;
	if (type != TTM_PL_SYSTEM) {
1371
		ret = (*man->func->init)(man, p_size);
1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384
		if (ret)
			return ret;
	}
	man->has_type = true;
	man->use_type = true;
	man->size = p_size;

	INIT_LIST_HEAD(&man->lru);

	return 0;
}
EXPORT_SYMBOL(ttm_bo_init_mm);

1385 1386 1387 1388 1389 1390 1391 1392 1393 1394
static void ttm_bo_global_kobj_release(struct kobject *kobj)
{
	struct ttm_bo_global *glob =
		container_of(kobj, struct ttm_bo_global, kobj);

	ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
	__free_page(glob->dummy_read_page);
	kfree(glob);
}

1395
void ttm_bo_global_release(struct drm_global_reference *ref)
1396 1397 1398 1399 1400 1401 1402 1403
{
	struct ttm_bo_global *glob = ref->object;

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

1404
int ttm_bo_global_init(struct drm_global_reference *ref)
1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426
{
	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;

	mutex_init(&glob->device_list_mutex);
	spin_lock_init(&glob->lru_lock);
	glob->mem_glob = bo_ref->mem_glob;
	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);

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

	INIT_LIST_HEAD(&glob->swap_lru);
	INIT_LIST_HEAD(&glob->device_list);

	ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
	ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
	if (unlikely(ret != 0)) {
J
Joe Perches 已提交
1427
		pr_err("Could not register buffer object swapout\n");
1428 1429 1430 1431 1432
		goto out_no_shrink;
	}

	atomic_set(&glob->bo_count, 0);

1433 1434
	ret = kobject_init_and_add(
		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446
	if (unlikely(ret != 0))
		kobject_put(&glob->kobj);
	return ret;
out_no_shrink:
	__free_page(glob->dummy_read_page);
out_no_drp:
	kfree(glob);
	return ret;
}
EXPORT_SYMBOL(ttm_bo_global_init);


1447 1448 1449 1450 1451
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;
1452
	struct ttm_bo_global *glob = bdev->glob;
1453 1454 1455 1456 1457 1458 1459

	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
1460 1461
				pr_err("DRM memory manager type %d is not clean\n",
				       i);
1462 1463 1464 1465 1466
			}
			man->has_type = false;
		}
	}

1467 1468 1469 1470
	mutex_lock(&glob->device_list_mutex);
	list_del(&bdev->device_list);
	mutex_unlock(&glob->device_list_mutex);

1471
	cancel_delayed_work_sync(&bdev->wq);
1472 1473 1474 1475

	while (ttm_bo_delayed_delete(bdev, true))
		;

1476
	spin_lock(&glob->lru_lock);
1477 1478 1479 1480 1481
	if (list_empty(&bdev->ddestroy))
		TTM_DEBUG("Delayed destroy list was clean\n");

	if (list_empty(&bdev->man[0].lru))
		TTM_DEBUG("Swap list was clean\n");
1482
	spin_unlock(&glob->lru_lock);
1483

1484
	drm_vma_offset_manager_destroy(&bdev->vma_manager);
1485 1486 1487 1488 1489 1490

	return ret;
}
EXPORT_SYMBOL(ttm_bo_device_release);

int ttm_bo_device_init(struct ttm_bo_device *bdev,
1491 1492
		       struct ttm_bo_global *glob,
		       struct ttm_bo_driver *driver,
1493
		       struct address_space *mapping,
D
Dave Airlie 已提交
1494
		       uint64_t file_page_offset,
D
Dave Airlie 已提交
1495
		       bool need_dma32)
1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506
{
	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.
	 */
1507
	ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1508
	if (unlikely(ret != 0))
1509
		goto out_no_sys;
1510

1511 1512
	drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
				    0x10000000);
1513 1514
	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
	INIT_LIST_HEAD(&bdev->ddestroy);
1515
	bdev->dev_mapping = mapping;
1516
	bdev->glob = glob;
D
Dave Airlie 已提交
1517
	bdev->need_dma32 = need_dma32;
1518 1519 1520
	mutex_lock(&glob->device_list_mutex);
	list_add_tail(&bdev->device_list, &glob->device_list);
	mutex_unlock(&glob->device_list_mutex);
1521 1522

	return 0;
1523
out_no_sys:
1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548
	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;
}

1549
void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1550 1551 1552
{
	struct ttm_bo_device *bdev = bo->bdev;

1553
	drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1554
	ttm_mem_io_free_vm(bo);
1555
}
1556 1557 1558 1559 1560 1561 1562 1563 1564

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);
1565
}
1566 1567


1568
EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1569 1570

int ttm_bo_wait(struct ttm_buffer_object *bo,
1571
		bool interruptible, bool no_wait)
1572
{
1573 1574 1575 1576 1577
	struct reservation_object_list *fobj;
	struct reservation_object *resv;
	struct fence *excl;
	long timeout = 15 * HZ;
	int i;
1578

1579 1580 1581 1582 1583 1584 1585
	resv = bo->resv;
	fobj = reservation_object_get_list(resv);
	excl = reservation_object_get_excl(resv);
	if (excl) {
		if (!fence_is_signaled(excl)) {
			if (no_wait)
				return -EBUSY;
1586

1587 1588
			timeout = fence_wait_timeout(excl,
						     interruptible, timeout);
1589
		}
1590
	}
1591

1592 1593 1594 1595
	for (i = 0; fobj && timeout > 0 && i < fobj->shared_count; ++i) {
		struct fence *fence;
		fence = rcu_dereference_protected(fobj->shared[i],
						reservation_object_held(resv));
1596

1597 1598 1599
		if (!fence_is_signaled(fence)) {
			if (no_wait)
				return -EBUSY;
M
Maarten Lankhorst 已提交
1600

1601 1602
			timeout = fence_wait_timeout(fence,
						     interruptible, timeout);
1603 1604
		}
	}
1605 1606 1607 1608 1609 1610 1611 1612 1613 1614

	if (timeout < 0)
		return timeout;

	if (timeout == 0)
		return -EBUSY;

	reservation_object_add_excl_fence(resv, NULL);
	clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
	return 0;
1615 1616 1617 1618 1619 1620 1621 1622
}
EXPORT_SYMBOL(ttm_bo_wait);

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

	/*
1623
	 * Using ttm_bo_reserve makes sure the lru lists are updated.
1624 1625
	 */

1626
	ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1627 1628
	if (unlikely(ret != 0))
		return ret;
1629
	ret = ttm_bo_wait(bo, true, no_wait);
1630 1631 1632 1633 1634
	if (likely(ret == 0))
		atomic_inc(&bo->cpu_writers);
	ttm_bo_unreserve(bo);
	return ret;
}
1635
EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1636 1637 1638

void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
{
1639
	atomic_dec(&bo->cpu_writers);
1640
}
1641
EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1642 1643 1644 1645 1646 1647 1648 1649

/**
 * A buffer object shrink method that tries to swap out the first
 * buffer object on the bo_global::swap_lru list.
 */

static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
{
1650 1651
	struct ttm_bo_global *glob =
	    container_of(shrink, struct ttm_bo_global, shrink);
1652 1653 1654 1655 1656
	struct ttm_buffer_object *bo;
	int ret = -EBUSY;
	int put_count;
	uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);

1657
	spin_lock(&glob->lru_lock);
1658
	list_for_each_entry(bo, &glob->swap_lru, swap) {
1659
		ret = __ttm_bo_reserve(bo, false, true, NULL);
1660 1661 1662
		if (!ret)
			break;
	}
1663

1664 1665 1666 1667
	if (ret) {
		spin_unlock(&glob->lru_lock);
		return ret;
	}
1668

1669
	kref_get(&bo->list_kref);
1670

1671 1672 1673 1674
	if (!list_empty(&bo->ddestroy)) {
		ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
		kref_put(&bo->list_kref, ttm_bo_release_list);
		return ret;
1675 1676 1677
	}

	put_count = ttm_bo_del_from_lru(bo);
1678
	spin_unlock(&glob->lru_lock);
1679

1680
	ttm_bo_list_ref_sub(bo, put_count, true);
1681 1682 1683 1684 1685

	/**
	 * Wait for GPU, then move to system cached.
	 */

1686
	ret = ttm_bo_wait(bo, false, false);
1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699

	if (unlikely(ret != 0))
		goto out;

	if ((bo->mem.placement & swap_placement) != swap_placement) {
		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;

		ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1700
					     false, false);
1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711
		if (unlikely(ret != 0))
			goto out;
	}

	ttm_bo_unmap_virtual(bo);

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

1712 1713 1714
	if (bo->bdev->driver->swap_notify)
		bo->bdev->driver->swap_notify(bo);

J
Jan Engelhardt 已提交
1715
	ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1716 1717 1718 1719 1720 1721 1722 1723
out:

	/**
	 *
	 * Unreserve without putting on LRU to avoid swapping out an
	 * already swapped buffer.
	 */

1724
	__ttm_bo_unreserve(bo);
1725 1726 1727 1728 1729 1730
	kref_put(&bo->list_kref, ttm_bo_release_list);
	return ret;
}

void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
{
1731
	while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1732 1733
		;
}
1734
EXPORT_SYMBOL(ttm_bo_swapout_all);
1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757

/**
 * 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;
1758
	ret = __ttm_bo_reserve(bo, true, false, NULL);
1759 1760
	if (unlikely(ret != 0))
		goto out_unlock;
1761
	__ttm_bo_unreserve(bo);
1762 1763 1764 1765 1766

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