ttm_bo.c 44.0 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
	int pos;
61

62 63 64 65 66 67
	pos = ffs(place->flags & TTM_PL_MASK_MEM);
	if (unlikely(!pos))
		return -EINVAL;

	*mem_type = pos - 1;
	return 0;
68 69
}

70
static void ttm_mem_type_debug(struct ttm_bo_device *bdev, int mem_type)
71
{
72
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
73
	struct drm_printer p = drm_debug_printer(TTM_PFX);
74

J
Joe Perches 已提交
75 76 77
	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);
78
	pr_err("    gpu_offset: 0x%08llX\n", man->gpu_offset);
J
Joe Perches 已提交
79 80 81
	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);
82
	if (mem_type != TTM_PL_SYSTEM)
83
		(*man->func->debug)(man, &p);
84 85 86 87 88 89 90
}

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
91 92 93
	pr_err("No space for %p (%lu pages, %luK, %luM)\n",
	       bo, bo->mem.num_pages, bo->mem.size >> 10,
	       bo->mem.size >> 20);
94
	for (i = 0; i < placement->num_placement; i++) {
95
		ret = ttm_mem_type_from_place(&placement->placement[i],
96 97 98
						&mem_type);
		if (ret)
			return;
J
Joe Perches 已提交
99
		pr_err("  placement[%d]=0x%08X (%d)\n",
100
		       i, placement->placement[i].flags, mem_type);
101
		ttm_mem_type_debug(bo->bdev, mem_type);
102 103 104
	}
}

105 106 107 108 109 110 111
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);

112 113
	return snprintf(buffer, PAGE_SIZE, "%d\n",
				atomic_read(&glob->bo_count));
114 115 116 117 118 119 120
}

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

121
static const struct sysfs_ops ttm_bo_global_ops = {
122 123 124 125 126 127 128 129 130
	.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
};

131 132 133 134 135 136 137 138 139 140 141

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;
142
	size_t acc_size = bo->acc_size;
143

144 145
	BUG_ON(kref_read(&bo->list_kref));
	BUG_ON(kref_read(&bo->kref));
146 147 148 149
	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));
150
	ttm_tt_destroy(bo->ttm);
151
	atomic_dec(&bo->glob->bo_count);
152
	dma_fence_put(bo->moving);
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
{
	struct ttm_bo_device *bdev = bo->bdev;
167
	struct ttm_mem_type_manager *man;
168

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

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

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

175 176
		man = &bdev->man[bo->mem.mem_type];
		list_add_tail(&bo->lru, &man->lru[bo->priority]);
177 178
		kref_get(&bo->list_kref);

179
		if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
180 181
			list_add_tail(&bo->swap,
				      &bo->glob->swap_lru[bo->priority]);
182 183 184 185
			kref_get(&bo->list_kref);
		}
	}
}
186
EXPORT_SYMBOL(ttm_bo_add_to_lru);
187

188 189 190 191 192 193
static void ttm_bo_ref_bug(struct kref *list_kref)
{
	BUG();
}

void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
194 195 196
{
	if (!list_empty(&bo->swap)) {
		list_del_init(&bo->swap);
197
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
198 199 200
	}
	if (!list_empty(&bo->lru)) {
		list_del_init(&bo->lru);
201
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
202 203
	}

204 205 206 207
	/*
	 * TODO: Add a driver hook to delete from
	 * driver-specific LRU's here.
	 */
208 209
}

210
void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
211
{
212
	spin_lock(&bo->glob->lru_lock);
213
	ttm_bo_del_from_lru(bo);
214
	spin_unlock(&bo->glob->lru_lock);
215
}
216
EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
217

218 219 220 221
void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
{
	lockdep_assert_held(&bo->resv->lock.base);

222
	ttm_bo_del_from_lru(bo);
223
	ttm_bo_add_to_lru(bo);
224 225 226
}
EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);

227 228 229 230 231 232
/*
 * 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;
233
	struct ttm_bo_global *glob = bo->glob;
234 235 236 237 238 239
	int ret = 0;
	uint32_t page_flags = 0;

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

D
Dave Airlie 已提交
240 241 242
	if (bdev->need_dma32)
		page_flags |= TTM_PAGE_FLAG_DMA32;

243 244 245 246 247
	switch (bo->type) {
	case ttm_bo_type_device:
		if (zero_alloc)
			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
	case ttm_bo_type_kernel:
248 249
		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
						      page_flags, glob->dummy_read_page);
250 251 252
		if (unlikely(bo->ttm == NULL))
			ret = -ENOMEM;
		break;
253 254 255 256 257 258 259 260 261 262
	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;
263
	default:
J
Joe Perches 已提交
264
		pr_err("Illegal buffer object type\n");
265 266 267 268 269 270 271 272 273
		ret = -EINVAL;
		break;
	}

	return ret;
}

static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
				  struct ttm_mem_reg *mem,
274
				  bool evict, bool interruptible,
275
				  bool no_wait_gpu)
276 277 278 279 280 281 282 283 284
{
	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 ||
285 286 287 288 289 290 291
	    ((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);
	}
292 293 294 295 296

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

297 298
	if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
		if (bo->ttm == NULL) {
299 300
			bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
			ret = ttm_bo_add_ttm(bo, zero);
301 302 303
			if (ret)
				goto out_err;
		}
304 305 306

		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
		if (ret)
307
			goto out_err;
308 309 310 311 312 313 314 315

		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) {
316
			if (bdev->driver->move_notify)
317
				bdev->driver->move_notify(bo, evict, mem);
318
			bo->mem = *mem;
319 320 321 322 323
			mem->mm_node = NULL;
			goto moved;
		}
	}

324
	if (bdev->driver->move_notify)
325
		bdev->driver->move_notify(bo, evict, mem);
326

327 328
	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
329
		ret = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, mem);
330 331
	else if (bdev->driver->move)
		ret = bdev->driver->move(bo, evict, interruptible,
332
					 no_wait_gpu, mem);
333
	else
334
		ret = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, mem);
335

336 337 338 339 340
	if (ret) {
		if (bdev->driver->move_notify) {
			struct ttm_mem_reg tmp_mem = *mem;
			*mem = bo->mem;
			bo->mem = tmp_mem;
341
			bdev->driver->move_notify(bo, false, mem);
342
			bo->mem = *mem;
343
			*mem = tmp_mem;
344
		}
345

346 347
		goto out_err;
	}
348

349 350
moved:
	if (bo->evicted) {
351 352 353 354 355
		if (bdev->driver->invalidate_caches) {
			ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
			if (ret)
				pr_err("Can not flush read caches\n");
		}
356 357 358 359
		bo->evicted = false;
	}

	if (bo->mem.mm_node) {
360
		bo->offset = (bo->mem.start << PAGE_SHIFT) +
361 362
		    bdev->man[bo->mem.mem_type].gpu_offset;
		bo->cur_placement = bo->mem.placement;
363 364
	} else
		bo->offset = 0;
365 366 367 368 369

	return 0;

out_err:
	new_man = &bdev->man[bo->mem.mem_type];
370
	if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
371 372 373 374 375 376 377
		ttm_tt_destroy(bo->ttm);
		bo->ttm = NULL;
	}

	return ret;
}

378
/**
379
 * Call bo::reserved.
380
 * Will release GPU memory type usage on destruction.
381 382 383
 * This is the place to put in driver specific hooks to release
 * driver private resources.
 * Will release the bo::reserved lock.
384 385 386 387
 */

static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
{
388
	if (bo->bdev->driver->move_notify)
389
		bo->bdev->driver->move_notify(bo, false, NULL);
390

391 392
	ttm_tt_destroy(bo->ttm);
	bo->ttm = NULL;
393
	ttm_bo_mem_put(bo, &bo->mem);
394

395
	ww_mutex_unlock (&bo->resv->lock);
396 397
}

398 399 400 401 402 403 404 405
static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
{
	int r;

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

	reservation_object_init(&bo->ttm_resv);
406
	BUG_ON(!reservation_object_trylock(&bo->ttm_resv));
407 408 409 410 411 412 413 414 415 416

	r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv);
	if (r) {
		reservation_object_unlock(&bo->ttm_resv);
		reservation_object_fini(&bo->ttm_resv);
	}

	return r;
}

417 418 419
static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
{
	struct reservation_object_list *fobj;
420
	struct dma_fence *fence;
421 422
	int i;

423 424
	fobj = reservation_object_get_list(&bo->ttm_resv);
	fence = reservation_object_get_excl(&bo->ttm_resv);
425
	if (fence && !fence->ops->signaled)
426
		dma_fence_enable_sw_signaling(fence);
427 428 429 430 431 432

	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)
433
			dma_fence_enable_sw_signaling(fence);
434 435 436
	}
}

437
static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
438 439
{
	struct ttm_bo_device *bdev = bo->bdev;
440
	struct ttm_bo_global *glob = bo->glob;
441 442
	int ret;

443 444 445 446 447 448 449 450 451 452 453
	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;
	}

454
	spin_lock(&glob->lru_lock);
455
	ret = __ttm_bo_reserve(bo, false, true, NULL);
M
Maarten Lankhorst 已提交
456
	if (!ret) {
457
		if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) {
458
			ttm_bo_del_from_lru(bo);
M
Maarten Lankhorst 已提交
459
			spin_unlock(&glob->lru_lock);
460
			if (bo->resv != &bo->ttm_resv) {
461
				reservation_object_unlock(&bo->ttm_resv);
462 463 464
				reservation_object_fini(&bo->ttm_resv);
			}

M
Maarten Lankhorst 已提交
465 466
			ttm_bo_cleanup_memtype_use(bo);
			return;
467 468 469
		}

		ttm_bo_flush_all_fences(bo);
470 471 472 473 474 475 476 477 478 479 480

		/*
		 * 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);
		}

481
		__ttm_bo_unreserve(bo);
482
	}
483 484
	if (bo->resv != &bo->ttm_resv)
		reservation_object_unlock(&bo->ttm_resv);
485

486
error:
487 488 489 490 491 492 493 494 495
	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);
}

/**
496
 * function ttm_bo_cleanup_refs_and_unlock
497 498 499
 * If bo idle, remove from delayed- and lru lists, and unref.
 * If not idle, do nothing.
 *
500 501 502
 * Must be called with lru_lock and reservation held, this function
 * will drop both before returning.
 *
503 504 505 506
 * @interruptible         Any sleeps should occur interruptibly.
 * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
 */

507 508 509
static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
					  bool interruptible,
					  bool no_wait_gpu)
510 511
{
	struct ttm_bo_global *glob = bo->glob;
512
	struct reservation_object *resv;
513
	int ret;
514

515 516 517 518 519 520 521 522 523
	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;
524

525
	if (ret && !no_wait_gpu) {
M
Maarten Lankhorst 已提交
526 527 528 529
		long lret;
		ww_mutex_unlock(&bo->resv->lock);
		spin_unlock(&glob->lru_lock);

530
		lret = reservation_object_wait_timeout_rcu(resv, true,
M
Maarten Lankhorst 已提交
531 532 533 534 535 536 537
							   interruptible,
							   30 * HZ);

		if (lret < 0)
			return lret;
		else if (lret == 0)
			return -EBUSY;
538

539
		spin_lock(&glob->lru_lock);
540
		ret = __ttm_bo_reserve(bo, false, true, NULL);
541

542 543 544 545 546 547 548 549 550 551 552 553
		/*
		 * 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;
		}
554
	}
555

556
	if (ret || unlikely(list_empty(&bo->ddestroy))) {
557
		__ttm_bo_unreserve(bo);
558
		spin_unlock(&glob->lru_lock);
559
		return ret;
560 561
	}

562
	ttm_bo_del_from_lru(bo);
563 564
	if (!list_empty(&bo->ddestroy) && (bo->resv != &bo->ttm_resv))
		reservation_object_fini(&bo->ttm_resv);
565
	list_del_init(&bo->ddestroy);
566
	kref_put(&bo->list_kref, ttm_bo_ref_bug);
567 568 569 570 571

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

	return 0;
572 573 574 575 576 577 578 579 580
}

/**
 * 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)
{
581
	struct ttm_bo_global *glob = bdev->glob;
582 583
	struct ttm_buffer_object *entry = NULL;
	int ret = 0;
584

585
	spin_lock(&glob->lru_lock);
586 587 588 589 590 591 592 593 594 595 596 597 598
	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);
599 600 601
			kref_get(&nentry->list_kref);
		}

602
		ret = __ttm_bo_reserve(entry, false, true, NULL);
603 604
		if (remove_all && ret) {
			spin_unlock(&glob->lru_lock);
605
			ret = __ttm_bo_reserve(entry, false, false, NULL);
606 607 608
			spin_lock(&glob->lru_lock);
		}

609 610 611 612 613 614
		if (!ret)
			ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
							     !remove_all);
		else
			spin_unlock(&glob->lru_lock);

615
		kref_put(&entry->list_kref, ttm_bo_release_list);
616 617 618 619
		entry = nentry;

		if (ret || !entry)
			goto out;
620

621
		spin_lock(&glob->lru_lock);
622
		if (list_empty(&entry->ddestroy))
623 624 625
			break;
	}

626 627 628 629 630
out_unlock:
	spin_unlock(&glob->lru_lock);
out:
	if (entry)
		kref_put(&entry->list_kref, ttm_bo_release_list);
631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649
	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;
650
	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
651

652
	drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
653 654 655
	ttm_mem_io_lock(man, false);
	ttm_mem_io_free_vm(bo);
	ttm_mem_io_unlock(man);
656
	ttm_bo_cleanup_refs_or_queue(bo);
657 658 659 660 661 662 663 664 665 666 667 668
	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);

669 670 671 672 673 674 675 676 677 678 679 680 681 682
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);

683
static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
684
			bool no_wait_gpu)
685 686 687
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_reg evict_mem;
688 689
	struct ttm_placement placement;
	int ret = 0;
690

691
	lockdep_assert_held(&bo->resv->lock.base);
692 693 694

	evict_mem = bo->mem;
	evict_mem.mm_node = NULL;
695 696
	evict_mem.bus.io_reserved_vm = false;
	evict_mem.bus.io_reserved_count = 0;
697

698 699
	placement.num_placement = 0;
	placement.num_busy_placement = 0;
700 701
	bdev->driver->evict_flags(bo, &placement);
	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
702
				no_wait_gpu);
703
	if (ret) {
704
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
705 706
			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
			       bo);
707 708
			ttm_bo_mem_space_debug(bo, &placement);
		}
709 710 711 712
		goto out;
	}

	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
713
				     no_wait_gpu);
714
	if (unlikely(ret)) {
715
		if (ret != -ERESTARTSYS)
J
Joe Perches 已提交
716
			pr_err("Buffer eviction failed\n");
717
		ttm_bo_mem_put(bo, &evict_mem);
718 719
		goto out;
	}
720 721 722 723 724
	bo->evicted = true;
out:
	return ret;
}

725 726 727 728 729 730 731 732 733 734 735 736 737 738
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);

739 740
static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
				uint32_t mem_type,
741
				const struct ttm_place *place,
742
				bool interruptible,
743
				bool no_wait_gpu)
744 745 746 747
{
	struct ttm_bo_global *glob = bdev->glob;
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	struct ttm_buffer_object *bo;
748
	int ret = -EBUSY;
749
	unsigned i;
750

751
	spin_lock(&glob->lru_lock);
752 753 754 755 756
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &man->lru[i], lru) {
			ret = __ttm_bo_reserve(bo, false, true, NULL);
			if (ret)
				continue;
757

758 759 760 761 762 763 764 765
			if (place && !bdev->driver->eviction_valuable(bo,
								      place)) {
				__ttm_bo_unreserve(bo);
				ret = -EBUSY;
				continue;
			}

			break;
766
		}
767

768 769
		if (!ret)
			break;
770 771 772
	}

	if (ret) {
773
		spin_unlock(&glob->lru_lock);
774
		return ret;
775 776
	}

777
	kref_get(&bo->list_kref);
778

779
	if (!list_empty(&bo->ddestroy)) {
780 781
		ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
						     no_wait_gpu);
782
		kref_put(&bo->list_kref, ttm_bo_release_list);
783
		return ret;
784 785
	}

786
	ttm_bo_del_from_lru(bo);
787
	spin_unlock(&glob->lru_lock);
788 789 790

	BUG_ON(ret != 0);

791
	ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
792
	ttm_bo_unreserve(bo);
793

794
	kref_put(&bo->list_kref, ttm_bo_release_list);
795 796 797
	return ret;
}

798 799
void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
{
800
	struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
801

802 803
	if (mem->mm_node)
		(*man->func->put_node)(man, mem);
804 805 806
}
EXPORT_SYMBOL(ttm_bo_mem_put);

807 808 809 810 811 812 813
/**
 * 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)
{
814
	struct dma_fence *fence;
815 816 817
	int ret;

	spin_lock(&man->move_lock);
818
	fence = dma_fence_get(man->move);
819 820 821 822 823 824 825 826 827
	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;

828
		dma_fence_put(bo->moving);
829 830 831 832 833 834
		bo->moving = fence;
	}

	return 0;
}

835 836 837 838
/**
 * 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.
 */
839 840
static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
					uint32_t mem_type,
841
					const struct ttm_place *place,
842
					struct ttm_mem_reg *mem,
843 844
					bool interruptible,
					bool no_wait_gpu)
845
{
846
	struct ttm_bo_device *bdev = bo->bdev;
847 848 849 850
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	int ret;

	do {
851
		ret = (*man->func->get_node)(man, bo, place, mem);
852 853
		if (unlikely(ret != 0))
			return ret;
854
		if (mem->mm_node)
855
			break;
856
		ret = ttm_mem_evict_first(bdev, mem_type, place,
857
					  interruptible, no_wait_gpu);
858 859 860 861
		if (unlikely(ret != 0))
			return ret;
	} while (1);
	mem->mem_type = mem_type;
862
	return ttm_bo_add_move_fence(bo, man, mem);
863 864
}

865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889
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;
}

890 891
static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
				 uint32_t mem_type,
892
				 const struct ttm_place *place,
893
				 uint32_t *masked_placement)
894 895 896
{
	uint32_t cur_flags = ttm_bo_type_flags(mem_type);

897
	if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
898 899
		return false;

900
	if ((place->flags & man->available_caching) == 0)
901 902
		return false;

903
	cur_flags |= (place->flags & man->available_caching);
904 905

	*masked_placement = cur_flags;
906 907 908 909 910 911 912 913 914 915 916 917
	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,
918 919
			struct ttm_placement *placement,
			struct ttm_mem_reg *mem,
920
			bool interruptible,
921
			bool no_wait_gpu)
922 923 924 925 926 927 928
{
	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;
929
	bool has_erestartsys = false;
930
	int i, ret;
931

932 933 934 935
	ret = reservation_object_reserve_shared(bo->resv);
	if (unlikely(ret))
		return ret;

936
	mem->mm_node = NULL;
937
	for (i = 0; i < placement->num_placement; ++i) {
938 939 940
		const struct ttm_place *place = &placement->placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
941 942
		if (ret)
			return ret;
943
		man = &bdev->man[mem_type];
944 945
		if (!man->has_type || !man->use_type)
			continue;
946

947
		type_ok = ttm_bo_mt_compatible(man, mem_type, place,
948
						&cur_flags);
949 950 951 952

		if (!type_ok)
			continue;

953
		type_found = true;
954 955
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
956 957 958 959
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
960
		ttm_flag_masked(&cur_flags, place->flags,
961
				~TTM_PL_MASK_MEMTYPE);
962

963 964 965
		if (mem_type == TTM_PL_SYSTEM)
			break;

966 967 968
		ret = (*man->func->get_node)(man, bo, place, mem);
		if (unlikely(ret))
			return ret;
969 970 971 972 973 974 975

		if (mem->mm_node) {
			ret = ttm_bo_add_move_fence(bo, man, mem);
			if (unlikely(ret)) {
				(*man->func->put_node)(man, mem);
				return ret;
			}
976
			break;
977
		}
978 979
	}

980
	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
981 982 983 984 985
		mem->mem_type = mem_type;
		mem->placement = cur_flags;
		return 0;
	}

986
	for (i = 0; i < placement->num_busy_placement; ++i) {
987 988 989
		const struct ttm_place *place = &placement->busy_placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
990 991
		if (ret)
			return ret;
992
		man = &bdev->man[mem_type];
993
		if (!man->has_type || !man->use_type)
994
			continue;
995
		if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
996 997
			continue;

998
		type_found = true;
999 1000
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
1001 1002 1003 1004
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
1005
		ttm_flag_masked(&cur_flags, place->flags,
1006
				~TTM_PL_MASK_MEMTYPE);
1007

1008 1009 1010 1011 1012 1013 1014
		if (mem_type == TTM_PL_SYSTEM) {
			mem->mem_type = mem_type;
			mem->placement = cur_flags;
			mem->mm_node = NULL;
			return 0;
		}

1015
		ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
1016
						interruptible, no_wait_gpu);
1017 1018 1019 1020
		if (ret == 0 && mem->mm_node) {
			mem->placement = cur_flags;
			return 0;
		}
1021 1022
		if (ret == -ERESTARTSYS)
			has_erestartsys = true;
1023
	}
1024 1025

	if (!type_found) {
1026
		pr_err(TTM_PFX "No compatible memory type found\n");
1027 1028 1029 1030
		return -EINVAL;
	}

	return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1031 1032 1033
}
EXPORT_SYMBOL(ttm_bo_mem_space);

1034
static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1035
			struct ttm_placement *placement,
1036
			bool interruptible,
1037
			bool no_wait_gpu)
1038 1039 1040 1041
{
	int ret = 0;
	struct ttm_mem_reg mem;

1042
	lockdep_assert_held(&bo->resv->lock.base);
1043 1044 1045 1046

	mem.num_pages = bo->num_pages;
	mem.size = mem.num_pages << PAGE_SHIFT;
	mem.page_alignment = bo->mem.page_alignment;
1047 1048
	mem.bus.io_reserved_vm = false;
	mem.bus.io_reserved_count = 0;
1049 1050 1051
	/*
	 * Determine where to move the buffer.
	 */
1052 1053
	ret = ttm_bo_mem_space(bo, placement, &mem,
			       interruptible, no_wait_gpu);
1054 1055
	if (ret)
		goto out_unlock;
1056 1057
	ret = ttm_bo_handle_move_mem(bo, &mem, false,
				     interruptible, no_wait_gpu);
1058
out_unlock:
1059 1060
	if (ret && mem.mm_node)
		ttm_bo_mem_put(bo, &mem);
1061 1062 1063
	return ret;
}

1064 1065 1066 1067
static bool ttm_bo_places_compat(const struct ttm_place *places,
				 unsigned num_placement,
				 struct ttm_mem_reg *mem,
				 uint32_t *new_flags)
1068
{
1069
	unsigned i;
1070

1071 1072
	for (i = 0; i < num_placement; i++) {
		const struct ttm_place *heap = &places[i];
1073

1074
		if (mem->mm_node && (mem->start < heap->fpfn ||
1075
		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1076 1077 1078
			continue;

		*new_flags = heap->flags;
1079
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1080 1081 1082
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
		    (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
		     (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1083
			return true;
1084
	}
1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101
	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;
1102 1103

	return false;
1104
}
1105
EXPORT_SYMBOL(ttm_bo_mem_compat);
1106

1107 1108
int ttm_bo_validate(struct ttm_buffer_object *bo,
			struct ttm_placement *placement,
1109
			bool interruptible,
1110
			bool no_wait_gpu)
1111 1112
{
	int ret;
1113
	uint32_t new_flags;
1114

1115
	lockdep_assert_held(&bo->resv->lock.base);
1116 1117 1118
	/*
	 * Check whether we need to move buffer.
	 */
1119
	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1120 1121
		ret = ttm_bo_move_buffer(bo, placement, interruptible,
					 no_wait_gpu);
1122
		if (ret)
1123
			return ret;
1124 1125 1126 1127 1128
	} else {
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the compatible memory placement flags to the active flags
		 */
1129
		ttm_flag_masked(&bo->mem.placement, new_flags,
1130
				~TTM_PL_MASK_MEMTYPE);
1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141
	}
	/*
	 * 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;
}
1142
EXPORT_SYMBOL(ttm_bo_validate);
1143

1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155
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,
			 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 *))
1156
{
1157
	int ret = 0;
1158
	unsigned long num_pages;
1159
	struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1160
	bool locked;
1161 1162 1163

	ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
	if (ret) {
J
Joe Perches 已提交
1164
		pr_err("Out of kernel memory\n");
1165 1166 1167 1168 1169 1170
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
		return -ENOMEM;
	}
1171 1172 1173

	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
	if (num_pages == 0) {
J
Joe Perches 已提交
1174
		pr_err("Illegal buffer object size\n");
1175 1176 1177 1178
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
1179
		ttm_mem_global_free(mem_glob, acc_size);
1180 1181 1182 1183 1184 1185 1186 1187 1188 1189
		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);
1190
	INIT_LIST_HEAD(&bo->io_reserve_lru);
1191
	mutex_init(&bo->wu_mutex);
1192
	bo->bdev = bdev;
1193
	bo->glob = bdev->glob;
1194 1195
	bo->type = type;
	bo->num_pages = num_pages;
1196
	bo->mem.size = num_pages << PAGE_SHIFT;
1197 1198 1199 1200
	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;
1201 1202
	bo->mem.bus.io_reserved_vm = false;
	bo->mem.bus.io_reserved_count = 0;
1203
	bo->moving = NULL;
1204
	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
J
Jan Engelhardt 已提交
1205
	bo->persistent_swap_storage = persistent_swap_storage;
1206
	bo->acc_size = acc_size;
1207
	bo->sg = sg;
1208 1209 1210 1211 1212 1213 1214
	if (resv) {
		bo->resv = resv;
		lockdep_assert_held(&bo->resv->lock.base);
	} else {
		bo->resv = &bo->ttm_resv;
		reservation_object_init(&bo->ttm_resv);
	}
1215
	atomic_inc(&bo->glob->bo_count);
1216
	drm_vma_node_reset(&bo->vma_node);
1217
	bo->priority = 0;
1218 1219 1220 1221 1222

	/*
	 * For ttm_bo_type_device buffers, allocate
	 * address space from the device.
	 */
1223 1224
	if (bo->type == ttm_bo_type_device ||
	    bo->type == ttm_bo_type_sg)
1225 1226
		ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
					 bo->mem.num_pages);
1227

1228 1229 1230 1231 1232 1233 1234
	/* 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);
	}
1235

1236 1237
	if (likely(!ret))
		ret = ttm_bo_validate(bo, placement, interruptible, false);
1238

1239
	if (unlikely(ret)) {
1240 1241 1242
		if (!resv)
			ttm_bo_unreserve(bo);

1243 1244 1245 1246 1247
		ttm_bo_unref(&bo);
		return ret;
	}

	if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1248 1249 1250 1251 1252
		spin_lock(&bo->glob->lru_lock);
		ttm_bo_add_to_lru(bo);
		spin_unlock(&bo->glob->lru_lock);
	}

1253 1254
	return ret;
}
1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283
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 *))
{
	int ret;

	ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
				   page_alignment, interruptible,
				   persistent_swap_storage, acc_size,
				   sg, resv, destroy);
	if (ret)
		return ret;

	if (!resv)
		ttm_bo_unreserve(bo);

	return 0;
}
1284
EXPORT_SYMBOL(ttm_bo_init);
1285

1286 1287 1288
size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
		       unsigned long bo_size,
		       unsigned struct_size)
1289
{
1290 1291
	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
	size_t size = 0;
1292

1293
	size += ttm_round_pot(struct_size);
F
Felix Kuehling 已提交
1294
	size += ttm_round_pot(npages * sizeof(void *));
1295 1296
	size += ttm_round_pot(sizeof(struct ttm_tt));
	return size;
1297
}
1298 1299 1300 1301 1302 1303 1304 1305 1306 1307
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
1308
	size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1309 1310 1311 1312
	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
	return size;
}
EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1313

1314 1315 1316 1317 1318 1319
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
1320
			struct file *persistent_swap_storage,
1321
			struct ttm_buffer_object **p_bo)
1322 1323
{
	struct ttm_buffer_object *bo;
1324
	size_t acc_size;
1325
	int ret;
1326 1327

	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1328
	if (unlikely(bo == NULL))
1329 1330
		return -ENOMEM;

1331
	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1332
	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1333
			  interruptible, persistent_swap_storage, acc_size,
1334
			  NULL, NULL, NULL);
1335 1336 1337 1338 1339
	if (likely(ret == 0))
		*p_bo = bo;

	return ret;
}
T
Thomas Hellstrom 已提交
1340
EXPORT_SYMBOL(ttm_bo_create);
1341 1342

static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1343
				   unsigned mem_type)
1344
{
1345
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1346
	struct ttm_bo_global *glob = bdev->glob;
1347
	struct dma_fence *fence;
1348
	int ret;
1349
	unsigned i;
1350 1351 1352 1353 1354

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

1355
	spin_lock(&glob->lru_lock);
1356 1357 1358 1359 1360
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		while (!list_empty(&man->lru[i])) {
			spin_unlock(&glob->lru_lock);
			ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false);
			if (ret)
1361
				return ret;
1362
			spin_lock(&glob->lru_lock);
1363
		}
1364
	}
1365
	spin_unlock(&glob->lru_lock);
1366 1367

	spin_lock(&man->move_lock);
1368
	fence = dma_fence_get(man->move);
1369 1370 1371
	spin_unlock(&man->move_lock);

	if (fence) {
1372 1373
		ret = dma_fence_wait(fence, false);
		dma_fence_put(fence);
1374 1375
		if (ret)
			return ret;
1376 1377
	}

1378 1379 1380 1381 1382
	return 0;
}

int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
{
R
Roel Kluin 已提交
1383
	struct ttm_mem_type_manager *man;
1384 1385 1386
	int ret = -EINVAL;

	if (mem_type >= TTM_NUM_MEM_TYPES) {
J
Joe Perches 已提交
1387
		pr_err("Illegal memory type %d\n", mem_type);
1388 1389
		return ret;
	}
R
Roel Kluin 已提交
1390
	man = &bdev->man[mem_type];
1391 1392

	if (!man->has_type) {
J
Joe Perches 已提交
1393 1394
		pr_err("Trying to take down uninitialized memory manager type %u\n",
		       mem_type);
1395 1396 1397 1398 1399 1400 1401 1402
		return ret;
	}

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

	ret = 0;
	if (mem_type > 0) {
1403 1404 1405 1406 1407
		ret = ttm_bo_force_list_clean(bdev, mem_type);
		if (ret) {
			pr_err("Cleanup eviction failed\n");
			return ret;
		}
1408

1409
		ret = (*man->func->takedown)(man);
1410 1411
	}

1412 1413 1414
	dma_fence_put(man->move);
	man->move = NULL;

1415 1416 1417 1418 1419 1420 1421 1422 1423
	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
1424
		pr_err("Illegal memory manager memory type %u\n", mem_type);
1425 1426 1427 1428
		return -EINVAL;
	}

	if (!man->has_type) {
J
Joe Perches 已提交
1429
		pr_err("Memory type %u has not been initialized\n", mem_type);
1430 1431 1432
		return 0;
	}

1433
	return ttm_bo_force_list_clean(bdev, mem_type);
1434 1435 1436 1437
}
EXPORT_SYMBOL(ttm_bo_evict_mm);

int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1438
			unsigned long p_size)
1439
{
H
Huang Rui 已提交
1440
	int ret;
1441
	struct ttm_mem_type_manager *man;
1442
	unsigned i;
1443

1444
	BUG_ON(type >= TTM_NUM_MEM_TYPES);
1445
	man = &bdev->man[type];
1446
	BUG_ON(man->has_type);
1447 1448 1449
	man->io_reserve_fastpath = true;
	man->use_io_reserve_lru = false;
	mutex_init(&man->io_reserve_mutex);
1450
	spin_lock_init(&man->move_lock);
1451
	INIT_LIST_HEAD(&man->io_reserve_lru);
1452 1453 1454 1455

	ret = bdev->driver->init_mem_type(bdev, type, man);
	if (ret)
		return ret;
1456
	man->bdev = bdev;
1457 1458

	if (type != TTM_PL_SYSTEM) {
1459
		ret = (*man->func->init)(man, p_size);
1460 1461 1462 1463 1464 1465 1466
		if (ret)
			return ret;
	}
	man->has_type = true;
	man->use_type = true;
	man->size = p_size;

1467 1468
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&man->lru[i]);
1469
	man->move = NULL;
1470 1471 1472 1473 1474

	return 0;
}
EXPORT_SYMBOL(ttm_bo_init_mm);

1475 1476 1477 1478 1479 1480 1481 1482 1483 1484
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);
}

1485
void ttm_bo_global_release(struct drm_global_reference *ref)
1486 1487 1488 1489 1490 1491 1492 1493
{
	struct ttm_bo_global *glob = ref->object;

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

1494
int ttm_bo_global_init(struct drm_global_reference *ref)
1495 1496 1497 1498 1499
{
	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;
1500
	unsigned i;
1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511

	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;
	}

1512 1513
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&glob->swap_lru[i]);
1514 1515 1516 1517 1518
	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
1519
		pr_err("Could not register buffer object swapout\n");
1520 1521 1522 1523 1524
		goto out_no_shrink;
	}

	atomic_set(&glob->bo_count, 0);

1525 1526
	ret = kobject_init_and_add(
		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538
	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);


1539 1540 1541 1542 1543
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;
1544
	struct ttm_bo_global *glob = bdev->glob;
1545 1546 1547 1548 1549 1550 1551

	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
1552 1553
				pr_err("DRM memory manager type %d is not clean\n",
				       i);
1554 1555 1556 1557 1558
			}
			man->has_type = false;
		}
	}

1559 1560 1561 1562
	mutex_lock(&glob->device_list_mutex);
	list_del(&bdev->device_list);
	mutex_unlock(&glob->device_list_mutex);

1563
	cancel_delayed_work_sync(&bdev->wq);
1564 1565 1566 1567

	while (ttm_bo_delayed_delete(bdev, true))
		;

1568
	spin_lock(&glob->lru_lock);
1569 1570 1571
	if (list_empty(&bdev->ddestroy))
		TTM_DEBUG("Delayed destroy list was clean\n");

1572 1573 1574
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		if (list_empty(&bdev->man[0].lru[0]))
			TTM_DEBUG("Swap list %d was clean\n", i);
1575
	spin_unlock(&glob->lru_lock);
1576

1577
	drm_vma_offset_manager_destroy(&bdev->vma_manager);
1578 1579 1580 1581 1582 1583

	return ret;
}
EXPORT_SYMBOL(ttm_bo_device_release);

int ttm_bo_device_init(struct ttm_bo_device *bdev,
1584 1585
		       struct ttm_bo_global *glob,
		       struct ttm_bo_driver *driver,
1586
		       struct address_space *mapping,
D
Dave Airlie 已提交
1587
		       uint64_t file_page_offset,
D
Dave Airlie 已提交
1588
		       bool need_dma32)
1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599
{
	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.
	 */
1600
	ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1601
	if (unlikely(ret != 0))
1602
		goto out_no_sys;
1603

1604 1605
	drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
				    0x10000000);
1606 1607
	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
	INIT_LIST_HEAD(&bdev->ddestroy);
1608
	bdev->dev_mapping = mapping;
1609
	bdev->glob = glob;
D
Dave Airlie 已提交
1610
	bdev->need_dma32 = need_dma32;
1611 1612 1613
	mutex_lock(&glob->device_list_mutex);
	list_add_tail(&bdev->device_list, &glob->device_list);
	mutex_unlock(&glob->device_list_mutex);
1614 1615

	return 0;
1616
out_no_sys:
1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641
	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;
}

1642
void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1643 1644 1645
{
	struct ttm_bo_device *bdev = bo->bdev;

1646
	drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1647
	ttm_mem_io_free_vm(bo);
1648
}
1649 1650 1651 1652 1653 1654 1655 1656 1657

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);
1658
}
1659 1660


1661
EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1662 1663

int ttm_bo_wait(struct ttm_buffer_object *bo,
1664
		bool interruptible, bool no_wait)
1665
{
C
Christian König 已提交
1666 1667 1668 1669 1670 1671 1672 1673
	long timeout = 15 * HZ;

	if (no_wait) {
		if (reservation_object_test_signaled_rcu(bo->resv, true))
			return 0;
		else
			return -EBUSY;
	}
1674

C
Christian König 已提交
1675 1676
	timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
						      interruptible, timeout);
1677 1678 1679 1680 1681 1682
	if (timeout < 0)
		return timeout;

	if (timeout == 0)
		return -EBUSY;

C
Christian König 已提交
1683
	reservation_object_add_excl_fence(bo->resv, NULL);
1684
	return 0;
1685 1686 1687 1688 1689 1690 1691 1692
}
EXPORT_SYMBOL(ttm_bo_wait);

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

	/*
1693
	 * Using ttm_bo_reserve makes sure the lru lists are updated.
1694 1695
	 */

1696
	ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1697 1698
	if (unlikely(ret != 0))
		return ret;
1699
	ret = ttm_bo_wait(bo, true, no_wait);
1700 1701 1702 1703 1704
	if (likely(ret == 0))
		atomic_inc(&bo->cpu_writers);
	ttm_bo_unreserve(bo);
	return ret;
}
1705
EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1706 1707 1708

void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
{
1709
	atomic_dec(&bo->cpu_writers);
1710
}
1711
EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1712 1713 1714 1715 1716 1717 1718 1719

/**
 * 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)
{
1720 1721
	struct ttm_bo_global *glob =
	    container_of(shrink, struct ttm_bo_global, shrink);
1722 1723
	struct ttm_buffer_object *bo;
	int ret = -EBUSY;
1724
	unsigned i;
1725

1726
	spin_lock(&glob->lru_lock);
1727 1728 1729 1730 1731 1732
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &glob->swap_lru[i], swap) {
			ret = __ttm_bo_reserve(bo, false, true, NULL);
			if (!ret)
				break;
		}
1733 1734 1735
		if (!ret)
			break;
	}
1736

1737 1738 1739 1740
	if (ret) {
		spin_unlock(&glob->lru_lock);
		return ret;
	}
1741

1742
	kref_get(&bo->list_kref);
1743

1744 1745 1746 1747
	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;
1748 1749
	}

1750
	ttm_bo_del_from_lru(bo);
1751
	spin_unlock(&glob->lru_lock);
1752 1753

	/**
1754
	 * Move to system cached
1755 1756
	 */

1757 1758
	if (bo->mem.mem_type != TTM_PL_SYSTEM ||
	    bo->ttm->caching_state != tt_cached) {
1759 1760 1761 1762 1763 1764 1765 1766
		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,
1767
					     false, false);
1768 1769 1770 1771
		if (unlikely(ret != 0))
			goto out;
	}

1772 1773 1774 1775 1776 1777 1778 1779
	/**
	 * Make sure BO is idle.
	 */

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

1780 1781 1782 1783 1784 1785 1786
	ttm_bo_unmap_virtual(bo);

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

1787 1788 1789
	if (bo->bdev->driver->swap_notify)
		bo->bdev->driver->swap_notify(bo);

J
Jan Engelhardt 已提交
1790
	ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1791 1792 1793 1794 1795 1796 1797 1798
out:

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

1799
	__ttm_bo_unreserve(bo);
1800 1801 1802 1803 1804 1805
	kref_put(&bo->list_kref, ttm_bo_release_list);
	return ret;
}

void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
{
1806
	while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1807 1808
		;
}
1809
EXPORT_SYMBOL(ttm_bo_swapout_all);
1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832

/**
 * 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;
1833
	ret = __ttm_bo_reserve(bo, true, false, NULL);
1834 1835
	if (unlikely(ret != 0))
		goto out_unlock;
1836
	__ttm_bo_unreserve(bo);
1837 1838 1839 1840 1841

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