ttm_bo.c 44.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
/**************************************************************************
 *
 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sub license, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 **************************************************************************/
/*
 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
 */

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

126 127 128 129 130 131 132 133 134 135 136

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

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

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

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

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

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

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

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

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

182 183 184 185 186 187
static void ttm_bo_ref_bug(struct kref *list_kref)
{
	BUG();
}

void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
188 189 190
{
	if (!list_empty(&bo->swap)) {
		list_del_init(&bo->swap);
191
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
192 193 194
	}
	if (!list_empty(&bo->lru)) {
		list_del_init(&bo->lru);
195
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
196 197
	}

198 199 200 201
	/*
	 * TODO: Add a driver hook to delete from
	 * driver-specific LRU's here.
	 */
202 203
}

204
void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
205
{
206
	spin_lock(&bo->glob->lru_lock);
207
	ttm_bo_del_from_lru(bo);
208
	spin_unlock(&bo->glob->lru_lock);
209
}
210
EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
211

212 213
void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
{
214
	reservation_object_assert_held(bo->resv);
215

216
	ttm_bo_del_from_lru(bo);
217
	ttm_bo_add_to_lru(bo);
218 219 220
}
EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);

221 222 223 224 225 226
/*
 * 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;
227
	struct ttm_bo_global *glob = bo->glob;
228 229 230
	int ret = 0;
	uint32_t page_flags = 0;

231
	reservation_object_assert_held(bo->resv);
232 233
	bo->ttm = NULL;

D
Dave Airlie 已提交
234 235 236
	if (bdev->need_dma32)
		page_flags |= TTM_PAGE_FLAG_DMA32;

237 238 239 240 241
	switch (bo->type) {
	case ttm_bo_type_device:
		if (zero_alloc)
			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
	case ttm_bo_type_kernel:
242 243
		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
						      page_flags, glob->dummy_read_page);
244 245 246
		if (unlikely(bo->ttm == NULL))
			ret = -ENOMEM;
		break;
247 248 249 250 251 252 253 254 255 256
	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;
257
	default:
J
Joe Perches 已提交
258
		pr_err("Illegal buffer object type\n");
259 260 261 262 263 264 265 266
		ret = -EINVAL;
		break;
	}

	return ret;
}

static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
267 268
				  struct ttm_mem_reg *mem, bool evict,
				  struct ttm_operation_ctx *ctx)
269 270 271 272 273 274 275 276 277
{
	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 ||
278 279 280 281 282 283 284
	    ((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);
	}
285 286 287 288 289

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

290 291
	if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
		if (bo->ttm == NULL) {
292 293
			bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
			ret = ttm_bo_add_ttm(bo, zero);
294 295 296
			if (ret)
				goto out_err;
		}
297 298 299

		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
		if (ret)
300
			goto out_err;
301 302

		if (mem->mem_type != TTM_PL_SYSTEM) {
303
			ret = ttm_tt_bind(bo->ttm, mem, ctx);
304 305 306 307 308
			if (ret)
				goto out_err;
		}

		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
309
			if (bdev->driver->move_notify)
310
				bdev->driver->move_notify(bo, evict, mem);
311
			bo->mem = *mem;
312 313 314 315 316
			mem->mm_node = NULL;
			goto moved;
		}
	}

317
	if (bdev->driver->move_notify)
318
		bdev->driver->move_notify(bo, evict, mem);
319

320 321
	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
322
		ret = ttm_bo_move_ttm(bo, ctx, mem);
323
	else if (bdev->driver->move)
324
		ret = bdev->driver->move(bo, evict, ctx, mem);
325
	else
326
		ret = ttm_bo_move_memcpy(bo, ctx, mem);
327

328 329 330 331 332
	if (ret) {
		if (bdev->driver->move_notify) {
			struct ttm_mem_reg tmp_mem = *mem;
			*mem = bo->mem;
			bo->mem = tmp_mem;
333
			bdev->driver->move_notify(bo, false, mem);
334
			bo->mem = *mem;
335
			*mem = tmp_mem;
336
		}
337

338 339
		goto out_err;
	}
340

341 342
moved:
	if (bo->evicted) {
343 344 345 346 347
		if (bdev->driver->invalidate_caches) {
			ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
			if (ret)
				pr_err("Can not flush read caches\n");
		}
348 349 350
		bo->evicted = false;
	}

C
Christian König 已提交
351
	if (bo->mem.mm_node)
352
		bo->offset = (bo->mem.start << PAGE_SHIFT) +
353
		    bdev->man[bo->mem.mem_type].gpu_offset;
C
Christian König 已提交
354
	else
355
		bo->offset = 0;
356

357
	ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
358 359 360 361
	return 0;

out_err:
	new_man = &bdev->man[bo->mem.mem_type];
362
	if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
363 364 365 366 367 368 369
		ttm_tt_destroy(bo->ttm);
		bo->ttm = NULL;
	}

	return ret;
}

370
/**
371
 * Call bo::reserved.
372
 * Will release GPU memory type usage on destruction.
373 374 375
 * This is the place to put in driver specific hooks to release
 * driver private resources.
 * Will release the bo::reserved lock.
376 377 378 379
 */

static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
{
380
	if (bo->bdev->driver->move_notify)
381
		bo->bdev->driver->move_notify(bo, false, NULL);
382

383 384
	ttm_tt_destroy(bo->ttm);
	bo->ttm = NULL;
385
	ttm_bo_mem_put(bo, &bo->mem);
386 387
}

388 389 390 391 392 393 394
static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
{
	int r;

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

395
	BUG_ON(!reservation_object_trylock(&bo->ttm_resv));
396 397

	r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv);
398
	if (r)
399 400 401 402 403
		reservation_object_unlock(&bo->ttm_resv);

	return r;
}

404 405 406
static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
{
	struct reservation_object_list *fobj;
407
	struct dma_fence *fence;
408 409
	int i;

410 411
	fobj = reservation_object_get_list(&bo->ttm_resv);
	fence = reservation_object_get_excl(&bo->ttm_resv);
412
	if (fence && !fence->ops->signaled)
413
		dma_fence_enable_sw_signaling(fence);
414 415 416 417 418 419

	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)
420
			dma_fence_enable_sw_signaling(fence);
421 422 423
	}
}

424
static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
425 426
{
	struct ttm_bo_device *bdev = bo->bdev;
427
	struct ttm_bo_global *glob = bo->glob;
428 429
	int ret;

430 431 432 433 434 435 436 437 438 439 440
	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;
	}

441
	spin_lock(&glob->lru_lock);
442
	ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY;
M
Maarten Lankhorst 已提交
443
	if (!ret) {
444
		if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) {
445
			ttm_bo_del_from_lru(bo);
M
Maarten Lankhorst 已提交
446
			spin_unlock(&glob->lru_lock);
447
			if (bo->resv != &bo->ttm_resv)
448
				reservation_object_unlock(&bo->ttm_resv);
449

M
Maarten Lankhorst 已提交
450
			ttm_bo_cleanup_memtype_use(bo);
451
			reservation_object_unlock(bo->resv);
M
Maarten Lankhorst 已提交
452
			return;
453 454 455
		}

		ttm_bo_flush_all_fences(bo);
456 457 458 459 460 461 462 463 464 465 466

		/*
		 * 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);
		}

467
		reservation_object_unlock(bo->resv);
468
	}
469 470
	if (bo->resv != &bo->ttm_resv)
		reservation_object_unlock(&bo->ttm_resv);
471

472
error:
473 474 475 476 477 478 479 480 481
	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);
}

/**
482
 * function ttm_bo_cleanup_refs
483 484 485
 * If bo idle, remove from delayed- and lru lists, and unref.
 * If not idle, do nothing.
 *
486
 * Must be called with lru_lock and reservation held, this function
487
 * will drop the lru lock and optionally the reservation lock before returning.
488
 *
489 490
 * @interruptible         Any sleeps should occur interruptibly.
 * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
491
 * @unlock_resv           Unlock the reservation lock as well.
492 493
 */

494 495 496
static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
			       bool interruptible, bool no_wait_gpu,
			       bool unlock_resv)
497 498
{
	struct ttm_bo_global *glob = bo->glob;
499
	struct reservation_object *resv;
500
	int ret;
501

502 503 504 505 506 507 508 509 510
	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;
511

512
	if (ret && !no_wait_gpu) {
M
Maarten Lankhorst 已提交
513
		long lret;
514

515 516
		if (unlock_resv)
			reservation_object_unlock(bo->resv);
M
Maarten Lankhorst 已提交
517 518
		spin_unlock(&glob->lru_lock);

519
		lret = reservation_object_wait_timeout_rcu(resv, true,
M
Maarten Lankhorst 已提交
520 521 522 523 524 525 526
							   interruptible,
							   30 * HZ);

		if (lret < 0)
			return lret;
		else if (lret == 0)
			return -EBUSY;
527

528
		spin_lock(&glob->lru_lock);
529 530 531 532 533 534 535 536 537
		if (unlock_resv && !reservation_object_trylock(bo->resv)) {
			/*
			 * We raced, and lost, someone else holds the reservation now,
			 * and is probably busy in ttm_bo_cleanup_memtype_use.
			 *
			 * Even if it's not the case, because we finished waiting any
			 * delayed destruction would succeed, so just return success
			 * here.
			 */
538 539 540
			spin_unlock(&glob->lru_lock);
			return 0;
		}
541
		ret = 0;
542
	}
543

544
	if (ret || unlikely(list_empty(&bo->ddestroy))) {
545 546
		if (unlock_resv)
			reservation_object_unlock(bo->resv);
547
		spin_unlock(&glob->lru_lock);
548
		return ret;
549 550
	}

551
	ttm_bo_del_from_lru(bo);
552
	list_del_init(&bo->ddestroy);
553
	kref_put(&bo->list_kref, ttm_bo_ref_bug);
554 555 556

	spin_unlock(&glob->lru_lock);
	ttm_bo_cleanup_memtype_use(bo);
557 558 559

	if (unlock_resv)
		reservation_object_unlock(bo->resv);
560 561

	return 0;
562 563 564 565 566 567
}

/**
 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 * encountered buffers.
 */
568
static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
569
{
570
	struct ttm_bo_global *glob = bdev->glob;
571 572
	struct list_head removed;
	bool empty;
573

574
	INIT_LIST_HEAD(&removed);
575

576 577 578
	spin_lock(&glob->lru_lock);
	while (!list_empty(&bdev->ddestroy)) {
		struct ttm_buffer_object *bo;
579

580 581 582 583
		bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
				      ddestroy);
		kref_get(&bo->list_kref);
		list_move_tail(&bo->ddestroy, &removed);
584

585 586 587
		if (remove_all || bo->resv != &bo->ttm_resv) {
			spin_unlock(&glob->lru_lock);
			reservation_object_lock(bo->resv, NULL);
588

589 590 591 592 593
			spin_lock(&glob->lru_lock);
			ttm_bo_cleanup_refs(bo, false, !remove_all, true);

		} else if (reservation_object_trylock(bo->resv)) {
			ttm_bo_cleanup_refs(bo, false, !remove_all, true);
594 595
		} else {
			spin_unlock(&glob->lru_lock);
596
		}
597

598
		kref_put(&bo->list_kref, ttm_bo_release_list);
599
		spin_lock(&glob->lru_lock);
600
	}
601 602
	list_splice_tail(&removed, &bdev->ddestroy);
	empty = list_empty(&bdev->ddestroy);
603
	spin_unlock(&glob->lru_lock);
604 605

	return empty;
606 607 608 609 610 611 612
}

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

613
	if (!ttm_bo_delayed_delete(bdev, false)) {
614 615 616 617 618 619 620 621 622 623
		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;
624
	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
625

626
	drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
627 628 629
	ttm_mem_io_lock(man, false);
	ttm_mem_io_free_vm(bo);
	ttm_mem_io_unlock(man);
630
	ttm_bo_cleanup_refs_or_queue(bo);
631 632 633 634 635 636 637 638 639 640 641 642
	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);

643 644 645 646 647 648 649 650 651 652 653 654 655 656
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);

657 658
static int ttm_bo_evict(struct ttm_buffer_object *bo,
			struct ttm_operation_ctx *ctx)
659 660 661
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_reg evict_mem;
662 663
	struct ttm_placement placement;
	int ret = 0;
664

665
	reservation_object_assert_held(bo->resv);
666 667 668

	evict_mem = bo->mem;
	evict_mem.mm_node = NULL;
669 670
	evict_mem.bus.io_reserved_vm = false;
	evict_mem.bus.io_reserved_count = 0;
671

672 673
	placement.num_placement = 0;
	placement.num_busy_placement = 0;
674
	bdev->driver->evict_flags(bo, &placement);
675
	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
676
	if (ret) {
677
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
678 679
			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
			       bo);
680 681
			ttm_bo_mem_space_debug(bo, &placement);
		}
682 683 684
		goto out;
	}

685
	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
686
	if (unlikely(ret)) {
687
		if (ret != -ERESTARTSYS)
J
Joe Perches 已提交
688
			pr_err("Buffer eviction failed\n");
689
		ttm_bo_mem_put(bo, &evict_mem);
690 691
		goto out;
	}
692 693 694 695 696
	bo->evicted = true;
out:
	return ret;
}

697 698 699 700 701 702 703 704 705 706 707 708 709 710
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);

711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738
/**
 * Check the target bo is allowable to be evicted or swapout, including cases:
 *
 * a. if share same reservation object with ctx->resv, have assumption
 * reservation objects should already be locked, so not lock again and
 * return true directly when either the opreation allow_reserved_eviction
 * or the target bo already is in delayed free list;
 *
 * b. Otherwise, trylock it.
 */
static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
			struct ttm_operation_ctx *ctx, bool *locked)
{
	bool ret = false;

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

	return ret;
}

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

751
	spin_lock(&glob->lru_lock);
752 753
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &man->lru[i], lru) {
754 755
			if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked))
				continue;
756

757 758
			if (place && !bdev->driver->eviction_valuable(bo,
								      place)) {
759 760
				if (locked)
					reservation_object_unlock(bo->resv);
761 762 763
				continue;
			}
			break;
764
		}
765

766 767
		/* If the inner loop terminated early, we have our candidate */
		if (&bo->lru != &man->lru[i])
768
			break;
769 770

		bo = NULL;
771 772
	}

773
	if (!bo) {
774
		spin_unlock(&glob->lru_lock);
775
		return -EBUSY;
776 777
	}

778
	kref_get(&bo->list_kref);
779

780
	if (!list_empty(&bo->ddestroy)) {
781 782
		ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
					  ctx->no_wait_gpu, locked);
783
		kref_put(&bo->list_kref, ttm_bo_release_list);
784
		return ret;
785 786
	}

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

790
	ret = ttm_bo_evict(bo, ctx);
791
	if (locked) {
792
		ttm_bo_unreserve(bo);
793 794
	} else {
		spin_lock(&glob->lru_lock);
795
		ttm_bo_add_to_lru(bo);
796 797
		spin_unlock(&glob->lru_lock);
	}
798

799
	kref_put(&bo->list_kref, ttm_bo_release_list);
800 801 802
	return ret;
}

803 804
void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
{
805
	struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
806

807 808
	if (mem->mm_node)
		(*man->func->put_node)(man, mem);
809 810 811
}
EXPORT_SYMBOL(ttm_bo_mem_put);

812 813 814 815 816 817 818
/**
 * 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)
{
819
	struct dma_fence *fence;
820 821 822
	int ret;

	spin_lock(&man->move_lock);
823
	fence = dma_fence_get(man->move);
824 825 826 827 828 829 830 831 832
	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;

833
		dma_fence_put(bo->moving);
834 835 836 837 838 839
		bo->moving = fence;
	}

	return 0;
}

840 841 842 843
/**
 * 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.
 */
844 845
static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
					uint32_t mem_type,
846
					const struct ttm_place *place,
847
					struct ttm_mem_reg *mem,
848
					struct ttm_operation_ctx *ctx)
849
{
850
	struct ttm_bo_device *bdev = bo->bdev;
851 852 853 854
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	int ret;

	do {
855
		ret = (*man->func->get_node)(man, bo, place, mem);
856 857
		if (unlikely(ret != 0))
			return ret;
858
		if (mem->mm_node)
859
			break;
R
Roger He 已提交
860
		ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
861 862 863 864
		if (unlikely(ret != 0))
			return ret;
	} while (1);
	mem->mem_type = mem_type;
865
	return ttm_bo_add_move_fence(bo, man, mem);
866 867
}

868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892
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;
}

893 894
static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
				 uint32_t mem_type,
895
				 const struct ttm_place *place,
896
				 uint32_t *masked_placement)
897 898 899
{
	uint32_t cur_flags = ttm_bo_type_flags(mem_type);

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

903
	if ((place->flags & man->available_caching) == 0)
904 905
		return false;

906
	cur_flags |= (place->flags & man->available_caching);
907 908

	*masked_placement = cur_flags;
909 910 911 912 913 914 915 916 917 918 919 920
	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,
921 922
			struct ttm_placement *placement,
			struct ttm_mem_reg *mem,
923
			struct ttm_operation_ctx *ctx)
924 925 926 927 928 929 930
{
	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;
931
	bool has_erestartsys = false;
932
	int i, ret;
933

934 935 936 937
	ret = reservation_object_reserve_shared(bo->resv);
	if (unlikely(ret))
		return ret;

938
	mem->mm_node = NULL;
939
	for (i = 0; i < placement->num_placement; ++i) {
940 941 942
		const struct ttm_place *place = &placement->placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
943 944
		if (ret)
			return ret;
945
		man = &bdev->man[mem_type];
946 947
		if (!man->has_type || !man->use_type)
			continue;
948

949
		type_ok = ttm_bo_mt_compatible(man, mem_type, place,
950
						&cur_flags);
951 952 953 954

		if (!type_ok)
			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
		if (mem_type == TTM_PL_SYSTEM)
			break;

968 969 970
		ret = (*man->func->get_node)(man, bo, place, mem);
		if (unlikely(ret))
			return ret;
971 972 973 974 975 976 977

		if (mem->mm_node) {
			ret = ttm_bo_add_move_fence(bo, man, mem);
			if (unlikely(ret)) {
				(*man->func->put_node)(man, mem);
				return ret;
			}
978
			break;
979
		}
980 981
	}

982
	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
983 984 985 986 987
		mem->mem_type = mem_type;
		mem->placement = cur_flags;
		return 0;
	}

988
	for (i = 0; i < placement->num_busy_placement; ++i) {
989 990 991
		const struct ttm_place *place = &placement->busy_placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
992 993
		if (ret)
			return ret;
994
		man = &bdev->man[mem_type];
995
		if (!man->has_type || !man->use_type)
996
			continue;
997
		if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
998 999
			continue;

1000
		type_found = true;
1001 1002
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
1003 1004 1005 1006
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
1007
		ttm_flag_masked(&cur_flags, place->flags,
1008
				~TTM_PL_MASK_MEMTYPE);
1009

1010 1011 1012 1013 1014 1015 1016
		if (mem_type == TTM_PL_SYSTEM) {
			mem->mem_type = mem_type;
			mem->placement = cur_flags;
			mem->mm_node = NULL;
			return 0;
		}

1017
		ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx);
1018 1019 1020 1021
		if (ret == 0 && mem->mm_node) {
			mem->placement = cur_flags;
			return 0;
		}
1022 1023
		if (ret == -ERESTARTSYS)
			has_erestartsys = true;
1024
	}
1025 1026

	if (!type_found) {
1027
		pr_err(TTM_PFX "No compatible memory type found\n");
1028 1029 1030 1031
		return -EINVAL;
	}

	return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1032 1033 1034
}
EXPORT_SYMBOL(ttm_bo_mem_space);

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

1042
	reservation_object_assert_held(bo->resv);
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
	ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1053 1054
	if (ret)
		goto out_unlock;
1055
	ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1056
out_unlock:
1057 1058
	if (ret && mem.mm_node)
		ttm_bo_mem_put(bo, &mem);
1059 1060 1061
	return ret;
}

1062 1063 1064 1065
static bool ttm_bo_places_compat(const struct ttm_place *places,
				 unsigned num_placement,
				 struct ttm_mem_reg *mem,
				 uint32_t *new_flags)
1066
{
1067
	unsigned i;
1068

1069 1070
	for (i = 0; i < num_placement; i++) {
		const struct ttm_place *heap = &places[i];
1071

1072
		if (mem->mm_node && (mem->start < heap->fpfn ||
1073
		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1074 1075 1076
			continue;

		*new_flags = heap->flags;
1077
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1078 1079 1080
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
		    (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
		     (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1081
			return true;
1082
	}
1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099
	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;
1100 1101

	return false;
1102
}
1103
EXPORT_SYMBOL(ttm_bo_mem_compat);
1104

1105
int ttm_bo_validate(struct ttm_buffer_object *bo,
1106 1107
		    struct ttm_placement *placement,
		    struct ttm_operation_ctx *ctx)
1108 1109
{
	int ret;
1110
	uint32_t new_flags;
1111

1112
	reservation_object_assert_held(bo->resv);
1113 1114 1115
	/*
	 * Check whether we need to move buffer.
	 */
1116
	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1117
		ret = ttm_bo_move_buffer(bo, placement, ctx);
1118
		if (ret)
1119
			return ret;
1120 1121 1122 1123 1124
	} else {
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the compatible memory placement flags to the active flags
		 */
1125
		ttm_flag_masked(&bo->mem.placement, new_flags,
1126
				~TTM_PL_MASK_MEMTYPE);
1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137
	}
	/*
	 * 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;
}
1138
EXPORT_SYMBOL(ttm_bo_validate);
1139

1140 1141 1142 1143 1144 1145
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,
1146
			 struct ttm_operation_ctx *ctx,
1147 1148 1149 1150 1151
			 struct file *persistent_swap_storage,
			 size_t acc_size,
			 struct sg_table *sg,
			 struct reservation_object *resv,
			 void (*destroy) (struct ttm_buffer_object *))
1152
{
1153
	int ret = 0;
1154
	unsigned long num_pages;
1155
	struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1156
	bool locked;
1157

1158
	ret = ttm_mem_global_alloc(mem_glob, acc_size, ctx);
1159
	if (ret) {
J
Joe Perches 已提交
1160
		pr_err("Out of kernel memory\n");
1161 1162 1163 1164 1165 1166
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
		return -ENOMEM;
	}
1167 1168 1169

	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
	if (num_pages == 0) {
J
Joe Perches 已提交
1170
		pr_err("Illegal buffer object size\n");
1171 1172 1173 1174
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
1175
		ttm_mem_global_free(mem_glob, acc_size);
1176 1177 1178 1179 1180 1181 1182 1183 1184 1185
		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);
1186
	INIT_LIST_HEAD(&bo->io_reserve_lru);
1187
	mutex_init(&bo->wu_mutex);
1188
	bo->bdev = bdev;
1189
	bo->glob = bdev->glob;
1190 1191
	bo->type = type;
	bo->num_pages = num_pages;
1192
	bo->mem.size = num_pages << PAGE_SHIFT;
1193 1194 1195 1196
	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;
1197 1198
	bo->mem.bus.io_reserved_vm = false;
	bo->mem.bus.io_reserved_count = 0;
1199
	bo->moving = NULL;
1200
	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
J
Jan Engelhardt 已提交
1201
	bo->persistent_swap_storage = persistent_swap_storage;
1202
	bo->acc_size = acc_size;
1203
	bo->sg = sg;
1204 1205
	if (resv) {
		bo->resv = resv;
1206
		reservation_object_assert_held(bo->resv);
1207 1208 1209
	} else {
		bo->resv = &bo->ttm_resv;
	}
1210
	reservation_object_init(&bo->ttm_resv);
1211
	atomic_inc(&bo->glob->bo_count);
1212
	drm_vma_node_reset(&bo->vma_node);
1213
	bo->priority = 0;
1214 1215 1216 1217 1218

	/*
	 * For ttm_bo_type_device buffers, allocate
	 * address space from the device.
	 */
1219 1220
	if (bo->type == ttm_bo_type_device ||
	    bo->type == ttm_bo_type_sg)
1221 1222
		ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
					 bo->mem.num_pages);
1223

1224 1225 1226 1227
	/* passed reservation objects should already be locked,
	 * since otherwise lockdep will be angered in radeon.
	 */
	if (!resv) {
1228
		locked = reservation_object_trylock(bo->resv);
1229 1230
		WARN_ON(!locked);
	}
1231

1232 1233
	if (likely(!ret))
		ret = ttm_bo_validate(bo, placement, ctx);
1234

1235
	if (unlikely(ret)) {
1236 1237 1238
		if (!resv)
			ttm_bo_unreserve(bo);

1239 1240 1241 1242 1243
		ttm_bo_unref(&bo);
		return ret;
	}

	if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1244 1245 1246 1247 1248
		spin_lock(&bo->glob->lru_lock);
		ttm_bo_add_to_lru(bo);
		spin_unlock(&bo->glob->lru_lock);
	}

1249 1250
	return ret;
}
1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265
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 *))
{
1266
	struct ttm_operation_ctx ctx = { interruptible, false };
1267 1268 1269
	int ret;

	ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1270
				   page_alignment, &ctx,
1271 1272 1273 1274 1275 1276 1277 1278 1279 1280
				   persistent_swap_storage, acc_size,
				   sg, resv, destroy);
	if (ret)
		return ret;

	if (!resv)
		ttm_bo_unreserve(bo);

	return 0;
}
1281
EXPORT_SYMBOL(ttm_bo_init);
1282

1283 1284 1285
size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
		       unsigned long bo_size,
		       unsigned struct_size)
1286
{
1287 1288
	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
	size_t size = 0;
1289

1290
	size += ttm_round_pot(struct_size);
F
Felix Kuehling 已提交
1291
	size += ttm_round_pot(npages * sizeof(void *));
1292 1293
	size += ttm_round_pot(sizeof(struct ttm_tt));
	return size;
1294
}
1295 1296 1297 1298 1299 1300 1301 1302 1303 1304
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
1305
	size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1306 1307 1308 1309
	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
	return size;
}
EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1310

1311 1312 1313 1314 1315 1316
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
1317
			struct file *persistent_swap_storage,
1318
			struct ttm_buffer_object **p_bo)
1319 1320
{
	struct ttm_buffer_object *bo;
1321
	size_t acc_size;
1322
	int ret;
1323 1324

	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1325
	if (unlikely(bo == NULL))
1326 1327
		return -ENOMEM;

1328
	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1329
	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1330
			  interruptible, persistent_swap_storage, acc_size,
1331
			  NULL, NULL, NULL);
1332 1333 1334 1335 1336
	if (likely(ret == 0))
		*p_bo = bo;

	return ret;
}
T
Thomas Hellstrom 已提交
1337
EXPORT_SYMBOL(ttm_bo_create);
1338 1339

static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1340
				   unsigned mem_type)
1341
{
1342
	struct ttm_operation_ctx ctx = { false, false };
1343
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1344
	struct ttm_bo_global *glob = bdev->glob;
1345
	struct dma_fence *fence;
1346
	int ret;
1347
	unsigned i;
1348 1349 1350 1351 1352

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

1353
	spin_lock(&glob->lru_lock);
1354 1355 1356
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		while (!list_empty(&man->lru[i])) {
			spin_unlock(&glob->lru_lock);
R
Roger He 已提交
1357
			ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx);
1358
			if (ret)
1359
				return ret;
1360
			spin_lock(&glob->lru_lock);
1361
		}
1362
	}
1363
	spin_unlock(&glob->lru_lock);
1364 1365

	spin_lock(&man->move_lock);
1366
	fence = dma_fence_get(man->move);
1367 1368 1369
	spin_unlock(&man->move_lock);

	if (fence) {
1370 1371
		ret = dma_fence_wait(fence, false);
		dma_fence_put(fence);
1372 1373
		if (ret)
			return ret;
1374 1375
	}

1376 1377 1378 1379 1380
	return 0;
}

int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
{
R
Roel Kluin 已提交
1381
	struct ttm_mem_type_manager *man;
1382 1383 1384
	int ret = -EINVAL;

	if (mem_type >= TTM_NUM_MEM_TYPES) {
J
Joe Perches 已提交
1385
		pr_err("Illegal memory type %d\n", mem_type);
1386 1387
		return ret;
	}
R
Roel Kluin 已提交
1388
	man = &bdev->man[mem_type];
1389 1390

	if (!man->has_type) {
J
Joe Perches 已提交
1391 1392
		pr_err("Trying to take down uninitialized memory manager type %u\n",
		       mem_type);
1393 1394 1395 1396 1397 1398 1399 1400
		return ret;
	}

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

	ret = 0;
	if (mem_type > 0) {
1401 1402 1403 1404 1405
		ret = ttm_bo_force_list_clean(bdev, mem_type);
		if (ret) {
			pr_err("Cleanup eviction failed\n");
			return ret;
		}
1406

1407
		ret = (*man->func->takedown)(man);
1408 1409
	}

1410 1411 1412
	dma_fence_put(man->move);
	man->move = NULL;

1413 1414 1415 1416 1417 1418 1419 1420 1421
	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
1422
		pr_err("Illegal memory manager memory type %u\n", mem_type);
1423 1424 1425 1426
		return -EINVAL;
	}

	if (!man->has_type) {
J
Joe Perches 已提交
1427
		pr_err("Memory type %u has not been initialized\n", mem_type);
1428 1429 1430
		return 0;
	}

1431
	return ttm_bo_force_list_clean(bdev, mem_type);
1432 1433 1434 1435
}
EXPORT_SYMBOL(ttm_bo_evict_mm);

int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1436
			unsigned long p_size)
1437
{
H
Huang Rui 已提交
1438
	int ret;
1439
	struct ttm_mem_type_manager *man;
1440
	unsigned i;
1441

1442
	BUG_ON(type >= TTM_NUM_MEM_TYPES);
1443
	man = &bdev->man[type];
1444
	BUG_ON(man->has_type);
1445 1446 1447
	man->io_reserve_fastpath = true;
	man->use_io_reserve_lru = false;
	mutex_init(&man->io_reserve_mutex);
1448
	spin_lock_init(&man->move_lock);
1449
	INIT_LIST_HEAD(&man->io_reserve_lru);
1450 1451 1452 1453

	ret = bdev->driver->init_mem_type(bdev, type, man);
	if (ret)
		return ret;
1454
	man->bdev = bdev;
1455 1456

	if (type != TTM_PL_SYSTEM) {
1457
		ret = (*man->func->init)(man, p_size);
1458 1459 1460 1461 1462 1463 1464
		if (ret)
			return ret;
	}
	man->has_type = true;
	man->use_type = true;
	man->size = p_size;

1465 1466
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&man->lru[i]);
1467
	man->move = NULL;
1468 1469 1470 1471 1472

	return 0;
}
EXPORT_SYMBOL(ttm_bo_init_mm);

1473 1474 1475 1476 1477 1478 1479 1480 1481
static void ttm_bo_global_kobj_release(struct kobject *kobj)
{
	struct ttm_bo_global *glob =
		container_of(kobj, struct ttm_bo_global, kobj);

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

1482
void ttm_bo_global_release(struct drm_global_reference *ref)
1483 1484 1485 1486 1487 1488 1489 1490
{
	struct ttm_bo_global *glob = ref->object;

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

1491
int ttm_bo_global_init(struct drm_global_reference *ref)
1492 1493 1494 1495 1496
{
	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;
1497
	unsigned i;
1498 1499 1500 1501

	mutex_init(&glob->device_list_mutex);
	spin_lock_init(&glob->lru_lock);
	glob->mem_glob = bo_ref->mem_glob;
1502
	glob->mem_glob->bo_glob = glob;
1503 1504 1505 1506 1507 1508 1509
	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);

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

1510 1511
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&glob->swap_lru[i]);
1512 1513 1514
	INIT_LIST_HEAD(&glob->device_list);
	atomic_set(&glob->bo_count, 0);

1515 1516
	ret = kobject_init_and_add(
		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1517 1518 1519 1520 1521 1522 1523 1524 1525 1526
	if (unlikely(ret != 0))
		kobject_put(&glob->kobj);
	return ret;
out_no_drp:
	kfree(glob);
	return ret;
}
EXPORT_SYMBOL(ttm_bo_global_init);


1527 1528 1529 1530 1531
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;
1532
	struct ttm_bo_global *glob = bdev->glob;
1533 1534 1535 1536 1537 1538 1539

	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
1540 1541
				pr_err("DRM memory manager type %d is not clean\n",
				       i);
1542 1543 1544 1545 1546
			}
			man->has_type = false;
		}
	}

1547 1548 1549 1550
	mutex_lock(&glob->device_list_mutex);
	list_del(&bdev->device_list);
	mutex_unlock(&glob->device_list_mutex);

1551
	cancel_delayed_work_sync(&bdev->wq);
1552

1553
	if (ttm_bo_delayed_delete(bdev, true))
1554
		pr_debug("Delayed destroy list was clean\n");
1555

1556
	spin_lock(&glob->lru_lock);
1557 1558
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		if (list_empty(&bdev->man[0].lru[0]))
1559
			pr_debug("Swap list %d was clean\n", i);
1560
	spin_unlock(&glob->lru_lock);
1561

1562
	drm_vma_offset_manager_destroy(&bdev->vma_manager);
1563 1564 1565 1566 1567 1568

	return ret;
}
EXPORT_SYMBOL(ttm_bo_device_release);

int ttm_bo_device_init(struct ttm_bo_device *bdev,
1569 1570
		       struct ttm_bo_global *glob,
		       struct ttm_bo_driver *driver,
1571
		       struct address_space *mapping,
D
Dave Airlie 已提交
1572
		       uint64_t file_page_offset,
D
Dave Airlie 已提交
1573
		       bool need_dma32)
1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584
{
	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.
	 */
1585
	ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1586
	if (unlikely(ret != 0))
1587
		goto out_no_sys;
1588

1589 1590
	drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
				    0x10000000);
1591 1592
	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
	INIT_LIST_HEAD(&bdev->ddestroy);
1593
	bdev->dev_mapping = mapping;
1594
	bdev->glob = glob;
D
Dave Airlie 已提交
1595
	bdev->need_dma32 = need_dma32;
1596 1597 1598
	mutex_lock(&glob->device_list_mutex);
	list_add_tail(&bdev->device_list, &glob->device_list);
	mutex_unlock(&glob->device_list_mutex);
1599 1600

	return 0;
1601
out_no_sys:
1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626
	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;
}

1627
void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1628 1629 1630
{
	struct ttm_bo_device *bdev = bo->bdev;

1631
	drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1632
	ttm_mem_io_free_vm(bo);
1633
}
1634 1635 1636 1637 1638 1639 1640 1641 1642

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);
1643
}
1644 1645


1646
EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1647 1648

int ttm_bo_wait(struct ttm_buffer_object *bo,
1649
		bool interruptible, bool no_wait)
1650
{
C
Christian König 已提交
1651 1652 1653 1654 1655 1656 1657 1658
	long timeout = 15 * HZ;

	if (no_wait) {
		if (reservation_object_test_signaled_rcu(bo->resv, true))
			return 0;
		else
			return -EBUSY;
	}
1659

C
Christian König 已提交
1660 1661
	timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
						      interruptible, timeout);
1662 1663 1664 1665 1666 1667
	if (timeout < 0)
		return timeout;

	if (timeout == 0)
		return -EBUSY;

C
Christian König 已提交
1668
	reservation_object_add_excl_fence(bo->resv, NULL);
1669
	return 0;
1670 1671 1672 1673 1674 1675 1676 1677
}
EXPORT_SYMBOL(ttm_bo_wait);

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

	/*
1678
	 * Using ttm_bo_reserve makes sure the lru lists are updated.
1679 1680
	 */

1681
	ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1682 1683
	if (unlikely(ret != 0))
		return ret;
1684
	ret = ttm_bo_wait(bo, true, no_wait);
1685 1686 1687 1688 1689
	if (likely(ret == 0))
		atomic_inc(&bo->cpu_writers);
	ttm_bo_unreserve(bo);
	return ret;
}
1690
EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1691 1692 1693

void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
{
1694
	atomic_dec(&bo->cpu_writers);
1695
}
1696
EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1697 1698 1699 1700 1701

/**
 * A buffer object shrink method that tries to swap out the first
 * buffer object on the bo_global::swap_lru list.
 */
1702
int ttm_bo_swapout(struct ttm_bo_global *glob, struct ttm_operation_ctx *ctx)
1703 1704 1705
{
	struct ttm_buffer_object *bo;
	int ret = -EBUSY;
1706
	bool locked;
1707
	unsigned i;
1708

1709
	spin_lock(&glob->lru_lock);
1710 1711
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1712 1713
			if (ttm_bo_evict_swapout_allowable(bo, ctx, &locked)) {
				ret = 0;
1714
				break;
1715
			}
1716
		}
1717 1718 1719
		if (!ret)
			break;
	}
1720

1721 1722 1723 1724
	if (ret) {
		spin_unlock(&glob->lru_lock);
		return ret;
	}
1725

1726
	kref_get(&bo->list_kref);
1727

1728
	if (!list_empty(&bo->ddestroy)) {
1729
		ret = ttm_bo_cleanup_refs(bo, false, false, true);
1730 1731
		kref_put(&bo->list_kref, ttm_bo_release_list);
		return ret;
1732 1733
	}

1734
	ttm_bo_del_from_lru(bo);
1735
	spin_unlock(&glob->lru_lock);
1736 1737

	/**
1738
	 * Move to system cached
1739 1740
	 */

1741 1742
	if (bo->mem.mem_type != TTM_PL_SYSTEM ||
	    bo->ttm->caching_state != tt_cached) {
1743
		struct ttm_operation_ctx ctx = { false, false };
1744 1745 1746 1747 1748 1749 1750
		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;

1751
		ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1752 1753 1754 1755
		if (unlikely(ret != 0))
			goto out;
	}

1756 1757 1758 1759 1760 1761 1762 1763
	/**
	 * Make sure BO is idle.
	 */

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

1764 1765 1766 1767 1768 1769 1770
	ttm_bo_unmap_virtual(bo);

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

1771 1772 1773
	if (bo->bdev->driver->swap_notify)
		bo->bdev->driver->swap_notify(bo);

J
Jan Engelhardt 已提交
1774
	ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1775 1776 1777 1778 1779 1780 1781 1782
out:

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

1783
	reservation_object_unlock(bo->resv);
1784 1785 1786
	kref_put(&bo->list_kref, ttm_bo_release_list);
	return ret;
}
1787
EXPORT_SYMBOL(ttm_bo_swapout);
1788 1789 1790

void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
{
1791 1792 1793 1794 1795 1796
	struct ttm_operation_ctx ctx = {
		.interruptible = false,
		.no_wait_gpu = false
	};

	while (ttm_bo_swapout(bdev->glob, &ctx) == 0)
1797 1798
		;
}
1799
EXPORT_SYMBOL(ttm_bo_swapout_all);
1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822

/**
 * 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;
1823 1824 1825
	ret = reservation_object_lock_interruptible(bo->resv, NULL);
	if (ret == -EINTR)
		ret = -ERESTARTSYS;
1826 1827
	if (unlikely(ret != 0))
		goto out_unlock;
1828
	reservation_object_unlock(bo->resv);
1829 1830 1831 1832 1833

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