ttm_bo.c 43.9 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
	reservation_object_fini(&bo->ttm_resv);
154
	mutex_destroy(&bo->wu_mutex);
155 156 157 158 159
	if (bo->destroy)
		bo->destroy(bo);
	else {
		kfree(bo);
	}
160
	ttm_mem_global_free(bdev->glob->mem_glob, acc_size);
161 162
}

163
void ttm_bo_add_to_lru(struct ttm_buffer_object *bo)
164 165
{
	struct ttm_bo_device *bdev = bo->bdev;
166
	struct ttm_mem_type_manager *man;
167

168
	lockdep_assert_held(&bo->resv->lock.base);
169 170 171 172 173

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

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

174 175
		man = &bdev->man[bo->mem.mem_type];
		list_add_tail(&bo->lru, &man->lru[bo->priority]);
176 177
		kref_get(&bo->list_kref);

178
		if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
179 180
			list_add_tail(&bo->swap,
				      &bo->glob->swap_lru[bo->priority]);
181 182 183 184
			kref_get(&bo->list_kref);
		}
	}
}
185
EXPORT_SYMBOL(ttm_bo_add_to_lru);
186

187 188 189 190 191 192
static void ttm_bo_ref_bug(struct kref *list_kref)
{
	BUG();
}

void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
193 194 195
{
	if (!list_empty(&bo->swap)) {
		list_del_init(&bo->swap);
196
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
197 198 199
	}
	if (!list_empty(&bo->lru)) {
		list_del_init(&bo->lru);
200
		kref_put(&bo->list_kref, ttm_bo_ref_bug);
201 202
	}

203 204 205 206
	/*
	 * TODO: Add a driver hook to delete from
	 * driver-specific LRU's here.
	 */
207 208
}

209
void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
210
{
211
	spin_lock(&bo->glob->lru_lock);
212
	ttm_bo_del_from_lru(bo);
213
	spin_unlock(&bo->glob->lru_lock);
214
}
215
EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
216

217 218 219 220
void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
{
	lockdep_assert_held(&bo->resv->lock.base);

221
	ttm_bo_del_from_lru(bo);
222
	ttm_bo_add_to_lru(bo);
223 224 225
}
EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);

226 227 228 229 230 231
/*
 * 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;
232
	struct ttm_bo_global *glob = bo->glob;
233 234 235 236 237 238
	int ret = 0;
	uint32_t page_flags = 0;

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

D
Dave Airlie 已提交
239 240 241
	if (bdev->need_dma32)
		page_flags |= TTM_PAGE_FLAG_DMA32;

242 243 244 245 246
	switch (bo->type) {
	case ttm_bo_type_device:
		if (zero_alloc)
			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
	case ttm_bo_type_kernel:
247 248
		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
						      page_flags, glob->dummy_read_page);
249 250 251
		if (unlikely(bo->ttm == NULL))
			ret = -ENOMEM;
		break;
252 253 254 255 256 257 258 259 260 261
	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;
262
	default:
J
Joe Perches 已提交
263
		pr_err("Illegal buffer object type\n");
264 265 266 267 268 269 270 271
		ret = -EINVAL;
		break;
	}

	return ret;
}

static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
272 273
				  struct ttm_mem_reg *mem, bool evict,
				  struct ttm_operation_ctx *ctx)
274 275 276 277 278 279 280 281 282
{
	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 ||
283 284 285 286 287 288 289
	    ((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);
	}
290 291 292 293 294

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

295 296
	if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
		if (bo->ttm == NULL) {
297 298
			bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
			ret = ttm_bo_add_ttm(bo, zero);
299 300 301
			if (ret)
				goto out_err;
		}
302 303 304

		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
		if (ret)
305
			goto out_err;
306 307 308 309 310 311 312 313

		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) {
314
			if (bdev->driver->move_notify)
315
				bdev->driver->move_notify(bo, evict, mem);
316
			bo->mem = *mem;
317 318 319 320 321
			mem->mm_node = NULL;
			goto moved;
		}
	}

322
	if (bdev->driver->move_notify)
323
		bdev->driver->move_notify(bo, evict, mem);
324

325 326
	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
327
		ret = ttm_bo_move_ttm(bo, ctx, mem);
328
	else if (bdev->driver->move)
329
		ret = bdev->driver->move(bo, evict, ctx, mem);
330
	else
331
		ret = ttm_bo_move_memcpy(bo, ctx, mem);
332

333 334 335 336 337
	if (ret) {
		if (bdev->driver->move_notify) {
			struct ttm_mem_reg tmp_mem = *mem;
			*mem = bo->mem;
			bo->mem = tmp_mem;
338
			bdev->driver->move_notify(bo, false, mem);
339
			bo->mem = *mem;
340
			*mem = tmp_mem;
341
		}
342

343 344
		goto out_err;
	}
345

346 347
moved:
	if (bo->evicted) {
348 349 350 351 352
		if (bdev->driver->invalidate_caches) {
			ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
			if (ret)
				pr_err("Can not flush read caches\n");
		}
353 354 355
		bo->evicted = false;
	}

C
Christian König 已提交
356
	if (bo->mem.mm_node)
357
		bo->offset = (bo->mem.start << PAGE_SHIFT) +
358
		    bdev->man[bo->mem.mem_type].gpu_offset;
C
Christian König 已提交
359
	else
360
		bo->offset = 0;
361

362
	ctx->bytes_moved += bo->num_pages << PAGE_SHIFT;
363 364 365 366
	return 0;

out_err:
	new_man = &bdev->man[bo->mem.mem_type];
367
	if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
368 369 370 371 372 373 374
		ttm_tt_destroy(bo->ttm);
		bo->ttm = NULL;
	}

	return ret;
}

375
/**
376
 * Call bo::reserved.
377
 * Will release GPU memory type usage on destruction.
378 379 380
 * This is the place to put in driver specific hooks to release
 * driver private resources.
 * Will release the bo::reserved lock.
381 382 383 384
 */

static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
{
385
	if (bo->bdev->driver->move_notify)
386
		bo->bdev->driver->move_notify(bo, false, NULL);
387

388 389
	ttm_tt_destroy(bo->ttm);
	bo->ttm = NULL;
390
	ttm_bo_mem_put(bo, &bo->mem);
391 392
}

393 394 395 396 397 398 399
static int ttm_bo_individualize_resv(struct ttm_buffer_object *bo)
{
	int r;

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

400
	BUG_ON(!reservation_object_trylock(&bo->ttm_resv));
401 402

	r = reservation_object_copy_fences(&bo->ttm_resv, bo->resv);
403
	if (r)
404 405 406 407 408
		reservation_object_unlock(&bo->ttm_resv);

	return r;
}

409 410 411
static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
{
	struct reservation_object_list *fobj;
412
	struct dma_fence *fence;
413 414
	int i;

415 416
	fobj = reservation_object_get_list(&bo->ttm_resv);
	fence = reservation_object_get_excl(&bo->ttm_resv);
417
	if (fence && !fence->ops->signaled)
418
		dma_fence_enable_sw_signaling(fence);
419 420 421 422 423 424

	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)
425
			dma_fence_enable_sw_signaling(fence);
426 427 428
	}
}

429
static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
430 431
{
	struct ttm_bo_device *bdev = bo->bdev;
432
	struct ttm_bo_global *glob = bo->glob;
433 434
	int ret;

435 436 437 438 439 440 441 442 443 444 445
	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;
	}

446
	spin_lock(&glob->lru_lock);
447
	ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY;
M
Maarten Lankhorst 已提交
448
	if (!ret) {
449
		if (reservation_object_test_signaled_rcu(&bo->ttm_resv, true)) {
450
			ttm_bo_del_from_lru(bo);
M
Maarten Lankhorst 已提交
451
			spin_unlock(&glob->lru_lock);
452
			if (bo->resv != &bo->ttm_resv)
453
				reservation_object_unlock(&bo->ttm_resv);
454

M
Maarten Lankhorst 已提交
455
			ttm_bo_cleanup_memtype_use(bo);
456
			reservation_object_unlock(bo->resv);
M
Maarten Lankhorst 已提交
457
			return;
458 459 460
		}

		ttm_bo_flush_all_fences(bo);
461 462 463 464 465 466 467 468 469 470 471

		/*
		 * 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);
		}

472
		reservation_object_unlock(bo->resv);
473
	}
474 475
	if (bo->resv != &bo->ttm_resv)
		reservation_object_unlock(&bo->ttm_resv);
476

477
error:
478 479 480 481 482 483 484 485 486
	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);
}

/**
487
 * function ttm_bo_cleanup_refs
488 489 490
 * If bo idle, remove from delayed- and lru lists, and unref.
 * If not idle, do nothing.
 *
491
 * Must be called with lru_lock and reservation held, this function
492
 * will drop the lru lock and optionally the reservation lock before returning.
493
 *
494 495
 * @interruptible         Any sleeps should occur interruptibly.
 * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
496
 * @unlock_resv           Unlock the reservation lock as well.
497 498
 */

499 500 501
static int ttm_bo_cleanup_refs(struct ttm_buffer_object *bo,
			       bool interruptible, bool no_wait_gpu,
			       bool unlock_resv)
502 503
{
	struct ttm_bo_global *glob = bo->glob;
504
	struct reservation_object *resv;
505
	int ret;
506

507 508 509 510 511 512 513 514 515
	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;
516

517
	if (ret && !no_wait_gpu) {
M
Maarten Lankhorst 已提交
518
		long lret;
519

520 521
		if (unlock_resv)
			reservation_object_unlock(bo->resv);
M
Maarten Lankhorst 已提交
522 523
		spin_unlock(&glob->lru_lock);

524
		lret = reservation_object_wait_timeout_rcu(resv, true,
M
Maarten Lankhorst 已提交
525 526 527 528 529 530 531
							   interruptible,
							   30 * HZ);

		if (lret < 0)
			return lret;
		else if (lret == 0)
			return -EBUSY;
532

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

549
	if (ret || unlikely(list_empty(&bo->ddestroy))) {
550 551
		if (unlock_resv)
			reservation_object_unlock(bo->resv);
552
		spin_unlock(&glob->lru_lock);
553
		return ret;
554 555
	}

556
	ttm_bo_del_from_lru(bo);
557
	list_del_init(&bo->ddestroy);
558
	kref_put(&bo->list_kref, ttm_bo_ref_bug);
559 560 561

	spin_unlock(&glob->lru_lock);
	ttm_bo_cleanup_memtype_use(bo);
562 563 564

	if (unlock_resv)
		reservation_object_unlock(bo->resv);
565 566

	return 0;
567 568 569 570 571 572
}

/**
 * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
 * encountered buffers.
 */
573
static bool ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
574
{
575
	struct ttm_bo_global *glob = bdev->glob;
576 577
	struct list_head removed;
	bool empty;
578

579
	INIT_LIST_HEAD(&removed);
580

581 582 583
	spin_lock(&glob->lru_lock);
	while (!list_empty(&bdev->ddestroy)) {
		struct ttm_buffer_object *bo;
584

585 586 587 588
		bo = list_first_entry(&bdev->ddestroy, struct ttm_buffer_object,
				      ddestroy);
		kref_get(&bo->list_kref);
		list_move_tail(&bo->ddestroy, &removed);
589

590 591 592
		if (remove_all || bo->resv != &bo->ttm_resv) {
			spin_unlock(&glob->lru_lock);
			reservation_object_lock(bo->resv, NULL);
593

594 595 596 597 598 599
			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);
		}
600

601
		kref_put(&bo->list_kref, ttm_bo_release_list);
602
		spin_lock(&glob->lru_lock);
603
	}
604 605
	list_splice_tail(&removed, &bdev->ddestroy);
	empty = list_empty(&bdev->ddestroy);
606
	spin_unlock(&glob->lru_lock);
607 608

	return empty;
609 610 611 612 613 614 615
}

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

616
	if (!ttm_bo_delayed_delete(bdev, false)) {
617 618 619 620 621 622 623 624 625 626
		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;
627
	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
628

629
	drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
630 631 632
	ttm_mem_io_lock(man, false);
	ttm_mem_io_free_vm(bo);
	ttm_mem_io_unlock(man);
633
	ttm_bo_cleanup_refs_or_queue(bo);
634 635 636 637 638 639 640 641 642 643 644 645
	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);

646 647 648 649 650 651 652 653 654 655 656 657 658 659
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);

660 661
static int ttm_bo_evict(struct ttm_buffer_object *bo,
			struct ttm_operation_ctx *ctx)
662 663 664
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_reg evict_mem;
665 666
	struct ttm_placement placement;
	int ret = 0;
667

668
	lockdep_assert_held(&bo->resv->lock.base);
669 670 671

	evict_mem = bo->mem;
	evict_mem.mm_node = NULL;
672 673
	evict_mem.bus.io_reserved_vm = false;
	evict_mem.bus.io_reserved_count = 0;
674

675 676
	placement.num_placement = 0;
	placement.num_busy_placement = 0;
677
	bdev->driver->evict_flags(bo, &placement);
678
	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, ctx);
679
	if (ret) {
680
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
681 682
			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
			       bo);
683 684
			ttm_bo_mem_space_debug(bo, &placement);
		}
685 686 687
		goto out;
	}

688
	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, ctx);
689
	if (unlikely(ret)) {
690
		if (ret != -ERESTARTSYS)
J
Joe Perches 已提交
691
			pr_err("Buffer eviction failed\n");
692
		ttm_bo_mem_put(bo, &evict_mem);
693 694
		goto out;
	}
695 696 697 698 699
	bo->evicted = true;
out:
	return ret;
}

700 701 702 703 704 705 706 707 708 709 710 711 712 713
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);

714
static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
715 716
			       uint32_t mem_type,
			       const struct ttm_place *place,
717
			       struct ttm_operation_ctx *ctx)
718 719 720
{
	struct ttm_bo_global *glob = bdev->glob;
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
721 722
	struct ttm_buffer_object *bo = NULL;
	bool locked = false;
723
	unsigned i;
724
	int ret;
725

726
	spin_lock(&glob->lru_lock);
727 728
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &man->lru[i], lru) {
R
Roger He 已提交
729 730 731
			if (bo->resv == ctx->resv) {
				if (!ctx->allow_reserved_eviction &&
				    list_empty(&bo->ddestroy))
732 733 734 735 736 737
					continue;
			} else {
				locked = reservation_object_trylock(bo->resv);
				if (!locked)
					continue;
			}
738

739 740
			if (place && !bdev->driver->eviction_valuable(bo,
								      place)) {
741 742
				if (locked)
					reservation_object_unlock(bo->resv);
743
				locked = false;
744 745 746
				continue;
			}
			break;
747
		}
748

749 750
		/* If the inner loop terminated early, we have our candidate */
		if (&bo->lru != &man->lru[i])
751
			break;
752 753

		bo = NULL;
754 755
	}

756
	if (!bo) {
757
		spin_unlock(&glob->lru_lock);
758
		return -EBUSY;
759 760
	}

761
	kref_get(&bo->list_kref);
762

763
	if (!list_empty(&bo->ddestroy)) {
764 765
		ret = ttm_bo_cleanup_refs(bo, ctx->interruptible,
					  ctx->no_wait_gpu, locked);
766
		kref_put(&bo->list_kref, ttm_bo_release_list);
767
		return ret;
768 769
	}

770
	ttm_bo_del_from_lru(bo);
771
	spin_unlock(&glob->lru_lock);
772

773
	ret = ttm_bo_evict(bo, ctx);
774
	if (locked) {
775
		ttm_bo_unreserve(bo);
776 777
	} else {
		spin_lock(&glob->lru_lock);
778
		ttm_bo_add_to_lru(bo);
779 780
		spin_unlock(&glob->lru_lock);
	}
781

782
	kref_put(&bo->list_kref, ttm_bo_release_list);
783 784 785
	return ret;
}

786 787
void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
{
788
	struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
789

790 791
	if (mem->mm_node)
		(*man->func->put_node)(man, mem);
792 793 794
}
EXPORT_SYMBOL(ttm_bo_mem_put);

795 796 797 798 799 800 801
/**
 * 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)
{
802
	struct dma_fence *fence;
803 804 805
	int ret;

	spin_lock(&man->move_lock);
806
	fence = dma_fence_get(man->move);
807 808 809 810 811 812 813 814 815
	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;

816
		dma_fence_put(bo->moving);
817 818 819 820 821 822
		bo->moving = fence;
	}

	return 0;
}

823 824 825 826
/**
 * 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.
 */
827 828
static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
					uint32_t mem_type,
829
					const struct ttm_place *place,
830
					struct ttm_mem_reg *mem,
831
					struct ttm_operation_ctx *ctx)
832
{
833
	struct ttm_bo_device *bdev = bo->bdev;
834 835 836 837
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	int ret;

	do {
838
		ret = (*man->func->get_node)(man, bo, place, mem);
839 840
		if (unlikely(ret != 0))
			return ret;
841
		if (mem->mm_node)
842
			break;
R
Roger He 已提交
843
		ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
844 845 846 847
		if (unlikely(ret != 0))
			return ret;
	} while (1);
	mem->mem_type = mem_type;
848
	return ttm_bo_add_move_fence(bo, man, mem);
849 850
}

851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875
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;
}

876 877
static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
				 uint32_t mem_type,
878
				 const struct ttm_place *place,
879
				 uint32_t *masked_placement)
880 881 882
{
	uint32_t cur_flags = ttm_bo_type_flags(mem_type);

883
	if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
884 885
		return false;

886
	if ((place->flags & man->available_caching) == 0)
887 888
		return false;

889
	cur_flags |= (place->flags & man->available_caching);
890 891

	*masked_placement = cur_flags;
892 893 894 895 896 897 898 899 900 901 902 903
	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,
904 905
			struct ttm_placement *placement,
			struct ttm_mem_reg *mem,
906
			struct ttm_operation_ctx *ctx)
907 908 909 910 911 912 913
{
	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;
914
	bool has_erestartsys = false;
915
	int i, ret;
916

917 918 919 920
	ret = reservation_object_reserve_shared(bo->resv);
	if (unlikely(ret))
		return ret;

921
	mem->mm_node = NULL;
922
	for (i = 0; i < placement->num_placement; ++i) {
923 924 925
		const struct ttm_place *place = &placement->placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
926 927
		if (ret)
			return ret;
928
		man = &bdev->man[mem_type];
929 930
		if (!man->has_type || !man->use_type)
			continue;
931

932
		type_ok = ttm_bo_mt_compatible(man, mem_type, place,
933
						&cur_flags);
934 935 936 937

		if (!type_ok)
			continue;

938
		type_found = true;
939 940
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
941 942 943 944
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
945
		ttm_flag_masked(&cur_flags, place->flags,
946
				~TTM_PL_MASK_MEMTYPE);
947

948 949 950
		if (mem_type == TTM_PL_SYSTEM)
			break;

951 952 953
		ret = (*man->func->get_node)(man, bo, place, mem);
		if (unlikely(ret))
			return ret;
954 955 956 957 958 959 960

		if (mem->mm_node) {
			ret = ttm_bo_add_move_fence(bo, man, mem);
			if (unlikely(ret)) {
				(*man->func->put_node)(man, mem);
				return ret;
			}
961
			break;
962
		}
963 964
	}

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

971
	for (i = 0; i < placement->num_busy_placement; ++i) {
972 973 974
		const struct ttm_place *place = &placement->busy_placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
975 976
		if (ret)
			return ret;
977
		man = &bdev->man[mem_type];
978
		if (!man->has_type || !man->use_type)
979
			continue;
980
		if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
981 982
			continue;

983
		type_found = true;
984 985
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
986 987 988 989
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
990
		ttm_flag_masked(&cur_flags, place->flags,
991
				~TTM_PL_MASK_MEMTYPE);
992

993 994 995 996 997 998 999
		if (mem_type == TTM_PL_SYSTEM) {
			mem->mem_type = mem_type;
			mem->placement = cur_flags;
			mem->mm_node = NULL;
			return 0;
		}

1000
		ret = ttm_bo_mem_force_space(bo, mem_type, place, mem, ctx);
1001 1002 1003 1004
		if (ret == 0 && mem->mm_node) {
			mem->placement = cur_flags;
			return 0;
		}
1005 1006
		if (ret == -ERESTARTSYS)
			has_erestartsys = true;
1007
	}
1008 1009

	if (!type_found) {
1010
		pr_err(TTM_PFX "No compatible memory type found\n");
1011 1012 1013 1014
		return -EINVAL;
	}

	return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
1015 1016 1017
}
EXPORT_SYMBOL(ttm_bo_mem_space);

1018
static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
1019 1020
			      struct ttm_placement *placement,
			      struct ttm_operation_ctx *ctx)
1021 1022 1023 1024
{
	int ret = 0;
	struct ttm_mem_reg mem;

1025
	lockdep_assert_held(&bo->resv->lock.base);
1026 1027 1028 1029

	mem.num_pages = bo->num_pages;
	mem.size = mem.num_pages << PAGE_SHIFT;
	mem.page_alignment = bo->mem.page_alignment;
1030 1031
	mem.bus.io_reserved_vm = false;
	mem.bus.io_reserved_count = 0;
1032 1033 1034
	/*
	 * Determine where to move the buffer.
	 */
1035
	ret = ttm_bo_mem_space(bo, placement, &mem, ctx);
1036 1037
	if (ret)
		goto out_unlock;
1038
	ret = ttm_bo_handle_move_mem(bo, &mem, false, ctx);
1039
out_unlock:
1040 1041
	if (ret && mem.mm_node)
		ttm_bo_mem_put(bo, &mem);
1042 1043 1044
	return ret;
}

1045 1046 1047 1048
static bool ttm_bo_places_compat(const struct ttm_place *places,
				 unsigned num_placement,
				 struct ttm_mem_reg *mem,
				 uint32_t *new_flags)
1049
{
1050
	unsigned i;
1051

1052 1053
	for (i = 0; i < num_placement; i++) {
		const struct ttm_place *heap = &places[i];
1054

1055
		if (mem->mm_node && (mem->start < heap->fpfn ||
1056
		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1057 1058 1059
			continue;

		*new_flags = heap->flags;
1060
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1061 1062 1063
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
		    (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
		     (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1064
			return true;
1065
	}
1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082
	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;
1083 1084

	return false;
1085
}
1086
EXPORT_SYMBOL(ttm_bo_mem_compat);
1087

1088
int ttm_bo_validate(struct ttm_buffer_object *bo,
1089 1090
		    struct ttm_placement *placement,
		    struct ttm_operation_ctx *ctx)
1091 1092
{
	int ret;
1093
	uint32_t new_flags;
1094

1095
	lockdep_assert_held(&bo->resv->lock.base);
1096 1097 1098
	/*
	 * Check whether we need to move buffer.
	 */
1099
	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1100
		ret = ttm_bo_move_buffer(bo, placement, ctx);
1101
		if (ret)
1102
			return ret;
1103 1104 1105 1106 1107
	} else {
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the compatible memory placement flags to the active flags
		 */
1108
		ttm_flag_masked(&bo->mem.placement, new_flags,
1109
				~TTM_PL_MASK_MEMTYPE);
1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120
	}
	/*
	 * 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;
}
1121
EXPORT_SYMBOL(ttm_bo_validate);
1122

1123 1124 1125 1126 1127 1128
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,
1129
			 struct ttm_operation_ctx *ctx,
1130 1131 1132 1133 1134
			 struct file *persistent_swap_storage,
			 size_t acc_size,
			 struct sg_table *sg,
			 struct reservation_object *resv,
			 void (*destroy) (struct ttm_buffer_object *))
1135
{
1136
	int ret = 0;
1137
	unsigned long num_pages;
1138
	struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1139
	bool locked;
1140 1141 1142

	ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
	if (ret) {
J
Joe Perches 已提交
1143
		pr_err("Out of kernel memory\n");
1144 1145 1146 1147 1148 1149
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
		return -ENOMEM;
	}
1150 1151 1152

	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
	if (num_pages == 0) {
J
Joe Perches 已提交
1153
		pr_err("Illegal buffer object size\n");
1154 1155 1156 1157
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
1158
		ttm_mem_global_free(mem_glob, acc_size);
1159 1160 1161 1162 1163 1164 1165 1166 1167 1168
		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);
1169
	INIT_LIST_HEAD(&bo->io_reserve_lru);
1170
	mutex_init(&bo->wu_mutex);
1171
	bo->bdev = bdev;
1172
	bo->glob = bdev->glob;
1173 1174
	bo->type = type;
	bo->num_pages = num_pages;
1175
	bo->mem.size = num_pages << PAGE_SHIFT;
1176 1177 1178 1179
	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;
1180 1181
	bo->mem.bus.io_reserved_vm = false;
	bo->mem.bus.io_reserved_count = 0;
1182
	bo->moving = NULL;
1183
	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
J
Jan Engelhardt 已提交
1184
	bo->persistent_swap_storage = persistent_swap_storage;
1185
	bo->acc_size = acc_size;
1186
	bo->sg = sg;
1187 1188 1189 1190 1191 1192
	if (resv) {
		bo->resv = resv;
		lockdep_assert_held(&bo->resv->lock.base);
	} else {
		bo->resv = &bo->ttm_resv;
	}
1193
	reservation_object_init(&bo->ttm_resv);
1194
	atomic_inc(&bo->glob->bo_count);
1195
	drm_vma_node_reset(&bo->vma_node);
1196
	bo->priority = 0;
1197 1198 1199 1200 1201

	/*
	 * For ttm_bo_type_device buffers, allocate
	 * address space from the device.
	 */
1202 1203
	if (bo->type == ttm_bo_type_device ||
	    bo->type == ttm_bo_type_sg)
1204 1205
		ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
					 bo->mem.num_pages);
1206

1207 1208 1209 1210 1211 1212 1213
	/* 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);
	}
1214

1215 1216
	if (likely(!ret))
		ret = ttm_bo_validate(bo, placement, ctx);
1217

1218
	if (unlikely(ret)) {
1219 1220 1221
		if (!resv)
			ttm_bo_unreserve(bo);

1222 1223 1224 1225 1226
		ttm_bo_unref(&bo);
		return ret;
	}

	if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1227 1228 1229 1230 1231
		spin_lock(&bo->glob->lru_lock);
		ttm_bo_add_to_lru(bo);
		spin_unlock(&bo->glob->lru_lock);
	}

1232 1233
	return ret;
}
1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248
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 *))
{
1249
	struct ttm_operation_ctx ctx = { interruptible, false };
1250 1251 1252
	int ret;

	ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1253
				   page_alignment, &ctx,
1254 1255 1256 1257 1258 1259 1260 1261 1262 1263
				   persistent_swap_storage, acc_size,
				   sg, resv, destroy);
	if (ret)
		return ret;

	if (!resv)
		ttm_bo_unreserve(bo);

	return 0;
}
1264
EXPORT_SYMBOL(ttm_bo_init);
1265

1266 1267 1268
size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
		       unsigned long bo_size,
		       unsigned struct_size)
1269
{
1270 1271
	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
	size_t size = 0;
1272

1273
	size += ttm_round_pot(struct_size);
F
Felix Kuehling 已提交
1274
	size += ttm_round_pot(npages * sizeof(void *));
1275 1276
	size += ttm_round_pot(sizeof(struct ttm_tt));
	return size;
1277
}
1278 1279 1280 1281 1282 1283 1284 1285 1286 1287
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
1288
	size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1289 1290 1291 1292
	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
	return size;
}
EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1293

1294 1295 1296 1297 1298 1299
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
1300
			struct file *persistent_swap_storage,
1301
			struct ttm_buffer_object **p_bo)
1302 1303
{
	struct ttm_buffer_object *bo;
1304
	size_t acc_size;
1305
	int ret;
1306 1307

	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1308
	if (unlikely(bo == NULL))
1309 1310
		return -ENOMEM;

1311
	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1312
	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1313
			  interruptible, persistent_swap_storage, acc_size,
1314
			  NULL, NULL, NULL);
1315 1316 1317 1318 1319
	if (likely(ret == 0))
		*p_bo = bo;

	return ret;
}
T
Thomas Hellstrom 已提交
1320
EXPORT_SYMBOL(ttm_bo_create);
1321 1322

static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1323
				   unsigned mem_type)
1324
{
1325
	struct ttm_operation_ctx ctx = { false, false };
1326
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1327
	struct ttm_bo_global *glob = bdev->glob;
1328
	struct dma_fence *fence;
1329
	int ret;
1330
	unsigned i;
1331 1332 1333 1334 1335

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

1336
	spin_lock(&glob->lru_lock);
1337 1338 1339
	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
1340
			ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx);
1341
			if (ret)
1342
				return ret;
1343
			spin_lock(&glob->lru_lock);
1344
		}
1345
	}
1346
	spin_unlock(&glob->lru_lock);
1347 1348

	spin_lock(&man->move_lock);
1349
	fence = dma_fence_get(man->move);
1350 1351 1352
	spin_unlock(&man->move_lock);

	if (fence) {
1353 1354
		ret = dma_fence_wait(fence, false);
		dma_fence_put(fence);
1355 1356
		if (ret)
			return ret;
1357 1358
	}

1359 1360 1361 1362 1363
	return 0;
}

int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
{
R
Roel Kluin 已提交
1364
	struct ttm_mem_type_manager *man;
1365 1366 1367
	int ret = -EINVAL;

	if (mem_type >= TTM_NUM_MEM_TYPES) {
J
Joe Perches 已提交
1368
		pr_err("Illegal memory type %d\n", mem_type);
1369 1370
		return ret;
	}
R
Roel Kluin 已提交
1371
	man = &bdev->man[mem_type];
1372 1373

	if (!man->has_type) {
J
Joe Perches 已提交
1374 1375
		pr_err("Trying to take down uninitialized memory manager type %u\n",
		       mem_type);
1376 1377 1378 1379 1380 1381 1382 1383
		return ret;
	}

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

	ret = 0;
	if (mem_type > 0) {
1384 1385 1386 1387 1388
		ret = ttm_bo_force_list_clean(bdev, mem_type);
		if (ret) {
			pr_err("Cleanup eviction failed\n");
			return ret;
		}
1389

1390
		ret = (*man->func->takedown)(man);
1391 1392
	}

1393 1394 1395
	dma_fence_put(man->move);
	man->move = NULL;

1396 1397 1398 1399 1400 1401 1402 1403 1404
	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
1405
		pr_err("Illegal memory manager memory type %u\n", mem_type);
1406 1407 1408 1409
		return -EINVAL;
	}

	if (!man->has_type) {
J
Joe Perches 已提交
1410
		pr_err("Memory type %u has not been initialized\n", mem_type);
1411 1412 1413
		return 0;
	}

1414
	return ttm_bo_force_list_clean(bdev, mem_type);
1415 1416 1417 1418
}
EXPORT_SYMBOL(ttm_bo_evict_mm);

int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1419
			unsigned long p_size)
1420
{
H
Huang Rui 已提交
1421
	int ret;
1422
	struct ttm_mem_type_manager *man;
1423
	unsigned i;
1424

1425
	BUG_ON(type >= TTM_NUM_MEM_TYPES);
1426
	man = &bdev->man[type];
1427
	BUG_ON(man->has_type);
1428 1429 1430
	man->io_reserve_fastpath = true;
	man->use_io_reserve_lru = false;
	mutex_init(&man->io_reserve_mutex);
1431
	spin_lock_init(&man->move_lock);
1432
	INIT_LIST_HEAD(&man->io_reserve_lru);
1433 1434 1435 1436

	ret = bdev->driver->init_mem_type(bdev, type, man);
	if (ret)
		return ret;
1437
	man->bdev = bdev;
1438 1439

	if (type != TTM_PL_SYSTEM) {
1440
		ret = (*man->func->init)(man, p_size);
1441 1442 1443 1444 1445 1446 1447
		if (ret)
			return ret;
	}
	man->has_type = true;
	man->use_type = true;
	man->size = p_size;

1448 1449
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&man->lru[i]);
1450
	man->move = NULL;
1451 1452 1453 1454 1455

	return 0;
}
EXPORT_SYMBOL(ttm_bo_init_mm);

1456 1457 1458 1459 1460 1461 1462 1463 1464 1465
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);
}

1466
void ttm_bo_global_release(struct drm_global_reference *ref)
1467 1468 1469 1470 1471 1472 1473 1474
{
	struct ttm_bo_global *glob = ref->object;

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

1475
int ttm_bo_global_init(struct drm_global_reference *ref)
1476 1477 1478 1479 1480
{
	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;
1481
	unsigned i;
1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492

	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;
	}

1493 1494
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
		INIT_LIST_HEAD(&glob->swap_lru[i]);
1495 1496 1497 1498 1499
	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
1500
		pr_err("Could not register buffer object swapout\n");
1501 1502 1503 1504 1505
		goto out_no_shrink;
	}

	atomic_set(&glob->bo_count, 0);

1506 1507
	ret = kobject_init_and_add(
		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519
	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);


1520 1521 1522 1523 1524
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;
1525
	struct ttm_bo_global *glob = bdev->glob;
1526 1527 1528 1529 1530 1531 1532

	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
1533 1534
				pr_err("DRM memory manager type %d is not clean\n",
				       i);
1535 1536 1537 1538 1539
			}
			man->has_type = false;
		}
	}

1540 1541 1542 1543
	mutex_lock(&glob->device_list_mutex);
	list_del(&bdev->device_list);
	mutex_unlock(&glob->device_list_mutex);

1544
	cancel_delayed_work_sync(&bdev->wq);
1545

1546
	if (ttm_bo_delayed_delete(bdev, true))
1547 1548
		TTM_DEBUG("Delayed destroy list was clean\n");

1549
	spin_lock(&glob->lru_lock);
1550 1551 1552
	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);
1553
	spin_unlock(&glob->lru_lock);
1554

1555
	drm_vma_offset_manager_destroy(&bdev->vma_manager);
1556 1557 1558 1559 1560 1561

	return ret;
}
EXPORT_SYMBOL(ttm_bo_device_release);

int ttm_bo_device_init(struct ttm_bo_device *bdev,
1562 1563
		       struct ttm_bo_global *glob,
		       struct ttm_bo_driver *driver,
1564
		       struct address_space *mapping,
D
Dave Airlie 已提交
1565
		       uint64_t file_page_offset,
D
Dave Airlie 已提交
1566
		       bool need_dma32)
1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577
{
	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.
	 */
1578
	ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1579
	if (unlikely(ret != 0))
1580
		goto out_no_sys;
1581

1582 1583
	drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
				    0x10000000);
1584 1585
	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
	INIT_LIST_HEAD(&bdev->ddestroy);
1586
	bdev->dev_mapping = mapping;
1587
	bdev->glob = glob;
D
Dave Airlie 已提交
1588
	bdev->need_dma32 = need_dma32;
1589 1590 1591
	mutex_lock(&glob->device_list_mutex);
	list_add_tail(&bdev->device_list, &glob->device_list);
	mutex_unlock(&glob->device_list_mutex);
1592 1593

	return 0;
1594
out_no_sys:
1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619
	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;
}

1620
void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1621 1622 1623
{
	struct ttm_bo_device *bdev = bo->bdev;

1624
	drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1625
	ttm_mem_io_free_vm(bo);
1626
}
1627 1628 1629 1630 1631 1632 1633 1634 1635

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);
1636
}
1637 1638


1639
EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1640 1641

int ttm_bo_wait(struct ttm_buffer_object *bo,
1642
		bool interruptible, bool no_wait)
1643
{
C
Christian König 已提交
1644 1645 1646 1647 1648 1649 1650 1651
	long timeout = 15 * HZ;

	if (no_wait) {
		if (reservation_object_test_signaled_rcu(bo->resv, true))
			return 0;
		else
			return -EBUSY;
	}
1652

C
Christian König 已提交
1653 1654
	timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
						      interruptible, timeout);
1655 1656 1657 1658 1659 1660
	if (timeout < 0)
		return timeout;

	if (timeout == 0)
		return -EBUSY;

C
Christian König 已提交
1661
	reservation_object_add_excl_fence(bo->resv, NULL);
1662
	return 0;
1663 1664 1665 1666 1667 1668 1669 1670
}
EXPORT_SYMBOL(ttm_bo_wait);

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

	/*
1671
	 * Using ttm_bo_reserve makes sure the lru lists are updated.
1672 1673
	 */

1674
	ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1675 1676
	if (unlikely(ret != 0))
		return ret;
1677
	ret = ttm_bo_wait(bo, true, no_wait);
1678 1679 1680 1681 1682
	if (likely(ret == 0))
		atomic_inc(&bo->cpu_writers);
	ttm_bo_unreserve(bo);
	return ret;
}
1683
EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1684 1685 1686

void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
{
1687
	atomic_dec(&bo->cpu_writers);
1688
}
1689
EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1690 1691 1692 1693 1694 1695 1696 1697

/**
 * 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)
{
1698 1699
	struct ttm_bo_global *glob =
	    container_of(shrink, struct ttm_bo_global, shrink);
1700 1701
	struct ttm_buffer_object *bo;
	int ret = -EBUSY;
1702
	unsigned i;
1703

1704
	spin_lock(&glob->lru_lock);
1705 1706
	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
		list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1707
			ret = reservation_object_trylock(bo->resv) ? 0 : -EBUSY;
1708 1709 1710
			if (!ret)
				break;
		}
1711 1712 1713
		if (!ret)
			break;
	}
1714

1715 1716 1717 1718
	if (ret) {
		spin_unlock(&glob->lru_lock);
		return ret;
	}
1719

1720
	kref_get(&bo->list_kref);
1721

1722
	if (!list_empty(&bo->ddestroy)) {
1723
		ret = ttm_bo_cleanup_refs(bo, false, false, true);
1724 1725
		kref_put(&bo->list_kref, ttm_bo_release_list);
		return ret;
1726 1727
	}

1728
	ttm_bo_del_from_lru(bo);
1729
	spin_unlock(&glob->lru_lock);
1730 1731

	/**
1732
	 * Move to system cached
1733 1734
	 */

1735 1736
	if (bo->mem.mem_type != TTM_PL_SYSTEM ||
	    bo->ttm->caching_state != tt_cached) {
1737
		struct ttm_operation_ctx ctx = { false, false };
1738 1739 1740 1741 1742 1743 1744
		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;

1745
		ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, &ctx);
1746 1747 1748 1749
		if (unlikely(ret != 0))
			goto out;
	}

1750 1751 1752 1753 1754 1755 1756 1757
	/**
	 * Make sure BO is idle.
	 */

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

1758 1759 1760 1761 1762 1763 1764
	ttm_bo_unmap_virtual(bo);

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

1765 1766 1767
	if (bo->bdev->driver->swap_notify)
		bo->bdev->driver->swap_notify(bo);

J
Jan Engelhardt 已提交
1768
	ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1769 1770 1771 1772 1773 1774 1775 1776
out:

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

1777
	reservation_object_unlock(bo->resv);
1778 1779 1780 1781 1782 1783
	kref_put(&bo->list_kref, ttm_bo_release_list);
	return ret;
}

void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
{
1784
	while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1785 1786
		;
}
1787
EXPORT_SYMBOL(ttm_bo_swapout_all);
1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810

/**
 * 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;
1811 1812 1813
	ret = reservation_object_lock_interruptible(bo->resv, NULL);
	if (ret == -EINTR)
		ret = -ERESTARTSYS;
1814 1815
	if (unlikely(ret != 0))
		goto out_unlock;
1816
	reservation_object_unlock(bo->resv);
1817 1818 1819 1820 1821

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