ttm_bo.c 41.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
/**************************************************************************
 *
 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sub license, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
 * USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 **************************************************************************/
/*
 * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
 */

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

33 34 35
#include <drm/ttm/ttm_module.h>
#include <drm/ttm/ttm_bo_driver.h>
#include <drm/ttm/ttm_placement.h>
36 37 38 39 40 41
#include <linux/jiffies.h>
#include <linux/slab.h>
#include <linux/sched.h>
#include <linux/mm.h>
#include <linux/file.h>
#include <linux/module.h>
A
Arun Sharma 已提交
42
#include <linux/atomic.h>
43
#include <linux/reservation.h>
44 45 46 47 48 49

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

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

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

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

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

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

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

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

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

104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
static ssize_t ttm_bo_global_show(struct kobject *kobj,
				  struct attribute *attr,
				  char *buffer)
{
	struct ttm_bo_global *glob =
		container_of(kobj, struct ttm_bo_global, kobj);

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

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

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

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

130 131 132 133 134 135 136 137 138 139 140

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

static void ttm_bo_release_list(struct kref *list_kref)
{
	struct ttm_buffer_object *bo =
	    container_of(list_kref, struct ttm_buffer_object, list_kref);
	struct ttm_bo_device *bdev = bo->bdev;
141
	size_t acc_size = bo->acc_size;
142 143 144 145 146 147 148 149 150 151

	BUG_ON(atomic_read(&bo->list_kref.refcount));
	BUG_ON(atomic_read(&bo->kref.refcount));
	BUG_ON(atomic_read(&bo->cpu_writers));
	BUG_ON(bo->mem.mm_node != NULL);
	BUG_ON(!list_empty(&bo->lru));
	BUG_ON(!list_empty(&bo->ddestroy));

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

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

169
	lockdep_assert_held(&bo->resv->lock.base);
170 171 172 173 174 175 176 177 178 179

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

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

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

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

187
int ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212
{
	int put_count = 0;

	if (!list_empty(&bo->swap)) {
		list_del_init(&bo->swap);
		++put_count;
	}
	if (!list_empty(&bo->lru)) {
		list_del_init(&bo->lru);
		++put_count;
	}

	/*
	 * TODO: Add a driver hook to delete from
	 * driver-specific LRU's here.
	 */

	return put_count;
}

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

213 214 215
void ttm_bo_list_ref_sub(struct ttm_buffer_object *bo, int count,
			 bool never_free)
{
216 217
	kref_sub(&bo->list_kref, count,
		 (never_free) ? ttm_bo_ref_bug : ttm_bo_release_list);
218 219
}

220
void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
221
{
222
	int put_count;
223

224 225 226 227
	spin_lock(&bo->glob->lru_lock);
	put_count = ttm_bo_del_from_lru(bo);
	spin_unlock(&bo->glob->lru_lock);
	ttm_bo_list_ref_sub(bo, put_count, true);
228
}
229
EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
230

231 232 233 234 235 236
/*
 * 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;
237
	struct ttm_bo_global *glob = bo->glob;
238 239 240 241 242 243
	int ret = 0;
	uint32_t page_flags = 0;

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

D
Dave Airlie 已提交
244 245 246
	if (bdev->need_dma32)
		page_flags |= TTM_PAGE_FLAG_DMA32;

247 248 249 250 251
	switch (bo->type) {
	case ttm_bo_type_device:
		if (zero_alloc)
			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
	case ttm_bo_type_kernel:
252 253
		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
						      page_flags, glob->dummy_read_page);
254 255 256
		if (unlikely(bo->ttm == NULL))
			ret = -ENOMEM;
		break;
257 258 259 260 261 262 263 264 265 266
	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;
267
	default:
J
Joe Perches 已提交
268
		pr_err("Illegal buffer object type\n");
269 270 271 272 273 274 275 276 277
		ret = -EINVAL;
		break;
	}

	return ret;
}

static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
				  struct ttm_mem_reg *mem,
278
				  bool evict, bool interruptible,
279
				  bool no_wait_gpu)
280 281 282 283 284 285 286 287 288
{
	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 ||
289 290 291 292 293 294 295
	    ((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);
	}
296 297 298 299 300

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

301 302
	if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
		if (bo->ttm == NULL) {
303 304
			bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
			ret = ttm_bo_add_ttm(bo, zero);
305 306 307
			if (ret)
				goto out_err;
		}
308 309 310

		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
		if (ret)
311
			goto out_err;
312 313 314 315 316 317 318 319

		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) {
320 321
			if (bdev->driver->move_notify)
				bdev->driver->move_notify(bo, mem);
322
			bo->mem = *mem;
323 324 325 326 327
			mem->mm_node = NULL;
			goto moved;
		}
	}

328 329 330
	if (bdev->driver->move_notify)
		bdev->driver->move_notify(bo, mem);

331 332
	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
333
		ret = ttm_bo_move_ttm(bo, evict, no_wait_gpu, mem);
334 335
	else if (bdev->driver->move)
		ret = bdev->driver->move(bo, evict, interruptible,
336
					 no_wait_gpu, mem);
337
	else
338
		ret = ttm_bo_move_memcpy(bo, evict, no_wait_gpu, mem);
339

340 341 342 343 344 345 346
	if (ret) {
		if (bdev->driver->move_notify) {
			struct ttm_mem_reg tmp_mem = *mem;
			*mem = bo->mem;
			bo->mem = tmp_mem;
			bdev->driver->move_notify(bo, mem);
			bo->mem = *mem;
347
			*mem = tmp_mem;
348
		}
349

350 351
		goto out_err;
	}
352

353 354
moved:
	if (bo->evicted) {
355 356 357 358 359
		if (bdev->driver->invalidate_caches) {
			ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
			if (ret)
				pr_err("Can not flush read caches\n");
		}
360 361 362 363
		bo->evicted = false;
	}

	if (bo->mem.mm_node) {
364
		bo->offset = (bo->mem.start << PAGE_SHIFT) +
365 366
		    bdev->man[bo->mem.mem_type].gpu_offset;
		bo->cur_placement = bo->mem.placement;
367 368
	} else
		bo->offset = 0;
369 370 371 372 373 374 375 376 377 378 379 380 381 382

	return 0;

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

	return ret;
}

383
/**
384
 * Call bo::reserved.
385
 * Will release GPU memory type usage on destruction.
386 387 388
 * This is the place to put in driver specific hooks to release
 * driver private resources.
 * Will release the bo::reserved lock.
389 390 391 392
 */

static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
{
393 394 395
	if (bo->bdev->driver->move_notify)
		bo->bdev->driver->move_notify(bo, NULL);

396 397 398 399 400
	if (bo->ttm) {
		ttm_tt_unbind(bo->ttm);
		ttm_tt_destroy(bo->ttm);
		bo->ttm = NULL;
	}
401
	ttm_bo_mem_put(bo, &bo->mem);
402

403
	ww_mutex_unlock (&bo->resv->lock);
404 405
}

406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425
static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
{
	struct reservation_object_list *fobj;
	struct fence *fence;
	int i;

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

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

		if (!fence->ops->signaled)
			fence_enable_sw_signaling(fence);
	}
}

426
static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
427 428
{
	struct ttm_bo_device *bdev = bo->bdev;
429
	struct ttm_bo_global *glob = bo->glob;
430
	int put_count;
431 432
	int ret;

433
	spin_lock(&glob->lru_lock);
434
	ret = __ttm_bo_reserve(bo, false, true, false, NULL);
435

M
Maarten Lankhorst 已提交
436
	if (!ret) {
437
		if (!ttm_bo_wait(bo, false, false, true)) {
M
Maarten Lankhorst 已提交
438
			put_count = ttm_bo_del_from_lru(bo);
439

M
Maarten Lankhorst 已提交
440 441
			spin_unlock(&glob->lru_lock);
			ttm_bo_cleanup_memtype_use(bo);
442

M
Maarten Lankhorst 已提交
443
			ttm_bo_list_ref_sub(bo, put_count, true);
444

M
Maarten Lankhorst 已提交
445
			return;
446 447
		} else
			ttm_bo_flush_all_fences(bo);
448 449 450 451 452 453 454 455 456 457 458

		/*
		 * 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);
		}

459
		__ttm_bo_unreserve(bo);
460
	}
461 462 463 464 465 466 467 468 469 470

	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);
}

/**
471
 * function ttm_bo_cleanup_refs_and_unlock
472 473 474
 * If bo idle, remove from delayed- and lru lists, and unref.
 * If not idle, do nothing.
 *
475 476 477
 * Must be called with lru_lock and reservation held, this function
 * will drop both before returning.
 *
478 479 480 481
 * @interruptible         Any sleeps should occur interruptibly.
 * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
 */

482 483 484
static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
					  bool interruptible,
					  bool no_wait_gpu)
485 486 487
{
	struct ttm_bo_global *glob = bo->glob;
	int put_count;
488
	int ret;
489

490
	ret = ttm_bo_wait(bo, false, false, true);
491

492
	if (ret && !no_wait_gpu) {
M
Maarten Lankhorst 已提交
493 494 495 496 497 498 499 500 501 502 503 504 505
		long lret;
		ww_mutex_unlock(&bo->resv->lock);
		spin_unlock(&glob->lru_lock);

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

		if (lret < 0)
			return lret;
		else if (lret == 0)
			return -EBUSY;
506

507
		spin_lock(&glob->lru_lock);
508
		ret = __ttm_bo_reserve(bo, false, true, false, NULL);
509

510 511 512 513 514 515 516 517 518 519 520 521
		/*
		 * We raced, and lost, someone else holds the reservation now,
		 * and is probably busy in ttm_bo_cleanup_memtype_use.
		 *
		 * Even if it's not the case, because we finished waiting any
		 * delayed destruction would succeed, so just return success
		 * here.
		 */
		if (ret) {
			spin_unlock(&glob->lru_lock);
			return 0;
		}
522 523 524 525 526 527 528 529

		/*
		 * remove sync_obj with ttm_bo_wait, the wait should be
		 * finished, and no new wait object should have been added.
		 */
		ret = ttm_bo_wait(bo, false, false, true);
		WARN_ON(ret);
	}
530

531
	if (ret || unlikely(list_empty(&bo->ddestroy))) {
532
		__ttm_bo_unreserve(bo);
533
		spin_unlock(&glob->lru_lock);
534
		return ret;
535 536
	}

537 538 539 540 541 542 543
	put_count = ttm_bo_del_from_lru(bo);
	list_del_init(&bo->ddestroy);
	++put_count;

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

544
	ttm_bo_list_ref_sub(bo, put_count, true);
545 546

	return 0;
547 548 549 550 551 552 553 554 555
}

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

static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
{
556
	struct ttm_bo_global *glob = bdev->glob;
557 558
	struct ttm_buffer_object *entry = NULL;
	int ret = 0;
559

560
	spin_lock(&glob->lru_lock);
561 562 563 564 565 566 567 568 569 570 571 572 573
	if (list_empty(&bdev->ddestroy))
		goto out_unlock;

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

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

		if (entry->ddestroy.next != &bdev->ddestroy) {
			nentry = list_first_entry(&entry->ddestroy,
				struct ttm_buffer_object, ddestroy);
574 575 576
			kref_get(&nentry->list_kref);
		}

577
		ret = __ttm_bo_reserve(entry, false, true, false, NULL);
578 579
		if (remove_all && ret) {
			spin_unlock(&glob->lru_lock);
580
			ret = __ttm_bo_reserve(entry, false, false,
581
					       false, NULL);
582 583 584
			spin_lock(&glob->lru_lock);
		}

585 586 587 588 589 590
		if (!ret)
			ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
							     !remove_all);
		else
			spin_unlock(&glob->lru_lock);

591
		kref_put(&entry->list_kref, ttm_bo_release_list);
592 593 594 595
		entry = nentry;

		if (ret || !entry)
			goto out;
596

597
		spin_lock(&glob->lru_lock);
598
		if (list_empty(&entry->ddestroy))
599 600 601
			break;
	}

602 603 604 605 606
out_unlock:
	spin_unlock(&glob->lru_lock);
out:
	if (entry)
		kref_put(&entry->list_kref, ttm_bo_release_list);
607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625
	return ret;
}

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

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

static void ttm_bo_release(struct kref *kref)
{
	struct ttm_buffer_object *bo =
	    container_of(kref, struct ttm_buffer_object, kref);
	struct ttm_bo_device *bdev = bo->bdev;
626
	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
627

628
	drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
629 630 631
	ttm_mem_io_lock(man, false);
	ttm_mem_io_free_vm(bo);
	ttm_mem_io_unlock(man);
632
	ttm_bo_cleanup_refs_or_queue(bo);
633 634 635 636 637 638 639 640 641 642 643 644
	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);

645 646 647 648 649 650 651 652 653 654 655 656 657 658
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);

659
static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
660
			bool no_wait_gpu)
661 662 663
{
	struct ttm_bo_device *bdev = bo->bdev;
	struct ttm_mem_reg evict_mem;
664 665
	struct ttm_placement placement;
	int ret = 0;
666

667
	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
668

669
	if (unlikely(ret != 0)) {
670
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
671
			pr_err("Failed to expire sync object before buffer eviction\n");
672
		}
673 674 675
		goto out;
	}

676
	lockdep_assert_held(&bo->resv->lock.base);
677 678 679

	evict_mem = bo->mem;
	evict_mem.mm_node = NULL;
680 681
	evict_mem.bus.io_reserved_vm = false;
	evict_mem.bus.io_reserved_count = 0;
682

683 684
	placement.num_placement = 0;
	placement.num_busy_placement = 0;
685 686
	bdev->driver->evict_flags(bo, &placement);
	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
687
				no_wait_gpu);
688
	if (ret) {
689
		if (ret != -ERESTARTSYS) {
J
Joe Perches 已提交
690 691
			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
			       bo);
692 693
			ttm_bo_mem_space_debug(bo, &placement);
		}
694 695 696 697
		goto out;
	}

	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
698
				     no_wait_gpu);
699
	if (ret) {
700
		if (ret != -ERESTARTSYS)
J
Joe Perches 已提交
701
			pr_err("Buffer eviction failed\n");
702
		ttm_bo_mem_put(bo, &evict_mem);
703 704
		goto out;
	}
705 706 707 708 709 710 711
	bo->evicted = true;
out:
	return ret;
}

static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
				uint32_t mem_type,
712
				bool interruptible,
713
				bool no_wait_gpu)
714 715 716 717
{
	struct ttm_bo_global *glob = bdev->glob;
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	struct ttm_buffer_object *bo;
718
	int ret = -EBUSY, put_count;
719

720
	spin_lock(&glob->lru_lock);
721
	list_for_each_entry(bo, &man->lru, lru) {
722
		ret = __ttm_bo_reserve(bo, false, true, false, NULL);
723 724 725 726 727
		if (!ret)
			break;
	}

	if (ret) {
728
		spin_unlock(&glob->lru_lock);
729
		return ret;
730 731
	}

732
	kref_get(&bo->list_kref);
733

734
	if (!list_empty(&bo->ddestroy)) {
735 736
		ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
						     no_wait_gpu);
737
		kref_put(&bo->list_kref, ttm_bo_release_list);
738
		return ret;
739 740
	}

741
	put_count = ttm_bo_del_from_lru(bo);
742
	spin_unlock(&glob->lru_lock);
743 744 745

	BUG_ON(ret != 0);

746
	ttm_bo_list_ref_sub(bo, put_count, true);
747

748
	ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
749
	ttm_bo_unreserve(bo);
750

751
	kref_put(&bo->list_kref, ttm_bo_release_list);
752 753 754
	return ret;
}

755 756
void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
{
757
	struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
758

759 760
	if (mem->mm_node)
		(*man->func->put_node)(man, mem);
761 762 763
}
EXPORT_SYMBOL(ttm_bo_mem_put);

764 765 766 767
/**
 * 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.
 */
768 769
static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
					uint32_t mem_type,
770
					const struct ttm_place *place,
771
					struct ttm_mem_reg *mem,
772 773
					bool interruptible,
					bool no_wait_gpu)
774
{
775
	struct ttm_bo_device *bdev = bo->bdev;
776 777 778 779
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
	int ret;

	do {
780
		ret = (*man->func->get_node)(man, bo, place, mem);
781 782
		if (unlikely(ret != 0))
			return ret;
783
		if (mem->mm_node)
784
			break;
785 786
		ret = ttm_mem_evict_first(bdev, mem_type,
					  interruptible, no_wait_gpu);
787 788 789
		if (unlikely(ret != 0))
			return ret;
	} while (1);
790
	if (mem->mm_node == NULL)
791 792 793 794 795
		return -ENOMEM;
	mem->mem_type = mem_type;
	return 0;
}

796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820
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;
}

821 822
static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
				 uint32_t mem_type,
823
				 const struct ttm_place *place,
824
				 uint32_t *masked_placement)
825 826 827
{
	uint32_t cur_flags = ttm_bo_type_flags(mem_type);

828
	if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
829 830
		return false;

831
	if ((place->flags & man->available_caching) == 0)
832 833
		return false;

834
	cur_flags |= (place->flags & man->available_caching);
835 836

	*masked_placement = cur_flags;
837 838 839 840 841 842 843 844 845 846 847 848
	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,
849 850
			struct ttm_placement *placement,
			struct ttm_mem_reg *mem,
851
			bool interruptible,
852
			bool no_wait_gpu)
853 854 855 856 857 858 859
{
	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;
860
	bool has_erestartsys = false;
861
	int i, ret;
862 863

	mem->mm_node = NULL;
864
	for (i = 0; i < placement->num_placement; ++i) {
865 866 867
		const struct ttm_place *place = &placement->placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
868 869
		if (ret)
			return ret;
870 871
		man = &bdev->man[mem_type];

872
		type_ok = ttm_bo_mt_compatible(man, mem_type, place,
873
						&cur_flags);
874 875 876 877

		if (!type_ok)
			continue;

878 879
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
880 881 882 883
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
884
		ttm_flag_masked(&cur_flags, place->flags,
885
				~TTM_PL_MASK_MEMTYPE);
886

887 888 889 890 891
		if (mem_type == TTM_PL_SYSTEM)
			break;

		if (man->has_type && man->use_type) {
			type_found = true;
892
			ret = (*man->func->get_node)(man, bo, place, mem);
893 894
			if (unlikely(ret))
				return ret;
895
		}
896
		if (mem->mm_node)
897 898 899
			break;
	}

900
	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
901 902 903 904 905 906 907 908
		mem->mem_type = mem_type;
		mem->placement = cur_flags;
		return 0;
	}

	if (!type_found)
		return -EINVAL;

909
	for (i = 0; i < placement->num_busy_placement; ++i) {
910 911 912
		const struct ttm_place *place = &placement->busy_placement[i];

		ret = ttm_mem_type_from_place(place, &mem_type);
913 914
		if (ret)
			return ret;
915 916 917
		man = &bdev->man[mem_type];
		if (!man->has_type)
			continue;
918
		if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
919 920
			continue;

921 922
		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
						  cur_flags);
923 924 925 926
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the memory placement flags to the current flags
		 */
927
		ttm_flag_masked(&cur_flags, place->flags,
928
				~TTM_PL_MASK_MEMTYPE);
929

930 931 932 933 934 935 936
		if (mem_type == TTM_PL_SYSTEM) {
			mem->mem_type = mem_type;
			mem->placement = cur_flags;
			mem->mm_node = NULL;
			return 0;
		}

937
		ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
938
						interruptible, no_wait_gpu);
939 940 941 942
		if (ret == 0 && mem->mm_node) {
			mem->placement = cur_flags;
			return 0;
		}
943 944
		if (ret == -ERESTARTSYS)
			has_erestartsys = true;
945
	}
946
	ret = (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
947 948 949 950
	return ret;
}
EXPORT_SYMBOL(ttm_bo_mem_space);

951
static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
952
			struct ttm_placement *placement,
953
			bool interruptible,
954
			bool no_wait_gpu)
955 956 957 958
{
	int ret = 0;
	struct ttm_mem_reg mem;

959
	lockdep_assert_held(&bo->resv->lock.base);
960 961 962 963 964 965

	/*
	 * FIXME: It's possible to pipeline buffer moves.
	 * Have the driver move function wait for idle when necessary,
	 * instead of doing it here.
	 */
966
	ret = ttm_bo_wait(bo, false, interruptible, no_wait_gpu);
967 968 969 970 971
	if (ret)
		return ret;
	mem.num_pages = bo->num_pages;
	mem.size = mem.num_pages << PAGE_SHIFT;
	mem.page_alignment = bo->mem.page_alignment;
972 973
	mem.bus.io_reserved_vm = false;
	mem.bus.io_reserved_count = 0;
974 975 976
	/*
	 * Determine where to move the buffer.
	 */
977 978
	ret = ttm_bo_mem_space(bo, placement, &mem,
			       interruptible, no_wait_gpu);
979 980
	if (ret)
		goto out_unlock;
981 982
	ret = ttm_bo_handle_move_mem(bo, &mem, false,
				     interruptible, no_wait_gpu);
983
out_unlock:
984 985
	if (ret && mem.mm_node)
		ttm_bo_mem_put(bo, &mem);
986 987 988
	return ret;
}

989 990 991
static bool ttm_bo_mem_compat(struct ttm_placement *placement,
			      struct ttm_mem_reg *mem,
			      uint32_t *new_flags)
992
{
993
	int i;
994

995
	for (i = 0; i < placement->num_placement; i++) {
996 997 998 999 1000 1001 1002
		const struct ttm_place *heap = &placement->placement[i];
		if (mem->mm_node && heap->lpfn != 0 &&
		    (mem->start < heap->fpfn ||
		     mem->start + mem->num_pages > heap->lpfn))
			continue;

		*new_flags = heap->flags;
1003 1004 1005 1006 1007 1008
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM))
			return true;
	}

	for (i = 0; i < placement->num_busy_placement; i++) {
1009 1010 1011 1012 1013 1014 1015
		const struct ttm_place *heap = &placement->busy_placement[i];
		if (mem->mm_node && heap->lpfn != 0 &&
		    (mem->start < heap->fpfn ||
		     mem->start + mem->num_pages > heap->lpfn))
			continue;

		*new_flags = heap->flags;
1016 1017 1018
		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
		    (*new_flags & mem->placement & TTM_PL_MASK_MEM))
			return true;
1019
	}
1020 1021

	return false;
1022 1023
}

1024 1025
int ttm_bo_validate(struct ttm_buffer_object *bo,
			struct ttm_placement *placement,
1026
			bool interruptible,
1027
			bool no_wait_gpu)
1028 1029
{
	int ret;
1030
	uint32_t new_flags;
1031

1032
	lockdep_assert_held(&bo->resv->lock.base);
1033 1034 1035
	/*
	 * Check whether we need to move buffer.
	 */
1036
	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1037 1038
		ret = ttm_bo_move_buffer(bo, placement, interruptible,
					 no_wait_gpu);
1039
		if (ret)
1040
			return ret;
1041 1042 1043 1044 1045
	} else {
		/*
		 * Use the access and other non-mapping-related flag bits from
		 * the compatible memory placement flags to the active flags
		 */
1046
		ttm_flag_masked(&bo->mem.placement, new_flags,
1047
				~TTM_PL_MASK_MEMTYPE);
1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
	}
	/*
	 * 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;
}
1059
EXPORT_SYMBOL(ttm_bo_validate);
1060

1061 1062 1063 1064 1065 1066 1067
int ttm_bo_init(struct ttm_bo_device *bdev,
		struct ttm_buffer_object *bo,
		unsigned long size,
		enum ttm_bo_type type,
		struct ttm_placement *placement,
		uint32_t page_alignment,
		bool interruptible,
J
Jan Engelhardt 已提交
1068
		struct file *persistent_swap_storage,
1069
		size_t acc_size,
1070
		struct sg_table *sg,
1071
		struct reservation_object *resv,
1072
		void (*destroy) (struct ttm_buffer_object *))
1073
{
1074
	int ret = 0;
1075
	unsigned long num_pages;
1076
	struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1077
	bool locked;
1078 1079 1080

	ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
	if (ret) {
J
Joe Perches 已提交
1081
		pr_err("Out of kernel memory\n");
1082 1083 1084 1085 1086 1087
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
		return -ENOMEM;
	}
1088 1089 1090

	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
	if (num_pages == 0) {
J
Joe Perches 已提交
1091
		pr_err("Illegal buffer object size\n");
1092 1093 1094 1095
		if (destroy)
			(*destroy)(bo);
		else
			kfree(bo);
1096
		ttm_mem_global_free(mem_glob, acc_size);
1097 1098 1099 1100 1101 1102 1103 1104 1105 1106
		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);
1107
	INIT_LIST_HEAD(&bo->io_reserve_lru);
1108
	mutex_init(&bo->wu_mutex);
1109
	bo->bdev = bdev;
1110
	bo->glob = bdev->glob;
1111 1112
	bo->type = type;
	bo->num_pages = num_pages;
1113
	bo->mem.size = num_pages << PAGE_SHIFT;
1114 1115 1116 1117
	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;
1118 1119
	bo->mem.bus.io_reserved_vm = false;
	bo->mem.bus.io_reserved_count = 0;
1120 1121
	bo->priv_flags = 0;
	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
J
Jan Engelhardt 已提交
1122
	bo->persistent_swap_storage = persistent_swap_storage;
1123
	bo->acc_size = acc_size;
1124
	bo->sg = sg;
1125 1126 1127 1128 1129 1130 1131
	if (resv) {
		bo->resv = resv;
		lockdep_assert_held(&bo->resv->lock.base);
	} else {
		bo->resv = &bo->ttm_resv;
		reservation_object_init(&bo->ttm_resv);
	}
1132
	atomic_inc(&bo->glob->bo_count);
1133
	drm_vma_node_reset(&bo->vma_node);
1134 1135 1136 1137 1138

	/*
	 * For ttm_bo_type_device buffers, allocate
	 * address space from the device.
	 */
1139 1140
	if (bo->type == ttm_bo_type_device ||
	    bo->type == ttm_bo_type_sg)
1141 1142
		ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
					 bo->mem.num_pages);
1143

1144 1145 1146 1147 1148 1149 1150
	/* 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);
	}
1151

1152 1153
	if (likely(!ret))
		ret = ttm_bo_validate(bo, placement, interruptible, false);
1154

1155 1156
	if (!resv)
		ttm_bo_unreserve(bo);
1157 1158 1159

	if (unlikely(ret))
		ttm_bo_unref(&bo);
1160 1161 1162

	return ret;
}
1163
EXPORT_SYMBOL(ttm_bo_init);
1164

1165 1166 1167
size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
		       unsigned long bo_size,
		       unsigned struct_size)
1168
{
1169 1170
	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
	size_t size = 0;
1171

1172 1173 1174 1175
	size += ttm_round_pot(struct_size);
	size += PAGE_ALIGN(npages * sizeof(void *));
	size += ttm_round_pot(sizeof(struct ttm_tt));
	return size;
1176
}
1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192
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);
	size += PAGE_ALIGN(npages * sizeof(void *));
	size += PAGE_ALIGN(npages * sizeof(dma_addr_t));
	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
	return size;
}
EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1193

1194 1195 1196 1197 1198 1199
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
1200
			struct file *persistent_swap_storage,
1201
			struct ttm_buffer_object **p_bo)
1202 1203
{
	struct ttm_buffer_object *bo;
1204
	size_t acc_size;
1205
	int ret;
1206 1207

	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1208
	if (unlikely(bo == NULL))
1209 1210
		return -ENOMEM;

1211
	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1212
	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1213
			  interruptible, persistent_swap_storage, acc_size,
1214
			  NULL, NULL, NULL);
1215 1216 1217 1218 1219
	if (likely(ret == 0))
		*p_bo = bo;

	return ret;
}
T
Thomas Hellstrom 已提交
1220
EXPORT_SYMBOL(ttm_bo_create);
1221 1222

static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1223
					unsigned mem_type, bool allow_errors)
1224
{
1225
	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1226
	struct ttm_bo_global *glob = bdev->glob;
1227 1228 1229 1230 1231 1232
	int ret;

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

1233
	spin_lock(&glob->lru_lock);
1234
	while (!list_empty(&man->lru)) {
1235
		spin_unlock(&glob->lru_lock);
1236
		ret = ttm_mem_evict_first(bdev, mem_type, false, false);
1237 1238 1239 1240
		if (ret) {
			if (allow_errors) {
				return ret;
			} else {
J
Joe Perches 已提交
1241
				pr_err("Cleanup eviction failed\n");
1242 1243
			}
		}
1244
		spin_lock(&glob->lru_lock);
1245
	}
1246
	spin_unlock(&glob->lru_lock);
1247 1248 1249 1250 1251
	return 0;
}

int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
{
R
Roel Kluin 已提交
1252
	struct ttm_mem_type_manager *man;
1253 1254 1255
	int ret = -EINVAL;

	if (mem_type >= TTM_NUM_MEM_TYPES) {
J
Joe Perches 已提交
1256
		pr_err("Illegal memory type %d\n", mem_type);
1257 1258
		return ret;
	}
R
Roel Kluin 已提交
1259
	man = &bdev->man[mem_type];
1260 1261

	if (!man->has_type) {
J
Joe Perches 已提交
1262 1263
		pr_err("Trying to take down uninitialized memory manager type %u\n",
		       mem_type);
1264 1265 1266 1267 1268 1269 1270 1271
		return ret;
	}

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

	ret = 0;
	if (mem_type > 0) {
1272
		ttm_bo_force_list_clean(bdev, mem_type, false);
1273

1274
		ret = (*man->func->takedown)(man);
1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285
	}

	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
1286
		pr_err("Illegal memory manager memory type %u\n", mem_type);
1287 1288 1289 1290
		return -EINVAL;
	}

	if (!man->has_type) {
J
Joe Perches 已提交
1291
		pr_err("Memory type %u has not been initialized\n", mem_type);
1292 1293 1294
		return 0;
	}

1295
	return ttm_bo_force_list_clean(bdev, mem_type, true);
1296 1297 1298 1299
}
EXPORT_SYMBOL(ttm_bo_evict_mm);

int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1300
			unsigned long p_size)
1301 1302 1303 1304
{
	int ret = -EINVAL;
	struct ttm_mem_type_manager *man;

1305
	BUG_ON(type >= TTM_NUM_MEM_TYPES);
1306
	man = &bdev->man[type];
1307
	BUG_ON(man->has_type);
1308 1309 1310 1311
	man->io_reserve_fastpath = true;
	man->use_io_reserve_lru = false;
	mutex_init(&man->io_reserve_mutex);
	INIT_LIST_HEAD(&man->io_reserve_lru);
1312 1313 1314 1315

	ret = bdev->driver->init_mem_type(bdev, type, man);
	if (ret)
		return ret;
1316
	man->bdev = bdev;
1317 1318 1319

	ret = 0;
	if (type != TTM_PL_SYSTEM) {
1320
		ret = (*man->func->init)(man, p_size);
1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333
		if (ret)
			return ret;
	}
	man->has_type = true;
	man->use_type = true;
	man->size = p_size;

	INIT_LIST_HEAD(&man->lru);

	return 0;
}
EXPORT_SYMBOL(ttm_bo_init_mm);

1334 1335 1336 1337 1338 1339 1340 1341 1342 1343
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);
}

1344
void ttm_bo_global_release(struct drm_global_reference *ref)
1345 1346 1347 1348 1349 1350 1351 1352
{
	struct ttm_bo_global *glob = ref->object;

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

1353
int ttm_bo_global_init(struct drm_global_reference *ref)
1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375
{
	struct ttm_bo_global_ref *bo_ref =
		container_of(ref, struct ttm_bo_global_ref, ref);
	struct ttm_bo_global *glob = ref->object;
	int ret;

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

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

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

	ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
	ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
	if (unlikely(ret != 0)) {
J
Joe Perches 已提交
1376
		pr_err("Could not register buffer object swapout\n");
1377 1378 1379 1380 1381
		goto out_no_shrink;
	}

	atomic_set(&glob->bo_count, 0);

1382 1383
	ret = kobject_init_and_add(
		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395
	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);


1396 1397 1398 1399 1400
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;
1401
	struct ttm_bo_global *glob = bdev->glob;
1402 1403 1404 1405 1406 1407 1408

	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
1409 1410
				pr_err("DRM memory manager type %d is not clean\n",
				       i);
1411 1412 1413 1414 1415
			}
			man->has_type = false;
		}
	}

1416 1417 1418 1419
	mutex_lock(&glob->device_list_mutex);
	list_del(&bdev->device_list);
	mutex_unlock(&glob->device_list_mutex);

1420
	cancel_delayed_work_sync(&bdev->wq);
1421 1422 1423 1424

	while (ttm_bo_delayed_delete(bdev, true))
		;

1425
	spin_lock(&glob->lru_lock);
1426 1427 1428 1429 1430
	if (list_empty(&bdev->ddestroy))
		TTM_DEBUG("Delayed destroy list was clean\n");

	if (list_empty(&bdev->man[0].lru))
		TTM_DEBUG("Swap list was clean\n");
1431
	spin_unlock(&glob->lru_lock);
1432

1433
	drm_vma_offset_manager_destroy(&bdev->vma_manager);
1434 1435 1436 1437 1438 1439

	return ret;
}
EXPORT_SYMBOL(ttm_bo_device_release);

int ttm_bo_device_init(struct ttm_bo_device *bdev,
1440 1441
		       struct ttm_bo_global *glob,
		       struct ttm_bo_driver *driver,
1442
		       struct address_space *mapping,
D
Dave Airlie 已提交
1443
		       uint64_t file_page_offset,
D
Dave Airlie 已提交
1444
		       bool need_dma32)
1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455
{
	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.
	 */
1456
	ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1457
	if (unlikely(ret != 0))
1458
		goto out_no_sys;
1459

1460 1461
	drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
				    0x10000000);
1462 1463
	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
	INIT_LIST_HEAD(&bdev->ddestroy);
1464
	bdev->dev_mapping = mapping;
1465
	bdev->glob = glob;
D
Dave Airlie 已提交
1466
	bdev->need_dma32 = need_dma32;
1467
	bdev->val_seq = 0;
1468 1469 1470
	mutex_lock(&glob->device_list_mutex);
	list_add_tail(&bdev->device_list, &glob->device_list);
	mutex_unlock(&glob->device_list_mutex);
1471 1472

	return 0;
1473
out_no_sys:
1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498
	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;
}

1499
void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1500 1501 1502
{
	struct ttm_bo_device *bdev = bo->bdev;

1503
	drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1504
	ttm_mem_io_free_vm(bo);
1505
}
1506 1507 1508 1509 1510 1511 1512 1513 1514

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);
1515
}
1516 1517


1518
EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1519 1520

int ttm_bo_wait(struct ttm_buffer_object *bo,
1521
		bool lazy, bool interruptible, bool no_wait)
1522
{
1523 1524 1525 1526 1527
	struct reservation_object_list *fobj;
	struct reservation_object *resv;
	struct fence *excl;
	long timeout = 15 * HZ;
	int i;
1528

1529 1530 1531 1532 1533 1534 1535
	resv = bo->resv;
	fobj = reservation_object_get_list(resv);
	excl = reservation_object_get_excl(resv);
	if (excl) {
		if (!fence_is_signaled(excl)) {
			if (no_wait)
				return -EBUSY;
1536

1537 1538
			timeout = fence_wait_timeout(excl,
						     interruptible, timeout);
1539
		}
1540
	}
1541

1542 1543 1544 1545
	for (i = 0; fobj && timeout > 0 && i < fobj->shared_count; ++i) {
		struct fence *fence;
		fence = rcu_dereference_protected(fobj->shared[i],
						reservation_object_held(resv));
1546

1547 1548 1549
		if (!fence_is_signaled(fence)) {
			if (no_wait)
				return -EBUSY;
M
Maarten Lankhorst 已提交
1550

1551 1552
			timeout = fence_wait_timeout(fence,
						     interruptible, timeout);
1553 1554
		}
	}
1555 1556 1557 1558 1559 1560 1561 1562 1563 1564

	if (timeout < 0)
		return timeout;

	if (timeout == 0)
		return -EBUSY;

	reservation_object_add_excl_fence(resv, NULL);
	clear_bit(TTM_BO_PRIV_FLAG_MOVING, &bo->priv_flags);
	return 0;
1565 1566 1567 1568 1569 1570 1571 1572
}
EXPORT_SYMBOL(ttm_bo_wait);

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

	/*
1573
	 * Using ttm_bo_reserve makes sure the lru lists are updated.
1574 1575
	 */

1576
	ret = ttm_bo_reserve(bo, true, no_wait, false, NULL);
1577 1578
	if (unlikely(ret != 0))
		return ret;
1579
	ret = ttm_bo_wait(bo, false, true, no_wait);
1580 1581 1582 1583 1584
	if (likely(ret == 0))
		atomic_inc(&bo->cpu_writers);
	ttm_bo_unreserve(bo);
	return ret;
}
1585
EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1586 1587 1588

void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
{
1589
	atomic_dec(&bo->cpu_writers);
1590
}
1591
EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1592 1593 1594 1595 1596 1597 1598 1599

/**
 * 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)
{
1600 1601
	struct ttm_bo_global *glob =
	    container_of(shrink, struct ttm_bo_global, shrink);
1602 1603 1604 1605 1606
	struct ttm_buffer_object *bo;
	int ret = -EBUSY;
	int put_count;
	uint32_t swap_placement = (TTM_PL_FLAG_CACHED | TTM_PL_FLAG_SYSTEM);

1607
	spin_lock(&glob->lru_lock);
1608
	list_for_each_entry(bo, &glob->swap_lru, swap) {
1609
		ret = __ttm_bo_reserve(bo, false, true, false, NULL);
1610 1611 1612
		if (!ret)
			break;
	}
1613

1614 1615 1616 1617
	if (ret) {
		spin_unlock(&glob->lru_lock);
		return ret;
	}
1618

1619
	kref_get(&bo->list_kref);
1620

1621 1622 1623 1624
	if (!list_empty(&bo->ddestroy)) {
		ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
		kref_put(&bo->list_kref, ttm_bo_release_list);
		return ret;
1625 1626 1627
	}

	put_count = ttm_bo_del_from_lru(bo);
1628
	spin_unlock(&glob->lru_lock);
1629

1630
	ttm_bo_list_ref_sub(bo, put_count, true);
1631 1632 1633 1634 1635

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

1636
	ret = ttm_bo_wait(bo, false, false, false);
1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649

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

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

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

		ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1650
					     false, false);
1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661
		if (unlikely(ret != 0))
			goto out;
	}

	ttm_bo_unmap_virtual(bo);

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

1662 1663 1664
	if (bo->bdev->driver->swap_notify)
		bo->bdev->driver->swap_notify(bo);

J
Jan Engelhardt 已提交
1665
	ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1666 1667 1668 1669 1670 1671 1672 1673
out:

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

1674
	__ttm_bo_unreserve(bo);
1675 1676 1677 1678 1679 1680
	kref_put(&bo->list_kref, ttm_bo_release_list);
	return ret;
}

void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
{
1681
	while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1682 1683
		;
}
1684
EXPORT_SYMBOL(ttm_bo_swapout_all);
1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707

/**
 * 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;
1708
	ret = __ttm_bo_reserve(bo, true, false, false, NULL);
1709 1710
	if (unlikely(ret != 0))
		goto out_unlock;
1711
	__ttm_bo_unreserve(bo);
1712 1713 1714 1715 1716

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