xref: /dragonfly/sys/dev/drm/ttm/ttm_bo.c (revision 58ff5dc8)
1 /**************************************************************************
2  *
3  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
4  * All Rights Reserved.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the
8  * "Software"), to deal in the Software without restriction, including
9  * without limitation the rights to use, copy, modify, merge, publish,
10  * distribute, sub license, and/or sell copies of the Software, and to
11  * permit persons to whom the Software is furnished to do so, subject to
12  * the following conditions:
13  *
14  * The above copyright notice and this permission notice (including the
15  * next paragraph) shall be included in all copies or substantial portions
16  * of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
22  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
23  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
24  * USE OR OTHER DEALINGS IN THE SOFTWARE.
25  *
26  **************************************************************************/
27 /*
28  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
29  */
30 
31 #define pr_fmt(fmt) "[TTM] " fmt
32 
33 #include <drm/ttm/ttm_module.h>
34 #include <drm/ttm/ttm_bo_driver.h>
35 #include <drm/ttm/ttm_placement.h>
36 #include <linux/jiffies.h>
37 #include <linux/slab.h>
38 #include <linux/sched.h>
39 #include <linux/mm.h>
40 #include <linux/file.h>
41 #include <linux/module.h>
42 #include <linux/atomic.h>
43 #include <linux/reservation.h>
44 
45 #define TTM_ASSERT_LOCKED(param)
46 #define TTM_DEBUG(fmt, arg...)
47 #define TTM_BO_HASH_ORDER 13
48 
49 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink);
50 static void ttm_bo_global_kobj_release(struct kobject *kobj);
51 
52 static struct attribute ttm_bo_count = {
53 	.name = "bo_count",
54 	.mode = S_IRUGO
55 };
56 
57 static inline int ttm_mem_type_from_place(const struct ttm_place *place,
58 					  uint32_t *mem_type)
59 {
60 	int pos;
61 
62 	pos = ffs(place->flags & TTM_PL_MASK_MEM);
63 	if (unlikely(!pos))
64 		return -EINVAL;
65 
66 	*mem_type = pos - 1;
67 	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 
74 	pr_err("    has_type: %d\n", man->has_type);
75 	pr_err("    use_type: %d\n", man->use_type);
76 	pr_err("    flags: 0x%08X\n", man->flags);
77 	pr_err("    gpu_offset: 0x%08lX\n", man->gpu_offset);
78 	pr_err("    size: %ju\n", man->size);
79 	pr_err("    available_caching: 0x%08X\n", man->available_caching);
80 	pr_err("    default_caching: 0x%08X\n", man->default_caching);
81 	if (mem_type != TTM_PL_SYSTEM)
82 		(*man->func->debug)(man, TTM_PFX);
83 }
84 
85 static void ttm_bo_mem_space_debug(struct ttm_buffer_object *bo,
86 					struct ttm_placement *placement)
87 {
88 	int i, ret, mem_type;
89 
90 	pr_err("No space for %p (%lu pages, %luK, %luM)\n",
91 	       bo, bo->mem.num_pages, bo->mem.size >> 10,
92 	       bo->mem.size >> 20);
93 	for (i = 0; i < placement->num_placement; i++) {
94 		ret = ttm_mem_type_from_place(&placement->placement[i],
95 						&mem_type);
96 		if (ret)
97 			return;
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 static ssize_t ttm_bo_global_show(struct kobject *kobj,
105 				  struct attribute *attr,
106 				  char *buffer)
107 {
108 	struct ttm_bo_global *glob =
109 		container_of(kobj, struct ttm_bo_global, kobj);
110 
111 	return snprintf(buffer, PAGE_SIZE, "%lu\n",
112 			(unsigned long) atomic_read(&glob->bo_count));
113 }
114 
115 static struct attribute *ttm_bo_global_attrs[] = {
116 	&ttm_bo_count,
117 	NULL
118 };
119 
120 static const struct sysfs_ops ttm_bo_global_ops = {
121 	.show = &ttm_bo_global_show
122 };
123 
124 static struct kobj_type ttm_bo_glob_kobj_type  = {
125 	.release = &ttm_bo_global_kobj_release,
126 	.sysfs_ops = &ttm_bo_global_ops,
127 	.default_attrs = ttm_bo_global_attrs
128 };
129 
130 
131 static inline uint32_t ttm_bo_type_flags(unsigned type)
132 {
133 	return 1 << (type);
134 }
135 
136 static void ttm_bo_release_list(struct kref *list_kref)
137 {
138 	struct ttm_buffer_object *bo =
139 	    container_of(list_kref, struct ttm_buffer_object, list_kref);
140 	struct ttm_bo_device *bdev = bo->bdev;
141 	size_t acc_size = bo->acc_size;
142 
143 	BUG_ON(kref_read(&bo->list_kref));
144 	BUG_ON(kref_read(&bo->kref));
145 	BUG_ON(atomic_read(&bo->cpu_writers));
146 	BUG_ON(bo->mem.mm_node != NULL);
147 	BUG_ON(!list_empty(&bo->lru));
148 	BUG_ON(!list_empty(&bo->ddestroy));
149 	ttm_tt_destroy(bo->ttm);
150 	atomic_dec(&bo->glob->bo_count);
151 	dma_fence_put(bo->moving);
152 	if (bo->resv == &bo->ttm_resv)
153 		reservation_object_fini(&bo->ttm_resv);
154 	mutex_destroy(&bo->wu_mutex);
155 	if (bo->destroy)
156 		bo->destroy(bo);
157 	else {
158 		kfree(bo);
159 	}
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 	if (!(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
171 
172 		BUG_ON(!list_empty(&bo->lru));
173 
174 		man = &bdev->man[bo->mem.mem_type];
175 		list_add_tail(&bo->lru, &man->lru[bo->priority]);
176 		kref_get(&bo->list_kref);
177 
178 		if (bo->ttm && !(bo->ttm->page_flags & TTM_PAGE_FLAG_SG)) {
179 			list_add_tail(&bo->swap,
180 				      &bo->glob->swap_lru[bo->priority]);
181 			kref_get(&bo->list_kref);
182 		}
183 	}
184 }
185 EXPORT_SYMBOL(ttm_bo_add_to_lru);
186 
187 static void ttm_bo_ref_bug(struct kref *list_kref)
188 {
189 	BUG();
190 }
191 
192 void ttm_bo_del_from_lru(struct ttm_buffer_object *bo)
193 {
194 	if (!list_empty(&bo->swap)) {
195 		list_del_init(&bo->swap);
196 		kref_put(&bo->list_kref, ttm_bo_ref_bug);
197 	}
198 	if (!list_empty(&bo->lru)) {
199 		list_del_init(&bo->lru);
200 		kref_put(&bo->list_kref, ttm_bo_ref_bug);
201 	}
202 
203 	/*
204 	 * TODO: Add a driver hook to delete from
205 	 * driver-specific LRU's here.
206 	 */
207 }
208 
209 void ttm_bo_del_sub_from_lru(struct ttm_buffer_object *bo)
210 {
211 	lockmgr(&bo->glob->lru_lock, LK_EXCLUSIVE);
212 	ttm_bo_del_from_lru(bo);
213 	lockmgr(&bo->glob->lru_lock, LK_RELEASE);
214 }
215 EXPORT_SYMBOL(ttm_bo_del_sub_from_lru);
216 
217 void ttm_bo_move_to_lru_tail(struct ttm_buffer_object *bo)
218 {
219 	lockdep_assert_held(&bo->resv->lock.base);
220 
221 	ttm_bo_del_from_lru(bo);
222 	ttm_bo_add_to_lru(bo);
223 }
224 EXPORT_SYMBOL(ttm_bo_move_to_lru_tail);
225 
226 /*
227  * Call bo->mutex locked.
228  */
229 static int ttm_bo_add_ttm(struct ttm_buffer_object *bo, bool zero_alloc)
230 {
231 	struct ttm_bo_device *bdev = bo->bdev;
232 	struct ttm_bo_global *glob = bo->glob;
233 	int ret = 0;
234 	uint32_t page_flags = 0;
235 
236 	TTM_ASSERT_LOCKED(&bo->mutex);
237 	bo->ttm = NULL;
238 
239 	if (bdev->need_dma32)
240 		page_flags |= TTM_PAGE_FLAG_DMA32;
241 
242 	switch (bo->type) {
243 	case ttm_bo_type_device:
244 		if (zero_alloc)
245 			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
246 	case ttm_bo_type_kernel:
247 		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
248 						      page_flags, glob->dummy_read_page);
249 		if (unlikely(bo->ttm == NULL))
250 			ret = -ENOMEM;
251 		break;
252 	case ttm_bo_type_sg:
253 		bo->ttm = bdev->driver->ttm_tt_create(bdev, bo->num_pages << PAGE_SHIFT,
254 						      page_flags | TTM_PAGE_FLAG_SG,
255 						      glob->dummy_read_page);
256 		if (unlikely(bo->ttm == NULL)) {
257 			ret = -ENOMEM;
258 			break;
259 		}
260 		bo->ttm->sg = bo->sg;
261 		break;
262 	default:
263 		pr_err("Illegal buffer object type\n");
264 		ret = -EINVAL;
265 		break;
266 	}
267 
268 	return ret;
269 }
270 
271 static int ttm_bo_handle_move_mem(struct ttm_buffer_object *bo,
272 				  struct ttm_mem_reg *mem,
273 				  bool evict, bool interruptible,
274 				  bool no_wait_gpu)
275 {
276 	struct ttm_bo_device *bdev = bo->bdev;
277 	bool old_is_pci = ttm_mem_reg_is_pci(bdev, &bo->mem);
278 	bool new_is_pci = ttm_mem_reg_is_pci(bdev, mem);
279 	struct ttm_mem_type_manager *old_man = &bdev->man[bo->mem.mem_type];
280 	struct ttm_mem_type_manager *new_man = &bdev->man[mem->mem_type];
281 	int ret = 0;
282 
283 	if (old_is_pci || new_is_pci ||
284 	    ((mem->placement & bo->mem.placement & TTM_PL_MASK_CACHING) == 0)) {
285 		ret = ttm_mem_io_lock(old_man, true);
286 		if (unlikely(ret != 0))
287 			goto out_err;
288 		ttm_bo_unmap_virtual_locked(bo);
289 		ttm_mem_io_unlock(old_man);
290 	}
291 
292 	/*
293 	 * Create and bind a ttm if required.
294 	 */
295 
296 	if (!(new_man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
297 		if (bo->ttm == NULL) {
298 			bool zero = !(old_man->flags & TTM_MEMTYPE_FLAG_FIXED);
299 			ret = ttm_bo_add_ttm(bo, zero);
300 			if (ret)
301 				goto out_err;
302 		}
303 
304 		ret = ttm_tt_set_placement_caching(bo->ttm, mem->placement);
305 		if (ret)
306 			goto out_err;
307 
308 		if (mem->mem_type != TTM_PL_SYSTEM) {
309 			ret = ttm_tt_bind(bo->ttm, mem);
310 			if (ret)
311 				goto out_err;
312 		}
313 
314 		if (bo->mem.mem_type == TTM_PL_SYSTEM) {
315 			if (bdev->driver->move_notify)
316 				bdev->driver->move_notify(bo, evict, mem);
317 			bo->mem = *mem;
318 			mem->mm_node = NULL;
319 			goto moved;
320 		}
321 	}
322 
323 	if (bdev->driver->move_notify)
324 		bdev->driver->move_notify(bo, evict, mem);
325 
326 	if (!(old_man->flags & TTM_MEMTYPE_FLAG_FIXED) &&
327 	    !(new_man->flags & TTM_MEMTYPE_FLAG_FIXED))
328 		ret = ttm_bo_move_ttm(bo, interruptible, no_wait_gpu, mem);
329 	else if (bdev->driver->move)
330 		ret = bdev->driver->move(bo, evict, interruptible,
331 					 no_wait_gpu, mem);
332 	else
333 		ret = ttm_bo_move_memcpy(bo, interruptible, no_wait_gpu, mem);
334 
335 	if (ret) {
336 		if (bdev->driver->move_notify) {
337 			struct ttm_mem_reg tmp_mem = *mem;
338 			*mem = bo->mem;
339 			bo->mem = tmp_mem;
340 			bdev->driver->move_notify(bo, false, mem);
341 			bo->mem = *mem;
342 			*mem = tmp_mem;
343 		}
344 
345 		goto out_err;
346 	}
347 
348 moved:
349 	if (bo->evicted) {
350 		if (bdev->driver->invalidate_caches) {
351 			ret = bdev->driver->invalidate_caches(bdev, bo->mem.placement);
352 			if (ret)
353 				pr_err("Can not flush read caches\n");
354 		}
355 		bo->evicted = false;
356 	}
357 
358 	if (bo->mem.mm_node) {
359 		bo->offset = (bo->mem.start << PAGE_SHIFT) +
360 		    bdev->man[bo->mem.mem_type].gpu_offset;
361 		bo->cur_placement = bo->mem.placement;
362 	} else
363 		bo->offset = 0;
364 
365 	return 0;
366 
367 out_err:
368 	new_man = &bdev->man[bo->mem.mem_type];
369 	if (new_man->flags & TTM_MEMTYPE_FLAG_FIXED) {
370 		ttm_tt_destroy(bo->ttm);
371 		bo->ttm = NULL;
372 	}
373 
374 	return ret;
375 }
376 
377 /**
378  * Call bo::reserved.
379  * Will release GPU memory type usage on destruction.
380  * This is the place to put in driver specific hooks to release
381  * driver private resources.
382  * Will release the bo::reserved lock.
383  */
384 
385 static void ttm_bo_cleanup_memtype_use(struct ttm_buffer_object *bo)
386 {
387 	if (bo->bdev->driver->move_notify)
388 		bo->bdev->driver->move_notify(bo, false, NULL);
389 
390 	ttm_tt_destroy(bo->ttm);
391 	bo->ttm = NULL;
392 	ttm_bo_mem_put(bo, &bo->mem);
393 
394 	ww_mutex_unlock (&bo->resv->lock);
395 }
396 
397 static void ttm_bo_flush_all_fences(struct ttm_buffer_object *bo)
398 {
399 	struct reservation_object_list *fobj;
400 	struct dma_fence *fence;
401 	int i;
402 
403 	fobj = reservation_object_get_list(bo->resv);
404 	fence = reservation_object_get_excl(bo->resv);
405 	if (fence && !fence->ops->signaled)
406 		dma_fence_enable_sw_signaling(fence);
407 
408 	for (i = 0; fobj && i < fobj->shared_count; ++i) {
409 		fence = rcu_dereference_protected(fobj->shared[i],
410 					reservation_object_held(bo->resv));
411 
412 		if (!fence->ops->signaled)
413 			dma_fence_enable_sw_signaling(fence);
414 	}
415 }
416 
417 static void ttm_bo_cleanup_refs_or_queue(struct ttm_buffer_object *bo)
418 {
419 	struct ttm_bo_device *bdev = bo->bdev;
420 	struct ttm_bo_global *glob = bo->glob;
421 	int ret;
422 
423 	lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
424 	ret = __ttm_bo_reserve(bo, false, true, NULL);
425 
426 	if (!ret) {
427 		if (!ttm_bo_wait(bo, false, true)) {
428 			ttm_bo_del_from_lru(bo);
429 			lockmgr(&glob->lru_lock, LK_RELEASE);
430 			ttm_bo_cleanup_memtype_use(bo);
431 
432 			return;
433 		} else
434 			ttm_bo_flush_all_fences(bo);
435 
436 		/*
437 		 * Make NO_EVICT bos immediately available to
438 		 * shrinkers, now that they are queued for
439 		 * destruction.
440 		 */
441 		if (bo->mem.placement & TTM_PL_FLAG_NO_EVICT) {
442 			bo->mem.placement &= ~TTM_PL_FLAG_NO_EVICT;
443 			ttm_bo_add_to_lru(bo);
444 		}
445 
446 		__ttm_bo_unreserve(bo);
447 	}
448 
449 	kref_get(&bo->list_kref);
450 	list_add_tail(&bo->ddestroy, &bdev->ddestroy);
451 	lockmgr(&glob->lru_lock, LK_RELEASE);
452 
453 	schedule_delayed_work(&bdev->wq,
454 			      ((HZ / 100) < 1) ? 1 : HZ / 100);
455 }
456 
457 /**
458  * function ttm_bo_cleanup_refs_and_unlock
459  * If bo idle, remove from delayed- and lru lists, and unref.
460  * If not idle, do nothing.
461  *
462  * Must be called with lru_lock and reservation held, this function
463  * will drop both before returning.
464  *
465  * @interruptible         Any sleeps should occur interruptibly.
466  * @no_wait_gpu           Never wait for gpu. Return -EBUSY instead.
467  */
468 
469 static int ttm_bo_cleanup_refs_and_unlock(struct ttm_buffer_object *bo,
470 					  bool interruptible,
471 					  bool no_wait_gpu)
472 {
473 	struct ttm_bo_global *glob = bo->glob;
474 	int ret;
475 
476 	ret = ttm_bo_wait(bo, false, true);
477 
478 	if (ret && !no_wait_gpu) {
479 		long lret;
480 		ww_mutex_unlock(&bo->resv->lock);
481 		lockmgr(&glob->lru_lock, LK_RELEASE);
482 
483 		lret = reservation_object_wait_timeout_rcu(bo->resv,
484 							   true,
485 							   interruptible,
486 							   30 * HZ);
487 
488 		if (lret < 0)
489 			return lret;
490 		else if (lret == 0)
491 			return -EBUSY;
492 
493 		lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
494 		ret = __ttm_bo_reserve(bo, false, true, NULL);
495 
496 		/*
497 		 * We raced, and lost, someone else holds the reservation now,
498 		 * and is probably busy in ttm_bo_cleanup_memtype_use.
499 		 *
500 		 * Even if it's not the case, because we finished waiting any
501 		 * delayed destruction would succeed, so just return success
502 		 * here.
503 		 */
504 		if (ret) {
505 			lockmgr(&glob->lru_lock, LK_RELEASE);
506 			return 0;
507 		}
508 
509 		/*
510 		 * remove sync_obj with ttm_bo_wait, the wait should be
511 		 * finished, and no new wait object should have been added.
512 		 */
513 		ret = ttm_bo_wait(bo, false, true);
514 		WARN_ON(ret);
515 	}
516 
517 	if (ret || unlikely(list_empty(&bo->ddestroy))) {
518 		__ttm_bo_unreserve(bo);
519 		lockmgr(&glob->lru_lock, LK_RELEASE);
520 		return ret;
521 	}
522 
523 	ttm_bo_del_from_lru(bo);
524 	list_del_init(&bo->ddestroy);
525 	kref_put(&bo->list_kref, ttm_bo_ref_bug);
526 
527 	lockmgr(&glob->lru_lock, LK_RELEASE);
528 	ttm_bo_cleanup_memtype_use(bo);
529 
530 	return 0;
531 }
532 
533 /**
534  * Traverse the delayed list, and call ttm_bo_cleanup_refs on all
535  * encountered buffers.
536  */
537 
538 static int ttm_bo_delayed_delete(struct ttm_bo_device *bdev, bool remove_all)
539 {
540 	struct ttm_bo_global *glob = bdev->glob;
541 	struct ttm_buffer_object *entry = NULL;
542 	int ret = 0;
543 
544 	lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
545 	if (list_empty(&bdev->ddestroy))
546 		goto out_unlock;
547 
548 	entry = list_first_entry(&bdev->ddestroy,
549 		struct ttm_buffer_object, ddestroy);
550 	kref_get(&entry->list_kref);
551 
552 	for (;;) {
553 		struct ttm_buffer_object *nentry = NULL;
554 
555 		if (entry->ddestroy.next != &bdev->ddestroy) {
556 			nentry = list_first_entry(&entry->ddestroy,
557 				struct ttm_buffer_object, ddestroy);
558 			kref_get(&nentry->list_kref);
559 		}
560 
561 		ret = __ttm_bo_reserve(entry, false, true, NULL);
562 		if (remove_all && ret) {
563 			lockmgr(&glob->lru_lock, LK_RELEASE);
564 			ret = __ttm_bo_reserve(entry, false, false, NULL);
565 			lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
566 		}
567 
568 		if (!ret)
569 			ret = ttm_bo_cleanup_refs_and_unlock(entry, false,
570 							     !remove_all);
571 		else
572 			lockmgr(&glob->lru_lock, LK_RELEASE);
573 
574 		kref_put(&entry->list_kref, ttm_bo_release_list);
575 		entry = nentry;
576 
577 		if (ret || !entry)
578 			goto out;
579 
580 		lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
581 		if (list_empty(&entry->ddestroy))
582 			break;
583 	}
584 
585 out_unlock:
586 	lockmgr(&glob->lru_lock, LK_RELEASE);
587 out:
588 	if (entry)
589 		kref_put(&entry->list_kref, ttm_bo_release_list);
590 	return ret;
591 }
592 
593 static void ttm_bo_delayed_workqueue(struct work_struct *work)
594 {
595 	struct ttm_bo_device *bdev =
596 	    container_of(work, struct ttm_bo_device, wq.work);
597 
598 	if (ttm_bo_delayed_delete(bdev, false)) {
599 		schedule_delayed_work(&bdev->wq,
600 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
601 	}
602 }
603 
604 static void ttm_bo_release(struct kref *kref)
605 {
606 	struct ttm_buffer_object *bo =
607 	    container_of(kref, struct ttm_buffer_object, kref);
608 	struct ttm_bo_device *bdev = bo->bdev;
609 	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
610 
611 	drm_vma_offset_remove(&bdev->vma_manager, &bo->vma_node);
612 	ttm_mem_io_lock(man, false);
613 	ttm_mem_io_free_vm(bo);
614 	ttm_mem_io_unlock(man);
615 	ttm_bo_cleanup_refs_or_queue(bo);
616 	kref_put(&bo->list_kref, ttm_bo_release_list);
617 }
618 
619 void ttm_bo_unref(struct ttm_buffer_object **p_bo)
620 {
621 	struct ttm_buffer_object *bo = *p_bo;
622 
623 	*p_bo = NULL;
624 	kref_put(&bo->kref, ttm_bo_release);
625 }
626 EXPORT_SYMBOL(ttm_bo_unref);
627 
628 int ttm_bo_lock_delayed_workqueue(struct ttm_bo_device *bdev)
629 {
630 	return cancel_delayed_work_sync(&bdev->wq);
631 }
632 EXPORT_SYMBOL(ttm_bo_lock_delayed_workqueue);
633 
634 void ttm_bo_unlock_delayed_workqueue(struct ttm_bo_device *bdev, int resched)
635 {
636 	if (resched)
637 		schedule_delayed_work(&bdev->wq,
638 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
639 }
640 EXPORT_SYMBOL(ttm_bo_unlock_delayed_workqueue);
641 
642 static int ttm_bo_evict(struct ttm_buffer_object *bo, bool interruptible,
643 			bool no_wait_gpu)
644 {
645 	struct ttm_bo_device *bdev = bo->bdev;
646 	struct ttm_mem_reg evict_mem;
647 	struct ttm_placement placement;
648 	int ret = 0;
649 
650 	lockdep_assert_held(&bo->resv->lock.base);
651 
652 	evict_mem = bo->mem;
653 	evict_mem.mm_node = NULL;
654 	evict_mem.bus.io_reserved_vm = false;
655 	evict_mem.bus.io_reserved_count = 0;
656 
657 	placement.num_placement = 0;
658 	placement.num_busy_placement = 0;
659 	bdev->driver->evict_flags(bo, &placement);
660 	ret = ttm_bo_mem_space(bo, &placement, &evict_mem, interruptible,
661 				no_wait_gpu);
662 	if (ret) {
663 		if (ret != -ERESTARTSYS) {
664 			pr_err("Failed to find memory space for buffer 0x%p eviction\n",
665 			       bo);
666 			ttm_bo_mem_space_debug(bo, &placement);
667 		}
668 		goto out;
669 	}
670 
671 	ret = ttm_bo_handle_move_mem(bo, &evict_mem, true, interruptible,
672 				     no_wait_gpu);
673 	if (unlikely(ret)) {
674 		if (ret != -ERESTARTSYS)
675 			pr_err("Buffer eviction failed\n");
676 		ttm_bo_mem_put(bo, &evict_mem);
677 		goto out;
678 	}
679 	bo->evicted = true;
680 out:
681 	return ret;
682 }
683 
684 bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
685 			      const struct ttm_place *place)
686 {
687 	/* Don't evict this BO if it's outside of the
688 	 * requested placement range
689 	 */
690 	if (place->fpfn >= (bo->mem.start + bo->mem.size) ||
691 	    (place->lpfn && place->lpfn <= bo->mem.start))
692 		return false;
693 
694 	return true;
695 }
696 EXPORT_SYMBOL(ttm_bo_eviction_valuable);
697 
698 static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
699 				uint32_t mem_type,
700 				const struct ttm_place *place,
701 				bool interruptible,
702 				bool no_wait_gpu)
703 {
704 	struct ttm_bo_global *glob = bdev->glob;
705 	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
706 	struct ttm_buffer_object *bo;
707 	int ret = -EBUSY;
708 	unsigned i;
709 
710 	lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
711 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
712 		list_for_each_entry(bo, &man->lru[i], lru) {
713 			ret = __ttm_bo_reserve(bo, false, true, NULL);
714 			if (ret)
715 				continue;
716 
717 			if (place && !bdev->driver->eviction_valuable(bo,
718 								      place)) {
719 				__ttm_bo_unreserve(bo);
720 				ret = -EBUSY;
721 				continue;
722 			}
723 
724 			break;
725 		}
726 
727 		if (!ret)
728 			break;
729 	}
730 
731 	if (ret) {
732 		lockmgr(&glob->lru_lock, LK_RELEASE);
733 		return ret;
734 	}
735 
736 	kref_get(&bo->list_kref);
737 
738 	if (!list_empty(&bo->ddestroy)) {
739 		ret = ttm_bo_cleanup_refs_and_unlock(bo, interruptible,
740 						     no_wait_gpu);
741 		kref_put(&bo->list_kref, ttm_bo_release_list);
742 		return ret;
743 	}
744 
745 	ttm_bo_del_from_lru(bo);
746 	lockmgr(&glob->lru_lock, LK_RELEASE);
747 
748 	BUG_ON(ret != 0);
749 
750 	ret = ttm_bo_evict(bo, interruptible, no_wait_gpu);
751 	ttm_bo_unreserve(bo);
752 
753 	kref_put(&bo->list_kref, ttm_bo_release_list);
754 	return ret;
755 }
756 
757 void ttm_bo_mem_put(struct ttm_buffer_object *bo, struct ttm_mem_reg *mem)
758 {
759 	struct ttm_mem_type_manager *man = &bo->bdev->man[mem->mem_type];
760 
761 	if (mem->mm_node)
762 		(*man->func->put_node)(man, mem);
763 }
764 EXPORT_SYMBOL(ttm_bo_mem_put);
765 
766 /**
767  * Add the last move fence to the BO and reserve a new shared slot.
768  */
769 static int ttm_bo_add_move_fence(struct ttm_buffer_object *bo,
770 				 struct ttm_mem_type_manager *man,
771 				 struct ttm_mem_reg *mem)
772 {
773 	struct dma_fence *fence;
774 	int ret;
775 
776 	lockmgr(&man->move_lock, LK_EXCLUSIVE);
777 	fence = dma_fence_get(man->move);
778 	lockmgr(&man->move_lock, LK_RELEASE);
779 
780 	if (fence) {
781 		reservation_object_add_shared_fence(bo->resv, fence);
782 
783 		ret = reservation_object_reserve_shared(bo->resv);
784 		if (unlikely(ret))
785 			return ret;
786 
787 		dma_fence_put(bo->moving);
788 		bo->moving = fence;
789 	}
790 
791 	return 0;
792 }
793 
794 /**
795  * Repeatedly evict memory from the LRU for @mem_type until we create enough
796  * space, or we've evicted everything and there isn't enough space.
797  */
798 static int ttm_bo_mem_force_space(struct ttm_buffer_object *bo,
799 					uint32_t mem_type,
800 					const struct ttm_place *place,
801 					struct ttm_mem_reg *mem,
802 					bool interruptible,
803 					bool no_wait_gpu)
804 {
805 	struct ttm_bo_device *bdev = bo->bdev;
806 	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
807 	int ret;
808 
809 	do {
810 		ret = (*man->func->get_node)(man, bo, place, mem);
811 		if (unlikely(ret != 0))
812 			return ret;
813 		if (mem->mm_node)
814 			break;
815 		ret = ttm_mem_evict_first(bdev, mem_type, place,
816 					  interruptible, no_wait_gpu);
817 		if (unlikely(ret != 0))
818 			return ret;
819 	} while (1);
820 	mem->mem_type = mem_type;
821 	return ttm_bo_add_move_fence(bo, man, mem);
822 }
823 
824 static uint32_t ttm_bo_select_caching(struct ttm_mem_type_manager *man,
825 				      uint32_t cur_placement,
826 				      uint32_t proposed_placement)
827 {
828 	uint32_t caching = proposed_placement & TTM_PL_MASK_CACHING;
829 	uint32_t result = proposed_placement & ~TTM_PL_MASK_CACHING;
830 
831 	/**
832 	 * Keep current caching if possible.
833 	 */
834 
835 	if ((cur_placement & caching) != 0)
836 		result |= (cur_placement & caching);
837 	else if ((man->default_caching & caching) != 0)
838 		result |= man->default_caching;
839 	else if ((TTM_PL_FLAG_CACHED & caching) != 0)
840 		result |= TTM_PL_FLAG_CACHED;
841 	else if ((TTM_PL_FLAG_WC & caching) != 0)
842 		result |= TTM_PL_FLAG_WC;
843 	else if ((TTM_PL_FLAG_UNCACHED & caching) != 0)
844 		result |= TTM_PL_FLAG_UNCACHED;
845 
846 	return result;
847 }
848 
849 static bool ttm_bo_mt_compatible(struct ttm_mem_type_manager *man,
850 				 uint32_t mem_type,
851 				 const struct ttm_place *place,
852 				 uint32_t *masked_placement)
853 {
854 	uint32_t cur_flags = ttm_bo_type_flags(mem_type);
855 
856 	if ((cur_flags & place->flags & TTM_PL_MASK_MEM) == 0)
857 		return false;
858 
859 	if ((place->flags & man->available_caching) == 0)
860 		return false;
861 
862 	cur_flags |= (place->flags & man->available_caching);
863 
864 	*masked_placement = cur_flags;
865 	return true;
866 }
867 
868 /**
869  * Creates space for memory region @mem according to its type.
870  *
871  * This function first searches for free space in compatible memory types in
872  * the priority order defined by the driver.  If free space isn't found, then
873  * ttm_bo_mem_force_space is attempted in priority order to evict and find
874  * space.
875  */
876 int ttm_bo_mem_space(struct ttm_buffer_object *bo,
877 			struct ttm_placement *placement,
878 			struct ttm_mem_reg *mem,
879 			bool interruptible,
880 			bool no_wait_gpu)
881 {
882 	struct ttm_bo_device *bdev = bo->bdev;
883 	struct ttm_mem_type_manager *man;
884 	uint32_t mem_type = TTM_PL_SYSTEM;
885 	uint32_t cur_flags = 0;
886 	bool type_found = false;
887 	bool type_ok = false;
888 	bool has_erestartsys = false;
889 	int i, ret;
890 
891 	ret = reservation_object_reserve_shared(bo->resv);
892 	if (unlikely(ret))
893 		return ret;
894 
895 	mem->mm_node = NULL;
896 	for (i = 0; i < placement->num_placement; ++i) {
897 		const struct ttm_place *place = &placement->placement[i];
898 
899 		ret = ttm_mem_type_from_place(place, &mem_type);
900 		if (ret)
901 			return ret;
902 		man = &bdev->man[mem_type];
903 		if (!man->has_type || !man->use_type)
904 			continue;
905 
906 		type_ok = ttm_bo_mt_compatible(man, mem_type, place,
907 						&cur_flags);
908 
909 		if (!type_ok)
910 			continue;
911 
912 		type_found = true;
913 		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
914 						  cur_flags);
915 		/*
916 		 * Use the access and other non-mapping-related flag bits from
917 		 * the memory placement flags to the current flags
918 		 */
919 		ttm_flag_masked(&cur_flags, place->flags,
920 				~TTM_PL_MASK_MEMTYPE);
921 
922 		if (mem_type == TTM_PL_SYSTEM)
923 			break;
924 
925 		ret = (*man->func->get_node)(man, bo, place, mem);
926 		if (unlikely(ret))
927 			return ret;
928 
929 		if (mem->mm_node) {
930 			ret = ttm_bo_add_move_fence(bo, man, mem);
931 			if (unlikely(ret)) {
932 				(*man->func->put_node)(man, mem);
933 				return ret;
934 			}
935 			break;
936 		}
937 	}
938 
939 	if ((type_ok && (mem_type == TTM_PL_SYSTEM)) || mem->mm_node) {
940 		mem->mem_type = mem_type;
941 		mem->placement = cur_flags;
942 		return 0;
943 	}
944 
945 	for (i = 0; i < placement->num_busy_placement; ++i) {
946 		const struct ttm_place *place = &placement->busy_placement[i];
947 
948 		ret = ttm_mem_type_from_place(place, &mem_type);
949 		if (ret)
950 			return ret;
951 		man = &bdev->man[mem_type];
952 		if (!man->has_type || !man->use_type)
953 			continue;
954 		if (!ttm_bo_mt_compatible(man, mem_type, place, &cur_flags))
955 			continue;
956 
957 		type_found = true;
958 		cur_flags = ttm_bo_select_caching(man, bo->mem.placement,
959 						  cur_flags);
960 		/*
961 		 * Use the access and other non-mapping-related flag bits from
962 		 * the memory placement flags to the current flags
963 		 */
964 		ttm_flag_masked(&cur_flags, place->flags,
965 				~TTM_PL_MASK_MEMTYPE);
966 
967 		if (mem_type == TTM_PL_SYSTEM) {
968 			mem->mem_type = mem_type;
969 			mem->placement = cur_flags;
970 			mem->mm_node = NULL;
971 			return 0;
972 		}
973 
974 		ret = ttm_bo_mem_force_space(bo, mem_type, place, mem,
975 						interruptible, no_wait_gpu);
976 		if (ret == 0 && mem->mm_node) {
977 			mem->placement = cur_flags;
978 			return 0;
979 		}
980 		if (ret == -ERESTARTSYS)
981 			has_erestartsys = true;
982 	}
983 
984 	if (!type_found) {
985 		pr_err(TTM_PFX "No compatible memory type found\n");
986 		return -EINVAL;
987 	}
988 
989 	return (has_erestartsys) ? -ERESTARTSYS : -ENOMEM;
990 }
991 EXPORT_SYMBOL(ttm_bo_mem_space);
992 
993 static int ttm_bo_move_buffer(struct ttm_buffer_object *bo,
994 			struct ttm_placement *placement,
995 			bool interruptible,
996 			bool no_wait_gpu)
997 {
998 	int ret = 0;
999 	struct ttm_mem_reg mem;
1000 
1001 	lockdep_assert_held(&bo->resv->lock.base);
1002 
1003 	mem.num_pages = bo->num_pages;
1004 	mem.size = mem.num_pages << PAGE_SHIFT;
1005 	mem.page_alignment = bo->mem.page_alignment;
1006 	mem.bus.io_reserved_vm = false;
1007 	mem.bus.io_reserved_count = 0;
1008 	/*
1009 	 * Determine where to move the buffer.
1010 	 */
1011 	ret = ttm_bo_mem_space(bo, placement, &mem,
1012 			       interruptible, no_wait_gpu);
1013 	if (ret)
1014 		goto out_unlock;
1015 	ret = ttm_bo_handle_move_mem(bo, &mem, false,
1016 				     interruptible, no_wait_gpu);
1017 out_unlock:
1018 	if (ret && mem.mm_node)
1019 		ttm_bo_mem_put(bo, &mem);
1020 	return ret;
1021 }
1022 
1023 static bool ttm_bo_places_compat(const struct ttm_place *places,
1024 				 unsigned num_placement,
1025 				 struct ttm_mem_reg *mem,
1026 				 uint32_t *new_flags)
1027 {
1028 	unsigned i;
1029 
1030 	for (i = 0; i < num_placement; i++) {
1031 		const struct ttm_place *heap = &places[i];
1032 
1033 		if (mem->mm_node && (mem->start < heap->fpfn ||
1034 		     (heap->lpfn != 0 && (mem->start + mem->num_pages) > heap->lpfn)))
1035 			continue;
1036 
1037 		*new_flags = heap->flags;
1038 		if ((*new_flags & mem->placement & TTM_PL_MASK_CACHING) &&
1039 		    (*new_flags & mem->placement & TTM_PL_MASK_MEM) &&
1040 		    (!(*new_flags & TTM_PL_FLAG_CONTIGUOUS) ||
1041 		     (mem->placement & TTM_PL_FLAG_CONTIGUOUS)))
1042 			return true;
1043 	}
1044 	return false;
1045 }
1046 
1047 bool ttm_bo_mem_compat(struct ttm_placement *placement,
1048 		       struct ttm_mem_reg *mem,
1049 		       uint32_t *new_flags)
1050 {
1051 	if (ttm_bo_places_compat(placement->placement, placement->num_placement,
1052 				 mem, new_flags))
1053 		return true;
1054 
1055 	if ((placement->busy_placement != placement->placement ||
1056 	     placement->num_busy_placement > placement->num_placement) &&
1057 	    ttm_bo_places_compat(placement->busy_placement,
1058 				 placement->num_busy_placement,
1059 				 mem, new_flags))
1060 		return true;
1061 
1062 	return false;
1063 }
1064 EXPORT_SYMBOL(ttm_bo_mem_compat);
1065 
1066 int ttm_bo_validate(struct ttm_buffer_object *bo,
1067 			struct ttm_placement *placement,
1068 			bool interruptible,
1069 			bool no_wait_gpu)
1070 {
1071 	int ret;
1072 	uint32_t new_flags;
1073 
1074 	lockdep_assert_held(&bo->resv->lock.base);
1075 	/*
1076 	 * Check whether we need to move buffer.
1077 	 */
1078 	if (!ttm_bo_mem_compat(placement, &bo->mem, &new_flags)) {
1079 		ret = ttm_bo_move_buffer(bo, placement, interruptible,
1080 					 no_wait_gpu);
1081 		if (ret)
1082 			return ret;
1083 	} else {
1084 		/*
1085 		 * Use the access and other non-mapping-related flag bits from
1086 		 * the compatible memory placement flags to the active flags
1087 		 */
1088 		ttm_flag_masked(&bo->mem.placement, new_flags,
1089 				~TTM_PL_MASK_MEMTYPE);
1090 	}
1091 	/*
1092 	 * We might need to add a TTM.
1093 	 */
1094 	if (bo->mem.mem_type == TTM_PL_SYSTEM && bo->ttm == NULL) {
1095 		ret = ttm_bo_add_ttm(bo, true);
1096 		if (ret)
1097 			return ret;
1098 	}
1099 	return 0;
1100 }
1101 EXPORT_SYMBOL(ttm_bo_validate);
1102 
1103 int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
1104 			 struct ttm_buffer_object *bo,
1105 			 unsigned long size,
1106 			 enum ttm_bo_type type,
1107 			 struct ttm_placement *placement,
1108 			 uint32_t page_alignment,
1109 			 bool interruptible,
1110 			 struct vm_object *persistent_swap_storage,
1111 			 size_t acc_size,
1112 			 struct sg_table *sg,
1113 			 struct reservation_object *resv,
1114 			 void (*destroy) (struct ttm_buffer_object *))
1115 {
1116 	int ret = 0;
1117 	unsigned long num_pages;
1118 	struct ttm_mem_global *mem_glob = bdev->glob->mem_glob;
1119 	bool locked;
1120 
1121 	ret = ttm_mem_global_alloc(mem_glob, acc_size, false, false);
1122 	if (ret) {
1123 		pr_err("Out of kernel memory\n");
1124 		if (destroy)
1125 			(*destroy)(bo);
1126 		else
1127 			kfree(bo);
1128 		return -ENOMEM;
1129 	}
1130 
1131 	num_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1132 	if (num_pages == 0) {
1133 		pr_err("Illegal buffer object size\n");
1134 		if (destroy)
1135 			(*destroy)(bo);
1136 		else
1137 			kfree(bo);
1138 		ttm_mem_global_free(mem_glob, acc_size);
1139 		return -EINVAL;
1140 	}
1141 	bo->destroy = destroy;
1142 
1143 	kref_init(&bo->kref);
1144 	kref_init(&bo->list_kref);
1145 	atomic_set(&bo->cpu_writers, 0);
1146 	INIT_LIST_HEAD(&bo->lru);
1147 	INIT_LIST_HEAD(&bo->ddestroy);
1148 	INIT_LIST_HEAD(&bo->swap);
1149 	INIT_LIST_HEAD(&bo->io_reserve_lru);
1150 	lockinit(&bo->wu_mutex, "ttmbwm", 0, LK_CANRECURSE);
1151 	bo->bdev = bdev;
1152 	bo->glob = bdev->glob;
1153 	bo->type = type;
1154 	bo->num_pages = num_pages;
1155 	bo->mem.size = num_pages << PAGE_SHIFT;
1156 	bo->mem.mem_type = TTM_PL_SYSTEM;
1157 	bo->mem.num_pages = bo->num_pages;
1158 	bo->mem.mm_node = NULL;
1159 	bo->mem.page_alignment = page_alignment;
1160 	bo->mem.bus.io_reserved_vm = false;
1161 	bo->mem.bus.io_reserved_count = 0;
1162 	bo->moving = NULL;
1163 	bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
1164 	bo->persistent_swap_storage = persistent_swap_storage;
1165 	bo->acc_size = acc_size;
1166 	bo->sg = sg;
1167 	if (resv) {
1168 		bo->resv = resv;
1169 		lockdep_assert_held(&bo->resv->lock.base);
1170 	} else {
1171 		bo->resv = &bo->ttm_resv;
1172 		reservation_object_init(&bo->ttm_resv);
1173 	}
1174 	atomic_inc(&bo->glob->bo_count);
1175 	drm_vma_node_reset(&bo->vma_node);
1176 	bo->priority = 0;
1177 
1178 	/*
1179 	 * For ttm_bo_type_device buffers, allocate
1180 	 * address space from the device.
1181 	 */
1182 	if (bo->type == ttm_bo_type_device ||
1183 	    bo->type == ttm_bo_type_sg)
1184 		ret = drm_vma_offset_add(&bdev->vma_manager, &bo->vma_node,
1185 					 bo->mem.num_pages);
1186 
1187 	/* passed reservation objects should already be locked,
1188 	 * since otherwise lockdep will be angered in radeon.
1189 	 */
1190 	if (!resv) {
1191 		locked = ww_mutex_trylock(&bo->resv->lock);
1192 		WARN_ON(!locked);
1193 	}
1194 
1195 	if (likely(!ret))
1196 		ret = ttm_bo_validate(bo, placement, interruptible, false);
1197 
1198 	if (unlikely(ret)) {
1199 		if (!resv)
1200 			ttm_bo_unreserve(bo);
1201 
1202 		ttm_bo_unref(&bo);
1203 		return ret;
1204 	}
1205 
1206 	if (resv && !(bo->mem.placement & TTM_PL_FLAG_NO_EVICT)) {
1207 		lockmgr(&bo->glob->lru_lock, LK_EXCLUSIVE);
1208 		ttm_bo_add_to_lru(bo);
1209 		lockmgr(&bo->glob->lru_lock, LK_RELEASE);
1210 	}
1211 
1212 	return ret;
1213 }
1214 EXPORT_SYMBOL(ttm_bo_init_reserved);
1215 
1216 int ttm_bo_init(struct ttm_bo_device *bdev,
1217 		struct ttm_buffer_object *bo,
1218 		unsigned long size,
1219 		enum ttm_bo_type type,
1220 		struct ttm_placement *placement,
1221 		uint32_t page_alignment,
1222 		bool interruptible,
1223 		struct vm_object *persistent_swap_storage,
1224 		size_t acc_size,
1225 		struct sg_table *sg,
1226 		struct reservation_object *resv,
1227 		void (*destroy) (struct ttm_buffer_object *))
1228 {
1229 	int ret;
1230 
1231 	ret = ttm_bo_init_reserved(bdev, bo, size, type, placement,
1232 				   page_alignment, interruptible,
1233 				   persistent_swap_storage, acc_size,
1234 				   sg, resv, destroy);
1235 	if (ret)
1236 		return ret;
1237 
1238 	if (!resv)
1239 		ttm_bo_unreserve(bo);
1240 
1241 	return 0;
1242 }
1243 EXPORT_SYMBOL(ttm_bo_init);
1244 
1245 size_t ttm_bo_acc_size(struct ttm_bo_device *bdev,
1246 		       unsigned long bo_size,
1247 		       unsigned struct_size)
1248 {
1249 	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1250 	size_t size = 0;
1251 
1252 	size += ttm_round_pot(struct_size);
1253 	size += ttm_round_pot(npages * sizeof(void *));
1254 	size += ttm_round_pot(sizeof(struct ttm_tt));
1255 	return size;
1256 }
1257 EXPORT_SYMBOL(ttm_bo_acc_size);
1258 
1259 size_t ttm_bo_dma_acc_size(struct ttm_bo_device *bdev,
1260 			   unsigned long bo_size,
1261 			   unsigned struct_size)
1262 {
1263 	unsigned npages = (PAGE_ALIGN(bo_size)) >> PAGE_SHIFT;
1264 	size_t size = 0;
1265 
1266 	size += ttm_round_pot(struct_size);
1267 	size += ttm_round_pot(npages * (2*sizeof(void *) + sizeof(dma_addr_t)));
1268 	size += ttm_round_pot(sizeof(struct ttm_dma_tt));
1269 	return size;
1270 }
1271 EXPORT_SYMBOL(ttm_bo_dma_acc_size);
1272 
1273 int ttm_bo_create(struct ttm_bo_device *bdev,
1274 			unsigned long size,
1275 			enum ttm_bo_type type,
1276 			struct ttm_placement *placement,
1277 			uint32_t page_alignment,
1278 			bool interruptible,
1279 			struct vm_object *persistent_swap_storage,
1280 			struct ttm_buffer_object **p_bo)
1281 {
1282 	struct ttm_buffer_object *bo;
1283 	size_t acc_size;
1284 	int ret;
1285 
1286 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
1287 	if (unlikely(bo == NULL))
1288 		return -ENOMEM;
1289 
1290 	acc_size = ttm_bo_acc_size(bdev, size, sizeof(struct ttm_buffer_object));
1291 	ret = ttm_bo_init(bdev, bo, size, type, placement, page_alignment,
1292 			  interruptible, persistent_swap_storage, acc_size,
1293 			  NULL, NULL, NULL);
1294 	if (likely(ret == 0))
1295 		*p_bo = bo;
1296 
1297 	return ret;
1298 }
1299 EXPORT_SYMBOL(ttm_bo_create);
1300 
1301 static int ttm_bo_force_list_clean(struct ttm_bo_device *bdev,
1302 				   unsigned mem_type)
1303 {
1304 	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1305 	struct ttm_bo_global *glob = bdev->glob;
1306 	struct dma_fence *fence;
1307 	int ret;
1308 	unsigned i;
1309 
1310 	/*
1311 	 * Can't use standard list traversal since we're unlocking.
1312 	 */
1313 
1314 	lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1315 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1316 		while (!list_empty(&man->lru[i])) {
1317 			lockmgr(&glob->lru_lock, LK_RELEASE);
1318 			ret = ttm_mem_evict_first(bdev, mem_type, NULL, false, false);
1319 			if (ret)
1320 				return ret;
1321 			lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1322 		}
1323 	}
1324 	lockmgr(&glob->lru_lock, LK_RELEASE);
1325 
1326 	lockmgr(&man->move_lock, LK_EXCLUSIVE);
1327 	fence = dma_fence_get(man->move);
1328 	lockmgr(&man->move_lock, LK_RELEASE);
1329 
1330 	if (fence) {
1331 		ret = dma_fence_wait(fence, false);
1332 		dma_fence_put(fence);
1333 		if (ret)
1334 			return ret;
1335 	}
1336 
1337 	return 0;
1338 }
1339 
1340 int ttm_bo_clean_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1341 {
1342 	struct ttm_mem_type_manager *man;
1343 	int ret = -EINVAL;
1344 
1345 	if (mem_type >= TTM_NUM_MEM_TYPES) {
1346 		pr_err("Illegal memory type %d\n", mem_type);
1347 		return ret;
1348 	}
1349 	man = &bdev->man[mem_type];
1350 
1351 	if (!man->has_type) {
1352 		pr_err("Trying to take down uninitialized memory manager type %u\n",
1353 		       mem_type);
1354 		return ret;
1355 	}
1356 
1357 	man->use_type = false;
1358 	man->has_type = false;
1359 
1360 	ret = 0;
1361 	if (mem_type > 0) {
1362 		ret = ttm_bo_force_list_clean(bdev, mem_type);
1363 		if (ret) {
1364 			pr_err("Cleanup eviction failed\n");
1365 			return ret;
1366 		}
1367 
1368 		ret = (*man->func->takedown)(man);
1369 	}
1370 
1371 	dma_fence_put(man->move);
1372 	man->move = NULL;
1373 
1374 	return ret;
1375 }
1376 EXPORT_SYMBOL(ttm_bo_clean_mm);
1377 
1378 int ttm_bo_evict_mm(struct ttm_bo_device *bdev, unsigned mem_type)
1379 {
1380 	struct ttm_mem_type_manager *man = &bdev->man[mem_type];
1381 
1382 	if (mem_type == 0 || mem_type >= TTM_NUM_MEM_TYPES) {
1383 		pr_err("Illegal memory manager memory type %u\n", mem_type);
1384 		return -EINVAL;
1385 	}
1386 
1387 	if (!man->has_type) {
1388 		pr_err("Memory type %u has not been initialized\n", mem_type);
1389 		return 0;
1390 	}
1391 
1392 	return ttm_bo_force_list_clean(bdev, mem_type);
1393 }
1394 EXPORT_SYMBOL(ttm_bo_evict_mm);
1395 
1396 int ttm_bo_init_mm(struct ttm_bo_device *bdev, unsigned type,
1397 			unsigned long p_size)
1398 {
1399 	int ret;
1400 	struct ttm_mem_type_manager *man;
1401 	unsigned i;
1402 
1403 	BUG_ON(type >= TTM_NUM_MEM_TYPES);
1404 	man = &bdev->man[type];
1405 	BUG_ON(man->has_type);
1406 	man->io_reserve_fastpath = true;
1407 	man->use_io_reserve_lru = false;
1408 	lockinit(&man->io_reserve_mutex, "ttmior", 0, 0);
1409 	lockinit(&man->move_lock, "ttmml", 0, 0);
1410 	INIT_LIST_HEAD(&man->io_reserve_lru);
1411 
1412 	ret = bdev->driver->init_mem_type(bdev, type, man);
1413 	if (ret)
1414 		return ret;
1415 	man->bdev = bdev;
1416 
1417 	if (type != TTM_PL_SYSTEM) {
1418 		ret = (*man->func->init)(man, p_size);
1419 		if (ret)
1420 			return ret;
1421 	}
1422 	man->has_type = true;
1423 	man->use_type = true;
1424 	man->size = p_size;
1425 
1426 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1427 		INIT_LIST_HEAD(&man->lru[i]);
1428 	man->move = NULL;
1429 
1430 	return 0;
1431 }
1432 EXPORT_SYMBOL(ttm_bo_init_mm);
1433 
1434 static void ttm_bo_global_kobj_release(struct kobject *kobj)
1435 {
1436 	struct ttm_bo_global *glob =
1437 		container_of(kobj, struct ttm_bo_global, kobj);
1438 
1439 	ttm_mem_unregister_shrink(glob->mem_glob, &glob->shrink);
1440 	__free_page(glob->dummy_read_page);
1441 	kfree(glob);
1442 }
1443 
1444 void ttm_bo_global_release(struct drm_global_reference *ref)
1445 {
1446 	struct ttm_bo_global *glob = ref->object;
1447 
1448 	kobject_del(&glob->kobj);
1449 	kobject_put(&glob->kobj);
1450 }
1451 EXPORT_SYMBOL(ttm_bo_global_release);
1452 
1453 int ttm_bo_global_init(struct drm_global_reference *ref)
1454 {
1455 	struct ttm_bo_global_ref *bo_ref =
1456 		container_of(ref, struct ttm_bo_global_ref, ref);
1457 	struct ttm_bo_global *glob = ref->object;
1458 	int ret;
1459 	unsigned i;
1460 
1461 	lockinit(&glob->device_list_mutex, "ttmdlm", 0, 0);
1462 	lockinit(&glob->lru_lock, "ttmlru", 0, 0);
1463 	glob->mem_glob = bo_ref->mem_glob;
1464 	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
1465 
1466 	if (unlikely(glob->dummy_read_page == NULL)) {
1467 		ret = -ENOMEM;
1468 		goto out_no_drp;
1469 	}
1470 
1471 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1472 		INIT_LIST_HEAD(&glob->swap_lru[i]);
1473 	INIT_LIST_HEAD(&glob->device_list);
1474 
1475 	ttm_mem_init_shrink(&glob->shrink, ttm_bo_swapout);
1476 	ret = ttm_mem_register_shrink(glob->mem_glob, &glob->shrink);
1477 	if (unlikely(ret != 0)) {
1478 		pr_err("Could not register buffer object swapout\n");
1479 		goto out_no_shrink;
1480 	}
1481 
1482 	atomic_set(&glob->bo_count, 0);
1483 
1484 	ret = kobject_init_and_add(
1485 		&glob->kobj, &ttm_bo_glob_kobj_type, ttm_get_kobj(), "buffer_objects");
1486 	if (unlikely(ret != 0))
1487 		kobject_put(&glob->kobj);
1488 	return ret;
1489 out_no_shrink:
1490 	__free_page(glob->dummy_read_page);
1491 out_no_drp:
1492 	kfree(glob);
1493 	return ret;
1494 }
1495 EXPORT_SYMBOL(ttm_bo_global_init);
1496 
1497 
1498 int ttm_bo_device_release(struct ttm_bo_device *bdev)
1499 {
1500 	int ret = 0;
1501 	unsigned i = TTM_NUM_MEM_TYPES;
1502 	struct ttm_mem_type_manager *man;
1503 	struct ttm_bo_global *glob = bdev->glob;
1504 
1505 	while (i--) {
1506 		man = &bdev->man[i];
1507 		if (man->has_type) {
1508 			man->use_type = false;
1509 			if ((i != TTM_PL_SYSTEM) && ttm_bo_clean_mm(bdev, i)) {
1510 				ret = -EBUSY;
1511 				pr_err("DRM memory manager type %d is not clean\n",
1512 				       i);
1513 			}
1514 			man->has_type = false;
1515 		}
1516 	}
1517 
1518 	mutex_lock(&glob->device_list_mutex);
1519 	list_del(&bdev->device_list);
1520 	mutex_unlock(&glob->device_list_mutex);
1521 
1522 	cancel_delayed_work_sync(&bdev->wq);
1523 
1524 	while (ttm_bo_delayed_delete(bdev, true))
1525 		;
1526 
1527 	lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1528 	if (list_empty(&bdev->ddestroy))
1529 		TTM_DEBUG("Delayed destroy list was clean\n");
1530 
1531 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
1532 		if (list_empty(&bdev->man[0].lru[0]))
1533 			TTM_DEBUG("Swap list %d was clean\n", i);
1534 	lockmgr(&glob->lru_lock, LK_RELEASE);
1535 
1536 	drm_vma_offset_manager_destroy(&bdev->vma_manager);
1537 
1538 	return ret;
1539 }
1540 EXPORT_SYMBOL(ttm_bo_device_release);
1541 
1542 int ttm_bo_device_init(struct ttm_bo_device *bdev,
1543 		       struct ttm_bo_global *glob,
1544 		       struct ttm_bo_driver *driver,
1545 		       struct address_space *mapping,
1546 		       uint64_t file_page_offset,
1547 		       bool need_dma32)
1548 {
1549 	int ret = -EINVAL;
1550 
1551 	bdev->driver = driver;
1552 
1553 	memset(bdev->man, 0, sizeof(bdev->man));
1554 
1555 	/*
1556 	 * Initialize the system memory buffer type.
1557 	 * Other types need to be driver / IOCTL initialized.
1558 	 */
1559 	ret = ttm_bo_init_mm(bdev, TTM_PL_SYSTEM, 0);
1560 	if (unlikely(ret != 0))
1561 		goto out_no_sys;
1562 
1563 	drm_vma_offset_manager_init(&bdev->vma_manager, file_page_offset,
1564 				    0x10000000);
1565 	INIT_DELAYED_WORK(&bdev->wq, ttm_bo_delayed_workqueue);
1566 	INIT_LIST_HEAD(&bdev->ddestroy);
1567 	/*
1568 	 * XXX DRAGONFLY - dev_mapping NULL atm, find other XXX DRAGONFLY
1569 	 * lines and fix when it no longer is in later API change.
1570 	 */
1571 	bdev->dev_mapping = mapping;
1572 	bdev->glob = glob;
1573 	bdev->need_dma32 = need_dma32;
1574 	mutex_lock(&glob->device_list_mutex);
1575 	list_add_tail(&bdev->device_list, &glob->device_list);
1576 	mutex_unlock(&glob->device_list_mutex);
1577 
1578 	return 0;
1579 out_no_sys:
1580 	return ret;
1581 }
1582 EXPORT_SYMBOL(ttm_bo_device_init);
1583 
1584 /*
1585  * buffer object vm functions.
1586  */
1587 
1588 bool ttm_mem_reg_is_pci(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
1589 {
1590 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
1591 
1592 	if (!(man->flags & TTM_MEMTYPE_FLAG_FIXED)) {
1593 		if (mem->mem_type == TTM_PL_SYSTEM)
1594 			return false;
1595 
1596 		if (man->flags & TTM_MEMTYPE_FLAG_CMA)
1597 			return false;
1598 
1599 		if (mem->placement & TTM_PL_FLAG_CACHED)
1600 			return false;
1601 	}
1602 	return true;
1603 }
1604 
1605 #ifdef __DragonFly__
1606 
1607 /*
1608  * XXX DRAGONFLY - device_mapping not yet implemented so
1609  * file_mapping is basically always NULL.  We have to properly
1610  * release the mmap, etc.
1611 */
1612 void ttm_bo_release_mmap(struct ttm_buffer_object *bo);
1613 
1614 /**
1615  * drm_vma_node_unmap() - Unmap offset node
1616  * @node: Offset node
1617  * @file_mapping: Address space to unmap @node from
1618  *
1619  * Unmap all userspace mappings for a given offset node. The mappings must be
1620  * associated with the @file_mapping address-space. If no offset exists or
1621  * the address-space is invalid, nothing is done.
1622  *
1623  * This call is unlocked. The caller must guarantee that drm_vma_offset_remove()
1624  * is not called on this node concurrently.
1625  */
1626 static inline void drm_vma_node_unmap(struct drm_vma_offset_node *node,
1627 				      struct address_space *file_mapping)
1628 {
1629 	struct ttm_buffer_object *bo = container_of(node, struct ttm_buffer_object, vma_node);
1630 
1631 	if (drm_vma_node_has_offset(node))
1632 		unmap_mapping_range(file_mapping,
1633 				    drm_vma_node_offset_addr(node),
1634 				    drm_vma_node_size(node) << PAGE_SHIFT, 1);
1635 	ttm_bo_release_mmap(bo);
1636 }
1637 #endif
1638 
1639 void ttm_bo_unmap_virtual_locked(struct ttm_buffer_object *bo)
1640 {
1641 	struct ttm_bo_device *bdev = bo->bdev;
1642 
1643 	drm_vma_node_unmap(&bo->vma_node, bdev->dev_mapping);
1644 	ttm_mem_io_free_vm(bo);
1645 }
1646 
1647 void ttm_bo_unmap_virtual(struct ttm_buffer_object *bo)
1648 {
1649 	struct ttm_bo_device *bdev = bo->bdev;
1650 	struct ttm_mem_type_manager *man = &bdev->man[bo->mem.mem_type];
1651 
1652 	ttm_mem_io_lock(man, false);
1653 	ttm_bo_unmap_virtual_locked(bo);
1654 	ttm_mem_io_unlock(man);
1655 }
1656 
1657 
1658 EXPORT_SYMBOL(ttm_bo_unmap_virtual);
1659 
1660 int ttm_bo_wait(struct ttm_buffer_object *bo,
1661 		bool interruptible, bool no_wait)
1662 {
1663 	long timeout = 15 * HZ;
1664 
1665 	if (no_wait) {
1666 		if (reservation_object_test_signaled_rcu(bo->resv, true))
1667 			return 0;
1668 		else
1669 			return -EBUSY;
1670 	}
1671 
1672 	timeout = reservation_object_wait_timeout_rcu(bo->resv, true,
1673 						      interruptible, timeout);
1674 	if (timeout < 0)
1675 		return timeout;
1676 
1677 	if (timeout == 0)
1678 		return -EBUSY;
1679 
1680 	reservation_object_add_excl_fence(bo->resv, NULL);
1681 	return 0;
1682 }
1683 EXPORT_SYMBOL(ttm_bo_wait);
1684 
1685 int ttm_bo_synccpu_write_grab(struct ttm_buffer_object *bo, bool no_wait)
1686 {
1687 	int ret = 0;
1688 
1689 	/*
1690 	 * Using ttm_bo_reserve makes sure the lru lists are updated.
1691 	 */
1692 
1693 	ret = ttm_bo_reserve(bo, true, no_wait, NULL);
1694 	if (unlikely(ret != 0))
1695 		return ret;
1696 	ret = ttm_bo_wait(bo, true, no_wait);
1697 	if (likely(ret == 0))
1698 		atomic_inc(&bo->cpu_writers);
1699 	ttm_bo_unreserve(bo);
1700 	return ret;
1701 }
1702 EXPORT_SYMBOL(ttm_bo_synccpu_write_grab);
1703 
1704 void ttm_bo_synccpu_write_release(struct ttm_buffer_object *bo)
1705 {
1706 	atomic_dec(&bo->cpu_writers);
1707 }
1708 EXPORT_SYMBOL(ttm_bo_synccpu_write_release);
1709 
1710 /**
1711  * A buffer object shrink method that tries to swap out the first
1712  * buffer object on the bo_global::swap_lru list.
1713  */
1714 
1715 static int ttm_bo_swapout(struct ttm_mem_shrink *shrink)
1716 {
1717 	struct ttm_bo_global *glob =
1718 	    container_of(shrink, struct ttm_bo_global, shrink);
1719 	struct ttm_buffer_object *bo;
1720 	int ret = -EBUSY;
1721 	unsigned i;
1722 
1723 	lockmgr(&glob->lru_lock, LK_EXCLUSIVE);
1724 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
1725 		list_for_each_entry(bo, &glob->swap_lru[i], swap) {
1726 			ret = __ttm_bo_reserve(bo, false, true, NULL);
1727 			if (!ret)
1728 				break;
1729 		}
1730 		if (!ret)
1731 			break;
1732 	}
1733 
1734 	if (ret) {
1735 		lockmgr(&glob->lru_lock, LK_RELEASE);
1736 		return ret;
1737 	}
1738 
1739 	kref_get(&bo->list_kref);
1740 
1741 	if (!list_empty(&bo->ddestroy)) {
1742 		ret = ttm_bo_cleanup_refs_and_unlock(bo, false, false);
1743 		kref_put(&bo->list_kref, ttm_bo_release_list);
1744 		return ret;
1745 	}
1746 
1747 	ttm_bo_del_from_lru(bo);
1748 	lockmgr(&glob->lru_lock, LK_RELEASE);
1749 
1750 	/**
1751 	 * Move to system cached
1752 	 */
1753 
1754 	if (bo->mem.mem_type != TTM_PL_SYSTEM ||
1755 	    bo->ttm->caching_state != tt_cached) {
1756 		struct ttm_mem_reg evict_mem;
1757 
1758 		evict_mem = bo->mem;
1759 		evict_mem.mm_node = NULL;
1760 		evict_mem.placement = TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED;
1761 		evict_mem.mem_type = TTM_PL_SYSTEM;
1762 
1763 		ret = ttm_bo_handle_move_mem(bo, &evict_mem, true,
1764 					     false, false);
1765 		if (unlikely(ret != 0))
1766 			goto out;
1767 	}
1768 
1769 	/**
1770 	 * Make sure BO is idle.
1771 	 */
1772 
1773 	ret = ttm_bo_wait(bo, false, false);
1774 	if (unlikely(ret != 0))
1775 		goto out;
1776 
1777 	ttm_bo_unmap_virtual(bo);
1778 
1779 	/**
1780 	 * Swap out. Buffer will be swapped in again as soon as
1781 	 * anyone tries to access a ttm page.
1782 	 */
1783 
1784 	if (bo->bdev->driver->swap_notify)
1785 		bo->bdev->driver->swap_notify(bo);
1786 
1787 	ret = ttm_tt_swapout(bo->ttm, bo->persistent_swap_storage);
1788 out:
1789 
1790 	/**
1791 	 *
1792 	 * Unreserve without putting on LRU to avoid swapping out an
1793 	 * already swapped buffer.
1794 	 */
1795 
1796 	__ttm_bo_unreserve(bo);
1797 	kref_put(&bo->list_kref, ttm_bo_release_list);
1798 	return ret;
1799 }
1800 
1801 void ttm_bo_swapout_all(struct ttm_bo_device *bdev)
1802 {
1803 	while (ttm_bo_swapout(&bdev->glob->shrink) == 0)
1804 		;
1805 }
1806 EXPORT_SYMBOL(ttm_bo_swapout_all);
1807 
1808 /**
1809  * ttm_bo_wait_unreserved - interruptible wait for a buffer object to become
1810  * unreserved
1811  *
1812  * @bo: Pointer to buffer
1813  */
1814 int ttm_bo_wait_unreserved(struct ttm_buffer_object *bo)
1815 {
1816 	int ret;
1817 
1818 	/*
1819 	 * In the absense of a wait_unlocked API,
1820 	 * Use the bo::wu_mutex to avoid triggering livelocks due to
1821 	 * concurrent use of this function. Note that this use of
1822 	 * bo::wu_mutex can go away if we change locking order to
1823 	 * mmap_sem -> bo::reserve.
1824 	 */
1825 	ret = mutex_lock_interruptible(&bo->wu_mutex);
1826 	if (unlikely(ret != 0))
1827 		return -ERESTARTSYS;
1828 	if (!ww_mutex_is_locked(&bo->resv->lock))
1829 		goto out_unlock;
1830 	ret = __ttm_bo_reserve(bo, true, false, NULL);
1831 	if (unlikely(ret != 0))
1832 		goto out_unlock;
1833 	__ttm_bo_unreserve(bo);
1834 
1835 out_unlock:
1836 	mutex_unlock(&bo->wu_mutex);
1837 	return ret;
1838 }
1839