xref: /linux/drivers/gpu/drm/drm_buddy.c (revision bf36793fa260cb68cc817f311f1f683788261796)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include <linux/kmemleak.h>
7 #include <linux/module.h>
8 #include <linux/sizes.h>
9 
10 #include <drm/drm_buddy.h>
11 
12 static struct kmem_cache *slab_blocks;
13 
14 static struct drm_buddy_block *drm_block_alloc(struct drm_buddy *mm,
15 					       struct drm_buddy_block *parent,
16 					       unsigned int order,
17 					       u64 offset)
18 {
19 	struct drm_buddy_block *block;
20 
21 	BUG_ON(order > DRM_BUDDY_MAX_ORDER);
22 
23 	block = kmem_cache_zalloc(slab_blocks, GFP_KERNEL);
24 	if (!block)
25 		return NULL;
26 
27 	block->header = offset;
28 	block->header |= order;
29 	block->parent = parent;
30 
31 	BUG_ON(block->header & DRM_BUDDY_HEADER_UNUSED);
32 	return block;
33 }
34 
35 static void drm_block_free(struct drm_buddy *mm,
36 			   struct drm_buddy_block *block)
37 {
38 	kmem_cache_free(slab_blocks, block);
39 }
40 
41 static void list_insert_sorted(struct drm_buddy *mm,
42 			       struct drm_buddy_block *block)
43 {
44 	struct drm_buddy_block *node;
45 	struct list_head *head;
46 
47 	head = &mm->free_list[drm_buddy_block_order(block)];
48 	if (list_empty(head)) {
49 		list_add(&block->link, head);
50 		return;
51 	}
52 
53 	list_for_each_entry(node, head, link)
54 		if (drm_buddy_block_offset(block) < drm_buddy_block_offset(node))
55 			break;
56 
57 	__list_add(&block->link, node->link.prev, &node->link);
58 }
59 
60 static void clear_reset(struct drm_buddy_block *block)
61 {
62 	block->header &= ~DRM_BUDDY_HEADER_CLEAR;
63 }
64 
65 static void mark_cleared(struct drm_buddy_block *block)
66 {
67 	block->header |= DRM_BUDDY_HEADER_CLEAR;
68 }
69 
70 static void mark_allocated(struct drm_buddy_block *block)
71 {
72 	block->header &= ~DRM_BUDDY_HEADER_STATE;
73 	block->header |= DRM_BUDDY_ALLOCATED;
74 
75 	list_del(&block->link);
76 }
77 
78 static void mark_free(struct drm_buddy *mm,
79 		      struct drm_buddy_block *block)
80 {
81 	block->header &= ~DRM_BUDDY_HEADER_STATE;
82 	block->header |= DRM_BUDDY_FREE;
83 
84 	list_insert_sorted(mm, block);
85 }
86 
87 static void mark_split(struct drm_buddy_block *block)
88 {
89 	block->header &= ~DRM_BUDDY_HEADER_STATE;
90 	block->header |= DRM_BUDDY_SPLIT;
91 
92 	list_del(&block->link);
93 }
94 
95 static inline bool overlaps(u64 s1, u64 e1, u64 s2, u64 e2)
96 {
97 	return s1 <= e2 && e1 >= s2;
98 }
99 
100 static inline bool contains(u64 s1, u64 e1, u64 s2, u64 e2)
101 {
102 	return s1 <= s2 && e1 >= e2;
103 }
104 
105 static struct drm_buddy_block *
106 __get_buddy(struct drm_buddy_block *block)
107 {
108 	struct drm_buddy_block *parent;
109 
110 	parent = block->parent;
111 	if (!parent)
112 		return NULL;
113 
114 	if (parent->left == block)
115 		return parent->right;
116 
117 	return parent->left;
118 }
119 
120 static unsigned int __drm_buddy_free(struct drm_buddy *mm,
121 				     struct drm_buddy_block *block,
122 				     bool force_merge)
123 {
124 	struct drm_buddy_block *parent;
125 	unsigned int order;
126 
127 	while ((parent = block->parent)) {
128 		struct drm_buddy_block *buddy;
129 
130 		buddy = __get_buddy(block);
131 
132 		if (!drm_buddy_block_is_free(buddy))
133 			break;
134 
135 		if (!force_merge) {
136 			/*
137 			 * Check the block and its buddy clear state and exit
138 			 * the loop if they both have the dissimilar state.
139 			 */
140 			if (drm_buddy_block_is_clear(block) !=
141 			    drm_buddy_block_is_clear(buddy))
142 				break;
143 
144 			if (drm_buddy_block_is_clear(block))
145 				mark_cleared(parent);
146 		}
147 
148 		list_del(&buddy->link);
149 		if (force_merge && drm_buddy_block_is_clear(buddy))
150 			mm->clear_avail -= drm_buddy_block_size(mm, buddy);
151 
152 		drm_block_free(mm, block);
153 		drm_block_free(mm, buddy);
154 
155 		block = parent;
156 	}
157 
158 	order = drm_buddy_block_order(block);
159 	mark_free(mm, block);
160 
161 	return order;
162 }
163 
164 static int __force_merge(struct drm_buddy *mm,
165 			 u64 start,
166 			 u64 end,
167 			 unsigned int min_order)
168 {
169 	unsigned int order;
170 	int i;
171 
172 	if (!min_order)
173 		return -ENOMEM;
174 
175 	if (min_order > mm->max_order)
176 		return -EINVAL;
177 
178 	for (i = min_order - 1; i >= 0; i--) {
179 		struct drm_buddy_block *block, *prev;
180 
181 		list_for_each_entry_safe_reverse(block, prev, &mm->free_list[i], link) {
182 			struct drm_buddy_block *buddy;
183 			u64 block_start, block_end;
184 
185 			if (!block->parent)
186 				continue;
187 
188 			block_start = drm_buddy_block_offset(block);
189 			block_end = block_start + drm_buddy_block_size(mm, block) - 1;
190 
191 			if (!contains(start, end, block_start, block_end))
192 				continue;
193 
194 			buddy = __get_buddy(block);
195 			if (!drm_buddy_block_is_free(buddy))
196 				continue;
197 
198 			WARN_ON(drm_buddy_block_is_clear(block) ==
199 				drm_buddy_block_is_clear(buddy));
200 
201 			/*
202 			 * If the prev block is same as buddy, don't access the
203 			 * block in the next iteration as we would free the
204 			 * buddy block as part of the free function.
205 			 */
206 			if (prev == buddy)
207 				prev = list_prev_entry(prev, link);
208 
209 			list_del(&block->link);
210 			if (drm_buddy_block_is_clear(block))
211 				mm->clear_avail -= drm_buddy_block_size(mm, block);
212 
213 			order = __drm_buddy_free(mm, block, true);
214 			if (order >= min_order)
215 				return 0;
216 		}
217 	}
218 
219 	return -ENOMEM;
220 }
221 
222 /**
223  * drm_buddy_init - init memory manager
224  *
225  * @mm: DRM buddy manager to initialize
226  * @size: size in bytes to manage
227  * @chunk_size: minimum page size in bytes for our allocations
228  *
229  * Initializes the memory manager and its resources.
230  *
231  * Returns:
232  * 0 on success, error code on failure.
233  */
234 int drm_buddy_init(struct drm_buddy *mm, u64 size, u64 chunk_size)
235 {
236 	unsigned int i;
237 	u64 offset;
238 
239 	if (size < chunk_size)
240 		return -EINVAL;
241 
242 	if (chunk_size < SZ_4K)
243 		return -EINVAL;
244 
245 	if (!is_power_of_2(chunk_size))
246 		return -EINVAL;
247 
248 	size = round_down(size, chunk_size);
249 
250 	mm->size = size;
251 	mm->avail = size;
252 	mm->clear_avail = 0;
253 	mm->chunk_size = chunk_size;
254 	mm->max_order = ilog2(size) - ilog2(chunk_size);
255 
256 	BUG_ON(mm->max_order > DRM_BUDDY_MAX_ORDER);
257 
258 	mm->free_list = kmalloc_array(mm->max_order + 1,
259 				      sizeof(struct list_head),
260 				      GFP_KERNEL);
261 	if (!mm->free_list)
262 		return -ENOMEM;
263 
264 	for (i = 0; i <= mm->max_order; ++i)
265 		INIT_LIST_HEAD(&mm->free_list[i]);
266 
267 	mm->n_roots = hweight64(size);
268 
269 	mm->roots = kmalloc_array(mm->n_roots,
270 				  sizeof(struct drm_buddy_block *),
271 				  GFP_KERNEL);
272 	if (!mm->roots)
273 		goto out_free_list;
274 
275 	offset = 0;
276 	i = 0;
277 
278 	/*
279 	 * Split into power-of-two blocks, in case we are given a size that is
280 	 * not itself a power-of-two.
281 	 */
282 	do {
283 		struct drm_buddy_block *root;
284 		unsigned int order;
285 		u64 root_size;
286 
287 		order = ilog2(size) - ilog2(chunk_size);
288 		root_size = chunk_size << order;
289 
290 		root = drm_block_alloc(mm, NULL, order, offset);
291 		if (!root)
292 			goto out_free_roots;
293 
294 		mark_free(mm, root);
295 
296 		BUG_ON(i > mm->max_order);
297 		BUG_ON(drm_buddy_block_size(mm, root) < chunk_size);
298 
299 		mm->roots[i] = root;
300 
301 		offset += root_size;
302 		size -= root_size;
303 		i++;
304 	} while (size);
305 
306 	return 0;
307 
308 out_free_roots:
309 	while (i--)
310 		drm_block_free(mm, mm->roots[i]);
311 	kfree(mm->roots);
312 out_free_list:
313 	kfree(mm->free_list);
314 	return -ENOMEM;
315 }
316 EXPORT_SYMBOL(drm_buddy_init);
317 
318 /**
319  * drm_buddy_fini - tear down the memory manager
320  *
321  * @mm: DRM buddy manager to free
322  *
323  * Cleanup memory manager resources and the freelist
324  */
325 void drm_buddy_fini(struct drm_buddy *mm)
326 {
327 	u64 root_size, size;
328 	unsigned int order;
329 	int i;
330 
331 	size = mm->size;
332 
333 	for (i = 0; i < mm->n_roots; ++i) {
334 		order = ilog2(size) - ilog2(mm->chunk_size);
335 		__force_merge(mm, 0, size, order);
336 
337 		WARN_ON(!drm_buddy_block_is_free(mm->roots[i]));
338 		drm_block_free(mm, mm->roots[i]);
339 
340 		root_size = mm->chunk_size << order;
341 		size -= root_size;
342 	}
343 
344 	WARN_ON(mm->avail != mm->size);
345 
346 	kfree(mm->roots);
347 	kfree(mm->free_list);
348 }
349 EXPORT_SYMBOL(drm_buddy_fini);
350 
351 static int split_block(struct drm_buddy *mm,
352 		       struct drm_buddy_block *block)
353 {
354 	unsigned int block_order = drm_buddy_block_order(block) - 1;
355 	u64 offset = drm_buddy_block_offset(block);
356 
357 	BUG_ON(!drm_buddy_block_is_free(block));
358 	BUG_ON(!drm_buddy_block_order(block));
359 
360 	block->left = drm_block_alloc(mm, block, block_order, offset);
361 	if (!block->left)
362 		return -ENOMEM;
363 
364 	block->right = drm_block_alloc(mm, block, block_order,
365 				       offset + (mm->chunk_size << block_order));
366 	if (!block->right) {
367 		drm_block_free(mm, block->left);
368 		return -ENOMEM;
369 	}
370 
371 	mark_free(mm, block->left);
372 	mark_free(mm, block->right);
373 
374 	if (drm_buddy_block_is_clear(block)) {
375 		mark_cleared(block->left);
376 		mark_cleared(block->right);
377 		clear_reset(block);
378 	}
379 
380 	mark_split(block);
381 
382 	return 0;
383 }
384 
385 /**
386  * drm_get_buddy - get buddy address
387  *
388  * @block: DRM buddy block
389  *
390  * Returns the corresponding buddy block for @block, or NULL
391  * if this is a root block and can't be merged further.
392  * Requires some kind of locking to protect against
393  * any concurrent allocate and free operations.
394  */
395 struct drm_buddy_block *
396 drm_get_buddy(struct drm_buddy_block *block)
397 {
398 	return __get_buddy(block);
399 }
400 EXPORT_SYMBOL(drm_get_buddy);
401 
402 /**
403  * drm_buddy_free_block - free a block
404  *
405  * @mm: DRM buddy manager
406  * @block: block to be freed
407  */
408 void drm_buddy_free_block(struct drm_buddy *mm,
409 			  struct drm_buddy_block *block)
410 {
411 	BUG_ON(!drm_buddy_block_is_allocated(block));
412 	mm->avail += drm_buddy_block_size(mm, block);
413 	if (drm_buddy_block_is_clear(block))
414 		mm->clear_avail += drm_buddy_block_size(mm, block);
415 
416 	__drm_buddy_free(mm, block, false);
417 }
418 EXPORT_SYMBOL(drm_buddy_free_block);
419 
420 static void __drm_buddy_free_list(struct drm_buddy *mm,
421 				  struct list_head *objects,
422 				  bool mark_clear,
423 				  bool mark_dirty)
424 {
425 	struct drm_buddy_block *block, *on;
426 
427 	WARN_ON(mark_dirty && mark_clear);
428 
429 	list_for_each_entry_safe(block, on, objects, link) {
430 		if (mark_clear)
431 			mark_cleared(block);
432 		else if (mark_dirty)
433 			clear_reset(block);
434 		drm_buddy_free_block(mm, block);
435 		cond_resched();
436 	}
437 	INIT_LIST_HEAD(objects);
438 }
439 
440 static void drm_buddy_free_list_internal(struct drm_buddy *mm,
441 					 struct list_head *objects)
442 {
443 	/*
444 	 * Don't touch the clear/dirty bit, since allocation is still internal
445 	 * at this point. For example we might have just failed part of the
446 	 * allocation.
447 	 */
448 	__drm_buddy_free_list(mm, objects, false, false);
449 }
450 
451 /**
452  * drm_buddy_free_list - free blocks
453  *
454  * @mm: DRM buddy manager
455  * @objects: input list head to free blocks
456  * @flags: optional flags like DRM_BUDDY_CLEARED
457  */
458 void drm_buddy_free_list(struct drm_buddy *mm,
459 			 struct list_head *objects,
460 			 unsigned int flags)
461 {
462 	bool mark_clear = flags & DRM_BUDDY_CLEARED;
463 
464 	__drm_buddy_free_list(mm, objects, mark_clear, !mark_clear);
465 }
466 EXPORT_SYMBOL(drm_buddy_free_list);
467 
468 static bool block_incompatible(struct drm_buddy_block *block, unsigned int flags)
469 {
470 	bool needs_clear = flags & DRM_BUDDY_CLEAR_ALLOCATION;
471 
472 	return needs_clear != drm_buddy_block_is_clear(block);
473 }
474 
475 static struct drm_buddy_block *
476 __alloc_range_bias(struct drm_buddy *mm,
477 		   u64 start, u64 end,
478 		   unsigned int order,
479 		   unsigned long flags,
480 		   bool fallback)
481 {
482 	u64 req_size = mm->chunk_size << order;
483 	struct drm_buddy_block *block;
484 	struct drm_buddy_block *buddy;
485 	LIST_HEAD(dfs);
486 	int err;
487 	int i;
488 
489 	end = end - 1;
490 
491 	for (i = 0; i < mm->n_roots; ++i)
492 		list_add_tail(&mm->roots[i]->tmp_link, &dfs);
493 
494 	do {
495 		u64 block_start;
496 		u64 block_end;
497 
498 		block = list_first_entry_or_null(&dfs,
499 						 struct drm_buddy_block,
500 						 tmp_link);
501 		if (!block)
502 			break;
503 
504 		list_del(&block->tmp_link);
505 
506 		if (drm_buddy_block_order(block) < order)
507 			continue;
508 
509 		block_start = drm_buddy_block_offset(block);
510 		block_end = block_start + drm_buddy_block_size(mm, block) - 1;
511 
512 		if (!overlaps(start, end, block_start, block_end))
513 			continue;
514 
515 		if (drm_buddy_block_is_allocated(block))
516 			continue;
517 
518 		if (block_start < start || block_end > end) {
519 			u64 adjusted_start = max(block_start, start);
520 			u64 adjusted_end = min(block_end, end);
521 
522 			if (round_down(adjusted_end + 1, req_size) <=
523 			    round_up(adjusted_start, req_size))
524 				continue;
525 		}
526 
527 		if (!fallback && block_incompatible(block, flags))
528 			continue;
529 
530 		if (contains(start, end, block_start, block_end) &&
531 		    order == drm_buddy_block_order(block)) {
532 			/*
533 			 * Find the free block within the range.
534 			 */
535 			if (drm_buddy_block_is_free(block))
536 				return block;
537 
538 			continue;
539 		}
540 
541 		if (!drm_buddy_block_is_split(block)) {
542 			err = split_block(mm, block);
543 			if (unlikely(err))
544 				goto err_undo;
545 		}
546 
547 		list_add(&block->right->tmp_link, &dfs);
548 		list_add(&block->left->tmp_link, &dfs);
549 	} while (1);
550 
551 	return ERR_PTR(-ENOSPC);
552 
553 err_undo:
554 	/*
555 	 * We really don't want to leave around a bunch of split blocks, since
556 	 * bigger is better, so make sure we merge everything back before we
557 	 * free the allocated blocks.
558 	 */
559 	buddy = __get_buddy(block);
560 	if (buddy &&
561 	    (drm_buddy_block_is_free(block) &&
562 	     drm_buddy_block_is_free(buddy)))
563 		__drm_buddy_free(mm, block, false);
564 	return ERR_PTR(err);
565 }
566 
567 static struct drm_buddy_block *
568 __drm_buddy_alloc_range_bias(struct drm_buddy *mm,
569 			     u64 start, u64 end,
570 			     unsigned int order,
571 			     unsigned long flags)
572 {
573 	struct drm_buddy_block *block;
574 	bool fallback = false;
575 
576 	block = __alloc_range_bias(mm, start, end, order,
577 				   flags, fallback);
578 	if (IS_ERR(block))
579 		return __alloc_range_bias(mm, start, end, order,
580 					  flags, !fallback);
581 
582 	return block;
583 }
584 
585 static struct drm_buddy_block *
586 get_maxblock(struct drm_buddy *mm, unsigned int order,
587 	     unsigned long flags)
588 {
589 	struct drm_buddy_block *max_block = NULL, *block = NULL;
590 	unsigned int i;
591 
592 	for (i = order; i <= mm->max_order; ++i) {
593 		struct drm_buddy_block *tmp_block;
594 
595 		list_for_each_entry_reverse(tmp_block, &mm->free_list[i], link) {
596 			if (block_incompatible(tmp_block, flags))
597 				continue;
598 
599 			block = tmp_block;
600 			break;
601 		}
602 
603 		if (!block)
604 			continue;
605 
606 		if (!max_block) {
607 			max_block = block;
608 			continue;
609 		}
610 
611 		if (drm_buddy_block_offset(block) >
612 		    drm_buddy_block_offset(max_block)) {
613 			max_block = block;
614 		}
615 	}
616 
617 	return max_block;
618 }
619 
620 static struct drm_buddy_block *
621 alloc_from_freelist(struct drm_buddy *mm,
622 		    unsigned int order,
623 		    unsigned long flags)
624 {
625 	struct drm_buddy_block *block = NULL;
626 	unsigned int tmp;
627 	int err;
628 
629 	if (flags & DRM_BUDDY_TOPDOWN_ALLOCATION) {
630 		block = get_maxblock(mm, order, flags);
631 		if (block)
632 			/* Store the obtained block order */
633 			tmp = drm_buddy_block_order(block);
634 	} else {
635 		for (tmp = order; tmp <= mm->max_order; ++tmp) {
636 			struct drm_buddy_block *tmp_block;
637 
638 			list_for_each_entry_reverse(tmp_block, &mm->free_list[tmp], link) {
639 				if (block_incompatible(tmp_block, flags))
640 					continue;
641 
642 				block = tmp_block;
643 				break;
644 			}
645 
646 			if (block)
647 				break;
648 		}
649 	}
650 
651 	if (!block) {
652 		/* Fallback method */
653 		for (tmp = order; tmp <= mm->max_order; ++tmp) {
654 			if (!list_empty(&mm->free_list[tmp])) {
655 				block = list_last_entry(&mm->free_list[tmp],
656 							struct drm_buddy_block,
657 							link);
658 				if (block)
659 					break;
660 			}
661 		}
662 
663 		if (!block)
664 			return ERR_PTR(-ENOSPC);
665 	}
666 
667 	BUG_ON(!drm_buddy_block_is_free(block));
668 
669 	while (tmp != order) {
670 		err = split_block(mm, block);
671 		if (unlikely(err))
672 			goto err_undo;
673 
674 		block = block->right;
675 		tmp--;
676 	}
677 	return block;
678 
679 err_undo:
680 	if (tmp != order)
681 		__drm_buddy_free(mm, block, false);
682 	return ERR_PTR(err);
683 }
684 
685 static int __alloc_range(struct drm_buddy *mm,
686 			 struct list_head *dfs,
687 			 u64 start, u64 size,
688 			 struct list_head *blocks,
689 			 u64 *total_allocated_on_err)
690 {
691 	struct drm_buddy_block *block;
692 	struct drm_buddy_block *buddy;
693 	u64 total_allocated = 0;
694 	LIST_HEAD(allocated);
695 	u64 end;
696 	int err;
697 
698 	end = start + size - 1;
699 
700 	do {
701 		u64 block_start;
702 		u64 block_end;
703 
704 		block = list_first_entry_or_null(dfs,
705 						 struct drm_buddy_block,
706 						 tmp_link);
707 		if (!block)
708 			break;
709 
710 		list_del(&block->tmp_link);
711 
712 		block_start = drm_buddy_block_offset(block);
713 		block_end = block_start + drm_buddy_block_size(mm, block) - 1;
714 
715 		if (!overlaps(start, end, block_start, block_end))
716 			continue;
717 
718 		if (drm_buddy_block_is_allocated(block)) {
719 			err = -ENOSPC;
720 			goto err_free;
721 		}
722 
723 		if (contains(start, end, block_start, block_end)) {
724 			if (drm_buddy_block_is_free(block)) {
725 				mark_allocated(block);
726 				total_allocated += drm_buddy_block_size(mm, block);
727 				mm->avail -= drm_buddy_block_size(mm, block);
728 				if (drm_buddy_block_is_clear(block))
729 					mm->clear_avail -= drm_buddy_block_size(mm, block);
730 				list_add_tail(&block->link, &allocated);
731 				continue;
732 			} else if (!mm->clear_avail) {
733 				err = -ENOSPC;
734 				goto err_free;
735 			}
736 		}
737 
738 		if (!drm_buddy_block_is_split(block)) {
739 			err = split_block(mm, block);
740 			if (unlikely(err))
741 				goto err_undo;
742 		}
743 
744 		list_add(&block->right->tmp_link, dfs);
745 		list_add(&block->left->tmp_link, dfs);
746 	} while (1);
747 
748 	if (total_allocated < size) {
749 		err = -ENOSPC;
750 		goto err_free;
751 	}
752 
753 	list_splice_tail(&allocated, blocks);
754 
755 	return 0;
756 
757 err_undo:
758 	/*
759 	 * We really don't want to leave around a bunch of split blocks, since
760 	 * bigger is better, so make sure we merge everything back before we
761 	 * free the allocated blocks.
762 	 */
763 	buddy = __get_buddy(block);
764 	if (buddy &&
765 	    (drm_buddy_block_is_free(block) &&
766 	     drm_buddy_block_is_free(buddy)))
767 		__drm_buddy_free(mm, block, false);
768 
769 err_free:
770 	if (err == -ENOSPC && total_allocated_on_err) {
771 		list_splice_tail(&allocated, blocks);
772 		*total_allocated_on_err = total_allocated;
773 	} else {
774 		drm_buddy_free_list_internal(mm, &allocated);
775 	}
776 
777 	return err;
778 }
779 
780 static int __drm_buddy_alloc_range(struct drm_buddy *mm,
781 				   u64 start,
782 				   u64 size,
783 				   u64 *total_allocated_on_err,
784 				   struct list_head *blocks)
785 {
786 	LIST_HEAD(dfs);
787 	int i;
788 
789 	for (i = 0; i < mm->n_roots; ++i)
790 		list_add_tail(&mm->roots[i]->tmp_link, &dfs);
791 
792 	return __alloc_range(mm, &dfs, start, size,
793 			     blocks, total_allocated_on_err);
794 }
795 
796 static int __alloc_contig_try_harder(struct drm_buddy *mm,
797 				     u64 size,
798 				     u64 min_block_size,
799 				     struct list_head *blocks)
800 {
801 	u64 rhs_offset, lhs_offset, lhs_size, filled;
802 	struct drm_buddy_block *block;
803 	struct list_head *list;
804 	LIST_HEAD(blocks_lhs);
805 	unsigned long pages;
806 	unsigned int order;
807 	u64 modify_size;
808 	int err;
809 
810 	modify_size = rounddown_pow_of_two(size);
811 	pages = modify_size >> ilog2(mm->chunk_size);
812 	order = fls(pages) - 1;
813 	if (order == 0)
814 		return -ENOSPC;
815 
816 	list = &mm->free_list[order];
817 	if (list_empty(list))
818 		return -ENOSPC;
819 
820 	list_for_each_entry_reverse(block, list, link) {
821 		/* Allocate blocks traversing RHS */
822 		rhs_offset = drm_buddy_block_offset(block);
823 		err =  __drm_buddy_alloc_range(mm, rhs_offset, size,
824 					       &filled, blocks);
825 		if (!err || err != -ENOSPC)
826 			return err;
827 
828 		lhs_size = max((size - filled), min_block_size);
829 		if (!IS_ALIGNED(lhs_size, min_block_size))
830 			lhs_size = round_up(lhs_size, min_block_size);
831 
832 		/* Allocate blocks traversing LHS */
833 		lhs_offset = drm_buddy_block_offset(block) - lhs_size;
834 		err =  __drm_buddy_alloc_range(mm, lhs_offset, lhs_size,
835 					       NULL, &blocks_lhs);
836 		if (!err) {
837 			list_splice(&blocks_lhs, blocks);
838 			return 0;
839 		} else if (err != -ENOSPC) {
840 			drm_buddy_free_list_internal(mm, blocks);
841 			return err;
842 		}
843 		/* Free blocks for the next iteration */
844 		drm_buddy_free_list_internal(mm, blocks);
845 	}
846 
847 	return -ENOSPC;
848 }
849 
850 /**
851  * drm_buddy_block_trim - free unused pages
852  *
853  * @mm: DRM buddy manager
854  * @start: start address to begin the trimming.
855  * @new_size: original size requested
856  * @blocks: Input and output list of allocated blocks.
857  * MUST contain single block as input to be trimmed.
858  * On success will contain the newly allocated blocks
859  * making up the @new_size. Blocks always appear in
860  * ascending order
861  *
862  * For contiguous allocation, we round up the size to the nearest
863  * power of two value, drivers consume *actual* size, so remaining
864  * portions are unused and can be optionally freed with this function
865  *
866  * Returns:
867  * 0 on success, error code on failure.
868  */
869 int drm_buddy_block_trim(struct drm_buddy *mm,
870 			 u64 *start,
871 			 u64 new_size,
872 			 struct list_head *blocks)
873 {
874 	struct drm_buddy_block *parent;
875 	struct drm_buddy_block *block;
876 	u64 block_start, block_end;
877 	LIST_HEAD(dfs);
878 	u64 new_start;
879 	int err;
880 
881 	if (!list_is_singular(blocks))
882 		return -EINVAL;
883 
884 	block = list_first_entry(blocks,
885 				 struct drm_buddy_block,
886 				 link);
887 
888 	block_start = drm_buddy_block_offset(block);
889 	block_end = block_start + drm_buddy_block_size(mm, block);
890 
891 	if (WARN_ON(!drm_buddy_block_is_allocated(block)))
892 		return -EINVAL;
893 
894 	if (new_size > drm_buddy_block_size(mm, block))
895 		return -EINVAL;
896 
897 	if (!new_size || !IS_ALIGNED(new_size, mm->chunk_size))
898 		return -EINVAL;
899 
900 	if (new_size == drm_buddy_block_size(mm, block))
901 		return 0;
902 
903 	new_start = block_start;
904 	if (start) {
905 		new_start = *start;
906 
907 		if (new_start < block_start)
908 			return -EINVAL;
909 
910 		if (!IS_ALIGNED(new_start, mm->chunk_size))
911 			return -EINVAL;
912 
913 		if (range_overflows(new_start, new_size, block_end))
914 			return -EINVAL;
915 	}
916 
917 	list_del(&block->link);
918 	mark_free(mm, block);
919 	mm->avail += drm_buddy_block_size(mm, block);
920 	if (drm_buddy_block_is_clear(block))
921 		mm->clear_avail += drm_buddy_block_size(mm, block);
922 
923 	/* Prevent recursively freeing this node */
924 	parent = block->parent;
925 	block->parent = NULL;
926 
927 	list_add(&block->tmp_link, &dfs);
928 	err =  __alloc_range(mm, &dfs, new_start, new_size, blocks, NULL);
929 	if (err) {
930 		mark_allocated(block);
931 		mm->avail -= drm_buddy_block_size(mm, block);
932 		if (drm_buddy_block_is_clear(block))
933 			mm->clear_avail -= drm_buddy_block_size(mm, block);
934 		list_add(&block->link, blocks);
935 	}
936 
937 	block->parent = parent;
938 	return err;
939 }
940 EXPORT_SYMBOL(drm_buddy_block_trim);
941 
942 static struct drm_buddy_block *
943 __drm_buddy_alloc_blocks(struct drm_buddy *mm,
944 			 u64 start, u64 end,
945 			 unsigned int order,
946 			 unsigned long flags)
947 {
948 	if (flags & DRM_BUDDY_RANGE_ALLOCATION)
949 		/* Allocate traversing within the range */
950 		return  __drm_buddy_alloc_range_bias(mm, start, end,
951 						     order, flags);
952 	else
953 		/* Allocate from freelist */
954 		return alloc_from_freelist(mm, order, flags);
955 }
956 
957 /**
958  * drm_buddy_alloc_blocks - allocate power-of-two blocks
959  *
960  * @mm: DRM buddy manager to allocate from
961  * @start: start of the allowed range for this block
962  * @end: end of the allowed range for this block
963  * @size: size of the allocation in bytes
964  * @min_block_size: alignment of the allocation
965  * @blocks: output list head to add allocated blocks
966  * @flags: DRM_BUDDY_*_ALLOCATION flags
967  *
968  * alloc_range_bias() called on range limitations, which traverses
969  * the tree and returns the desired block.
970  *
971  * alloc_from_freelist() called when *no* range restrictions
972  * are enforced, which picks the block from the freelist.
973  *
974  * Returns:
975  * 0 on success, error code on failure.
976  */
977 int drm_buddy_alloc_blocks(struct drm_buddy *mm,
978 			   u64 start, u64 end, u64 size,
979 			   u64 min_block_size,
980 			   struct list_head *blocks,
981 			   unsigned long flags)
982 {
983 	struct drm_buddy_block *block = NULL;
984 	u64 original_size, original_min_size;
985 	unsigned int min_order, order;
986 	LIST_HEAD(allocated);
987 	unsigned long pages;
988 	int err;
989 
990 	if (size < mm->chunk_size)
991 		return -EINVAL;
992 
993 	if (min_block_size < mm->chunk_size)
994 		return -EINVAL;
995 
996 	if (!is_power_of_2(min_block_size))
997 		return -EINVAL;
998 
999 	if (!IS_ALIGNED(start | end | size, mm->chunk_size))
1000 		return -EINVAL;
1001 
1002 	if (end > mm->size)
1003 		return -EINVAL;
1004 
1005 	if (range_overflows(start, size, mm->size))
1006 		return -EINVAL;
1007 
1008 	/* Actual range allocation */
1009 	if (start + size == end) {
1010 		if (!IS_ALIGNED(start | end, min_block_size))
1011 			return -EINVAL;
1012 
1013 		return __drm_buddy_alloc_range(mm, start, size, NULL, blocks);
1014 	}
1015 
1016 	original_size = size;
1017 	original_min_size = min_block_size;
1018 
1019 	/* Roundup the size to power of 2 */
1020 	if (flags & DRM_BUDDY_CONTIGUOUS_ALLOCATION) {
1021 		size = roundup_pow_of_two(size);
1022 		min_block_size = size;
1023 	/* Align size value to min_block_size */
1024 	} else if (!IS_ALIGNED(size, min_block_size)) {
1025 		size = round_up(size, min_block_size);
1026 	}
1027 
1028 	pages = size >> ilog2(mm->chunk_size);
1029 	order = fls(pages) - 1;
1030 	min_order = ilog2(min_block_size) - ilog2(mm->chunk_size);
1031 
1032 	do {
1033 		order = min(order, (unsigned int)fls(pages) - 1);
1034 		BUG_ON(order > mm->max_order);
1035 		BUG_ON(order < min_order);
1036 
1037 		do {
1038 			block = __drm_buddy_alloc_blocks(mm, start,
1039 							 end,
1040 							 order,
1041 							 flags);
1042 			if (!IS_ERR(block))
1043 				break;
1044 
1045 			if (order-- == min_order) {
1046 				/* Try allocation through force merge method */
1047 				if (mm->clear_avail &&
1048 				    !__force_merge(mm, start, end, min_order)) {
1049 					block = __drm_buddy_alloc_blocks(mm, start,
1050 									 end,
1051 									 min_order,
1052 									 flags);
1053 					if (!IS_ERR(block)) {
1054 						order = min_order;
1055 						break;
1056 					}
1057 				}
1058 
1059 				/*
1060 				 * Try contiguous block allocation through
1061 				 * try harder method.
1062 				 */
1063 				if (flags & DRM_BUDDY_CONTIGUOUS_ALLOCATION &&
1064 				    !(flags & DRM_BUDDY_RANGE_ALLOCATION))
1065 					return __alloc_contig_try_harder(mm,
1066 									 original_size,
1067 									 original_min_size,
1068 									 blocks);
1069 				err = -ENOSPC;
1070 				goto err_free;
1071 			}
1072 		} while (1);
1073 
1074 		mark_allocated(block);
1075 		mm->avail -= drm_buddy_block_size(mm, block);
1076 		if (drm_buddy_block_is_clear(block))
1077 			mm->clear_avail -= drm_buddy_block_size(mm, block);
1078 		kmemleak_update_trace(block);
1079 		list_add_tail(&block->link, &allocated);
1080 
1081 		pages -= BIT(order);
1082 
1083 		if (!pages)
1084 			break;
1085 	} while (1);
1086 
1087 	/* Trim the allocated block to the required size */
1088 	if (!(flags & DRM_BUDDY_TRIM_DISABLE) &&
1089 	    original_size != size) {
1090 		struct list_head *trim_list;
1091 		LIST_HEAD(temp);
1092 		u64 trim_size;
1093 
1094 		trim_list = &allocated;
1095 		trim_size = original_size;
1096 
1097 		if (!list_is_singular(&allocated)) {
1098 			block = list_last_entry(&allocated, typeof(*block), link);
1099 			list_move(&block->link, &temp);
1100 			trim_list = &temp;
1101 			trim_size = drm_buddy_block_size(mm, block) -
1102 				(size - original_size);
1103 		}
1104 
1105 		drm_buddy_block_trim(mm,
1106 				     NULL,
1107 				     trim_size,
1108 				     trim_list);
1109 
1110 		if (!list_empty(&temp))
1111 			list_splice_tail(trim_list, &allocated);
1112 	}
1113 
1114 	list_splice_tail(&allocated, blocks);
1115 	return 0;
1116 
1117 err_free:
1118 	drm_buddy_free_list_internal(mm, &allocated);
1119 	return err;
1120 }
1121 EXPORT_SYMBOL(drm_buddy_alloc_blocks);
1122 
1123 /**
1124  * drm_buddy_block_print - print block information
1125  *
1126  * @mm: DRM buddy manager
1127  * @block: DRM buddy block
1128  * @p: DRM printer to use
1129  */
1130 void drm_buddy_block_print(struct drm_buddy *mm,
1131 			   struct drm_buddy_block *block,
1132 			   struct drm_printer *p)
1133 {
1134 	u64 start = drm_buddy_block_offset(block);
1135 	u64 size = drm_buddy_block_size(mm, block);
1136 
1137 	drm_printf(p, "%#018llx-%#018llx: %llu\n", start, start + size, size);
1138 }
1139 EXPORT_SYMBOL(drm_buddy_block_print);
1140 
1141 /**
1142  * drm_buddy_print - print allocator state
1143  *
1144  * @mm: DRM buddy manager
1145  * @p: DRM printer to use
1146  */
1147 void drm_buddy_print(struct drm_buddy *mm, struct drm_printer *p)
1148 {
1149 	int order;
1150 
1151 	drm_printf(p, "chunk_size: %lluKiB, total: %lluMiB, free: %lluMiB, clear_free: %lluMiB\n",
1152 		   mm->chunk_size >> 10, mm->size >> 20, mm->avail >> 20, mm->clear_avail >> 20);
1153 
1154 	for (order = mm->max_order; order >= 0; order--) {
1155 		struct drm_buddy_block *block;
1156 		u64 count = 0, free;
1157 
1158 		list_for_each_entry(block, &mm->free_list[order], link) {
1159 			BUG_ON(!drm_buddy_block_is_free(block));
1160 			count++;
1161 		}
1162 
1163 		drm_printf(p, "order-%2d ", order);
1164 
1165 		free = count * (mm->chunk_size << order);
1166 		if (free < SZ_1M)
1167 			drm_printf(p, "free: %8llu KiB", free >> 10);
1168 		else
1169 			drm_printf(p, "free: %8llu MiB", free >> 20);
1170 
1171 		drm_printf(p, ", blocks: %llu\n", count);
1172 	}
1173 }
1174 EXPORT_SYMBOL(drm_buddy_print);
1175 
1176 static void drm_buddy_module_exit(void)
1177 {
1178 	kmem_cache_destroy(slab_blocks);
1179 }
1180 
1181 static int __init drm_buddy_module_init(void)
1182 {
1183 	slab_blocks = KMEM_CACHE(drm_buddy_block, 0);
1184 	if (!slab_blocks)
1185 		return -ENOMEM;
1186 
1187 	return 0;
1188 }
1189 
1190 module_init(drm_buddy_module_init);
1191 module_exit(drm_buddy_module_exit);
1192 
1193 MODULE_DESCRIPTION("DRM Buddy Allocator");
1194 MODULE_LICENSE("Dual MIT/GPL");
1195