Loading...
1/**************************************************************************
2 *
3 * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., 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
29/*
30 * Generic simple memory manager implementation. Intended to be used as a base
31 * class implementation for more advanced memory managers.
32 *
33 * Note that the algorithm used is quite simple and there might be substantial
34 * performance gains if a smarter free list is implemented. Currently it is just an
35 * unordered stack of free regions. This could easily be improved if an RB-tree
36 * is used instead. At least if we expect heavy fragmentation.
37 *
38 * Aligned allocations can also see improvement.
39 *
40 * Authors:
41 * Thomas Hellström <thomas-at-tungstengraphics-dot-com>
42 */
43
44#include "drmP.h"
45#include "drm_mm.h"
46#include <linux/slab.h>
47#include <linux/seq_file.h>
48
49#define MM_UNUSED_TARGET 4
50
51static struct drm_mm_node *drm_mm_kmalloc(struct drm_mm *mm, int atomic)
52{
53 struct drm_mm_node *child;
54
55 if (atomic)
56 child = kzalloc(sizeof(*child), GFP_ATOMIC);
57 else
58 child = kzalloc(sizeof(*child), GFP_KERNEL);
59
60 if (unlikely(child == NULL)) {
61 spin_lock(&mm->unused_lock);
62 if (list_empty(&mm->unused_nodes))
63 child = NULL;
64 else {
65 child =
66 list_entry(mm->unused_nodes.next,
67 struct drm_mm_node, node_list);
68 list_del(&child->node_list);
69 --mm->num_unused;
70 }
71 spin_unlock(&mm->unused_lock);
72 }
73 return child;
74}
75
76/* drm_mm_pre_get() - pre allocate drm_mm_node structure
77 * drm_mm: memory manager struct we are pre-allocating for
78 *
79 * Returns 0 on success or -ENOMEM if allocation fails.
80 */
81int drm_mm_pre_get(struct drm_mm *mm)
82{
83 struct drm_mm_node *node;
84
85 spin_lock(&mm->unused_lock);
86 while (mm->num_unused < MM_UNUSED_TARGET) {
87 spin_unlock(&mm->unused_lock);
88 node = kzalloc(sizeof(*node), GFP_KERNEL);
89 spin_lock(&mm->unused_lock);
90
91 if (unlikely(node == NULL)) {
92 int ret = (mm->num_unused < 2) ? -ENOMEM : 0;
93 spin_unlock(&mm->unused_lock);
94 return ret;
95 }
96 ++mm->num_unused;
97 list_add_tail(&node->node_list, &mm->unused_nodes);
98 }
99 spin_unlock(&mm->unused_lock);
100 return 0;
101}
102EXPORT_SYMBOL(drm_mm_pre_get);
103
104static inline unsigned long drm_mm_hole_node_start(struct drm_mm_node *hole_node)
105{
106 return hole_node->start + hole_node->size;
107}
108
109static inline unsigned long drm_mm_hole_node_end(struct drm_mm_node *hole_node)
110{
111 struct drm_mm_node *next_node =
112 list_entry(hole_node->node_list.next, struct drm_mm_node,
113 node_list);
114
115 return next_node->start;
116}
117
118static void drm_mm_insert_helper(struct drm_mm_node *hole_node,
119 struct drm_mm_node *node,
120 unsigned long size, unsigned alignment)
121{
122 struct drm_mm *mm = hole_node->mm;
123 unsigned long tmp = 0, wasted = 0;
124 unsigned long hole_start = drm_mm_hole_node_start(hole_node);
125 unsigned long hole_end = drm_mm_hole_node_end(hole_node);
126
127 BUG_ON(!hole_node->hole_follows || node->allocated);
128
129 if (alignment)
130 tmp = hole_start % alignment;
131
132 if (!tmp) {
133 hole_node->hole_follows = 0;
134 list_del_init(&hole_node->hole_stack);
135 } else
136 wasted = alignment - tmp;
137
138 node->start = hole_start + wasted;
139 node->size = size;
140 node->mm = mm;
141 node->allocated = 1;
142
143 INIT_LIST_HEAD(&node->hole_stack);
144 list_add(&node->node_list, &hole_node->node_list);
145
146 BUG_ON(node->start + node->size > hole_end);
147
148 if (node->start + node->size < hole_end) {
149 list_add(&node->hole_stack, &mm->hole_stack);
150 node->hole_follows = 1;
151 } else {
152 node->hole_follows = 0;
153 }
154}
155
156struct drm_mm_node *drm_mm_get_block_generic(struct drm_mm_node *hole_node,
157 unsigned long size,
158 unsigned alignment,
159 int atomic)
160{
161 struct drm_mm_node *node;
162
163 node = drm_mm_kmalloc(hole_node->mm, atomic);
164 if (unlikely(node == NULL))
165 return NULL;
166
167 drm_mm_insert_helper(hole_node, node, size, alignment);
168
169 return node;
170}
171EXPORT_SYMBOL(drm_mm_get_block_generic);
172
173/**
174 * Search for free space and insert a preallocated memory node. Returns
175 * -ENOSPC if no suitable free area is available. The preallocated memory node
176 * must be cleared.
177 */
178int drm_mm_insert_node(struct drm_mm *mm, struct drm_mm_node *node,
179 unsigned long size, unsigned alignment)
180{
181 struct drm_mm_node *hole_node;
182
183 hole_node = drm_mm_search_free(mm, size, alignment, 0);
184 if (!hole_node)
185 return -ENOSPC;
186
187 drm_mm_insert_helper(hole_node, node, size, alignment);
188
189 return 0;
190}
191EXPORT_SYMBOL(drm_mm_insert_node);
192
193static void drm_mm_insert_helper_range(struct drm_mm_node *hole_node,
194 struct drm_mm_node *node,
195 unsigned long size, unsigned alignment,
196 unsigned long start, unsigned long end)
197{
198 struct drm_mm *mm = hole_node->mm;
199 unsigned long tmp = 0, wasted = 0;
200 unsigned long hole_start = drm_mm_hole_node_start(hole_node);
201 unsigned long hole_end = drm_mm_hole_node_end(hole_node);
202
203 BUG_ON(!hole_node->hole_follows || node->allocated);
204
205 if (hole_start < start)
206 wasted += start - hole_start;
207 if (alignment)
208 tmp = (hole_start + wasted) % alignment;
209
210 if (tmp)
211 wasted += alignment - tmp;
212
213 if (!wasted) {
214 hole_node->hole_follows = 0;
215 list_del_init(&hole_node->hole_stack);
216 }
217
218 node->start = hole_start + wasted;
219 node->size = size;
220 node->mm = mm;
221 node->allocated = 1;
222
223 INIT_LIST_HEAD(&node->hole_stack);
224 list_add(&node->node_list, &hole_node->node_list);
225
226 BUG_ON(node->start + node->size > hole_end);
227 BUG_ON(node->start + node->size > end);
228
229 if (node->start + node->size < hole_end) {
230 list_add(&node->hole_stack, &mm->hole_stack);
231 node->hole_follows = 1;
232 } else {
233 node->hole_follows = 0;
234 }
235}
236
237struct drm_mm_node *drm_mm_get_block_range_generic(struct drm_mm_node *hole_node,
238 unsigned long size,
239 unsigned alignment,
240 unsigned long start,
241 unsigned long end,
242 int atomic)
243{
244 struct drm_mm_node *node;
245
246 node = drm_mm_kmalloc(hole_node->mm, atomic);
247 if (unlikely(node == NULL))
248 return NULL;
249
250 drm_mm_insert_helper_range(hole_node, node, size, alignment,
251 start, end);
252
253 return node;
254}
255EXPORT_SYMBOL(drm_mm_get_block_range_generic);
256
257/**
258 * Search for free space and insert a preallocated memory node. Returns
259 * -ENOSPC if no suitable free area is available. This is for range
260 * restricted allocations. The preallocated memory node must be cleared.
261 */
262int drm_mm_insert_node_in_range(struct drm_mm *mm, struct drm_mm_node *node,
263 unsigned long size, unsigned alignment,
264 unsigned long start, unsigned long end)
265{
266 struct drm_mm_node *hole_node;
267
268 hole_node = drm_mm_search_free_in_range(mm, size, alignment,
269 start, end, 0);
270 if (!hole_node)
271 return -ENOSPC;
272
273 drm_mm_insert_helper_range(hole_node, node, size, alignment,
274 start, end);
275
276 return 0;
277}
278EXPORT_SYMBOL(drm_mm_insert_node_in_range);
279
280/**
281 * Remove a memory node from the allocator.
282 */
283void drm_mm_remove_node(struct drm_mm_node *node)
284{
285 struct drm_mm *mm = node->mm;
286 struct drm_mm_node *prev_node;
287
288 BUG_ON(node->scanned_block || node->scanned_prev_free
289 || node->scanned_next_free);
290
291 prev_node =
292 list_entry(node->node_list.prev, struct drm_mm_node, node_list);
293
294 if (node->hole_follows) {
295 BUG_ON(drm_mm_hole_node_start(node)
296 == drm_mm_hole_node_end(node));
297 list_del(&node->hole_stack);
298 } else
299 BUG_ON(drm_mm_hole_node_start(node)
300 != drm_mm_hole_node_end(node));
301
302 if (!prev_node->hole_follows) {
303 prev_node->hole_follows = 1;
304 list_add(&prev_node->hole_stack, &mm->hole_stack);
305 } else
306 list_move(&prev_node->hole_stack, &mm->hole_stack);
307
308 list_del(&node->node_list);
309 node->allocated = 0;
310}
311EXPORT_SYMBOL(drm_mm_remove_node);
312
313/*
314 * Remove a memory node from the allocator and free the allocated struct
315 * drm_mm_node. Only to be used on a struct drm_mm_node obtained by one of the
316 * drm_mm_get_block functions.
317 */
318void drm_mm_put_block(struct drm_mm_node *node)
319{
320
321 struct drm_mm *mm = node->mm;
322
323 drm_mm_remove_node(node);
324
325 spin_lock(&mm->unused_lock);
326 if (mm->num_unused < MM_UNUSED_TARGET) {
327 list_add(&node->node_list, &mm->unused_nodes);
328 ++mm->num_unused;
329 } else
330 kfree(node);
331 spin_unlock(&mm->unused_lock);
332}
333EXPORT_SYMBOL(drm_mm_put_block);
334
335static int check_free_hole(unsigned long start, unsigned long end,
336 unsigned long size, unsigned alignment)
337{
338 unsigned wasted = 0;
339
340 if (end - start < size)
341 return 0;
342
343 if (alignment) {
344 unsigned tmp = start % alignment;
345 if (tmp)
346 wasted = alignment - tmp;
347 }
348
349 if (end >= start + size + wasted) {
350 return 1;
351 }
352
353 return 0;
354}
355
356struct drm_mm_node *drm_mm_search_free(const struct drm_mm *mm,
357 unsigned long size,
358 unsigned alignment, int best_match)
359{
360 struct drm_mm_node *entry;
361 struct drm_mm_node *best;
362 unsigned long best_size;
363
364 BUG_ON(mm->scanned_blocks);
365
366 best = NULL;
367 best_size = ~0UL;
368
369 list_for_each_entry(entry, &mm->hole_stack, hole_stack) {
370 BUG_ON(!entry->hole_follows);
371 if (!check_free_hole(drm_mm_hole_node_start(entry),
372 drm_mm_hole_node_end(entry),
373 size, alignment))
374 continue;
375
376 if (!best_match)
377 return entry;
378
379 if (entry->size < best_size) {
380 best = entry;
381 best_size = entry->size;
382 }
383 }
384
385 return best;
386}
387EXPORT_SYMBOL(drm_mm_search_free);
388
389struct drm_mm_node *drm_mm_search_free_in_range(const struct drm_mm *mm,
390 unsigned long size,
391 unsigned alignment,
392 unsigned long start,
393 unsigned long end,
394 int best_match)
395{
396 struct drm_mm_node *entry;
397 struct drm_mm_node *best;
398 unsigned long best_size;
399
400 BUG_ON(mm->scanned_blocks);
401
402 best = NULL;
403 best_size = ~0UL;
404
405 list_for_each_entry(entry, &mm->hole_stack, hole_stack) {
406 unsigned long adj_start = drm_mm_hole_node_start(entry) < start ?
407 start : drm_mm_hole_node_start(entry);
408 unsigned long adj_end = drm_mm_hole_node_end(entry) > end ?
409 end : drm_mm_hole_node_end(entry);
410
411 BUG_ON(!entry->hole_follows);
412 if (!check_free_hole(adj_start, adj_end, size, alignment))
413 continue;
414
415 if (!best_match)
416 return entry;
417
418 if (entry->size < best_size) {
419 best = entry;
420 best_size = entry->size;
421 }
422 }
423
424 return best;
425}
426EXPORT_SYMBOL(drm_mm_search_free_in_range);
427
428/**
429 * Moves an allocation. To be used with embedded struct drm_mm_node.
430 */
431void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
432{
433 list_replace(&old->node_list, &new->node_list);
434 list_replace(&old->hole_stack, &new->hole_stack);
435 new->hole_follows = old->hole_follows;
436 new->mm = old->mm;
437 new->start = old->start;
438 new->size = old->size;
439
440 old->allocated = 0;
441 new->allocated = 1;
442}
443EXPORT_SYMBOL(drm_mm_replace_node);
444
445/**
446 * Initializa lru scanning.
447 *
448 * This simply sets up the scanning routines with the parameters for the desired
449 * hole.
450 *
451 * Warning: As long as the scan list is non-empty, no other operations than
452 * adding/removing nodes to/from the scan list are allowed.
453 */
454void drm_mm_init_scan(struct drm_mm *mm, unsigned long size,
455 unsigned alignment)
456{
457 mm->scan_alignment = alignment;
458 mm->scan_size = size;
459 mm->scanned_blocks = 0;
460 mm->scan_hit_start = 0;
461 mm->scan_hit_size = 0;
462 mm->scan_check_range = 0;
463 mm->prev_scanned_node = NULL;
464}
465EXPORT_SYMBOL(drm_mm_init_scan);
466
467/**
468 * Initializa lru scanning.
469 *
470 * This simply sets up the scanning routines with the parameters for the desired
471 * hole. This version is for range-restricted scans.
472 *
473 * Warning: As long as the scan list is non-empty, no other operations than
474 * adding/removing nodes to/from the scan list are allowed.
475 */
476void drm_mm_init_scan_with_range(struct drm_mm *mm, unsigned long size,
477 unsigned alignment,
478 unsigned long start,
479 unsigned long end)
480{
481 mm->scan_alignment = alignment;
482 mm->scan_size = size;
483 mm->scanned_blocks = 0;
484 mm->scan_hit_start = 0;
485 mm->scan_hit_size = 0;
486 mm->scan_start = start;
487 mm->scan_end = end;
488 mm->scan_check_range = 1;
489 mm->prev_scanned_node = NULL;
490}
491EXPORT_SYMBOL(drm_mm_init_scan_with_range);
492
493/**
494 * Add a node to the scan list that might be freed to make space for the desired
495 * hole.
496 *
497 * Returns non-zero, if a hole has been found, zero otherwise.
498 */
499int drm_mm_scan_add_block(struct drm_mm_node *node)
500{
501 struct drm_mm *mm = node->mm;
502 struct drm_mm_node *prev_node;
503 unsigned long hole_start, hole_end;
504 unsigned long adj_start;
505 unsigned long adj_end;
506
507 mm->scanned_blocks++;
508
509 BUG_ON(node->scanned_block);
510 node->scanned_block = 1;
511
512 prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
513 node_list);
514
515 node->scanned_preceeds_hole = prev_node->hole_follows;
516 prev_node->hole_follows = 1;
517 list_del(&node->node_list);
518 node->node_list.prev = &prev_node->node_list;
519 node->node_list.next = &mm->prev_scanned_node->node_list;
520 mm->prev_scanned_node = node;
521
522 hole_start = drm_mm_hole_node_start(prev_node);
523 hole_end = drm_mm_hole_node_end(prev_node);
524 if (mm->scan_check_range) {
525 adj_start = hole_start < mm->scan_start ?
526 mm->scan_start : hole_start;
527 adj_end = hole_end > mm->scan_end ?
528 mm->scan_end : hole_end;
529 } else {
530 adj_start = hole_start;
531 adj_end = hole_end;
532 }
533
534 if (check_free_hole(adj_start , adj_end,
535 mm->scan_size, mm->scan_alignment)) {
536 mm->scan_hit_start = hole_start;
537 mm->scan_hit_size = hole_end;
538
539 return 1;
540 }
541
542 return 0;
543}
544EXPORT_SYMBOL(drm_mm_scan_add_block);
545
546/**
547 * Remove a node from the scan list.
548 *
549 * Nodes _must_ be removed in the exact same order from the scan list as they
550 * have been added, otherwise the internal state of the memory manager will be
551 * corrupted.
552 *
553 * When the scan list is empty, the selected memory nodes can be freed. An
554 * immediately following drm_mm_search_free with best_match = 0 will then return
555 * the just freed block (because its at the top of the free_stack list).
556 *
557 * Returns one if this block should be evicted, zero otherwise. Will always
558 * return zero when no hole has been found.
559 */
560int drm_mm_scan_remove_block(struct drm_mm_node *node)
561{
562 struct drm_mm *mm = node->mm;
563 struct drm_mm_node *prev_node;
564
565 mm->scanned_blocks--;
566
567 BUG_ON(!node->scanned_block);
568 node->scanned_block = 0;
569
570 prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
571 node_list);
572
573 prev_node->hole_follows = node->scanned_preceeds_hole;
574 INIT_LIST_HEAD(&node->node_list);
575 list_add(&node->node_list, &prev_node->node_list);
576
577 /* Only need to check for containement because start&size for the
578 * complete resulting free block (not just the desired part) is
579 * stored. */
580 if (node->start >= mm->scan_hit_start &&
581 node->start + node->size
582 <= mm->scan_hit_start + mm->scan_hit_size) {
583 return 1;
584 }
585
586 return 0;
587}
588EXPORT_SYMBOL(drm_mm_scan_remove_block);
589
590int drm_mm_clean(struct drm_mm * mm)
591{
592 struct list_head *head = &mm->head_node.node_list;
593
594 return (head->next->next == head);
595}
596EXPORT_SYMBOL(drm_mm_clean);
597
598int drm_mm_init(struct drm_mm * mm, unsigned long start, unsigned long size)
599{
600 INIT_LIST_HEAD(&mm->hole_stack);
601 INIT_LIST_HEAD(&mm->unused_nodes);
602 mm->num_unused = 0;
603 mm->scanned_blocks = 0;
604 spin_lock_init(&mm->unused_lock);
605
606 /* Clever trick to avoid a special case in the free hole tracking. */
607 INIT_LIST_HEAD(&mm->head_node.node_list);
608 INIT_LIST_HEAD(&mm->head_node.hole_stack);
609 mm->head_node.hole_follows = 1;
610 mm->head_node.scanned_block = 0;
611 mm->head_node.scanned_prev_free = 0;
612 mm->head_node.scanned_next_free = 0;
613 mm->head_node.mm = mm;
614 mm->head_node.start = start + size;
615 mm->head_node.size = start - mm->head_node.start;
616 list_add_tail(&mm->head_node.hole_stack, &mm->hole_stack);
617
618 return 0;
619}
620EXPORT_SYMBOL(drm_mm_init);
621
622void drm_mm_takedown(struct drm_mm * mm)
623{
624 struct drm_mm_node *entry, *next;
625
626 if (!list_empty(&mm->head_node.node_list)) {
627 DRM_ERROR("Memory manager not clean. Delaying takedown\n");
628 return;
629 }
630
631 spin_lock(&mm->unused_lock);
632 list_for_each_entry_safe(entry, next, &mm->unused_nodes, node_list) {
633 list_del(&entry->node_list);
634 kfree(entry);
635 --mm->num_unused;
636 }
637 spin_unlock(&mm->unused_lock);
638
639 BUG_ON(mm->num_unused != 0);
640}
641EXPORT_SYMBOL(drm_mm_takedown);
642
643void drm_mm_debug_table(struct drm_mm *mm, const char *prefix)
644{
645 struct drm_mm_node *entry;
646 unsigned long total_used = 0, total_free = 0, total = 0;
647 unsigned long hole_start, hole_end, hole_size;
648
649 hole_start = drm_mm_hole_node_start(&mm->head_node);
650 hole_end = drm_mm_hole_node_end(&mm->head_node);
651 hole_size = hole_end - hole_start;
652 if (hole_size)
653 printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: free\n",
654 prefix, hole_start, hole_end,
655 hole_size);
656 total_free += hole_size;
657
658 drm_mm_for_each_node(entry, mm) {
659 printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: used\n",
660 prefix, entry->start, entry->start + entry->size,
661 entry->size);
662 total_used += entry->size;
663
664 if (entry->hole_follows) {
665 hole_start = drm_mm_hole_node_start(entry);
666 hole_end = drm_mm_hole_node_end(entry);
667 hole_size = hole_end - hole_start;
668 printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: free\n",
669 prefix, hole_start, hole_end,
670 hole_size);
671 total_free += hole_size;
672 }
673 }
674 total = total_free + total_used;
675
676 printk(KERN_DEBUG "%s total: %lu, used %lu free %lu\n", prefix, total,
677 total_used, total_free);
678}
679EXPORT_SYMBOL(drm_mm_debug_table);
680
681#if defined(CONFIG_DEBUG_FS)
682int drm_mm_dump_table(struct seq_file *m, struct drm_mm *mm)
683{
684 struct drm_mm_node *entry;
685 unsigned long total_used = 0, total_free = 0, total = 0;
686 unsigned long hole_start, hole_end, hole_size;
687
688 hole_start = drm_mm_hole_node_start(&mm->head_node);
689 hole_end = drm_mm_hole_node_end(&mm->head_node);
690 hole_size = hole_end - hole_start;
691 if (hole_size)
692 seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: free\n",
693 hole_start, hole_end, hole_size);
694 total_free += hole_size;
695
696 drm_mm_for_each_node(entry, mm) {
697 seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: used\n",
698 entry->start, entry->start + entry->size,
699 entry->size);
700 total_used += entry->size;
701 if (entry->hole_follows) {
702 hole_start = drm_mm_hole_node_start(entry);
703 hole_end = drm_mm_hole_node_end(entry);
704 hole_size = hole_end - hole_start;
705 seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: free\n",
706 hole_start, hole_end, hole_size);
707 total_free += hole_size;
708 }
709 }
710 total = total_free + total_used;
711
712 seq_printf(m, "total: %lu, used %lu free %lu\n", total, total_used, total_free);
713 return 0;
714}
715EXPORT_SYMBOL(drm_mm_dump_table);
716#endif
1/**************************************************************************
2 *
3 * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA.
4 * Copyright 2016 Intel Corporation
5 * All Rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25 * USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 *
28 **************************************************************************/
29
30/*
31 * Generic simple memory manager implementation. Intended to be used as a base
32 * class implementation for more advanced memory managers.
33 *
34 * Note that the algorithm used is quite simple and there might be substantial
35 * performance gains if a smarter free list is implemented. Currently it is
36 * just an unordered stack of free regions. This could easily be improved if
37 * an RB-tree is used instead. At least if we expect heavy fragmentation.
38 *
39 * Aligned allocations can also see improvement.
40 *
41 * Authors:
42 * Thomas Hellström <thomas-at-tungstengraphics-dot-com>
43 */
44
45#include <linux/export.h>
46#include <linux/interval_tree_generic.h>
47#include <linux/seq_file.h>
48#include <linux/slab.h>
49#include <linux/stacktrace.h>
50
51#include <drm/drm_mm.h>
52
53/**
54 * DOC: Overview
55 *
56 * drm_mm provides a simple range allocator. The drivers are free to use the
57 * resource allocator from the linux core if it suits them, the upside of drm_mm
58 * is that it's in the DRM core. Which means that it's easier to extend for
59 * some of the crazier special purpose needs of gpus.
60 *
61 * The main data struct is &drm_mm, allocations are tracked in &drm_mm_node.
62 * Drivers are free to embed either of them into their own suitable
63 * datastructures. drm_mm itself will not do any memory allocations of its own,
64 * so if drivers choose not to embed nodes they need to still allocate them
65 * themselves.
66 *
67 * The range allocator also supports reservation of preallocated blocks. This is
68 * useful for taking over initial mode setting configurations from the firmware,
69 * where an object needs to be created which exactly matches the firmware's
70 * scanout target. As long as the range is still free it can be inserted anytime
71 * after the allocator is initialized, which helps with avoiding looped
72 * dependencies in the driver load sequence.
73 *
74 * drm_mm maintains a stack of most recently freed holes, which of all
75 * simplistic datastructures seems to be a fairly decent approach to clustering
76 * allocations and avoiding too much fragmentation. This means free space
77 * searches are O(num_holes). Given that all the fancy features drm_mm supports
78 * something better would be fairly complex and since gfx thrashing is a fairly
79 * steep cliff not a real concern. Removing a node again is O(1).
80 *
81 * drm_mm supports a few features: Alignment and range restrictions can be
82 * supplied. Furthermore every &drm_mm_node has a color value (which is just an
83 * opaque unsigned long) which in conjunction with a driver callback can be used
84 * to implement sophisticated placement restrictions. The i915 DRM driver uses
85 * this to implement guard pages between incompatible caching domains in the
86 * graphics TT.
87 *
88 * Two behaviors are supported for searching and allocating: bottom-up and
89 * top-down. The default is bottom-up. Top-down allocation can be used if the
90 * memory area has different restrictions, or just to reduce fragmentation.
91 *
92 * Finally iteration helpers to walk all nodes and all holes are provided as are
93 * some basic allocator dumpers for debugging.
94 *
95 * Note that this range allocator is not thread-safe, drivers need to protect
96 * modifications with their own locking. The idea behind this is that for a full
97 * memory manager additional data needs to be protected anyway, hence internal
98 * locking would be fully redundant.
99 */
100
101#ifdef CONFIG_DRM_DEBUG_MM
102#include <linux/stackdepot.h>
103
104#define STACKDEPTH 32
105#define BUFSZ 4096
106
107static noinline void save_stack(struct drm_mm_node *node)
108{
109 unsigned long entries[STACKDEPTH];
110 unsigned int n;
111
112 n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
113
114 /* May be called under spinlock, so avoid sleeping */
115 node->stack = stack_depot_save(entries, n, GFP_NOWAIT);
116}
117
118static void show_leaks(struct drm_mm *mm)
119{
120 struct drm_mm_node *node;
121 unsigned long *entries;
122 unsigned int nr_entries;
123 char *buf;
124
125 buf = kmalloc(BUFSZ, GFP_KERNEL);
126 if (!buf)
127 return;
128
129 list_for_each_entry(node, drm_mm_nodes(mm), node_list) {
130 if (!node->stack) {
131 DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
132 node->start, node->size);
133 continue;
134 }
135
136 nr_entries = stack_depot_fetch(node->stack, &entries);
137 stack_trace_snprint(buf, BUFSZ, entries, nr_entries, 0);
138 DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
139 node->start, node->size, buf);
140 }
141
142 kfree(buf);
143}
144
145#undef STACKDEPTH
146#undef BUFSZ
147#else
148static void save_stack(struct drm_mm_node *node) { }
149static void show_leaks(struct drm_mm *mm) { }
150#endif
151
152#define START(node) ((node)->start)
153#define LAST(node) ((node)->start + (node)->size - 1)
154
155INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
156 u64, __subtree_last,
157 START, LAST, static inline, drm_mm_interval_tree)
158
159struct drm_mm_node *
160__drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last)
161{
162 return drm_mm_interval_tree_iter_first((struct rb_root_cached *)&mm->interval_tree,
163 start, last) ?: (struct drm_mm_node *)&mm->head_node;
164}
165EXPORT_SYMBOL(__drm_mm_interval_first);
166
167static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
168 struct drm_mm_node *node)
169{
170 struct drm_mm *mm = hole_node->mm;
171 struct rb_node **link, *rb;
172 struct drm_mm_node *parent;
173 bool leftmost;
174
175 node->__subtree_last = LAST(node);
176
177 if (drm_mm_node_allocated(hole_node)) {
178 rb = &hole_node->rb;
179 while (rb) {
180 parent = rb_entry(rb, struct drm_mm_node, rb);
181 if (parent->__subtree_last >= node->__subtree_last)
182 break;
183
184 parent->__subtree_last = node->__subtree_last;
185 rb = rb_parent(rb);
186 }
187
188 rb = &hole_node->rb;
189 link = &hole_node->rb.rb_right;
190 leftmost = false;
191 } else {
192 rb = NULL;
193 link = &mm->interval_tree.rb_root.rb_node;
194 leftmost = true;
195 }
196
197 while (*link) {
198 rb = *link;
199 parent = rb_entry(rb, struct drm_mm_node, rb);
200 if (parent->__subtree_last < node->__subtree_last)
201 parent->__subtree_last = node->__subtree_last;
202 if (node->start < parent->start) {
203 link = &parent->rb.rb_left;
204 } else {
205 link = &parent->rb.rb_right;
206 leftmost = false;
207 }
208 }
209
210 rb_link_node(&node->rb, rb, link);
211 rb_insert_augmented_cached(&node->rb, &mm->interval_tree, leftmost,
212 &drm_mm_interval_tree_augment);
213}
214
215#define HOLE_SIZE(NODE) ((NODE)->hole_size)
216#define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
217
218static u64 rb_to_hole_size(struct rb_node *rb)
219{
220 return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
221}
222
223static void insert_hole_size(struct rb_root_cached *root,
224 struct drm_mm_node *node)
225{
226 struct rb_node **link = &root->rb_root.rb_node, *rb = NULL;
227 u64 x = node->hole_size;
228 bool first = true;
229
230 while (*link) {
231 rb = *link;
232 if (x > rb_to_hole_size(rb)) {
233 link = &rb->rb_left;
234 } else {
235 link = &rb->rb_right;
236 first = false;
237 }
238 }
239
240 rb_link_node(&node->rb_hole_size, rb, link);
241 rb_insert_color_cached(&node->rb_hole_size, root, first);
242}
243
244RB_DECLARE_CALLBACKS_MAX(static, augment_callbacks,
245 struct drm_mm_node, rb_hole_addr,
246 u64, subtree_max_hole, HOLE_SIZE)
247
248static void insert_hole_addr(struct rb_root *root, struct drm_mm_node *node)
249{
250 struct rb_node **link = &root->rb_node, *rb_parent = NULL;
251 u64 start = HOLE_ADDR(node), subtree_max_hole = node->subtree_max_hole;
252 struct drm_mm_node *parent;
253
254 while (*link) {
255 rb_parent = *link;
256 parent = rb_entry(rb_parent, struct drm_mm_node, rb_hole_addr);
257 if (parent->subtree_max_hole < subtree_max_hole)
258 parent->subtree_max_hole = subtree_max_hole;
259 if (start < HOLE_ADDR(parent))
260 link = &parent->rb_hole_addr.rb_left;
261 else
262 link = &parent->rb_hole_addr.rb_right;
263 }
264
265 rb_link_node(&node->rb_hole_addr, rb_parent, link);
266 rb_insert_augmented(&node->rb_hole_addr, root, &augment_callbacks);
267}
268
269static void add_hole(struct drm_mm_node *node)
270{
271 struct drm_mm *mm = node->mm;
272
273 node->hole_size =
274 __drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
275 node->subtree_max_hole = node->hole_size;
276 DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
277
278 insert_hole_size(&mm->holes_size, node);
279 insert_hole_addr(&mm->holes_addr, node);
280
281 list_add(&node->hole_stack, &mm->hole_stack);
282}
283
284static void rm_hole(struct drm_mm_node *node)
285{
286 DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
287
288 list_del(&node->hole_stack);
289 rb_erase_cached(&node->rb_hole_size, &node->mm->holes_size);
290 rb_erase_augmented(&node->rb_hole_addr, &node->mm->holes_addr,
291 &augment_callbacks);
292 node->hole_size = 0;
293 node->subtree_max_hole = 0;
294
295 DRM_MM_BUG_ON(drm_mm_hole_follows(node));
296}
297
298static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb)
299{
300 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size);
301}
302
303static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb)
304{
305 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr);
306}
307
308static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size)
309{
310 struct rb_node *rb = mm->holes_size.rb_root.rb_node;
311 struct drm_mm_node *best = NULL;
312
313 do {
314 struct drm_mm_node *node =
315 rb_entry(rb, struct drm_mm_node, rb_hole_size);
316
317 if (size <= node->hole_size) {
318 best = node;
319 rb = rb->rb_right;
320 } else {
321 rb = rb->rb_left;
322 }
323 } while (rb);
324
325 return best;
326}
327
328static bool usable_hole_addr(struct rb_node *rb, u64 size)
329{
330 return rb && rb_hole_addr_to_node(rb)->subtree_max_hole >= size;
331}
332
333static struct drm_mm_node *find_hole_addr(struct drm_mm *mm, u64 addr, u64 size)
334{
335 struct rb_node *rb = mm->holes_addr.rb_node;
336 struct drm_mm_node *node = NULL;
337
338 while (rb) {
339 u64 hole_start;
340
341 if (!usable_hole_addr(rb, size))
342 break;
343
344 node = rb_hole_addr_to_node(rb);
345 hole_start = __drm_mm_hole_node_start(node);
346
347 if (addr < hole_start)
348 rb = node->rb_hole_addr.rb_left;
349 else if (addr > hole_start + node->hole_size)
350 rb = node->rb_hole_addr.rb_right;
351 else
352 break;
353 }
354
355 return node;
356}
357
358static struct drm_mm_node *
359first_hole(struct drm_mm *mm,
360 u64 start, u64 end, u64 size,
361 enum drm_mm_insert_mode mode)
362{
363 switch (mode) {
364 default:
365 case DRM_MM_INSERT_BEST:
366 return best_hole(mm, size);
367
368 case DRM_MM_INSERT_LOW:
369 return find_hole_addr(mm, start, size);
370
371 case DRM_MM_INSERT_HIGH:
372 return find_hole_addr(mm, end, size);
373
374 case DRM_MM_INSERT_EVICT:
375 return list_first_entry_or_null(&mm->hole_stack,
376 struct drm_mm_node,
377 hole_stack);
378 }
379}
380
381/**
382 * DECLARE_NEXT_HOLE_ADDR - macro to declare next hole functions
383 * @name: name of function to declare
384 * @first: first rb member to traverse (either rb_left or rb_right).
385 * @last: last rb member to traverse (either rb_right or rb_left).
386 *
387 * This macro declares a function to return the next hole of the addr rb tree.
388 * While traversing the tree we take the searched size into account and only
389 * visit branches with potential big enough holes.
390 */
391
392#define DECLARE_NEXT_HOLE_ADDR(name, first, last) \
393static struct drm_mm_node *name(struct drm_mm_node *entry, u64 size) \
394{ \
395 struct rb_node *parent, *node = &entry->rb_hole_addr; \
396 \
397 if (!entry || RB_EMPTY_NODE(node)) \
398 return NULL; \
399 \
400 if (usable_hole_addr(node->first, size)) { \
401 node = node->first; \
402 while (usable_hole_addr(node->last, size)) \
403 node = node->last; \
404 return rb_hole_addr_to_node(node); \
405 } \
406 \
407 while ((parent = rb_parent(node)) && node == parent->first) \
408 node = parent; \
409 \
410 return rb_hole_addr_to_node(parent); \
411}
412
413DECLARE_NEXT_HOLE_ADDR(next_hole_high_addr, rb_left, rb_right)
414DECLARE_NEXT_HOLE_ADDR(next_hole_low_addr, rb_right, rb_left)
415
416static struct drm_mm_node *
417next_hole(struct drm_mm *mm,
418 struct drm_mm_node *node,
419 u64 size,
420 enum drm_mm_insert_mode mode)
421{
422 switch (mode) {
423 default:
424 case DRM_MM_INSERT_BEST:
425 return rb_hole_size_to_node(rb_prev(&node->rb_hole_size));
426
427 case DRM_MM_INSERT_LOW:
428 return next_hole_low_addr(node, size);
429
430 case DRM_MM_INSERT_HIGH:
431 return next_hole_high_addr(node, size);
432
433 case DRM_MM_INSERT_EVICT:
434 node = list_next_entry(node, hole_stack);
435 return &node->hole_stack == &mm->hole_stack ? NULL : node;
436 }
437}
438
439/**
440 * drm_mm_reserve_node - insert an pre-initialized node
441 * @mm: drm_mm allocator to insert @node into
442 * @node: drm_mm_node to insert
443 *
444 * This functions inserts an already set-up &drm_mm_node into the allocator,
445 * meaning that start, size and color must be set by the caller. All other
446 * fields must be cleared to 0. This is useful to initialize the allocator with
447 * preallocated objects which must be set-up before the range allocator can be
448 * set-up, e.g. when taking over a firmware framebuffer.
449 *
450 * Returns:
451 * 0 on success, -ENOSPC if there's no hole where @node is.
452 */
453int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
454{
455 struct drm_mm_node *hole;
456 u64 hole_start, hole_end;
457 u64 adj_start, adj_end;
458 u64 end;
459
460 end = node->start + node->size;
461 if (unlikely(end <= node->start))
462 return -ENOSPC;
463
464 /* Find the relevant hole to add our node to */
465 hole = find_hole_addr(mm, node->start, 0);
466 if (!hole)
467 return -ENOSPC;
468
469 adj_start = hole_start = __drm_mm_hole_node_start(hole);
470 adj_end = hole_end = hole_start + hole->hole_size;
471
472 if (mm->color_adjust)
473 mm->color_adjust(hole, node->color, &adj_start, &adj_end);
474
475 if (adj_start > node->start || adj_end < end)
476 return -ENOSPC;
477
478 node->mm = mm;
479
480 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
481 list_add(&node->node_list, &hole->node_list);
482 drm_mm_interval_tree_add_node(hole, node);
483 node->hole_size = 0;
484
485 rm_hole(hole);
486 if (node->start > hole_start)
487 add_hole(hole);
488 if (end < hole_end)
489 add_hole(node);
490
491 save_stack(node);
492 return 0;
493}
494EXPORT_SYMBOL(drm_mm_reserve_node);
495
496static u64 rb_to_hole_size_or_zero(struct rb_node *rb)
497{
498 return rb ? rb_to_hole_size(rb) : 0;
499}
500
501/**
502 * drm_mm_insert_node_in_range - ranged search for space and insert @node
503 * @mm: drm_mm to allocate from
504 * @node: preallocate node to insert
505 * @size: size of the allocation
506 * @alignment: alignment of the allocation
507 * @color: opaque tag value to use for this node
508 * @range_start: start of the allowed range for this node
509 * @range_end: end of the allowed range for this node
510 * @mode: fine-tune the allocation search and placement
511 *
512 * The preallocated @node must be cleared to 0.
513 *
514 * Returns:
515 * 0 on success, -ENOSPC if there's no suitable hole.
516 */
517int drm_mm_insert_node_in_range(struct drm_mm * const mm,
518 struct drm_mm_node * const node,
519 u64 size, u64 alignment,
520 unsigned long color,
521 u64 range_start, u64 range_end,
522 enum drm_mm_insert_mode mode)
523{
524 struct drm_mm_node *hole;
525 u64 remainder_mask;
526 bool once;
527
528 DRM_MM_BUG_ON(range_start > range_end);
529
530 if (unlikely(size == 0 || range_end - range_start < size))
531 return -ENOSPC;
532
533 if (rb_to_hole_size_or_zero(rb_first_cached(&mm->holes_size)) < size)
534 return -ENOSPC;
535
536 if (alignment <= 1)
537 alignment = 0;
538
539 once = mode & DRM_MM_INSERT_ONCE;
540 mode &= ~DRM_MM_INSERT_ONCE;
541
542 remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
543 for (hole = first_hole(mm, range_start, range_end, size, mode);
544 hole;
545 hole = once ? NULL : next_hole(mm, hole, size, mode)) {
546 u64 hole_start = __drm_mm_hole_node_start(hole);
547 u64 hole_end = hole_start + hole->hole_size;
548 u64 adj_start, adj_end;
549 u64 col_start, col_end;
550
551 if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end)
552 break;
553
554 if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start)
555 break;
556
557 col_start = hole_start;
558 col_end = hole_end;
559 if (mm->color_adjust)
560 mm->color_adjust(hole, color, &col_start, &col_end);
561
562 adj_start = max(col_start, range_start);
563 adj_end = min(col_end, range_end);
564
565 if (adj_end <= adj_start || adj_end - adj_start < size)
566 continue;
567
568 if (mode == DRM_MM_INSERT_HIGH)
569 adj_start = adj_end - size;
570
571 if (alignment) {
572 u64 rem;
573
574 if (likely(remainder_mask))
575 rem = adj_start & remainder_mask;
576 else
577 div64_u64_rem(adj_start, alignment, &rem);
578 if (rem) {
579 adj_start -= rem;
580 if (mode != DRM_MM_INSERT_HIGH)
581 adj_start += alignment;
582
583 if (adj_start < max(col_start, range_start) ||
584 min(col_end, range_end) - adj_start < size)
585 continue;
586
587 if (adj_end <= adj_start ||
588 adj_end - adj_start < size)
589 continue;
590 }
591 }
592
593 node->mm = mm;
594 node->size = size;
595 node->start = adj_start;
596 node->color = color;
597 node->hole_size = 0;
598
599 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
600 list_add(&node->node_list, &hole->node_list);
601 drm_mm_interval_tree_add_node(hole, node);
602
603 rm_hole(hole);
604 if (adj_start > hole_start)
605 add_hole(hole);
606 if (adj_start + size < hole_end)
607 add_hole(node);
608
609 save_stack(node);
610 return 0;
611 }
612
613 return -ENOSPC;
614}
615EXPORT_SYMBOL(drm_mm_insert_node_in_range);
616
617static inline bool drm_mm_node_scanned_block(const struct drm_mm_node *node)
618{
619 return test_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
620}
621
622/**
623 * drm_mm_remove_node - Remove a memory node from the allocator.
624 * @node: drm_mm_node to remove
625 *
626 * This just removes a node from its drm_mm allocator. The node does not need to
627 * be cleared again before it can be re-inserted into this or any other drm_mm
628 * allocator. It is a bug to call this function on a unallocated node.
629 */
630void drm_mm_remove_node(struct drm_mm_node *node)
631{
632 struct drm_mm *mm = node->mm;
633 struct drm_mm_node *prev_node;
634
635 DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
636 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
637
638 prev_node = list_prev_entry(node, node_list);
639
640 if (drm_mm_hole_follows(node))
641 rm_hole(node);
642
643 drm_mm_interval_tree_remove(node, &mm->interval_tree);
644 list_del(&node->node_list);
645
646 if (drm_mm_hole_follows(prev_node))
647 rm_hole(prev_node);
648 add_hole(prev_node);
649
650 clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
651}
652EXPORT_SYMBOL(drm_mm_remove_node);
653
654/**
655 * drm_mm_replace_node - move an allocation from @old to @new
656 * @old: drm_mm_node to remove from the allocator
657 * @new: drm_mm_node which should inherit @old's allocation
658 *
659 * This is useful for when drivers embed the drm_mm_node structure and hence
660 * can't move allocations by reassigning pointers. It's a combination of remove
661 * and insert with the guarantee that the allocation start will match.
662 */
663void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
664{
665 struct drm_mm *mm = old->mm;
666
667 DRM_MM_BUG_ON(!drm_mm_node_allocated(old));
668
669 *new = *old;
670
671 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &new->flags);
672 list_replace(&old->node_list, &new->node_list);
673 rb_replace_node_cached(&old->rb, &new->rb, &mm->interval_tree);
674
675 if (drm_mm_hole_follows(old)) {
676 list_replace(&old->hole_stack, &new->hole_stack);
677 rb_replace_node_cached(&old->rb_hole_size,
678 &new->rb_hole_size,
679 &mm->holes_size);
680 rb_replace_node(&old->rb_hole_addr,
681 &new->rb_hole_addr,
682 &mm->holes_addr);
683 }
684
685 clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT, &old->flags);
686}
687EXPORT_SYMBOL(drm_mm_replace_node);
688
689/**
690 * DOC: lru scan roster
691 *
692 * Very often GPUs need to have continuous allocations for a given object. When
693 * evicting objects to make space for a new one it is therefore not most
694 * efficient when we simply start to select all objects from the tail of an LRU
695 * until there's a suitable hole: Especially for big objects or nodes that
696 * otherwise have special allocation constraints there's a good chance we evict
697 * lots of (smaller) objects unnecessarily.
698 *
699 * The DRM range allocator supports this use-case through the scanning
700 * interfaces. First a scan operation needs to be initialized with
701 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
702 * objects to the roster, probably by walking an LRU list, but this can be
703 * freely implemented. Eviction candiates are added using
704 * drm_mm_scan_add_block() until a suitable hole is found or there are no
705 * further evictable objects. Eviction roster metadata is tracked in &struct
706 * drm_mm_scan.
707 *
708 * The driver must walk through all objects again in exactly the reverse
709 * order to restore the allocator state. Note that while the allocator is used
710 * in the scan mode no other operation is allowed.
711 *
712 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
713 * reported true) in the scan, and any overlapping nodes after color adjustment
714 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
715 * since freeing a node is also O(1) the overall complexity is
716 * O(scanned_objects). So like the free stack which needs to be walked before a
717 * scan operation even begins this is linear in the number of objects. It
718 * doesn't seem to hurt too badly.
719 */
720
721/**
722 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
723 * @scan: scan state
724 * @mm: drm_mm to scan
725 * @size: size of the allocation
726 * @alignment: alignment of the allocation
727 * @color: opaque tag value to use for the allocation
728 * @start: start of the allowed range for the allocation
729 * @end: end of the allowed range for the allocation
730 * @mode: fine-tune the allocation search and placement
731 *
732 * This simply sets up the scanning routines with the parameters for the desired
733 * hole.
734 *
735 * Warning:
736 * As long as the scan list is non-empty, no other operations than
737 * adding/removing nodes to/from the scan list are allowed.
738 */
739void drm_mm_scan_init_with_range(struct drm_mm_scan *scan,
740 struct drm_mm *mm,
741 u64 size,
742 u64 alignment,
743 unsigned long color,
744 u64 start,
745 u64 end,
746 enum drm_mm_insert_mode mode)
747{
748 DRM_MM_BUG_ON(start >= end);
749 DRM_MM_BUG_ON(!size || size > end - start);
750 DRM_MM_BUG_ON(mm->scan_active);
751
752 scan->mm = mm;
753
754 if (alignment <= 1)
755 alignment = 0;
756
757 scan->color = color;
758 scan->alignment = alignment;
759 scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
760 scan->size = size;
761 scan->mode = mode;
762
763 DRM_MM_BUG_ON(end <= start);
764 scan->range_start = start;
765 scan->range_end = end;
766
767 scan->hit_start = U64_MAX;
768 scan->hit_end = 0;
769}
770EXPORT_SYMBOL(drm_mm_scan_init_with_range);
771
772/**
773 * drm_mm_scan_add_block - add a node to the scan list
774 * @scan: the active drm_mm scanner
775 * @node: drm_mm_node to add
776 *
777 * Add a node to the scan list that might be freed to make space for the desired
778 * hole.
779 *
780 * Returns:
781 * True if a hole has been found, false otherwise.
782 */
783bool drm_mm_scan_add_block(struct drm_mm_scan *scan,
784 struct drm_mm_node *node)
785{
786 struct drm_mm *mm = scan->mm;
787 struct drm_mm_node *hole;
788 u64 hole_start, hole_end;
789 u64 col_start, col_end;
790 u64 adj_start, adj_end;
791
792 DRM_MM_BUG_ON(node->mm != mm);
793 DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
794 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
795 __set_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
796 mm->scan_active++;
797
798 /* Remove this block from the node_list so that we enlarge the hole
799 * (distance between the end of our previous node and the start of
800 * or next), without poisoning the link so that we can restore it
801 * later in drm_mm_scan_remove_block().
802 */
803 hole = list_prev_entry(node, node_list);
804 DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node);
805 __list_del_entry(&node->node_list);
806
807 hole_start = __drm_mm_hole_node_start(hole);
808 hole_end = __drm_mm_hole_node_end(hole);
809
810 col_start = hole_start;
811 col_end = hole_end;
812 if (mm->color_adjust)
813 mm->color_adjust(hole, scan->color, &col_start, &col_end);
814
815 adj_start = max(col_start, scan->range_start);
816 adj_end = min(col_end, scan->range_end);
817 if (adj_end <= adj_start || adj_end - adj_start < scan->size)
818 return false;
819
820 if (scan->mode == DRM_MM_INSERT_HIGH)
821 adj_start = adj_end - scan->size;
822
823 if (scan->alignment) {
824 u64 rem;
825
826 if (likely(scan->remainder_mask))
827 rem = adj_start & scan->remainder_mask;
828 else
829 div64_u64_rem(adj_start, scan->alignment, &rem);
830 if (rem) {
831 adj_start -= rem;
832 if (scan->mode != DRM_MM_INSERT_HIGH)
833 adj_start += scan->alignment;
834 if (adj_start < max(col_start, scan->range_start) ||
835 min(col_end, scan->range_end) - adj_start < scan->size)
836 return false;
837
838 if (adj_end <= adj_start ||
839 adj_end - adj_start < scan->size)
840 return false;
841 }
842 }
843
844 scan->hit_start = adj_start;
845 scan->hit_end = adj_start + scan->size;
846
847 DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end);
848 DRM_MM_BUG_ON(scan->hit_start < hole_start);
849 DRM_MM_BUG_ON(scan->hit_end > hole_end);
850
851 return true;
852}
853EXPORT_SYMBOL(drm_mm_scan_add_block);
854
855/**
856 * drm_mm_scan_remove_block - remove a node from the scan list
857 * @scan: the active drm_mm scanner
858 * @node: drm_mm_node to remove
859 *
860 * Nodes **must** be removed in exactly the reverse order from the scan list as
861 * they have been added (e.g. using list_add() as they are added and then
862 * list_for_each() over that eviction list to remove), otherwise the internal
863 * state of the memory manager will be corrupted.
864 *
865 * When the scan list is empty, the selected memory nodes can be freed. An
866 * immediately following drm_mm_insert_node_in_range_generic() or one of the
867 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
868 * the just freed block (because it's at the top of the free_stack list).
869 *
870 * Returns:
871 * True if this block should be evicted, false otherwise. Will always
872 * return false when no hole has been found.
873 */
874bool drm_mm_scan_remove_block(struct drm_mm_scan *scan,
875 struct drm_mm_node *node)
876{
877 struct drm_mm_node *prev_node;
878
879 DRM_MM_BUG_ON(node->mm != scan->mm);
880 DRM_MM_BUG_ON(!drm_mm_node_scanned_block(node));
881 __clear_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
882
883 DRM_MM_BUG_ON(!node->mm->scan_active);
884 node->mm->scan_active--;
885
886 /* During drm_mm_scan_add_block() we decoupled this node leaving
887 * its pointers intact. Now that the caller is walking back along
888 * the eviction list we can restore this block into its rightful
889 * place on the full node_list. To confirm that the caller is walking
890 * backwards correctly we check that prev_node->next == node->next,
891 * i.e. both believe the same node should be on the other side of the
892 * hole.
893 */
894 prev_node = list_prev_entry(node, node_list);
895 DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) !=
896 list_next_entry(node, node_list));
897 list_add(&node->node_list, &prev_node->node_list);
898
899 return (node->start + node->size > scan->hit_start &&
900 node->start < scan->hit_end);
901}
902EXPORT_SYMBOL(drm_mm_scan_remove_block);
903
904/**
905 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
906 * @scan: drm_mm scan with target hole
907 *
908 * After completing an eviction scan and removing the selected nodes, we may
909 * need to remove a few more nodes from either side of the target hole if
910 * mm.color_adjust is being used.
911 *
912 * Returns:
913 * A node to evict, or NULL if there are no overlapping nodes.
914 */
915struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan)
916{
917 struct drm_mm *mm = scan->mm;
918 struct drm_mm_node *hole;
919 u64 hole_start, hole_end;
920
921 DRM_MM_BUG_ON(list_empty(&mm->hole_stack));
922
923 if (!mm->color_adjust)
924 return NULL;
925
926 /*
927 * The hole found during scanning should ideally be the first element
928 * in the hole_stack list, but due to side-effects in the driver it
929 * may not be.
930 */
931 list_for_each_entry(hole, &mm->hole_stack, hole_stack) {
932 hole_start = __drm_mm_hole_node_start(hole);
933 hole_end = hole_start + hole->hole_size;
934
935 if (hole_start <= scan->hit_start &&
936 hole_end >= scan->hit_end)
937 break;
938 }
939
940 /* We should only be called after we found the hole previously */
941 DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack);
942 if (unlikely(&hole->hole_stack == &mm->hole_stack))
943 return NULL;
944
945 DRM_MM_BUG_ON(hole_start > scan->hit_start);
946 DRM_MM_BUG_ON(hole_end < scan->hit_end);
947
948 mm->color_adjust(hole, scan->color, &hole_start, &hole_end);
949 if (hole_start > scan->hit_start)
950 return hole;
951 if (hole_end < scan->hit_end)
952 return list_next_entry(hole, node_list);
953
954 return NULL;
955}
956EXPORT_SYMBOL(drm_mm_scan_color_evict);
957
958/**
959 * drm_mm_init - initialize a drm-mm allocator
960 * @mm: the drm_mm structure to initialize
961 * @start: start of the range managed by @mm
962 * @size: end of the range managed by @mm
963 *
964 * Note that @mm must be cleared to 0 before calling this function.
965 */
966void drm_mm_init(struct drm_mm *mm, u64 start, u64 size)
967{
968 DRM_MM_BUG_ON(start + size <= start);
969
970 mm->color_adjust = NULL;
971
972 INIT_LIST_HEAD(&mm->hole_stack);
973 mm->interval_tree = RB_ROOT_CACHED;
974 mm->holes_size = RB_ROOT_CACHED;
975 mm->holes_addr = RB_ROOT;
976
977 /* Clever trick to avoid a special case in the free hole tracking. */
978 INIT_LIST_HEAD(&mm->head_node.node_list);
979 mm->head_node.flags = 0;
980 mm->head_node.mm = mm;
981 mm->head_node.start = start + size;
982 mm->head_node.size = -size;
983 add_hole(&mm->head_node);
984
985 mm->scan_active = 0;
986}
987EXPORT_SYMBOL(drm_mm_init);
988
989/**
990 * drm_mm_takedown - clean up a drm_mm allocator
991 * @mm: drm_mm allocator to clean up
992 *
993 * Note that it is a bug to call this function on an allocator which is not
994 * clean.
995 */
996void drm_mm_takedown(struct drm_mm *mm)
997{
998 if (WARN(!drm_mm_clean(mm),
999 "Memory manager not clean during takedown.\n"))
1000 show_leaks(mm);
1001}
1002EXPORT_SYMBOL(drm_mm_takedown);
1003
1004static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry)
1005{
1006 u64 start, size;
1007
1008 size = entry->hole_size;
1009 if (size) {
1010 start = drm_mm_hole_node_start(entry);
1011 drm_printf(p, "%#018llx-%#018llx: %llu: free\n",
1012 start, start + size, size);
1013 }
1014
1015 return size;
1016}
1017/**
1018 * drm_mm_print - print allocator state
1019 * @mm: drm_mm allocator to print
1020 * @p: DRM printer to use
1021 */
1022void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p)
1023{
1024 const struct drm_mm_node *entry;
1025 u64 total_used = 0, total_free = 0, total = 0;
1026
1027 total_free += drm_mm_dump_hole(p, &mm->head_node);
1028
1029 drm_mm_for_each_node(entry, mm) {
1030 drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start,
1031 entry->start + entry->size, entry->size);
1032 total_used += entry->size;
1033 total_free += drm_mm_dump_hole(p, entry);
1034 }
1035 total = total_free + total_used;
1036
1037 drm_printf(p, "total: %llu, used %llu free %llu\n", total,
1038 total_used, total_free);
1039}
1040EXPORT_SYMBOL(drm_mm_print);