Loading...
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);
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 <drm/drmP.h>
45#include <drm/drm_mm.h>
46#include <linux/slab.h>
47#include <linux/seq_file.h>
48#include <linux/export.h>
49#include <linux/interval_tree_generic.h>
50
51/**
52 * DOC: Overview
53 *
54 * drm_mm provides a simple range allocator. The drivers are free to use the
55 * resource allocator from the linux core if it suits them, the upside of drm_mm
56 * is that it's in the DRM core. Which means that it's easier to extend for
57 * some of the crazier special purpose needs of gpus.
58 *
59 * The main data struct is &drm_mm, allocations are tracked in &drm_mm_node.
60 * Drivers are free to embed either of them into their own suitable
61 * datastructures. drm_mm itself will not do any allocations of its own, so if
62 * drivers choose not to embed nodes they need to still allocate them
63 * themselves.
64 *
65 * The range allocator also supports reservation of preallocated blocks. This is
66 * useful for taking over initial mode setting configurations from the firmware,
67 * where an object needs to be created which exactly matches the firmware's
68 * scanout target. As long as the range is still free it can be inserted anytime
69 * after the allocator is initialized, which helps with avoiding looped
70 * depencies in the driver load sequence.
71 *
72 * drm_mm maintains a stack of most recently freed holes, which of all
73 * simplistic datastructures seems to be a fairly decent approach to clustering
74 * allocations and avoiding too much fragmentation. This means free space
75 * searches are O(num_holes). Given that all the fancy features drm_mm supports
76 * something better would be fairly complex and since gfx thrashing is a fairly
77 * steep cliff not a real concern. Removing a node again is O(1).
78 *
79 * drm_mm supports a few features: Alignment and range restrictions can be
80 * supplied. Further more every &drm_mm_node has a color value (which is just an
81 * opaqua unsigned long) which in conjunction with a driver callback can be used
82 * to implement sophisticated placement restrictions. The i915 DRM driver uses
83 * this to implement guard pages between incompatible caching domains in the
84 * graphics TT.
85 *
86 * Two behaviors are supported for searching and allocating: bottom-up and top-down.
87 * The default is bottom-up. Top-down allocation can be used if the memory area
88 * has different restrictions, or just to reduce fragmentation.
89 *
90 * Finally iteration helpers to walk all nodes and all holes are provided as are
91 * some basic allocator dumpers for debugging.
92 */
93
94static struct drm_mm_node *drm_mm_search_free_generic(const struct drm_mm *mm,
95 u64 size,
96 unsigned alignment,
97 unsigned long color,
98 enum drm_mm_search_flags flags);
99static struct drm_mm_node *drm_mm_search_free_in_range_generic(const struct drm_mm *mm,
100 u64 size,
101 unsigned alignment,
102 unsigned long color,
103 u64 start,
104 u64 end,
105 enum drm_mm_search_flags flags);
106
107#ifdef CONFIG_DRM_DEBUG_MM
108#include <linux/stackdepot.h>
109
110#define STACKDEPTH 32
111#define BUFSZ 4096
112
113static noinline void save_stack(struct drm_mm_node *node)
114{
115 unsigned long entries[STACKDEPTH];
116 struct stack_trace trace = {
117 .entries = entries,
118 .max_entries = STACKDEPTH,
119 .skip = 1
120 };
121
122 save_stack_trace(&trace);
123 if (trace.nr_entries != 0 &&
124 trace.entries[trace.nr_entries-1] == ULONG_MAX)
125 trace.nr_entries--;
126
127 /* May be called under spinlock, so avoid sleeping */
128 node->stack = depot_save_stack(&trace, GFP_NOWAIT);
129}
130
131static void show_leaks(struct drm_mm *mm)
132{
133 struct drm_mm_node *node;
134 unsigned long entries[STACKDEPTH];
135 char *buf;
136
137 buf = kmalloc(BUFSZ, GFP_KERNEL);
138 if (!buf)
139 return;
140
141 list_for_each_entry(node, &mm->head_node.node_list, node_list) {
142 struct stack_trace trace = {
143 .entries = entries,
144 .max_entries = STACKDEPTH
145 };
146
147 if (!node->stack) {
148 DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
149 node->start, node->size);
150 continue;
151 }
152
153 depot_fetch_stack(node->stack, &trace);
154 snprint_stack_trace(buf, BUFSZ, &trace, 0);
155 DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
156 node->start, node->size, buf);
157 }
158
159 kfree(buf);
160}
161
162#undef STACKDEPTH
163#undef BUFSZ
164#else
165static void save_stack(struct drm_mm_node *node) { }
166static void show_leaks(struct drm_mm *mm) { }
167#endif
168
169#define START(node) ((node)->start)
170#define LAST(node) ((node)->start + (node)->size - 1)
171
172INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
173 u64, __subtree_last,
174 START, LAST, static inline, drm_mm_interval_tree)
175
176struct drm_mm_node *
177__drm_mm_interval_first(struct drm_mm *mm, u64 start, u64 last)
178{
179 return drm_mm_interval_tree_iter_first(&mm->interval_tree,
180 start, last);
181}
182EXPORT_SYMBOL(__drm_mm_interval_first);
183
184static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
185 struct drm_mm_node *node)
186{
187 struct drm_mm *mm = hole_node->mm;
188 struct rb_node **link, *rb;
189 struct drm_mm_node *parent;
190
191 node->__subtree_last = LAST(node);
192
193 if (hole_node->allocated) {
194 rb = &hole_node->rb;
195 while (rb) {
196 parent = rb_entry(rb, struct drm_mm_node, rb);
197 if (parent->__subtree_last >= node->__subtree_last)
198 break;
199
200 parent->__subtree_last = node->__subtree_last;
201 rb = rb_parent(rb);
202 }
203
204 rb = &hole_node->rb;
205 link = &hole_node->rb.rb_right;
206 } else {
207 rb = NULL;
208 link = &mm->interval_tree.rb_node;
209 }
210
211 while (*link) {
212 rb = *link;
213 parent = rb_entry(rb, struct drm_mm_node, rb);
214 if (parent->__subtree_last < node->__subtree_last)
215 parent->__subtree_last = node->__subtree_last;
216 if (node->start < parent->start)
217 link = &parent->rb.rb_left;
218 else
219 link = &parent->rb.rb_right;
220 }
221
222 rb_link_node(&node->rb, rb, link);
223 rb_insert_augmented(&node->rb,
224 &mm->interval_tree,
225 &drm_mm_interval_tree_augment);
226}
227
228static void drm_mm_insert_helper(struct drm_mm_node *hole_node,
229 struct drm_mm_node *node,
230 u64 size, unsigned alignment,
231 unsigned long color,
232 enum drm_mm_allocator_flags flags)
233{
234 struct drm_mm *mm = hole_node->mm;
235 u64 hole_start = drm_mm_hole_node_start(hole_node);
236 u64 hole_end = drm_mm_hole_node_end(hole_node);
237 u64 adj_start = hole_start;
238 u64 adj_end = hole_end;
239
240 BUG_ON(node->allocated);
241
242 if (mm->color_adjust)
243 mm->color_adjust(hole_node, color, &adj_start, &adj_end);
244
245 if (flags & DRM_MM_CREATE_TOP)
246 adj_start = adj_end - size;
247
248 if (alignment) {
249 u64 tmp = adj_start;
250 unsigned rem;
251
252 rem = do_div(tmp, alignment);
253 if (rem) {
254 if (flags & DRM_MM_CREATE_TOP)
255 adj_start -= rem;
256 else
257 adj_start += alignment - rem;
258 }
259 }
260
261 BUG_ON(adj_start < hole_start);
262 BUG_ON(adj_end > hole_end);
263
264 if (adj_start == hole_start) {
265 hole_node->hole_follows = 0;
266 list_del(&hole_node->hole_stack);
267 }
268
269 node->start = adj_start;
270 node->size = size;
271 node->mm = mm;
272 node->color = color;
273 node->allocated = 1;
274
275 list_add(&node->node_list, &hole_node->node_list);
276
277 drm_mm_interval_tree_add_node(hole_node, node);
278
279 BUG_ON(node->start + node->size > adj_end);
280
281 node->hole_follows = 0;
282 if (__drm_mm_hole_node_start(node) < hole_end) {
283 list_add(&node->hole_stack, &mm->hole_stack);
284 node->hole_follows = 1;
285 }
286
287 save_stack(node);
288}
289
290/**
291 * drm_mm_reserve_node - insert an pre-initialized node
292 * @mm: drm_mm allocator to insert @node into
293 * @node: drm_mm_node to insert
294 *
295 * This functions inserts an already set-up drm_mm_node into the allocator,
296 * meaning that start, size and color must be set by the caller. This is useful
297 * to initialize the allocator with preallocated objects which must be set-up
298 * before the range allocator can be set-up, e.g. when taking over a firmware
299 * framebuffer.
300 *
301 * Returns:
302 * 0 on success, -ENOSPC if there's no hole where @node is.
303 */
304int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
305{
306 u64 end = node->start + node->size;
307 struct drm_mm_node *hole;
308 u64 hole_start, hole_end;
309 u64 adj_start, adj_end;
310
311 if (WARN_ON(node->size == 0))
312 return -EINVAL;
313
314 end = node->start + node->size;
315
316 /* Find the relevant hole to add our node to */
317 hole = drm_mm_interval_tree_iter_first(&mm->interval_tree,
318 node->start, ~(u64)0);
319 if (hole) {
320 if (hole->start < end)
321 return -ENOSPC;
322 } else {
323 hole = list_entry(&mm->head_node.node_list,
324 typeof(*hole), node_list);
325 }
326
327 hole = list_last_entry(&hole->node_list, typeof(*hole), node_list);
328 if (!hole->hole_follows)
329 return -ENOSPC;
330
331 adj_start = hole_start = __drm_mm_hole_node_start(hole);
332 adj_end = hole_end = __drm_mm_hole_node_end(hole);
333
334 if (mm->color_adjust)
335 mm->color_adjust(hole, node->color, &adj_start, &adj_end);
336
337 if (adj_start > node->start || adj_end < end)
338 return -ENOSPC;
339
340 node->mm = mm;
341 node->allocated = 1;
342
343 list_add(&node->node_list, &hole->node_list);
344
345 drm_mm_interval_tree_add_node(hole, node);
346
347 if (node->start == hole_start) {
348 hole->hole_follows = 0;
349 list_del(&hole->hole_stack);
350 }
351
352 node->hole_follows = 0;
353 if (end != hole_end) {
354 list_add(&node->hole_stack, &mm->hole_stack);
355 node->hole_follows = 1;
356 }
357
358 save_stack(node);
359
360 return 0;
361}
362EXPORT_SYMBOL(drm_mm_reserve_node);
363
364/**
365 * drm_mm_insert_node_generic - search for space and insert @node
366 * @mm: drm_mm to allocate from
367 * @node: preallocate node to insert
368 * @size: size of the allocation
369 * @alignment: alignment of the allocation
370 * @color: opaque tag value to use for this node
371 * @sflags: flags to fine-tune the allocation search
372 * @aflags: flags to fine-tune the allocation behavior
373 *
374 * The preallocated node must be cleared to 0.
375 *
376 * Returns:
377 * 0 on success, -ENOSPC if there's no suitable hole.
378 */
379int drm_mm_insert_node_generic(struct drm_mm *mm, struct drm_mm_node *node,
380 u64 size, unsigned alignment,
381 unsigned long color,
382 enum drm_mm_search_flags sflags,
383 enum drm_mm_allocator_flags aflags)
384{
385 struct drm_mm_node *hole_node;
386
387 if (WARN_ON(size == 0))
388 return -EINVAL;
389
390 hole_node = drm_mm_search_free_generic(mm, size, alignment,
391 color, sflags);
392 if (!hole_node)
393 return -ENOSPC;
394
395 drm_mm_insert_helper(hole_node, node, size, alignment, color, aflags);
396 return 0;
397}
398EXPORT_SYMBOL(drm_mm_insert_node_generic);
399
400static void drm_mm_insert_helper_range(struct drm_mm_node *hole_node,
401 struct drm_mm_node *node,
402 u64 size, unsigned alignment,
403 unsigned long color,
404 u64 start, u64 end,
405 enum drm_mm_allocator_flags flags)
406{
407 struct drm_mm *mm = hole_node->mm;
408 u64 hole_start = drm_mm_hole_node_start(hole_node);
409 u64 hole_end = drm_mm_hole_node_end(hole_node);
410 u64 adj_start = hole_start;
411 u64 adj_end = hole_end;
412
413 BUG_ON(!hole_node->hole_follows || node->allocated);
414
415 if (adj_start < start)
416 adj_start = start;
417 if (adj_end > end)
418 adj_end = end;
419
420 if (mm->color_adjust)
421 mm->color_adjust(hole_node, color, &adj_start, &adj_end);
422
423 if (flags & DRM_MM_CREATE_TOP)
424 adj_start = adj_end - size;
425
426 if (alignment) {
427 u64 tmp = adj_start;
428 unsigned rem;
429
430 rem = do_div(tmp, alignment);
431 if (rem) {
432 if (flags & DRM_MM_CREATE_TOP)
433 adj_start -= rem;
434 else
435 adj_start += alignment - rem;
436 }
437 }
438
439 if (adj_start == hole_start) {
440 hole_node->hole_follows = 0;
441 list_del(&hole_node->hole_stack);
442 }
443
444 node->start = adj_start;
445 node->size = size;
446 node->mm = mm;
447 node->color = color;
448 node->allocated = 1;
449
450 list_add(&node->node_list, &hole_node->node_list);
451
452 drm_mm_interval_tree_add_node(hole_node, node);
453
454 BUG_ON(node->start < start);
455 BUG_ON(node->start < adj_start);
456 BUG_ON(node->start + node->size > adj_end);
457 BUG_ON(node->start + node->size > end);
458
459 node->hole_follows = 0;
460 if (__drm_mm_hole_node_start(node) < hole_end) {
461 list_add(&node->hole_stack, &mm->hole_stack);
462 node->hole_follows = 1;
463 }
464
465 save_stack(node);
466}
467
468/**
469 * drm_mm_insert_node_in_range_generic - ranged search for space and insert @node
470 * @mm: drm_mm to allocate from
471 * @node: preallocate node to insert
472 * @size: size of the allocation
473 * @alignment: alignment of the allocation
474 * @color: opaque tag value to use for this node
475 * @start: start of the allowed range for this node
476 * @end: end of the allowed range for this node
477 * @sflags: flags to fine-tune the allocation search
478 * @aflags: flags to fine-tune the allocation behavior
479 *
480 * The preallocated node must be cleared to 0.
481 *
482 * Returns:
483 * 0 on success, -ENOSPC if there's no suitable hole.
484 */
485int drm_mm_insert_node_in_range_generic(struct drm_mm *mm, struct drm_mm_node *node,
486 u64 size, unsigned alignment,
487 unsigned long color,
488 u64 start, u64 end,
489 enum drm_mm_search_flags sflags,
490 enum drm_mm_allocator_flags aflags)
491{
492 struct drm_mm_node *hole_node;
493
494 if (WARN_ON(size == 0))
495 return -EINVAL;
496
497 hole_node = drm_mm_search_free_in_range_generic(mm,
498 size, alignment, color,
499 start, end, sflags);
500 if (!hole_node)
501 return -ENOSPC;
502
503 drm_mm_insert_helper_range(hole_node, node,
504 size, alignment, color,
505 start, end, aflags);
506 return 0;
507}
508EXPORT_SYMBOL(drm_mm_insert_node_in_range_generic);
509
510/**
511 * drm_mm_remove_node - Remove a memory node from the allocator.
512 * @node: drm_mm_node to remove
513 *
514 * This just removes a node from its drm_mm allocator. The node does not need to
515 * be cleared again before it can be re-inserted into this or any other drm_mm
516 * allocator. It is a bug to call this function on a un-allocated node.
517 */
518void drm_mm_remove_node(struct drm_mm_node *node)
519{
520 struct drm_mm *mm = node->mm;
521 struct drm_mm_node *prev_node;
522
523 if (WARN_ON(!node->allocated))
524 return;
525
526 BUG_ON(node->scanned_block || node->scanned_prev_free
527 || node->scanned_next_free);
528
529 prev_node =
530 list_entry(node->node_list.prev, struct drm_mm_node, node_list);
531
532 if (node->hole_follows) {
533 BUG_ON(__drm_mm_hole_node_start(node) ==
534 __drm_mm_hole_node_end(node));
535 list_del(&node->hole_stack);
536 } else
537 BUG_ON(__drm_mm_hole_node_start(node) !=
538 __drm_mm_hole_node_end(node));
539
540
541 if (!prev_node->hole_follows) {
542 prev_node->hole_follows = 1;
543 list_add(&prev_node->hole_stack, &mm->hole_stack);
544 } else
545 list_move(&prev_node->hole_stack, &mm->hole_stack);
546
547 drm_mm_interval_tree_remove(node, &mm->interval_tree);
548 list_del(&node->node_list);
549 node->allocated = 0;
550}
551EXPORT_SYMBOL(drm_mm_remove_node);
552
553static int check_free_hole(u64 start, u64 end, u64 size, unsigned alignment)
554{
555 if (end - start < size)
556 return 0;
557
558 if (alignment) {
559 u64 tmp = start;
560 unsigned rem;
561
562 rem = do_div(tmp, alignment);
563 if (rem)
564 start += alignment - rem;
565 }
566
567 return end >= start + size;
568}
569
570static struct drm_mm_node *drm_mm_search_free_generic(const struct drm_mm *mm,
571 u64 size,
572 unsigned alignment,
573 unsigned long color,
574 enum drm_mm_search_flags flags)
575{
576 struct drm_mm_node *entry;
577 struct drm_mm_node *best;
578 u64 adj_start;
579 u64 adj_end;
580 u64 best_size;
581
582 BUG_ON(mm->scanned_blocks);
583
584 best = NULL;
585 best_size = ~0UL;
586
587 __drm_mm_for_each_hole(entry, mm, adj_start, adj_end,
588 flags & DRM_MM_SEARCH_BELOW) {
589 u64 hole_size = adj_end - adj_start;
590
591 if (mm->color_adjust) {
592 mm->color_adjust(entry, color, &adj_start, &adj_end);
593 if (adj_end <= adj_start)
594 continue;
595 }
596
597 if (!check_free_hole(adj_start, adj_end, size, alignment))
598 continue;
599
600 if (!(flags & DRM_MM_SEARCH_BEST))
601 return entry;
602
603 if (hole_size < best_size) {
604 best = entry;
605 best_size = hole_size;
606 }
607 }
608
609 return best;
610}
611
612static struct drm_mm_node *drm_mm_search_free_in_range_generic(const struct drm_mm *mm,
613 u64 size,
614 unsigned alignment,
615 unsigned long color,
616 u64 start,
617 u64 end,
618 enum drm_mm_search_flags flags)
619{
620 struct drm_mm_node *entry;
621 struct drm_mm_node *best;
622 u64 adj_start;
623 u64 adj_end;
624 u64 best_size;
625
626 BUG_ON(mm->scanned_blocks);
627
628 best = NULL;
629 best_size = ~0UL;
630
631 __drm_mm_for_each_hole(entry, mm, adj_start, adj_end,
632 flags & DRM_MM_SEARCH_BELOW) {
633 u64 hole_size = adj_end - adj_start;
634
635 if (adj_start < start)
636 adj_start = start;
637 if (adj_end > end)
638 adj_end = end;
639
640 if (mm->color_adjust) {
641 mm->color_adjust(entry, color, &adj_start, &adj_end);
642 if (adj_end <= adj_start)
643 continue;
644 }
645
646 if (!check_free_hole(adj_start, adj_end, size, alignment))
647 continue;
648
649 if (!(flags & DRM_MM_SEARCH_BEST))
650 return entry;
651
652 if (hole_size < best_size) {
653 best = entry;
654 best_size = hole_size;
655 }
656 }
657
658 return best;
659}
660
661/**
662 * drm_mm_replace_node - move an allocation from @old to @new
663 * @old: drm_mm_node to remove from the allocator
664 * @new: drm_mm_node which should inherit @old's allocation
665 *
666 * This is useful for when drivers embed the drm_mm_node structure and hence
667 * can't move allocations by reassigning pointers. It's a combination of remove
668 * and insert with the guarantee that the allocation start will match.
669 */
670void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
671{
672 list_replace(&old->node_list, &new->node_list);
673 list_replace(&old->hole_stack, &new->hole_stack);
674 rb_replace_node(&old->rb, &new->rb, &old->mm->interval_tree);
675 new->hole_follows = old->hole_follows;
676 new->mm = old->mm;
677 new->start = old->start;
678 new->size = old->size;
679 new->color = old->color;
680 new->__subtree_last = old->__subtree_last;
681
682 old->allocated = 0;
683 new->allocated = 1;
684}
685EXPORT_SYMBOL(drm_mm_replace_node);
686
687/**
688 * DOC: lru scan roaster
689 *
690 * Very often GPUs need to have continuous allocations for a given object. When
691 * evicting objects to make space for a new one it is therefore not most
692 * efficient when we simply start to select all objects from the tail of an LRU
693 * until there's a suitable hole: Especially for big objects or nodes that
694 * otherwise have special allocation constraints there's a good chance we evict
695 * lots of (smaller) objects unecessarily.
696 *
697 * The DRM range allocator supports this use-case through the scanning
698 * interfaces. First a scan operation needs to be initialized with
699 * drm_mm_init_scan() or drm_mm_init_scan_with_range(). The the driver adds
700 * objects to the roaster (probably by walking an LRU list, but this can be
701 * freely implemented) until a suitable hole is found or there's no further
702 * evitable object.
703 *
704 * The the driver must walk through all objects again in exactly the reverse
705 * order to restore the allocator state. Note that while the allocator is used
706 * in the scan mode no other operation is allowed.
707 *
708 * Finally the driver evicts all objects selected in the scan. Adding and
709 * removing an object is O(1), and since freeing a node is also O(1) the overall
710 * complexity is O(scanned_objects). So like the free stack which needs to be
711 * walked before a scan operation even begins this is linear in the number of
712 * objects. It doesn't seem to hurt badly.
713 */
714
715/**
716 * drm_mm_init_scan - initialize lru scanning
717 * @mm: drm_mm to scan
718 * @size: size of the allocation
719 * @alignment: alignment of the allocation
720 * @color: opaque tag value to use for the allocation
721 *
722 * This simply sets up the scanning routines with the parameters for the desired
723 * hole. Note that there's no need to specify allocation flags, since they only
724 * change the place a node is allocated from within a suitable hole.
725 *
726 * Warning:
727 * As long as the scan list is non-empty, no other operations than
728 * adding/removing nodes to/from the scan list are allowed.
729 */
730void drm_mm_init_scan(struct drm_mm *mm,
731 u64 size,
732 unsigned alignment,
733 unsigned long color)
734{
735 mm->scan_color = color;
736 mm->scan_alignment = alignment;
737 mm->scan_size = size;
738 mm->scanned_blocks = 0;
739 mm->scan_hit_start = 0;
740 mm->scan_hit_end = 0;
741 mm->scan_check_range = 0;
742 mm->prev_scanned_node = NULL;
743}
744EXPORT_SYMBOL(drm_mm_init_scan);
745
746/**
747 * drm_mm_init_scan - initialize range-restricted lru scanning
748 * @mm: drm_mm to scan
749 * @size: size of the allocation
750 * @alignment: alignment of the allocation
751 * @color: opaque tag value to use for the allocation
752 * @start: start of the allowed range for the allocation
753 * @end: end of the allowed range for the allocation
754 *
755 * This simply sets up the scanning routines with the parameters for the desired
756 * hole. Note that there's no need to specify allocation flags, since they only
757 * change the place a node is allocated from within a suitable hole.
758 *
759 * Warning:
760 * As long as the scan list is non-empty, no other operations than
761 * adding/removing nodes to/from the scan list are allowed.
762 */
763void drm_mm_init_scan_with_range(struct drm_mm *mm,
764 u64 size,
765 unsigned alignment,
766 unsigned long color,
767 u64 start,
768 u64 end)
769{
770 mm->scan_color = color;
771 mm->scan_alignment = alignment;
772 mm->scan_size = size;
773 mm->scanned_blocks = 0;
774 mm->scan_hit_start = 0;
775 mm->scan_hit_end = 0;
776 mm->scan_start = start;
777 mm->scan_end = end;
778 mm->scan_check_range = 1;
779 mm->prev_scanned_node = NULL;
780}
781EXPORT_SYMBOL(drm_mm_init_scan_with_range);
782
783/**
784 * drm_mm_scan_add_block - add a node to the scan list
785 * @node: drm_mm_node to add
786 *
787 * Add a node to the scan list that might be freed to make space for the desired
788 * hole.
789 *
790 * Returns:
791 * True if a hole has been found, false otherwise.
792 */
793bool drm_mm_scan_add_block(struct drm_mm_node *node)
794{
795 struct drm_mm *mm = node->mm;
796 struct drm_mm_node *prev_node;
797 u64 hole_start, hole_end;
798 u64 adj_start, adj_end;
799
800 mm->scanned_blocks++;
801
802 BUG_ON(node->scanned_block);
803 node->scanned_block = 1;
804
805 prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
806 node_list);
807
808 node->scanned_preceeds_hole = prev_node->hole_follows;
809 prev_node->hole_follows = 1;
810 list_del(&node->node_list);
811 node->node_list.prev = &prev_node->node_list;
812 node->node_list.next = &mm->prev_scanned_node->node_list;
813 mm->prev_scanned_node = node;
814
815 adj_start = hole_start = drm_mm_hole_node_start(prev_node);
816 adj_end = hole_end = drm_mm_hole_node_end(prev_node);
817
818 if (mm->scan_check_range) {
819 if (adj_start < mm->scan_start)
820 adj_start = mm->scan_start;
821 if (adj_end > mm->scan_end)
822 adj_end = mm->scan_end;
823 }
824
825 if (mm->color_adjust)
826 mm->color_adjust(prev_node, mm->scan_color,
827 &adj_start, &adj_end);
828
829 if (check_free_hole(adj_start, adj_end,
830 mm->scan_size, mm->scan_alignment)) {
831 mm->scan_hit_start = hole_start;
832 mm->scan_hit_end = hole_end;
833 return true;
834 }
835
836 return false;
837}
838EXPORT_SYMBOL(drm_mm_scan_add_block);
839
840/**
841 * drm_mm_scan_remove_block - remove a node from the scan list
842 * @node: drm_mm_node to remove
843 *
844 * Nodes _must_ be removed in the exact same order from the scan list as they
845 * have been added, otherwise the internal state of the memory manager will be
846 * corrupted.
847 *
848 * When the scan list is empty, the selected memory nodes can be freed. An
849 * immediately following drm_mm_search_free with !DRM_MM_SEARCH_BEST will then
850 * return the just freed block (because its at the top of the free_stack list).
851 *
852 * Returns:
853 * True if this block should be evicted, false otherwise. Will always
854 * return false when no hole has been found.
855 */
856bool drm_mm_scan_remove_block(struct drm_mm_node *node)
857{
858 struct drm_mm *mm = node->mm;
859 struct drm_mm_node *prev_node;
860
861 mm->scanned_blocks--;
862
863 BUG_ON(!node->scanned_block);
864 node->scanned_block = 0;
865
866 prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
867 node_list);
868
869 prev_node->hole_follows = node->scanned_preceeds_hole;
870 list_add(&node->node_list, &prev_node->node_list);
871
872 return (drm_mm_hole_node_end(node) > mm->scan_hit_start &&
873 node->start < mm->scan_hit_end);
874}
875EXPORT_SYMBOL(drm_mm_scan_remove_block);
876
877/**
878 * drm_mm_clean - checks whether an allocator is clean
879 * @mm: drm_mm allocator to check
880 *
881 * Returns:
882 * True if the allocator is completely free, false if there's still a node
883 * allocated in it.
884 */
885bool drm_mm_clean(struct drm_mm * mm)
886{
887 struct list_head *head = &mm->head_node.node_list;
888
889 return (head->next->next == head);
890}
891EXPORT_SYMBOL(drm_mm_clean);
892
893/**
894 * drm_mm_init - initialize a drm-mm allocator
895 * @mm: the drm_mm structure to initialize
896 * @start: start of the range managed by @mm
897 * @size: end of the range managed by @mm
898 *
899 * Note that @mm must be cleared to 0 before calling this function.
900 */
901void drm_mm_init(struct drm_mm * mm, u64 start, u64 size)
902{
903 INIT_LIST_HEAD(&mm->hole_stack);
904 mm->scanned_blocks = 0;
905
906 /* Clever trick to avoid a special case in the free hole tracking. */
907 INIT_LIST_HEAD(&mm->head_node.node_list);
908 mm->head_node.allocated = 0;
909 mm->head_node.hole_follows = 1;
910 mm->head_node.scanned_block = 0;
911 mm->head_node.scanned_prev_free = 0;
912 mm->head_node.scanned_next_free = 0;
913 mm->head_node.mm = mm;
914 mm->head_node.start = start + size;
915 mm->head_node.size = start - mm->head_node.start;
916 list_add_tail(&mm->head_node.hole_stack, &mm->hole_stack);
917
918 mm->interval_tree = RB_ROOT;
919
920 mm->color_adjust = NULL;
921}
922EXPORT_SYMBOL(drm_mm_init);
923
924/**
925 * drm_mm_takedown - clean up a drm_mm allocator
926 * @mm: drm_mm allocator to clean up
927 *
928 * Note that it is a bug to call this function on an allocator which is not
929 * clean.
930 */
931void drm_mm_takedown(struct drm_mm *mm)
932{
933 if (WARN(!list_empty(&mm->head_node.node_list),
934 "Memory manager not clean during takedown.\n"))
935 show_leaks(mm);
936
937}
938EXPORT_SYMBOL(drm_mm_takedown);
939
940static u64 drm_mm_debug_hole(struct drm_mm_node *entry,
941 const char *prefix)
942{
943 u64 hole_start, hole_end, hole_size;
944
945 if (entry->hole_follows) {
946 hole_start = drm_mm_hole_node_start(entry);
947 hole_end = drm_mm_hole_node_end(entry);
948 hole_size = hole_end - hole_start;
949 pr_debug("%s %#llx-%#llx: %llu: free\n", prefix, hole_start,
950 hole_end, hole_size);
951 return hole_size;
952 }
953
954 return 0;
955}
956
957/**
958 * drm_mm_debug_table - dump allocator state to dmesg
959 * @mm: drm_mm allocator to dump
960 * @prefix: prefix to use for dumping to dmesg
961 */
962void drm_mm_debug_table(struct drm_mm *mm, const char *prefix)
963{
964 struct drm_mm_node *entry;
965 u64 total_used = 0, total_free = 0, total = 0;
966
967 total_free += drm_mm_debug_hole(&mm->head_node, prefix);
968
969 drm_mm_for_each_node(entry, mm) {
970 pr_debug("%s %#llx-%#llx: %llu: used\n", prefix, entry->start,
971 entry->start + entry->size, entry->size);
972 total_used += entry->size;
973 total_free += drm_mm_debug_hole(entry, prefix);
974 }
975 total = total_free + total_used;
976
977 pr_debug("%s total: %llu, used %llu free %llu\n", prefix, total,
978 total_used, total_free);
979}
980EXPORT_SYMBOL(drm_mm_debug_table);
981
982#if defined(CONFIG_DEBUG_FS)
983static u64 drm_mm_dump_hole(struct seq_file *m, struct drm_mm_node *entry)
984{
985 u64 hole_start, hole_end, hole_size;
986
987 if (entry->hole_follows) {
988 hole_start = drm_mm_hole_node_start(entry);
989 hole_end = drm_mm_hole_node_end(entry);
990 hole_size = hole_end - hole_start;
991 seq_printf(m, "%#018llx-%#018llx: %llu: free\n", hole_start,
992 hole_end, hole_size);
993 return hole_size;
994 }
995
996 return 0;
997}
998
999/**
1000 * drm_mm_dump_table - dump allocator state to a seq_file
1001 * @m: seq_file to dump to
1002 * @mm: drm_mm allocator to dump
1003 */
1004int drm_mm_dump_table(struct seq_file *m, struct drm_mm *mm)
1005{
1006 struct drm_mm_node *entry;
1007 u64 total_used = 0, total_free = 0, total = 0;
1008
1009 total_free += drm_mm_dump_hole(m, &mm->head_node);
1010
1011 drm_mm_for_each_node(entry, mm) {
1012 seq_printf(m, "%#018llx-%#018llx: %llu: used\n", entry->start,
1013 entry->start + entry->size, entry->size);
1014 total_used += entry->size;
1015 total_free += drm_mm_dump_hole(m, entry);
1016 }
1017 total = total_free + total_used;
1018
1019 seq_printf(m, "total: %llu, used %llu free %llu\n", total,
1020 total_used, total_free);
1021 return 0;
1022}
1023EXPORT_SYMBOL(drm_mm_dump_table);
1024#endif