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 <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
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 <drm/drmP.h>
46#include <drm/drm_mm.h>
47#include <linux/slab.h>
48#include <linux/seq_file.h>
49#include <linux/export.h>
50#include <linux/interval_tree_generic.h>
51
52/**
53 * DOC: Overview
54 *
55 * drm_mm provides a simple range allocator. The drivers are free to use the
56 * resource allocator from the linux core if it suits them, the upside of drm_mm
57 * is that it's in the DRM core. Which means that it's easier to extend for
58 * some of the crazier special purpose needs of gpus.
59 *
60 * The main data struct is &drm_mm, allocations are tracked in &drm_mm_node.
61 * Drivers are free to embed either of them into their own suitable
62 * datastructures. drm_mm itself will not do any memory allocations of its own,
63 * so if drivers choose not to embed nodes they need to still allocate them
64 * themselves.
65 *
66 * The range allocator also supports reservation of preallocated blocks. This is
67 * useful for taking over initial mode setting configurations from the firmware,
68 * where an object needs to be created which exactly matches the firmware's
69 * scanout target. As long as the range is still free it can be inserted anytime
70 * after the allocator is initialized, which helps with avoiding looped
71 * dependencies in the driver load sequence.
72 *
73 * drm_mm maintains a stack of most recently freed holes, which of all
74 * simplistic datastructures seems to be a fairly decent approach to clustering
75 * allocations and avoiding too much fragmentation. This means free space
76 * searches are O(num_holes). Given that all the fancy features drm_mm supports
77 * something better would be fairly complex and since gfx thrashing is a fairly
78 * steep cliff not a real concern. Removing a node again is O(1).
79 *
80 * drm_mm supports a few features: Alignment and range restrictions can be
81 * supplied. Furthermore every &drm_mm_node has a color value (which is just an
82 * opaque unsigned long) which in conjunction with a driver callback can be used
83 * to implement sophisticated placement restrictions. The i915 DRM driver uses
84 * this to implement guard pages between incompatible caching domains in the
85 * graphics TT.
86 *
87 * Two behaviors are supported for searching and allocating: bottom-up and
88 * top-down. The default is bottom-up. Top-down allocation can be used if the
89 * memory area has different restrictions, or just to reduce fragmentation.
90 *
91 * Finally iteration helpers to walk all nodes and all holes are provided as are
92 * some basic allocator dumpers for debugging.
93 *
94 * Note that this range allocator is not thread-safe, drivers need to protect
95 * modifications with their own locking. The idea behind this is that for a full
96 * memory manager additional data needs to be protected anyway, hence internal
97 * locking would be fully redundant.
98 */
99
100#ifdef CONFIG_DRM_DEBUG_MM
101#include <linux/stackdepot.h>
102
103#define STACKDEPTH 32
104#define BUFSZ 4096
105
106static noinline void save_stack(struct drm_mm_node *node)
107{
108 unsigned long entries[STACKDEPTH];
109 struct stack_trace trace = {
110 .entries = entries,
111 .max_entries = STACKDEPTH,
112 .skip = 1
113 };
114
115 save_stack_trace(&trace);
116 if (trace.nr_entries != 0 &&
117 trace.entries[trace.nr_entries-1] == ULONG_MAX)
118 trace.nr_entries--;
119
120 /* May be called under spinlock, so avoid sleeping */
121 node->stack = depot_save_stack(&trace, GFP_NOWAIT);
122}
123
124static void show_leaks(struct drm_mm *mm)
125{
126 struct drm_mm_node *node;
127 unsigned long entries[STACKDEPTH];
128 char *buf;
129
130 buf = kmalloc(BUFSZ, GFP_KERNEL);
131 if (!buf)
132 return;
133
134 list_for_each_entry(node, drm_mm_nodes(mm), node_list) {
135 struct stack_trace trace = {
136 .entries = entries,
137 .max_entries = STACKDEPTH
138 };
139
140 if (!node->stack) {
141 DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
142 node->start, node->size);
143 continue;
144 }
145
146 depot_fetch_stack(node->stack, &trace);
147 snprint_stack_trace(buf, BUFSZ, &trace, 0);
148 DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
149 node->start, node->size, buf);
150 }
151
152 kfree(buf);
153}
154
155#undef STACKDEPTH
156#undef BUFSZ
157#else
158static void save_stack(struct drm_mm_node *node) { }
159static void show_leaks(struct drm_mm *mm) { }
160#endif
161
162#define START(node) ((node)->start)
163#define LAST(node) ((node)->start + (node)->size - 1)
164
165INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
166 u64, __subtree_last,
167 START, LAST, static inline, drm_mm_interval_tree)
168
169struct drm_mm_node *
170__drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last)
171{
172 return drm_mm_interval_tree_iter_first((struct rb_root_cached *)&mm->interval_tree,
173 start, last) ?: (struct drm_mm_node *)&mm->head_node;
174}
175EXPORT_SYMBOL(__drm_mm_interval_first);
176
177static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
178 struct drm_mm_node *node)
179{
180 struct drm_mm *mm = hole_node->mm;
181 struct rb_node **link, *rb;
182 struct drm_mm_node *parent;
183 bool leftmost;
184
185 node->__subtree_last = LAST(node);
186
187 if (hole_node->allocated) {
188 rb = &hole_node->rb;
189 while (rb) {
190 parent = rb_entry(rb, struct drm_mm_node, rb);
191 if (parent->__subtree_last >= node->__subtree_last)
192 break;
193
194 parent->__subtree_last = node->__subtree_last;
195 rb = rb_parent(rb);
196 }
197
198 rb = &hole_node->rb;
199 link = &hole_node->rb.rb_right;
200 leftmost = false;
201 } else {
202 rb = NULL;
203 link = &mm->interval_tree.rb_root.rb_node;
204 leftmost = true;
205 }
206
207 while (*link) {
208 rb = *link;
209 parent = rb_entry(rb, struct drm_mm_node, rb);
210 if (parent->__subtree_last < node->__subtree_last)
211 parent->__subtree_last = node->__subtree_last;
212 if (node->start < parent->start) {
213 link = &parent->rb.rb_left;
214 } else {
215 link = &parent->rb.rb_right;
216 leftmost = false;
217 }
218 }
219
220 rb_link_node(&node->rb, rb, link);
221 rb_insert_augmented_cached(&node->rb, &mm->interval_tree, leftmost,
222 &drm_mm_interval_tree_augment);
223}
224
225#define RB_INSERT(root, member, expr) do { \
226 struct rb_node **link = &root.rb_node, *rb = NULL; \
227 u64 x = expr(node); \
228 while (*link) { \
229 rb = *link; \
230 if (x < expr(rb_entry(rb, struct drm_mm_node, member))) \
231 link = &rb->rb_left; \
232 else \
233 link = &rb->rb_right; \
234 } \
235 rb_link_node(&node->member, rb, link); \
236 rb_insert_color(&node->member, &root); \
237} while (0)
238
239#define HOLE_SIZE(NODE) ((NODE)->hole_size)
240#define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
241
242static void add_hole(struct drm_mm_node *node)
243{
244 struct drm_mm *mm = node->mm;
245
246 node->hole_size =
247 __drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
248 DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
249
250 RB_INSERT(mm->holes_size, rb_hole_size, HOLE_SIZE);
251 RB_INSERT(mm->holes_addr, rb_hole_addr, HOLE_ADDR);
252
253 list_add(&node->hole_stack, &mm->hole_stack);
254}
255
256static void rm_hole(struct drm_mm_node *node)
257{
258 DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
259
260 list_del(&node->hole_stack);
261 rb_erase(&node->rb_hole_size, &node->mm->holes_size);
262 rb_erase(&node->rb_hole_addr, &node->mm->holes_addr);
263 node->hole_size = 0;
264
265 DRM_MM_BUG_ON(drm_mm_hole_follows(node));
266}
267
268static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb)
269{
270 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size);
271}
272
273static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb)
274{
275 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr);
276}
277
278static inline u64 rb_hole_size(struct rb_node *rb)
279{
280 return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
281}
282
283static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size)
284{
285 struct rb_node *best = NULL;
286 struct rb_node **link = &mm->holes_size.rb_node;
287
288 while (*link) {
289 struct rb_node *rb = *link;
290
291 if (size <= rb_hole_size(rb)) {
292 link = &rb->rb_left;
293 best = rb;
294 } else {
295 link = &rb->rb_right;
296 }
297 }
298
299 return rb_hole_size_to_node(best);
300}
301
302static struct drm_mm_node *find_hole(struct drm_mm *mm, u64 addr)
303{
304 struct drm_mm_node *node = NULL;
305 struct rb_node **link = &mm->holes_addr.rb_node;
306
307 while (*link) {
308 u64 hole_start;
309
310 node = rb_hole_addr_to_node(*link);
311 hole_start = __drm_mm_hole_node_start(node);
312
313 if (addr < hole_start)
314 link = &node->rb_hole_addr.rb_left;
315 else if (addr > hole_start + node->hole_size)
316 link = &node->rb_hole_addr.rb_right;
317 else
318 break;
319 }
320
321 return node;
322}
323
324static struct drm_mm_node *
325first_hole(struct drm_mm *mm,
326 u64 start, u64 end, u64 size,
327 enum drm_mm_insert_mode mode)
328{
329 if (RB_EMPTY_ROOT(&mm->holes_size))
330 return NULL;
331
332 switch (mode) {
333 default:
334 case DRM_MM_INSERT_BEST:
335 return best_hole(mm, size);
336
337 case DRM_MM_INSERT_LOW:
338 return find_hole(mm, start);
339
340 case DRM_MM_INSERT_HIGH:
341 return find_hole(mm, end);
342
343 case DRM_MM_INSERT_EVICT:
344 return list_first_entry_or_null(&mm->hole_stack,
345 struct drm_mm_node,
346 hole_stack);
347 }
348}
349
350static struct drm_mm_node *
351next_hole(struct drm_mm *mm,
352 struct drm_mm_node *node,
353 enum drm_mm_insert_mode mode)
354{
355 switch (mode) {
356 default:
357 case DRM_MM_INSERT_BEST:
358 return rb_hole_size_to_node(rb_next(&node->rb_hole_size));
359
360 case DRM_MM_INSERT_LOW:
361 return rb_hole_addr_to_node(rb_next(&node->rb_hole_addr));
362
363 case DRM_MM_INSERT_HIGH:
364 return rb_hole_addr_to_node(rb_prev(&node->rb_hole_addr));
365
366 case DRM_MM_INSERT_EVICT:
367 node = list_next_entry(node, hole_stack);
368 return &node->hole_stack == &mm->hole_stack ? NULL : node;
369 }
370}
371
372/**
373 * drm_mm_reserve_node - insert an pre-initialized node
374 * @mm: drm_mm allocator to insert @node into
375 * @node: drm_mm_node to insert
376 *
377 * This functions inserts an already set-up &drm_mm_node into the allocator,
378 * meaning that start, size and color must be set by the caller. All other
379 * fields must be cleared to 0. This is useful to initialize the allocator with
380 * preallocated objects which must be set-up before the range allocator can be
381 * set-up, e.g. when taking over a firmware framebuffer.
382 *
383 * Returns:
384 * 0 on success, -ENOSPC if there's no hole where @node is.
385 */
386int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
387{
388 u64 end = node->start + node->size;
389 struct drm_mm_node *hole;
390 u64 hole_start, hole_end;
391 u64 adj_start, adj_end;
392
393 end = node->start + node->size;
394 if (unlikely(end <= node->start))
395 return -ENOSPC;
396
397 /* Find the relevant hole to add our node to */
398 hole = find_hole(mm, node->start);
399 if (!hole)
400 return -ENOSPC;
401
402 adj_start = hole_start = __drm_mm_hole_node_start(hole);
403 adj_end = hole_end = hole_start + hole->hole_size;
404
405 if (mm->color_adjust)
406 mm->color_adjust(hole, node->color, &adj_start, &adj_end);
407
408 if (adj_start > node->start || adj_end < end)
409 return -ENOSPC;
410
411 node->mm = mm;
412
413 list_add(&node->node_list, &hole->node_list);
414 drm_mm_interval_tree_add_node(hole, node);
415 node->allocated = true;
416 node->hole_size = 0;
417
418 rm_hole(hole);
419 if (node->start > hole_start)
420 add_hole(hole);
421 if (end < hole_end)
422 add_hole(node);
423
424 save_stack(node);
425 return 0;
426}
427EXPORT_SYMBOL(drm_mm_reserve_node);
428
429/**
430 * drm_mm_insert_node_in_range - ranged search for space and insert @node
431 * @mm: drm_mm to allocate from
432 * @node: preallocate node to insert
433 * @size: size of the allocation
434 * @alignment: alignment of the allocation
435 * @color: opaque tag value to use for this node
436 * @range_start: start of the allowed range for this node
437 * @range_end: end of the allowed range for this node
438 * @mode: fine-tune the allocation search and placement
439 *
440 * The preallocated @node must be cleared to 0.
441 *
442 * Returns:
443 * 0 on success, -ENOSPC if there's no suitable hole.
444 */
445int drm_mm_insert_node_in_range(struct drm_mm * const mm,
446 struct drm_mm_node * const node,
447 u64 size, u64 alignment,
448 unsigned long color,
449 u64 range_start, u64 range_end,
450 enum drm_mm_insert_mode mode)
451{
452 struct drm_mm_node *hole;
453 u64 remainder_mask;
454
455 DRM_MM_BUG_ON(range_start >= range_end);
456
457 if (unlikely(size == 0 || range_end - range_start < size))
458 return -ENOSPC;
459
460 if (alignment <= 1)
461 alignment = 0;
462
463 remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
464 for (hole = first_hole(mm, range_start, range_end, size, mode); hole;
465 hole = next_hole(mm, hole, mode)) {
466 u64 hole_start = __drm_mm_hole_node_start(hole);
467 u64 hole_end = hole_start + hole->hole_size;
468 u64 adj_start, adj_end;
469 u64 col_start, col_end;
470
471 if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end)
472 break;
473
474 if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start)
475 break;
476
477 col_start = hole_start;
478 col_end = hole_end;
479 if (mm->color_adjust)
480 mm->color_adjust(hole, color, &col_start, &col_end);
481
482 adj_start = max(col_start, range_start);
483 adj_end = min(col_end, range_end);
484
485 if (adj_end <= adj_start || adj_end - adj_start < size)
486 continue;
487
488 if (mode == DRM_MM_INSERT_HIGH)
489 adj_start = adj_end - size;
490
491 if (alignment) {
492 u64 rem;
493
494 if (likely(remainder_mask))
495 rem = adj_start & remainder_mask;
496 else
497 div64_u64_rem(adj_start, alignment, &rem);
498 if (rem) {
499 adj_start -= rem;
500 if (mode != DRM_MM_INSERT_HIGH)
501 adj_start += alignment;
502
503 if (adj_start < max(col_start, range_start) ||
504 min(col_end, range_end) - adj_start < size)
505 continue;
506
507 if (adj_end <= adj_start ||
508 adj_end - adj_start < size)
509 continue;
510 }
511 }
512
513 node->mm = mm;
514 node->size = size;
515 node->start = adj_start;
516 node->color = color;
517 node->hole_size = 0;
518
519 list_add(&node->node_list, &hole->node_list);
520 drm_mm_interval_tree_add_node(hole, node);
521 node->allocated = true;
522
523 rm_hole(hole);
524 if (adj_start > hole_start)
525 add_hole(hole);
526 if (adj_start + size < hole_end)
527 add_hole(node);
528
529 save_stack(node);
530 return 0;
531 }
532
533 return -ENOSPC;
534}
535EXPORT_SYMBOL(drm_mm_insert_node_in_range);
536
537/**
538 * drm_mm_remove_node - Remove a memory node from the allocator.
539 * @node: drm_mm_node to remove
540 *
541 * This just removes a node from its drm_mm allocator. The node does not need to
542 * be cleared again before it can be re-inserted into this or any other drm_mm
543 * allocator. It is a bug to call this function on a unallocated node.
544 */
545void drm_mm_remove_node(struct drm_mm_node *node)
546{
547 struct drm_mm *mm = node->mm;
548 struct drm_mm_node *prev_node;
549
550 DRM_MM_BUG_ON(!node->allocated);
551 DRM_MM_BUG_ON(node->scanned_block);
552
553 prev_node = list_prev_entry(node, node_list);
554
555 if (drm_mm_hole_follows(node))
556 rm_hole(node);
557
558 drm_mm_interval_tree_remove(node, &mm->interval_tree);
559 list_del(&node->node_list);
560 node->allocated = false;
561
562 if (drm_mm_hole_follows(prev_node))
563 rm_hole(prev_node);
564 add_hole(prev_node);
565}
566EXPORT_SYMBOL(drm_mm_remove_node);
567
568/**
569 * drm_mm_replace_node - move an allocation from @old to @new
570 * @old: drm_mm_node to remove from the allocator
571 * @new: drm_mm_node which should inherit @old's allocation
572 *
573 * This is useful for when drivers embed the drm_mm_node structure and hence
574 * can't move allocations by reassigning pointers. It's a combination of remove
575 * and insert with the guarantee that the allocation start will match.
576 */
577void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
578{
579 struct drm_mm *mm = old->mm;
580
581 DRM_MM_BUG_ON(!old->allocated);
582
583 *new = *old;
584
585 list_replace(&old->node_list, &new->node_list);
586 rb_replace_node_cached(&old->rb, &new->rb, &mm->interval_tree);
587
588 if (drm_mm_hole_follows(old)) {
589 list_replace(&old->hole_stack, &new->hole_stack);
590 rb_replace_node(&old->rb_hole_size,
591 &new->rb_hole_size,
592 &mm->holes_size);
593 rb_replace_node(&old->rb_hole_addr,
594 &new->rb_hole_addr,
595 &mm->holes_addr);
596 }
597
598 old->allocated = false;
599 new->allocated = true;
600}
601EXPORT_SYMBOL(drm_mm_replace_node);
602
603/**
604 * DOC: lru scan roster
605 *
606 * Very often GPUs need to have continuous allocations for a given object. When
607 * evicting objects to make space for a new one it is therefore not most
608 * efficient when we simply start to select all objects from the tail of an LRU
609 * until there's a suitable hole: Especially for big objects or nodes that
610 * otherwise have special allocation constraints there's a good chance we evict
611 * lots of (smaller) objects unnecessarily.
612 *
613 * The DRM range allocator supports this use-case through the scanning
614 * interfaces. First a scan operation needs to be initialized with
615 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
616 * objects to the roster, probably by walking an LRU list, but this can be
617 * freely implemented. Eviction candiates are added using
618 * drm_mm_scan_add_block() until a suitable hole is found or there are no
619 * further evictable objects. Eviction roster metadata is tracked in &struct
620 * drm_mm_scan.
621 *
622 * The driver must walk through all objects again in exactly the reverse
623 * order to restore the allocator state. Note that while the allocator is used
624 * in the scan mode no other operation is allowed.
625 *
626 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
627 * reported true) in the scan, and any overlapping nodes after color adjustment
628 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
629 * since freeing a node is also O(1) the overall complexity is
630 * O(scanned_objects). So like the free stack which needs to be walked before a
631 * scan operation even begins this is linear in the number of objects. It
632 * doesn't seem to hurt too badly.
633 */
634
635/**
636 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
637 * @scan: scan state
638 * @mm: drm_mm to scan
639 * @size: size of the allocation
640 * @alignment: alignment of the allocation
641 * @color: opaque tag value to use for the allocation
642 * @start: start of the allowed range for the allocation
643 * @end: end of the allowed range for the allocation
644 * @mode: fine-tune the allocation search and placement
645 *
646 * This simply sets up the scanning routines with the parameters for the desired
647 * hole.
648 *
649 * Warning:
650 * As long as the scan list is non-empty, no other operations than
651 * adding/removing nodes to/from the scan list are allowed.
652 */
653void drm_mm_scan_init_with_range(struct drm_mm_scan *scan,
654 struct drm_mm *mm,
655 u64 size,
656 u64 alignment,
657 unsigned long color,
658 u64 start,
659 u64 end,
660 enum drm_mm_insert_mode mode)
661{
662 DRM_MM_BUG_ON(start >= end);
663 DRM_MM_BUG_ON(!size || size > end - start);
664 DRM_MM_BUG_ON(mm->scan_active);
665
666 scan->mm = mm;
667
668 if (alignment <= 1)
669 alignment = 0;
670
671 scan->color = color;
672 scan->alignment = alignment;
673 scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
674 scan->size = size;
675 scan->mode = mode;
676
677 DRM_MM_BUG_ON(end <= start);
678 scan->range_start = start;
679 scan->range_end = end;
680
681 scan->hit_start = U64_MAX;
682 scan->hit_end = 0;
683}
684EXPORT_SYMBOL(drm_mm_scan_init_with_range);
685
686/**
687 * drm_mm_scan_add_block - add a node to the scan list
688 * @scan: the active drm_mm scanner
689 * @node: drm_mm_node to add
690 *
691 * Add a node to the scan list that might be freed to make space for the desired
692 * hole.
693 *
694 * Returns:
695 * True if a hole has been found, false otherwise.
696 */
697bool drm_mm_scan_add_block(struct drm_mm_scan *scan,
698 struct drm_mm_node *node)
699{
700 struct drm_mm *mm = scan->mm;
701 struct drm_mm_node *hole;
702 u64 hole_start, hole_end;
703 u64 col_start, col_end;
704 u64 adj_start, adj_end;
705
706 DRM_MM_BUG_ON(node->mm != mm);
707 DRM_MM_BUG_ON(!node->allocated);
708 DRM_MM_BUG_ON(node->scanned_block);
709 node->scanned_block = true;
710 mm->scan_active++;
711
712 /* Remove this block from the node_list so that we enlarge the hole
713 * (distance between the end of our previous node and the start of
714 * or next), without poisoning the link so that we can restore it
715 * later in drm_mm_scan_remove_block().
716 */
717 hole = list_prev_entry(node, node_list);
718 DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node);
719 __list_del_entry(&node->node_list);
720
721 hole_start = __drm_mm_hole_node_start(hole);
722 hole_end = __drm_mm_hole_node_end(hole);
723
724 col_start = hole_start;
725 col_end = hole_end;
726 if (mm->color_adjust)
727 mm->color_adjust(hole, scan->color, &col_start, &col_end);
728
729 adj_start = max(col_start, scan->range_start);
730 adj_end = min(col_end, scan->range_end);
731 if (adj_end <= adj_start || adj_end - adj_start < scan->size)
732 return false;
733
734 if (scan->mode == DRM_MM_INSERT_HIGH)
735 adj_start = adj_end - scan->size;
736
737 if (scan->alignment) {
738 u64 rem;
739
740 if (likely(scan->remainder_mask))
741 rem = adj_start & scan->remainder_mask;
742 else
743 div64_u64_rem(adj_start, scan->alignment, &rem);
744 if (rem) {
745 adj_start -= rem;
746 if (scan->mode != DRM_MM_INSERT_HIGH)
747 adj_start += scan->alignment;
748 if (adj_start < max(col_start, scan->range_start) ||
749 min(col_end, scan->range_end) - adj_start < scan->size)
750 return false;
751
752 if (adj_end <= adj_start ||
753 adj_end - adj_start < scan->size)
754 return false;
755 }
756 }
757
758 scan->hit_start = adj_start;
759 scan->hit_end = adj_start + scan->size;
760
761 DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end);
762 DRM_MM_BUG_ON(scan->hit_start < hole_start);
763 DRM_MM_BUG_ON(scan->hit_end > hole_end);
764
765 return true;
766}
767EXPORT_SYMBOL(drm_mm_scan_add_block);
768
769/**
770 * drm_mm_scan_remove_block - remove a node from the scan list
771 * @scan: the active drm_mm scanner
772 * @node: drm_mm_node to remove
773 *
774 * Nodes **must** be removed in exactly the reverse order from the scan list as
775 * they have been added (e.g. using list_add() as they are added and then
776 * list_for_each() over that eviction list to remove), otherwise the internal
777 * state of the memory manager will be corrupted.
778 *
779 * When the scan list is empty, the selected memory nodes can be freed. An
780 * immediately following drm_mm_insert_node_in_range_generic() or one of the
781 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
782 * the just freed block (because its at the top of the free_stack list).
783 *
784 * Returns:
785 * True if this block should be evicted, false otherwise. Will always
786 * return false when no hole has been found.
787 */
788bool drm_mm_scan_remove_block(struct drm_mm_scan *scan,
789 struct drm_mm_node *node)
790{
791 struct drm_mm_node *prev_node;
792
793 DRM_MM_BUG_ON(node->mm != scan->mm);
794 DRM_MM_BUG_ON(!node->scanned_block);
795 node->scanned_block = false;
796
797 DRM_MM_BUG_ON(!node->mm->scan_active);
798 node->mm->scan_active--;
799
800 /* During drm_mm_scan_add_block() we decoupled this node leaving
801 * its pointers intact. Now that the caller is walking back along
802 * the eviction list we can restore this block into its rightful
803 * place on the full node_list. To confirm that the caller is walking
804 * backwards correctly we check that prev_node->next == node->next,
805 * i.e. both believe the same node should be on the other side of the
806 * hole.
807 */
808 prev_node = list_prev_entry(node, node_list);
809 DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) !=
810 list_next_entry(node, node_list));
811 list_add(&node->node_list, &prev_node->node_list);
812
813 return (node->start + node->size > scan->hit_start &&
814 node->start < scan->hit_end);
815}
816EXPORT_SYMBOL(drm_mm_scan_remove_block);
817
818/**
819 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
820 * @scan: drm_mm scan with target hole
821 *
822 * After completing an eviction scan and removing the selected nodes, we may
823 * need to remove a few more nodes from either side of the target hole if
824 * mm.color_adjust is being used.
825 *
826 * Returns:
827 * A node to evict, or NULL if there are no overlapping nodes.
828 */
829struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan)
830{
831 struct drm_mm *mm = scan->mm;
832 struct drm_mm_node *hole;
833 u64 hole_start, hole_end;
834
835 DRM_MM_BUG_ON(list_empty(&mm->hole_stack));
836
837 if (!mm->color_adjust)
838 return NULL;
839
840 /*
841 * The hole found during scanning should ideally be the first element
842 * in the hole_stack list, but due to side-effects in the driver it
843 * may not be.
844 */
845 list_for_each_entry(hole, &mm->hole_stack, hole_stack) {
846 hole_start = __drm_mm_hole_node_start(hole);
847 hole_end = hole_start + hole->hole_size;
848
849 if (hole_start <= scan->hit_start &&
850 hole_end >= scan->hit_end)
851 break;
852 }
853
854 /* We should only be called after we found the hole previously */
855 DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack);
856 if (unlikely(&hole->hole_stack == &mm->hole_stack))
857 return NULL;
858
859 DRM_MM_BUG_ON(hole_start > scan->hit_start);
860 DRM_MM_BUG_ON(hole_end < scan->hit_end);
861
862 mm->color_adjust(hole, scan->color, &hole_start, &hole_end);
863 if (hole_start > scan->hit_start)
864 return hole;
865 if (hole_end < scan->hit_end)
866 return list_next_entry(hole, node_list);
867
868 return NULL;
869}
870EXPORT_SYMBOL(drm_mm_scan_color_evict);
871
872/**
873 * drm_mm_init - initialize a drm-mm allocator
874 * @mm: the drm_mm structure to initialize
875 * @start: start of the range managed by @mm
876 * @size: end of the range managed by @mm
877 *
878 * Note that @mm must be cleared to 0 before calling this function.
879 */
880void drm_mm_init(struct drm_mm *mm, u64 start, u64 size)
881{
882 DRM_MM_BUG_ON(start + size <= start);
883
884 mm->color_adjust = NULL;
885
886 INIT_LIST_HEAD(&mm->hole_stack);
887 mm->interval_tree = RB_ROOT_CACHED;
888 mm->holes_size = RB_ROOT;
889 mm->holes_addr = RB_ROOT;
890
891 /* Clever trick to avoid a special case in the free hole tracking. */
892 INIT_LIST_HEAD(&mm->head_node.node_list);
893 mm->head_node.allocated = false;
894 mm->head_node.mm = mm;
895 mm->head_node.start = start + size;
896 mm->head_node.size = -size;
897 add_hole(&mm->head_node);
898
899 mm->scan_active = 0;
900}
901EXPORT_SYMBOL(drm_mm_init);
902
903/**
904 * drm_mm_takedown - clean up a drm_mm allocator
905 * @mm: drm_mm allocator to clean up
906 *
907 * Note that it is a bug to call this function on an allocator which is not
908 * clean.
909 */
910void drm_mm_takedown(struct drm_mm *mm)
911{
912 if (WARN(!drm_mm_clean(mm),
913 "Memory manager not clean during takedown.\n"))
914 show_leaks(mm);
915}
916EXPORT_SYMBOL(drm_mm_takedown);
917
918static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry)
919{
920 u64 start, size;
921
922 size = entry->hole_size;
923 if (size) {
924 start = drm_mm_hole_node_start(entry);
925 drm_printf(p, "%#018llx-%#018llx: %llu: free\n",
926 start, start + size, size);
927 }
928
929 return size;
930}
931/**
932 * drm_mm_print - print allocator state
933 * @mm: drm_mm allocator to print
934 * @p: DRM printer to use
935 */
936void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p)
937{
938 const struct drm_mm_node *entry;
939 u64 total_used = 0, total_free = 0, total = 0;
940
941 total_free += drm_mm_dump_hole(p, &mm->head_node);
942
943 drm_mm_for_each_node(entry, mm) {
944 drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start,
945 entry->start + entry->size, entry->size);
946 total_used += entry->size;
947 total_free += drm_mm_dump_hole(p, entry);
948 }
949 total = total_free + total_used;
950
951 drm_printf(p, "total: %llu, used %llu free %llu\n", total,
952 total_used, total_free);
953}
954EXPORT_SYMBOL(drm_mm_print);