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 <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);
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
50/**
51 * DOC: Overview
52 *
53 * drm_mm provides a simple range allocator. The drivers are free to use the
54 * resource allocator from the linux core if it suits them, the upside of drm_mm
55 * is that it's in the DRM core. Which means that it's easier to extend for
56 * some of the crazier special purpose needs of gpus.
57 *
58 * The main data struct is &drm_mm, allocations are tracked in &drm_mm_node.
59 * Drivers are free to embed either of them into their own suitable
60 * datastructures. drm_mm itself will not do any allocations of its own, so if
61 * drivers choose not to embed nodes they need to still allocate them
62 * themselves.
63 *
64 * The range allocator also supports reservation of preallocated blocks. This is
65 * useful for taking over initial mode setting configurations from the firmware,
66 * where an object needs to be created which exactly matches the firmware's
67 * scanout target. As long as the range is still free it can be inserted anytime
68 * after the allocator is initialized, which helps with avoiding looped
69 * depencies in the driver load sequence.
70 *
71 * drm_mm maintains a stack of most recently freed holes, which of all
72 * simplistic datastructures seems to be a fairly decent approach to clustering
73 * allocations and avoiding too much fragmentation. This means free space
74 * searches are O(num_holes). Given that all the fancy features drm_mm supports
75 * something better would be fairly complex and since gfx thrashing is a fairly
76 * steep cliff not a real concern. Removing a node again is O(1).
77 *
78 * drm_mm supports a few features: Alignment and range restrictions can be
79 * supplied. Further more every &drm_mm_node has a color value (which is just an
80 * opaqua unsigned long) which in conjunction with a driver callback can be used
81 * to implement sophisticated placement restrictions. The i915 DRM driver uses
82 * this to implement guard pages between incompatible caching domains in the
83 * graphics TT.
84 *
85 * Two behaviors are supported for searching and allocating: bottom-up and top-down.
86 * The default is bottom-up. Top-down allocation can be used if the memory area
87 * has different restrictions, or just to reduce fragmentation.
88 *
89 * Finally iteration helpers to walk all nodes and all holes are provided as are
90 * some basic allocator dumpers for debugging.
91 */
92
93static struct drm_mm_node *drm_mm_search_free_generic(const struct drm_mm *mm,
94 unsigned long size,
95 unsigned alignment,
96 unsigned long color,
97 enum drm_mm_search_flags flags);
98static struct drm_mm_node *drm_mm_search_free_in_range_generic(const struct drm_mm *mm,
99 unsigned long size,
100 unsigned alignment,
101 unsigned long color,
102 unsigned long start,
103 unsigned long end,
104 enum drm_mm_search_flags flags);
105
106static void drm_mm_insert_helper(struct drm_mm_node *hole_node,
107 struct drm_mm_node *node,
108 unsigned long size, unsigned alignment,
109 unsigned long color,
110 enum drm_mm_allocator_flags flags)
111{
112 struct drm_mm *mm = hole_node->mm;
113 unsigned long hole_start = drm_mm_hole_node_start(hole_node);
114 unsigned long hole_end = drm_mm_hole_node_end(hole_node);
115 unsigned long adj_start = hole_start;
116 unsigned long adj_end = hole_end;
117
118 BUG_ON(node->allocated);
119
120 if (mm->color_adjust)
121 mm->color_adjust(hole_node, color, &adj_start, &adj_end);
122
123 if (flags & DRM_MM_CREATE_TOP)
124 adj_start = adj_end - size;
125
126 if (alignment) {
127 unsigned tmp = adj_start % alignment;
128 if (tmp) {
129 if (flags & DRM_MM_CREATE_TOP)
130 adj_start -= tmp;
131 else
132 adj_start += alignment - tmp;
133 }
134 }
135
136 BUG_ON(adj_start < hole_start);
137 BUG_ON(adj_end > hole_end);
138
139 if (adj_start == hole_start) {
140 hole_node->hole_follows = 0;
141 list_del(&hole_node->hole_stack);
142 }
143
144 node->start = adj_start;
145 node->size = size;
146 node->mm = mm;
147 node->color = color;
148 node->allocated = 1;
149
150 INIT_LIST_HEAD(&node->hole_stack);
151 list_add(&node->node_list, &hole_node->node_list);
152
153 BUG_ON(node->start + node->size > adj_end);
154
155 node->hole_follows = 0;
156 if (__drm_mm_hole_node_start(node) < hole_end) {
157 list_add(&node->hole_stack, &mm->hole_stack);
158 node->hole_follows = 1;
159 }
160}
161
162/**
163 * drm_mm_reserve_node - insert an pre-initialized node
164 * @mm: drm_mm allocator to insert @node into
165 * @node: drm_mm_node to insert
166 *
167 * This functions inserts an already set-up drm_mm_node into the allocator,
168 * meaning that start, size and color must be set by the caller. This is useful
169 * to initialize the allocator with preallocated objects which must be set-up
170 * before the range allocator can be set-up, e.g. when taking over a firmware
171 * framebuffer.
172 *
173 * Returns:
174 * 0 on success, -ENOSPC if there's no hole where @node is.
175 */
176int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
177{
178 struct drm_mm_node *hole;
179 unsigned long end = node->start + node->size;
180 unsigned long hole_start;
181 unsigned long hole_end;
182
183 BUG_ON(node == NULL);
184
185 /* Find the relevant hole to add our node to */
186 drm_mm_for_each_hole(hole, mm, hole_start, hole_end) {
187 if (hole_start > node->start || hole_end < end)
188 continue;
189
190 node->mm = mm;
191 node->allocated = 1;
192
193 INIT_LIST_HEAD(&node->hole_stack);
194 list_add(&node->node_list, &hole->node_list);
195
196 if (node->start == hole_start) {
197 hole->hole_follows = 0;
198 list_del_init(&hole->hole_stack);
199 }
200
201 node->hole_follows = 0;
202 if (end != hole_end) {
203 list_add(&node->hole_stack, &mm->hole_stack);
204 node->hole_follows = 1;
205 }
206
207 return 0;
208 }
209
210 return -ENOSPC;
211}
212EXPORT_SYMBOL(drm_mm_reserve_node);
213
214/**
215 * drm_mm_insert_node_generic - search for space and insert @node
216 * @mm: drm_mm to allocate from
217 * @node: preallocate node to insert
218 * @size: size of the allocation
219 * @alignment: alignment of the allocation
220 * @color: opaque tag value to use for this node
221 * @sflags: flags to fine-tune the allocation search
222 * @aflags: flags to fine-tune the allocation behavior
223 *
224 * The preallocated node must be cleared to 0.
225 *
226 * Returns:
227 * 0 on success, -ENOSPC if there's no suitable hole.
228 */
229int drm_mm_insert_node_generic(struct drm_mm *mm, struct drm_mm_node *node,
230 unsigned long size, unsigned alignment,
231 unsigned long color,
232 enum drm_mm_search_flags sflags,
233 enum drm_mm_allocator_flags aflags)
234{
235 struct drm_mm_node *hole_node;
236
237 hole_node = drm_mm_search_free_generic(mm, size, alignment,
238 color, sflags);
239 if (!hole_node)
240 return -ENOSPC;
241
242 drm_mm_insert_helper(hole_node, node, size, alignment, color, aflags);
243 return 0;
244}
245EXPORT_SYMBOL(drm_mm_insert_node_generic);
246
247static void drm_mm_insert_helper_range(struct drm_mm_node *hole_node,
248 struct drm_mm_node *node,
249 unsigned long size, unsigned alignment,
250 unsigned long color,
251 unsigned long start, unsigned long end,
252 enum drm_mm_allocator_flags flags)
253{
254 struct drm_mm *mm = hole_node->mm;
255 unsigned long hole_start = drm_mm_hole_node_start(hole_node);
256 unsigned long hole_end = drm_mm_hole_node_end(hole_node);
257 unsigned long adj_start = hole_start;
258 unsigned long adj_end = hole_end;
259
260 BUG_ON(!hole_node->hole_follows || node->allocated);
261
262 if (adj_start < start)
263 adj_start = start;
264 if (adj_end > end)
265 adj_end = end;
266
267 if (flags & DRM_MM_CREATE_TOP)
268 adj_start = adj_end - size;
269
270 if (mm->color_adjust)
271 mm->color_adjust(hole_node, color, &adj_start, &adj_end);
272
273 if (alignment) {
274 unsigned tmp = adj_start % alignment;
275 if (tmp) {
276 if (flags & DRM_MM_CREATE_TOP)
277 adj_start -= tmp;
278 else
279 adj_start += alignment - tmp;
280 }
281 }
282
283 if (adj_start == hole_start) {
284 hole_node->hole_follows = 0;
285 list_del(&hole_node->hole_stack);
286 }
287
288 node->start = adj_start;
289 node->size = size;
290 node->mm = mm;
291 node->color = color;
292 node->allocated = 1;
293
294 INIT_LIST_HEAD(&node->hole_stack);
295 list_add(&node->node_list, &hole_node->node_list);
296
297 BUG_ON(node->start < start);
298 BUG_ON(node->start < adj_start);
299 BUG_ON(node->start + node->size > adj_end);
300 BUG_ON(node->start + node->size > end);
301
302 node->hole_follows = 0;
303 if (__drm_mm_hole_node_start(node) < hole_end) {
304 list_add(&node->hole_stack, &mm->hole_stack);
305 node->hole_follows = 1;
306 }
307}
308
309/**
310 * drm_mm_insert_node_in_range_generic - ranged search for space and insert @node
311 * @mm: drm_mm to allocate from
312 * @node: preallocate node to insert
313 * @size: size of the allocation
314 * @alignment: alignment of the allocation
315 * @color: opaque tag value to use for this node
316 * @start: start of the allowed range for this node
317 * @end: end of the allowed range for this node
318 * @sflags: flags to fine-tune the allocation search
319 * @aflags: flags to fine-tune the allocation behavior
320 *
321 * The preallocated node must be cleared to 0.
322 *
323 * Returns:
324 * 0 on success, -ENOSPC if there's no suitable hole.
325 */
326int drm_mm_insert_node_in_range_generic(struct drm_mm *mm, struct drm_mm_node *node,
327 unsigned long size, unsigned alignment,
328 unsigned long color,
329 unsigned long start, unsigned long end,
330 enum drm_mm_search_flags sflags,
331 enum drm_mm_allocator_flags aflags)
332{
333 struct drm_mm_node *hole_node;
334
335 hole_node = drm_mm_search_free_in_range_generic(mm,
336 size, alignment, color,
337 start, end, sflags);
338 if (!hole_node)
339 return -ENOSPC;
340
341 drm_mm_insert_helper_range(hole_node, node,
342 size, alignment, color,
343 start, end, aflags);
344 return 0;
345}
346EXPORT_SYMBOL(drm_mm_insert_node_in_range_generic);
347
348/**
349 * drm_mm_remove_node - Remove a memory node from the allocator.
350 * @node: drm_mm_node to remove
351 *
352 * This just removes a node from its drm_mm allocator. The node does not need to
353 * be cleared again before it can be re-inserted into this or any other drm_mm
354 * allocator. It is a bug to call this function on a un-allocated node.
355 */
356void drm_mm_remove_node(struct drm_mm_node *node)
357{
358 struct drm_mm *mm = node->mm;
359 struct drm_mm_node *prev_node;
360
361 if (WARN_ON(!node->allocated))
362 return;
363
364 BUG_ON(node->scanned_block || node->scanned_prev_free
365 || node->scanned_next_free);
366
367 prev_node =
368 list_entry(node->node_list.prev, struct drm_mm_node, node_list);
369
370 if (node->hole_follows) {
371 BUG_ON(__drm_mm_hole_node_start(node) ==
372 __drm_mm_hole_node_end(node));
373 list_del(&node->hole_stack);
374 } else
375 BUG_ON(__drm_mm_hole_node_start(node) !=
376 __drm_mm_hole_node_end(node));
377
378
379 if (!prev_node->hole_follows) {
380 prev_node->hole_follows = 1;
381 list_add(&prev_node->hole_stack, &mm->hole_stack);
382 } else
383 list_move(&prev_node->hole_stack, &mm->hole_stack);
384
385 list_del(&node->node_list);
386 node->allocated = 0;
387}
388EXPORT_SYMBOL(drm_mm_remove_node);
389
390static int check_free_hole(unsigned long start, unsigned long end,
391 unsigned long size, unsigned alignment)
392{
393 if (end - start < size)
394 return 0;
395
396 if (alignment) {
397 unsigned tmp = start % alignment;
398 if (tmp)
399 start += alignment - tmp;
400 }
401
402 return end >= start + size;
403}
404
405static struct drm_mm_node *drm_mm_search_free_generic(const struct drm_mm *mm,
406 unsigned long size,
407 unsigned alignment,
408 unsigned long color,
409 enum drm_mm_search_flags flags)
410{
411 struct drm_mm_node *entry;
412 struct drm_mm_node *best;
413 unsigned long adj_start;
414 unsigned long adj_end;
415 unsigned long best_size;
416
417 BUG_ON(mm->scanned_blocks);
418
419 best = NULL;
420 best_size = ~0UL;
421
422 __drm_mm_for_each_hole(entry, mm, adj_start, adj_end,
423 flags & DRM_MM_SEARCH_BELOW) {
424 unsigned long hole_size = adj_end - adj_start;
425
426 if (mm->color_adjust) {
427 mm->color_adjust(entry, color, &adj_start, &adj_end);
428 if (adj_end <= adj_start)
429 continue;
430 }
431
432 if (!check_free_hole(adj_start, adj_end, size, alignment))
433 continue;
434
435 if (!(flags & DRM_MM_SEARCH_BEST))
436 return entry;
437
438 if (hole_size < best_size) {
439 best = entry;
440 best_size = hole_size;
441 }
442 }
443
444 return best;
445}
446
447static struct drm_mm_node *drm_mm_search_free_in_range_generic(const struct drm_mm *mm,
448 unsigned long size,
449 unsigned alignment,
450 unsigned long color,
451 unsigned long start,
452 unsigned long end,
453 enum drm_mm_search_flags flags)
454{
455 struct drm_mm_node *entry;
456 struct drm_mm_node *best;
457 unsigned long adj_start;
458 unsigned long adj_end;
459 unsigned long best_size;
460
461 BUG_ON(mm->scanned_blocks);
462
463 best = NULL;
464 best_size = ~0UL;
465
466 __drm_mm_for_each_hole(entry, mm, adj_start, adj_end,
467 flags & DRM_MM_SEARCH_BELOW) {
468 unsigned long hole_size = adj_end - adj_start;
469
470 if (adj_start < start)
471 adj_start = start;
472 if (adj_end > end)
473 adj_end = end;
474
475 if (mm->color_adjust) {
476 mm->color_adjust(entry, color, &adj_start, &adj_end);
477 if (adj_end <= adj_start)
478 continue;
479 }
480
481 if (!check_free_hole(adj_start, adj_end, size, alignment))
482 continue;
483
484 if (!(flags & DRM_MM_SEARCH_BEST))
485 return entry;
486
487 if (hole_size < best_size) {
488 best = entry;
489 best_size = hole_size;
490 }
491 }
492
493 return best;
494}
495
496/**
497 * drm_mm_replace_node - move an allocation from @old to @new
498 * @old: drm_mm_node to remove from the allocator
499 * @new: drm_mm_node which should inherit @old's allocation
500 *
501 * This is useful for when drivers embed the drm_mm_node structure and hence
502 * can't move allocations by reassigning pointers. It's a combination of remove
503 * and insert with the guarantee that the allocation start will match.
504 */
505void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new)
506{
507 list_replace(&old->node_list, &new->node_list);
508 list_replace(&old->hole_stack, &new->hole_stack);
509 new->hole_follows = old->hole_follows;
510 new->mm = old->mm;
511 new->start = old->start;
512 new->size = old->size;
513 new->color = old->color;
514
515 old->allocated = 0;
516 new->allocated = 1;
517}
518EXPORT_SYMBOL(drm_mm_replace_node);
519
520/**
521 * DOC: lru scan roaster
522 *
523 * Very often GPUs need to have continuous allocations for a given object. When
524 * evicting objects to make space for a new one it is therefore not most
525 * efficient when we simply start to select all objects from the tail of an LRU
526 * until there's a suitable hole: Especially for big objects or nodes that
527 * otherwise have special allocation constraints there's a good chance we evict
528 * lots of (smaller) objects unecessarily.
529 *
530 * The DRM range allocator supports this use-case through the scanning
531 * interfaces. First a scan operation needs to be initialized with
532 * drm_mm_init_scan() or drm_mm_init_scan_with_range(). The the driver adds
533 * objects to the roaster (probably by walking an LRU list, but this can be
534 * freely implemented) until a suitable hole is found or there's no further
535 * evitable object.
536 *
537 * The the driver must walk through all objects again in exactly the reverse
538 * order to restore the allocator state. Note that while the allocator is used
539 * in the scan mode no other operation is allowed.
540 *
541 * Finally the driver evicts all objects selected in the scan. Adding and
542 * removing an object is O(1), and since freeing a node is also O(1) the overall
543 * complexity is O(scanned_objects). So like the free stack which needs to be
544 * walked before a scan operation even begins this is linear in the number of
545 * objects. It doesn't seem to hurt badly.
546 */
547
548/**
549 * drm_mm_init_scan - initialize lru scanning
550 * @mm: drm_mm to scan
551 * @size: size of the allocation
552 * @alignment: alignment of the allocation
553 * @color: opaque tag value to use for the allocation
554 *
555 * This simply sets up the scanning routines with the parameters for the desired
556 * hole. Note that there's no need to specify allocation flags, since they only
557 * change the place a node is allocated from within a suitable hole.
558 *
559 * Warning:
560 * As long as the scan list is non-empty, no other operations than
561 * adding/removing nodes to/from the scan list are allowed.
562 */
563void drm_mm_init_scan(struct drm_mm *mm,
564 unsigned long size,
565 unsigned alignment,
566 unsigned long color)
567{
568 mm->scan_color = color;
569 mm->scan_alignment = alignment;
570 mm->scan_size = size;
571 mm->scanned_blocks = 0;
572 mm->scan_hit_start = 0;
573 mm->scan_hit_end = 0;
574 mm->scan_check_range = 0;
575 mm->prev_scanned_node = NULL;
576}
577EXPORT_SYMBOL(drm_mm_init_scan);
578
579/**
580 * drm_mm_init_scan - initialize range-restricted lru scanning
581 * @mm: drm_mm to scan
582 * @size: size of the allocation
583 * @alignment: alignment of the allocation
584 * @color: opaque tag value to use for the allocation
585 * @start: start of the allowed range for the allocation
586 * @end: end of the allowed range for the allocation
587 *
588 * This simply sets up the scanning routines with the parameters for the desired
589 * hole. Note that there's no need to specify allocation flags, since they only
590 * change the place a node is allocated from within a suitable hole.
591 *
592 * Warning:
593 * As long as the scan list is non-empty, no other operations than
594 * adding/removing nodes to/from the scan list are allowed.
595 */
596void drm_mm_init_scan_with_range(struct drm_mm *mm,
597 unsigned long size,
598 unsigned alignment,
599 unsigned long color,
600 unsigned long start,
601 unsigned long end)
602{
603 mm->scan_color = color;
604 mm->scan_alignment = alignment;
605 mm->scan_size = size;
606 mm->scanned_blocks = 0;
607 mm->scan_hit_start = 0;
608 mm->scan_hit_end = 0;
609 mm->scan_start = start;
610 mm->scan_end = end;
611 mm->scan_check_range = 1;
612 mm->prev_scanned_node = NULL;
613}
614EXPORT_SYMBOL(drm_mm_init_scan_with_range);
615
616/**
617 * drm_mm_scan_add_block - add a node to the scan list
618 * @node: drm_mm_node to add
619 *
620 * Add a node to the scan list that might be freed to make space for the desired
621 * hole.
622 *
623 * Returns:
624 * True if a hole has been found, false otherwise.
625 */
626bool drm_mm_scan_add_block(struct drm_mm_node *node)
627{
628 struct drm_mm *mm = node->mm;
629 struct drm_mm_node *prev_node;
630 unsigned long hole_start, hole_end;
631 unsigned long adj_start, adj_end;
632
633 mm->scanned_blocks++;
634
635 BUG_ON(node->scanned_block);
636 node->scanned_block = 1;
637
638 prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
639 node_list);
640
641 node->scanned_preceeds_hole = prev_node->hole_follows;
642 prev_node->hole_follows = 1;
643 list_del(&node->node_list);
644 node->node_list.prev = &prev_node->node_list;
645 node->node_list.next = &mm->prev_scanned_node->node_list;
646 mm->prev_scanned_node = node;
647
648 adj_start = hole_start = drm_mm_hole_node_start(prev_node);
649 adj_end = hole_end = drm_mm_hole_node_end(prev_node);
650
651 if (mm->scan_check_range) {
652 if (adj_start < mm->scan_start)
653 adj_start = mm->scan_start;
654 if (adj_end > mm->scan_end)
655 adj_end = mm->scan_end;
656 }
657
658 if (mm->color_adjust)
659 mm->color_adjust(prev_node, mm->scan_color,
660 &adj_start, &adj_end);
661
662 if (check_free_hole(adj_start, adj_end,
663 mm->scan_size, mm->scan_alignment)) {
664 mm->scan_hit_start = hole_start;
665 mm->scan_hit_end = hole_end;
666 return true;
667 }
668
669 return false;
670}
671EXPORT_SYMBOL(drm_mm_scan_add_block);
672
673/**
674 * drm_mm_scan_remove_block - remove a node from the scan list
675 * @node: drm_mm_node to remove
676 *
677 * Nodes _must_ be removed in the exact same order from the scan list as they
678 * have been added, otherwise the internal state of the memory manager will be
679 * corrupted.
680 *
681 * When the scan list is empty, the selected memory nodes can be freed. An
682 * immediately following drm_mm_search_free with !DRM_MM_SEARCH_BEST will then
683 * return the just freed block (because its at the top of the free_stack list).
684 *
685 * Returns:
686 * True if this block should be evicted, false otherwise. Will always
687 * return false when no hole has been found.
688 */
689bool drm_mm_scan_remove_block(struct drm_mm_node *node)
690{
691 struct drm_mm *mm = node->mm;
692 struct drm_mm_node *prev_node;
693
694 mm->scanned_blocks--;
695
696 BUG_ON(!node->scanned_block);
697 node->scanned_block = 0;
698
699 prev_node = list_entry(node->node_list.prev, struct drm_mm_node,
700 node_list);
701
702 prev_node->hole_follows = node->scanned_preceeds_hole;
703 list_add(&node->node_list, &prev_node->node_list);
704
705 return (drm_mm_hole_node_end(node) > mm->scan_hit_start &&
706 node->start < mm->scan_hit_end);
707}
708EXPORT_SYMBOL(drm_mm_scan_remove_block);
709
710/**
711 * drm_mm_clean - checks whether an allocator is clean
712 * @mm: drm_mm allocator to check
713 *
714 * Returns:
715 * True if the allocator is completely free, false if there's still a node
716 * allocated in it.
717 */
718bool drm_mm_clean(struct drm_mm * mm)
719{
720 struct list_head *head = &mm->head_node.node_list;
721
722 return (head->next->next == head);
723}
724EXPORT_SYMBOL(drm_mm_clean);
725
726/**
727 * drm_mm_init - initialize a drm-mm allocator
728 * @mm: the drm_mm structure to initialize
729 * @start: start of the range managed by @mm
730 * @size: end of the range managed by @mm
731 *
732 * Note that @mm must be cleared to 0 before calling this function.
733 */
734void drm_mm_init(struct drm_mm * mm, unsigned long start, unsigned long size)
735{
736 INIT_LIST_HEAD(&mm->hole_stack);
737 mm->scanned_blocks = 0;
738
739 /* Clever trick to avoid a special case in the free hole tracking. */
740 INIT_LIST_HEAD(&mm->head_node.node_list);
741 INIT_LIST_HEAD(&mm->head_node.hole_stack);
742 mm->head_node.hole_follows = 1;
743 mm->head_node.scanned_block = 0;
744 mm->head_node.scanned_prev_free = 0;
745 mm->head_node.scanned_next_free = 0;
746 mm->head_node.mm = mm;
747 mm->head_node.start = start + size;
748 mm->head_node.size = start - mm->head_node.start;
749 list_add_tail(&mm->head_node.hole_stack, &mm->hole_stack);
750
751 mm->color_adjust = NULL;
752}
753EXPORT_SYMBOL(drm_mm_init);
754
755/**
756 * drm_mm_takedown - clean up a drm_mm allocator
757 * @mm: drm_mm allocator to clean up
758 *
759 * Note that it is a bug to call this function on an allocator which is not
760 * clean.
761 */
762void drm_mm_takedown(struct drm_mm * mm)
763{
764 WARN(!list_empty(&mm->head_node.node_list),
765 "Memory manager not clean during takedown.\n");
766}
767EXPORT_SYMBOL(drm_mm_takedown);
768
769static unsigned long drm_mm_debug_hole(struct drm_mm_node *entry,
770 const char *prefix)
771{
772 unsigned long hole_start, hole_end, hole_size;
773
774 if (entry->hole_follows) {
775 hole_start = drm_mm_hole_node_start(entry);
776 hole_end = drm_mm_hole_node_end(entry);
777 hole_size = hole_end - hole_start;
778 printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: free\n",
779 prefix, hole_start, hole_end,
780 hole_size);
781 return hole_size;
782 }
783
784 return 0;
785}
786
787/**
788 * drm_mm_debug_table - dump allocator state to dmesg
789 * @mm: drm_mm allocator to dump
790 * @prefix: prefix to use for dumping to dmesg
791 */
792void drm_mm_debug_table(struct drm_mm *mm, const char *prefix)
793{
794 struct drm_mm_node *entry;
795 unsigned long total_used = 0, total_free = 0, total = 0;
796
797 total_free += drm_mm_debug_hole(&mm->head_node, prefix);
798
799 drm_mm_for_each_node(entry, mm) {
800 printk(KERN_DEBUG "%s 0x%08lx-0x%08lx: %8lu: used\n",
801 prefix, entry->start, entry->start + entry->size,
802 entry->size);
803 total_used += entry->size;
804 total_free += drm_mm_debug_hole(entry, prefix);
805 }
806 total = total_free + total_used;
807
808 printk(KERN_DEBUG "%s total: %lu, used %lu free %lu\n", prefix, total,
809 total_used, total_free);
810}
811EXPORT_SYMBOL(drm_mm_debug_table);
812
813#if defined(CONFIG_DEBUG_FS)
814static unsigned long drm_mm_dump_hole(struct seq_file *m, struct drm_mm_node *entry)
815{
816 unsigned long hole_start, hole_end, hole_size;
817
818 if (entry->hole_follows) {
819 hole_start = drm_mm_hole_node_start(entry);
820 hole_end = drm_mm_hole_node_end(entry);
821 hole_size = hole_end - hole_start;
822 seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: free\n",
823 hole_start, hole_end, hole_size);
824 return hole_size;
825 }
826
827 return 0;
828}
829
830/**
831 * drm_mm_dump_table - dump allocator state to a seq_file
832 * @m: seq_file to dump to
833 * @mm: drm_mm allocator to dump
834 */
835int drm_mm_dump_table(struct seq_file *m, struct drm_mm *mm)
836{
837 struct drm_mm_node *entry;
838 unsigned long total_used = 0, total_free = 0, total = 0;
839
840 total_free += drm_mm_dump_hole(m, &mm->head_node);
841
842 drm_mm_for_each_node(entry, mm) {
843 seq_printf(m, "0x%08lx-0x%08lx: 0x%08lx: used\n",
844 entry->start, entry->start + entry->size,
845 entry->size);
846 total_used += entry->size;
847 total_free += drm_mm_dump_hole(m, entry);
848 }
849 total = total_free + total_used;
850
851 seq_printf(m, "total: %lu, used %lu free %lu\n", total, total_used, total_free);
852 return 0;
853}
854EXPORT_SYMBOL(drm_mm_dump_table);
855#endif