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