Merge tag 'pci-v6.16-fixes-3' of git://git.kernel.org/pub/scm/linux/kernel/git/pci/pci
[linux-2.6-block.git] / drivers / gpu / drm / drm_mm.c
CommitLineData
3a1bd924
TH
1/**************************************************************************
2 *
3 * Copyright 2006 Tungsten Graphics, Inc., Bismarck, ND., USA.
ba004e39 4 * Copyright 2016 Intel Corporation
3a1bd924
TH
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
ba004e39
CW
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.
3a1bd924
TH
38 *
39 * Aligned allocations can also see improvement.
40 *
41 * Authors:
96de0e25 42 * Thomas Hellström <thomas-at-tungstengraphics-dot-com>
3a1bd924
TH
43 */
44
2d1a8a48 45#include <linux/export.h>
202b52b7 46#include <linux/interval_tree_generic.h>
0500c04e
SR
47#include <linux/seq_file.h>
48#include <linux/slab.h>
49#include <linux/stacktrace.h>
50
51#include <drm/drm_mm.h>
1d58420b 52
93110be6
DV
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
05fc0321
DV
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
93110be6
DV
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
ba004e39 72 * dependencies in the driver load sequence.
93110be6
DV
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
05fc0321 82 * supplied. Furthermore every &drm_mm_node has a color value (which is just an
ba004e39 83 * opaque unsigned long) which in conjunction with a driver callback can be used
93110be6
DV
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 *
ba004e39
CW
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.
62347f9e 91 *
93110be6
DV
92 * Finally iteration helpers to walk all nodes and all holes are provided as are
93 * some basic allocator dumpers for debugging.
1c9bd1ed
DV
94 *
95 * Note that this range allocator is not thread-safe, drivers need to protect
21be9154 96 * modifications with their own locking. The idea behind this is that for a full
1c9bd1ed
DV
97 * memory manager additional data needs to be protected anyway, hence internal
98 * locking would be fully redundant.
93110be6
DV
99 */
100
5705670d 101#ifdef CONFIG_DRM_DEBUG_MM
93ce75fa
CW
102#include <linux/stackdepot.h>
103
5705670d
CW
104#define STACKDEPTH 32
105#define BUFSZ 4096
106
107static noinline void save_stack(struct drm_mm_node *node)
108{
109 unsigned long entries[STACKDEPTH];
487f3c7f 110 unsigned int n;
5705670d 111
487f3c7f 112 n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
5705670d
CW
113
114 /* May be called under spinlock, so avoid sleeping */
487f3c7f 115 node->stack = stack_depot_save(entries, n, GFP_NOWAIT);
5705670d
CW
116}
117
118static void show_leaks(struct drm_mm *mm)
119{
120 struct drm_mm_node *node;
5705670d
CW
121 char *buf;
122
123 buf = kmalloc(BUFSZ, GFP_KERNEL);
124 if (!buf)
125 return;
126
2bc98c86 127 list_for_each_entry(node, drm_mm_nodes(mm), node_list) {
5705670d
CW
128 if (!node->stack) {
129 DRM_ERROR("node [%08llx + %08llx]: unknown owner\n",
130 node->start, node->size);
131 continue;
132 }
133
0f68d45e 134 stack_depot_snprint(node->stack, buf, BUFSZ, 0);
5705670d
CW
135 DRM_ERROR("node [%08llx + %08llx]: inserted at\n%s",
136 node->start, node->size, buf);
137 }
138
139 kfree(buf);
140}
141
142#undef STACKDEPTH
143#undef BUFSZ
144#else
145static void save_stack(struct drm_mm_node *node) { }
146static void show_leaks(struct drm_mm *mm) { }
147#endif
148
202b52b7
CW
149#define START(node) ((node)->start)
150#define LAST(node) ((node)->start + (node)->size - 1)
151
152INTERVAL_TREE_DEFINE(struct drm_mm_node, rb,
153 u64, __subtree_last,
53bd7c1c 154 START, LAST, static inline __maybe_unused, drm_mm_interval_tree)
202b52b7
CW
155
156struct drm_mm_node *
45b186f1 157__drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last)
202b52b7 158{
f808c13f 159 return drm_mm_interval_tree_iter_first((struct rb_root_cached *)&mm->interval_tree,
bbba9693 160 start, last) ?: (struct drm_mm_node *)&mm->head_node;
202b52b7 161}
522e85dd 162EXPORT_SYMBOL(__drm_mm_interval_first);
202b52b7
CW
163
164static void drm_mm_interval_tree_add_node(struct drm_mm_node *hole_node,
165 struct drm_mm_node *node)
166{
167 struct drm_mm *mm = hole_node->mm;
168 struct rb_node **link, *rb;
169 struct drm_mm_node *parent;
8a194945 170 bool leftmost;
202b52b7
CW
171
172 node->__subtree_last = LAST(node);
173
71724f70 174 if (drm_mm_node_allocated(hole_node)) {
202b52b7
CW
175 rb = &hole_node->rb;
176 while (rb) {
177 parent = rb_entry(rb, struct drm_mm_node, rb);
178 if (parent->__subtree_last >= node->__subtree_last)
179 break;
180
181 parent->__subtree_last = node->__subtree_last;
182 rb = rb_parent(rb);
183 }
184
185 rb = &hole_node->rb;
186 link = &hole_node->rb.rb_right;
f808c13f 187 leftmost = false;
202b52b7
CW
188 } else {
189 rb = NULL;
f808c13f 190 link = &mm->interval_tree.rb_root.rb_node;
8a194945 191 leftmost = true;
202b52b7
CW
192 }
193
194 while (*link) {
195 rb = *link;
196 parent = rb_entry(rb, struct drm_mm_node, rb);
197 if (parent->__subtree_last < node->__subtree_last)
198 parent->__subtree_last = node->__subtree_last;
8a194945 199 if (node->start < parent->start) {
202b52b7 200 link = &parent->rb.rb_left;
8a194945 201 } else {
202b52b7 202 link = &parent->rb.rb_right;
8a194945 203 leftmost = false;
f808c13f 204 }
202b52b7
CW
205 }
206
207 rb_link_node(&node->rb, rb, link);
f808c13f
DB
208 rb_insert_augmented_cached(&node->rb, &mm->interval_tree, leftmost,
209 &drm_mm_interval_tree_augment);
202b52b7
CW
210}
211
4e64e553
CW
212#define HOLE_SIZE(NODE) ((NODE)->hole_size)
213#define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE))
214
2f7e8769
CW
215static u64 rb_to_hole_size(struct rb_node *rb)
216{
217 return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size;
218}
219
220static void insert_hole_size(struct rb_root_cached *root,
221 struct drm_mm_node *node)
222{
223 struct rb_node **link = &root->rb_root.rb_node, *rb = NULL;
224 u64 x = node->hole_size;
225 bool first = true;
226
227 while (*link) {
228 rb = *link;
229 if (x > rb_to_hole_size(rb)) {
230 link = &rb->rb_left;
231 } else {
232 link = &rb->rb_right;
233 first = false;
234 }
235 }
236
237 rb_link_node(&node->rb_hole_size, rb, link);
238 rb_insert_color_cached(&node->rb_hole_size, root, first);
239}
240
0cdea445
ND
241RB_DECLARE_CALLBACKS_MAX(static, augment_callbacks,
242 struct drm_mm_node, rb_hole_addr,
243 u64, subtree_max_hole, HOLE_SIZE)
244
245static void insert_hole_addr(struct rb_root *root, struct drm_mm_node *node)
246{
247 struct rb_node **link = &root->rb_node, *rb_parent = NULL;
248 u64 start = HOLE_ADDR(node), subtree_max_hole = node->subtree_max_hole;
249 struct drm_mm_node *parent;
250
251 while (*link) {
252 rb_parent = *link;
253 parent = rb_entry(rb_parent, struct drm_mm_node, rb_hole_addr);
254 if (parent->subtree_max_hole < subtree_max_hole)
255 parent->subtree_max_hole = subtree_max_hole;
256 if (start < HOLE_ADDR(parent))
257 link = &parent->rb_hole_addr.rb_left;
258 else
259 link = &parent->rb_hole_addr.rb_right;
260 }
261
262 rb_link_node(&node->rb_hole_addr, rb_parent, link);
263 rb_insert_augmented(&node->rb_hole_addr, root, &augment_callbacks);
264}
265
4e64e553 266static void add_hole(struct drm_mm_node *node)
3a1bd924 267{
4e64e553 268 struct drm_mm *mm = node->mm;
ea7b1dd4 269
4e64e553
CW
270 node->hole_size =
271 __drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node);
0cdea445 272 node->subtree_max_hole = node->hole_size;
4e64e553 273 DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
b0b7af18 274
2f7e8769 275 insert_hole_size(&mm->holes_size, node);
0cdea445 276 insert_hole_addr(&mm->holes_addr, node);
1d58420b 277
4e64e553
CW
278 list_add(&node->hole_stack, &mm->hole_stack);
279}
adb040b8 280
4e64e553
CW
281static void rm_hole(struct drm_mm_node *node)
282{
283 DRM_MM_BUG_ON(!drm_mm_hole_follows(node));
62347f9e 284
4e64e553 285 list_del(&node->hole_stack);
2f7e8769 286 rb_erase_cached(&node->rb_hole_size, &node->mm->holes_size);
0cdea445
ND
287 rb_erase_augmented(&node->rb_hole_addr, &node->mm->holes_addr,
288 &augment_callbacks);
4e64e553 289 node->hole_size = 0;
0cdea445 290 node->subtree_max_hole = 0;
440fd528 291
4e64e553
CW
292 DRM_MM_BUG_ON(drm_mm_hole_follows(node));
293}
294
295static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb)
296{
297 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size);
298}
299
300static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb)
301{
302 return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr);
303}
304
4e64e553
CW
305static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size)
306{
2f7e8769
CW
307 struct rb_node *rb = mm->holes_size.rb_root.rb_node;
308 struct drm_mm_node *best = NULL;
4e64e553 309
2f7e8769
CW
310 do {
311 struct drm_mm_node *node =
312 rb_entry(rb, struct drm_mm_node, rb_hole_size);
4e64e553 313
2f7e8769
CW
314 if (size <= node->hole_size) {
315 best = node;
316 rb = rb->rb_right;
4e64e553 317 } else {
2f7e8769 318 rb = rb->rb_left;
62347f9e 319 }
2f7e8769 320 } while (rb);
6b9d89b4 321
2f7e8769 322 return best;
4e64e553
CW
323}
324
5fad79fd
CK
325static bool usable_hole_addr(struct rb_node *rb, u64 size)
326{
327 return rb && rb_hole_addr_to_node(rb)->subtree_max_hole >= size;
328}
329
271e7dec 330static struct drm_mm_node *find_hole_addr(struct drm_mm *mm, u64 addr, u64 size)
4e64e553 331{
2f7e8769 332 struct rb_node *rb = mm->holes_addr.rb_node;
4e64e553 333 struct drm_mm_node *node = NULL;
4e64e553 334
2f7e8769 335 while (rb) {
4e64e553
CW
336 u64 hole_start;
337
5fad79fd 338 if (!usable_hole_addr(rb, size))
271e7dec
CK
339 break;
340
2f7e8769 341 node = rb_hole_addr_to_node(rb);
4e64e553
CW
342 hole_start = __drm_mm_hole_node_start(node);
343
344 if (addr < hole_start)
2f7e8769 345 rb = node->rb_hole_addr.rb_left;
4e64e553 346 else if (addr > hole_start + node->hole_size)
2f7e8769 347 rb = node->rb_hole_addr.rb_right;
4e64e553
CW
348 else
349 break;
6b9d89b4 350 }
ea7b1dd4 351
4e64e553
CW
352 return node;
353}
3a1bd924 354
4e64e553
CW
355static struct drm_mm_node *
356first_hole(struct drm_mm *mm,
357 u64 start, u64 end, u64 size,
358 enum drm_mm_insert_mode mode)
359{
4e64e553
CW
360 switch (mode) {
361 default:
362 case DRM_MM_INSERT_BEST:
363 return best_hole(mm, size);
202b52b7 364
4e64e553 365 case DRM_MM_INSERT_LOW:
271e7dec 366 return find_hole_addr(mm, start, size);
ea7b1dd4 367
4e64e553 368 case DRM_MM_INSERT_HIGH:
271e7dec 369 return find_hole_addr(mm, end, size);
4e64e553
CW
370
371 case DRM_MM_INSERT_EVICT:
372 return list_first_entry_or_null(&mm->hole_stack,
373 struct drm_mm_node,
374 hole_stack);
1d58420b 375 }
4e64e553 376}
5705670d 377
0cdea445 378/**
5fad79fd
CK
379 * DECLARE_NEXT_HOLE_ADDR - macro to declare next hole functions
380 * @name: name of function to declare
381 * @first: first rb member to traverse (either rb_left or rb_right).
382 * @last: last rb member to traverse (either rb_right or rb_left).
0cdea445 383 *
5fad79fd
CK
384 * This macro declares a function to return the next hole of the addr rb tree.
385 * While traversing the tree we take the searched size into account and only
386 * visit branches with potential big enough holes.
0cdea445 387 */
0cdea445 388
5fad79fd
CK
389#define DECLARE_NEXT_HOLE_ADDR(name, first, last) \
390static struct drm_mm_node *name(struct drm_mm_node *entry, u64 size) \
391{ \
392 struct rb_node *parent, *node = &entry->rb_hole_addr; \
393 \
394 if (!entry || RB_EMPTY_NODE(node)) \
395 return NULL; \
396 \
397 if (usable_hole_addr(node->first, size)) { \
398 node = node->first; \
399 while (usable_hole_addr(node->last, size)) \
400 node = node->last; \
401 return rb_hole_addr_to_node(node); \
402 } \
403 \
404 while ((parent = rb_parent(node)) && node == parent->first) \
405 node = parent; \
406 \
407 return rb_hole_addr_to_node(parent); \
0cdea445
ND
408}
409
5fad79fd
CK
410DECLARE_NEXT_HOLE_ADDR(next_hole_high_addr, rb_left, rb_right)
411DECLARE_NEXT_HOLE_ADDR(next_hole_low_addr, rb_right, rb_left)
0cdea445 412
4e64e553
CW
413static struct drm_mm_node *
414next_hole(struct drm_mm *mm,
415 struct drm_mm_node *node,
0cdea445 416 u64 size,
4e64e553
CW
417 enum drm_mm_insert_mode mode)
418{
419 switch (mode) {
420 default:
421 case DRM_MM_INSERT_BEST:
2f7e8769 422 return rb_hole_size_to_node(rb_prev(&node->rb_hole_size));
4e64e553
CW
423
424 case DRM_MM_INSERT_LOW:
0cdea445 425 return next_hole_low_addr(node, size);
4e64e553
CW
426
427 case DRM_MM_INSERT_HIGH:
0cdea445 428 return next_hole_high_addr(node, size);
4e64e553
CW
429
430 case DRM_MM_INSERT_EVICT:
431 node = list_next_entry(node, hole_stack);
432 return &node->hole_stack == &mm->hole_stack ? NULL : node;
433 }
9fc935de
DV
434}
435
e18c0412
DV
436/**
437 * drm_mm_reserve_node - insert an pre-initialized node
438 * @mm: drm_mm allocator to insert @node into
439 * @node: drm_mm_node to insert
440 *
05fc0321
DV
441 * This functions inserts an already set-up &drm_mm_node into the allocator,
442 * meaning that start, size and color must be set by the caller. All other
443 * fields must be cleared to 0. This is useful to initialize the allocator with
444 * preallocated objects which must be set-up before the range allocator can be
445 * set-up, e.g. when taking over a firmware framebuffer.
e18c0412
DV
446 *
447 * Returns:
448 * 0 on success, -ENOSPC if there's no hole where @node is.
449 */
338710e7 450int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node)
5973c7ee 451{
b3a070cc 452 struct drm_mm_node *hole;
202b52b7 453 u64 hole_start, hole_end;
2db86dfc 454 u64 adj_start, adj_end;
0d1650fa 455 u64 end;
338710e7 456
b80d3942 457 end = node->start + node->size;
c820186d
CW
458 if (unlikely(end <= node->start))
459 return -ENOSPC;
b80d3942 460
338710e7 461 /* Find the relevant hole to add our node to */
271e7dec 462 hole = find_hole_addr(mm, node->start, 0);
4e64e553 463 if (!hole)
202b52b7 464 return -ENOSPC;
5973c7ee 465
2db86dfc 466 adj_start = hole_start = __drm_mm_hole_node_start(hole);
4e64e553 467 adj_end = hole_end = hole_start + hole->hole_size;
2db86dfc
CW
468
469 if (mm->color_adjust)
470 mm->color_adjust(hole, node->color, &adj_start, &adj_end);
471
472 if (adj_start > node->start || adj_end < end)
202b52b7 473 return -ENOSPC;
5973c7ee 474
202b52b7 475 node->mm = mm;
5973c7ee 476
3dda22d3 477 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
202b52b7 478 list_add(&node->node_list, &hole->node_list);
202b52b7 479 drm_mm_interval_tree_add_node(hole, node);
4e64e553 480 node->hole_size = 0;
202b52b7 481
4e64e553
CW
482 rm_hole(hole);
483 if (node->start > hole_start)
484 add_hole(hole);
485 if (end < hole_end)
486 add_hole(node);
5973c7ee 487
5705670d 488 save_stack(node);
202b52b7 489 return 0;
5973c7ee 490}
338710e7 491EXPORT_SYMBOL(drm_mm_reserve_node);
5973c7ee 492
2f7e8769
CW
493static u64 rb_to_hole_size_or_zero(struct rb_node *rb)
494{
495 return rb ? rb_to_hole_size(rb) : 0;
496}
497
b0b7af18 498/**
4e64e553 499 * drm_mm_insert_node_in_range - ranged search for space and insert @node
e18c0412
DV
500 * @mm: drm_mm to allocate from
501 * @node: preallocate node to insert
502 * @size: size of the allocation
503 * @alignment: alignment of the allocation
504 * @color: opaque tag value to use for this node
4e64e553
CW
505 * @range_start: start of the allowed range for this node
506 * @range_end: end of the allowed range for this node
507 * @mode: fine-tune the allocation search and placement
e18c0412 508 *
05fc0321 509 * The preallocated @node must be cleared to 0.
e18c0412
DV
510 *
511 * Returns:
512 * 0 on success, -ENOSPC if there's no suitable hole.
3a1bd924 513 */
4e64e553
CW
514int drm_mm_insert_node_in_range(struct drm_mm * const mm,
515 struct drm_mm_node * const node,
516 u64 size, u64 alignment,
517 unsigned long color,
518 u64 range_start, u64 range_end,
519 enum drm_mm_insert_mode mode)
3a1bd924 520{
4e64e553
CW
521 struct drm_mm_node *hole;
522 u64 remainder_mask;
83bc4ec3 523 bool once;
b0b7af18 524
c1a495a5 525 DRM_MM_BUG_ON(range_start > range_end);
aafdcfd3 526
4e64e553 527 if (unlikely(size == 0 || range_end - range_start < size))
b0b7af18
DV
528 return -ENOSPC;
529
2f7e8769
CW
530 if (rb_to_hole_size_or_zero(rb_first_cached(&mm->holes_size)) < size)
531 return -ENOSPC;
532
4e64e553
CW
533 if (alignment <= 1)
534 alignment = 0;
535
83bc4ec3
CW
536 once = mode & DRM_MM_INSERT_ONCE;
537 mode &= ~DRM_MM_INSERT_ONCE;
538
4e64e553 539 remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
83bc4ec3
CW
540 for (hole = first_hole(mm, range_start, range_end, size, mode);
541 hole;
0cdea445 542 hole = once ? NULL : next_hole(mm, hole, size, mode)) {
4e64e553
CW
543 u64 hole_start = __drm_mm_hole_node_start(hole);
544 u64 hole_end = hole_start + hole->hole_size;
545 u64 adj_start, adj_end;
546 u64 col_start, col_end;
547
548 if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end)
549 break;
550
551 if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start)
552 break;
553
554 col_start = hole_start;
555 col_end = hole_end;
556 if (mm->color_adjust)
557 mm->color_adjust(hole, color, &col_start, &col_end);
558
559 adj_start = max(col_start, range_start);
560 adj_end = min(col_end, range_end);
561
562 if (adj_end <= adj_start || adj_end - adj_start < size)
563 continue;
564
565 if (mode == DRM_MM_INSERT_HIGH)
566 adj_start = adj_end - size;
567
568 if (alignment) {
569 u64 rem;
570
571 if (likely(remainder_mask))
572 rem = adj_start & remainder_mask;
573 else
574 div64_u64_rem(adj_start, alignment, &rem);
575 if (rem) {
576 adj_start -= rem;
577 if (mode != DRM_MM_INSERT_HIGH)
578 adj_start += alignment;
579
580 if (adj_start < max(col_start, range_start) ||
581 min(col_end, range_end) - adj_start < size)
582 continue;
583
584 if (adj_end <= adj_start ||
585 adj_end - adj_start < size)
586 continue;
587 }
588 }
589
590 node->mm = mm;
591 node->size = size;
592 node->start = adj_start;
593 node->color = color;
594 node->hole_size = 0;
595
3dda22d3 596 __set_bit(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
4e64e553
CW
597 list_add(&node->node_list, &hole->node_list);
598 drm_mm_interval_tree_add_node(hole, node);
4e64e553
CW
599
600 rm_hole(hole);
601 if (adj_start > hole_start)
602 add_hole(hole);
603 if (adj_start + size < hole_end)
604 add_hole(node);
605
606 save_stack(node);
607 return 0;
608 }
609
2713778c 610 return -ENOSPC;
b0b7af18 611}
4e64e553 612EXPORT_SYMBOL(drm_mm_insert_node_in_range);
b8103450 613
461ed4aa 614static inline __maybe_unused bool drm_mm_node_scanned_block(const struct drm_mm_node *node)
71724f70 615{
4ee92c71 616 return test_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
71724f70
CW
617}
618
b0b7af18 619/**
e18c0412
DV
620 * drm_mm_remove_node - Remove a memory node from the allocator.
621 * @node: drm_mm_node to remove
622 *
623 * This just removes a node from its drm_mm allocator. The node does not need to
624 * be cleared again before it can be re-inserted into this or any other drm_mm
ba004e39 625 * allocator. It is a bug to call this function on a unallocated node.
b0b7af18
DV
626 */
627void drm_mm_remove_node(struct drm_mm_node *node)
628{
ea7b1dd4
DV
629 struct drm_mm *mm = node->mm;
630 struct drm_mm_node *prev_node;
3a1bd924 631
71724f70
CW
632 DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
633 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
3a1bd924 634
4e64e553 635 prev_node = list_prev_entry(node, node_list);
9e8944ab 636
4e64e553
CW
637 if (drm_mm_hole_follows(node))
638 rm_hole(node);
ea7b1dd4 639
202b52b7 640 drm_mm_interval_tree_remove(node, &mm->interval_tree);
ea7b1dd4 641 list_del(&node->node_list);
7a6b2896 642
4e64e553
CW
643 if (drm_mm_hole_follows(prev_node))
644 rm_hole(prev_node);
645 add_hole(prev_node);
3dda22d3
CW
646
647 clear_bit_unlock(DRM_MM_NODE_ALLOCATED_BIT, &node->flags);
a2e68e92 648}
4e64e553 649EXPORT_SYMBOL(drm_mm_remove_node);
a2e68e92 650
93110be6 651/**
05fc0321 652 * DOC: lru scan roster
93110be6
DV
653 *
654 * Very often GPUs need to have continuous allocations for a given object. When
655 * evicting objects to make space for a new one it is therefore not most
656 * efficient when we simply start to select all objects from the tail of an LRU
657 * until there's a suitable hole: Especially for big objects or nodes that
658 * otherwise have special allocation constraints there's a good chance we evict
ba004e39 659 * lots of (smaller) objects unnecessarily.
93110be6
DV
660 *
661 * The DRM range allocator supports this use-case through the scanning
662 * interfaces. First a scan operation needs to be initialized with
9a71e277 663 * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds
05fc0321 664 * objects to the roster, probably by walking an LRU list, but this can be
0ae865ef 665 * freely implemented. Eviction candidates are added using
05fc0321 666 * drm_mm_scan_add_block() until a suitable hole is found or there are no
940eba2d
DV
667 * further evictable objects. Eviction roster metadata is tracked in &struct
668 * drm_mm_scan.
93110be6 669 *
ba004e39 670 * The driver must walk through all objects again in exactly the reverse
93110be6
DV
671 * order to restore the allocator state. Note that while the allocator is used
672 * in the scan mode no other operation is allowed.
673 *
3fa489da
CW
674 * Finally the driver evicts all objects selected (drm_mm_scan_remove_block()
675 * reported true) in the scan, and any overlapping nodes after color adjustment
05fc0321 676 * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and
3fa489da
CW
677 * since freeing a node is also O(1) the overall complexity is
678 * O(scanned_objects). So like the free stack which needs to be walked before a
679 * scan operation even begins this is linear in the number of objects. It
680 * doesn't seem to hurt too badly.
93110be6
DV
681 */
682
d935cc61 683/**
9a71e277
CW
684 * drm_mm_scan_init_with_range - initialize range-restricted lru scanning
685 * @scan: scan state
e18c0412
DV
686 * @mm: drm_mm to scan
687 * @size: size of the allocation
688 * @alignment: alignment of the allocation
689 * @color: opaque tag value to use for the allocation
690 * @start: start of the allowed range for the allocation
691 * @end: end of the allowed range for the allocation
4e64e553 692 * @mode: fine-tune the allocation search and placement
d935cc61
DV
693 *
694 * This simply sets up the scanning routines with the parameters for the desired
0b04d474 695 * hole.
d935cc61 696 *
e18c0412
DV
697 * Warning:
698 * As long as the scan list is non-empty, no other operations than
d935cc61
DV
699 * adding/removing nodes to/from the scan list are allowed.
700 */
9a71e277
CW
701void drm_mm_scan_init_with_range(struct drm_mm_scan *scan,
702 struct drm_mm *mm,
440fd528 703 u64 size,
71733207 704 u64 alignment,
6b9d89b4 705 unsigned long color,
440fd528 706 u64 start,
0b04d474 707 u64 end,
4e64e553 708 enum drm_mm_insert_mode mode)
d935cc61 709{
6259a56b
CW
710 DRM_MM_BUG_ON(start >= end);
711 DRM_MM_BUG_ON(!size || size > end - start);
9a71e277
CW
712 DRM_MM_BUG_ON(mm->scan_active);
713
714 scan->mm = mm;
715
9a956b15
CW
716 if (alignment <= 1)
717 alignment = 0;
718
9a71e277
CW
719 scan->color = color;
720 scan->alignment = alignment;
9a956b15 721 scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0;
9a71e277 722 scan->size = size;
4e64e553 723 scan->mode = mode;
9a71e277
CW
724
725 DRM_MM_BUG_ON(end <= start);
726 scan->range_start = start;
727 scan->range_end = end;
6259a56b 728
9a71e277
CW
729 scan->hit_start = U64_MAX;
730 scan->hit_end = 0;
d935cc61 731}
9a71e277 732EXPORT_SYMBOL(drm_mm_scan_init_with_range);
d935cc61 733
709ea971 734/**
e18c0412 735 * drm_mm_scan_add_block - add a node to the scan list
9b8b75de 736 * @scan: the active drm_mm scanner
e18c0412
DV
737 * @node: drm_mm_node to add
738 *
709ea971
DV
739 * Add a node to the scan list that might be freed to make space for the desired
740 * hole.
741 *
e18c0412
DV
742 * Returns:
743 * True if a hole has been found, false otherwise.
709ea971 744 */
9a71e277
CW
745bool drm_mm_scan_add_block(struct drm_mm_scan *scan,
746 struct drm_mm_node *node)
709ea971 747{
9a71e277 748 struct drm_mm *mm = scan->mm;
4a6c156f 749 struct drm_mm_node *hole;
440fd528 750 u64 hole_start, hole_end;
268c6498 751 u64 col_start, col_end;
440fd528 752 u64 adj_start, adj_end;
709ea971 753
9a71e277 754 DRM_MM_BUG_ON(node->mm != mm);
71724f70
CW
755 DRM_MM_BUG_ON(!drm_mm_node_allocated(node));
756 DRM_MM_BUG_ON(drm_mm_node_scanned_block(node));
4ee92c71 757 __set_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
9a71e277 758 mm->scan_active++;
709ea971 759
f29051f1
CW
760 /* Remove this block from the node_list so that we enlarge the hole
761 * (distance between the end of our previous node and the start of
762 * or next), without poisoning the link so that we can restore it
763 * later in drm_mm_scan_remove_block().
764 */
4a6c156f 765 hole = list_prev_entry(node, node_list);
f29051f1
CW
766 DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node);
767 __list_del_entry(&node->node_list);
709ea971 768
268c6498
CW
769 hole_start = __drm_mm_hole_node_start(hole);
770 hole_end = __drm_mm_hole_node_end(hole);
d935cc61 771
268c6498
CW
772 col_start = hole_start;
773 col_end = hole_end;
901593f2 774 if (mm->color_adjust)
268c6498
CW
775 mm->color_adjust(hole, scan->color, &col_start, &col_end);
776
777 adj_start = max(col_start, scan->range_start);
778 adj_end = min(col_end, scan->range_end);
0b04d474
CW
779 if (adj_end <= adj_start || adj_end - adj_start < scan->size)
780 return false;
781
4e64e553 782 if (scan->mode == DRM_MM_INSERT_HIGH)
0b04d474
CW
783 adj_start = adj_end - scan->size;
784
785 if (scan->alignment) {
786 u64 rem;
787
9a956b15
CW
788 if (likely(scan->remainder_mask))
789 rem = adj_start & scan->remainder_mask;
790 else
791 div64_u64_rem(adj_start, scan->alignment, &rem);
0b04d474
CW
792 if (rem) {
793 adj_start -= rem;
4e64e553 794 if (scan->mode != DRM_MM_INSERT_HIGH)
0b04d474
CW
795 adj_start += scan->alignment;
796 if (adj_start < max(col_start, scan->range_start) ||
797 min(col_end, scan->range_end) - adj_start < scan->size)
798 return false;
799
800 if (adj_end <= adj_start ||
801 adj_end - adj_start < scan->size)
802 return false;
803 }
804 }
901593f2 805
3fa489da
CW
806 scan->hit_start = adj_start;
807 scan->hit_end = adj_start + scan->size;
709ea971 808
0b04d474
CW
809 DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end);
810 DRM_MM_BUG_ON(scan->hit_start < hole_start);
811 DRM_MM_BUG_ON(scan->hit_end > hole_end);
812
813 return true;
709ea971
DV
814}
815EXPORT_SYMBOL(drm_mm_scan_add_block);
816
817/**
e18c0412 818 * drm_mm_scan_remove_block - remove a node from the scan list
9b8b75de 819 * @scan: the active drm_mm scanner
e18c0412 820 * @node: drm_mm_node to remove
709ea971 821 *
05fc0321
DV
822 * Nodes **must** be removed in exactly the reverse order from the scan list as
823 * they have been added (e.g. using list_add() as they are added and then
824 * list_for_each() over that eviction list to remove), otherwise the internal
ba004e39 825 * state of the memory manager will be corrupted.
709ea971
DV
826 *
827 * When the scan list is empty, the selected memory nodes can be freed. An
05fc0321
DV
828 * immediately following drm_mm_insert_node_in_range_generic() or one of the
829 * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return
1e55a53a 830 * the just freed block (because it's at the top of the free_stack list).
709ea971 831 *
e18c0412
DV
832 * Returns:
833 * True if this block should be evicted, false otherwise. Will always
834 * return false when no hole has been found.
709ea971 835 */
9a71e277
CW
836bool drm_mm_scan_remove_block(struct drm_mm_scan *scan,
837 struct drm_mm_node *node)
709ea971 838{
ea7b1dd4 839 struct drm_mm_node *prev_node;
709ea971 840
9a71e277 841 DRM_MM_BUG_ON(node->mm != scan->mm);
71724f70 842 DRM_MM_BUG_ON(!drm_mm_node_scanned_block(node));
4ee92c71 843 __clear_bit(DRM_MM_NODE_SCANNED_BIT, &node->flags);
709ea971 844
9a71e277
CW
845 DRM_MM_BUG_ON(!node->mm->scan_active);
846 node->mm->scan_active--;
847
f29051f1
CW
848 /* During drm_mm_scan_add_block() we decoupled this node leaving
849 * its pointers intact. Now that the caller is walking back along
850 * the eviction list we can restore this block into its rightful
851 * place on the full node_list. To confirm that the caller is walking
852 * backwards correctly we check that prev_node->next == node->next,
853 * i.e. both believe the same node should be on the other side of the
854 * hole.
855 */
9a71e277 856 prev_node = list_prev_entry(node, node_list);
f29051f1
CW
857 DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) !=
858 list_next_entry(node, node_list));
ea7b1dd4 859 list_add(&node->node_list, &prev_node->node_list);
709ea971 860
0b04d474 861 return (node->start + node->size > scan->hit_start &&
9a71e277 862 node->start < scan->hit_end);
709ea971
DV
863}
864EXPORT_SYMBOL(drm_mm_scan_remove_block);
865
3fa489da
CW
866/**
867 * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole
868 * @scan: drm_mm scan with target hole
869 *
870 * After completing an eviction scan and removing the selected nodes, we may
871 * need to remove a few more nodes from either side of the target hole if
872 * mm.color_adjust is being used.
873 *
874 * Returns:
875 * A node to evict, or NULL if there are no overlapping nodes.
876 */
877struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan)
878{
879 struct drm_mm *mm = scan->mm;
880 struct drm_mm_node *hole;
881 u64 hole_start, hole_end;
882
883 DRM_MM_BUG_ON(list_empty(&mm->hole_stack));
884
885 if (!mm->color_adjust)
886 return NULL;
887
b8ff1802
CW
888 /*
889 * The hole found during scanning should ideally be the first element
890 * in the hole_stack list, but due to side-effects in the driver it
891 * may not be.
892 */
893 list_for_each_entry(hole, &mm->hole_stack, hole_stack) {
894 hole_start = __drm_mm_hole_node_start(hole);
895 hole_end = hole_start + hole->hole_size;
896
897 if (hole_start <= scan->hit_start &&
898 hole_end >= scan->hit_end)
899 break;
900 }
901
902 /* We should only be called after we found the hole previously */
903 DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack);
904 if (unlikely(&hole->hole_stack == &mm->hole_stack))
905 return NULL;
3fa489da
CW
906
907 DRM_MM_BUG_ON(hole_start > scan->hit_start);
908 DRM_MM_BUG_ON(hole_end < scan->hit_end);
909
910 mm->color_adjust(hole, scan->color, &hole_start, &hole_end);
911 if (hole_start > scan->hit_start)
912 return hole;
913 if (hole_end < scan->hit_end)
914 return list_next_entry(hole, node_list);
915
916 return NULL;
917}
918EXPORT_SYMBOL(drm_mm_scan_color_evict);
919
e18c0412
DV
920/**
921 * drm_mm_init - initialize a drm-mm allocator
922 * @mm: the drm_mm structure to initialize
923 * @start: start of the range managed by @mm
924 * @size: end of the range managed by @mm
925 *
926 * Note that @mm must be cleared to 0 before calling this function.
927 */
45b186f1 928void drm_mm_init(struct drm_mm *mm, u64 start, u64 size)
1d58420b 929{
6259a56b
CW
930 DRM_MM_BUG_ON(start + size <= start);
931
4e64e553
CW
932 mm->color_adjust = NULL;
933
ea7b1dd4 934 INIT_LIST_HEAD(&mm->hole_stack);
f808c13f 935 mm->interval_tree = RB_ROOT_CACHED;
2f7e8769 936 mm->holes_size = RB_ROOT_CACHED;
4e64e553 937 mm->holes_addr = RB_ROOT;
3a1bd924 938
ea7b1dd4
DV
939 /* Clever trick to avoid a special case in the free hole tracking. */
940 INIT_LIST_HEAD(&mm->head_node.node_list);
4ee92c71 941 mm->head_node.flags = 0;
ea7b1dd4
DV
942 mm->head_node.mm = mm;
943 mm->head_node.start = start + size;
4e64e553
CW
944 mm->head_node.size = -size;
945 add_hole(&mm->head_node);
ea7b1dd4 946
4e64e553 947 mm->scan_active = 0;
2dba5eb1
VB
948
949#ifdef CONFIG_DRM_DEBUG_MM
950 stack_depot_init();
951#endif
3a1bd924 952}
673a394b 953EXPORT_SYMBOL(drm_mm_init);
3a1bd924 954
e18c0412
DV
955/**
956 * drm_mm_takedown - clean up a drm_mm allocator
957 * @mm: drm_mm allocator to clean up
958 *
959 * Note that it is a bug to call this function on an allocator which is not
960 * clean.
961 */
5705670d 962void drm_mm_takedown(struct drm_mm *mm)
3a1bd924 963{
ac9bb7b7 964 if (WARN(!drm_mm_clean(mm),
5705670d
CW
965 "Memory manager not clean during takedown.\n"))
966 show_leaks(mm);
3a1bd924 967}
f453ba04 968EXPORT_SYMBOL(drm_mm_takedown);
fa8a1238 969
b5c3714f 970static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry)
99d7e48e 971{
4e64e553
CW
972 u64 start, size;
973
974 size = entry->hole_size;
975 if (size) {
976 start = drm_mm_hole_node_start(entry);
977 drm_printf(p, "%#018llx-%#018llx: %llu: free\n",
978 start, start + size, size);
3a359f0b
DV
979 }
980
4e64e553 981 return size;
3a359f0b 982}
e18c0412 983/**
b5c3714f
DV
984 * drm_mm_print - print allocator state
985 * @mm: drm_mm allocator to print
986 * @p: DRM printer to use
e18c0412 987 */
b5c3714f 988void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p)
3a359f0b 989{
45b186f1 990 const struct drm_mm_node *entry;
440fd528 991 u64 total_used = 0, total_free = 0, total = 0;
3a359f0b 992
b5c3714f 993 total_free += drm_mm_dump_hole(p, &mm->head_node);
ea7b1dd4
DV
994
995 drm_mm_for_each_node(entry, mm) {
b5c3714f 996 drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start,
440fd528 997 entry->start + entry->size, entry->size);
ea7b1dd4 998 total_used += entry->size;
b5c3714f 999 total_free += drm_mm_dump_hole(p, entry);
fa8a1238 1000 }
ea7b1dd4
DV
1001 total = total_free + total_used;
1002
b5c3714f 1003 drm_printf(p, "total: %llu, used %llu free %llu\n", total,
440fd528 1004 total_used, total_free);
fa8a1238 1005}
b5c3714f 1006EXPORT_SYMBOL(drm_mm_print);