]>
Commit | Line | Data |
---|---|---|
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 | ||
760285e7 DH |
45 | #include <drm/drmP.h> |
46 | #include <drm/drm_mm.h> | |
1d58420b | 47 | #include <linux/slab.h> |
fa8a1238 | 48 | #include <linux/seq_file.h> |
2d1a8a48 | 49 | #include <linux/export.h> |
202b52b7 | 50 | #include <linux/interval_tree_generic.h> |
1d58420b | 51 | |
93110be6 DV |
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 | |
05fc0321 DV |
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 | |
93110be6 DV |
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 | |
ba004e39 | 71 | * dependencies in the driver load sequence. |
93110be6 DV |
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 | |
05fc0321 | 81 | * supplied. Furthermore every &drm_mm_node has a color value (which is just an |
ba004e39 | 82 | * opaque unsigned long) which in conjunction with a driver callback can be used |
93110be6 DV |
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 | * | |
ba004e39 CW |
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. | |
62347f9e | 90 | * |
93110be6 DV |
91 | * Finally iteration helpers to walk all nodes and all holes are provided as are |
92 | * some basic allocator dumpers for debugging. | |
1c9bd1ed DV |
93 | * |
94 | * Note that this range allocator is not thread-safe, drivers need to protect | |
95 | * modifications with their on 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. | |
93110be6 DV |
98 | */ |
99 | ||
5705670d | 100 | #ifdef CONFIG_DRM_DEBUG_MM |
93ce75fa CW |
101 | #include <linux/stackdepot.h> |
102 | ||
5705670d CW |
103 | #define STACKDEPTH 32 |
104 | #define BUFSZ 4096 | |
105 | ||
106 | static 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 | ||
124 | static 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 | ||
2bc98c86 | 134 | list_for_each_entry(node, drm_mm_nodes(mm), node_list) { |
5705670d CW |
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 | |
158 | static void save_stack(struct drm_mm_node *node) { } | |
159 | static void show_leaks(struct drm_mm *mm) { } | |
160 | #endif | |
161 | ||
202b52b7 CW |
162 | #define START(node) ((node)->start) |
163 | #define LAST(node) ((node)->start + (node)->size - 1) | |
164 | ||
165 | INTERVAL_TREE_DEFINE(struct drm_mm_node, rb, | |
166 | u64, __subtree_last, | |
167 | START, LAST, static inline, drm_mm_interval_tree) | |
168 | ||
169 | struct drm_mm_node * | |
45b186f1 | 170 | __drm_mm_interval_first(const struct drm_mm *mm, u64 start, u64 last) |
202b52b7 | 171 | { |
f808c13f | 172 | return drm_mm_interval_tree_iter_first((struct rb_root_cached *)&mm->interval_tree, |
bbba9693 | 173 | start, last) ?: (struct drm_mm_node *)&mm->head_node; |
202b52b7 | 174 | } |
522e85dd | 175 | EXPORT_SYMBOL(__drm_mm_interval_first); |
202b52b7 CW |
176 | |
177 | static 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; | |
f808c13f | 183 | bool leftmost = true; |
202b52b7 CW |
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; | |
f808c13f | 200 | leftmost = false; |
202b52b7 CW |
201 | } else { |
202 | rb = NULL; | |
f808c13f | 203 | link = &mm->interval_tree.rb_root.rb_node; |
202b52b7 CW |
204 | } |
205 | ||
206 | while (*link) { | |
207 | rb = *link; | |
208 | parent = rb_entry(rb, struct drm_mm_node, rb); | |
209 | if (parent->__subtree_last < node->__subtree_last) | |
210 | parent->__subtree_last = node->__subtree_last; | |
211 | if (node->start < parent->start) | |
212 | link = &parent->rb.rb_left; | |
f808c13f | 213 | else { |
202b52b7 | 214 | link = &parent->rb.rb_right; |
f808c13f DB |
215 | leftmost = true; |
216 | } | |
202b52b7 CW |
217 | } |
218 | ||
219 | rb_link_node(&node->rb, rb, link); | |
f808c13f DB |
220 | rb_insert_augmented_cached(&node->rb, &mm->interval_tree, leftmost, |
221 | &drm_mm_interval_tree_augment); | |
202b52b7 CW |
222 | } |
223 | ||
4e64e553 CW |
224 | #define RB_INSERT(root, member, expr) do { \ |
225 | struct rb_node **link = &root.rb_node, *rb = NULL; \ | |
226 | u64 x = expr(node); \ | |
227 | while (*link) { \ | |
228 | rb = *link; \ | |
229 | if (x < expr(rb_entry(rb, struct drm_mm_node, member))) \ | |
230 | link = &rb->rb_left; \ | |
231 | else \ | |
232 | link = &rb->rb_right; \ | |
233 | } \ | |
234 | rb_link_node(&node->member, rb, link); \ | |
235 | rb_insert_color(&node->member, &root); \ | |
236 | } while (0) | |
237 | ||
238 | #define HOLE_SIZE(NODE) ((NODE)->hole_size) | |
239 | #define HOLE_ADDR(NODE) (__drm_mm_hole_node_start(NODE)) | |
240 | ||
241 | static void add_hole(struct drm_mm_node *node) | |
3a1bd924 | 242 | { |
4e64e553 | 243 | struct drm_mm *mm = node->mm; |
ea7b1dd4 | 244 | |
4e64e553 CW |
245 | node->hole_size = |
246 | __drm_mm_hole_node_end(node) - __drm_mm_hole_node_start(node); | |
247 | DRM_MM_BUG_ON(!drm_mm_hole_follows(node)); | |
b0b7af18 | 248 | |
4e64e553 CW |
249 | RB_INSERT(mm->holes_size, rb_hole_size, HOLE_SIZE); |
250 | RB_INSERT(mm->holes_addr, rb_hole_addr, HOLE_ADDR); | |
1d58420b | 251 | |
4e64e553 CW |
252 | list_add(&node->hole_stack, &mm->hole_stack); |
253 | } | |
adb040b8 | 254 | |
4e64e553 CW |
255 | static void rm_hole(struct drm_mm_node *node) |
256 | { | |
257 | DRM_MM_BUG_ON(!drm_mm_hole_follows(node)); | |
62347f9e | 258 | |
4e64e553 CW |
259 | list_del(&node->hole_stack); |
260 | rb_erase(&node->rb_hole_size, &node->mm->holes_size); | |
261 | rb_erase(&node->rb_hole_addr, &node->mm->holes_addr); | |
262 | node->hole_size = 0; | |
440fd528 | 263 | |
4e64e553 CW |
264 | DRM_MM_BUG_ON(drm_mm_hole_follows(node)); |
265 | } | |
266 | ||
267 | static inline struct drm_mm_node *rb_hole_size_to_node(struct rb_node *rb) | |
268 | { | |
269 | return rb_entry_safe(rb, struct drm_mm_node, rb_hole_size); | |
270 | } | |
271 | ||
272 | static inline struct drm_mm_node *rb_hole_addr_to_node(struct rb_node *rb) | |
273 | { | |
274 | return rb_entry_safe(rb, struct drm_mm_node, rb_hole_addr); | |
275 | } | |
276 | ||
277 | static inline u64 rb_hole_size(struct rb_node *rb) | |
278 | { | |
279 | return rb_entry(rb, struct drm_mm_node, rb_hole_size)->hole_size; | |
280 | } | |
281 | ||
282 | static struct drm_mm_node *best_hole(struct drm_mm *mm, u64 size) | |
283 | { | |
284 | struct rb_node *best = NULL; | |
285 | struct rb_node **link = &mm->holes_size.rb_node; | |
286 | ||
287 | while (*link) { | |
288 | struct rb_node *rb = *link; | |
289 | ||
290 | if (size <= rb_hole_size(rb)) { | |
291 | link = &rb->rb_left; | |
292 | best = rb; | |
293 | } else { | |
294 | link = &rb->rb_right; | |
62347f9e | 295 | } |
6b9d89b4 CW |
296 | } |
297 | ||
4e64e553 CW |
298 | return rb_hole_size_to_node(best); |
299 | } | |
300 | ||
301 | static struct drm_mm_node *find_hole(struct drm_mm *mm, u64 addr) | |
302 | { | |
303 | struct drm_mm_node *node = NULL; | |
304 | struct rb_node **link = &mm->holes_addr.rb_node; | |
305 | ||
306 | while (*link) { | |
307 | u64 hole_start; | |
308 | ||
309 | node = rb_hole_addr_to_node(*link); | |
310 | hole_start = __drm_mm_hole_node_start(node); | |
311 | ||
312 | if (addr < hole_start) | |
313 | link = &node->rb_hole_addr.rb_left; | |
314 | else if (addr > hole_start + node->hole_size) | |
315 | link = &node->rb_hole_addr.rb_right; | |
316 | else | |
317 | break; | |
6b9d89b4 | 318 | } |
ea7b1dd4 | 319 | |
4e64e553 CW |
320 | return node; |
321 | } | |
3a1bd924 | 322 | |
4e64e553 CW |
323 | static struct drm_mm_node * |
324 | first_hole(struct drm_mm *mm, | |
325 | u64 start, u64 end, u64 size, | |
326 | enum drm_mm_insert_mode mode) | |
327 | { | |
328 | if (RB_EMPTY_ROOT(&mm->holes_size)) | |
329 | return NULL; | |
ea7b1dd4 | 330 | |
4e64e553 CW |
331 | switch (mode) { |
332 | default: | |
333 | case DRM_MM_INSERT_BEST: | |
334 | return best_hole(mm, size); | |
202b52b7 | 335 | |
4e64e553 CW |
336 | case DRM_MM_INSERT_LOW: |
337 | return find_hole(mm, start); | |
ea7b1dd4 | 338 | |
4e64e553 CW |
339 | case DRM_MM_INSERT_HIGH: |
340 | return find_hole(mm, end); | |
341 | ||
342 | case DRM_MM_INSERT_EVICT: | |
343 | return list_first_entry_or_null(&mm->hole_stack, | |
344 | struct drm_mm_node, | |
345 | hole_stack); | |
1d58420b | 346 | } |
4e64e553 | 347 | } |
5705670d | 348 | |
4e64e553 CW |
349 | static struct drm_mm_node * |
350 | next_hole(struct drm_mm *mm, | |
351 | struct drm_mm_node *node, | |
352 | enum drm_mm_insert_mode mode) | |
353 | { | |
354 | switch (mode) { | |
355 | default: | |
356 | case DRM_MM_INSERT_BEST: | |
357 | return rb_hole_size_to_node(rb_next(&node->rb_hole_size)); | |
358 | ||
359 | case DRM_MM_INSERT_LOW: | |
360 | return rb_hole_addr_to_node(rb_next(&node->rb_hole_addr)); | |
361 | ||
362 | case DRM_MM_INSERT_HIGH: | |
363 | return rb_hole_addr_to_node(rb_prev(&node->rb_hole_addr)); | |
364 | ||
365 | case DRM_MM_INSERT_EVICT: | |
366 | node = list_next_entry(node, hole_stack); | |
367 | return &node->hole_stack == &mm->hole_stack ? NULL : node; | |
368 | } | |
9fc935de DV |
369 | } |
370 | ||
e18c0412 DV |
371 | /** |
372 | * drm_mm_reserve_node - insert an pre-initialized node | |
373 | * @mm: drm_mm allocator to insert @node into | |
374 | * @node: drm_mm_node to insert | |
375 | * | |
05fc0321 DV |
376 | * This functions inserts an already set-up &drm_mm_node into the allocator, |
377 | * meaning that start, size and color must be set by the caller. All other | |
378 | * fields must be cleared to 0. This is useful to initialize the allocator with | |
379 | * preallocated objects which must be set-up before the range allocator can be | |
380 | * set-up, e.g. when taking over a firmware framebuffer. | |
e18c0412 DV |
381 | * |
382 | * Returns: | |
383 | * 0 on success, -ENOSPC if there's no hole where @node is. | |
384 | */ | |
338710e7 | 385 | int drm_mm_reserve_node(struct drm_mm *mm, struct drm_mm_node *node) |
5973c7ee | 386 | { |
202b52b7 | 387 | u64 end = node->start + node->size; |
b3a070cc | 388 | struct drm_mm_node *hole; |
202b52b7 | 389 | u64 hole_start, hole_end; |
2db86dfc | 390 | u64 adj_start, adj_end; |
338710e7 | 391 | |
b80d3942 | 392 | end = node->start + node->size; |
c820186d CW |
393 | if (unlikely(end <= node->start)) |
394 | return -ENOSPC; | |
b80d3942 | 395 | |
338710e7 | 396 | /* Find the relevant hole to add our node to */ |
4e64e553 CW |
397 | hole = find_hole(mm, node->start); |
398 | if (!hole) | |
202b52b7 | 399 | return -ENOSPC; |
5973c7ee | 400 | |
2db86dfc | 401 | adj_start = hole_start = __drm_mm_hole_node_start(hole); |
4e64e553 | 402 | adj_end = hole_end = hole_start + hole->hole_size; |
2db86dfc CW |
403 | |
404 | if (mm->color_adjust) | |
405 | mm->color_adjust(hole, node->color, &adj_start, &adj_end); | |
406 | ||
407 | if (adj_start > node->start || adj_end < end) | |
202b52b7 | 408 | return -ENOSPC; |
5973c7ee | 409 | |
202b52b7 | 410 | node->mm = mm; |
5973c7ee | 411 | |
202b52b7 | 412 | list_add(&node->node_list, &hole->node_list); |
202b52b7 | 413 | drm_mm_interval_tree_add_node(hole, node); |
4e64e553 CW |
414 | node->allocated = true; |
415 | node->hole_size = 0; | |
202b52b7 | 416 | |
4e64e553 CW |
417 | rm_hole(hole); |
418 | if (node->start > hole_start) | |
419 | add_hole(hole); | |
420 | if (end < hole_end) | |
421 | add_hole(node); | |
5973c7ee | 422 | |
5705670d | 423 | save_stack(node); |
202b52b7 | 424 | return 0; |
5973c7ee | 425 | } |
338710e7 | 426 | EXPORT_SYMBOL(drm_mm_reserve_node); |
5973c7ee | 427 | |
b0b7af18 | 428 | /** |
4e64e553 | 429 | * drm_mm_insert_node_in_range - ranged search for space and insert @node |
e18c0412 DV |
430 | * @mm: drm_mm to allocate from |
431 | * @node: preallocate node to insert | |
432 | * @size: size of the allocation | |
433 | * @alignment: alignment of the allocation | |
434 | * @color: opaque tag value to use for this node | |
4e64e553 CW |
435 | * @range_start: start of the allowed range for this node |
436 | * @range_end: end of the allowed range for this node | |
437 | * @mode: fine-tune the allocation search and placement | |
e18c0412 | 438 | * |
05fc0321 | 439 | * The preallocated @node must be cleared to 0. |
e18c0412 DV |
440 | * |
441 | * Returns: | |
442 | * 0 on success, -ENOSPC if there's no suitable hole. | |
3a1bd924 | 443 | */ |
4e64e553 CW |
444 | int drm_mm_insert_node_in_range(struct drm_mm * const mm, |
445 | struct drm_mm_node * const node, | |
446 | u64 size, u64 alignment, | |
447 | unsigned long color, | |
448 | u64 range_start, u64 range_end, | |
449 | enum drm_mm_insert_mode mode) | |
3a1bd924 | 450 | { |
4e64e553 CW |
451 | struct drm_mm_node *hole; |
452 | u64 remainder_mask; | |
b0b7af18 | 453 | |
4e64e553 | 454 | DRM_MM_BUG_ON(range_start >= range_end); |
aafdcfd3 | 455 | |
4e64e553 | 456 | if (unlikely(size == 0 || range_end - range_start < size)) |
b0b7af18 DV |
457 | return -ENOSPC; |
458 | ||
4e64e553 CW |
459 | if (alignment <= 1) |
460 | alignment = 0; | |
461 | ||
462 | remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0; | |
463 | for (hole = first_hole(mm, range_start, range_end, size, mode); hole; | |
464 | hole = next_hole(mm, hole, mode)) { | |
465 | u64 hole_start = __drm_mm_hole_node_start(hole); | |
466 | u64 hole_end = hole_start + hole->hole_size; | |
467 | u64 adj_start, adj_end; | |
468 | u64 col_start, col_end; | |
469 | ||
470 | if (mode == DRM_MM_INSERT_LOW && hole_start >= range_end) | |
471 | break; | |
472 | ||
473 | if (mode == DRM_MM_INSERT_HIGH && hole_end <= range_start) | |
474 | break; | |
475 | ||
476 | col_start = hole_start; | |
477 | col_end = hole_end; | |
478 | if (mm->color_adjust) | |
479 | mm->color_adjust(hole, color, &col_start, &col_end); | |
480 | ||
481 | adj_start = max(col_start, range_start); | |
482 | adj_end = min(col_end, range_end); | |
483 | ||
484 | if (adj_end <= adj_start || adj_end - adj_start < size) | |
485 | continue; | |
486 | ||
487 | if (mode == DRM_MM_INSERT_HIGH) | |
488 | adj_start = adj_end - size; | |
489 | ||
490 | if (alignment) { | |
491 | u64 rem; | |
492 | ||
493 | if (likely(remainder_mask)) | |
494 | rem = adj_start & remainder_mask; | |
495 | else | |
496 | div64_u64_rem(adj_start, alignment, &rem); | |
497 | if (rem) { | |
498 | adj_start -= rem; | |
499 | if (mode != DRM_MM_INSERT_HIGH) | |
500 | adj_start += alignment; | |
501 | ||
502 | if (adj_start < max(col_start, range_start) || | |
503 | min(col_end, range_end) - adj_start < size) | |
504 | continue; | |
505 | ||
506 | if (adj_end <= adj_start || | |
507 | adj_end - adj_start < size) | |
508 | continue; | |
509 | } | |
510 | } | |
511 | ||
512 | node->mm = mm; | |
513 | node->size = size; | |
514 | node->start = adj_start; | |
515 | node->color = color; | |
516 | node->hole_size = 0; | |
517 | ||
518 | list_add(&node->node_list, &hole->node_list); | |
519 | drm_mm_interval_tree_add_node(hole, node); | |
520 | node->allocated = true; | |
521 | ||
522 | rm_hole(hole); | |
523 | if (adj_start > hole_start) | |
524 | add_hole(hole); | |
525 | if (adj_start + size < hole_end) | |
526 | add_hole(node); | |
527 | ||
528 | save_stack(node); | |
529 | return 0; | |
530 | } | |
531 | ||
532 | return -ENOSPC; | |
b0b7af18 | 533 | } |
4e64e553 | 534 | EXPORT_SYMBOL(drm_mm_insert_node_in_range); |
b8103450 | 535 | |
b0b7af18 | 536 | /** |
e18c0412 DV |
537 | * drm_mm_remove_node - Remove a memory node from the allocator. |
538 | * @node: drm_mm_node to remove | |
539 | * | |
540 | * This just removes a node from its drm_mm allocator. The node does not need to | |
541 | * be cleared again before it can be re-inserted into this or any other drm_mm | |
ba004e39 | 542 | * allocator. It is a bug to call this function on a unallocated node. |
b0b7af18 DV |
543 | */ |
544 | void drm_mm_remove_node(struct drm_mm_node *node) | |
545 | { | |
ea7b1dd4 DV |
546 | struct drm_mm *mm = node->mm; |
547 | struct drm_mm_node *prev_node; | |
3a1bd924 | 548 | |
b3ee963f | 549 | DRM_MM_BUG_ON(!node->allocated); |
f29051f1 | 550 | DRM_MM_BUG_ON(node->scanned_block); |
3a1bd924 | 551 | |
4e64e553 | 552 | prev_node = list_prev_entry(node, node_list); |
9e8944ab | 553 | |
4e64e553 CW |
554 | if (drm_mm_hole_follows(node)) |
555 | rm_hole(node); | |
ea7b1dd4 | 556 | |
202b52b7 | 557 | drm_mm_interval_tree_remove(node, &mm->interval_tree); |
ea7b1dd4 | 558 | list_del(&node->node_list); |
4e64e553 | 559 | node->allocated = false; |
7a6b2896 | 560 | |
4e64e553 CW |
561 | if (drm_mm_hole_follows(prev_node)) |
562 | rm_hole(prev_node); | |
563 | add_hole(prev_node); | |
a2e68e92 | 564 | } |
4e64e553 | 565 | EXPORT_SYMBOL(drm_mm_remove_node); |
a2e68e92 | 566 | |
b0b7af18 | 567 | /** |
e18c0412 DV |
568 | * drm_mm_replace_node - move an allocation from @old to @new |
569 | * @old: drm_mm_node to remove from the allocator | |
570 | * @new: drm_mm_node which should inherit @old's allocation | |
571 | * | |
572 | * This is useful for when drivers embed the drm_mm_node structure and hence | |
573 | * can't move allocations by reassigning pointers. It's a combination of remove | |
574 | * and insert with the guarantee that the allocation start will match. | |
b0b7af18 DV |
575 | */ |
576 | void drm_mm_replace_node(struct drm_mm_node *old, struct drm_mm_node *new) | |
577 | { | |
338f1d9d CW |
578 | struct drm_mm *mm = old->mm; |
579 | ||
b3ee963f CW |
580 | DRM_MM_BUG_ON(!old->allocated); |
581 | ||
4e64e553 CW |
582 | *new = *old; |
583 | ||
b0b7af18 | 584 | list_replace(&old->node_list, &new->node_list); |
338f1d9d | 585 | rb_replace_node_cached(&old->rb, &new->rb, &mm->interval_tree); |
4e64e553 CW |
586 | |
587 | if (drm_mm_hole_follows(old)) { | |
588 | list_replace(&old->hole_stack, &new->hole_stack); | |
589 | rb_replace_node(&old->rb_hole_size, | |
590 | &new->rb_hole_size, | |
338f1d9d | 591 | &mm->holes_size); |
4e64e553 CW |
592 | rb_replace_node(&old->rb_hole_addr, |
593 | &new->rb_hole_addr, | |
338f1d9d | 594 | &mm->holes_addr); |
4e64e553 CW |
595 | } |
596 | ||
597 | old->allocated = false; | |
598 | new->allocated = true; | |
b0b7af18 DV |
599 | } |
600 | EXPORT_SYMBOL(drm_mm_replace_node); | |
601 | ||
93110be6 | 602 | /** |
05fc0321 | 603 | * DOC: lru scan roster |
93110be6 DV |
604 | * |
605 | * Very often GPUs need to have continuous allocations for a given object. When | |
606 | * evicting objects to make space for a new one it is therefore not most | |
607 | * efficient when we simply start to select all objects from the tail of an LRU | |
608 | * until there's a suitable hole: Especially for big objects or nodes that | |
609 | * otherwise have special allocation constraints there's a good chance we evict | |
ba004e39 | 610 | * lots of (smaller) objects unnecessarily. |
93110be6 DV |
611 | * |
612 | * The DRM range allocator supports this use-case through the scanning | |
613 | * interfaces. First a scan operation needs to be initialized with | |
9a71e277 | 614 | * drm_mm_scan_init() or drm_mm_scan_init_with_range(). The driver adds |
05fc0321 DV |
615 | * objects to the roster, probably by walking an LRU list, but this can be |
616 | * freely implemented. Eviction candiates are added using | |
617 | * drm_mm_scan_add_block() until a suitable hole is found or there are no | |
940eba2d DV |
618 | * further evictable objects. Eviction roster metadata is tracked in &struct |
619 | * drm_mm_scan. | |
93110be6 | 620 | * |
ba004e39 | 621 | * The driver must walk through all objects again in exactly the reverse |
93110be6 DV |
622 | * order to restore the allocator state. Note that while the allocator is used |
623 | * in the scan mode no other operation is allowed. | |
624 | * | |
3fa489da CW |
625 | * Finally the driver evicts all objects selected (drm_mm_scan_remove_block() |
626 | * reported true) in the scan, and any overlapping nodes after color adjustment | |
05fc0321 | 627 | * (drm_mm_scan_color_evict()). Adding and removing an object is O(1), and |
3fa489da CW |
628 | * since freeing a node is also O(1) the overall complexity is |
629 | * O(scanned_objects). So like the free stack which needs to be walked before a | |
630 | * scan operation even begins this is linear in the number of objects. It | |
631 | * doesn't seem to hurt too badly. | |
93110be6 DV |
632 | */ |
633 | ||
d935cc61 | 634 | /** |
9a71e277 CW |
635 | * drm_mm_scan_init_with_range - initialize range-restricted lru scanning |
636 | * @scan: scan state | |
e18c0412 DV |
637 | * @mm: drm_mm to scan |
638 | * @size: size of the allocation | |
639 | * @alignment: alignment of the allocation | |
640 | * @color: opaque tag value to use for the allocation | |
641 | * @start: start of the allowed range for the allocation | |
642 | * @end: end of the allowed range for the allocation | |
4e64e553 | 643 | * @mode: fine-tune the allocation search and placement |
d935cc61 DV |
644 | * |
645 | * This simply sets up the scanning routines with the parameters for the desired | |
0b04d474 | 646 | * hole. |
d935cc61 | 647 | * |
e18c0412 DV |
648 | * Warning: |
649 | * As long as the scan list is non-empty, no other operations than | |
d935cc61 DV |
650 | * adding/removing nodes to/from the scan list are allowed. |
651 | */ | |
9a71e277 CW |
652 | void drm_mm_scan_init_with_range(struct drm_mm_scan *scan, |
653 | struct drm_mm *mm, | |
440fd528 | 654 | u64 size, |
71733207 | 655 | u64 alignment, |
6b9d89b4 | 656 | unsigned long color, |
440fd528 | 657 | u64 start, |
0b04d474 | 658 | u64 end, |
4e64e553 | 659 | enum drm_mm_insert_mode mode) |
d935cc61 | 660 | { |
6259a56b CW |
661 | DRM_MM_BUG_ON(start >= end); |
662 | DRM_MM_BUG_ON(!size || size > end - start); | |
9a71e277 CW |
663 | DRM_MM_BUG_ON(mm->scan_active); |
664 | ||
665 | scan->mm = mm; | |
666 | ||
9a956b15 CW |
667 | if (alignment <= 1) |
668 | alignment = 0; | |
669 | ||
9a71e277 CW |
670 | scan->color = color; |
671 | scan->alignment = alignment; | |
9a956b15 | 672 | scan->remainder_mask = is_power_of_2(alignment) ? alignment - 1 : 0; |
9a71e277 | 673 | scan->size = size; |
4e64e553 | 674 | scan->mode = mode; |
9a71e277 CW |
675 | |
676 | DRM_MM_BUG_ON(end <= start); | |
677 | scan->range_start = start; | |
678 | scan->range_end = end; | |
6259a56b | 679 | |
9a71e277 CW |
680 | scan->hit_start = U64_MAX; |
681 | scan->hit_end = 0; | |
d935cc61 | 682 | } |
9a71e277 | 683 | EXPORT_SYMBOL(drm_mm_scan_init_with_range); |
d935cc61 | 684 | |
709ea971 | 685 | /** |
e18c0412 | 686 | * drm_mm_scan_add_block - add a node to the scan list |
9b8b75de | 687 | * @scan: the active drm_mm scanner |
e18c0412 DV |
688 | * @node: drm_mm_node to add |
689 | * | |
709ea971 DV |
690 | * Add a node to the scan list that might be freed to make space for the desired |
691 | * hole. | |
692 | * | |
e18c0412 DV |
693 | * Returns: |
694 | * True if a hole has been found, false otherwise. | |
709ea971 | 695 | */ |
9a71e277 CW |
696 | bool drm_mm_scan_add_block(struct drm_mm_scan *scan, |
697 | struct drm_mm_node *node) | |
709ea971 | 698 | { |
9a71e277 | 699 | struct drm_mm *mm = scan->mm; |
4a6c156f | 700 | struct drm_mm_node *hole; |
440fd528 | 701 | u64 hole_start, hole_end; |
268c6498 | 702 | u64 col_start, col_end; |
440fd528 | 703 | u64 adj_start, adj_end; |
709ea971 | 704 | |
9a71e277 CW |
705 | DRM_MM_BUG_ON(node->mm != mm); |
706 | DRM_MM_BUG_ON(!node->allocated); | |
b3ee963f | 707 | DRM_MM_BUG_ON(node->scanned_block); |
0b04d474 | 708 | node->scanned_block = true; |
9a71e277 | 709 | mm->scan_active++; |
709ea971 | 710 | |
f29051f1 CW |
711 | /* Remove this block from the node_list so that we enlarge the hole |
712 | * (distance between the end of our previous node and the start of | |
713 | * or next), without poisoning the link so that we can restore it | |
714 | * later in drm_mm_scan_remove_block(). | |
715 | */ | |
4a6c156f | 716 | hole = list_prev_entry(node, node_list); |
f29051f1 CW |
717 | DRM_MM_BUG_ON(list_next_entry(hole, node_list) != node); |
718 | __list_del_entry(&node->node_list); | |
709ea971 | 719 | |
268c6498 CW |
720 | hole_start = __drm_mm_hole_node_start(hole); |
721 | hole_end = __drm_mm_hole_node_end(hole); | |
d935cc61 | 722 | |
268c6498 CW |
723 | col_start = hole_start; |
724 | col_end = hole_end; | |
901593f2 | 725 | if (mm->color_adjust) |
268c6498 CW |
726 | mm->color_adjust(hole, scan->color, &col_start, &col_end); |
727 | ||
728 | adj_start = max(col_start, scan->range_start); | |
729 | adj_end = min(col_end, scan->range_end); | |
0b04d474 CW |
730 | if (adj_end <= adj_start || adj_end - adj_start < scan->size) |
731 | return false; | |
732 | ||
4e64e553 | 733 | if (scan->mode == DRM_MM_INSERT_HIGH) |
0b04d474 CW |
734 | adj_start = adj_end - scan->size; |
735 | ||
736 | if (scan->alignment) { | |
737 | u64 rem; | |
738 | ||
9a956b15 CW |
739 | if (likely(scan->remainder_mask)) |
740 | rem = adj_start & scan->remainder_mask; | |
741 | else | |
742 | div64_u64_rem(adj_start, scan->alignment, &rem); | |
0b04d474 CW |
743 | if (rem) { |
744 | adj_start -= rem; | |
4e64e553 | 745 | if (scan->mode != DRM_MM_INSERT_HIGH) |
0b04d474 CW |
746 | adj_start += scan->alignment; |
747 | if (adj_start < max(col_start, scan->range_start) || | |
748 | min(col_end, scan->range_end) - adj_start < scan->size) | |
749 | return false; | |
750 | ||
751 | if (adj_end <= adj_start || | |
752 | adj_end - adj_start < scan->size) | |
753 | return false; | |
754 | } | |
755 | } | |
901593f2 | 756 | |
3fa489da CW |
757 | scan->hit_start = adj_start; |
758 | scan->hit_end = adj_start + scan->size; | |
709ea971 | 759 | |
0b04d474 CW |
760 | DRM_MM_BUG_ON(scan->hit_start >= scan->hit_end); |
761 | DRM_MM_BUG_ON(scan->hit_start < hole_start); | |
762 | DRM_MM_BUG_ON(scan->hit_end > hole_end); | |
763 | ||
764 | return true; | |
709ea971 DV |
765 | } |
766 | EXPORT_SYMBOL(drm_mm_scan_add_block); | |
767 | ||
768 | /** | |
e18c0412 | 769 | * drm_mm_scan_remove_block - remove a node from the scan list |
9b8b75de | 770 | * @scan: the active drm_mm scanner |
e18c0412 | 771 | * @node: drm_mm_node to remove |
709ea971 | 772 | * |
05fc0321 DV |
773 | * Nodes **must** be removed in exactly the reverse order from the scan list as |
774 | * they have been added (e.g. using list_add() as they are added and then | |
775 | * list_for_each() over that eviction list to remove), otherwise the internal | |
ba004e39 | 776 | * state of the memory manager will be corrupted. |
709ea971 DV |
777 | * |
778 | * When the scan list is empty, the selected memory nodes can be freed. An | |
05fc0321 DV |
779 | * immediately following drm_mm_insert_node_in_range_generic() or one of the |
780 | * simpler versions of that function with !DRM_MM_SEARCH_BEST will then return | |
781 | * the just freed block (because its at the top of the free_stack list). | |
709ea971 | 782 | * |
e18c0412 DV |
783 | * Returns: |
784 | * True if this block should be evicted, false otherwise. Will always | |
785 | * return false when no hole has been found. | |
709ea971 | 786 | */ |
9a71e277 CW |
787 | bool drm_mm_scan_remove_block(struct drm_mm_scan *scan, |
788 | struct drm_mm_node *node) | |
709ea971 | 789 | { |
ea7b1dd4 | 790 | struct drm_mm_node *prev_node; |
709ea971 | 791 | |
9a71e277 | 792 | DRM_MM_BUG_ON(node->mm != scan->mm); |
b3ee963f | 793 | DRM_MM_BUG_ON(!node->scanned_block); |
0b04d474 | 794 | node->scanned_block = false; |
709ea971 | 795 | |
9a71e277 CW |
796 | DRM_MM_BUG_ON(!node->mm->scan_active); |
797 | node->mm->scan_active--; | |
798 | ||
f29051f1 CW |
799 | /* During drm_mm_scan_add_block() we decoupled this node leaving |
800 | * its pointers intact. Now that the caller is walking back along | |
801 | * the eviction list we can restore this block into its rightful | |
802 | * place on the full node_list. To confirm that the caller is walking | |
803 | * backwards correctly we check that prev_node->next == node->next, | |
804 | * i.e. both believe the same node should be on the other side of the | |
805 | * hole. | |
806 | */ | |
9a71e277 | 807 | prev_node = list_prev_entry(node, node_list); |
f29051f1 CW |
808 | DRM_MM_BUG_ON(list_next_entry(prev_node, node_list) != |
809 | list_next_entry(node, node_list)); | |
ea7b1dd4 | 810 | list_add(&node->node_list, &prev_node->node_list); |
709ea971 | 811 | |
0b04d474 | 812 | return (node->start + node->size > scan->hit_start && |
9a71e277 | 813 | node->start < scan->hit_end); |
709ea971 DV |
814 | } |
815 | EXPORT_SYMBOL(drm_mm_scan_remove_block); | |
816 | ||
3fa489da CW |
817 | /** |
818 | * drm_mm_scan_color_evict - evict overlapping nodes on either side of hole | |
819 | * @scan: drm_mm scan with target hole | |
820 | * | |
821 | * After completing an eviction scan and removing the selected nodes, we may | |
822 | * need to remove a few more nodes from either side of the target hole if | |
823 | * mm.color_adjust is being used. | |
824 | * | |
825 | * Returns: | |
826 | * A node to evict, or NULL if there are no overlapping nodes. | |
827 | */ | |
828 | struct drm_mm_node *drm_mm_scan_color_evict(struct drm_mm_scan *scan) | |
829 | { | |
830 | struct drm_mm *mm = scan->mm; | |
831 | struct drm_mm_node *hole; | |
832 | u64 hole_start, hole_end; | |
833 | ||
834 | DRM_MM_BUG_ON(list_empty(&mm->hole_stack)); | |
835 | ||
836 | if (!mm->color_adjust) | |
837 | return NULL; | |
838 | ||
5f0b3b50 CW |
839 | /* |
840 | * The hole found during scanning should ideally be the first element | |
841 | * in the hole_stack list, but due to side-effects in the driver it | |
842 | * may not be. | |
843 | */ | |
844 | list_for_each_entry(hole, &mm->hole_stack, hole_stack) { | |
845 | hole_start = __drm_mm_hole_node_start(hole); | |
846 | hole_end = hole_start + hole->hole_size; | |
847 | ||
848 | if (hole_start <= scan->hit_start && | |
849 | hole_end >= scan->hit_end) | |
850 | break; | |
851 | } | |
852 | ||
853 | /* We should only be called after we found the hole previously */ | |
854 | DRM_MM_BUG_ON(&hole->hole_stack == &mm->hole_stack); | |
855 | if (unlikely(&hole->hole_stack == &mm->hole_stack)) | |
856 | return NULL; | |
3fa489da CW |
857 | |
858 | DRM_MM_BUG_ON(hole_start > scan->hit_start); | |
859 | DRM_MM_BUG_ON(hole_end < scan->hit_end); | |
860 | ||
861 | mm->color_adjust(hole, scan->color, &hole_start, &hole_end); | |
862 | if (hole_start > scan->hit_start) | |
863 | return hole; | |
864 | if (hole_end < scan->hit_end) | |
865 | return list_next_entry(hole, node_list); | |
866 | ||
867 | return NULL; | |
868 | } | |
869 | EXPORT_SYMBOL(drm_mm_scan_color_evict); | |
870 | ||
e18c0412 DV |
871 | /** |
872 | * drm_mm_init - initialize a drm-mm allocator | |
873 | * @mm: the drm_mm structure to initialize | |
874 | * @start: start of the range managed by @mm | |
875 | * @size: end of the range managed by @mm | |
876 | * | |
877 | * Note that @mm must be cleared to 0 before calling this function. | |
878 | */ | |
45b186f1 | 879 | void drm_mm_init(struct drm_mm *mm, u64 start, u64 size) |
1d58420b | 880 | { |
6259a56b CW |
881 | DRM_MM_BUG_ON(start + size <= start); |
882 | ||
4e64e553 CW |
883 | mm->color_adjust = NULL; |
884 | ||
ea7b1dd4 | 885 | INIT_LIST_HEAD(&mm->hole_stack); |
f808c13f | 886 | mm->interval_tree = RB_ROOT_CACHED; |
4e64e553 CW |
887 | mm->holes_size = RB_ROOT; |
888 | mm->holes_addr = RB_ROOT; | |
3a1bd924 | 889 | |
ea7b1dd4 DV |
890 | /* Clever trick to avoid a special case in the free hole tracking. */ |
891 | INIT_LIST_HEAD(&mm->head_node.node_list); | |
4e64e553 | 892 | mm->head_node.allocated = false; |
ea7b1dd4 DV |
893 | mm->head_node.mm = mm; |
894 | mm->head_node.start = start + size; | |
4e64e553 CW |
895 | mm->head_node.size = -size; |
896 | add_hole(&mm->head_node); | |
ea7b1dd4 | 897 | |
4e64e553 | 898 | mm->scan_active = 0; |
3a1bd924 | 899 | } |
673a394b | 900 | EXPORT_SYMBOL(drm_mm_init); |
3a1bd924 | 901 | |
e18c0412 DV |
902 | /** |
903 | * drm_mm_takedown - clean up a drm_mm allocator | |
904 | * @mm: drm_mm allocator to clean up | |
905 | * | |
906 | * Note that it is a bug to call this function on an allocator which is not | |
907 | * clean. | |
908 | */ | |
5705670d | 909 | void drm_mm_takedown(struct drm_mm *mm) |
3a1bd924 | 910 | { |
ac9bb7b7 | 911 | if (WARN(!drm_mm_clean(mm), |
5705670d CW |
912 | "Memory manager not clean during takedown.\n")) |
913 | show_leaks(mm); | |
3a1bd924 | 914 | } |
f453ba04 | 915 | EXPORT_SYMBOL(drm_mm_takedown); |
fa8a1238 | 916 | |
b5c3714f | 917 | static u64 drm_mm_dump_hole(struct drm_printer *p, const struct drm_mm_node *entry) |
99d7e48e | 918 | { |
4e64e553 CW |
919 | u64 start, size; |
920 | ||
921 | size = entry->hole_size; | |
922 | if (size) { | |
923 | start = drm_mm_hole_node_start(entry); | |
924 | drm_printf(p, "%#018llx-%#018llx: %llu: free\n", | |
925 | start, start + size, size); | |
3a359f0b DV |
926 | } |
927 | ||
4e64e553 | 928 | return size; |
3a359f0b | 929 | } |
e18c0412 | 930 | /** |
b5c3714f DV |
931 | * drm_mm_print - print allocator state |
932 | * @mm: drm_mm allocator to print | |
933 | * @p: DRM printer to use | |
e18c0412 | 934 | */ |
b5c3714f | 935 | void drm_mm_print(const struct drm_mm *mm, struct drm_printer *p) |
3a359f0b | 936 | { |
45b186f1 | 937 | const struct drm_mm_node *entry; |
440fd528 | 938 | u64 total_used = 0, total_free = 0, total = 0; |
3a359f0b | 939 | |
b5c3714f | 940 | total_free += drm_mm_dump_hole(p, &mm->head_node); |
ea7b1dd4 DV |
941 | |
942 | drm_mm_for_each_node(entry, mm) { | |
b5c3714f | 943 | drm_printf(p, "%#018llx-%#018llx: %llu: used\n", entry->start, |
440fd528 | 944 | entry->start + entry->size, entry->size); |
ea7b1dd4 | 945 | total_used += entry->size; |
b5c3714f | 946 | total_free += drm_mm_dump_hole(p, entry); |
fa8a1238 | 947 | } |
ea7b1dd4 DV |
948 | total = total_free + total_used; |
949 | ||
b5c3714f | 950 | drm_printf(p, "total: %llu, used %llu free %llu\n", total, |
440fd528 | 951 | total_used, total_free); |
fa8a1238 | 952 | } |
b5c3714f | 953 | EXPORT_SYMBOL(drm_mm_print); |