1 /* SPDX-License-Identifier: GPL-2.0-only */
2 /*
3 * Fast and scalable bitmaps.
4 *
5 * Copyright (C) 2016 Facebook
6 * Copyright (C) 2013-2014 Jens Axboe
7 */
8
9 #ifndef __LINUX_SCALE_BITMAP_H
10 #define __LINUX_SCALE_BITMAP_H
11
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14
15 struct seq_file;
16
17 /**
18 * struct sbitmap_word - Word in a &struct sbitmap.
19 */
20 struct sbitmap_word {
21 /**
22 * @depth: Number of bits being used in @word/@cleared
23 */
24 unsigned long depth;
25
26 /**
27 * @word: word holding free bits
28 */
29 unsigned long word ____cacheline_aligned_in_smp;
30
31 /**
32 * @cleared: word holding cleared bits
33 */
34 unsigned long cleared ____cacheline_aligned_in_smp;
35
36 /**
37 * @swap_lock: Held while swapping word <-> cleared
38 */
39 spinlock_t swap_lock;
40 } ____cacheline_aligned_in_smp;
41
42 /**
43 * struct sbitmap - Scalable bitmap.
44 *
45 * A &struct sbitmap is spread over multiple cachelines to avoid ping-pong. This
46 * trades off higher memory usage for better scalability.
47 */
48 struct sbitmap {
49 /**
50 * @depth: Number of bits used in the whole bitmap.
51 */
52 unsigned int depth;
53
54 /**
55 * @shift: log2(number of bits used per word)
56 */
57 unsigned int shift;
58
59 /**
60 * @map_nr: Number of words (cachelines) being used for the bitmap.
61 */
62 unsigned int map_nr;
63
64 /**
65 * @map: Allocated bitmap.
66 */
67 struct sbitmap_word *map;
68 };
69
70 #define SBQ_WAIT_QUEUES 8
71 #define SBQ_WAKE_BATCH 8
72
73 /**
74 * struct sbq_wait_state - Wait queue in a &struct sbitmap_queue.
75 */
76 struct sbq_wait_state {
77 /**
78 * @wait_cnt: Number of frees remaining before we wake up.
79 */
80 atomic_t wait_cnt;
81
82 /**
83 * @wait: Wait queue.
84 */
85 wait_queue_head_t wait;
86 } ____cacheline_aligned_in_smp;
87
88 /**
89 * struct sbitmap_queue - Scalable bitmap with the added ability to wait on free
90 * bits.
91 *
92 * A &struct sbitmap_queue uses multiple wait queues and rolling wakeups to
93 * avoid contention on the wait queue spinlock. This ensures that we don't hit a
94 * scalability wall when we run out of free bits and have to start putting tasks
95 * to sleep.
96 */
97 struct sbitmap_queue {
98 /**
99 * @sb: Scalable bitmap.
100 */
101 struct sbitmap sb;
102
103 /*
104 * @alloc_hint: Cache of last successfully allocated or freed bit.
105 *
106 * This is per-cpu, which allows multiple users to stick to different
107 * cachelines until the map is exhausted.
108 */
109 unsigned int __percpu *alloc_hint;
110
111 /**
112 * @wake_batch: Number of bits which must be freed before we wake up any
113 * waiters.
114 */
115 unsigned int wake_batch;
116
117 /**
118 * @wake_index: Next wait queue in @ws to wake up.
119 */
120 atomic_t wake_index;
121
122 /**
123 * @ws: Wait queues.
124 */
125 struct sbq_wait_state *ws;
126
127 /*
128 * @ws_active: count of currently active ws waitqueues
129 */
130 atomic_t ws_active;
131
132 /**
133 * @round_robin: Allocate bits in strict round-robin order.
134 */
135 bool round_robin;
136
137 /**
138 * @min_shallow_depth: The minimum shallow depth which may be passed to
139 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow().
140 */
141 unsigned int min_shallow_depth;
142 };
143
144 /**
145 * sbitmap_init_node() - Initialize a &struct sbitmap on a specific memory node.
146 * @sb: Bitmap to initialize.
147 * @depth: Number of bits to allocate.
148 * @shift: Use 2^@shift bits per word in the bitmap; if a negative number if
149 * given, a good default is chosen.
150 * @flags: Allocation flags.
151 * @node: Memory node to allocate on.
152 *
153 * Return: Zero on success or negative errno on failure.
154 */
155 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
156 gfp_t flags, int node);
157
158 /**
159 * sbitmap_free() - Free memory used by a &struct sbitmap.
160 * @sb: Bitmap to free.
161 */
162 static inline void sbitmap_free(struct sbitmap *sb)
163 {
164 kfree(sb->map);
165 sb->map = NULL;
166 }
167
168 /**
169 * sbitmap_resize() - Resize a &struct sbitmap.
170 * @sb: Bitmap to resize.
171 * @depth: New number of bits to resize to.
172 *
173 * Doesn't reallocate anything. It's up to the caller to ensure that the new
174 * depth doesn't exceed the depth that the sb was initialized with.
175 */
176 void sbitmap_resize(struct sbitmap *sb, unsigned int depth);
177
178 /**
179 * sbitmap_get() - Try to allocate a free bit from a &struct sbitmap.
180 * @sb: Bitmap to allocate from.
181 * @alloc_hint: Hint for where to start searching for a free bit.
182 * @round_robin: If true, be stricter about allocation order; always allocate
183 * starting from the last allocated bit. This is less efficient
184 * than the default behavior (false).
185 *
186 * This operation provides acquire barrier semantics if it succeeds.
187 *
188 * Return: Non-negative allocated bit number if successful, -1 otherwise.
189 */
190 int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin);
191
192 /**
193 * sbitmap_get_shallow() - Try to allocate a free bit from a &struct sbitmap,
194 * limiting the depth used from each word.
195 * @sb: Bitmap to allocate from.
196 * @alloc_hint: Hint for where to start searching for a free bit.
197 * @shallow_depth: The maximum number of bits to allocate from a single word.
198 *
199 * This rather specific operation allows for having multiple users with
200 * different allocation limits. E.g., there can be a high-priority class that
201 * uses sbitmap_get() and a low-priority class that uses sbitmap_get_shallow()
202 * with a @shallow_depth of (1 << (@sb->shift - 1)). Then, the low-priority
203 * class can only allocate half of the total bits in the bitmap, preventing it
204 * from starving out the high-priority class.
205 *
206 * Return: Non-negative allocated bit number if successful, -1 otherwise.
207 */
208 int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint,
209 unsigned long shallow_depth);
210
211 /**
212 * sbitmap_any_bit_set() - Check for a set bit in a &struct sbitmap.
213 * @sb: Bitmap to check.
214 *
215 * Return: true if any bit in the bitmap is set, false otherwise.
216 */
217 bool sbitmap_any_bit_set(const struct sbitmap *sb);
218
219 /**
220 * sbitmap_any_bit_clear() - Check for an unset bit in a &struct
221 * sbitmap.
222 * @sb: Bitmap to check.
223 *
224 * Return: true if any bit in the bitmap is clear, false otherwise.
225 */
226 bool sbitmap_any_bit_clear(const struct sbitmap *sb);
227
228 #define SB_NR_TO_INDEX(sb, bitnr) ((bitnr) >> (sb)->shift)
229 #define SB_NR_TO_BIT(sb, bitnr) ((bitnr) & ((1U << (sb)->shift) - 1U))
230
231 typedef bool (*sb_for_each_fn)(struct sbitmap *, unsigned int, void *);
232
233 /**
234 * __sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
235 * @start: Where to start the iteration.
236 * @sb: Bitmap to iterate over.
237 * @fn: Callback. Should return true to continue or false to break early.
238 * @data: Pointer to pass to callback.
239 *
240 * This is inline even though it's non-trivial so that the function calls to the
241 * callback will hopefully get optimized away.
242 */
243 static inline void __sbitmap_for_each_set(struct sbitmap *sb,
244 unsigned int start,
245 sb_for_each_fn fn, void *data)
246 {
247 unsigned int index;
248 unsigned int nr;
249 unsigned int scanned = 0;
250
251 if (start >= sb->depth)
252 start = 0;
253 index = SB_NR_TO_INDEX(sb, start);
254 nr = SB_NR_TO_BIT(sb, start);
255
256 while (scanned < sb->depth) {
257 unsigned long word;
258 unsigned int depth = min_t(unsigned int,
259 sb->map[index].depth - nr,
260 sb->depth - scanned);
261
262 scanned += depth;
263 word = sb->map[index].word & ~sb->map[index].cleared;
264 if (!word)
265 goto next;
266
267 /*
268 * On the first iteration of the outer loop, we need to add the
269 * bit offset back to the size of the word for find_next_bit().
270 * On all other iterations, nr is zero, so this is a noop.
271 */
272 depth += nr;
273 while (1) {
274 nr = find_next_bit(&word, depth, nr);
275 if (nr >= depth)
276 break;
277 if (!fn(sb, (index << sb->shift) + nr, data))
278 return;
279
280 nr++;
281 }
282 next:
283 nr = 0;
284 if (++index >= sb->map_nr)
285 index = 0;
286 }
287 }
288
289 /**
290 * sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
291 * @sb: Bitmap to iterate over.
292 * @fn: Callback. Should return true to continue or false to break early.
293 * @data: Pointer to pass to callback.
294 */
295 static inline void sbitmap_for_each_set(struct sbitmap *sb, sb_for_each_fn fn,
296 void *data)
297 {
298 __sbitmap_for_each_set(sb, 0, fn, data);
299 }
300
301 static inline unsigned long *__sbitmap_word(struct sbitmap *sb,
302 unsigned int bitnr)
303 {
304 return &sb->map[SB_NR_TO_INDEX(sb, bitnr)].word;
305 }
306
307 /* Helpers equivalent to the operations in asm/bitops.h and linux/bitmap.h */
308
309 static inline void sbitmap_set_bit(struct sbitmap *sb, unsigned int bitnr)
310 {
311 set_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
312 }
313
314 static inline void sbitmap_clear_bit(struct sbitmap *sb, unsigned int bitnr)
315 {
316 clear_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
317 }
318
319 /*
320 * This one is special, since it doesn't actually clear the bit, rather it
321 * sets the corresponding bit in the ->cleared mask instead. Paired with
322 * the caller doing sbitmap_deferred_clear() if a given index is full, which
323 * will clear the previously freed entries in the corresponding ->word.
324 */
325 static inline void sbitmap_deferred_clear_bit(struct sbitmap *sb, unsigned int bitnr)
326 {
327 unsigned long *addr = &sb->map[SB_NR_TO_INDEX(sb, bitnr)].cleared;
328
329 set_bit(SB_NR_TO_BIT(sb, bitnr), addr);
330 }
331
332 static inline void sbitmap_clear_bit_unlock(struct sbitmap *sb,
333 unsigned int bitnr)
334 {
335 clear_bit_unlock(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
336 }
337
338 static inline int sbitmap_test_bit(struct sbitmap *sb, unsigned int bitnr)
339 {
340 return test_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
341 }
342
343 /**
344 * sbitmap_show() - Dump &struct sbitmap information to a &struct seq_file.
345 * @sb: Bitmap to show.
346 * @m: struct seq_file to write to.
347 *
348 * This is intended for debugging. The format may change at any time.
349 */
350 void sbitmap_show(struct sbitmap *sb, struct seq_file *m);
351
352 /**
353 * sbitmap_bitmap_show() - Write a hex dump of a &struct sbitmap to a &struct
354 * seq_file.
355 * @sb: Bitmap to show.
356 * @m: struct seq_file to write to.
357 *
358 * This is intended for debugging. The output isn't guaranteed to be internally
359 * consistent.
360 */
361 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m);
362
363 /**
364 * sbitmap_queue_init_node() - Initialize a &struct sbitmap_queue on a specific
365 * memory node.
366 * @sbq: Bitmap queue to initialize.
367 * @depth: See sbitmap_init_node().
368 * @shift: See sbitmap_init_node().
369 * @round_robin: See sbitmap_get().
370 * @flags: Allocation flags.
371 * @node: Memory node to allocate on.
372 *
373 * Return: Zero on success or negative errno on failure.
374 */
375 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
376 int shift, bool round_robin, gfp_t flags, int node);
377
378 /**
379 * sbitmap_queue_free() - Free memory used by a &struct sbitmap_queue.
380 *
381 * @sbq: Bitmap queue to free.
382 */
383 static inline void sbitmap_queue_free(struct sbitmap_queue *sbq)
384 {
385 kfree(sbq->ws);
386 free_percpu(sbq->alloc_hint);
387 sbitmap_free(&sbq->sb);
388 }
389
390 /**
391 * sbitmap_queue_resize() - Resize a &struct sbitmap_queue.
392 * @sbq: Bitmap queue to resize.
393 * @depth: New number of bits to resize to.
394 *
395 * Like sbitmap_resize(), this doesn't reallocate anything. It has to do
396 * some extra work on the &struct sbitmap_queue, so it's not safe to just
397 * resize the underlying &struct sbitmap.
398 */
399 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth);
400
401 /**
402 * __sbitmap_queue_get() - Try to allocate a free bit from a &struct
403 * sbitmap_queue with preemption already disabled.
404 * @sbq: Bitmap queue to allocate from.
405 *
406 * Return: Non-negative allocated bit number if successful, -1 otherwise.
407 */
408 int __sbitmap_queue_get(struct sbitmap_queue *sbq);
409
410 /**
411 * __sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
412 * sbitmap_queue, limiting the depth used from each word, with preemption
413 * already disabled.
414 * @sbq: Bitmap queue to allocate from.
415 * @shallow_depth: The maximum number of bits to allocate from a single word.
416 * See sbitmap_get_shallow().
417 *
418 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after
419 * initializing @sbq.
420 *
421 * Return: Non-negative allocated bit number if successful, -1 otherwise.
422 */
423 int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
424 unsigned int shallow_depth);
425
426 /**
427 * sbitmap_queue_get() - Try to allocate a free bit from a &struct
428 * sbitmap_queue.
429 * @sbq: Bitmap queue to allocate from.
430 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to
431 * sbitmap_queue_clear()).
432 *
433 * Return: Non-negative allocated bit number if successful, -1 otherwise.
434 */
435 static inline int sbitmap_queue_get(struct sbitmap_queue *sbq,
436 unsigned int *cpu)
437 {
438 int nr;
439
440 *cpu = get_cpu();
441 nr = __sbitmap_queue_get(sbq);
442 put_cpu();
443 return nr;
444 }
445
446 /**
447 * sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
448 * sbitmap_queue, limiting the depth used from each word.
449 * @sbq: Bitmap queue to allocate from.
450 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to
451 * sbitmap_queue_clear()).
452 * @shallow_depth: The maximum number of bits to allocate from a single word.
453 * See sbitmap_get_shallow().
454 *
455 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after
456 * initializing @sbq.
457 *
458 * Return: Non-negative allocated bit number if successful, -1 otherwise.
459 */
460 static inline int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
461 unsigned int *cpu,
462 unsigned int shallow_depth)
463 {
464 int nr;
465
466 *cpu = get_cpu();
467 nr = __sbitmap_queue_get_shallow(sbq, shallow_depth);
468 put_cpu();
469 return nr;
470 }
471
472 /**
473 * sbitmap_queue_min_shallow_depth() - Inform a &struct sbitmap_queue of the
474 * minimum shallow depth that will be used.
475 * @sbq: Bitmap queue in question.
476 * @min_shallow_depth: The minimum shallow depth that will be passed to
477 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow().
478 *
479 * sbitmap_queue_clear() batches wakeups as an optimization. The batch size
480 * depends on the depth of the bitmap. Since the shallow allocation functions
481 * effectively operate with a different depth, the shallow depth must be taken
482 * into account when calculating the batch size. This function must be called
483 * with the minimum shallow depth that will be used. Failure to do so can result
484 * in missed wakeups.
485 */
486 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
487 unsigned int min_shallow_depth);
488
489 /**
490 * sbitmap_queue_clear() - Free an allocated bit and wake up waiters on a
491 * &struct sbitmap_queue.
492 * @sbq: Bitmap to free from.
493 * @nr: Bit number to free.
494 * @cpu: CPU the bit was allocated on.
495 */
496 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
497 unsigned int cpu);
498
499 static inline int sbq_index_inc(int index)
500 {
501 return (index + 1) & (SBQ_WAIT_QUEUES - 1);
502 }
503
504 static inline void sbq_index_atomic_inc(atomic_t *index)
505 {
506 int old = atomic_read(index);
507 int new = sbq_index_inc(old);
508 atomic_cmpxchg(index, old, new);
509 }
510
511 /**
512 * sbq_wait_ptr() - Get the next wait queue to use for a &struct
513 * sbitmap_queue.
514 * @sbq: Bitmap queue to wait on.
515 * @wait_index: A counter per "user" of @sbq.
516 */
517 static inline struct sbq_wait_state *sbq_wait_ptr(struct sbitmap_queue *sbq,
518 atomic_t *wait_index)
519 {
520 struct sbq_wait_state *ws;
521
522 ws = &sbq->ws[atomic_read(wait_index)];
523 sbq_index_atomic_inc(wait_index);
524 return ws;
525 }
526
527 /**
528 * sbitmap_queue_wake_all() - Wake up everything waiting on a &struct
529 * sbitmap_queue.
530 * @sbq: Bitmap queue to wake up.
531 */
532 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq);
533
534 /**
535 * sbitmap_queue_wake_up() - Wake up some of waiters in one waitqueue
536 * on a &struct sbitmap_queue.
537 * @sbq: Bitmap queue to wake up.
538 */
539 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq);
540
541 /**
542 * sbitmap_queue_show() - Dump &struct sbitmap_queue information to a &struct
543 * seq_file.
544 * @sbq: Bitmap queue to show.
545 * @m: struct seq_file to write to.
546 *
547 * This is intended for debugging. The format may change at any time.
548 */
549 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m);
550
551 struct sbq_wait {
552 struct sbitmap_queue *sbq; /* if set, sbq_wait is accounted */
553 struct wait_queue_entry wait;
554 };
555
556 #define DEFINE_SBQ_WAIT(name) \
557 struct sbq_wait name = { \
558 .sbq = NULL, \
559 .wait = { \
560 .private = current, \
561 .func = autoremove_wake_function, \
562 .entry = LIST_HEAD_INIT((name).wait.entry), \
563 } \
564 }
565
566 /*
567 * Wrapper around prepare_to_wait_exclusive(), which maintains some extra
568 * internal state.
569 */
570 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
571 struct sbq_wait_state *ws,
572 struct sbq_wait *sbq_wait, int state);
573
574 /*
575 * Must be paired with sbitmap_prepare_to_wait().
576 */
577 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
578 struct sbq_wait *sbq_wait);
579
580 /*
581 * Wrapper around add_wait_queue(), which maintains some extra internal state
582 */
583 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
584 struct sbq_wait_state *ws,
585 struct sbq_wait *sbq_wait);
586
587 /*
588 * Must be paired with sbitmap_add_wait_queue()
589 */
590 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait);
591
592 #endif /* __LINUX_SCALE_BITMAP_H */