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/atomic.h>
13#include <linux/bitops.h>
14#include <linux/cache.h>
15#include <linux/list.h>
16#include <linux/log2.h>
17#include <linux/minmax.h>
18#include <linux/percpu.h>
19#include <linux/slab.h>
20#include <linux/smp.h>
21#include <linux/types.h>
22#include <linux/wait.h>
23
24struct seq_file;
25
26/**
27 * struct sbitmap_word - Word in a &struct sbitmap.
28 */
29struct sbitmap_word {
30 /**
31 * @word: word holding free bits
32 */
33 unsigned long word;
34
35 /**
36 * @cleared: word holding cleared bits
37 */
38 unsigned long cleared ____cacheline_aligned_in_smp;
39
40 /**
41 * @swap_lock: serializes simultaneous updates of ->word and ->cleared
42 */
43 raw_spinlock_t swap_lock;
44} ____cacheline_aligned_in_smp;
45
46/**
47 * struct sbitmap - Scalable bitmap.
48 *
49 * A &struct sbitmap is spread over multiple cachelines to avoid ping-pong. This
50 * trades off higher memory usage for better scalability.
51 */
52struct sbitmap {
53 /**
54 * @depth: Number of bits used in the whole bitmap.
55 */
56 unsigned int depth;
57
58 /**
59 * @shift: log2(number of bits used per word)
60 */
61 unsigned int shift;
62
63 /**
64 * @map_nr: Number of words (cachelines) being used for the bitmap.
65 */
66 unsigned int map_nr;
67
68 /**
69 * @round_robin: Allocate bits in strict round-robin order.
70 */
71 bool round_robin;
72
73 /**
74 * @map: Allocated bitmap.
75 */
76 struct sbitmap_word *map;
77
78 /*
79 * @alloc_hint: Cache of last successfully allocated or freed bit.
80 *
81 * This is per-cpu, which allows multiple users to stick to different
82 * cachelines until the map is exhausted.
83 */
84 unsigned int __percpu *alloc_hint;
85};
86
87#define SBQ_WAIT_QUEUES 8
88#define SBQ_WAKE_BATCH 8
89
90/**
91 * struct sbq_wait_state - Wait queue in a &struct sbitmap_queue.
92 */
93struct sbq_wait_state {
94 /**
95 * @wait: Wait queue.
96 */
97 wait_queue_head_t wait;
98} ____cacheline_aligned_in_smp;
99
100/**
101 * struct sbitmap_queue - Scalable bitmap with the added ability to wait on free
102 * bits.
103 *
104 * A &struct sbitmap_queue uses multiple wait queues and rolling wakeups to
105 * avoid contention on the wait queue spinlock. This ensures that we don't hit a
106 * scalability wall when we run out of free bits and have to start putting tasks
107 * to sleep.
108 */
109struct sbitmap_queue {
110 /**
111 * @sb: Scalable bitmap.
112 */
113 struct sbitmap sb;
114
115 /**
116 * @wake_batch: Number of bits which must be freed before we wake up any
117 * waiters.
118 */
119 unsigned int wake_batch;
120
121 /**
122 * @wake_index: Next wait queue in @ws to wake up.
123 */
124 atomic_t wake_index;
125
126 /**
127 * @ws: Wait queues.
128 */
129 struct sbq_wait_state *ws;
130
131 /*
132 * @ws_active: count of currently active ws waitqueues
133 */
134 atomic_t ws_active;
135
136 /**
137 * @min_shallow_depth: The minimum shallow depth which may be passed to
138 * sbitmap_queue_get_shallow()
139 */
140 unsigned int min_shallow_depth;
141
142 /**
143 * @completion_cnt: Number of bits cleared passed to the
144 * wakeup function.
145 */
146 atomic_t completion_cnt;
147
148 /**
149 * @wakeup_cnt: Number of thread wake ups issued.
150 */
151 atomic_t wakeup_cnt;
152};
153
154/**
155 * sbitmap_init_node() - Initialize a &struct sbitmap on a specific memory node.
156 * @sb: Bitmap to initialize.
157 * @depth: Number of bits to allocate.
158 * @shift: Use 2^@shift bits per word in the bitmap; if a negative number if
159 * given, a good default is chosen.
160 * @flags: Allocation flags.
161 * @node: Memory node to allocate on.
162 * @round_robin: If true, be stricter about allocation order; always allocate
163 * starting from the last allocated bit. This is less efficient
164 * than the default behavior (false).
165 * @alloc_hint: If true, apply percpu hint for where to start searching for
166 * a free bit.
167 *
168 * Return: Zero on success or negative errno on failure.
169 */
170int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
171 gfp_t flags, int node, bool round_robin, bool alloc_hint);
172
173/* sbitmap internal helper */
174static inline unsigned int __map_depth(const struct sbitmap *sb, int index)
175{
176 if (index == sb->map_nr - 1)
177 return sb->depth - (index << sb->shift);
178 return 1U << sb->shift;
179}
180
181/**
182 * sbitmap_free() - Free memory used by a &struct sbitmap.
183 * @sb: Bitmap to free.
184 */
185static inline void sbitmap_free(struct sbitmap *sb)
186{
187 free_percpu(pdata: sb->alloc_hint);
188 kvfree(addr: sb->map);
189 sb->map = NULL;
190}
191
192/**
193 * sbitmap_resize() - Resize a &struct sbitmap.
194 * @sb: Bitmap to resize.
195 * @depth: New number of bits to resize to.
196 *
197 * Doesn't reallocate anything. It's up to the caller to ensure that the new
198 * depth doesn't exceed the depth that the sb was initialized with.
199 */
200void sbitmap_resize(struct sbitmap *sb, unsigned int depth);
201
202/**
203 * sbitmap_get() - Try to allocate a free bit from a &struct sbitmap.
204 * @sb: Bitmap to allocate from.
205 *
206 * This operation provides acquire barrier semantics if it succeeds.
207 *
208 * Return: Non-negative allocated bit number if successful, -1 otherwise.
209 */
210int sbitmap_get(struct sbitmap *sb);
211
212/**
213 * sbitmap_any_bit_set() - Check for a set bit in a &struct sbitmap.
214 * @sb: Bitmap to check.
215 *
216 * Return: true if any bit in the bitmap is set, false otherwise.
217 */
218bool sbitmap_any_bit_set(const struct sbitmap *sb);
219
220#define SB_NR_TO_INDEX(sb, bitnr) ((bitnr) >> (sb)->shift)
221#define SB_NR_TO_BIT(sb, bitnr) ((bitnr) & ((1U << (sb)->shift) - 1U))
222
223typedef bool (*sb_for_each_fn)(struct sbitmap *, unsigned int, void *);
224
225/**
226 * __sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
227 * @start: Where to start the iteration.
228 * @sb: Bitmap to iterate over.
229 * @fn: Callback. Should return true to continue or false to break early.
230 * @data: Pointer to pass to callback.
231 *
232 * This is inline even though it's non-trivial so that the function calls to the
233 * callback will hopefully get optimized away.
234 */
235static inline void __sbitmap_for_each_set(struct sbitmap *sb,
236 unsigned int start,
237 sb_for_each_fn fn, void *data)
238{
239 unsigned int index;
240 unsigned int nr;
241 unsigned int scanned = 0;
242
243 if (start >= sb->depth)
244 start = 0;
245 index = SB_NR_TO_INDEX(sb, start);
246 nr = SB_NR_TO_BIT(sb, start);
247
248 while (scanned < sb->depth) {
249 unsigned long word;
250 unsigned int depth = min_t(unsigned int,
251 __map_depth(sb, index) - nr,
252 sb->depth - scanned);
253
254 scanned += depth;
255 word = sb->map[index].word & ~sb->map[index].cleared;
256 if (!word)
257 goto next;
258
259 /*
260 * On the first iteration of the outer loop, we need to add the
261 * bit offset back to the size of the word for find_next_bit().
262 * On all other iterations, nr is zero, so this is a noop.
263 */
264 depth += nr;
265 while (1) {
266 nr = find_next_bit(addr: &word, size: depth, offset: nr);
267 if (nr >= depth)
268 break;
269 if (!fn(sb, (index << sb->shift) + nr, data))
270 return;
271
272 nr++;
273 }
274next:
275 nr = 0;
276 if (++index >= sb->map_nr)
277 index = 0;
278 }
279}
280
281/**
282 * sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
283 * @sb: Bitmap to iterate over.
284 * @fn: Callback. Should return true to continue or false to break early.
285 * @data: Pointer to pass to callback.
286 */
287static inline void sbitmap_for_each_set(struct sbitmap *sb, sb_for_each_fn fn,
288 void *data)
289{
290 __sbitmap_for_each_set(sb, start: 0, fn, data);
291}
292
293static inline unsigned long *__sbitmap_word(struct sbitmap *sb,
294 unsigned int bitnr)
295{
296 return &sb->map[SB_NR_TO_INDEX(sb, bitnr)].word;
297}
298
299/* Helpers equivalent to the operations in asm/bitops.h and linux/bitmap.h */
300
301static inline void sbitmap_set_bit(struct sbitmap *sb, unsigned int bitnr)
302{
303 set_bit(SB_NR_TO_BIT(sb, bitnr), addr: __sbitmap_word(sb, bitnr));
304}
305
306static inline void sbitmap_clear_bit(struct sbitmap *sb, unsigned int bitnr)
307{
308 clear_bit(SB_NR_TO_BIT(sb, bitnr), addr: __sbitmap_word(sb, bitnr));
309}
310
311/*
312 * This one is special, since it doesn't actually clear the bit, rather it
313 * sets the corresponding bit in the ->cleared mask instead. Paired with
314 * the caller doing sbitmap_deferred_clear() if a given index is full, which
315 * will clear the previously freed entries in the corresponding ->word.
316 */
317static inline void sbitmap_deferred_clear_bit(struct sbitmap *sb, unsigned int bitnr)
318{
319 unsigned long *addr = &sb->map[SB_NR_TO_INDEX(sb, bitnr)].cleared;
320
321 set_bit(SB_NR_TO_BIT(sb, bitnr), addr);
322}
323
324/*
325 * Pair of sbitmap_get, and this one applies both cleared bit and
326 * allocation hint.
327 */
328static inline void sbitmap_put(struct sbitmap *sb, unsigned int bitnr)
329{
330 sbitmap_deferred_clear_bit(sb, bitnr);
331
332 if (likely(sb->alloc_hint && !sb->round_robin && bitnr < sb->depth))
333 *raw_cpu_ptr(sb->alloc_hint) = bitnr;
334}
335
336static inline int sbitmap_test_bit(struct sbitmap *sb, unsigned int bitnr)
337{
338 return test_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
339}
340
341static inline int sbitmap_calculate_shift(unsigned int depth)
342{
343 int shift = ilog2(BITS_PER_LONG);
344
345 /*
346 * If the bitmap is small, shrink the number of bits per word so
347 * we spread over a few cachelines, at least. If less than 4
348 * bits, just forget about it, it's not going to work optimally
349 * anyway.
350 */
351 if (depth >= 4) {
352 while ((4U << shift) > depth)
353 shift--;
354 }
355
356 return shift;
357}
358
359/**
360 * sbitmap_show() - Dump &struct sbitmap information to a &struct seq_file.
361 * @sb: Bitmap to show.
362 * @m: struct seq_file to write to.
363 *
364 * This is intended for debugging. The format may change at any time.
365 */
366void sbitmap_show(struct sbitmap *sb, struct seq_file *m);
367
368
369/**
370 * sbitmap_weight() - Return how many set and not cleared bits in a &struct
371 * sbitmap.
372 * @sb: Bitmap to check.
373 *
374 * Return: How many set and not cleared bits set
375 */
376unsigned int sbitmap_weight(const struct sbitmap *sb);
377
378/**
379 * sbitmap_bitmap_show() - Write a hex dump of a &struct sbitmap to a &struct
380 * seq_file.
381 * @sb: Bitmap to show.
382 * @m: struct seq_file to write to.
383 *
384 * This is intended for debugging. The output isn't guaranteed to be internally
385 * consistent.
386 */
387void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m);
388
389/**
390 * sbitmap_queue_init_node() - Initialize a &struct sbitmap_queue on a specific
391 * memory node.
392 * @sbq: Bitmap queue to initialize.
393 * @depth: See sbitmap_init_node().
394 * @shift: See sbitmap_init_node().
395 * @round_robin: See sbitmap_get().
396 * @flags: Allocation flags.
397 * @node: Memory node to allocate on.
398 *
399 * Return: Zero on success or negative errno on failure.
400 */
401int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
402 int shift, bool round_robin, gfp_t flags, int node);
403
404/**
405 * sbitmap_queue_free() - Free memory used by a &struct sbitmap_queue.
406 *
407 * @sbq: Bitmap queue to free.
408 */
409static inline void sbitmap_queue_free(struct sbitmap_queue *sbq)
410{
411 kfree(objp: sbq->ws);
412 sbitmap_free(sb: &sbq->sb);
413}
414
415/**
416 * sbitmap_queue_recalculate_wake_batch() - Recalculate wake batch
417 * @sbq: Bitmap queue to recalculate wake batch.
418 * @users: Number of shares.
419 *
420 * Like sbitmap_queue_update_wake_batch(), this will calculate wake batch
421 * by depth. This interface is for HCTX shared tags or queue shared tags.
422 */
423void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
424 unsigned int users);
425
426/**
427 * sbitmap_queue_resize() - Resize a &struct sbitmap_queue.
428 * @sbq: Bitmap queue to resize.
429 * @depth: New number of bits to resize to.
430 *
431 * Like sbitmap_resize(), this doesn't reallocate anything. It has to do
432 * some extra work on the &struct sbitmap_queue, so it's not safe to just
433 * resize the underlying &struct sbitmap.
434 */
435void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth);
436
437/**
438 * __sbitmap_queue_get() - Try to allocate a free bit from a &struct
439 * sbitmap_queue with preemption already disabled.
440 * @sbq: Bitmap queue to allocate from.
441 *
442 * Return: Non-negative allocated bit number if successful, -1 otherwise.
443 */
444int __sbitmap_queue_get(struct sbitmap_queue *sbq);
445
446/**
447 * __sbitmap_queue_get_batch() - Try to allocate a batch of free bits
448 * @sbq: Bitmap queue to allocate from.
449 * @nr_tags: number of tags requested
450 * @offset: offset to add to returned bits
451 *
452 * Return: Mask of allocated tags, 0 if none are found. Each tag allocated is
453 * a bit in the mask returned, and the caller must add @offset to the value to
454 * get the absolute tag value.
455 */
456unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
457 unsigned int *offset);
458
459/**
460 * sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
461 * sbitmap_queue, limiting the depth used from each word, with preemption
462 * already disabled.
463 * @sbq: Bitmap queue to allocate from.
464 * @shallow_depth: The maximum number of bits to allocate from the queue.
465 * See sbitmap_get_shallow().
466 *
467 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after
468 * initializing @sbq.
469 *
470 * Return: Non-negative allocated bit number if successful, -1 otherwise.
471 */
472int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
473 unsigned int shallow_depth);
474
475/**
476 * sbitmap_queue_get() - Try to allocate a free bit from a &struct
477 * sbitmap_queue.
478 * @sbq: Bitmap queue to allocate from.
479 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to
480 * sbitmap_queue_clear()).
481 *
482 * Return: Non-negative allocated bit number if successful, -1 otherwise.
483 */
484static inline int sbitmap_queue_get(struct sbitmap_queue *sbq,
485 unsigned int *cpu)
486{
487 int nr;
488
489 *cpu = get_cpu();
490 nr = __sbitmap_queue_get(sbq);
491 put_cpu();
492 return nr;
493}
494
495/**
496 * sbitmap_queue_min_shallow_depth() - Inform a &struct sbitmap_queue of the
497 * minimum shallow depth that will be used.
498 * @sbq: Bitmap queue in question.
499 * @min_shallow_depth: The minimum shallow depth that will be passed to
500 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow().
501 *
502 * sbitmap_queue_clear() batches wakeups as an optimization. The batch size
503 * depends on the depth of the bitmap. Since the shallow allocation functions
504 * effectively operate with a different depth, the shallow depth must be taken
505 * into account when calculating the batch size. This function must be called
506 * with the minimum shallow depth that will be used. Failure to do so can result
507 * in missed wakeups.
508 */
509void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
510 unsigned int min_shallow_depth);
511
512/**
513 * sbitmap_queue_clear() - Free an allocated bit and wake up waiters on a
514 * &struct sbitmap_queue.
515 * @sbq: Bitmap to free from.
516 * @nr: Bit number to free.
517 * @cpu: CPU the bit was allocated on.
518 */
519void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
520 unsigned int cpu);
521
522/**
523 * sbitmap_queue_clear_batch() - Free a batch of allocated bits
524 * &struct sbitmap_queue.
525 * @sbq: Bitmap to free from.
526 * @offset: offset for each tag in array
527 * @tags: array of tags
528 * @nr_tags: number of tags in array
529 */
530void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
531 int *tags, int nr_tags);
532
533static inline int sbq_index_inc(int index)
534{
535 return (index + 1) & (SBQ_WAIT_QUEUES - 1);
536}
537
538static inline void sbq_index_atomic_inc(atomic_t *index)
539{
540 int old = atomic_read(v: index);
541 int new = sbq_index_inc(index: old);
542 atomic_cmpxchg(v: index, old, new);
543}
544
545/**
546 * sbq_wait_ptr() - Get the next wait queue to use for a &struct
547 * sbitmap_queue.
548 * @sbq: Bitmap queue to wait on.
549 * @wait_index: A counter per "user" of @sbq.
550 */
551static inline struct sbq_wait_state *sbq_wait_ptr(struct sbitmap_queue *sbq,
552 atomic_t *wait_index)
553{
554 struct sbq_wait_state *ws;
555
556 ws = &sbq->ws[atomic_read(v: wait_index)];
557 sbq_index_atomic_inc(index: wait_index);
558 return ws;
559}
560
561/**
562 * sbitmap_queue_wake_all() - Wake up everything waiting on a &struct
563 * sbitmap_queue.
564 * @sbq: Bitmap queue to wake up.
565 */
566void sbitmap_queue_wake_all(struct sbitmap_queue *sbq);
567
568/**
569 * sbitmap_queue_wake_up() - Wake up some of waiters in one waitqueue
570 * on a &struct sbitmap_queue.
571 * @sbq: Bitmap queue to wake up.
572 * @nr: Number of bits cleared.
573 */
574void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr);
575
576/**
577 * sbitmap_queue_show() - Dump &struct sbitmap_queue information to a &struct
578 * seq_file.
579 * @sbq: Bitmap queue to show.
580 * @m: struct seq_file to write to.
581 *
582 * This is intended for debugging. The format may change at any time.
583 */
584void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m);
585
586struct sbq_wait {
587 struct sbitmap_queue *sbq; /* if set, sbq_wait is accounted */
588 struct wait_queue_entry wait;
589};
590
591#define DEFINE_SBQ_WAIT(name) \
592 struct sbq_wait name = { \
593 .sbq = NULL, \
594 .wait = { \
595 .private = current, \
596 .func = autoremove_wake_function, \
597 .entry = LIST_HEAD_INIT((name).wait.entry), \
598 } \
599 }
600
601/*
602 * Wrapper around prepare_to_wait_exclusive(), which maintains some extra
603 * internal state.
604 */
605void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
606 struct sbq_wait_state *ws,
607 struct sbq_wait *sbq_wait, int state);
608
609/*
610 * Must be paired with sbitmap_prepare_to_wait().
611 */
612void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
613 struct sbq_wait *sbq_wait);
614
615/*
616 * Wrapper around add_wait_queue(), which maintains some extra internal state
617 */
618void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
619 struct sbq_wait_state *ws,
620 struct sbq_wait *sbq_wait);
621
622/*
623 * Must be paired with sbitmap_add_wait_queue()
624 */
625void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait);
626
627#endif /* __LINUX_SCALE_BITMAP_H */
628