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
24 struct seq_file;
25
26 /**
27 * struct sbitmap_word - Word in a &struct sbitmap.
28 */
29 struct 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 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 */
52 struct 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 */
93 struct 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 */
109 struct 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 */
170 int 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 */
__map_depth(const struct sbitmap * sb,int index)174 static 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 */
sbitmap_free(struct sbitmap * sb)185 static inline void sbitmap_free(struct sbitmap *sb)
186 {
187 free_percpu(sb->alloc_hint);
188 kvfree(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 */
200 void 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 */
210 int sbitmap_get(struct sbitmap *sb);
211
212 /**
213 * sbitmap_get_shallow() - Try to allocate a free bit from a &struct sbitmap,
214 * limiting the depth used from each word.
215 * @sb: Bitmap to allocate from.
216 * @shallow_depth: The maximum number of bits to allocate from a single word.
217 *
218 * This rather specific operation allows for having multiple users with
219 * different allocation limits. E.g., there can be a high-priority class that
220 * uses sbitmap_get() and a low-priority class that uses sbitmap_get_shallow()
221 * with a @shallow_depth of (1 << (@sb->shift - 1)). Then, the low-priority
222 * class can only allocate half of the total bits in the bitmap, preventing it
223 * from starving out the high-priority class.
224 *
225 * Return: Non-negative allocated bit number if successful, -1 otherwise.
226 */
227 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth);
228
229 /**
230 * sbitmap_any_bit_set() - Check for a set bit in a &struct sbitmap.
231 * @sb: Bitmap to check.
232 *
233 * Return: true if any bit in the bitmap is set, false otherwise.
234 */
235 bool sbitmap_any_bit_set(const struct sbitmap *sb);
236
237 #define SB_NR_TO_INDEX(sb, bitnr) ((bitnr) >> (sb)->shift)
238 #define SB_NR_TO_BIT(sb, bitnr) ((bitnr) & ((1U << (sb)->shift) - 1U))
239
240 typedef bool (*sb_for_each_fn)(struct sbitmap *, unsigned int, void *);
241
242 /**
243 * __sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
244 * @start: Where to start the iteration.
245 * @sb: Bitmap to iterate over.
246 * @fn: Callback. Should return true to continue or false to break early.
247 * @data: Pointer to pass to callback.
248 *
249 * This is inline even though it's non-trivial so that the function calls to the
250 * callback will hopefully get optimized away.
251 */
__sbitmap_for_each_set(struct sbitmap * sb,unsigned int start,sb_for_each_fn fn,void * data)252 static inline void __sbitmap_for_each_set(struct sbitmap *sb,
253 unsigned int start,
254 sb_for_each_fn fn, void *data)
255 {
256 unsigned int index;
257 unsigned int nr;
258 unsigned int scanned = 0;
259
260 if (start >= sb->depth)
261 start = 0;
262 index = SB_NR_TO_INDEX(sb, start);
263 nr = SB_NR_TO_BIT(sb, start);
264
265 while (scanned < sb->depth) {
266 unsigned long word;
267 unsigned int depth = min_t(unsigned int,
268 __map_depth(sb, index) - nr,
269 sb->depth - scanned);
270
271 scanned += depth;
272 word = sb->map[index].word & ~sb->map[index].cleared;
273 if (!word)
274 goto next;
275
276 /*
277 * On the first iteration of the outer loop, we need to add the
278 * bit offset back to the size of the word for find_next_bit().
279 * On all other iterations, nr is zero, so this is a noop.
280 */
281 depth += nr;
282 while (1) {
283 nr = find_next_bit(&word, depth, nr);
284 if (nr >= depth)
285 break;
286 if (!fn(sb, (index << sb->shift) + nr, data))
287 return;
288
289 nr++;
290 }
291 next:
292 nr = 0;
293 if (++index >= sb->map_nr)
294 index = 0;
295 }
296 }
297
298 /**
299 * sbitmap_for_each_set() - Iterate over each set bit in a &struct sbitmap.
300 * @sb: Bitmap to iterate over.
301 * @fn: Callback. Should return true to continue or false to break early.
302 * @data: Pointer to pass to callback.
303 */
sbitmap_for_each_set(struct sbitmap * sb,sb_for_each_fn fn,void * data)304 static inline void sbitmap_for_each_set(struct sbitmap *sb, sb_for_each_fn fn,
305 void *data)
306 {
307 __sbitmap_for_each_set(sb, 0, fn, data);
308 }
309
__sbitmap_word(struct sbitmap * sb,unsigned int bitnr)310 static inline unsigned long *__sbitmap_word(struct sbitmap *sb,
311 unsigned int bitnr)
312 {
313 return &sb->map[SB_NR_TO_INDEX(sb, bitnr)].word;
314 }
315
316 /* Helpers equivalent to the operations in asm/bitops.h and linux/bitmap.h */
317
sbitmap_set_bit(struct sbitmap * sb,unsigned int bitnr)318 static inline void sbitmap_set_bit(struct sbitmap *sb, unsigned int bitnr)
319 {
320 set_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
321 }
322
sbitmap_clear_bit(struct sbitmap * sb,unsigned int bitnr)323 static inline void sbitmap_clear_bit(struct sbitmap *sb, unsigned int bitnr)
324 {
325 clear_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
326 }
327
328 /*
329 * This one is special, since it doesn't actually clear the bit, rather it
330 * sets the corresponding bit in the ->cleared mask instead. Paired with
331 * the caller doing sbitmap_deferred_clear() if a given index is full, which
332 * will clear the previously freed entries in the corresponding ->word.
333 */
sbitmap_deferred_clear_bit(struct sbitmap * sb,unsigned int bitnr)334 static inline void sbitmap_deferred_clear_bit(struct sbitmap *sb, unsigned int bitnr)
335 {
336 unsigned long *addr = &sb->map[SB_NR_TO_INDEX(sb, bitnr)].cleared;
337
338 set_bit(SB_NR_TO_BIT(sb, bitnr), addr);
339 }
340
341 /*
342 * Pair of sbitmap_get, and this one applies both cleared bit and
343 * allocation hint.
344 */
sbitmap_put(struct sbitmap * sb,unsigned int bitnr)345 static inline void sbitmap_put(struct sbitmap *sb, unsigned int bitnr)
346 {
347 sbitmap_deferred_clear_bit(sb, bitnr);
348
349 if (likely(sb->alloc_hint && !sb->round_robin && bitnr < sb->depth))
350 *raw_cpu_ptr(sb->alloc_hint) = bitnr;
351 }
352
sbitmap_test_bit(struct sbitmap * sb,unsigned int bitnr)353 static inline int sbitmap_test_bit(struct sbitmap *sb, unsigned int bitnr)
354 {
355 return test_bit(SB_NR_TO_BIT(sb, bitnr), __sbitmap_word(sb, bitnr));
356 }
357
sbitmap_calculate_shift(unsigned int depth)358 static inline int sbitmap_calculate_shift(unsigned int depth)
359 {
360 int shift = ilog2(BITS_PER_LONG);
361
362 /*
363 * If the bitmap is small, shrink the number of bits per word so
364 * we spread over a few cachelines, at least. If less than 4
365 * bits, just forget about it, it's not going to work optimally
366 * anyway.
367 */
368 if (depth >= 4) {
369 while ((4U << shift) > depth)
370 shift--;
371 }
372
373 return shift;
374 }
375
376 /**
377 * sbitmap_show() - Dump &struct sbitmap information to a &struct seq_file.
378 * @sb: Bitmap to show.
379 * @m: struct seq_file to write to.
380 *
381 * This is intended for debugging. The format may change at any time.
382 */
383 void sbitmap_show(struct sbitmap *sb, struct seq_file *m);
384
385
386 /**
387 * sbitmap_weight() - Return how many set and not cleared bits in a &struct
388 * sbitmap.
389 * @sb: Bitmap to check.
390 *
391 * Return: How many set and not cleared bits set
392 */
393 unsigned int sbitmap_weight(const struct sbitmap *sb);
394
395 /**
396 * sbitmap_bitmap_show() - Write a hex dump of a &struct sbitmap to a &struct
397 * seq_file.
398 * @sb: Bitmap to show.
399 * @m: struct seq_file to write to.
400 *
401 * This is intended for debugging. The output isn't guaranteed to be internally
402 * consistent.
403 */
404 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m);
405
406 /**
407 * sbitmap_queue_init_node() - Initialize a &struct sbitmap_queue on a specific
408 * memory node.
409 * @sbq: Bitmap queue to initialize.
410 * @depth: See sbitmap_init_node().
411 * @shift: See sbitmap_init_node().
412 * @round_robin: See sbitmap_get().
413 * @flags: Allocation flags.
414 * @node: Memory node to allocate on.
415 *
416 * Return: Zero on success or negative errno on failure.
417 */
418 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
419 int shift, bool round_robin, gfp_t flags, int node);
420
421 /**
422 * sbitmap_queue_free() - Free memory used by a &struct sbitmap_queue.
423 *
424 * @sbq: Bitmap queue to free.
425 */
sbitmap_queue_free(struct sbitmap_queue * sbq)426 static inline void sbitmap_queue_free(struct sbitmap_queue *sbq)
427 {
428 kfree(sbq->ws);
429 sbitmap_free(&sbq->sb);
430 }
431
432 /**
433 * sbitmap_queue_recalculate_wake_batch() - Recalculate wake batch
434 * @sbq: Bitmap queue to recalculate wake batch.
435 * @users: Number of shares.
436 *
437 * Like sbitmap_queue_update_wake_batch(), this will calculate wake batch
438 * by depth. This interface is for HCTX shared tags or queue shared tags.
439 */
440 void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
441 unsigned int users);
442
443 /**
444 * sbitmap_queue_resize() - Resize a &struct sbitmap_queue.
445 * @sbq: Bitmap queue to resize.
446 * @depth: New number of bits to resize to.
447 *
448 * Like sbitmap_resize(), this doesn't reallocate anything. It has to do
449 * some extra work on the &struct sbitmap_queue, so it's not safe to just
450 * resize the underlying &struct sbitmap.
451 */
452 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth);
453
454 /**
455 * __sbitmap_queue_get() - Try to allocate a free bit from a &struct
456 * sbitmap_queue with preemption already disabled.
457 * @sbq: Bitmap queue to allocate from.
458 *
459 * Return: Non-negative allocated bit number if successful, -1 otherwise.
460 */
461 int __sbitmap_queue_get(struct sbitmap_queue *sbq);
462
463 /**
464 * __sbitmap_queue_get_batch() - Try to allocate a batch of free bits
465 * @sbq: Bitmap queue to allocate from.
466 * @nr_tags: number of tags requested
467 * @offset: offset to add to returned bits
468 *
469 * Return: Mask of allocated tags, 0 if none are found. Each tag allocated is
470 * a bit in the mask returned, and the caller must add @offset to the value to
471 * get the absolute tag value.
472 */
473 unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
474 unsigned int *offset);
475
476 /**
477 * sbitmap_queue_get_shallow() - Try to allocate a free bit from a &struct
478 * sbitmap_queue, limiting the depth used from each word, with preemption
479 * already disabled.
480 * @sbq: Bitmap queue to allocate from.
481 * @shallow_depth: The maximum number of bits to allocate from a single word.
482 * See sbitmap_get_shallow().
483 *
484 * If you call this, make sure to call sbitmap_queue_min_shallow_depth() after
485 * initializing @sbq.
486 *
487 * Return: Non-negative allocated bit number if successful, -1 otherwise.
488 */
489 int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
490 unsigned int shallow_depth);
491
492 /**
493 * sbitmap_queue_get() - Try to allocate a free bit from a &struct
494 * sbitmap_queue.
495 * @sbq: Bitmap queue to allocate from.
496 * @cpu: Output parameter; will contain the CPU we ran on (e.g., to be passed to
497 * sbitmap_queue_clear()).
498 *
499 * Return: Non-negative allocated bit number if successful, -1 otherwise.
500 */
sbitmap_queue_get(struct sbitmap_queue * sbq,unsigned int * cpu)501 static inline int sbitmap_queue_get(struct sbitmap_queue *sbq,
502 unsigned int *cpu)
503 {
504 int nr;
505
506 *cpu = get_cpu();
507 nr = __sbitmap_queue_get(sbq);
508 put_cpu();
509 return nr;
510 }
511
512 /**
513 * sbitmap_queue_min_shallow_depth() - Inform a &struct sbitmap_queue of the
514 * minimum shallow depth that will be used.
515 * @sbq: Bitmap queue in question.
516 * @min_shallow_depth: The minimum shallow depth that will be passed to
517 * sbitmap_queue_get_shallow() or __sbitmap_queue_get_shallow().
518 *
519 * sbitmap_queue_clear() batches wakeups as an optimization. The batch size
520 * depends on the depth of the bitmap. Since the shallow allocation functions
521 * effectively operate with a different depth, the shallow depth must be taken
522 * into account when calculating the batch size. This function must be called
523 * with the minimum shallow depth that will be used. Failure to do so can result
524 * in missed wakeups.
525 */
526 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
527 unsigned int min_shallow_depth);
528
529 /**
530 * sbitmap_queue_clear() - Free an allocated bit and wake up waiters on a
531 * &struct sbitmap_queue.
532 * @sbq: Bitmap to free from.
533 * @nr: Bit number to free.
534 * @cpu: CPU the bit was allocated on.
535 */
536 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
537 unsigned int cpu);
538
539 /**
540 * sbitmap_queue_clear_batch() - Free a batch of allocated bits
541 * &struct sbitmap_queue.
542 * @sbq: Bitmap to free from.
543 * @offset: offset for each tag in array
544 * @tags: array of tags
545 * @nr_tags: number of tags in array
546 */
547 void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
548 int *tags, int nr_tags);
549
sbq_index_inc(int index)550 static inline int sbq_index_inc(int index)
551 {
552 return (index + 1) & (SBQ_WAIT_QUEUES - 1);
553 }
554
sbq_index_atomic_inc(atomic_t * index)555 static inline void sbq_index_atomic_inc(atomic_t *index)
556 {
557 int old = atomic_read(index);
558 int new = sbq_index_inc(old);
559 atomic_cmpxchg(index, old, new);
560 }
561
562 /**
563 * sbq_wait_ptr() - Get the next wait queue to use for a &struct
564 * sbitmap_queue.
565 * @sbq: Bitmap queue to wait on.
566 * @wait_index: A counter per "user" of @sbq.
567 */
sbq_wait_ptr(struct sbitmap_queue * sbq,atomic_t * wait_index)568 static inline struct sbq_wait_state *sbq_wait_ptr(struct sbitmap_queue *sbq,
569 atomic_t *wait_index)
570 {
571 struct sbq_wait_state *ws;
572
573 ws = &sbq->ws[atomic_read(wait_index)];
574 sbq_index_atomic_inc(wait_index);
575 return ws;
576 }
577
578 /**
579 * sbitmap_queue_wake_all() - Wake up everything waiting on a &struct
580 * sbitmap_queue.
581 * @sbq: Bitmap queue to wake up.
582 */
583 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq);
584
585 /**
586 * sbitmap_queue_wake_up() - Wake up some of waiters in one waitqueue
587 * on a &struct sbitmap_queue.
588 * @sbq: Bitmap queue to wake up.
589 * @nr: Number of bits cleared.
590 */
591 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr);
592
593 /**
594 * sbitmap_queue_show() - Dump &struct sbitmap_queue information to a &struct
595 * seq_file.
596 * @sbq: Bitmap queue to show.
597 * @m: struct seq_file to write to.
598 *
599 * This is intended for debugging. The format may change at any time.
600 */
601 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m);
602
603 struct sbq_wait {
604 struct sbitmap_queue *sbq; /* if set, sbq_wait is accounted */
605 struct wait_queue_entry wait;
606 };
607
608 #define DEFINE_SBQ_WAIT(name) \
609 struct sbq_wait name = { \
610 .sbq = NULL, \
611 .wait = { \
612 .private = current, \
613 .func = autoremove_wake_function, \
614 .entry = LIST_HEAD_INIT((name).wait.entry), \
615 } \
616 }
617
618 /*
619 * Wrapper around prepare_to_wait_exclusive(), which maintains some extra
620 * internal state.
621 */
622 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
623 struct sbq_wait_state *ws,
624 struct sbq_wait *sbq_wait, int state);
625
626 /*
627 * Must be paired with sbitmap_prepare_to_wait().
628 */
629 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
630 struct sbq_wait *sbq_wait);
631
632 /*
633 * Wrapper around add_wait_queue(), which maintains some extra internal state
634 */
635 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
636 struct sbq_wait_state *ws,
637 struct sbq_wait *sbq_wait);
638
639 /*
640 * Must be paired with sbitmap_add_wait_queue()
641 */
642 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait);
643
644 #endif /* __LINUX_SCALE_BITMAP_H */
645