xref: /openbmc/linux/lib/sbitmap.c (revision f1591a8b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2016 Facebook
4  * Copyright (C) 2013-2014 Jens Axboe
5  */
6 
7 #include <linux/sched.h>
8 #include <linux/random.h>
9 #include <linux/sbitmap.h>
10 #include <linux/seq_file.h>
11 
12 static int init_alloc_hint(struct sbitmap *sb, gfp_t flags)
13 {
14 	unsigned depth = sb->depth;
15 
16 	sb->alloc_hint = alloc_percpu_gfp(unsigned int, flags);
17 	if (!sb->alloc_hint)
18 		return -ENOMEM;
19 
20 	if (depth && !sb->round_robin) {
21 		int i;
22 
23 		for_each_possible_cpu(i)
24 			*per_cpu_ptr(sb->alloc_hint, i) = get_random_u32_below(depth);
25 	}
26 	return 0;
27 }
28 
29 static inline unsigned update_alloc_hint_before_get(struct sbitmap *sb,
30 						    unsigned int depth)
31 {
32 	unsigned hint;
33 
34 	hint = this_cpu_read(*sb->alloc_hint);
35 	if (unlikely(hint >= depth)) {
36 		hint = depth ? get_random_u32_below(depth) : 0;
37 		this_cpu_write(*sb->alloc_hint, hint);
38 	}
39 
40 	return hint;
41 }
42 
43 static inline void update_alloc_hint_after_get(struct sbitmap *sb,
44 					       unsigned int depth,
45 					       unsigned int hint,
46 					       unsigned int nr)
47 {
48 	if (nr == -1) {
49 		/* If the map is full, a hint won't do us much good. */
50 		this_cpu_write(*sb->alloc_hint, 0);
51 	} else if (nr == hint || unlikely(sb->round_robin)) {
52 		/* Only update the hint if we used it. */
53 		hint = nr + 1;
54 		if (hint >= depth - 1)
55 			hint = 0;
56 		this_cpu_write(*sb->alloc_hint, hint);
57 	}
58 }
59 
60 /*
61  * See if we have deferred clears that we can batch move
62  */
63 static inline bool sbitmap_deferred_clear(struct sbitmap_word *map)
64 {
65 	unsigned long mask;
66 
67 	if (!READ_ONCE(map->cleared))
68 		return false;
69 
70 	/*
71 	 * First get a stable cleared mask, setting the old mask to 0.
72 	 */
73 	mask = xchg(&map->cleared, 0);
74 
75 	/*
76 	 * Now clear the masked bits in our free word
77 	 */
78 	atomic_long_andnot(mask, (atomic_long_t *)&map->word);
79 	BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(map->word));
80 	return true;
81 }
82 
83 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
84 		      gfp_t flags, int node, bool round_robin,
85 		      bool alloc_hint)
86 {
87 	unsigned int bits_per_word;
88 
89 	if (shift < 0)
90 		shift = sbitmap_calculate_shift(depth);
91 
92 	bits_per_word = 1U << shift;
93 	if (bits_per_word > BITS_PER_LONG)
94 		return -EINVAL;
95 
96 	sb->shift = shift;
97 	sb->depth = depth;
98 	sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
99 	sb->round_robin = round_robin;
100 
101 	if (depth == 0) {
102 		sb->map = NULL;
103 		return 0;
104 	}
105 
106 	if (alloc_hint) {
107 		if (init_alloc_hint(sb, flags))
108 			return -ENOMEM;
109 	} else {
110 		sb->alloc_hint = NULL;
111 	}
112 
113 	sb->map = kvzalloc_node(sb->map_nr * sizeof(*sb->map), flags, node);
114 	if (!sb->map) {
115 		free_percpu(sb->alloc_hint);
116 		return -ENOMEM;
117 	}
118 
119 	return 0;
120 }
121 EXPORT_SYMBOL_GPL(sbitmap_init_node);
122 
123 void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
124 {
125 	unsigned int bits_per_word = 1U << sb->shift;
126 	unsigned int i;
127 
128 	for (i = 0; i < sb->map_nr; i++)
129 		sbitmap_deferred_clear(&sb->map[i]);
130 
131 	sb->depth = depth;
132 	sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
133 }
134 EXPORT_SYMBOL_GPL(sbitmap_resize);
135 
136 static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
137 			      unsigned int hint, bool wrap)
138 {
139 	int nr;
140 
141 	/* don't wrap if starting from 0 */
142 	wrap = wrap && hint;
143 
144 	while (1) {
145 		nr = find_next_zero_bit(word, depth, hint);
146 		if (unlikely(nr >= depth)) {
147 			/*
148 			 * We started with an offset, and we didn't reset the
149 			 * offset to 0 in a failure case, so start from 0 to
150 			 * exhaust the map.
151 			 */
152 			if (hint && wrap) {
153 				hint = 0;
154 				continue;
155 			}
156 			return -1;
157 		}
158 
159 		if (!test_and_set_bit_lock(nr, word))
160 			break;
161 
162 		hint = nr + 1;
163 		if (hint >= depth - 1)
164 			hint = 0;
165 	}
166 
167 	return nr;
168 }
169 
170 static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
171 				     unsigned int alloc_hint)
172 {
173 	struct sbitmap_word *map = &sb->map[index];
174 	int nr;
175 
176 	do {
177 		nr = __sbitmap_get_word(&map->word, __map_depth(sb, index),
178 					alloc_hint, !sb->round_robin);
179 		if (nr != -1)
180 			break;
181 		if (!sbitmap_deferred_clear(map))
182 			break;
183 	} while (1);
184 
185 	return nr;
186 }
187 
188 static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
189 {
190 	unsigned int i, index;
191 	int nr = -1;
192 
193 	index = SB_NR_TO_INDEX(sb, alloc_hint);
194 
195 	/*
196 	 * Unless we're doing round robin tag allocation, just use the
197 	 * alloc_hint to find the right word index. No point in looping
198 	 * twice in find_next_zero_bit() for that case.
199 	 */
200 	if (sb->round_robin)
201 		alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
202 	else
203 		alloc_hint = 0;
204 
205 	for (i = 0; i < sb->map_nr; i++) {
206 		nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
207 		if (nr != -1) {
208 			nr += index << sb->shift;
209 			break;
210 		}
211 
212 		/* Jump to next index. */
213 		alloc_hint = 0;
214 		if (++index >= sb->map_nr)
215 			index = 0;
216 	}
217 
218 	return nr;
219 }
220 
221 int sbitmap_get(struct sbitmap *sb)
222 {
223 	int nr;
224 	unsigned int hint, depth;
225 
226 	if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
227 		return -1;
228 
229 	depth = READ_ONCE(sb->depth);
230 	hint = update_alloc_hint_before_get(sb, depth);
231 	nr = __sbitmap_get(sb, hint);
232 	update_alloc_hint_after_get(sb, depth, hint, nr);
233 
234 	return nr;
235 }
236 EXPORT_SYMBOL_GPL(sbitmap_get);
237 
238 static int __sbitmap_get_shallow(struct sbitmap *sb,
239 				 unsigned int alloc_hint,
240 				 unsigned long shallow_depth)
241 {
242 	unsigned int i, index;
243 	int nr = -1;
244 
245 	index = SB_NR_TO_INDEX(sb, alloc_hint);
246 	alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
247 
248 	for (i = 0; i < sb->map_nr; i++) {
249 again:
250 		nr = __sbitmap_get_word(&sb->map[index].word,
251 					min_t(unsigned int,
252 					      __map_depth(sb, index),
253 					      shallow_depth),
254 					alloc_hint, true);
255 		if (nr != -1) {
256 			nr += index << sb->shift;
257 			break;
258 		}
259 
260 		if (sbitmap_deferred_clear(&sb->map[index]))
261 			goto again;
262 
263 		/* Jump to next index. */
264 		alloc_hint = 0;
265 		if (++index >= sb->map_nr)
266 			index = 0;
267 	}
268 
269 	return nr;
270 }
271 
272 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
273 {
274 	int nr;
275 	unsigned int hint, depth;
276 
277 	if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
278 		return -1;
279 
280 	depth = READ_ONCE(sb->depth);
281 	hint = update_alloc_hint_before_get(sb, depth);
282 	nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
283 	update_alloc_hint_after_get(sb, depth, hint, nr);
284 
285 	return nr;
286 }
287 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
288 
289 bool sbitmap_any_bit_set(const struct sbitmap *sb)
290 {
291 	unsigned int i;
292 
293 	for (i = 0; i < sb->map_nr; i++) {
294 		if (sb->map[i].word & ~sb->map[i].cleared)
295 			return true;
296 	}
297 	return false;
298 }
299 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
300 
301 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
302 {
303 	unsigned int i, weight = 0;
304 
305 	for (i = 0; i < sb->map_nr; i++) {
306 		const struct sbitmap_word *word = &sb->map[i];
307 		unsigned int word_depth = __map_depth(sb, i);
308 
309 		if (set)
310 			weight += bitmap_weight(&word->word, word_depth);
311 		else
312 			weight += bitmap_weight(&word->cleared, word_depth);
313 	}
314 	return weight;
315 }
316 
317 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
318 {
319 	return __sbitmap_weight(sb, false);
320 }
321 
322 unsigned int sbitmap_weight(const struct sbitmap *sb)
323 {
324 	return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
325 }
326 EXPORT_SYMBOL_GPL(sbitmap_weight);
327 
328 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
329 {
330 	seq_printf(m, "depth=%u\n", sb->depth);
331 	seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
332 	seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
333 	seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
334 	seq_printf(m, "map_nr=%u\n", sb->map_nr);
335 }
336 EXPORT_SYMBOL_GPL(sbitmap_show);
337 
338 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
339 {
340 	if ((offset & 0xf) == 0) {
341 		if (offset != 0)
342 			seq_putc(m, '\n');
343 		seq_printf(m, "%08x:", offset);
344 	}
345 	if ((offset & 0x1) == 0)
346 		seq_putc(m, ' ');
347 	seq_printf(m, "%02x", byte);
348 }
349 
350 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
351 {
352 	u8 byte = 0;
353 	unsigned int byte_bits = 0;
354 	unsigned int offset = 0;
355 	int i;
356 
357 	for (i = 0; i < sb->map_nr; i++) {
358 		unsigned long word = READ_ONCE(sb->map[i].word);
359 		unsigned long cleared = READ_ONCE(sb->map[i].cleared);
360 		unsigned int word_bits = __map_depth(sb, i);
361 
362 		word &= ~cleared;
363 
364 		while (word_bits > 0) {
365 			unsigned int bits = min(8 - byte_bits, word_bits);
366 
367 			byte |= (word & (BIT(bits) - 1)) << byte_bits;
368 			byte_bits += bits;
369 			if (byte_bits == 8) {
370 				emit_byte(m, offset, byte);
371 				byte = 0;
372 				byte_bits = 0;
373 				offset++;
374 			}
375 			word >>= bits;
376 			word_bits -= bits;
377 		}
378 	}
379 	if (byte_bits) {
380 		emit_byte(m, offset, byte);
381 		offset++;
382 	}
383 	if (offset)
384 		seq_putc(m, '\n');
385 }
386 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
387 
388 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
389 					unsigned int depth)
390 {
391 	unsigned int wake_batch;
392 	unsigned int shallow_depth;
393 
394 	/*
395 	 * For each batch, we wake up one queue. We need to make sure that our
396 	 * batch size is small enough that the full depth of the bitmap,
397 	 * potentially limited by a shallow depth, is enough to wake up all of
398 	 * the queues.
399 	 *
400 	 * Each full word of the bitmap has bits_per_word bits, and there might
401 	 * be a partial word. There are depth / bits_per_word full words and
402 	 * depth % bits_per_word bits left over. In bitwise arithmetic:
403 	 *
404 	 * bits_per_word = 1 << shift
405 	 * depth / bits_per_word = depth >> shift
406 	 * depth % bits_per_word = depth & ((1 << shift) - 1)
407 	 *
408 	 * Each word can be limited to sbq->min_shallow_depth bits.
409 	 */
410 	shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
411 	depth = ((depth >> sbq->sb.shift) * shallow_depth +
412 		 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
413 	wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
414 			     SBQ_WAKE_BATCH);
415 
416 	return wake_batch;
417 }
418 
419 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
420 			    int shift, bool round_robin, gfp_t flags, int node)
421 {
422 	int ret;
423 	int i;
424 
425 	ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
426 				round_robin, true);
427 	if (ret)
428 		return ret;
429 
430 	sbq->min_shallow_depth = UINT_MAX;
431 	sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
432 	atomic_set(&sbq->wake_index, 0);
433 	atomic_set(&sbq->ws_active, 0);
434 	atomic_set(&sbq->completion_cnt, 0);
435 	atomic_set(&sbq->wakeup_cnt, 0);
436 
437 	sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
438 	if (!sbq->ws) {
439 		sbitmap_free(&sbq->sb);
440 		return -ENOMEM;
441 	}
442 
443 	for (i = 0; i < SBQ_WAIT_QUEUES; i++)
444 		init_waitqueue_head(&sbq->ws[i].wait);
445 
446 	return 0;
447 }
448 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
449 
450 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
451 					    unsigned int depth)
452 {
453 	unsigned int wake_batch;
454 
455 	wake_batch = sbq_calc_wake_batch(sbq, depth);
456 	if (sbq->wake_batch != wake_batch)
457 		WRITE_ONCE(sbq->wake_batch, wake_batch);
458 }
459 
460 void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
461 					    unsigned int users)
462 {
463 	unsigned int wake_batch;
464 	unsigned int min_batch;
465 	unsigned int depth = (sbq->sb.depth + users - 1) / users;
466 
467 	min_batch = sbq->sb.depth >= (4 * SBQ_WAIT_QUEUES) ? 4 : 1;
468 
469 	wake_batch = clamp_val(depth / SBQ_WAIT_QUEUES,
470 			min_batch, SBQ_WAKE_BATCH);
471 
472 	WRITE_ONCE(sbq->wake_batch, wake_batch);
473 }
474 EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
475 
476 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
477 {
478 	sbitmap_queue_update_wake_batch(sbq, depth);
479 	sbitmap_resize(&sbq->sb, depth);
480 }
481 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
482 
483 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
484 {
485 	return sbitmap_get(&sbq->sb);
486 }
487 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
488 
489 unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
490 					unsigned int *offset)
491 {
492 	struct sbitmap *sb = &sbq->sb;
493 	unsigned int hint, depth;
494 	unsigned long index, nr;
495 	int i;
496 
497 	if (unlikely(sb->round_robin))
498 		return 0;
499 
500 	depth = READ_ONCE(sb->depth);
501 	hint = update_alloc_hint_before_get(sb, depth);
502 
503 	index = SB_NR_TO_INDEX(sb, hint);
504 
505 	for (i = 0; i < sb->map_nr; i++) {
506 		struct sbitmap_word *map = &sb->map[index];
507 		unsigned long get_mask;
508 		unsigned int map_depth = __map_depth(sb, index);
509 
510 		sbitmap_deferred_clear(map);
511 		if (map->word == (1UL << (map_depth - 1)) - 1)
512 			goto next;
513 
514 		nr = find_first_zero_bit(&map->word, map_depth);
515 		if (nr + nr_tags <= map_depth) {
516 			atomic_long_t *ptr = (atomic_long_t *) &map->word;
517 			unsigned long val;
518 
519 			get_mask = ((1UL << nr_tags) - 1) << nr;
520 			val = READ_ONCE(map->word);
521 			do {
522 				if ((val & ~get_mask) != val)
523 					goto next;
524 			} while (!atomic_long_try_cmpxchg(ptr, &val,
525 							  get_mask | val));
526 			get_mask = (get_mask & ~val) >> nr;
527 			if (get_mask) {
528 				*offset = nr + (index << sb->shift);
529 				update_alloc_hint_after_get(sb, depth, hint,
530 							*offset + nr_tags - 1);
531 				return get_mask;
532 			}
533 		}
534 next:
535 		/* Jump to next index. */
536 		if (++index >= sb->map_nr)
537 			index = 0;
538 	}
539 
540 	return 0;
541 }
542 
543 int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
544 			      unsigned int shallow_depth)
545 {
546 	WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
547 
548 	return sbitmap_get_shallow(&sbq->sb, shallow_depth);
549 }
550 EXPORT_SYMBOL_GPL(sbitmap_queue_get_shallow);
551 
552 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
553 				     unsigned int min_shallow_depth)
554 {
555 	sbq->min_shallow_depth = min_shallow_depth;
556 	sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
557 }
558 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
559 
560 static void __sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
561 {
562 	int i, wake_index;
563 
564 	if (!atomic_read(&sbq->ws_active))
565 		return;
566 
567 	wake_index = atomic_read(&sbq->wake_index);
568 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
569 		struct sbq_wait_state *ws = &sbq->ws[wake_index];
570 
571 		/*
572 		 * Advance the index before checking the current queue.
573 		 * It improves fairness, by ensuring the queue doesn't
574 		 * need to be fully emptied before trying to wake up
575 		 * from the next one.
576 		 */
577 		wake_index = sbq_index_inc(wake_index);
578 
579 		/*
580 		 * It is sufficient to wake up at least one waiter to
581 		 * guarantee forward progress.
582 		 */
583 		if (waitqueue_active(&ws->wait) &&
584 		    wake_up_nr(&ws->wait, nr))
585 			break;
586 	}
587 
588 	if (wake_index != atomic_read(&sbq->wake_index))
589 		atomic_set(&sbq->wake_index, wake_index);
590 }
591 
592 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
593 {
594 	unsigned int wake_batch = READ_ONCE(sbq->wake_batch);
595 	unsigned int wakeups;
596 
597 	if (!atomic_read(&sbq->ws_active))
598 		return;
599 
600 	atomic_add(nr, &sbq->completion_cnt);
601 	wakeups = atomic_read(&sbq->wakeup_cnt);
602 
603 	do {
604 		if (atomic_read(&sbq->completion_cnt) - wakeups < wake_batch)
605 			return;
606 	} while (!atomic_try_cmpxchg(&sbq->wakeup_cnt,
607 				     &wakeups, wakeups + wake_batch));
608 
609 	__sbitmap_queue_wake_up(sbq, wake_batch);
610 }
611 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
612 
613 static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
614 {
615 	if (likely(!sb->round_robin && tag < sb->depth))
616 		data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
617 }
618 
619 void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
620 				int *tags, int nr_tags)
621 {
622 	struct sbitmap *sb = &sbq->sb;
623 	unsigned long *addr = NULL;
624 	unsigned long mask = 0;
625 	int i;
626 
627 	smp_mb__before_atomic();
628 	for (i = 0; i < nr_tags; i++) {
629 		const int tag = tags[i] - offset;
630 		unsigned long *this_addr;
631 
632 		/* since we're clearing a batch, skip the deferred map */
633 		this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
634 		if (!addr) {
635 			addr = this_addr;
636 		} else if (addr != this_addr) {
637 			atomic_long_andnot(mask, (atomic_long_t *) addr);
638 			mask = 0;
639 			addr = this_addr;
640 		}
641 		mask |= (1UL << SB_NR_TO_BIT(sb, tag));
642 	}
643 
644 	if (mask)
645 		atomic_long_andnot(mask, (atomic_long_t *) addr);
646 
647 	smp_mb__after_atomic();
648 	sbitmap_queue_wake_up(sbq, nr_tags);
649 	sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
650 					tags[nr_tags - 1] - offset);
651 }
652 
653 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
654 			 unsigned int cpu)
655 {
656 	/*
657 	 * Once the clear bit is set, the bit may be allocated out.
658 	 *
659 	 * Orders READ/WRITE on the associated instance(such as request
660 	 * of blk_mq) by this bit for avoiding race with re-allocation,
661 	 * and its pair is the memory barrier implied in __sbitmap_get_word.
662 	 *
663 	 * One invariant is that the clear bit has to be zero when the bit
664 	 * is in use.
665 	 */
666 	smp_mb__before_atomic();
667 	sbitmap_deferred_clear_bit(&sbq->sb, nr);
668 
669 	/*
670 	 * Pairs with the memory barrier in set_current_state() to ensure the
671 	 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
672 	 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
673 	 * waiter. See the comment on waitqueue_active().
674 	 */
675 	smp_mb__after_atomic();
676 	sbitmap_queue_wake_up(sbq, 1);
677 	sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
678 }
679 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
680 
681 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
682 {
683 	int i, wake_index;
684 
685 	/*
686 	 * Pairs with the memory barrier in set_current_state() like in
687 	 * sbitmap_queue_wake_up().
688 	 */
689 	smp_mb();
690 	wake_index = atomic_read(&sbq->wake_index);
691 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
692 		struct sbq_wait_state *ws = &sbq->ws[wake_index];
693 
694 		if (waitqueue_active(&ws->wait))
695 			wake_up(&ws->wait);
696 
697 		wake_index = sbq_index_inc(wake_index);
698 	}
699 }
700 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
701 
702 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
703 {
704 	bool first;
705 	int i;
706 
707 	sbitmap_show(&sbq->sb, m);
708 
709 	seq_puts(m, "alloc_hint={");
710 	first = true;
711 	for_each_possible_cpu(i) {
712 		if (!first)
713 			seq_puts(m, ", ");
714 		first = false;
715 		seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
716 	}
717 	seq_puts(m, "}\n");
718 
719 	seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
720 	seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
721 	seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
722 
723 	seq_puts(m, "ws={\n");
724 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
725 		struct sbq_wait_state *ws = &sbq->ws[i];
726 		seq_printf(m, "\t{.wait=%s},\n",
727 			   waitqueue_active(&ws->wait) ? "active" : "inactive");
728 	}
729 	seq_puts(m, "}\n");
730 
731 	seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
732 	seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
733 }
734 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
735 
736 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
737 			    struct sbq_wait_state *ws,
738 			    struct sbq_wait *sbq_wait)
739 {
740 	if (!sbq_wait->sbq) {
741 		sbq_wait->sbq = sbq;
742 		atomic_inc(&sbq->ws_active);
743 		add_wait_queue(&ws->wait, &sbq_wait->wait);
744 	}
745 }
746 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
747 
748 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
749 {
750 	list_del_init(&sbq_wait->wait.entry);
751 	if (sbq_wait->sbq) {
752 		atomic_dec(&sbq_wait->sbq->ws_active);
753 		sbq_wait->sbq = NULL;
754 	}
755 }
756 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
757 
758 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
759 			     struct sbq_wait_state *ws,
760 			     struct sbq_wait *sbq_wait, int state)
761 {
762 	if (!sbq_wait->sbq) {
763 		atomic_inc(&sbq->ws_active);
764 		sbq_wait->sbq = sbq;
765 	}
766 	prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
767 }
768 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
769 
770 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
771 			 struct sbq_wait *sbq_wait)
772 {
773 	finish_wait(&ws->wait, &sbq_wait->wait);
774 	if (sbq_wait->sbq) {
775 		atomic_dec(&sbq->ws_active);
776 		sbq_wait->sbq = NULL;
777 	}
778 }
779 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);
780