1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2016 Facebook
4 * Copyright (C) 2013-2014 Jens Axboe
7 #include <linux/sched.h>
8 #include <linux/random.h>
9 #include <linux/sbitmap.h>
10 #include <linux/seq_file.h>
12 static int init_alloc_hint(struct sbitmap *sb, gfp_t flags)
14 unsigned depth = sb->depth;
16 sb->alloc_hint = alloc_percpu_gfp(unsigned int, flags);
20 if (depth && !sb->round_robin) {
23 for_each_possible_cpu(i)
24 *per_cpu_ptr(sb->alloc_hint, i) = prandom_u32() % depth;
29 static inline unsigned update_alloc_hint_before_get(struct sbitmap *sb,
34 hint = this_cpu_read(*sb->alloc_hint);
35 if (unlikely(hint >= depth)) {
36 hint = depth ? prandom_u32() % depth : 0;
37 this_cpu_write(*sb->alloc_hint, hint);
43 static inline void update_alloc_hint_after_get(struct sbitmap *sb,
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. */
54 if (hint >= depth - 1)
56 this_cpu_write(*sb->alloc_hint, hint);
61 * See if we have deferred clears that we can batch move
63 static inline bool sbitmap_deferred_clear(struct sbitmap_word *map)
67 if (!READ_ONCE(map->cleared))
71 * First get a stable cleared mask, setting the old mask to 0.
73 mask = xchg(&map->cleared, 0);
76 * Now clear the masked bits in our free word
78 atomic_long_andnot(mask, (atomic_long_t *)&map->word);
79 BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(map->word));
83 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
84 gfp_t flags, int node, bool round_robin,
87 unsigned int bits_per_word;
91 shift = sbitmap_calculate_shift(depth);
93 bits_per_word = 1U << shift;
94 if (bits_per_word > BITS_PER_LONG)
99 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
100 sb->round_robin = round_robin;
108 if (init_alloc_hint(sb, flags))
111 sb->alloc_hint = NULL;
114 sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node);
116 free_percpu(sb->alloc_hint);
120 for (i = 0; i < sb->map_nr; i++) {
121 sb->map[i].depth = min(depth, bits_per_word);
122 depth -= sb->map[i].depth;
126 EXPORT_SYMBOL_GPL(sbitmap_init_node);
128 void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
130 unsigned int bits_per_word = 1U << sb->shift;
133 for (i = 0; i < sb->map_nr; i++)
134 sbitmap_deferred_clear(&sb->map[i]);
137 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
139 for (i = 0; i < sb->map_nr; i++) {
140 sb->map[i].depth = min(depth, bits_per_word);
141 depth -= sb->map[i].depth;
144 EXPORT_SYMBOL_GPL(sbitmap_resize);
146 static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
147 unsigned int hint, bool wrap)
151 /* don't wrap if starting from 0 */
155 nr = find_next_zero_bit(word, depth, hint);
156 if (unlikely(nr >= depth)) {
158 * We started with an offset, and we didn't reset the
159 * offset to 0 in a failure case, so start from 0 to
169 if (!test_and_set_bit_lock(nr, word))
173 if (hint >= depth - 1)
180 static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
181 unsigned int alloc_hint)
183 struct sbitmap_word *map = &sb->map[index];
187 nr = __sbitmap_get_word(&map->word, map->depth, alloc_hint,
191 if (!sbitmap_deferred_clear(map))
198 static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
200 unsigned int i, index;
203 index = SB_NR_TO_INDEX(sb, alloc_hint);
206 * Unless we're doing round robin tag allocation, just use the
207 * alloc_hint to find the right word index. No point in looping
208 * twice in find_next_zero_bit() for that case.
211 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
215 for (i = 0; i < sb->map_nr; i++) {
216 nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
218 nr += index << sb->shift;
222 /* Jump to next index. */
224 if (++index >= sb->map_nr)
231 int sbitmap_get(struct sbitmap *sb)
234 unsigned int hint, depth;
236 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
239 depth = READ_ONCE(sb->depth);
240 hint = update_alloc_hint_before_get(sb, depth);
241 nr = __sbitmap_get(sb, hint);
242 update_alloc_hint_after_get(sb, depth, hint, nr);
246 EXPORT_SYMBOL_GPL(sbitmap_get);
248 static int __sbitmap_get_shallow(struct sbitmap *sb,
249 unsigned int alloc_hint,
250 unsigned long shallow_depth)
252 unsigned int i, index;
255 index = SB_NR_TO_INDEX(sb, alloc_hint);
257 for (i = 0; i < sb->map_nr; i++) {
259 nr = __sbitmap_get_word(&sb->map[index].word,
260 min(sb->map[index].depth, shallow_depth),
261 SB_NR_TO_BIT(sb, alloc_hint), true);
263 nr += index << sb->shift;
267 if (sbitmap_deferred_clear(&sb->map[index]))
270 /* Jump to next index. */
272 alloc_hint = index << sb->shift;
274 if (index >= sb->map_nr) {
283 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
286 unsigned int hint, depth;
288 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
291 depth = READ_ONCE(sb->depth);
292 hint = update_alloc_hint_before_get(sb, depth);
293 nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
294 update_alloc_hint_after_get(sb, depth, hint, nr);
298 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
300 bool sbitmap_any_bit_set(const struct sbitmap *sb)
304 for (i = 0; i < sb->map_nr; i++) {
305 if (sb->map[i].word & ~sb->map[i].cleared)
310 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
312 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
314 unsigned int i, weight = 0;
316 for (i = 0; i < sb->map_nr; i++) {
317 const struct sbitmap_word *word = &sb->map[i];
320 weight += bitmap_weight(&word->word, word->depth);
322 weight += bitmap_weight(&word->cleared, word->depth);
327 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
329 return __sbitmap_weight(sb, false);
332 unsigned int sbitmap_weight(const struct sbitmap *sb)
334 return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
336 EXPORT_SYMBOL_GPL(sbitmap_weight);
338 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
340 seq_printf(m, "depth=%u\n", sb->depth);
341 seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
342 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
343 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
344 seq_printf(m, "map_nr=%u\n", sb->map_nr);
346 EXPORT_SYMBOL_GPL(sbitmap_show);
348 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
350 if ((offset & 0xf) == 0) {
353 seq_printf(m, "%08x:", offset);
355 if ((offset & 0x1) == 0)
357 seq_printf(m, "%02x", byte);
360 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
363 unsigned int byte_bits = 0;
364 unsigned int offset = 0;
367 for (i = 0; i < sb->map_nr; i++) {
368 unsigned long word = READ_ONCE(sb->map[i].word);
369 unsigned long cleared = READ_ONCE(sb->map[i].cleared);
370 unsigned int word_bits = READ_ONCE(sb->map[i].depth);
374 while (word_bits > 0) {
375 unsigned int bits = min(8 - byte_bits, word_bits);
377 byte |= (word & (BIT(bits) - 1)) << byte_bits;
379 if (byte_bits == 8) {
380 emit_byte(m, offset, byte);
390 emit_byte(m, offset, byte);
396 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
398 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
401 unsigned int wake_batch;
402 unsigned int shallow_depth;
405 * For each batch, we wake up one queue. We need to make sure that our
406 * batch size is small enough that the full depth of the bitmap,
407 * potentially limited by a shallow depth, is enough to wake up all of
410 * Each full word of the bitmap has bits_per_word bits, and there might
411 * be a partial word. There are depth / bits_per_word full words and
412 * depth % bits_per_word bits left over. In bitwise arithmetic:
414 * bits_per_word = 1 << shift
415 * depth / bits_per_word = depth >> shift
416 * depth % bits_per_word = depth & ((1 << shift) - 1)
418 * Each word can be limited to sbq->min_shallow_depth bits.
420 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
421 depth = ((depth >> sbq->sb.shift) * shallow_depth +
422 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
423 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
429 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
430 int shift, bool round_robin, gfp_t flags, int node)
435 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
440 sbq->min_shallow_depth = UINT_MAX;
441 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
442 atomic_set(&sbq->wake_index, 0);
443 atomic_set(&sbq->ws_active, 0);
445 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
447 sbitmap_free(&sbq->sb);
451 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
452 init_waitqueue_head(&sbq->ws[i].wait);
453 atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
458 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
460 static inline void __sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
461 unsigned int wake_batch)
465 if (sbq->wake_batch != wake_batch) {
466 WRITE_ONCE(sbq->wake_batch, wake_batch);
468 * Pairs with the memory barrier in sbitmap_queue_wake_up()
469 * to ensure that the batch size is updated before the wait
473 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
474 atomic_set(&sbq->ws[i].wait_cnt, 1);
478 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
481 unsigned int wake_batch;
483 wake_batch = sbq_calc_wake_batch(sbq, depth);
484 __sbitmap_queue_update_wake_batch(sbq, wake_batch);
487 void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
490 unsigned int wake_batch;
492 wake_batch = clamp_val((sbq->sb.depth + users - 1) /
493 users, 4, SBQ_WAKE_BATCH);
494 __sbitmap_queue_update_wake_batch(sbq, wake_batch);
496 EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
498 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
500 sbitmap_queue_update_wake_batch(sbq, depth);
501 sbitmap_resize(&sbq->sb, depth);
503 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
505 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
507 return sbitmap_get(&sbq->sb);
509 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
511 unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
512 unsigned int *offset)
514 struct sbitmap *sb = &sbq->sb;
515 unsigned int hint, depth;
516 unsigned long index, nr;
519 if (unlikely(sb->round_robin))
522 depth = READ_ONCE(sb->depth);
523 hint = update_alloc_hint_before_get(sb, depth);
525 index = SB_NR_TO_INDEX(sb, hint);
527 for (i = 0; i < sb->map_nr; i++) {
528 struct sbitmap_word *map = &sb->map[index];
529 unsigned long get_mask;
531 sbitmap_deferred_clear(map);
532 if (map->word == (1UL << (map->depth - 1)) - 1)
535 nr = find_first_zero_bit(&map->word, map->depth);
536 if (nr + nr_tags <= map->depth) {
537 atomic_long_t *ptr = (atomic_long_t *) &map->word;
538 int map_tags = min_t(int, nr_tags, map->depth);
539 unsigned long val, ret;
541 get_mask = ((1UL << map_tags) - 1) << nr;
543 val = READ_ONCE(map->word);
544 ret = atomic_long_cmpxchg(ptr, val, get_mask | val);
545 } while (ret != val);
546 get_mask = (get_mask & ~ret) >> nr;
548 *offset = nr + (index << sb->shift);
549 update_alloc_hint_after_get(sb, depth, hint,
550 *offset + map_tags - 1);
554 /* Jump to next index. */
555 if (++index >= sb->map_nr)
562 int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
563 unsigned int shallow_depth)
565 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
567 return sbitmap_get_shallow(&sbq->sb, shallow_depth);
569 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow);
571 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
572 unsigned int min_shallow_depth)
574 sbq->min_shallow_depth = min_shallow_depth;
575 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
577 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
579 static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
583 if (!atomic_read(&sbq->ws_active))
586 wake_index = atomic_read(&sbq->wake_index);
587 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
588 struct sbq_wait_state *ws = &sbq->ws[wake_index];
590 if (waitqueue_active(&ws->wait)) {
591 if (wake_index != atomic_read(&sbq->wake_index))
592 atomic_set(&sbq->wake_index, wake_index);
596 wake_index = sbq_index_inc(wake_index);
602 static bool __sbq_wake_up(struct sbitmap_queue *sbq)
604 struct sbq_wait_state *ws;
605 unsigned int wake_batch;
608 ws = sbq_wake_ptr(sbq);
612 wait_cnt = atomic_dec_return(&ws->wait_cnt);
616 wake_batch = READ_ONCE(sbq->wake_batch);
619 * Pairs with the memory barrier in sbitmap_queue_resize() to
620 * ensure that we see the batch size update before the wait
623 smp_mb__before_atomic();
626 * For concurrent callers of this, the one that failed the
627 * atomic_cmpxhcg() race should call this function again
628 * to wakeup a new batch on a different 'ws'.
630 ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch);
631 if (ret == wait_cnt) {
632 sbq_index_atomic_inc(&sbq->wake_index);
633 wake_up_nr(&ws->wait, wake_batch);
643 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
645 while (__sbq_wake_up(sbq))
648 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
650 static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
652 if (likely(!sb->round_robin && tag < sb->depth))
653 data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
656 void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
657 int *tags, int nr_tags)
659 struct sbitmap *sb = &sbq->sb;
660 unsigned long *addr = NULL;
661 unsigned long mask = 0;
664 smp_mb__before_atomic();
665 for (i = 0; i < nr_tags; i++) {
666 const int tag = tags[i] - offset;
667 unsigned long *this_addr;
669 /* since we're clearing a batch, skip the deferred map */
670 this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
673 } else if (addr != this_addr) {
674 atomic_long_andnot(mask, (atomic_long_t *) addr);
678 mask |= (1UL << SB_NR_TO_BIT(sb, tag));
682 atomic_long_andnot(mask, (atomic_long_t *) addr);
684 smp_mb__after_atomic();
685 sbitmap_queue_wake_up(sbq);
686 sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
687 tags[nr_tags - 1] - offset);
690 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
694 * Once the clear bit is set, the bit may be allocated out.
696 * Orders READ/WRITE on the associated instance(such as request
697 * of blk_mq) by this bit for avoiding race with re-allocation,
698 * and its pair is the memory barrier implied in __sbitmap_get_word.
700 * One invariant is that the clear bit has to be zero when the bit
703 smp_mb__before_atomic();
704 sbitmap_deferred_clear_bit(&sbq->sb, nr);
707 * Pairs with the memory barrier in set_current_state() to ensure the
708 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
709 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
710 * waiter. See the comment on waitqueue_active().
712 smp_mb__after_atomic();
713 sbitmap_queue_wake_up(sbq);
714 sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
716 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
718 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
723 * Pairs with the memory barrier in set_current_state() like in
724 * sbitmap_queue_wake_up().
727 wake_index = atomic_read(&sbq->wake_index);
728 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
729 struct sbq_wait_state *ws = &sbq->ws[wake_index];
731 if (waitqueue_active(&ws->wait))
734 wake_index = sbq_index_inc(wake_index);
737 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
739 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
744 sbitmap_show(&sbq->sb, m);
746 seq_puts(m, "alloc_hint={");
748 for_each_possible_cpu(i) {
752 seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
756 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
757 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
758 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
760 seq_puts(m, "ws={\n");
761 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
762 struct sbq_wait_state *ws = &sbq->ws[i];
764 seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
765 atomic_read(&ws->wait_cnt),
766 waitqueue_active(&ws->wait) ? "active" : "inactive");
770 seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
771 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
773 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
775 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
776 struct sbq_wait_state *ws,
777 struct sbq_wait *sbq_wait)
779 if (!sbq_wait->sbq) {
781 atomic_inc(&sbq->ws_active);
782 add_wait_queue(&ws->wait, &sbq_wait->wait);
785 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
787 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
789 list_del_init(&sbq_wait->wait.entry);
791 atomic_dec(&sbq_wait->sbq->ws_active);
792 sbq_wait->sbq = NULL;
795 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
797 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
798 struct sbq_wait_state *ws,
799 struct sbq_wait *sbq_wait, int state)
801 if (!sbq_wait->sbq) {
802 atomic_inc(&sbq->ws_active);
805 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
807 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
809 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
810 struct sbq_wait *sbq_wait)
812 finish_wait(&ws->wait, &sbq_wait->wait);
814 atomic_dec(&sbq->ws_active);
815 sbq_wait->sbq = NULL;
818 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);