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) = get_random_u32_below(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 ? get_random_u32_below(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;
90 shift = sbitmap_calculate_shift(depth);
92 bits_per_word = 1U << shift;
93 if (bits_per_word > BITS_PER_LONG)
98 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
99 sb->round_robin = round_robin;
107 if (init_alloc_hint(sb, flags))
110 sb->alloc_hint = NULL;
113 sb->map = kvzalloc_node(sb->map_nr * sizeof(*sb->map), flags, node);
115 free_percpu(sb->alloc_hint);
121 EXPORT_SYMBOL_GPL(sbitmap_init_node);
123 void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
125 unsigned int bits_per_word = 1U << sb->shift;
128 for (i = 0; i < sb->map_nr; i++)
129 sbitmap_deferred_clear(&sb->map[i]);
132 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
134 EXPORT_SYMBOL_GPL(sbitmap_resize);
136 static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
137 unsigned int hint, bool wrap)
141 /* don't wrap if starting from 0 */
145 nr = find_next_zero_bit(word, depth, hint);
146 if (unlikely(nr >= depth)) {
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
159 if (!test_and_set_bit_lock(nr, word))
163 if (hint >= depth - 1)
170 static int sbitmap_find_bit_in_word(struct sbitmap_word *map,
172 unsigned int alloc_hint,
178 nr = __sbitmap_get_word(&map->word, depth,
182 if (!sbitmap_deferred_clear(map))
189 static int sbitmap_find_bit(struct sbitmap *sb,
192 unsigned int alloc_hint,
198 for (i = 0; i < sb->map_nr; i++) {
199 nr = sbitmap_find_bit_in_word(&sb->map[index],
201 __map_depth(sb, index),
206 nr += index << sb->shift;
210 /* Jump to next index. */
212 if (++index >= sb->map_nr)
219 static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
223 index = SB_NR_TO_INDEX(sb, alloc_hint);
226 * Unless we're doing round robin tag allocation, just use the
227 * alloc_hint to find the right word index. No point in looping
228 * twice in find_next_zero_bit() for that case.
231 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
235 return sbitmap_find_bit(sb, UINT_MAX, index, alloc_hint,
239 int sbitmap_get(struct sbitmap *sb)
242 unsigned int hint, depth;
244 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
247 depth = READ_ONCE(sb->depth);
248 hint = update_alloc_hint_before_get(sb, depth);
249 nr = __sbitmap_get(sb, hint);
250 update_alloc_hint_after_get(sb, depth, hint, nr);
254 EXPORT_SYMBOL_GPL(sbitmap_get);
256 static int __sbitmap_get_shallow(struct sbitmap *sb,
257 unsigned int alloc_hint,
258 unsigned long shallow_depth)
262 index = SB_NR_TO_INDEX(sb, alloc_hint);
263 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
265 return sbitmap_find_bit(sb, shallow_depth, index, alloc_hint, true);
268 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
271 unsigned int hint, depth;
273 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
276 depth = READ_ONCE(sb->depth);
277 hint = update_alloc_hint_before_get(sb, depth);
278 nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
279 update_alloc_hint_after_get(sb, depth, hint, nr);
283 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
285 bool sbitmap_any_bit_set(const struct sbitmap *sb)
289 for (i = 0; i < sb->map_nr; i++) {
290 if (sb->map[i].word & ~sb->map[i].cleared)
295 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
297 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
299 unsigned int i, weight = 0;
301 for (i = 0; i < sb->map_nr; i++) {
302 const struct sbitmap_word *word = &sb->map[i];
303 unsigned int word_depth = __map_depth(sb, i);
306 weight += bitmap_weight(&word->word, word_depth);
308 weight += bitmap_weight(&word->cleared, word_depth);
313 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
315 return __sbitmap_weight(sb, false);
318 unsigned int sbitmap_weight(const struct sbitmap *sb)
320 return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
322 EXPORT_SYMBOL_GPL(sbitmap_weight);
324 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
326 seq_printf(m, "depth=%u\n", sb->depth);
327 seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
328 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
329 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
330 seq_printf(m, "map_nr=%u\n", sb->map_nr);
332 EXPORT_SYMBOL_GPL(sbitmap_show);
334 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
336 if ((offset & 0xf) == 0) {
339 seq_printf(m, "%08x:", offset);
341 if ((offset & 0x1) == 0)
343 seq_printf(m, "%02x", byte);
346 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
349 unsigned int byte_bits = 0;
350 unsigned int offset = 0;
353 for (i = 0; i < sb->map_nr; i++) {
354 unsigned long word = READ_ONCE(sb->map[i].word);
355 unsigned long cleared = READ_ONCE(sb->map[i].cleared);
356 unsigned int word_bits = __map_depth(sb, i);
360 while (word_bits > 0) {
361 unsigned int bits = min(8 - byte_bits, word_bits);
363 byte |= (word & (BIT(bits) - 1)) << byte_bits;
365 if (byte_bits == 8) {
366 emit_byte(m, offset, byte);
376 emit_byte(m, offset, byte);
382 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
384 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
387 unsigned int wake_batch;
388 unsigned int shallow_depth;
391 * For each batch, we wake up one queue. We need to make sure that our
392 * batch size is small enough that the full depth of the bitmap,
393 * potentially limited by a shallow depth, is enough to wake up all of
396 * Each full word of the bitmap has bits_per_word bits, and there might
397 * be a partial word. There are depth / bits_per_word full words and
398 * depth % bits_per_word bits left over. In bitwise arithmetic:
400 * bits_per_word = 1 << shift
401 * depth / bits_per_word = depth >> shift
402 * depth % bits_per_word = depth & ((1 << shift) - 1)
404 * Each word can be limited to sbq->min_shallow_depth bits.
406 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
407 depth = ((depth >> sbq->sb.shift) * shallow_depth +
408 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
409 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
415 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
416 int shift, bool round_robin, gfp_t flags, int node)
421 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
426 sbq->min_shallow_depth = UINT_MAX;
427 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
428 atomic_set(&sbq->wake_index, 0);
429 atomic_set(&sbq->ws_active, 0);
430 atomic_set(&sbq->completion_cnt, 0);
431 atomic_set(&sbq->wakeup_cnt, 0);
433 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
435 sbitmap_free(&sbq->sb);
439 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
440 init_waitqueue_head(&sbq->ws[i].wait);
444 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
446 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
449 unsigned int wake_batch;
451 wake_batch = sbq_calc_wake_batch(sbq, depth);
452 if (sbq->wake_batch != wake_batch)
453 WRITE_ONCE(sbq->wake_batch, wake_batch);
456 void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
459 unsigned int wake_batch;
460 unsigned int depth = (sbq->sb.depth + users - 1) / users;
462 wake_batch = clamp_val(depth / SBQ_WAIT_QUEUES,
465 WRITE_ONCE(sbq->wake_batch, wake_batch);
467 EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
469 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
471 sbitmap_queue_update_wake_batch(sbq, depth);
472 sbitmap_resize(&sbq->sb, depth);
474 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
476 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
478 return sbitmap_get(&sbq->sb);
480 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
482 unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
483 unsigned int *offset)
485 struct sbitmap *sb = &sbq->sb;
486 unsigned int hint, depth;
487 unsigned long index, nr;
490 if (unlikely(sb->round_robin))
493 depth = READ_ONCE(sb->depth);
494 hint = update_alloc_hint_before_get(sb, depth);
496 index = SB_NR_TO_INDEX(sb, hint);
498 for (i = 0; i < sb->map_nr; i++) {
499 struct sbitmap_word *map = &sb->map[index];
500 unsigned long get_mask;
501 unsigned int map_depth = __map_depth(sb, index);
503 sbitmap_deferred_clear(map);
504 if (map->word == (1UL << (map_depth - 1)) - 1)
507 nr = find_first_zero_bit(&map->word, map_depth);
508 if (nr + nr_tags <= map_depth) {
509 atomic_long_t *ptr = (atomic_long_t *) &map->word;
512 get_mask = ((1UL << nr_tags) - 1) << nr;
513 val = READ_ONCE(map->word);
514 while (!atomic_long_try_cmpxchg(ptr, &val,
517 get_mask = (get_mask & ~val) >> nr;
519 *offset = nr + (index << sb->shift);
520 update_alloc_hint_after_get(sb, depth, hint,
521 *offset + nr_tags - 1);
526 /* Jump to next index. */
527 if (++index >= sb->map_nr)
534 int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
535 unsigned int shallow_depth)
537 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
539 return sbitmap_get_shallow(&sbq->sb, shallow_depth);
541 EXPORT_SYMBOL_GPL(sbitmap_queue_get_shallow);
543 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
544 unsigned int min_shallow_depth)
546 sbq->min_shallow_depth = min_shallow_depth;
547 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
549 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
551 static void __sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
555 if (!atomic_read(&sbq->ws_active))
558 wake_index = atomic_read(&sbq->wake_index);
559 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
560 struct sbq_wait_state *ws = &sbq->ws[wake_index];
563 * Advance the index before checking the current queue.
564 * It improves fairness, by ensuring the queue doesn't
565 * need to be fully emptied before trying to wake up
568 wake_index = sbq_index_inc(wake_index);
571 * It is sufficient to wake up at least one waiter to
572 * guarantee forward progress.
574 if (waitqueue_active(&ws->wait) &&
575 wake_up_nr(&ws->wait, nr))
579 if (wake_index != atomic_read(&sbq->wake_index))
580 atomic_set(&sbq->wake_index, wake_index);
583 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
585 unsigned int wake_batch = READ_ONCE(sbq->wake_batch);
586 unsigned int wakeups;
588 if (!atomic_read(&sbq->ws_active))
591 atomic_add(nr, &sbq->completion_cnt);
592 wakeups = atomic_read(&sbq->wakeup_cnt);
595 if (atomic_read(&sbq->completion_cnt) - wakeups < wake_batch)
597 } while (!atomic_try_cmpxchg(&sbq->wakeup_cnt,
598 &wakeups, wakeups + wake_batch));
600 __sbitmap_queue_wake_up(sbq, wake_batch);
602 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
604 static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
606 if (likely(!sb->round_robin && tag < sb->depth))
607 data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
610 void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
611 int *tags, int nr_tags)
613 struct sbitmap *sb = &sbq->sb;
614 unsigned long *addr = NULL;
615 unsigned long mask = 0;
618 smp_mb__before_atomic();
619 for (i = 0; i < nr_tags; i++) {
620 const int tag = tags[i] - offset;
621 unsigned long *this_addr;
623 /* since we're clearing a batch, skip the deferred map */
624 this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
627 } else if (addr != this_addr) {
628 atomic_long_andnot(mask, (atomic_long_t *) addr);
632 mask |= (1UL << SB_NR_TO_BIT(sb, tag));
636 atomic_long_andnot(mask, (atomic_long_t *) addr);
638 smp_mb__after_atomic();
639 sbitmap_queue_wake_up(sbq, nr_tags);
640 sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
641 tags[nr_tags - 1] - offset);
644 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
648 * Once the clear bit is set, the bit may be allocated out.
650 * Orders READ/WRITE on the associated instance(such as request
651 * of blk_mq) by this bit for avoiding race with re-allocation,
652 * and its pair is the memory barrier implied in __sbitmap_get_word.
654 * One invariant is that the clear bit has to be zero when the bit
657 smp_mb__before_atomic();
658 sbitmap_deferred_clear_bit(&sbq->sb, nr);
661 * Pairs with the memory barrier in set_current_state() to ensure the
662 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
663 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
664 * waiter. See the comment on waitqueue_active().
666 smp_mb__after_atomic();
667 sbitmap_queue_wake_up(sbq, 1);
668 sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
670 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
672 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
677 * Pairs with the memory barrier in set_current_state() like in
678 * sbitmap_queue_wake_up().
681 wake_index = atomic_read(&sbq->wake_index);
682 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
683 struct sbq_wait_state *ws = &sbq->ws[wake_index];
685 if (waitqueue_active(&ws->wait))
688 wake_index = sbq_index_inc(wake_index);
691 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
693 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
698 sbitmap_show(&sbq->sb, m);
700 seq_puts(m, "alloc_hint={");
702 for_each_possible_cpu(i) {
706 seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
710 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
711 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
712 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
714 seq_puts(m, "ws={\n");
715 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
716 struct sbq_wait_state *ws = &sbq->ws[i];
717 seq_printf(m, "\t{.wait=%s},\n",
718 waitqueue_active(&ws->wait) ? "active" : "inactive");
722 seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
723 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
725 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
727 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
728 struct sbq_wait_state *ws,
729 struct sbq_wait *sbq_wait)
731 if (!sbq_wait->sbq) {
733 atomic_inc(&sbq->ws_active);
734 add_wait_queue(&ws->wait, &sbq_wait->wait);
737 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
739 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
741 list_del_init(&sbq_wait->wait.entry);
743 atomic_dec(&sbq_wait->sbq->ws_active);
744 sbq_wait->sbq = NULL;
747 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
749 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
750 struct sbq_wait_state *ws,
751 struct sbq_wait *sbq_wait, int state)
753 if (!sbq_wait->sbq) {
754 atomic_inc(&sbq->ws_active);
757 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
759 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
761 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
762 struct sbq_wait *sbq_wait)
764 finish_wait(&ws->wait, &sbq_wait->wait);
766 atomic_dec(&sbq->ws_active);
767 sbq_wait->sbq = NULL;
770 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);