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_max(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_max(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_index(struct sbitmap *sb, int index,
171 unsigned int alloc_hint)
173 struct sbitmap_word *map = &sb->map[index];
177 nr = __sbitmap_get_word(&map->word, __map_depth(sb, index),
178 alloc_hint, !sb->round_robin);
181 if (!sbitmap_deferred_clear(map))
188 static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
190 unsigned int i, index;
193 index = SB_NR_TO_INDEX(sb, alloc_hint);
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.
201 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
205 for (i = 0; i < sb->map_nr; i++) {
206 nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
208 nr += index << sb->shift;
212 /* Jump to next index. */
214 if (++index >= sb->map_nr)
221 int sbitmap_get(struct sbitmap *sb)
224 unsigned int hint, depth;
226 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
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);
236 EXPORT_SYMBOL_GPL(sbitmap_get);
238 static int __sbitmap_get_shallow(struct sbitmap *sb,
239 unsigned int alloc_hint,
240 unsigned long shallow_depth)
242 unsigned int i, index;
245 index = SB_NR_TO_INDEX(sb, alloc_hint);
247 for (i = 0; i < sb->map_nr; i++) {
249 nr = __sbitmap_get_word(&sb->map[index].word,
251 __map_depth(sb, index),
253 SB_NR_TO_BIT(sb, alloc_hint), true);
255 nr += index << sb->shift;
259 if (sbitmap_deferred_clear(&sb->map[index]))
262 /* Jump to next index. */
264 alloc_hint = index << sb->shift;
266 if (index >= sb->map_nr) {
275 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
278 unsigned int hint, depth;
280 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
283 depth = READ_ONCE(sb->depth);
284 hint = update_alloc_hint_before_get(sb, depth);
285 nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
286 update_alloc_hint_after_get(sb, depth, hint, nr);
290 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
292 bool sbitmap_any_bit_set(const struct sbitmap *sb)
296 for (i = 0; i < sb->map_nr; i++) {
297 if (sb->map[i].word & ~sb->map[i].cleared)
302 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
304 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
306 unsigned int i, weight = 0;
308 for (i = 0; i < sb->map_nr; i++) {
309 const struct sbitmap_word *word = &sb->map[i];
310 unsigned int word_depth = __map_depth(sb, i);
313 weight += bitmap_weight(&word->word, word_depth);
315 weight += bitmap_weight(&word->cleared, word_depth);
320 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
322 return __sbitmap_weight(sb, false);
325 unsigned int sbitmap_weight(const struct sbitmap *sb)
327 return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
329 EXPORT_SYMBOL_GPL(sbitmap_weight);
331 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
333 seq_printf(m, "depth=%u\n", sb->depth);
334 seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
335 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
336 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
337 seq_printf(m, "map_nr=%u\n", sb->map_nr);
339 EXPORT_SYMBOL_GPL(sbitmap_show);
341 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
343 if ((offset & 0xf) == 0) {
346 seq_printf(m, "%08x:", offset);
348 if ((offset & 0x1) == 0)
350 seq_printf(m, "%02x", byte);
353 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
356 unsigned int byte_bits = 0;
357 unsigned int offset = 0;
360 for (i = 0; i < sb->map_nr; i++) {
361 unsigned long word = READ_ONCE(sb->map[i].word);
362 unsigned long cleared = READ_ONCE(sb->map[i].cleared);
363 unsigned int word_bits = __map_depth(sb, i);
367 while (word_bits > 0) {
368 unsigned int bits = min(8 - byte_bits, word_bits);
370 byte |= (word & (BIT(bits) - 1)) << byte_bits;
372 if (byte_bits == 8) {
373 emit_byte(m, offset, byte);
383 emit_byte(m, offset, byte);
389 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
391 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
394 unsigned int wake_batch;
395 unsigned int shallow_depth;
398 * For each batch, we wake up one queue. We need to make sure that our
399 * batch size is small enough that the full depth of the bitmap,
400 * potentially limited by a shallow depth, is enough to wake up all of
403 * Each full word of the bitmap has bits_per_word bits, and there might
404 * be a partial word. There are depth / bits_per_word full words and
405 * depth % bits_per_word bits left over. In bitwise arithmetic:
407 * bits_per_word = 1 << shift
408 * depth / bits_per_word = depth >> shift
409 * depth % bits_per_word = depth & ((1 << shift) - 1)
411 * Each word can be limited to sbq->min_shallow_depth bits.
413 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
414 depth = ((depth >> sbq->sb.shift) * shallow_depth +
415 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
416 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
422 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
423 int shift, bool round_robin, gfp_t flags, int node)
428 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
433 sbq->min_shallow_depth = UINT_MAX;
434 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
435 atomic_set(&sbq->wake_index, 0);
436 atomic_set(&sbq->ws_active, 0);
438 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
440 sbitmap_free(&sbq->sb);
444 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
445 init_waitqueue_head(&sbq->ws[i].wait);
446 atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
451 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
453 static inline void __sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
454 unsigned int wake_batch)
458 if (sbq->wake_batch != wake_batch) {
459 WRITE_ONCE(sbq->wake_batch, wake_batch);
461 * Pairs with the memory barrier in sbitmap_queue_wake_up()
462 * to ensure that the batch size is updated before the wait
466 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
467 atomic_set(&sbq->ws[i].wait_cnt, 1);
471 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
474 unsigned int wake_batch;
476 wake_batch = sbq_calc_wake_batch(sbq, depth);
477 __sbitmap_queue_update_wake_batch(sbq, wake_batch);
480 void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
483 unsigned int wake_batch;
484 unsigned int min_batch;
485 unsigned int depth = (sbq->sb.depth + users - 1) / users;
487 min_batch = sbq->sb.depth >= (4 * SBQ_WAIT_QUEUES) ? 4 : 1;
489 wake_batch = clamp_val(depth / SBQ_WAIT_QUEUES,
490 min_batch, SBQ_WAKE_BATCH);
491 __sbitmap_queue_update_wake_batch(sbq, wake_batch);
493 EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
495 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
497 sbitmap_queue_update_wake_batch(sbq, depth);
498 sbitmap_resize(&sbq->sb, depth);
500 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
502 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
504 return sbitmap_get(&sbq->sb);
506 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
508 unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
509 unsigned int *offset)
511 struct sbitmap *sb = &sbq->sb;
512 unsigned int hint, depth;
513 unsigned long index, nr;
516 if (unlikely(sb->round_robin))
519 depth = READ_ONCE(sb->depth);
520 hint = update_alloc_hint_before_get(sb, depth);
522 index = SB_NR_TO_INDEX(sb, hint);
524 for (i = 0; i < sb->map_nr; i++) {
525 struct sbitmap_word *map = &sb->map[index];
526 unsigned long get_mask;
527 unsigned int map_depth = __map_depth(sb, index);
529 sbitmap_deferred_clear(map);
530 if (map->word == (1UL << (map_depth - 1)) - 1)
533 nr = find_first_zero_bit(&map->word, map_depth);
534 if (nr + nr_tags <= map_depth) {
535 atomic_long_t *ptr = (atomic_long_t *) &map->word;
538 get_mask = ((1UL << nr_tags) - 1) << nr;
539 val = READ_ONCE(map->word);
541 if ((val & ~get_mask) != val)
543 } while (!atomic_long_try_cmpxchg(ptr, &val,
545 get_mask = (get_mask & ~val) >> nr;
547 *offset = nr + (index << sb->shift);
548 update_alloc_hint_after_get(sb, depth, hint,
549 *offset + nr_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) && atomic_read(&ws->wait_cnt)) {
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, int *nr)
604 struct sbq_wait_state *ws;
605 unsigned int wake_batch;
606 int wait_cnt, cur, sub;
612 ws = sbq_wake_ptr(sbq);
616 cur = atomic_read(&ws->wait_cnt);
619 * For concurrent callers of this, callers should call this
620 * function again to wakeup a new batch on a different 'ws'.
625 wait_cnt = cur - sub;
626 } while (!atomic_try_cmpxchg(&ws->wait_cnt, &cur, wait_cnt));
629 * If we decremented queue without waiters, retry to avoid lost
633 return !waitqueue_active(&ws->wait);
638 * When wait_cnt == 0, we have to be particularly careful as we are
639 * responsible to reset wait_cnt regardless whether we've actually
640 * woken up anybody. But in case we didn't wakeup anybody, we still
643 ret = !waitqueue_active(&ws->wait);
644 wake_batch = READ_ONCE(sbq->wake_batch);
647 * Wake up first in case that concurrent callers decrease wait_cnt
648 * while waitqueue is empty.
650 wake_up_nr(&ws->wait, wake_batch);
653 * Pairs with the memory barrier in sbitmap_queue_resize() to
654 * ensure that we see the batch size update before the wait
657 * Also pairs with the implicit barrier between decrementing wait_cnt
658 * and checking for waitqueue_active() to make sure waitqueue_active()
659 * sees result of the wakeup if atomic_dec_return() has seen the result
662 smp_mb__before_atomic();
665 * Increase wake_index before updating wait_cnt, otherwise concurrent
666 * callers can see valid wait_cnt in old waitqueue, which can cause
667 * invalid wakeup on the old waitqueue.
669 sbq_index_atomic_inc(&sbq->wake_index);
670 atomic_set(&ws->wait_cnt, wake_batch);
675 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
677 while (__sbq_wake_up(sbq, &nr))
680 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
682 static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
684 if (likely(!sb->round_robin && tag < sb->depth))
685 data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
688 void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
689 int *tags, int nr_tags)
691 struct sbitmap *sb = &sbq->sb;
692 unsigned long *addr = NULL;
693 unsigned long mask = 0;
696 smp_mb__before_atomic();
697 for (i = 0; i < nr_tags; i++) {
698 const int tag = tags[i] - offset;
699 unsigned long *this_addr;
701 /* since we're clearing a batch, skip the deferred map */
702 this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
705 } else if (addr != this_addr) {
706 atomic_long_andnot(mask, (atomic_long_t *) addr);
710 mask |= (1UL << SB_NR_TO_BIT(sb, tag));
714 atomic_long_andnot(mask, (atomic_long_t *) addr);
716 smp_mb__after_atomic();
717 sbitmap_queue_wake_up(sbq, nr_tags);
718 sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
719 tags[nr_tags - 1] - offset);
722 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
726 * Once the clear bit is set, the bit may be allocated out.
728 * Orders READ/WRITE on the associated instance(such as request
729 * of blk_mq) by this bit for avoiding race with re-allocation,
730 * and its pair is the memory barrier implied in __sbitmap_get_word.
732 * One invariant is that the clear bit has to be zero when the bit
735 smp_mb__before_atomic();
736 sbitmap_deferred_clear_bit(&sbq->sb, nr);
739 * Pairs with the memory barrier in set_current_state() to ensure the
740 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
741 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
742 * waiter. See the comment on waitqueue_active().
744 smp_mb__after_atomic();
745 sbitmap_queue_wake_up(sbq, 1);
746 sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
748 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
750 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
755 * Pairs with the memory barrier in set_current_state() like in
756 * sbitmap_queue_wake_up().
759 wake_index = atomic_read(&sbq->wake_index);
760 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
761 struct sbq_wait_state *ws = &sbq->ws[wake_index];
763 if (waitqueue_active(&ws->wait))
766 wake_index = sbq_index_inc(wake_index);
769 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
771 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
776 sbitmap_show(&sbq->sb, m);
778 seq_puts(m, "alloc_hint={");
780 for_each_possible_cpu(i) {
784 seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
788 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
789 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
790 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
792 seq_puts(m, "ws={\n");
793 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
794 struct sbq_wait_state *ws = &sbq->ws[i];
796 seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
797 atomic_read(&ws->wait_cnt),
798 waitqueue_active(&ws->wait) ? "active" : "inactive");
802 seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
803 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
805 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
807 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
808 struct sbq_wait_state *ws,
809 struct sbq_wait *sbq_wait)
811 if (!sbq_wait->sbq) {
813 atomic_inc(&sbq->ws_active);
814 add_wait_queue(&ws->wait, &sbq_wait->wait);
817 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
819 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
821 list_del_init(&sbq_wait->wait.entry);
823 atomic_dec(&sbq_wait->sbq->ws_active);
824 sbq_wait->sbq = NULL;
827 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
829 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
830 struct sbq_wait_state *ws,
831 struct sbq_wait *sbq_wait, int state)
833 if (!sbq_wait->sbq) {
834 atomic_inc(&sbq->ws_active);
837 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
839 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
841 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
842 struct sbq_wait *sbq_wait)
844 finish_wait(&ws->wait, &sbq_wait->wait);
846 atomic_dec(&sbq->ws_active);
847 sbq_wait->sbq = NULL;
850 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);