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 = ilog2(BITS_PER_LONG);
93 * If the bitmap is small, shrink the number of bits per word so
94 * we spread over a few cachelines, at least. If less than 4
95 * bits, just forget about it, it's not going to work optimally
99 while ((4U << shift) > depth)
103 bits_per_word = 1U << shift;
104 if (bits_per_word > BITS_PER_LONG)
109 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
110 sb->round_robin = round_robin;
118 if (init_alloc_hint(sb, flags))
121 sb->alloc_hint = NULL;
124 sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node);
126 free_percpu(sb->alloc_hint);
130 for (i = 0; i < sb->map_nr; i++) {
131 sb->map[i].depth = min(depth, bits_per_word);
132 depth -= sb->map[i].depth;
136 EXPORT_SYMBOL_GPL(sbitmap_init_node);
138 void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
140 unsigned int bits_per_word = 1U << sb->shift;
143 for (i = 0; i < sb->map_nr; i++)
144 sbitmap_deferred_clear(&sb->map[i]);
147 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
149 for (i = 0; i < sb->map_nr; i++) {
150 sb->map[i].depth = min(depth, bits_per_word);
151 depth -= sb->map[i].depth;
154 EXPORT_SYMBOL_GPL(sbitmap_resize);
156 static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
157 unsigned int hint, bool wrap)
161 /* don't wrap if starting from 0 */
165 nr = find_next_zero_bit(word, depth, hint);
166 if (unlikely(nr >= depth)) {
168 * We started with an offset, and we didn't reset the
169 * offset to 0 in a failure case, so start from 0 to
179 if (!test_and_set_bit_lock(nr, word))
183 if (hint >= depth - 1)
190 static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
191 unsigned int alloc_hint)
193 struct sbitmap_word *map = &sb->map[index];
197 nr = __sbitmap_get_word(&map->word, map->depth, alloc_hint,
201 if (!sbitmap_deferred_clear(map))
208 static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
210 unsigned int i, index;
213 index = SB_NR_TO_INDEX(sb, alloc_hint);
216 * Unless we're doing round robin tag allocation, just use the
217 * alloc_hint to find the right word index. No point in looping
218 * twice in find_next_zero_bit() for that case.
221 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
225 for (i = 0; i < sb->map_nr; i++) {
226 nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
228 nr += index << sb->shift;
232 /* Jump to next index. */
234 if (++index >= sb->map_nr)
241 int sbitmap_get(struct sbitmap *sb)
244 unsigned int hint, depth;
246 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
249 depth = READ_ONCE(sb->depth);
250 hint = update_alloc_hint_before_get(sb, depth);
251 nr = __sbitmap_get(sb, hint);
252 update_alloc_hint_after_get(sb, depth, hint, nr);
256 EXPORT_SYMBOL_GPL(sbitmap_get);
258 static int __sbitmap_get_shallow(struct sbitmap *sb,
259 unsigned int alloc_hint,
260 unsigned long shallow_depth)
262 unsigned int i, index;
265 index = SB_NR_TO_INDEX(sb, alloc_hint);
267 for (i = 0; i < sb->map_nr; i++) {
269 nr = __sbitmap_get_word(&sb->map[index].word,
270 min(sb->map[index].depth, shallow_depth),
271 SB_NR_TO_BIT(sb, alloc_hint), true);
273 nr += index << sb->shift;
277 if (sbitmap_deferred_clear(&sb->map[index]))
280 /* Jump to next index. */
282 alloc_hint = index << sb->shift;
284 if (index >= sb->map_nr) {
293 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
296 unsigned int hint, depth;
298 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
301 depth = READ_ONCE(sb->depth);
302 hint = update_alloc_hint_before_get(sb, depth);
303 nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
304 update_alloc_hint_after_get(sb, depth, hint, nr);
308 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
310 bool sbitmap_any_bit_set(const struct sbitmap *sb)
314 for (i = 0; i < sb->map_nr; i++) {
315 if (sb->map[i].word & ~sb->map[i].cleared)
320 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
322 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
324 unsigned int i, weight = 0;
326 for (i = 0; i < sb->map_nr; i++) {
327 const struct sbitmap_word *word = &sb->map[i];
330 weight += bitmap_weight(&word->word, word->depth);
332 weight += bitmap_weight(&word->cleared, word->depth);
337 static unsigned int sbitmap_weight(const struct sbitmap *sb)
339 return __sbitmap_weight(sb, true);
342 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
344 return __sbitmap_weight(sb, false);
347 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
349 seq_printf(m, "depth=%u\n", sb->depth);
350 seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb));
351 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
352 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
353 seq_printf(m, "map_nr=%u\n", sb->map_nr);
355 EXPORT_SYMBOL_GPL(sbitmap_show);
357 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
359 if ((offset & 0xf) == 0) {
362 seq_printf(m, "%08x:", offset);
364 if ((offset & 0x1) == 0)
366 seq_printf(m, "%02x", byte);
369 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
372 unsigned int byte_bits = 0;
373 unsigned int offset = 0;
376 for (i = 0; i < sb->map_nr; i++) {
377 unsigned long word = READ_ONCE(sb->map[i].word);
378 unsigned long cleared = READ_ONCE(sb->map[i].cleared);
379 unsigned int word_bits = READ_ONCE(sb->map[i].depth);
383 while (word_bits > 0) {
384 unsigned int bits = min(8 - byte_bits, word_bits);
386 byte |= (word & (BIT(bits) - 1)) << byte_bits;
388 if (byte_bits == 8) {
389 emit_byte(m, offset, byte);
399 emit_byte(m, offset, byte);
405 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
407 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
410 unsigned int wake_batch;
411 unsigned int shallow_depth;
414 * For each batch, we wake up one queue. We need to make sure that our
415 * batch size is small enough that the full depth of the bitmap,
416 * potentially limited by a shallow depth, is enough to wake up all of
419 * Each full word of the bitmap has bits_per_word bits, and there might
420 * be a partial word. There are depth / bits_per_word full words and
421 * depth % bits_per_word bits left over. In bitwise arithmetic:
423 * bits_per_word = 1 << shift
424 * depth / bits_per_word = depth >> shift
425 * depth % bits_per_word = depth & ((1 << shift) - 1)
427 * Each word can be limited to sbq->min_shallow_depth bits.
429 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
430 depth = ((depth >> sbq->sb.shift) * shallow_depth +
431 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
432 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
438 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
439 int shift, bool round_robin, gfp_t flags, int node)
444 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
449 sbq->min_shallow_depth = UINT_MAX;
450 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
451 atomic_set(&sbq->wake_index, 0);
452 atomic_set(&sbq->ws_active, 0);
454 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
456 sbitmap_free(&sbq->sb);
460 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
461 init_waitqueue_head(&sbq->ws[i].wait);
462 atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
467 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
469 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
472 unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth);
475 if (sbq->wake_batch != wake_batch) {
476 WRITE_ONCE(sbq->wake_batch, wake_batch);
478 * Pairs with the memory barrier in sbitmap_queue_wake_up()
479 * to ensure that the batch size is updated before the wait
483 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
484 atomic_set(&sbq->ws[i].wait_cnt, 1);
488 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
490 sbitmap_queue_update_wake_batch(sbq, depth);
491 sbitmap_resize(&sbq->sb, depth);
493 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
495 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
497 return sbitmap_get(&sbq->sb);
499 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
501 int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
502 unsigned int shallow_depth)
504 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
506 return sbitmap_get_shallow(&sbq->sb, shallow_depth);
508 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow);
510 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
511 unsigned int min_shallow_depth)
513 sbq->min_shallow_depth = min_shallow_depth;
514 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
516 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
518 static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
522 if (!atomic_read(&sbq->ws_active))
525 wake_index = atomic_read(&sbq->wake_index);
526 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
527 struct sbq_wait_state *ws = &sbq->ws[wake_index];
529 if (waitqueue_active(&ws->wait)) {
530 if (wake_index != atomic_read(&sbq->wake_index))
531 atomic_set(&sbq->wake_index, wake_index);
535 wake_index = sbq_index_inc(wake_index);
541 static bool __sbq_wake_up(struct sbitmap_queue *sbq)
543 struct sbq_wait_state *ws;
544 unsigned int wake_batch;
547 ws = sbq_wake_ptr(sbq);
551 wait_cnt = atomic_dec_return(&ws->wait_cnt);
555 wake_batch = READ_ONCE(sbq->wake_batch);
558 * Pairs with the memory barrier in sbitmap_queue_resize() to
559 * ensure that we see the batch size update before the wait
562 smp_mb__before_atomic();
565 * For concurrent callers of this, the one that failed the
566 * atomic_cmpxhcg() race should call this function again
567 * to wakeup a new batch on a different 'ws'.
569 ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch);
570 if (ret == wait_cnt) {
571 sbq_index_atomic_inc(&sbq->wake_index);
572 wake_up_nr(&ws->wait, wake_batch);
582 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
584 while (__sbq_wake_up(sbq))
587 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
589 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
593 * Once the clear bit is set, the bit may be allocated out.
595 * Orders READ/WRITE on the asssociated instance(such as request
596 * of blk_mq) by this bit for avoiding race with re-allocation,
597 * and its pair is the memory barrier implied in __sbitmap_get_word.
599 * One invariant is that the clear bit has to be zero when the bit
602 smp_mb__before_atomic();
603 sbitmap_deferred_clear_bit(&sbq->sb, nr);
606 * Pairs with the memory barrier in set_current_state() to ensure the
607 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
608 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
609 * waiter. See the comment on waitqueue_active().
611 smp_mb__after_atomic();
612 sbitmap_queue_wake_up(sbq);
614 if (likely(!sbq->sb.round_robin && nr < sbq->sb.depth))
615 *per_cpu_ptr(sbq->sb.alloc_hint, cpu) = nr;
617 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
619 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
624 * Pairs with the memory barrier in set_current_state() like in
625 * sbitmap_queue_wake_up().
628 wake_index = atomic_read(&sbq->wake_index);
629 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
630 struct sbq_wait_state *ws = &sbq->ws[wake_index];
632 if (waitqueue_active(&ws->wait))
635 wake_index = sbq_index_inc(wake_index);
638 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
640 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
645 sbitmap_show(&sbq->sb, m);
647 seq_puts(m, "alloc_hint={");
649 for_each_possible_cpu(i) {
653 seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
657 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
658 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
659 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
661 seq_puts(m, "ws={\n");
662 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
663 struct sbq_wait_state *ws = &sbq->ws[i];
665 seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
666 atomic_read(&ws->wait_cnt),
667 waitqueue_active(&ws->wait) ? "active" : "inactive");
671 seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
672 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
674 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
676 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
677 struct sbq_wait_state *ws,
678 struct sbq_wait *sbq_wait)
680 if (!sbq_wait->sbq) {
682 atomic_inc(&sbq->ws_active);
683 add_wait_queue(&ws->wait, &sbq_wait->wait);
686 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
688 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
690 list_del_init(&sbq_wait->wait.entry);
692 atomic_dec(&sbq_wait->sbq->ws_active);
693 sbq_wait->sbq = NULL;
696 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
698 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
699 struct sbq_wait_state *ws,
700 struct sbq_wait *sbq_wait, int state)
702 if (!sbq_wait->sbq) {
703 atomic_inc(&sbq->ws_active);
706 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
708 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
710 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
711 struct sbq_wait *sbq_wait)
713 finish_wait(&ws->wait, &sbq_wait->wait);
715 atomic_dec(&sbq->ws_active);
716 sbq_wait->sbq = NULL;
719 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);