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>
13 * See if we have deferred clears that we can batch move
15 static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
17 unsigned long mask, val;
21 spin_lock_irqsave(&sb->map[index].swap_lock, flags);
23 if (!sb->map[index].cleared)
27 * First get a stable cleared mask, setting the old mask to 0.
29 mask = xchg(&sb->map[index].cleared, 0);
32 * Now clear the masked bits in our free word
35 val = sb->map[index].word;
36 } while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val);
40 spin_unlock_irqrestore(&sb->map[index].swap_lock, flags);
44 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
45 gfp_t flags, int node)
47 unsigned int bits_per_word;
51 shift = ilog2(BITS_PER_LONG);
53 * If the bitmap is small, shrink the number of bits per word so
54 * we spread over a few cachelines, at least. If less than 4
55 * bits, just forget about it, it's not going to work optimally
59 while ((4U << shift) > depth)
63 bits_per_word = 1U << shift;
64 if (bits_per_word > BITS_PER_LONG)
69 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
76 sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node);
80 for (i = 0; i < sb->map_nr; i++) {
81 sb->map[i].depth = min(depth, bits_per_word);
82 depth -= sb->map[i].depth;
83 spin_lock_init(&sb->map[i].swap_lock);
87 EXPORT_SYMBOL_GPL(sbitmap_init_node);
89 void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
91 unsigned int bits_per_word = 1U << sb->shift;
94 for (i = 0; i < sb->map_nr; i++)
95 sbitmap_deferred_clear(sb, i);
98 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
100 for (i = 0; i < sb->map_nr; i++) {
101 sb->map[i].depth = min(depth, bits_per_word);
102 depth -= sb->map[i].depth;
105 EXPORT_SYMBOL_GPL(sbitmap_resize);
107 static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
108 unsigned int hint, bool wrap)
110 unsigned int orig_hint = hint;
114 nr = find_next_zero_bit(word, depth, hint);
115 if (unlikely(nr >= depth)) {
117 * We started with an offset, and we didn't reset the
118 * offset to 0 in a failure case, so start from 0 to
121 if (orig_hint && hint && wrap) {
122 hint = orig_hint = 0;
128 if (!test_and_set_bit_lock(nr, word))
132 if (hint >= depth - 1)
139 static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
140 unsigned int alloc_hint, bool round_robin)
145 nr = __sbitmap_get_word(&sb->map[index].word,
146 sb->map[index].depth, alloc_hint,
150 if (!sbitmap_deferred_clear(sb, index))
157 int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
159 unsigned int i, index;
162 index = SB_NR_TO_INDEX(sb, alloc_hint);
165 * Unless we're doing round robin tag allocation, just use the
166 * alloc_hint to find the right word index. No point in looping
167 * twice in find_next_zero_bit() for that case.
170 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
174 for (i = 0; i < sb->map_nr; i++) {
175 nr = sbitmap_find_bit_in_index(sb, index, alloc_hint,
178 nr += index << sb->shift;
182 /* Jump to next index. */
184 if (++index >= sb->map_nr)
190 EXPORT_SYMBOL_GPL(sbitmap_get);
192 int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint,
193 unsigned long shallow_depth)
195 unsigned int i, index;
198 index = SB_NR_TO_INDEX(sb, alloc_hint);
200 for (i = 0; i < sb->map_nr; i++) {
202 nr = __sbitmap_get_word(&sb->map[index].word,
203 min(sb->map[index].depth, shallow_depth),
204 SB_NR_TO_BIT(sb, alloc_hint), true);
206 nr += index << sb->shift;
210 if (sbitmap_deferred_clear(sb, index))
213 /* Jump to next index. */
215 alloc_hint = index << sb->shift;
217 if (index >= sb->map_nr) {
225 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
227 bool sbitmap_any_bit_set(const struct sbitmap *sb)
231 for (i = 0; i < sb->map_nr; i++) {
232 if (sb->map[i].word & ~sb->map[i].cleared)
237 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
239 bool sbitmap_any_bit_clear(const struct sbitmap *sb)
243 for (i = 0; i < sb->map_nr; i++) {
244 const struct sbitmap_word *word = &sb->map[i];
245 unsigned long mask = word->word & ~word->cleared;
248 ret = find_first_zero_bit(&mask, word->depth);
249 if (ret < word->depth)
254 EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear);
256 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
258 unsigned int i, weight = 0;
260 for (i = 0; i < sb->map_nr; i++) {
261 const struct sbitmap_word *word = &sb->map[i];
264 weight += bitmap_weight(&word->word, word->depth);
266 weight += bitmap_weight(&word->cleared, word->depth);
271 static unsigned int sbitmap_weight(const struct sbitmap *sb)
273 return __sbitmap_weight(sb, true);
276 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
278 return __sbitmap_weight(sb, false);
281 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
283 seq_printf(m, "depth=%u\n", sb->depth);
284 seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb));
285 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
286 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
287 seq_printf(m, "map_nr=%u\n", sb->map_nr);
289 EXPORT_SYMBOL_GPL(sbitmap_show);
291 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
293 if ((offset & 0xf) == 0) {
296 seq_printf(m, "%08x:", offset);
298 if ((offset & 0x1) == 0)
300 seq_printf(m, "%02x", byte);
303 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
306 unsigned int byte_bits = 0;
307 unsigned int offset = 0;
310 for (i = 0; i < sb->map_nr; i++) {
311 unsigned long word = READ_ONCE(sb->map[i].word);
312 unsigned int word_bits = READ_ONCE(sb->map[i].depth);
314 while (word_bits > 0) {
315 unsigned int bits = min(8 - byte_bits, word_bits);
317 byte |= (word & (BIT(bits) - 1)) << byte_bits;
319 if (byte_bits == 8) {
320 emit_byte(m, offset, byte);
330 emit_byte(m, offset, byte);
336 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
338 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
341 unsigned int wake_batch;
342 unsigned int shallow_depth;
345 * For each batch, we wake up one queue. We need to make sure that our
346 * batch size is small enough that the full depth of the bitmap,
347 * potentially limited by a shallow depth, is enough to wake up all of
350 * Each full word of the bitmap has bits_per_word bits, and there might
351 * be a partial word. There are depth / bits_per_word full words and
352 * depth % bits_per_word bits left over. In bitwise arithmetic:
354 * bits_per_word = 1 << shift
355 * depth / bits_per_word = depth >> shift
356 * depth % bits_per_word = depth & ((1 << shift) - 1)
358 * Each word can be limited to sbq->min_shallow_depth bits.
360 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
361 depth = ((depth >> sbq->sb.shift) * shallow_depth +
362 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
363 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
369 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
370 int shift, bool round_robin, gfp_t flags, int node)
375 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node);
379 sbq->alloc_hint = alloc_percpu_gfp(unsigned int, flags);
380 if (!sbq->alloc_hint) {
381 sbitmap_free(&sbq->sb);
385 if (depth && !round_robin) {
386 for_each_possible_cpu(i)
387 *per_cpu_ptr(sbq->alloc_hint, i) = prandom_u32() % depth;
390 sbq->min_shallow_depth = UINT_MAX;
391 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
392 atomic_set(&sbq->wake_index, 0);
393 atomic_set(&sbq->ws_active, 0);
395 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
397 free_percpu(sbq->alloc_hint);
398 sbitmap_free(&sbq->sb);
402 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
403 init_waitqueue_head(&sbq->ws[i].wait);
404 atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
407 sbq->round_robin = round_robin;
410 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
412 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
415 unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth);
418 if (sbq->wake_batch != wake_batch) {
419 WRITE_ONCE(sbq->wake_batch, wake_batch);
421 * Pairs with the memory barrier in sbitmap_queue_wake_up()
422 * to ensure that the batch size is updated before the wait
426 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
427 atomic_set(&sbq->ws[i].wait_cnt, 1);
431 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
433 sbitmap_queue_update_wake_batch(sbq, depth);
434 sbitmap_resize(&sbq->sb, depth);
436 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
438 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
440 unsigned int hint, depth;
443 hint = this_cpu_read(*sbq->alloc_hint);
444 depth = READ_ONCE(sbq->sb.depth);
445 if (unlikely(hint >= depth)) {
446 hint = depth ? prandom_u32() % depth : 0;
447 this_cpu_write(*sbq->alloc_hint, hint);
449 nr = sbitmap_get(&sbq->sb, hint, sbq->round_robin);
452 /* If the map is full, a hint won't do us much good. */
453 this_cpu_write(*sbq->alloc_hint, 0);
454 } else if (nr == hint || unlikely(sbq->round_robin)) {
455 /* Only update the hint if we used it. */
457 if (hint >= depth - 1)
459 this_cpu_write(*sbq->alloc_hint, hint);
464 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
466 int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
467 unsigned int shallow_depth)
469 unsigned int hint, depth;
472 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
474 hint = this_cpu_read(*sbq->alloc_hint);
475 depth = READ_ONCE(sbq->sb.depth);
476 if (unlikely(hint >= depth)) {
477 hint = depth ? prandom_u32() % depth : 0;
478 this_cpu_write(*sbq->alloc_hint, hint);
480 nr = sbitmap_get_shallow(&sbq->sb, hint, shallow_depth);
483 /* If the map is full, a hint won't do us much good. */
484 this_cpu_write(*sbq->alloc_hint, 0);
485 } else if (nr == hint || unlikely(sbq->round_robin)) {
486 /* Only update the hint if we used it. */
488 if (hint >= depth - 1)
490 this_cpu_write(*sbq->alloc_hint, hint);
495 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow);
497 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
498 unsigned int min_shallow_depth)
500 sbq->min_shallow_depth = min_shallow_depth;
501 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
503 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
505 static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
509 if (!atomic_read(&sbq->ws_active))
512 wake_index = atomic_read(&sbq->wake_index);
513 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
514 struct sbq_wait_state *ws = &sbq->ws[wake_index];
516 if (waitqueue_active(&ws->wait)) {
517 if (wake_index != atomic_read(&sbq->wake_index))
518 atomic_set(&sbq->wake_index, wake_index);
522 wake_index = sbq_index_inc(wake_index);
528 static bool __sbq_wake_up(struct sbitmap_queue *sbq)
530 struct sbq_wait_state *ws;
531 unsigned int wake_batch;
534 ws = sbq_wake_ptr(sbq);
538 wait_cnt = atomic_dec_return(&ws->wait_cnt);
542 wake_batch = READ_ONCE(sbq->wake_batch);
545 * Pairs with the memory barrier in sbitmap_queue_resize() to
546 * ensure that we see the batch size update before the wait
549 smp_mb__before_atomic();
552 * For concurrent callers of this, the one that failed the
553 * atomic_cmpxhcg() race should call this function again
554 * to wakeup a new batch on a different 'ws'.
556 ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch);
557 if (ret == wait_cnt) {
558 sbq_index_atomic_inc(&sbq->wake_index);
559 wake_up_nr(&ws->wait, wake_batch);
569 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
571 while (__sbq_wake_up(sbq))
574 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
576 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
580 * Once the clear bit is set, the bit may be allocated out.
582 * Orders READ/WRITE on the asssociated instance(such as request
583 * of blk_mq) by this bit for avoiding race with re-allocation,
584 * and its pair is the memory barrier implied in __sbitmap_get_word.
586 * One invariant is that the clear bit has to be zero when the bit
589 smp_mb__before_atomic();
590 sbitmap_deferred_clear_bit(&sbq->sb, nr);
593 * Pairs with the memory barrier in set_current_state() to ensure the
594 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
595 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
596 * waiter. See the comment on waitqueue_active().
598 smp_mb__after_atomic();
599 sbitmap_queue_wake_up(sbq);
601 if (likely(!sbq->round_robin && nr < sbq->sb.depth))
602 *per_cpu_ptr(sbq->alloc_hint, cpu) = nr;
604 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
606 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
611 * Pairs with the memory barrier in set_current_state() like in
612 * sbitmap_queue_wake_up().
615 wake_index = atomic_read(&sbq->wake_index);
616 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
617 struct sbq_wait_state *ws = &sbq->ws[wake_index];
619 if (waitqueue_active(&ws->wait))
622 wake_index = sbq_index_inc(wake_index);
625 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
627 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
632 sbitmap_show(&sbq->sb, m);
634 seq_puts(m, "alloc_hint={");
636 for_each_possible_cpu(i) {
640 seq_printf(m, "%u", *per_cpu_ptr(sbq->alloc_hint, i));
644 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
645 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
646 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
648 seq_puts(m, "ws={\n");
649 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
650 struct sbq_wait_state *ws = &sbq->ws[i];
652 seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
653 atomic_read(&ws->wait_cnt),
654 waitqueue_active(&ws->wait) ? "active" : "inactive");
658 seq_printf(m, "round_robin=%d\n", sbq->round_robin);
659 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
661 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
663 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
664 struct sbq_wait_state *ws,
665 struct sbq_wait *sbq_wait)
667 if (!sbq_wait->sbq) {
669 atomic_inc(&sbq->ws_active);
671 add_wait_queue(&ws->wait, &sbq_wait->wait);
673 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
675 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
677 list_del_init(&sbq_wait->wait.entry);
679 atomic_dec(&sbq_wait->sbq->ws_active);
680 sbq_wait->sbq = NULL;
683 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
685 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
686 struct sbq_wait_state *ws,
687 struct sbq_wait *sbq_wait, int state)
689 if (!sbq_wait->sbq) {
690 atomic_inc(&sbq->ws_active);
693 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
695 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
697 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
698 struct sbq_wait *sbq_wait)
700 finish_wait(&ws->wait, &sbq_wait->wait);
702 atomic_dec(&sbq->ws_active);
703 sbq_wait->sbq = NULL;
706 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);