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_word
*map
)
19 if (!READ_ONCE(map
->cleared
))
23 * First get a stable cleared mask, setting the old mask to 0.
25 mask
= xchg(&map
->cleared
, 0);
28 * Now clear the masked bits in our free word
30 atomic_long_andnot(mask
, (atomic_long_t
*)&map
->word
);
31 BUILD_BUG_ON(sizeof(atomic_long_t
) != sizeof(map
->word
));
35 int sbitmap_init_node(struct sbitmap
*sb
, unsigned int depth
, int shift
,
36 gfp_t flags
, int node
)
38 unsigned int bits_per_word
;
42 shift
= ilog2(BITS_PER_LONG
);
44 * If the bitmap is small, shrink the number of bits per word so
45 * we spread over a few cachelines, at least. If less than 4
46 * bits, just forget about it, it's not going to work optimally
50 while ((4U << shift
) > depth
)
54 bits_per_word
= 1U << shift
;
55 if (bits_per_word
> BITS_PER_LONG
)
60 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
67 sb
->map
= kcalloc_node(sb
->map_nr
, sizeof(*sb
->map
), flags
, node
);
71 for (i
= 0; i
< sb
->map_nr
; i
++) {
72 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
73 depth
-= sb
->map
[i
].depth
;
77 EXPORT_SYMBOL_GPL(sbitmap_init_node
);
79 void sbitmap_resize(struct sbitmap
*sb
, unsigned int depth
)
81 unsigned int bits_per_word
= 1U << sb
->shift
;
84 for (i
= 0; i
< sb
->map_nr
; i
++)
85 sbitmap_deferred_clear(&sb
->map
[i
]);
88 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
90 for (i
= 0; i
< sb
->map_nr
; i
++) {
91 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
92 depth
-= sb
->map
[i
].depth
;
95 EXPORT_SYMBOL_GPL(sbitmap_resize
);
97 static int __sbitmap_get_word(unsigned long *word
, unsigned long depth
,
98 unsigned int hint
, bool wrap
)
102 /* don't wrap if starting from 0 */
106 nr
= find_next_zero_bit(word
, depth
, hint
);
107 if (unlikely(nr
>= depth
)) {
109 * We started with an offset, and we didn't reset the
110 * offset to 0 in a failure case, so start from 0 to
120 if (!test_and_set_bit_lock(nr
, word
))
124 if (hint
>= depth
- 1)
131 static int sbitmap_find_bit_in_index(struct sbitmap
*sb
, int index
,
132 unsigned int alloc_hint
, bool round_robin
)
134 struct sbitmap_word
*map
= &sb
->map
[index
];
138 nr
= __sbitmap_get_word(&map
->word
, map
->depth
, alloc_hint
,
142 if (!sbitmap_deferred_clear(map
))
149 int sbitmap_get(struct sbitmap
*sb
, unsigned int alloc_hint
, bool round_robin
)
151 unsigned int i
, index
;
154 index
= SB_NR_TO_INDEX(sb
, alloc_hint
);
157 * Unless we're doing round robin tag allocation, just use the
158 * alloc_hint to find the right word index. No point in looping
159 * twice in find_next_zero_bit() for that case.
162 alloc_hint
= SB_NR_TO_BIT(sb
, alloc_hint
);
166 for (i
= 0; i
< sb
->map_nr
; i
++) {
167 nr
= sbitmap_find_bit_in_index(sb
, index
, alloc_hint
,
170 nr
+= index
<< sb
->shift
;
174 /* Jump to next index. */
176 if (++index
>= sb
->map_nr
)
182 EXPORT_SYMBOL_GPL(sbitmap_get
);
184 int sbitmap_get_shallow(struct sbitmap
*sb
, unsigned int alloc_hint
,
185 unsigned long shallow_depth
)
187 unsigned int i
, index
;
190 index
= SB_NR_TO_INDEX(sb
, alloc_hint
);
192 for (i
= 0; i
< sb
->map_nr
; i
++) {
194 nr
= __sbitmap_get_word(&sb
->map
[index
].word
,
195 min(sb
->map
[index
].depth
, shallow_depth
),
196 SB_NR_TO_BIT(sb
, alloc_hint
), true);
198 nr
+= index
<< sb
->shift
;
202 if (sbitmap_deferred_clear(&sb
->map
[index
]))
205 /* Jump to next index. */
207 alloc_hint
= index
<< sb
->shift
;
209 if (index
>= sb
->map_nr
) {
217 EXPORT_SYMBOL_GPL(sbitmap_get_shallow
);
219 bool sbitmap_any_bit_set(const struct sbitmap
*sb
)
223 for (i
= 0; i
< sb
->map_nr
; i
++) {
224 if (sb
->map
[i
].word
& ~sb
->map
[i
].cleared
)
229 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set
);
231 static unsigned int __sbitmap_weight(const struct sbitmap
*sb
, bool set
)
233 unsigned int i
, weight
= 0;
235 for (i
= 0; i
< sb
->map_nr
; i
++) {
236 const struct sbitmap_word
*word
= &sb
->map
[i
];
239 weight
+= bitmap_weight(&word
->word
, word
->depth
);
241 weight
+= bitmap_weight(&word
->cleared
, word
->depth
);
246 static unsigned int sbitmap_weight(const struct sbitmap
*sb
)
248 return __sbitmap_weight(sb
, true);
251 static unsigned int sbitmap_cleared(const struct sbitmap
*sb
)
253 return __sbitmap_weight(sb
, false);
256 void sbitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
258 seq_printf(m
, "depth=%u\n", sb
->depth
);
259 seq_printf(m
, "busy=%u\n", sbitmap_weight(sb
) - sbitmap_cleared(sb
));
260 seq_printf(m
, "cleared=%u\n", sbitmap_cleared(sb
));
261 seq_printf(m
, "bits_per_word=%u\n", 1U << sb
->shift
);
262 seq_printf(m
, "map_nr=%u\n", sb
->map_nr
);
264 EXPORT_SYMBOL_GPL(sbitmap_show
);
266 static inline void emit_byte(struct seq_file
*m
, unsigned int offset
, u8 byte
)
268 if ((offset
& 0xf) == 0) {
271 seq_printf(m
, "%08x:", offset
);
273 if ((offset
& 0x1) == 0)
275 seq_printf(m
, "%02x", byte
);
278 void sbitmap_bitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
281 unsigned int byte_bits
= 0;
282 unsigned int offset
= 0;
285 for (i
= 0; i
< sb
->map_nr
; i
++) {
286 unsigned long word
= READ_ONCE(sb
->map
[i
].word
);
287 unsigned long cleared
= READ_ONCE(sb
->map
[i
].cleared
);
288 unsigned int word_bits
= READ_ONCE(sb
->map
[i
].depth
);
292 while (word_bits
> 0) {
293 unsigned int bits
= min(8 - byte_bits
, word_bits
);
295 byte
|= (word
& (BIT(bits
) - 1)) << byte_bits
;
297 if (byte_bits
== 8) {
298 emit_byte(m
, offset
, byte
);
308 emit_byte(m
, offset
, byte
);
314 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show
);
316 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue
*sbq
,
319 unsigned int wake_batch
;
320 unsigned int shallow_depth
;
323 * For each batch, we wake up one queue. We need to make sure that our
324 * batch size is small enough that the full depth of the bitmap,
325 * potentially limited by a shallow depth, is enough to wake up all of
328 * Each full word of the bitmap has bits_per_word bits, and there might
329 * be a partial word. There are depth / bits_per_word full words and
330 * depth % bits_per_word bits left over. In bitwise arithmetic:
332 * bits_per_word = 1 << shift
333 * depth / bits_per_word = depth >> shift
334 * depth % bits_per_word = depth & ((1 << shift) - 1)
336 * Each word can be limited to sbq->min_shallow_depth bits.
338 shallow_depth
= min(1U << sbq
->sb
.shift
, sbq
->min_shallow_depth
);
339 depth
= ((depth
>> sbq
->sb
.shift
) * shallow_depth
+
340 min(depth
& ((1U << sbq
->sb
.shift
) - 1), shallow_depth
));
341 wake_batch
= clamp_t(unsigned int, depth
/ SBQ_WAIT_QUEUES
, 1,
347 int sbitmap_queue_init_node(struct sbitmap_queue
*sbq
, unsigned int depth
,
348 int shift
, bool round_robin
, gfp_t flags
, int node
)
353 ret
= sbitmap_init_node(&sbq
->sb
, depth
, shift
, flags
, node
);
357 sbq
->alloc_hint
= alloc_percpu_gfp(unsigned int, flags
);
358 if (!sbq
->alloc_hint
) {
359 sbitmap_free(&sbq
->sb
);
363 if (depth
&& !round_robin
) {
364 for_each_possible_cpu(i
)
365 *per_cpu_ptr(sbq
->alloc_hint
, i
) = prandom_u32() % depth
;
368 sbq
->min_shallow_depth
= UINT_MAX
;
369 sbq
->wake_batch
= sbq_calc_wake_batch(sbq
, depth
);
370 atomic_set(&sbq
->wake_index
, 0);
371 atomic_set(&sbq
->ws_active
, 0);
373 sbq
->ws
= kzalloc_node(SBQ_WAIT_QUEUES
* sizeof(*sbq
->ws
), flags
, node
);
375 free_percpu(sbq
->alloc_hint
);
376 sbitmap_free(&sbq
->sb
);
380 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
381 init_waitqueue_head(&sbq
->ws
[i
].wait
);
382 atomic_set(&sbq
->ws
[i
].wait_cnt
, sbq
->wake_batch
);
385 sbq
->round_robin
= round_robin
;
388 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node
);
390 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue
*sbq
,
393 unsigned int wake_batch
= sbq_calc_wake_batch(sbq
, depth
);
396 if (sbq
->wake_batch
!= wake_batch
) {
397 WRITE_ONCE(sbq
->wake_batch
, wake_batch
);
399 * Pairs with the memory barrier in sbitmap_queue_wake_up()
400 * to ensure that the batch size is updated before the wait
404 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++)
405 atomic_set(&sbq
->ws
[i
].wait_cnt
, 1);
409 void sbitmap_queue_resize(struct sbitmap_queue
*sbq
, unsigned int depth
)
411 sbitmap_queue_update_wake_batch(sbq
, depth
);
412 sbitmap_resize(&sbq
->sb
, depth
);
414 EXPORT_SYMBOL_GPL(sbitmap_queue_resize
);
416 int __sbitmap_queue_get(struct sbitmap_queue
*sbq
)
418 unsigned int hint
, depth
;
421 hint
= this_cpu_read(*sbq
->alloc_hint
);
422 depth
= READ_ONCE(sbq
->sb
.depth
);
423 if (unlikely(hint
>= depth
)) {
424 hint
= depth
? prandom_u32() % depth
: 0;
425 this_cpu_write(*sbq
->alloc_hint
, hint
);
427 nr
= sbitmap_get(&sbq
->sb
, hint
, sbq
->round_robin
);
430 /* If the map is full, a hint won't do us much good. */
431 this_cpu_write(*sbq
->alloc_hint
, 0);
432 } else if (nr
== hint
|| unlikely(sbq
->round_robin
)) {
433 /* Only update the hint if we used it. */
435 if (hint
>= depth
- 1)
437 this_cpu_write(*sbq
->alloc_hint
, hint
);
442 EXPORT_SYMBOL_GPL(__sbitmap_queue_get
);
444 int __sbitmap_queue_get_shallow(struct sbitmap_queue
*sbq
,
445 unsigned int shallow_depth
)
447 unsigned int hint
, depth
;
450 WARN_ON_ONCE(shallow_depth
< sbq
->min_shallow_depth
);
452 hint
= this_cpu_read(*sbq
->alloc_hint
);
453 depth
= READ_ONCE(sbq
->sb
.depth
);
454 if (unlikely(hint
>= depth
)) {
455 hint
= depth
? prandom_u32() % depth
: 0;
456 this_cpu_write(*sbq
->alloc_hint
, hint
);
458 nr
= sbitmap_get_shallow(&sbq
->sb
, hint
, shallow_depth
);
461 /* If the map is full, a hint won't do us much good. */
462 this_cpu_write(*sbq
->alloc_hint
, 0);
463 } else if (nr
== hint
|| unlikely(sbq
->round_robin
)) {
464 /* Only update the hint if we used it. */
466 if (hint
>= depth
- 1)
468 this_cpu_write(*sbq
->alloc_hint
, hint
);
473 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow
);
475 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue
*sbq
,
476 unsigned int min_shallow_depth
)
478 sbq
->min_shallow_depth
= min_shallow_depth
;
479 sbitmap_queue_update_wake_batch(sbq
, sbq
->sb
.depth
);
481 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth
);
483 static struct sbq_wait_state
*sbq_wake_ptr(struct sbitmap_queue
*sbq
)
487 if (!atomic_read(&sbq
->ws_active
))
490 wake_index
= atomic_read(&sbq
->wake_index
);
491 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
492 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
494 if (waitqueue_active(&ws
->wait
)) {
495 if (wake_index
!= atomic_read(&sbq
->wake_index
))
496 atomic_set(&sbq
->wake_index
, wake_index
);
500 wake_index
= sbq_index_inc(wake_index
);
506 static bool __sbq_wake_up(struct sbitmap_queue
*sbq
)
508 struct sbq_wait_state
*ws
;
509 unsigned int wake_batch
;
512 ws
= sbq_wake_ptr(sbq
);
516 wait_cnt
= atomic_dec_return(&ws
->wait_cnt
);
520 wake_batch
= READ_ONCE(sbq
->wake_batch
);
523 * Pairs with the memory barrier in sbitmap_queue_resize() to
524 * ensure that we see the batch size update before the wait
527 smp_mb__before_atomic();
530 * For concurrent callers of this, the one that failed the
531 * atomic_cmpxhcg() race should call this function again
532 * to wakeup a new batch on a different 'ws'.
534 ret
= atomic_cmpxchg(&ws
->wait_cnt
, wait_cnt
, wake_batch
);
535 if (ret
== wait_cnt
) {
536 sbq_index_atomic_inc(&sbq
->wake_index
);
537 wake_up_nr(&ws
->wait
, wake_batch
);
547 void sbitmap_queue_wake_up(struct sbitmap_queue
*sbq
)
549 while (__sbq_wake_up(sbq
))
552 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up
);
554 void sbitmap_queue_clear(struct sbitmap_queue
*sbq
, unsigned int nr
,
558 * Once the clear bit is set, the bit may be allocated out.
560 * Orders READ/WRITE on the asssociated instance(such as request
561 * of blk_mq) by this bit for avoiding race with re-allocation,
562 * and its pair is the memory barrier implied in __sbitmap_get_word.
564 * One invariant is that the clear bit has to be zero when the bit
567 smp_mb__before_atomic();
568 sbitmap_deferred_clear_bit(&sbq
->sb
, nr
);
571 * Pairs with the memory barrier in set_current_state() to ensure the
572 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
573 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
574 * waiter. See the comment on waitqueue_active().
576 smp_mb__after_atomic();
577 sbitmap_queue_wake_up(sbq
);
579 if (likely(!sbq
->round_robin
&& nr
< sbq
->sb
.depth
))
580 *per_cpu_ptr(sbq
->alloc_hint
, cpu
) = nr
;
582 EXPORT_SYMBOL_GPL(sbitmap_queue_clear
);
584 void sbitmap_queue_wake_all(struct sbitmap_queue
*sbq
)
589 * Pairs with the memory barrier in set_current_state() like in
590 * sbitmap_queue_wake_up().
593 wake_index
= atomic_read(&sbq
->wake_index
);
594 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
595 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
597 if (waitqueue_active(&ws
->wait
))
600 wake_index
= sbq_index_inc(wake_index
);
603 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all
);
605 void sbitmap_queue_show(struct sbitmap_queue
*sbq
, struct seq_file
*m
)
610 sbitmap_show(&sbq
->sb
, m
);
612 seq_puts(m
, "alloc_hint={");
614 for_each_possible_cpu(i
) {
618 seq_printf(m
, "%u", *per_cpu_ptr(sbq
->alloc_hint
, i
));
622 seq_printf(m
, "wake_batch=%u\n", sbq
->wake_batch
);
623 seq_printf(m
, "wake_index=%d\n", atomic_read(&sbq
->wake_index
));
624 seq_printf(m
, "ws_active=%d\n", atomic_read(&sbq
->ws_active
));
626 seq_puts(m
, "ws={\n");
627 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
628 struct sbq_wait_state
*ws
= &sbq
->ws
[i
];
630 seq_printf(m
, "\t{.wait_cnt=%d, .wait=%s},\n",
631 atomic_read(&ws
->wait_cnt
),
632 waitqueue_active(&ws
->wait
) ? "active" : "inactive");
636 seq_printf(m
, "round_robin=%d\n", sbq
->round_robin
);
637 seq_printf(m
, "min_shallow_depth=%u\n", sbq
->min_shallow_depth
);
639 EXPORT_SYMBOL_GPL(sbitmap_queue_show
);
641 void sbitmap_add_wait_queue(struct sbitmap_queue
*sbq
,
642 struct sbq_wait_state
*ws
,
643 struct sbq_wait
*sbq_wait
)
645 if (!sbq_wait
->sbq
) {
647 atomic_inc(&sbq
->ws_active
);
648 add_wait_queue(&ws
->wait
, &sbq_wait
->wait
);
651 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue
);
653 void sbitmap_del_wait_queue(struct sbq_wait
*sbq_wait
)
655 list_del_init(&sbq_wait
->wait
.entry
);
657 atomic_dec(&sbq_wait
->sbq
->ws_active
);
658 sbq_wait
->sbq
= NULL
;
661 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue
);
663 void sbitmap_prepare_to_wait(struct sbitmap_queue
*sbq
,
664 struct sbq_wait_state
*ws
,
665 struct sbq_wait
*sbq_wait
, int state
)
667 if (!sbq_wait
->sbq
) {
668 atomic_inc(&sbq
->ws_active
);
671 prepare_to_wait_exclusive(&ws
->wait
, &sbq_wait
->wait
, state
);
673 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait
);
675 void sbitmap_finish_wait(struct sbitmap_queue
*sbq
, struct sbq_wait_state
*ws
,
676 struct sbq_wait
*sbq_wait
)
678 finish_wait(&ws
->wait
, &sbq_wait
->wait
);
680 atomic_dec(&sbq
->ws_active
);
681 sbq_wait
->sbq
= NULL
;
684 EXPORT_SYMBOL_GPL(sbitmap_finish_wait
);