2 * Copyright (C) 2016 Facebook
3 * Copyright (C) 2013-2014 Jens Axboe
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public
7 * License v2 as published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12 * General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program. If not, see <https://www.gnu.org/licenses/>.
18 #include <linux/sched.h>
19 #include <linux/random.h>
20 #include <linux/sbitmap.h>
21 #include <linux/seq_file.h>
23 int sbitmap_init_node(struct sbitmap
*sb
, unsigned int depth
, int shift
,
24 gfp_t flags
, int node
)
26 unsigned int bits_per_word
;
30 shift
= ilog2(BITS_PER_LONG
);
32 * If the bitmap is small, shrink the number of bits per word so
33 * we spread over a few cachelines, at least. If less than 4
34 * bits, just forget about it, it's not going to work optimally
38 while ((4U << shift
) > depth
)
42 bits_per_word
= 1U << shift
;
43 if (bits_per_word
> BITS_PER_LONG
)
48 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
55 sb
->map
= kcalloc_node(sb
->map_nr
, sizeof(*sb
->map
), flags
, node
);
59 for (i
= 0; i
< sb
->map_nr
; i
++) {
60 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
61 depth
-= sb
->map
[i
].depth
;
65 EXPORT_SYMBOL_GPL(sbitmap_init_node
);
67 void sbitmap_resize(struct sbitmap
*sb
, unsigned int depth
)
69 unsigned int bits_per_word
= 1U << sb
->shift
;
73 sb
->map_nr
= DIV_ROUND_UP(sb
->depth
, bits_per_word
);
75 for (i
= 0; i
< sb
->map_nr
; i
++) {
76 sb
->map
[i
].depth
= min(depth
, bits_per_word
);
77 depth
-= sb
->map
[i
].depth
;
80 EXPORT_SYMBOL_GPL(sbitmap_resize
);
82 static int __sbitmap_get_word(unsigned long *word
, unsigned long depth
,
83 unsigned int hint
, bool wrap
)
85 unsigned int orig_hint
= hint
;
89 nr
= find_next_zero_bit(word
, depth
, hint
);
90 if (unlikely(nr
>= depth
)) {
92 * We started with an offset, and we didn't reset the
93 * offset to 0 in a failure case, so start from 0 to
96 if (orig_hint
&& hint
&& wrap
) {
103 if (!test_and_set_bit_lock(nr
, word
))
107 if (hint
>= depth
- 1)
114 int sbitmap_get(struct sbitmap
*sb
, unsigned int alloc_hint
, bool round_robin
)
116 unsigned int i
, index
;
119 index
= SB_NR_TO_INDEX(sb
, alloc_hint
);
121 for (i
= 0; i
< sb
->map_nr
; i
++) {
122 nr
= __sbitmap_get_word(&sb
->map
[index
].word
,
123 sb
->map
[index
].depth
,
124 SB_NR_TO_BIT(sb
, alloc_hint
),
127 nr
+= index
<< sb
->shift
;
131 /* Jump to next index. */
133 alloc_hint
= index
<< sb
->shift
;
135 if (index
>= sb
->map_nr
) {
143 EXPORT_SYMBOL_GPL(sbitmap_get
);
145 int sbitmap_get_shallow(struct sbitmap
*sb
, unsigned int alloc_hint
,
146 unsigned long shallow_depth
)
148 unsigned int i
, index
;
151 index
= SB_NR_TO_INDEX(sb
, alloc_hint
);
153 for (i
= 0; i
< sb
->map_nr
; i
++) {
154 nr
= __sbitmap_get_word(&sb
->map
[index
].word
,
155 min(sb
->map
[index
].depth
, shallow_depth
),
156 SB_NR_TO_BIT(sb
, alloc_hint
), true);
158 nr
+= index
<< sb
->shift
;
162 /* Jump to next index. */
164 alloc_hint
= index
<< sb
->shift
;
166 if (index
>= sb
->map_nr
) {
174 EXPORT_SYMBOL_GPL(sbitmap_get_shallow
);
176 bool sbitmap_any_bit_set(const struct sbitmap
*sb
)
180 for (i
= 0; i
< sb
->map_nr
; i
++) {
186 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set
);
188 bool sbitmap_any_bit_clear(const struct sbitmap
*sb
)
192 for (i
= 0; i
< sb
->map_nr
; i
++) {
193 const struct sbitmap_word
*word
= &sb
->map
[i
];
196 ret
= find_first_zero_bit(&word
->word
, word
->depth
);
197 if (ret
< word
->depth
)
202 EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear
);
204 unsigned int sbitmap_weight(const struct sbitmap
*sb
)
206 unsigned int i
, weight
= 0;
208 for (i
= 0; i
< sb
->map_nr
; i
++) {
209 const struct sbitmap_word
*word
= &sb
->map
[i
];
211 weight
+= bitmap_weight(&word
->word
, word
->depth
);
215 EXPORT_SYMBOL_GPL(sbitmap_weight
);
217 void sbitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
219 seq_printf(m
, "depth=%u\n", sb
->depth
);
220 seq_printf(m
, "busy=%u\n", sbitmap_weight(sb
));
221 seq_printf(m
, "bits_per_word=%u\n", 1U << sb
->shift
);
222 seq_printf(m
, "map_nr=%u\n", sb
->map_nr
);
224 EXPORT_SYMBOL_GPL(sbitmap_show
);
226 static inline void emit_byte(struct seq_file
*m
, unsigned int offset
, u8 byte
)
228 if ((offset
& 0xf) == 0) {
231 seq_printf(m
, "%08x:", offset
);
233 if ((offset
& 0x1) == 0)
235 seq_printf(m
, "%02x", byte
);
238 void sbitmap_bitmap_show(struct sbitmap
*sb
, struct seq_file
*m
)
241 unsigned int byte_bits
= 0;
242 unsigned int offset
= 0;
245 for (i
= 0; i
< sb
->map_nr
; i
++) {
246 unsigned long word
= READ_ONCE(sb
->map
[i
].word
);
247 unsigned int word_bits
= READ_ONCE(sb
->map
[i
].depth
);
249 while (word_bits
> 0) {
250 unsigned int bits
= min(8 - byte_bits
, word_bits
);
252 byte
|= (word
& (BIT(bits
) - 1)) << byte_bits
;
254 if (byte_bits
== 8) {
255 emit_byte(m
, offset
, byte
);
265 emit_byte(m
, offset
, byte
);
271 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show
);
273 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue
*sbq
,
276 unsigned int wake_batch
;
277 unsigned int shallow_depth
;
280 * For each batch, we wake up one queue. We need to make sure that our
281 * batch size is small enough that the full depth of the bitmap,
282 * potentially limited by a shallow depth, is enough to wake up all of
285 * Each full word of the bitmap has bits_per_word bits, and there might
286 * be a partial word. There are depth / bits_per_word full words and
287 * depth % bits_per_word bits left over. In bitwise arithmetic:
289 * bits_per_word = 1 << shift
290 * depth / bits_per_word = depth >> shift
291 * depth % bits_per_word = depth & ((1 << shift) - 1)
293 * Each word can be limited to sbq->min_shallow_depth bits.
295 shallow_depth
= min(1U << sbq
->sb
.shift
, sbq
->min_shallow_depth
);
296 depth
= ((depth
>> sbq
->sb
.shift
) * shallow_depth
+
297 min(depth
& ((1U << sbq
->sb
.shift
) - 1), shallow_depth
));
298 wake_batch
= clamp_t(unsigned int, depth
/ SBQ_WAIT_QUEUES
, 1,
304 int sbitmap_queue_init_node(struct sbitmap_queue
*sbq
, unsigned int depth
,
305 int shift
, bool round_robin
, gfp_t flags
, int node
)
310 ret
= sbitmap_init_node(&sbq
->sb
, depth
, shift
, flags
, node
);
314 sbq
->alloc_hint
= alloc_percpu_gfp(unsigned int, flags
);
315 if (!sbq
->alloc_hint
) {
316 sbitmap_free(&sbq
->sb
);
320 if (depth
&& !round_robin
) {
321 for_each_possible_cpu(i
)
322 *per_cpu_ptr(sbq
->alloc_hint
, i
) = prandom_u32() % depth
;
325 sbq
->min_shallow_depth
= UINT_MAX
;
326 sbq
->wake_batch
= sbq_calc_wake_batch(sbq
, depth
);
327 atomic_set(&sbq
->wake_index
, 0);
329 sbq
->ws
= kzalloc_node(SBQ_WAIT_QUEUES
* sizeof(*sbq
->ws
), flags
, node
);
331 free_percpu(sbq
->alloc_hint
);
332 sbitmap_free(&sbq
->sb
);
336 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
337 init_waitqueue_head(&sbq
->ws
[i
].wait
);
338 atomic_set(&sbq
->ws
[i
].wait_cnt
, sbq
->wake_batch
);
341 sbq
->round_robin
= round_robin
;
344 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node
);
346 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue
*sbq
,
349 unsigned int wake_batch
= sbq_calc_wake_batch(sbq
, depth
);
352 if (sbq
->wake_batch
!= wake_batch
) {
353 WRITE_ONCE(sbq
->wake_batch
, wake_batch
);
355 * Pairs with the memory barrier in sbitmap_queue_wake_up()
356 * to ensure that the batch size is updated before the wait
360 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++)
361 atomic_set(&sbq
->ws
[i
].wait_cnt
, 1);
365 void sbitmap_queue_resize(struct sbitmap_queue
*sbq
, unsigned int depth
)
367 sbitmap_queue_update_wake_batch(sbq
, depth
);
368 sbitmap_resize(&sbq
->sb
, depth
);
370 EXPORT_SYMBOL_GPL(sbitmap_queue_resize
);
372 int __sbitmap_queue_get(struct sbitmap_queue
*sbq
)
374 unsigned int hint
, depth
;
377 hint
= this_cpu_read(*sbq
->alloc_hint
);
378 depth
= READ_ONCE(sbq
->sb
.depth
);
379 if (unlikely(hint
>= depth
)) {
380 hint
= depth
? prandom_u32() % depth
: 0;
381 this_cpu_write(*sbq
->alloc_hint
, hint
);
383 nr
= sbitmap_get(&sbq
->sb
, hint
, sbq
->round_robin
);
386 /* If the map is full, a hint won't do us much good. */
387 this_cpu_write(*sbq
->alloc_hint
, 0);
388 } else if (nr
== hint
|| unlikely(sbq
->round_robin
)) {
389 /* Only update the hint if we used it. */
391 if (hint
>= depth
- 1)
393 this_cpu_write(*sbq
->alloc_hint
, hint
);
398 EXPORT_SYMBOL_GPL(__sbitmap_queue_get
);
400 int __sbitmap_queue_get_shallow(struct sbitmap_queue
*sbq
,
401 unsigned int shallow_depth
)
403 unsigned int hint
, depth
;
406 WARN_ON_ONCE(shallow_depth
< sbq
->min_shallow_depth
);
408 hint
= this_cpu_read(*sbq
->alloc_hint
);
409 depth
= READ_ONCE(sbq
->sb
.depth
);
410 if (unlikely(hint
>= depth
)) {
411 hint
= depth
? prandom_u32() % depth
: 0;
412 this_cpu_write(*sbq
->alloc_hint
, hint
);
414 nr
= sbitmap_get_shallow(&sbq
->sb
, hint
, shallow_depth
);
417 /* If the map is full, a hint won't do us much good. */
418 this_cpu_write(*sbq
->alloc_hint
, 0);
419 } else if (nr
== hint
|| unlikely(sbq
->round_robin
)) {
420 /* Only update the hint if we used it. */
422 if (hint
>= depth
- 1)
424 this_cpu_write(*sbq
->alloc_hint
, hint
);
429 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow
);
431 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue
*sbq
,
432 unsigned int min_shallow_depth
)
434 sbq
->min_shallow_depth
= min_shallow_depth
;
435 sbitmap_queue_update_wake_batch(sbq
, sbq
->sb
.depth
);
437 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth
);
439 static struct sbq_wait_state
*sbq_wake_ptr(struct sbitmap_queue
*sbq
)
443 wake_index
= atomic_read(&sbq
->wake_index
);
444 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
445 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
447 if (waitqueue_active(&ws
->wait
)) {
448 int o
= atomic_read(&sbq
->wake_index
);
451 atomic_cmpxchg(&sbq
->wake_index
, o
, wake_index
);
455 wake_index
= sbq_index_inc(wake_index
);
461 static bool __sbq_wake_up(struct sbitmap_queue
*sbq
)
463 struct sbq_wait_state
*ws
;
464 unsigned int wake_batch
;
467 ws
= sbq_wake_ptr(sbq
);
471 wait_cnt
= atomic_dec_return(&ws
->wait_cnt
);
475 wake_batch
= READ_ONCE(sbq
->wake_batch
);
478 * Pairs with the memory barrier in sbitmap_queue_resize() to
479 * ensure that we see the batch size update before the wait
482 smp_mb__before_atomic();
485 * For concurrent callers of this, the one that failed the
486 * atomic_cmpxhcg() race should call this function again
487 * to wakeup a new batch on a different 'ws'.
489 ret
= atomic_cmpxchg(&ws
->wait_cnt
, wait_cnt
, wake_batch
);
490 if (ret
== wait_cnt
) {
491 sbq_index_atomic_inc(&sbq
->wake_index
);
492 wake_up_nr(&ws
->wait
, wake_batch
);
502 void sbitmap_queue_wake_up(struct sbitmap_queue
*sbq
)
504 while (__sbq_wake_up(sbq
))
507 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up
);
509 void sbitmap_queue_clear(struct sbitmap_queue
*sbq
, unsigned int nr
,
512 sbitmap_clear_bit_unlock(&sbq
->sb
, nr
);
514 * Pairs with the memory barrier in set_current_state() to ensure the
515 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
516 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
517 * waiter. See the comment on waitqueue_active().
519 smp_mb__after_atomic();
520 sbitmap_queue_wake_up(sbq
);
522 if (likely(!sbq
->round_robin
&& nr
< sbq
->sb
.depth
))
523 *per_cpu_ptr(sbq
->alloc_hint
, cpu
) = nr
;
525 EXPORT_SYMBOL_GPL(sbitmap_queue_clear
);
527 void sbitmap_queue_wake_all(struct sbitmap_queue
*sbq
)
532 * Pairs with the memory barrier in set_current_state() like in
533 * sbitmap_queue_wake_up().
536 wake_index
= atomic_read(&sbq
->wake_index
);
537 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
538 struct sbq_wait_state
*ws
= &sbq
->ws
[wake_index
];
540 if (waitqueue_active(&ws
->wait
))
543 wake_index
= sbq_index_inc(wake_index
);
546 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all
);
548 void sbitmap_queue_show(struct sbitmap_queue
*sbq
, struct seq_file
*m
)
553 sbitmap_show(&sbq
->sb
, m
);
555 seq_puts(m
, "alloc_hint={");
557 for_each_possible_cpu(i
) {
561 seq_printf(m
, "%u", *per_cpu_ptr(sbq
->alloc_hint
, i
));
565 seq_printf(m
, "wake_batch=%u\n", sbq
->wake_batch
);
566 seq_printf(m
, "wake_index=%d\n", atomic_read(&sbq
->wake_index
));
568 seq_puts(m
, "ws={\n");
569 for (i
= 0; i
< SBQ_WAIT_QUEUES
; i
++) {
570 struct sbq_wait_state
*ws
= &sbq
->ws
[i
];
572 seq_printf(m
, "\t{.wait_cnt=%d, .wait=%s},\n",
573 atomic_read(&ws
->wait_cnt
),
574 waitqueue_active(&ws
->wait
) ? "active" : "inactive");
578 seq_printf(m
, "round_robin=%d\n", sbq
->round_robin
);
579 seq_printf(m
, "min_shallow_depth=%u\n", sbq
->min_shallow_depth
);
581 EXPORT_SYMBOL_GPL(sbitmap_queue_show
);