sbitmap: remove unnecessary calculation of alloc_hint in __sbitmap_get_shallow
[linux-block.git] / lib / sbitmap.c
CommitLineData
0fc479b1 1// SPDX-License-Identifier: GPL-2.0-only
88459642
OS
2/*
3 * Copyright (C) 2016 Facebook
4 * Copyright (C) 2013-2014 Jens Axboe
88459642
OS
5 */
6
af8601ad 7#include <linux/sched.h>
98d95416 8#include <linux/random.h>
88459642 9#include <linux/sbitmap.h>
24af1ccf 10#include <linux/seq_file.h>
88459642 11
c548e62b 12static int init_alloc_hint(struct sbitmap *sb, gfp_t flags)
bf2c4282 13{
c548e62b 14 unsigned depth = sb->depth;
bf2c4282 15
c548e62b
ML
16 sb->alloc_hint = alloc_percpu_gfp(unsigned int, flags);
17 if (!sb->alloc_hint)
bf2c4282
ML
18 return -ENOMEM;
19
c548e62b 20 if (depth && !sb->round_robin) {
bf2c4282
ML
21 int i;
22
23 for_each_possible_cpu(i)
8032bf12 24 *per_cpu_ptr(sb->alloc_hint, i) = get_random_u32_below(depth);
bf2c4282 25 }
bf2c4282
ML
26 return 0;
27}
28
c548e62b 29static inline unsigned update_alloc_hint_before_get(struct sbitmap *sb,
bf2c4282
ML
30 unsigned int depth)
31{
32 unsigned hint;
33
c548e62b 34 hint = this_cpu_read(*sb->alloc_hint);
bf2c4282 35 if (unlikely(hint >= depth)) {
8032bf12 36 hint = depth ? get_random_u32_below(depth) : 0;
c548e62b 37 this_cpu_write(*sb->alloc_hint, hint);
bf2c4282
ML
38 }
39
40 return hint;
41}
42
c548e62b 43static inline void update_alloc_hint_after_get(struct sbitmap *sb,
bf2c4282
ML
44 unsigned int depth,
45 unsigned int hint,
46 unsigned int nr)
47{
48 if (nr == -1) {
49 /* If the map is full, a hint won't do us much good. */
c548e62b
ML
50 this_cpu_write(*sb->alloc_hint, 0);
51 } else if (nr == hint || unlikely(sb->round_robin)) {
bf2c4282
ML
52 /* Only update the hint if we used it. */
53 hint = nr + 1;
54 if (hint >= depth - 1)
55 hint = 0;
c548e62b 56 this_cpu_write(*sb->alloc_hint, hint);
bf2c4282
ML
57 }
58}
59
b2dbff1b
JA
60/*
61 * See if we have deferred clears that we can batch move
62 */
b78beea0 63static inline bool sbitmap_deferred_clear(struct sbitmap_word *map)
b2dbff1b 64{
c3250c8d 65 unsigned long mask;
b2dbff1b 66
661d4f55
PB
67 if (!READ_ONCE(map->cleared))
68 return false;
b2dbff1b
JA
69
70 /*
71 * First get a stable cleared mask, setting the old mask to 0.
72 */
b78beea0 73 mask = xchg(&map->cleared, 0);
b2dbff1b
JA
74
75 /*
76 * Now clear the masked bits in our free word
77 */
c3250c8d
PB
78 atomic_long_andnot(mask, (atomic_long_t *)&map->word);
79 BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(map->word));
661d4f55 80 return true;
b2dbff1b
JA
81}
82
88459642 83int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
c548e62b
ML
84 gfp_t flags, int node, bool round_robin,
85 bool alloc_hint)
88459642
OS
86{
87 unsigned int bits_per_word;
88459642 88
2d13b1ea
ML
89 if (shift < 0)
90 shift = sbitmap_calculate_shift(depth);
91
88459642
OS
92 bits_per_word = 1U << shift;
93 if (bits_per_word > BITS_PER_LONG)
94 return -EINVAL;
95
96 sb->shift = shift;
97 sb->depth = depth;
98 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
efe1f3a1 99 sb->round_robin = round_robin;
88459642
OS
100
101 if (depth == 0) {
102 sb->map = NULL;
103 return 0;
104 }
105
c548e62b
ML
106 if (alloc_hint) {
107 if (init_alloc_hint(sb, flags))
108 return -ENOMEM;
109 } else {
110 sb->alloc_hint = NULL;
111 }
112
863a66cd 113 sb->map = kvzalloc_node(sb->map_nr * sizeof(*sb->map), flags, node);
c548e62b
ML
114 if (!sb->map) {
115 free_percpu(sb->alloc_hint);
88459642 116 return -ENOMEM;
c548e62b 117 }
88459642 118
88459642
OS
119 return 0;
120}
121EXPORT_SYMBOL_GPL(sbitmap_init_node);
122
123void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
124{
125 unsigned int bits_per_word = 1U << sb->shift;
126 unsigned int i;
127
b2dbff1b 128 for (i = 0; i < sb->map_nr; i++)
b78beea0 129 sbitmap_deferred_clear(&sb->map[i]);
b2dbff1b 130
88459642
OS
131 sb->depth = depth;
132 sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
88459642
OS
133}
134EXPORT_SYMBOL_GPL(sbitmap_resize);
135
c05e6673
OS
136static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
137 unsigned int hint, bool wrap)
88459642 138{
88459642
OS
139 int nr;
140
0eff1f1a
PB
141 /* don't wrap if starting from 0 */
142 wrap = wrap && hint;
143
88459642 144 while (1) {
c05e6673
OS
145 nr = find_next_zero_bit(word, depth, hint);
146 if (unlikely(nr >= depth)) {
88459642
OS
147 /*
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
150 * exhaust the map.
151 */
0eff1f1a
PB
152 if (hint && wrap) {
153 hint = 0;
88459642
OS
154 continue;
155 }
156 return -1;
157 }
158
4ace53f1 159 if (!test_and_set_bit_lock(nr, word))
88459642
OS
160 break;
161
162 hint = nr + 1;
c05e6673 163 if (hint >= depth - 1)
88459642
OS
164 hint = 0;
165 }
166
167 return nr;
168}
169
ea86ea2c 170static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
efe1f3a1 171 unsigned int alloc_hint)
ea86ea2c 172{
b78beea0 173 struct sbitmap_word *map = &sb->map[index];
ea86ea2c
JA
174 int nr;
175
176 do {
3301bc53
ML
177 nr = __sbitmap_get_word(&map->word, __map_depth(sb, index),
178 alloc_hint, !sb->round_robin);
ea86ea2c
JA
179 if (nr != -1)
180 break;
b78beea0 181 if (!sbitmap_deferred_clear(map))
ea86ea2c
JA
182 break;
183 } while (1);
184
185 return nr;
186}
187
c548e62b 188static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
88459642
OS
189{
190 unsigned int i, index;
191 int nr = -1;
192
193 index = SB_NR_TO_INDEX(sb, alloc_hint);
194
27fae429
JA
195 /*
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.
199 */
efe1f3a1 200 if (sb->round_robin)
27fae429
JA
201 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
202 else
203 alloc_hint = 0;
204
88459642 205 for (i = 0; i < sb->map_nr; i++) {
efe1f3a1 206 nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
88459642
OS
207 if (nr != -1) {
208 nr += index << sb->shift;
209 break;
210 }
211
212 /* Jump to next index. */
27fae429
JA
213 alloc_hint = 0;
214 if (++index >= sb->map_nr)
88459642 215 index = 0;
88459642
OS
216 }
217
218 return nr;
219}
c548e62b
ML
220
221int sbitmap_get(struct sbitmap *sb)
222{
223 int nr;
224 unsigned int hint, depth;
225
226 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
227 return -1;
228
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);
233
234 return nr;
235}
88459642
OS
236EXPORT_SYMBOL_GPL(sbitmap_get);
237
c548e62b
ML
238static int __sbitmap_get_shallow(struct sbitmap *sb,
239 unsigned int alloc_hint,
240 unsigned long shallow_depth)
c05e6673
OS
241{
242 unsigned int i, index;
243 int nr = -1;
244
245 index = SB_NR_TO_INDEX(sb, alloc_hint);
f1591a8b 246 alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
c05e6673
OS
247
248 for (i = 0; i < sb->map_nr; i++) {
b2dbff1b 249again:
c05e6673 250 nr = __sbitmap_get_word(&sb->map[index].word,
3301bc53
ML
251 min_t(unsigned int,
252 __map_depth(sb, index),
253 shallow_depth),
f1591a8b 254 alloc_hint, true);
c05e6673
OS
255 if (nr != -1) {
256 nr += index << sb->shift;
257 break;
258 }
259
b78beea0 260 if (sbitmap_deferred_clear(&sb->map[index]))
b2dbff1b
JA
261 goto again;
262
c05e6673 263 /* Jump to next index. */
f1591a8b
KS
264 alloc_hint = 0;
265 if (++index >= sb->map_nr)
c05e6673 266 index = 0;
c05e6673
OS
267 }
268
269 return nr;
270}
c548e62b
ML
271
272int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
273{
274 int nr;
275 unsigned int hint, depth;
276
277 if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
278 return -1;
279
280 depth = READ_ONCE(sb->depth);
281 hint = update_alloc_hint_before_get(sb, depth);
282 nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
283 update_alloc_hint_after_get(sb, depth, hint, nr);
284
285 return nr;
286}
c05e6673
OS
287EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
288
88459642
OS
289bool sbitmap_any_bit_set(const struct sbitmap *sb)
290{
291 unsigned int i;
292
293 for (i = 0; i < sb->map_nr; i++) {
b2dbff1b 294 if (sb->map[i].word & ~sb->map[i].cleared)
88459642
OS
295 return true;
296 }
297 return false;
298}
299EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
300
ea86ea2c 301static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
88459642 302{
60658e0d 303 unsigned int i, weight = 0;
88459642
OS
304
305 for (i = 0; i < sb->map_nr; i++) {
306 const struct sbitmap_word *word = &sb->map[i];
3301bc53 307 unsigned int word_depth = __map_depth(sb, i);
88459642 308
ea86ea2c 309 if (set)
3301bc53 310 weight += bitmap_weight(&word->word, word_depth);
ea86ea2c 311 else
3301bc53 312 weight += bitmap_weight(&word->cleared, word_depth);
88459642
OS
313 }
314 return weight;
315}
ea86ea2c 316
cbb9950b 317static unsigned int sbitmap_cleared(const struct sbitmap *sb)
ea86ea2c 318{
cbb9950b 319 return __sbitmap_weight(sb, false);
ea86ea2c
JA
320}
321
cbb9950b 322unsigned int sbitmap_weight(const struct sbitmap *sb)
ea86ea2c 323{
cbb9950b 324 return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
ea86ea2c 325}
cbb9950b 326EXPORT_SYMBOL_GPL(sbitmap_weight);
88459642 327
24af1ccf
OS
328void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
329{
330 seq_printf(m, "depth=%u\n", sb->depth);
cbb9950b 331 seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
ea86ea2c 332 seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
24af1ccf
OS
333 seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
334 seq_printf(m, "map_nr=%u\n", sb->map_nr);
335}
336EXPORT_SYMBOL_GPL(sbitmap_show);
337
338static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
339{
340 if ((offset & 0xf) == 0) {
341 if (offset != 0)
342 seq_putc(m, '\n');
343 seq_printf(m, "%08x:", offset);
344 }
345 if ((offset & 0x1) == 0)
346 seq_putc(m, ' ');
347 seq_printf(m, "%02x", byte);
348}
349
350void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
351{
352 u8 byte = 0;
353 unsigned int byte_bits = 0;
354 unsigned int offset = 0;
355 int i;
356
357 for (i = 0; i < sb->map_nr; i++) {
358 unsigned long word = READ_ONCE(sb->map[i].word);
6bf0eb55 359 unsigned long cleared = READ_ONCE(sb->map[i].cleared);
3301bc53 360 unsigned int word_bits = __map_depth(sb, i);
24af1ccf 361
6bf0eb55
JG
362 word &= ~cleared;
363
24af1ccf
OS
364 while (word_bits > 0) {
365 unsigned int bits = min(8 - byte_bits, word_bits);
366
367 byte |= (word & (BIT(bits) - 1)) << byte_bits;
368 byte_bits += bits;
369 if (byte_bits == 8) {
370 emit_byte(m, offset, byte);
371 byte = 0;
372 byte_bits = 0;
373 offset++;
374 }
375 word >>= bits;
376 word_bits -= bits;
377 }
378 }
379 if (byte_bits) {
380 emit_byte(m, offset, byte);
381 offset++;
382 }
383 if (offset)
384 seq_putc(m, '\n');
385}
386EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
387
a3275539
OS
388static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
389 unsigned int depth)
88459642
OS
390{
391 unsigned int wake_batch;
a3275539 392 unsigned int shallow_depth;
88459642
OS
393
394 /*
395 * For each batch, we wake up one queue. We need to make sure that our
a3275539
OS
396 * batch size is small enough that the full depth of the bitmap,
397 * potentially limited by a shallow depth, is enough to wake up all of
398 * the queues.
399 *
400 * Each full word of the bitmap has bits_per_word bits, and there might
401 * be a partial word. There are depth / bits_per_word full words and
402 * depth % bits_per_word bits left over. In bitwise arithmetic:
403 *
404 * bits_per_word = 1 << shift
405 * depth / bits_per_word = depth >> shift
406 * depth % bits_per_word = depth & ((1 << shift) - 1)
407 *
408 * Each word can be limited to sbq->min_shallow_depth bits.
88459642 409 */
a3275539
OS
410 shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
411 depth = ((depth >> sbq->sb.shift) * shallow_depth +
412 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
413 wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
414 SBQ_WAKE_BATCH);
88459642
OS
415
416 return wake_batch;
417}
418
419int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
f4a644db 420 int shift, bool round_robin, gfp_t flags, int node)
88459642
OS
421{
422 int ret;
423 int i;
424
efe1f3a1 425 ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
c548e62b 426 round_robin, true);
88459642
OS
427 if (ret)
428 return ret;
429
a3275539
OS
430 sbq->min_shallow_depth = UINT_MAX;
431 sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
88459642 432 atomic_set(&sbq->wake_index, 0);
5d2ee712 433 atomic_set(&sbq->ws_active, 0);
4f8126bb
GKB
434 atomic_set(&sbq->completion_cnt, 0);
435 atomic_set(&sbq->wakeup_cnt, 0);
88459642 436
48e28166 437 sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
88459642
OS
438 if (!sbq->ws) {
439 sbitmap_free(&sbq->sb);
440 return -ENOMEM;
441 }
442
4f8126bb 443 for (i = 0; i < SBQ_WAIT_QUEUES; i++)
88459642 444 init_waitqueue_head(&sbq->ws[i].wait);
f4a644db 445
88459642
OS
446 return 0;
447}
448EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
449
180dccb0
LQ
450static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
451 unsigned int depth)
452{
453 unsigned int wake_batch;
454
455 wake_batch = sbq_calc_wake_batch(sbq, depth);
4f8126bb
GKB
456 if (sbq->wake_batch != wake_batch)
457 WRITE_ONCE(sbq->wake_batch, wake_batch);
180dccb0
LQ
458}
459
460void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
461 unsigned int users)
462{
463 unsigned int wake_batch;
10825410
LQ
464 unsigned int min_batch;
465 unsigned int depth = (sbq->sb.depth + users - 1) / users;
180dccb0 466
10825410
LQ
467 min_batch = sbq->sb.depth >= (4 * SBQ_WAIT_QUEUES) ? 4 : 1;
468
469 wake_batch = clamp_val(depth / SBQ_WAIT_QUEUES,
470 min_batch, SBQ_WAKE_BATCH);
4f8126bb
GKB
471
472 WRITE_ONCE(sbq->wake_batch, wake_batch);
180dccb0
LQ
473}
474EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
475
a3275539
OS
476void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
477{
478 sbitmap_queue_update_wake_batch(sbq, depth);
88459642
OS
479 sbitmap_resize(&sbq->sb, depth);
480}
481EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
482
f4a644db 483int __sbitmap_queue_get(struct sbitmap_queue *sbq)
40aabb67 484{
c548e62b 485 return sbitmap_get(&sbq->sb);
40aabb67
OS
486}
487EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
488
9672b0d4
JA
489unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
490 unsigned int *offset)
491{
492 struct sbitmap *sb = &sbq->sb;
493 unsigned int hint, depth;
494 unsigned long index, nr;
495 int i;
496
497 if (unlikely(sb->round_robin))
498 return 0;
499
500 depth = READ_ONCE(sb->depth);
501 hint = update_alloc_hint_before_get(sb, depth);
502
503 index = SB_NR_TO_INDEX(sb, hint);
504
505 for (i = 0; i < sb->map_nr; i++) {
506 struct sbitmap_word *map = &sb->map[index];
507 unsigned long get_mask;
3301bc53 508 unsigned int map_depth = __map_depth(sb, index);
9672b0d4
JA
509
510 sbitmap_deferred_clear(map);
3301bc53 511 if (map->word == (1UL << (map_depth - 1)) - 1)
fbb564a5 512 goto next;
9672b0d4 513
3301bc53
ML
514 nr = find_first_zero_bit(&map->word, map_depth);
515 if (nr + nr_tags <= map_depth) {
9672b0d4 516 atomic_long_t *ptr = (atomic_long_t *) &map->word;
c35227d4 517 unsigned long val;
9672b0d4 518
ddbfc34f 519 get_mask = ((1UL << nr_tags) - 1) << nr;
c35227d4 520 val = READ_ONCE(map->word);
9672b0d4 521 do {
fbb564a5 522 if ((val & ~get_mask) != val)
523 goto next;
c35227d4
UB
524 } while (!atomic_long_try_cmpxchg(ptr, &val,
525 get_mask | val));
526 get_mask = (get_mask & ~val) >> nr;
9672b0d4
JA
527 if (get_mask) {
528 *offset = nr + (index << sb->shift);
529 update_alloc_hint_after_get(sb, depth, hint,
ddbfc34f 530 *offset + nr_tags - 1);
9672b0d4
JA
531 return get_mask;
532 }
533 }
fbb564a5 534next:
9672b0d4
JA
535 /* Jump to next index. */
536 if (++index >= sb->map_nr)
537 index = 0;
538 }
539
540 return 0;
541}
542
3f607293
JG
543int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
544 unsigned int shallow_depth)
c05e6673 545{
61445b56
OS
546 WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
547
c548e62b 548 return sbitmap_get_shallow(&sbq->sb, shallow_depth);
c05e6673 549}
3f607293 550EXPORT_SYMBOL_GPL(sbitmap_queue_get_shallow);
c05e6673 551
a3275539
OS
552void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
553 unsigned int min_shallow_depth)
554{
555 sbq->min_shallow_depth = min_shallow_depth;
556 sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
557}
558EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
559
26edb30d 560static void __sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
88459642
OS
561{
562 int i, wake_index;
563
5d2ee712 564 if (!atomic_read(&sbq->ws_active))
26edb30d 565 return;
5d2ee712 566
88459642
OS
567 wake_index = atomic_read(&sbq->wake_index);
568 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
569 struct sbq_wait_state *ws = &sbq->ws[wake_index];
570
976570b4
GKB
571 /*
572 * Advance the index before checking the current queue.
573 * It improves fairness, by ensuring the queue doesn't
574 * need to be fully emptied before trying to wake up
575 * from the next one.
576 */
88459642 577 wake_index = sbq_index_inc(wake_index);
976570b4 578
26edb30d
GKB
579 /*
580 * It is sufficient to wake up at least one waiter to
581 * guarantee forward progress.
582 */
583 if (waitqueue_active(&ws->wait) &&
584 wake_up_nr(&ws->wait, nr))
585 break;
88459642
OS
586 }
587
26edb30d
GKB
588 if (wake_index != atomic_read(&sbq->wake_index))
589 atomic_set(&sbq->wake_index, wake_index);
88459642
OS
590}
591
4f8126bb 592void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
88459642 593{
4f8126bb 594 unsigned int wake_batch = READ_ONCE(sbq->wake_batch);
4f8126bb 595 unsigned int wakeups;
88459642 596
4f8126bb
GKB
597 if (!atomic_read(&sbq->ws_active))
598 return;
4acb8341 599
4f8126bb
GKB
600 atomic_add(nr, &sbq->completion_cnt);
601 wakeups = atomic_read(&sbq->wakeup_cnt);
88459642 602
4acb8341 603 do {
4f8126bb
GKB
604 if (atomic_read(&sbq->completion_cnt) - wakeups < wake_batch)
605 return;
4f8126bb
GKB
606 } while (!atomic_try_cmpxchg(&sbq->wakeup_cnt,
607 &wakeups, wakeups + wake_batch));
c854ab57 608
26edb30d 609 __sbitmap_queue_wake_up(sbq, wake_batch);
88459642 610}
bce1b56c 611EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
88459642 612
1aec5e4a
JA
613static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
614{
615 if (likely(!sb->round_robin && tag < sb->depth))
9f8b93a7 616 data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
1aec5e4a
JA
617}
618
619void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
620 int *tags, int nr_tags)
621{
622 struct sbitmap *sb = &sbq->sb;
623 unsigned long *addr = NULL;
624 unsigned long mask = 0;
625 int i;
626
627 smp_mb__before_atomic();
628 for (i = 0; i < nr_tags; i++) {
629 const int tag = tags[i] - offset;
630 unsigned long *this_addr;
631
632 /* since we're clearing a batch, skip the deferred map */
633 this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
634 if (!addr) {
635 addr = this_addr;
636 } else if (addr != this_addr) {
637 atomic_long_andnot(mask, (atomic_long_t *) addr);
638 mask = 0;
639 addr = this_addr;
640 }
641 mask |= (1UL << SB_NR_TO_BIT(sb, tag));
642 }
643
644 if (mask)
645 atomic_long_andnot(mask, (atomic_long_t *) addr);
646
647 smp_mb__after_atomic();
4acb8341 648 sbitmap_queue_wake_up(sbq, nr_tags);
1aec5e4a
JA
649 sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
650 tags[nr_tags - 1] - offset);
651}
652
40aabb67 653void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
f4a644db 654 unsigned int cpu)
88459642 655{
e6d1fa58
ML
656 /*
657 * Once the clear bit is set, the bit may be allocated out.
658 *
9dbbc3b9 659 * Orders READ/WRITE on the associated instance(such as request
e6d1fa58
ML
660 * of blk_mq) by this bit for avoiding race with re-allocation,
661 * and its pair is the memory barrier implied in __sbitmap_get_word.
662 *
663 * One invariant is that the clear bit has to be zero when the bit
664 * is in use.
665 */
666 smp_mb__before_atomic();
ea86ea2c
JA
667 sbitmap_deferred_clear_bit(&sbq->sb, nr);
668
e6fc4649
ML
669 /*
670 * Pairs with the memory barrier in set_current_state() to ensure the
671 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
672 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
673 * waiter. See the comment on waitqueue_active().
674 */
675 smp_mb__after_atomic();
4acb8341 676 sbitmap_queue_wake_up(sbq, 1);
1aec5e4a 677 sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
88459642
OS
678}
679EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
680
681void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
682{
683 int i, wake_index;
684
685 /*
f66227de 686 * Pairs with the memory barrier in set_current_state() like in
e6fc4649 687 * sbitmap_queue_wake_up().
88459642
OS
688 */
689 smp_mb();
690 wake_index = atomic_read(&sbq->wake_index);
691 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
692 struct sbq_wait_state *ws = &sbq->ws[wake_index];
693
694 if (waitqueue_active(&ws->wait))
695 wake_up(&ws->wait);
696
697 wake_index = sbq_index_inc(wake_index);
698 }
699}
700EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
24af1ccf
OS
701
702void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
703{
704 bool first;
705 int i;
706
707 sbitmap_show(&sbq->sb, m);
708
709 seq_puts(m, "alloc_hint={");
710 first = true;
711 for_each_possible_cpu(i) {
712 if (!first)
713 seq_puts(m, ", ");
714 first = false;
c548e62b 715 seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
24af1ccf
OS
716 }
717 seq_puts(m, "}\n");
718
719 seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
720 seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
5d2ee712 721 seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
24af1ccf
OS
722
723 seq_puts(m, "ws={\n");
724 for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
725 struct sbq_wait_state *ws = &sbq->ws[i];
4f8126bb 726 seq_printf(m, "\t{.wait=%s},\n",
24af1ccf
OS
727 waitqueue_active(&ws->wait) ? "active" : "inactive");
728 }
729 seq_puts(m, "}\n");
730
efe1f3a1 731 seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
a3275539 732 seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
24af1ccf
OS
733}
734EXPORT_SYMBOL_GPL(sbitmap_queue_show);
5d2ee712 735
9f6b7ef6
JA
736void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
737 struct sbq_wait_state *ws,
738 struct sbq_wait *sbq_wait)
739{
740 if (!sbq_wait->sbq) {
741 sbq_wait->sbq = sbq;
742 atomic_inc(&sbq->ws_active);
df034c93 743 add_wait_queue(&ws->wait, &sbq_wait->wait);
9f6b7ef6 744 }
9f6b7ef6
JA
745}
746EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
747
748void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
749{
750 list_del_init(&sbq_wait->wait.entry);
751 if (sbq_wait->sbq) {
752 atomic_dec(&sbq_wait->sbq->ws_active);
753 sbq_wait->sbq = NULL;
754 }
755}
756EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
757
5d2ee712
JA
758void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
759 struct sbq_wait_state *ws,
760 struct sbq_wait *sbq_wait, int state)
761{
9f6b7ef6 762 if (!sbq_wait->sbq) {
5d2ee712 763 atomic_inc(&sbq->ws_active);
9f6b7ef6 764 sbq_wait->sbq = sbq;
5d2ee712
JA
765 }
766 prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
767}
768EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
769
770void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
771 struct sbq_wait *sbq_wait)
772{
773 finish_wait(&ws->wait, &sbq_wait->wait);
9f6b7ef6 774 if (sbq_wait->sbq) {
5d2ee712 775 atomic_dec(&sbq->ws_active);
9f6b7ef6 776 sbq_wait->sbq = NULL;
5d2ee712
JA
777 }
778}
779EXPORT_SYMBOL_GPL(sbitmap_finish_wait);