Commit | Line | Data |
---|---|---|
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 | |
b2dbff1b JA |
12 | /* |
13 | * See if we have deferred clears that we can batch move | |
14 | */ | |
15 | static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index) | |
16 | { | |
17 | unsigned long mask, val; | |
b2dbff1b | 18 | bool ret = false; |
fe76fc6a | 19 | unsigned long flags; |
b2dbff1b | 20 | |
fe76fc6a | 21 | spin_lock_irqsave(&sb->map[index].swap_lock, flags); |
b2dbff1b JA |
22 | |
23 | if (!sb->map[index].cleared) | |
24 | goto out_unlock; | |
25 | ||
26 | /* | |
27 | * First get a stable cleared mask, setting the old mask to 0. | |
28 | */ | |
29 | do { | |
30 | mask = sb->map[index].cleared; | |
31 | } while (cmpxchg(&sb->map[index].cleared, mask, 0) != mask); | |
32 | ||
33 | /* | |
34 | * Now clear the masked bits in our free word | |
35 | */ | |
36 | do { | |
37 | val = sb->map[index].word; | |
38 | } while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val); | |
39 | ||
40 | ret = true; | |
41 | out_unlock: | |
fe76fc6a | 42 | spin_unlock_irqrestore(&sb->map[index].swap_lock, flags); |
b2dbff1b JA |
43 | return ret; |
44 | } | |
45 | ||
88459642 OS |
46 | int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift, |
47 | gfp_t flags, int node) | |
48 | { | |
49 | unsigned int bits_per_word; | |
50 | unsigned int i; | |
51 | ||
52 | if (shift < 0) { | |
53 | shift = ilog2(BITS_PER_LONG); | |
54 | /* | |
55 | * If the bitmap is small, shrink the number of bits per word so | |
56 | * we spread over a few cachelines, at least. If less than 4 | |
57 | * bits, just forget about it, it's not going to work optimally | |
58 | * anyway. | |
59 | */ | |
60 | if (depth >= 4) { | |
61 | while ((4U << shift) > depth) | |
62 | shift--; | |
63 | } | |
64 | } | |
65 | bits_per_word = 1U << shift; | |
66 | if (bits_per_word > BITS_PER_LONG) | |
67 | return -EINVAL; | |
68 | ||
69 | sb->shift = shift; | |
70 | sb->depth = depth; | |
71 | sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); | |
72 | ||
73 | if (depth == 0) { | |
74 | sb->map = NULL; | |
75 | return 0; | |
76 | } | |
77 | ||
590b5b7d | 78 | sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node); |
88459642 OS |
79 | if (!sb->map) |
80 | return -ENOMEM; | |
81 | ||
82 | for (i = 0; i < sb->map_nr; i++) { | |
83 | sb->map[i].depth = min(depth, bits_per_word); | |
84 | depth -= sb->map[i].depth; | |
ea86ea2c | 85 | spin_lock_init(&sb->map[i].swap_lock); |
88459642 OS |
86 | } |
87 | return 0; | |
88 | } | |
89 | EXPORT_SYMBOL_GPL(sbitmap_init_node); | |
90 | ||
91 | void sbitmap_resize(struct sbitmap *sb, unsigned int depth) | |
92 | { | |
93 | unsigned int bits_per_word = 1U << sb->shift; | |
94 | unsigned int i; | |
95 | ||
b2dbff1b JA |
96 | for (i = 0; i < sb->map_nr; i++) |
97 | sbitmap_deferred_clear(sb, i); | |
98 | ||
88459642 OS |
99 | sb->depth = depth; |
100 | sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); | |
101 | ||
102 | for (i = 0; i < sb->map_nr; i++) { | |
103 | sb->map[i].depth = min(depth, bits_per_word); | |
104 | depth -= sb->map[i].depth; | |
105 | } | |
106 | } | |
107 | EXPORT_SYMBOL_GPL(sbitmap_resize); | |
108 | ||
c05e6673 OS |
109 | static int __sbitmap_get_word(unsigned long *word, unsigned long depth, |
110 | unsigned int hint, bool wrap) | |
88459642 OS |
111 | { |
112 | unsigned int orig_hint = hint; | |
113 | int nr; | |
114 | ||
115 | while (1) { | |
c05e6673 OS |
116 | nr = find_next_zero_bit(word, depth, hint); |
117 | if (unlikely(nr >= depth)) { | |
88459642 OS |
118 | /* |
119 | * We started with an offset, and we didn't reset the | |
120 | * offset to 0 in a failure case, so start from 0 to | |
121 | * exhaust the map. | |
122 | */ | |
123 | if (orig_hint && hint && wrap) { | |
124 | hint = orig_hint = 0; | |
125 | continue; | |
126 | } | |
127 | return -1; | |
128 | } | |
129 | ||
4ace53f1 | 130 | if (!test_and_set_bit_lock(nr, word)) |
88459642 OS |
131 | break; |
132 | ||
133 | hint = nr + 1; | |
c05e6673 | 134 | if (hint >= depth - 1) |
88459642 OS |
135 | hint = 0; |
136 | } | |
137 | ||
138 | return nr; | |
139 | } | |
140 | ||
ea86ea2c JA |
141 | static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index, |
142 | unsigned int alloc_hint, bool round_robin) | |
143 | { | |
144 | int nr; | |
145 | ||
146 | do { | |
147 | nr = __sbitmap_get_word(&sb->map[index].word, | |
148 | sb->map[index].depth, alloc_hint, | |
149 | !round_robin); | |
150 | if (nr != -1) | |
151 | break; | |
152 | if (!sbitmap_deferred_clear(sb, index)) | |
153 | break; | |
154 | } while (1); | |
155 | ||
156 | return nr; | |
157 | } | |
158 | ||
88459642 OS |
159 | int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin) |
160 | { | |
161 | unsigned int i, index; | |
162 | int nr = -1; | |
163 | ||
164 | index = SB_NR_TO_INDEX(sb, alloc_hint); | |
165 | ||
27fae429 JA |
166 | /* |
167 | * Unless we're doing round robin tag allocation, just use the | |
168 | * alloc_hint to find the right word index. No point in looping | |
169 | * twice in find_next_zero_bit() for that case. | |
170 | */ | |
171 | if (round_robin) | |
172 | alloc_hint = SB_NR_TO_BIT(sb, alloc_hint); | |
173 | else | |
174 | alloc_hint = 0; | |
175 | ||
88459642 | 176 | for (i = 0; i < sb->map_nr; i++) { |
ea86ea2c JA |
177 | nr = sbitmap_find_bit_in_index(sb, index, alloc_hint, |
178 | round_robin); | |
88459642 OS |
179 | if (nr != -1) { |
180 | nr += index << sb->shift; | |
181 | break; | |
182 | } | |
183 | ||
184 | /* Jump to next index. */ | |
27fae429 JA |
185 | alloc_hint = 0; |
186 | if (++index >= sb->map_nr) | |
88459642 | 187 | index = 0; |
88459642 OS |
188 | } |
189 | ||
190 | return nr; | |
191 | } | |
192 | EXPORT_SYMBOL_GPL(sbitmap_get); | |
193 | ||
c05e6673 OS |
194 | int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint, |
195 | unsigned long shallow_depth) | |
196 | { | |
197 | unsigned int i, index; | |
198 | int nr = -1; | |
199 | ||
200 | index = SB_NR_TO_INDEX(sb, alloc_hint); | |
201 | ||
202 | for (i = 0; i < sb->map_nr; i++) { | |
b2dbff1b | 203 | again: |
c05e6673 OS |
204 | nr = __sbitmap_get_word(&sb->map[index].word, |
205 | min(sb->map[index].depth, shallow_depth), | |
206 | SB_NR_TO_BIT(sb, alloc_hint), true); | |
207 | if (nr != -1) { | |
208 | nr += index << sb->shift; | |
209 | break; | |
210 | } | |
211 | ||
b2dbff1b JA |
212 | if (sbitmap_deferred_clear(sb, index)) |
213 | goto again; | |
214 | ||
c05e6673 OS |
215 | /* Jump to next index. */ |
216 | index++; | |
217 | alloc_hint = index << sb->shift; | |
218 | ||
219 | if (index >= sb->map_nr) { | |
220 | index = 0; | |
221 | alloc_hint = 0; | |
222 | } | |
223 | } | |
224 | ||
225 | return nr; | |
226 | } | |
227 | EXPORT_SYMBOL_GPL(sbitmap_get_shallow); | |
228 | ||
88459642 OS |
229 | bool sbitmap_any_bit_set(const struct sbitmap *sb) |
230 | { | |
231 | unsigned int i; | |
232 | ||
233 | for (i = 0; i < sb->map_nr; i++) { | |
b2dbff1b | 234 | if (sb->map[i].word & ~sb->map[i].cleared) |
88459642 OS |
235 | return true; |
236 | } | |
237 | return false; | |
238 | } | |
239 | EXPORT_SYMBOL_GPL(sbitmap_any_bit_set); | |
240 | ||
241 | bool sbitmap_any_bit_clear(const struct sbitmap *sb) | |
242 | { | |
243 | unsigned int i; | |
244 | ||
245 | for (i = 0; i < sb->map_nr; i++) { | |
246 | const struct sbitmap_word *word = &sb->map[i]; | |
b2dbff1b | 247 | unsigned long mask = word->word & ~word->cleared; |
88459642 OS |
248 | unsigned long ret; |
249 | ||
b2dbff1b | 250 | ret = find_first_zero_bit(&mask, word->depth); |
88459642 OS |
251 | if (ret < word->depth) |
252 | return true; | |
253 | } | |
254 | return false; | |
255 | } | |
256 | EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear); | |
257 | ||
ea86ea2c | 258 | static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set) |
88459642 | 259 | { |
60658e0d | 260 | unsigned int i, weight = 0; |
88459642 OS |
261 | |
262 | for (i = 0; i < sb->map_nr; i++) { | |
263 | const struct sbitmap_word *word = &sb->map[i]; | |
264 | ||
ea86ea2c JA |
265 | if (set) |
266 | weight += bitmap_weight(&word->word, word->depth); | |
267 | else | |
268 | weight += bitmap_weight(&word->cleared, word->depth); | |
88459642 OS |
269 | } |
270 | return weight; | |
271 | } | |
ea86ea2c JA |
272 | |
273 | static unsigned int sbitmap_weight(const struct sbitmap *sb) | |
274 | { | |
275 | return __sbitmap_weight(sb, true); | |
276 | } | |
277 | ||
278 | static unsigned int sbitmap_cleared(const struct sbitmap *sb) | |
279 | { | |
280 | return __sbitmap_weight(sb, false); | |
281 | } | |
88459642 | 282 | |
24af1ccf OS |
283 | void sbitmap_show(struct sbitmap *sb, struct seq_file *m) |
284 | { | |
285 | seq_printf(m, "depth=%u\n", sb->depth); | |
ea86ea2c JA |
286 | seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb)); |
287 | seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb)); | |
24af1ccf OS |
288 | seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift); |
289 | seq_printf(m, "map_nr=%u\n", sb->map_nr); | |
290 | } | |
291 | EXPORT_SYMBOL_GPL(sbitmap_show); | |
292 | ||
293 | static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte) | |
294 | { | |
295 | if ((offset & 0xf) == 0) { | |
296 | if (offset != 0) | |
297 | seq_putc(m, '\n'); | |
298 | seq_printf(m, "%08x:", offset); | |
299 | } | |
300 | if ((offset & 0x1) == 0) | |
301 | seq_putc(m, ' '); | |
302 | seq_printf(m, "%02x", byte); | |
303 | } | |
304 | ||
305 | void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m) | |
306 | { | |
307 | u8 byte = 0; | |
308 | unsigned int byte_bits = 0; | |
309 | unsigned int offset = 0; | |
310 | int i; | |
311 | ||
312 | for (i = 0; i < sb->map_nr; i++) { | |
313 | unsigned long word = READ_ONCE(sb->map[i].word); | |
314 | unsigned int word_bits = READ_ONCE(sb->map[i].depth); | |
315 | ||
316 | while (word_bits > 0) { | |
317 | unsigned int bits = min(8 - byte_bits, word_bits); | |
318 | ||
319 | byte |= (word & (BIT(bits) - 1)) << byte_bits; | |
320 | byte_bits += bits; | |
321 | if (byte_bits == 8) { | |
322 | emit_byte(m, offset, byte); | |
323 | byte = 0; | |
324 | byte_bits = 0; | |
325 | offset++; | |
326 | } | |
327 | word >>= bits; | |
328 | word_bits -= bits; | |
329 | } | |
330 | } | |
331 | if (byte_bits) { | |
332 | emit_byte(m, offset, byte); | |
333 | offset++; | |
334 | } | |
335 | if (offset) | |
336 | seq_putc(m, '\n'); | |
337 | } | |
338 | EXPORT_SYMBOL_GPL(sbitmap_bitmap_show); | |
339 | ||
a3275539 OS |
340 | static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq, |
341 | unsigned int depth) | |
88459642 OS |
342 | { |
343 | unsigned int wake_batch; | |
a3275539 | 344 | unsigned int shallow_depth; |
88459642 OS |
345 | |
346 | /* | |
347 | * For each batch, we wake up one queue. We need to make sure that our | |
a3275539 OS |
348 | * batch size is small enough that the full depth of the bitmap, |
349 | * potentially limited by a shallow depth, is enough to wake up all of | |
350 | * the queues. | |
351 | * | |
352 | * Each full word of the bitmap has bits_per_word bits, and there might | |
353 | * be a partial word. There are depth / bits_per_word full words and | |
354 | * depth % bits_per_word bits left over. In bitwise arithmetic: | |
355 | * | |
356 | * bits_per_word = 1 << shift | |
357 | * depth / bits_per_word = depth >> shift | |
358 | * depth % bits_per_word = depth & ((1 << shift) - 1) | |
359 | * | |
360 | * Each word can be limited to sbq->min_shallow_depth bits. | |
88459642 | 361 | */ |
a3275539 OS |
362 | shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth); |
363 | depth = ((depth >> sbq->sb.shift) * shallow_depth + | |
364 | min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth)); | |
365 | wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1, | |
366 | SBQ_WAKE_BATCH); | |
88459642 OS |
367 | |
368 | return wake_batch; | |
369 | } | |
370 | ||
371 | int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth, | |
f4a644db | 372 | int shift, bool round_robin, gfp_t flags, int node) |
88459642 OS |
373 | { |
374 | int ret; | |
375 | int i; | |
376 | ||
377 | ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node); | |
378 | if (ret) | |
379 | return ret; | |
380 | ||
40aabb67 OS |
381 | sbq->alloc_hint = alloc_percpu_gfp(unsigned int, flags); |
382 | if (!sbq->alloc_hint) { | |
383 | sbitmap_free(&sbq->sb); | |
384 | return -ENOMEM; | |
385 | } | |
386 | ||
98d95416 OS |
387 | if (depth && !round_robin) { |
388 | for_each_possible_cpu(i) | |
389 | *per_cpu_ptr(sbq->alloc_hint, i) = prandom_u32() % depth; | |
390 | } | |
391 | ||
a3275539 OS |
392 | sbq->min_shallow_depth = UINT_MAX; |
393 | sbq->wake_batch = sbq_calc_wake_batch(sbq, depth); | |
88459642 | 394 | atomic_set(&sbq->wake_index, 0); |
5d2ee712 | 395 | atomic_set(&sbq->ws_active, 0); |
88459642 | 396 | |
48e28166 | 397 | sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node); |
88459642 | 398 | if (!sbq->ws) { |
40aabb67 | 399 | free_percpu(sbq->alloc_hint); |
88459642 OS |
400 | sbitmap_free(&sbq->sb); |
401 | return -ENOMEM; | |
402 | } | |
403 | ||
404 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
405 | init_waitqueue_head(&sbq->ws[i].wait); | |
406 | atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch); | |
407 | } | |
f4a644db OS |
408 | |
409 | sbq->round_robin = round_robin; | |
88459642 OS |
410 | return 0; |
411 | } | |
412 | EXPORT_SYMBOL_GPL(sbitmap_queue_init_node); | |
413 | ||
a3275539 OS |
414 | static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq, |
415 | unsigned int depth) | |
88459642 | 416 | { |
a3275539 | 417 | unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth); |
6c0ca7ae OS |
418 | int i; |
419 | ||
420 | if (sbq->wake_batch != wake_batch) { | |
421 | WRITE_ONCE(sbq->wake_batch, wake_batch); | |
422 | /* | |
e6fc4649 ML |
423 | * Pairs with the memory barrier in sbitmap_queue_wake_up() |
424 | * to ensure that the batch size is updated before the wait | |
425 | * counts. | |
6c0ca7ae | 426 | */ |
a0934fd2 | 427 | smp_mb(); |
6c0ca7ae OS |
428 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) |
429 | atomic_set(&sbq->ws[i].wait_cnt, 1); | |
430 | } | |
a3275539 OS |
431 | } |
432 | ||
433 | void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth) | |
434 | { | |
435 | sbitmap_queue_update_wake_batch(sbq, depth); | |
88459642 OS |
436 | sbitmap_resize(&sbq->sb, depth); |
437 | } | |
438 | EXPORT_SYMBOL_GPL(sbitmap_queue_resize); | |
439 | ||
f4a644db | 440 | int __sbitmap_queue_get(struct sbitmap_queue *sbq) |
40aabb67 | 441 | { |
05fd095d | 442 | unsigned int hint, depth; |
40aabb67 OS |
443 | int nr; |
444 | ||
445 | hint = this_cpu_read(*sbq->alloc_hint); | |
05fd095d OS |
446 | depth = READ_ONCE(sbq->sb.depth); |
447 | if (unlikely(hint >= depth)) { | |
448 | hint = depth ? prandom_u32() % depth : 0; | |
449 | this_cpu_write(*sbq->alloc_hint, hint); | |
450 | } | |
f4a644db | 451 | nr = sbitmap_get(&sbq->sb, hint, sbq->round_robin); |
40aabb67 OS |
452 | |
453 | if (nr == -1) { | |
454 | /* If the map is full, a hint won't do us much good. */ | |
455 | this_cpu_write(*sbq->alloc_hint, 0); | |
f4a644db | 456 | } else if (nr == hint || unlikely(sbq->round_robin)) { |
40aabb67 OS |
457 | /* Only update the hint if we used it. */ |
458 | hint = nr + 1; | |
05fd095d | 459 | if (hint >= depth - 1) |
40aabb67 OS |
460 | hint = 0; |
461 | this_cpu_write(*sbq->alloc_hint, hint); | |
462 | } | |
463 | ||
464 | return nr; | |
465 | } | |
466 | EXPORT_SYMBOL_GPL(__sbitmap_queue_get); | |
467 | ||
c05e6673 OS |
468 | int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq, |
469 | unsigned int shallow_depth) | |
470 | { | |
471 | unsigned int hint, depth; | |
472 | int nr; | |
473 | ||
61445b56 OS |
474 | WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth); |
475 | ||
c05e6673 OS |
476 | hint = this_cpu_read(*sbq->alloc_hint); |
477 | depth = READ_ONCE(sbq->sb.depth); | |
478 | if (unlikely(hint >= depth)) { | |
479 | hint = depth ? prandom_u32() % depth : 0; | |
480 | this_cpu_write(*sbq->alloc_hint, hint); | |
481 | } | |
482 | nr = sbitmap_get_shallow(&sbq->sb, hint, shallow_depth); | |
483 | ||
484 | if (nr == -1) { | |
485 | /* If the map is full, a hint won't do us much good. */ | |
486 | this_cpu_write(*sbq->alloc_hint, 0); | |
487 | } else if (nr == hint || unlikely(sbq->round_robin)) { | |
488 | /* Only update the hint if we used it. */ | |
489 | hint = nr + 1; | |
490 | if (hint >= depth - 1) | |
491 | hint = 0; | |
492 | this_cpu_write(*sbq->alloc_hint, hint); | |
493 | } | |
494 | ||
495 | return nr; | |
496 | } | |
497 | EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow); | |
498 | ||
a3275539 OS |
499 | void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq, |
500 | unsigned int min_shallow_depth) | |
501 | { | |
502 | sbq->min_shallow_depth = min_shallow_depth; | |
503 | sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth); | |
504 | } | |
505 | EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth); | |
506 | ||
88459642 OS |
507 | static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq) |
508 | { | |
509 | int i, wake_index; | |
510 | ||
5d2ee712 JA |
511 | if (!atomic_read(&sbq->ws_active)) |
512 | return NULL; | |
513 | ||
88459642 OS |
514 | wake_index = atomic_read(&sbq->wake_index); |
515 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
516 | struct sbq_wait_state *ws = &sbq->ws[wake_index]; | |
517 | ||
518 | if (waitqueue_active(&ws->wait)) { | |
519 | int o = atomic_read(&sbq->wake_index); | |
520 | ||
521 | if (wake_index != o) | |
522 | atomic_cmpxchg(&sbq->wake_index, o, wake_index); | |
523 | return ws; | |
524 | } | |
525 | ||
526 | wake_index = sbq_index_inc(wake_index); | |
527 | } | |
528 | ||
529 | return NULL; | |
530 | } | |
531 | ||
c854ab57 | 532 | static bool __sbq_wake_up(struct sbitmap_queue *sbq) |
88459642 OS |
533 | { |
534 | struct sbq_wait_state *ws; | |
6c0ca7ae | 535 | unsigned int wake_batch; |
88459642 OS |
536 | int wait_cnt; |
537 | ||
88459642 OS |
538 | ws = sbq_wake_ptr(sbq); |
539 | if (!ws) | |
c854ab57 | 540 | return false; |
88459642 OS |
541 | |
542 | wait_cnt = atomic_dec_return(&ws->wait_cnt); | |
6c0ca7ae | 543 | if (wait_cnt <= 0) { |
c854ab57 JA |
544 | int ret; |
545 | ||
6c0ca7ae | 546 | wake_batch = READ_ONCE(sbq->wake_batch); |
c854ab57 | 547 | |
6c0ca7ae OS |
548 | /* |
549 | * Pairs with the memory barrier in sbitmap_queue_resize() to | |
550 | * ensure that we see the batch size update before the wait | |
551 | * count is reset. | |
552 | */ | |
553 | smp_mb__before_atomic(); | |
c854ab57 | 554 | |
6c0ca7ae | 555 | /* |
c854ab57 JA |
556 | * For concurrent callers of this, the one that failed the |
557 | * atomic_cmpxhcg() race should call this function again | |
558 | * to wakeup a new batch on a different 'ws'. | |
6c0ca7ae | 559 | */ |
c854ab57 JA |
560 | ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch); |
561 | if (ret == wait_cnt) { | |
562 | sbq_index_atomic_inc(&sbq->wake_index); | |
563 | wake_up_nr(&ws->wait, wake_batch); | |
564 | return false; | |
565 | } | |
566 | ||
567 | return true; | |
88459642 | 568 | } |
c854ab57 JA |
569 | |
570 | return false; | |
571 | } | |
572 | ||
e6fc4649 | 573 | void sbitmap_queue_wake_up(struct sbitmap_queue *sbq) |
c854ab57 JA |
574 | { |
575 | while (__sbq_wake_up(sbq)) | |
576 | ; | |
88459642 | 577 | } |
e6fc4649 | 578 | EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up); |
88459642 | 579 | |
40aabb67 | 580 | void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr, |
f4a644db | 581 | unsigned int cpu) |
88459642 | 582 | { |
e6d1fa58 ML |
583 | /* |
584 | * Once the clear bit is set, the bit may be allocated out. | |
585 | * | |
586 | * Orders READ/WRITE on the asssociated instance(such as request | |
587 | * of blk_mq) by this bit for avoiding race with re-allocation, | |
588 | * and its pair is the memory barrier implied in __sbitmap_get_word. | |
589 | * | |
590 | * One invariant is that the clear bit has to be zero when the bit | |
591 | * is in use. | |
592 | */ | |
593 | smp_mb__before_atomic(); | |
ea86ea2c JA |
594 | sbitmap_deferred_clear_bit(&sbq->sb, nr); |
595 | ||
e6fc4649 ML |
596 | /* |
597 | * Pairs with the memory barrier in set_current_state() to ensure the | |
598 | * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker | |
599 | * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the | |
600 | * waiter. See the comment on waitqueue_active(). | |
601 | */ | |
602 | smp_mb__after_atomic(); | |
603 | sbitmap_queue_wake_up(sbq); | |
604 | ||
5c64a8df | 605 | if (likely(!sbq->round_robin && nr < sbq->sb.depth)) |
40aabb67 | 606 | *per_cpu_ptr(sbq->alloc_hint, cpu) = nr; |
88459642 OS |
607 | } |
608 | EXPORT_SYMBOL_GPL(sbitmap_queue_clear); | |
609 | ||
610 | void sbitmap_queue_wake_all(struct sbitmap_queue *sbq) | |
611 | { | |
612 | int i, wake_index; | |
613 | ||
614 | /* | |
f66227de | 615 | * Pairs with the memory barrier in set_current_state() like in |
e6fc4649 | 616 | * sbitmap_queue_wake_up(). |
88459642 OS |
617 | */ |
618 | smp_mb(); | |
619 | wake_index = atomic_read(&sbq->wake_index); | |
620 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
621 | struct sbq_wait_state *ws = &sbq->ws[wake_index]; | |
622 | ||
623 | if (waitqueue_active(&ws->wait)) | |
624 | wake_up(&ws->wait); | |
625 | ||
626 | wake_index = sbq_index_inc(wake_index); | |
627 | } | |
628 | } | |
629 | EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all); | |
24af1ccf OS |
630 | |
631 | void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m) | |
632 | { | |
633 | bool first; | |
634 | int i; | |
635 | ||
636 | sbitmap_show(&sbq->sb, m); | |
637 | ||
638 | seq_puts(m, "alloc_hint={"); | |
639 | first = true; | |
640 | for_each_possible_cpu(i) { | |
641 | if (!first) | |
642 | seq_puts(m, ", "); | |
643 | first = false; | |
644 | seq_printf(m, "%u", *per_cpu_ptr(sbq->alloc_hint, i)); | |
645 | } | |
646 | seq_puts(m, "}\n"); | |
647 | ||
648 | seq_printf(m, "wake_batch=%u\n", sbq->wake_batch); | |
649 | seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index)); | |
5d2ee712 | 650 | seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active)); |
24af1ccf OS |
651 | |
652 | seq_puts(m, "ws={\n"); | |
653 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
654 | struct sbq_wait_state *ws = &sbq->ws[i]; | |
655 | ||
656 | seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n", | |
657 | atomic_read(&ws->wait_cnt), | |
658 | waitqueue_active(&ws->wait) ? "active" : "inactive"); | |
659 | } | |
660 | seq_puts(m, "}\n"); | |
661 | ||
662 | seq_printf(m, "round_robin=%d\n", sbq->round_robin); | |
a3275539 | 663 | seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth); |
24af1ccf OS |
664 | } |
665 | EXPORT_SYMBOL_GPL(sbitmap_queue_show); | |
5d2ee712 | 666 | |
9f6b7ef6 JA |
667 | void sbitmap_add_wait_queue(struct sbitmap_queue *sbq, |
668 | struct sbq_wait_state *ws, | |
669 | struct sbq_wait *sbq_wait) | |
670 | { | |
671 | if (!sbq_wait->sbq) { | |
672 | sbq_wait->sbq = sbq; | |
673 | atomic_inc(&sbq->ws_active); | |
674 | } | |
675 | add_wait_queue(&ws->wait, &sbq_wait->wait); | |
676 | } | |
677 | EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue); | |
678 | ||
679 | void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait) | |
680 | { | |
681 | list_del_init(&sbq_wait->wait.entry); | |
682 | if (sbq_wait->sbq) { | |
683 | atomic_dec(&sbq_wait->sbq->ws_active); | |
684 | sbq_wait->sbq = NULL; | |
685 | } | |
686 | } | |
687 | EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue); | |
688 | ||
5d2ee712 JA |
689 | void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq, |
690 | struct sbq_wait_state *ws, | |
691 | struct sbq_wait *sbq_wait, int state) | |
692 | { | |
9f6b7ef6 | 693 | if (!sbq_wait->sbq) { |
5d2ee712 | 694 | atomic_inc(&sbq->ws_active); |
9f6b7ef6 | 695 | sbq_wait->sbq = sbq; |
5d2ee712 JA |
696 | } |
697 | prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state); | |
698 | } | |
699 | EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait); | |
700 | ||
701 | void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws, | |
702 | struct sbq_wait *sbq_wait) | |
703 | { | |
704 | finish_wait(&ws->wait, &sbq_wait->wait); | |
9f6b7ef6 | 705 | if (sbq_wait->sbq) { |
5d2ee712 | 706 | atomic_dec(&sbq->ws_active); |
9f6b7ef6 | 707 | sbq_wait->sbq = NULL; |
5d2ee712 JA |
708 | } |
709 | } | |
710 | EXPORT_SYMBOL_GPL(sbitmap_finish_wait); |