]>
Commit | Line | Data |
---|---|---|
88459642 OS |
1 | /* |
2 | * Copyright (C) 2016 Facebook | |
3 | * Copyright (C) 2013-2014 Jens Axboe | |
4 | * | |
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. | |
8 | * | |
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. | |
13 | * | |
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/>. | |
16 | */ | |
17 | ||
af8601ad | 18 | #include <linux/sched.h> |
98d95416 | 19 | #include <linux/random.h> |
88459642 | 20 | #include <linux/sbitmap.h> |
24af1ccf | 21 | #include <linux/seq_file.h> |
88459642 | 22 | |
b2dbff1b JA |
23 | /* |
24 | * See if we have deferred clears that we can batch move | |
25 | */ | |
26 | static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index) | |
27 | { | |
28 | unsigned long mask, val; | |
b2dbff1b JA |
29 | bool ret = false; |
30 | ||
37198768 | 31 | spin_lock_bh(&sb->map[index].swap_lock); |
b2dbff1b JA |
32 | |
33 | if (!sb->map[index].cleared) | |
34 | goto out_unlock; | |
35 | ||
36 | /* | |
37 | * First get a stable cleared mask, setting the old mask to 0. | |
38 | */ | |
39 | do { | |
40 | mask = sb->map[index].cleared; | |
41 | } while (cmpxchg(&sb->map[index].cleared, mask, 0) != mask); | |
42 | ||
43 | /* | |
44 | * Now clear the masked bits in our free word | |
45 | */ | |
46 | do { | |
47 | val = sb->map[index].word; | |
48 | } while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val); | |
49 | ||
50 | ret = true; | |
51 | out_unlock: | |
37198768 | 52 | spin_unlock_bh(&sb->map[index].swap_lock); |
b2dbff1b JA |
53 | return ret; |
54 | } | |
55 | ||
88459642 OS |
56 | int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift, |
57 | gfp_t flags, int node) | |
58 | { | |
59 | unsigned int bits_per_word; | |
60 | unsigned int i; | |
61 | ||
62 | if (shift < 0) { | |
63 | shift = ilog2(BITS_PER_LONG); | |
64 | /* | |
65 | * If the bitmap is small, shrink the number of bits per word so | |
66 | * we spread over a few cachelines, at least. If less than 4 | |
67 | * bits, just forget about it, it's not going to work optimally | |
68 | * anyway. | |
69 | */ | |
70 | if (depth >= 4) { | |
71 | while ((4U << shift) > depth) | |
72 | shift--; | |
73 | } | |
74 | } | |
75 | bits_per_word = 1U << shift; | |
76 | if (bits_per_word > BITS_PER_LONG) | |
77 | return -EINVAL; | |
78 | ||
79 | sb->shift = shift; | |
80 | sb->depth = depth; | |
81 | sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); | |
82 | ||
83 | if (depth == 0) { | |
84 | sb->map = NULL; | |
85 | return 0; | |
86 | } | |
87 | ||
590b5b7d | 88 | sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node); |
88459642 OS |
89 | if (!sb->map) |
90 | return -ENOMEM; | |
91 | ||
92 | for (i = 0; i < sb->map_nr; i++) { | |
93 | sb->map[i].depth = min(depth, bits_per_word); | |
94 | depth -= sb->map[i].depth; | |
ea86ea2c | 95 | spin_lock_init(&sb->map[i].swap_lock); |
88459642 OS |
96 | } |
97 | return 0; | |
98 | } | |
99 | EXPORT_SYMBOL_GPL(sbitmap_init_node); | |
100 | ||
101 | void sbitmap_resize(struct sbitmap *sb, unsigned int depth) | |
102 | { | |
103 | unsigned int bits_per_word = 1U << sb->shift; | |
104 | unsigned int i; | |
105 | ||
b2dbff1b JA |
106 | for (i = 0; i < sb->map_nr; i++) |
107 | sbitmap_deferred_clear(sb, i); | |
108 | ||
88459642 OS |
109 | sb->depth = depth; |
110 | sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); | |
111 | ||
112 | for (i = 0; i < sb->map_nr; i++) { | |
113 | sb->map[i].depth = min(depth, bits_per_word); | |
114 | depth -= sb->map[i].depth; | |
115 | } | |
116 | } | |
117 | EXPORT_SYMBOL_GPL(sbitmap_resize); | |
118 | ||
c05e6673 OS |
119 | static int __sbitmap_get_word(unsigned long *word, unsigned long depth, |
120 | unsigned int hint, bool wrap) | |
88459642 OS |
121 | { |
122 | unsigned int orig_hint = hint; | |
123 | int nr; | |
124 | ||
125 | while (1) { | |
c05e6673 OS |
126 | nr = find_next_zero_bit(word, depth, hint); |
127 | if (unlikely(nr >= depth)) { | |
88459642 OS |
128 | /* |
129 | * We started with an offset, and we didn't reset the | |
130 | * offset to 0 in a failure case, so start from 0 to | |
131 | * exhaust the map. | |
132 | */ | |
133 | if (orig_hint && hint && wrap) { | |
134 | hint = orig_hint = 0; | |
135 | continue; | |
136 | } | |
137 | return -1; | |
138 | } | |
139 | ||
4ace53f1 | 140 | if (!test_and_set_bit_lock(nr, word)) |
88459642 OS |
141 | break; |
142 | ||
143 | hint = nr + 1; | |
c05e6673 | 144 | if (hint >= depth - 1) |
88459642 OS |
145 | hint = 0; |
146 | } | |
147 | ||
148 | return nr; | |
149 | } | |
150 | ||
ea86ea2c JA |
151 | static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index, |
152 | unsigned int alloc_hint, bool round_robin) | |
153 | { | |
154 | int nr; | |
155 | ||
156 | do { | |
157 | nr = __sbitmap_get_word(&sb->map[index].word, | |
158 | sb->map[index].depth, alloc_hint, | |
159 | !round_robin); | |
160 | if (nr != -1) | |
161 | break; | |
162 | if (!sbitmap_deferred_clear(sb, index)) | |
163 | break; | |
164 | } while (1); | |
165 | ||
166 | return nr; | |
167 | } | |
168 | ||
88459642 OS |
169 | int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin) |
170 | { | |
171 | unsigned int i, index; | |
172 | int nr = -1; | |
173 | ||
174 | index = SB_NR_TO_INDEX(sb, alloc_hint); | |
175 | ||
27fae429 JA |
176 | /* |
177 | * Unless we're doing round robin tag allocation, just use the | |
178 | * alloc_hint to find the right word index. No point in looping | |
179 | * twice in find_next_zero_bit() for that case. | |
180 | */ | |
181 | if (round_robin) | |
182 | alloc_hint = SB_NR_TO_BIT(sb, alloc_hint); | |
183 | else | |
184 | alloc_hint = 0; | |
185 | ||
88459642 | 186 | for (i = 0; i < sb->map_nr; i++) { |
ea86ea2c JA |
187 | nr = sbitmap_find_bit_in_index(sb, index, alloc_hint, |
188 | round_robin); | |
88459642 OS |
189 | if (nr != -1) { |
190 | nr += index << sb->shift; | |
191 | break; | |
192 | } | |
193 | ||
194 | /* Jump to next index. */ | |
27fae429 JA |
195 | alloc_hint = 0; |
196 | if (++index >= sb->map_nr) | |
88459642 | 197 | index = 0; |
88459642 OS |
198 | } |
199 | ||
200 | return nr; | |
201 | } | |
202 | EXPORT_SYMBOL_GPL(sbitmap_get); | |
203 | ||
c05e6673 OS |
204 | int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint, |
205 | unsigned long shallow_depth) | |
206 | { | |
207 | unsigned int i, index; | |
208 | int nr = -1; | |
209 | ||
210 | index = SB_NR_TO_INDEX(sb, alloc_hint); | |
211 | ||
212 | for (i = 0; i < sb->map_nr; i++) { | |
b2dbff1b | 213 | again: |
c05e6673 OS |
214 | nr = __sbitmap_get_word(&sb->map[index].word, |
215 | min(sb->map[index].depth, shallow_depth), | |
216 | SB_NR_TO_BIT(sb, alloc_hint), true); | |
217 | if (nr != -1) { | |
218 | nr += index << sb->shift; | |
219 | break; | |
220 | } | |
221 | ||
b2dbff1b JA |
222 | if (sbitmap_deferred_clear(sb, index)) |
223 | goto again; | |
224 | ||
c05e6673 OS |
225 | /* Jump to next index. */ |
226 | index++; | |
227 | alloc_hint = index << sb->shift; | |
228 | ||
229 | if (index >= sb->map_nr) { | |
230 | index = 0; | |
231 | alloc_hint = 0; | |
232 | } | |
233 | } | |
234 | ||
235 | return nr; | |
236 | } | |
237 | EXPORT_SYMBOL_GPL(sbitmap_get_shallow); | |
238 | ||
88459642 OS |
239 | bool sbitmap_any_bit_set(const struct sbitmap *sb) |
240 | { | |
241 | unsigned int i; | |
242 | ||
243 | for (i = 0; i < sb->map_nr; i++) { | |
b2dbff1b | 244 | if (sb->map[i].word & ~sb->map[i].cleared) |
88459642 OS |
245 | return true; |
246 | } | |
247 | return false; | |
248 | } | |
249 | EXPORT_SYMBOL_GPL(sbitmap_any_bit_set); | |
250 | ||
251 | bool sbitmap_any_bit_clear(const struct sbitmap *sb) | |
252 | { | |
253 | unsigned int i; | |
254 | ||
255 | for (i = 0; i < sb->map_nr; i++) { | |
256 | const struct sbitmap_word *word = &sb->map[i]; | |
b2dbff1b | 257 | unsigned long mask = word->word & ~word->cleared; |
88459642 OS |
258 | unsigned long ret; |
259 | ||
b2dbff1b | 260 | ret = find_first_zero_bit(&mask, word->depth); |
88459642 OS |
261 | if (ret < word->depth) |
262 | return true; | |
263 | } | |
264 | return false; | |
265 | } | |
266 | EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear); | |
267 | ||
ea86ea2c | 268 | static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set) |
88459642 | 269 | { |
60658e0d | 270 | unsigned int i, weight = 0; |
88459642 OS |
271 | |
272 | for (i = 0; i < sb->map_nr; i++) { | |
273 | const struct sbitmap_word *word = &sb->map[i]; | |
274 | ||
ea86ea2c JA |
275 | if (set) |
276 | weight += bitmap_weight(&word->word, word->depth); | |
277 | else | |
278 | weight += bitmap_weight(&word->cleared, word->depth); | |
88459642 OS |
279 | } |
280 | return weight; | |
281 | } | |
ea86ea2c JA |
282 | |
283 | static unsigned int sbitmap_weight(const struct sbitmap *sb) | |
284 | { | |
285 | return __sbitmap_weight(sb, true); | |
286 | } | |
287 | ||
288 | static unsigned int sbitmap_cleared(const struct sbitmap *sb) | |
289 | { | |
290 | return __sbitmap_weight(sb, false); | |
291 | } | |
88459642 | 292 | |
24af1ccf OS |
293 | void sbitmap_show(struct sbitmap *sb, struct seq_file *m) |
294 | { | |
295 | seq_printf(m, "depth=%u\n", sb->depth); | |
ea86ea2c JA |
296 | seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb)); |
297 | seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb)); | |
24af1ccf OS |
298 | seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift); |
299 | seq_printf(m, "map_nr=%u\n", sb->map_nr); | |
300 | } | |
301 | EXPORT_SYMBOL_GPL(sbitmap_show); | |
302 | ||
303 | static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte) | |
304 | { | |
305 | if ((offset & 0xf) == 0) { | |
306 | if (offset != 0) | |
307 | seq_putc(m, '\n'); | |
308 | seq_printf(m, "%08x:", offset); | |
309 | } | |
310 | if ((offset & 0x1) == 0) | |
311 | seq_putc(m, ' '); | |
312 | seq_printf(m, "%02x", byte); | |
313 | } | |
314 | ||
315 | void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m) | |
316 | { | |
317 | u8 byte = 0; | |
318 | unsigned int byte_bits = 0; | |
319 | unsigned int offset = 0; | |
320 | int i; | |
321 | ||
322 | for (i = 0; i < sb->map_nr; i++) { | |
323 | unsigned long word = READ_ONCE(sb->map[i].word); | |
324 | unsigned int word_bits = READ_ONCE(sb->map[i].depth); | |
325 | ||
326 | while (word_bits > 0) { | |
327 | unsigned int bits = min(8 - byte_bits, word_bits); | |
328 | ||
329 | byte |= (word & (BIT(bits) - 1)) << byte_bits; | |
330 | byte_bits += bits; | |
331 | if (byte_bits == 8) { | |
332 | emit_byte(m, offset, byte); | |
333 | byte = 0; | |
334 | byte_bits = 0; | |
335 | offset++; | |
336 | } | |
337 | word >>= bits; | |
338 | word_bits -= bits; | |
339 | } | |
340 | } | |
341 | if (byte_bits) { | |
342 | emit_byte(m, offset, byte); | |
343 | offset++; | |
344 | } | |
345 | if (offset) | |
346 | seq_putc(m, '\n'); | |
347 | } | |
348 | EXPORT_SYMBOL_GPL(sbitmap_bitmap_show); | |
349 | ||
a3275539 OS |
350 | static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq, |
351 | unsigned int depth) | |
88459642 OS |
352 | { |
353 | unsigned int wake_batch; | |
a3275539 | 354 | unsigned int shallow_depth; |
88459642 OS |
355 | |
356 | /* | |
357 | * For each batch, we wake up one queue. We need to make sure that our | |
a3275539 OS |
358 | * batch size is small enough that the full depth of the bitmap, |
359 | * potentially limited by a shallow depth, is enough to wake up all of | |
360 | * the queues. | |
361 | * | |
362 | * Each full word of the bitmap has bits_per_word bits, and there might | |
363 | * be a partial word. There are depth / bits_per_word full words and | |
364 | * depth % bits_per_word bits left over. In bitwise arithmetic: | |
365 | * | |
366 | * bits_per_word = 1 << shift | |
367 | * depth / bits_per_word = depth >> shift | |
368 | * depth % bits_per_word = depth & ((1 << shift) - 1) | |
369 | * | |
370 | * Each word can be limited to sbq->min_shallow_depth bits. | |
88459642 | 371 | */ |
a3275539 OS |
372 | shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth); |
373 | depth = ((depth >> sbq->sb.shift) * shallow_depth + | |
374 | min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth)); | |
375 | wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1, | |
376 | SBQ_WAKE_BATCH); | |
88459642 OS |
377 | |
378 | return wake_batch; | |
379 | } | |
380 | ||
381 | int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth, | |
f4a644db | 382 | int shift, bool round_robin, gfp_t flags, int node) |
88459642 OS |
383 | { |
384 | int ret; | |
385 | int i; | |
386 | ||
387 | ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node); | |
388 | if (ret) | |
389 | return ret; | |
390 | ||
40aabb67 OS |
391 | sbq->alloc_hint = alloc_percpu_gfp(unsigned int, flags); |
392 | if (!sbq->alloc_hint) { | |
393 | sbitmap_free(&sbq->sb); | |
394 | return -ENOMEM; | |
395 | } | |
396 | ||
98d95416 OS |
397 | if (depth && !round_robin) { |
398 | for_each_possible_cpu(i) | |
399 | *per_cpu_ptr(sbq->alloc_hint, i) = prandom_u32() % depth; | |
400 | } | |
401 | ||
a3275539 OS |
402 | sbq->min_shallow_depth = UINT_MAX; |
403 | sbq->wake_batch = sbq_calc_wake_batch(sbq, depth); | |
88459642 | 404 | atomic_set(&sbq->wake_index, 0); |
5d2ee712 | 405 | atomic_set(&sbq->ws_active, 0); |
88459642 | 406 | |
48e28166 | 407 | sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node); |
88459642 | 408 | if (!sbq->ws) { |
40aabb67 | 409 | free_percpu(sbq->alloc_hint); |
88459642 OS |
410 | sbitmap_free(&sbq->sb); |
411 | return -ENOMEM; | |
412 | } | |
413 | ||
414 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
415 | init_waitqueue_head(&sbq->ws[i].wait); | |
416 | atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch); | |
417 | } | |
f4a644db OS |
418 | |
419 | sbq->round_robin = round_robin; | |
88459642 OS |
420 | return 0; |
421 | } | |
422 | EXPORT_SYMBOL_GPL(sbitmap_queue_init_node); | |
423 | ||
a3275539 OS |
424 | static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq, |
425 | unsigned int depth) | |
88459642 | 426 | { |
a3275539 | 427 | unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth); |
6c0ca7ae OS |
428 | int i; |
429 | ||
430 | if (sbq->wake_batch != wake_batch) { | |
431 | WRITE_ONCE(sbq->wake_batch, wake_batch); | |
432 | /* | |
e6fc4649 ML |
433 | * Pairs with the memory barrier in sbitmap_queue_wake_up() |
434 | * to ensure that the batch size is updated before the wait | |
435 | * counts. | |
6c0ca7ae OS |
436 | */ |
437 | smp_mb__before_atomic(); | |
438 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) | |
439 | atomic_set(&sbq->ws[i].wait_cnt, 1); | |
440 | } | |
a3275539 OS |
441 | } |
442 | ||
443 | void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth) | |
444 | { | |
445 | sbitmap_queue_update_wake_batch(sbq, depth); | |
88459642 OS |
446 | sbitmap_resize(&sbq->sb, depth); |
447 | } | |
448 | EXPORT_SYMBOL_GPL(sbitmap_queue_resize); | |
449 | ||
f4a644db | 450 | int __sbitmap_queue_get(struct sbitmap_queue *sbq) |
40aabb67 | 451 | { |
05fd095d | 452 | unsigned int hint, depth; |
40aabb67 OS |
453 | int nr; |
454 | ||
455 | hint = this_cpu_read(*sbq->alloc_hint); | |
05fd095d OS |
456 | depth = READ_ONCE(sbq->sb.depth); |
457 | if (unlikely(hint >= depth)) { | |
458 | hint = depth ? prandom_u32() % depth : 0; | |
459 | this_cpu_write(*sbq->alloc_hint, hint); | |
460 | } | |
f4a644db | 461 | nr = sbitmap_get(&sbq->sb, hint, sbq->round_robin); |
40aabb67 OS |
462 | |
463 | if (nr == -1) { | |
464 | /* If the map is full, a hint won't do us much good. */ | |
465 | this_cpu_write(*sbq->alloc_hint, 0); | |
f4a644db | 466 | } else if (nr == hint || unlikely(sbq->round_robin)) { |
40aabb67 OS |
467 | /* Only update the hint if we used it. */ |
468 | hint = nr + 1; | |
05fd095d | 469 | if (hint >= depth - 1) |
40aabb67 OS |
470 | hint = 0; |
471 | this_cpu_write(*sbq->alloc_hint, hint); | |
472 | } | |
473 | ||
474 | return nr; | |
475 | } | |
476 | EXPORT_SYMBOL_GPL(__sbitmap_queue_get); | |
477 | ||
c05e6673 OS |
478 | int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq, |
479 | unsigned int shallow_depth) | |
480 | { | |
481 | unsigned int hint, depth; | |
482 | int nr; | |
483 | ||
61445b56 OS |
484 | WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth); |
485 | ||
c05e6673 OS |
486 | hint = this_cpu_read(*sbq->alloc_hint); |
487 | depth = READ_ONCE(sbq->sb.depth); | |
488 | if (unlikely(hint >= depth)) { | |
489 | hint = depth ? prandom_u32() % depth : 0; | |
490 | this_cpu_write(*sbq->alloc_hint, hint); | |
491 | } | |
492 | nr = sbitmap_get_shallow(&sbq->sb, hint, shallow_depth); | |
493 | ||
494 | if (nr == -1) { | |
495 | /* If the map is full, a hint won't do us much good. */ | |
496 | this_cpu_write(*sbq->alloc_hint, 0); | |
497 | } else if (nr == hint || unlikely(sbq->round_robin)) { | |
498 | /* Only update the hint if we used it. */ | |
499 | hint = nr + 1; | |
500 | if (hint >= depth - 1) | |
501 | hint = 0; | |
502 | this_cpu_write(*sbq->alloc_hint, hint); | |
503 | } | |
504 | ||
505 | return nr; | |
506 | } | |
507 | EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow); | |
508 | ||
a3275539 OS |
509 | void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq, |
510 | unsigned int min_shallow_depth) | |
511 | { | |
512 | sbq->min_shallow_depth = min_shallow_depth; | |
513 | sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth); | |
514 | } | |
515 | EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth); | |
516 | ||
88459642 OS |
517 | static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq) |
518 | { | |
519 | int i, wake_index; | |
520 | ||
5d2ee712 JA |
521 | if (!atomic_read(&sbq->ws_active)) |
522 | return NULL; | |
523 | ||
88459642 OS |
524 | wake_index = atomic_read(&sbq->wake_index); |
525 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
526 | struct sbq_wait_state *ws = &sbq->ws[wake_index]; | |
527 | ||
528 | if (waitqueue_active(&ws->wait)) { | |
529 | int o = atomic_read(&sbq->wake_index); | |
530 | ||
531 | if (wake_index != o) | |
532 | atomic_cmpxchg(&sbq->wake_index, o, wake_index); | |
533 | return ws; | |
534 | } | |
535 | ||
536 | wake_index = sbq_index_inc(wake_index); | |
537 | } | |
538 | ||
539 | return NULL; | |
540 | } | |
541 | ||
c854ab57 | 542 | static bool __sbq_wake_up(struct sbitmap_queue *sbq) |
88459642 OS |
543 | { |
544 | struct sbq_wait_state *ws; | |
6c0ca7ae | 545 | unsigned int wake_batch; |
88459642 OS |
546 | int wait_cnt; |
547 | ||
88459642 OS |
548 | ws = sbq_wake_ptr(sbq); |
549 | if (!ws) | |
c854ab57 | 550 | return false; |
88459642 OS |
551 | |
552 | wait_cnt = atomic_dec_return(&ws->wait_cnt); | |
6c0ca7ae | 553 | if (wait_cnt <= 0) { |
c854ab57 JA |
554 | int ret; |
555 | ||
6c0ca7ae | 556 | wake_batch = READ_ONCE(sbq->wake_batch); |
c854ab57 | 557 | |
6c0ca7ae OS |
558 | /* |
559 | * Pairs with the memory barrier in sbitmap_queue_resize() to | |
560 | * ensure that we see the batch size update before the wait | |
561 | * count is reset. | |
562 | */ | |
563 | smp_mb__before_atomic(); | |
c854ab57 | 564 | |
6c0ca7ae | 565 | /* |
c854ab57 JA |
566 | * For concurrent callers of this, the one that failed the |
567 | * atomic_cmpxhcg() race should call this function again | |
568 | * to wakeup a new batch on a different 'ws'. | |
6c0ca7ae | 569 | */ |
c854ab57 JA |
570 | ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch); |
571 | if (ret == wait_cnt) { | |
572 | sbq_index_atomic_inc(&sbq->wake_index); | |
573 | wake_up_nr(&ws->wait, wake_batch); | |
574 | return false; | |
575 | } | |
576 | ||
577 | return true; | |
88459642 | 578 | } |
c854ab57 JA |
579 | |
580 | return false; | |
581 | } | |
582 | ||
e6fc4649 | 583 | void sbitmap_queue_wake_up(struct sbitmap_queue *sbq) |
c854ab57 JA |
584 | { |
585 | while (__sbq_wake_up(sbq)) | |
586 | ; | |
88459642 | 587 | } |
e6fc4649 | 588 | EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up); |
88459642 | 589 | |
40aabb67 | 590 | void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr, |
f4a644db | 591 | unsigned int cpu) |
88459642 | 592 | { |
ea86ea2c JA |
593 | sbitmap_deferred_clear_bit(&sbq->sb, nr); |
594 | ||
e6fc4649 ML |
595 | /* |
596 | * Pairs with the memory barrier in set_current_state() to ensure the | |
597 | * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker | |
598 | * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the | |
599 | * waiter. See the comment on waitqueue_active(). | |
600 | */ | |
601 | smp_mb__after_atomic(); | |
602 | sbitmap_queue_wake_up(sbq); | |
603 | ||
5c64a8df | 604 | if (likely(!sbq->round_robin && nr < sbq->sb.depth)) |
40aabb67 | 605 | *per_cpu_ptr(sbq->alloc_hint, cpu) = nr; |
88459642 OS |
606 | } |
607 | EXPORT_SYMBOL_GPL(sbitmap_queue_clear); | |
608 | ||
609 | void sbitmap_queue_wake_all(struct sbitmap_queue *sbq) | |
610 | { | |
611 | int i, wake_index; | |
612 | ||
613 | /* | |
f66227de | 614 | * Pairs with the memory barrier in set_current_state() like in |
e6fc4649 | 615 | * sbitmap_queue_wake_up(). |
88459642 OS |
616 | */ |
617 | smp_mb(); | |
618 | wake_index = atomic_read(&sbq->wake_index); | |
619 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
620 | struct sbq_wait_state *ws = &sbq->ws[wake_index]; | |
621 | ||
622 | if (waitqueue_active(&ws->wait)) | |
623 | wake_up(&ws->wait); | |
624 | ||
625 | wake_index = sbq_index_inc(wake_index); | |
626 | } | |
627 | } | |
628 | EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all); | |
24af1ccf OS |
629 | |
630 | void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m) | |
631 | { | |
632 | bool first; | |
633 | int i; | |
634 | ||
635 | sbitmap_show(&sbq->sb, m); | |
636 | ||
637 | seq_puts(m, "alloc_hint={"); | |
638 | first = true; | |
639 | for_each_possible_cpu(i) { | |
640 | if (!first) | |
641 | seq_puts(m, ", "); | |
642 | first = false; | |
643 | seq_printf(m, "%u", *per_cpu_ptr(sbq->alloc_hint, i)); | |
644 | } | |
645 | seq_puts(m, "}\n"); | |
646 | ||
647 | seq_printf(m, "wake_batch=%u\n", sbq->wake_batch); | |
648 | seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index)); | |
5d2ee712 | 649 | seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active)); |
24af1ccf OS |
650 | |
651 | seq_puts(m, "ws={\n"); | |
652 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
653 | struct sbq_wait_state *ws = &sbq->ws[i]; | |
654 | ||
655 | seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n", | |
656 | atomic_read(&ws->wait_cnt), | |
657 | waitqueue_active(&ws->wait) ? "active" : "inactive"); | |
658 | } | |
659 | seq_puts(m, "}\n"); | |
660 | ||
661 | seq_printf(m, "round_robin=%d\n", sbq->round_robin); | |
a3275539 | 662 | seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth); |
24af1ccf OS |
663 | } |
664 | EXPORT_SYMBOL_GPL(sbitmap_queue_show); | |
5d2ee712 | 665 | |
9f6b7ef6 JA |
666 | void sbitmap_add_wait_queue(struct sbitmap_queue *sbq, |
667 | struct sbq_wait_state *ws, | |
668 | struct sbq_wait *sbq_wait) | |
669 | { | |
670 | if (!sbq_wait->sbq) { | |
671 | sbq_wait->sbq = sbq; | |
672 | atomic_inc(&sbq->ws_active); | |
673 | } | |
674 | add_wait_queue(&ws->wait, &sbq_wait->wait); | |
675 | } | |
676 | EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue); | |
677 | ||
678 | void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait) | |
679 | { | |
680 | list_del_init(&sbq_wait->wait.entry); | |
681 | if (sbq_wait->sbq) { | |
682 | atomic_dec(&sbq_wait->sbq->ws_active); | |
683 | sbq_wait->sbq = NULL; | |
684 | } | |
685 | } | |
686 | EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue); | |
687 | ||
5d2ee712 JA |
688 | void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq, |
689 | struct sbq_wait_state *ws, | |
690 | struct sbq_wait *sbq_wait, int state) | |
691 | { | |
9f6b7ef6 | 692 | if (!sbq_wait->sbq) { |
5d2ee712 | 693 | atomic_inc(&sbq->ws_active); |
9f6b7ef6 | 694 | sbq_wait->sbq = sbq; |
5d2ee712 JA |
695 | } |
696 | prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state); | |
697 | } | |
698 | EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait); | |
699 | ||
700 | void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws, | |
701 | struct sbq_wait *sbq_wait) | |
702 | { | |
703 | finish_wait(&ws->wait, &sbq_wait->wait); | |
9f6b7ef6 | 704 | if (sbq_wait->sbq) { |
5d2ee712 | 705 | atomic_dec(&sbq->ws_active); |
9f6b7ef6 | 706 | sbq_wait->sbq = NULL; |
5d2ee712 JA |
707 | } |
708 | } | |
709 | EXPORT_SYMBOL_GPL(sbitmap_finish_wait); |