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; | |
29 | unsigned long __maybe_unused flags; | |
30 | bool ret = false; | |
31 | ||
32 | /* Silence bogus lockdep warning */ | |
33 | #if defined(CONFIG_LOCKDEP) | |
34 | local_irq_save(flags); | |
35 | #endif | |
36 | spin_lock(&sb->map[index].swap_lock); | |
37 | ||
38 | if (!sb->map[index].cleared) | |
39 | goto out_unlock; | |
40 | ||
41 | /* | |
42 | * First get a stable cleared mask, setting the old mask to 0. | |
43 | */ | |
44 | do { | |
45 | mask = sb->map[index].cleared; | |
46 | } while (cmpxchg(&sb->map[index].cleared, mask, 0) != mask); | |
47 | ||
48 | /* | |
49 | * Now clear the masked bits in our free word | |
50 | */ | |
51 | do { | |
52 | val = sb->map[index].word; | |
53 | } while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val); | |
54 | ||
55 | ret = true; | |
56 | out_unlock: | |
57 | spin_unlock(&sb->map[index].swap_lock); | |
58 | #if defined(CONFIG_LOCKDEP) | |
59 | local_irq_restore(flags); | |
60 | #endif | |
61 | return ret; | |
62 | } | |
63 | ||
88459642 OS |
64 | int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift, |
65 | gfp_t flags, int node) | |
66 | { | |
67 | unsigned int bits_per_word; | |
68 | unsigned int i; | |
69 | ||
70 | if (shift < 0) { | |
71 | shift = ilog2(BITS_PER_LONG); | |
72 | /* | |
73 | * If the bitmap is small, shrink the number of bits per word so | |
74 | * we spread over a few cachelines, at least. If less than 4 | |
75 | * bits, just forget about it, it's not going to work optimally | |
76 | * anyway. | |
77 | */ | |
78 | if (depth >= 4) { | |
79 | while ((4U << shift) > depth) | |
80 | shift--; | |
81 | } | |
82 | } | |
83 | bits_per_word = 1U << shift; | |
84 | if (bits_per_word > BITS_PER_LONG) | |
85 | return -EINVAL; | |
86 | ||
87 | sb->shift = shift; | |
88 | sb->depth = depth; | |
89 | sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); | |
90 | ||
91 | if (depth == 0) { | |
92 | sb->map = NULL; | |
93 | return 0; | |
94 | } | |
95 | ||
590b5b7d | 96 | sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node); |
88459642 OS |
97 | if (!sb->map) |
98 | return -ENOMEM; | |
99 | ||
100 | for (i = 0; i < sb->map_nr; i++) { | |
101 | sb->map[i].depth = min(depth, bits_per_word); | |
102 | depth -= sb->map[i].depth; | |
ea86ea2c | 103 | spin_lock_init(&sb->map[i].swap_lock); |
88459642 OS |
104 | } |
105 | return 0; | |
106 | } | |
107 | EXPORT_SYMBOL_GPL(sbitmap_init_node); | |
108 | ||
109 | void sbitmap_resize(struct sbitmap *sb, unsigned int depth) | |
110 | { | |
111 | unsigned int bits_per_word = 1U << sb->shift; | |
112 | unsigned int i; | |
113 | ||
b2dbff1b JA |
114 | for (i = 0; i < sb->map_nr; i++) |
115 | sbitmap_deferred_clear(sb, i); | |
116 | ||
88459642 OS |
117 | sb->depth = depth; |
118 | sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word); | |
119 | ||
120 | for (i = 0; i < sb->map_nr; i++) { | |
121 | sb->map[i].depth = min(depth, bits_per_word); | |
122 | depth -= sb->map[i].depth; | |
123 | } | |
124 | } | |
125 | EXPORT_SYMBOL_GPL(sbitmap_resize); | |
126 | ||
c05e6673 OS |
127 | static int __sbitmap_get_word(unsigned long *word, unsigned long depth, |
128 | unsigned int hint, bool wrap) | |
88459642 OS |
129 | { |
130 | unsigned int orig_hint = hint; | |
131 | int nr; | |
132 | ||
133 | while (1) { | |
c05e6673 OS |
134 | nr = find_next_zero_bit(word, depth, hint); |
135 | if (unlikely(nr >= depth)) { | |
88459642 OS |
136 | /* |
137 | * We started with an offset, and we didn't reset the | |
138 | * offset to 0 in a failure case, so start from 0 to | |
139 | * exhaust the map. | |
140 | */ | |
141 | if (orig_hint && hint && wrap) { | |
142 | hint = orig_hint = 0; | |
143 | continue; | |
144 | } | |
145 | return -1; | |
146 | } | |
147 | ||
4ace53f1 | 148 | if (!test_and_set_bit_lock(nr, word)) |
88459642 OS |
149 | break; |
150 | ||
151 | hint = nr + 1; | |
c05e6673 | 152 | if (hint >= depth - 1) |
88459642 OS |
153 | hint = 0; |
154 | } | |
155 | ||
156 | return nr; | |
157 | } | |
158 | ||
ea86ea2c JA |
159 | static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index, |
160 | unsigned int alloc_hint, bool round_robin) | |
161 | { | |
162 | int nr; | |
163 | ||
164 | do { | |
165 | nr = __sbitmap_get_word(&sb->map[index].word, | |
166 | sb->map[index].depth, alloc_hint, | |
167 | !round_robin); | |
168 | if (nr != -1) | |
169 | break; | |
170 | if (!sbitmap_deferred_clear(sb, index)) | |
171 | break; | |
172 | } while (1); | |
173 | ||
174 | return nr; | |
175 | } | |
176 | ||
88459642 OS |
177 | int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin) |
178 | { | |
179 | unsigned int i, index; | |
180 | int nr = -1; | |
181 | ||
182 | index = SB_NR_TO_INDEX(sb, alloc_hint); | |
183 | ||
27fae429 JA |
184 | /* |
185 | * Unless we're doing round robin tag allocation, just use the | |
186 | * alloc_hint to find the right word index. No point in looping | |
187 | * twice in find_next_zero_bit() for that case. | |
188 | */ | |
189 | if (round_robin) | |
190 | alloc_hint = SB_NR_TO_BIT(sb, alloc_hint); | |
191 | else | |
192 | alloc_hint = 0; | |
193 | ||
88459642 | 194 | for (i = 0; i < sb->map_nr; i++) { |
ea86ea2c JA |
195 | nr = sbitmap_find_bit_in_index(sb, index, alloc_hint, |
196 | round_robin); | |
88459642 OS |
197 | if (nr != -1) { |
198 | nr += index << sb->shift; | |
199 | break; | |
200 | } | |
201 | ||
202 | /* Jump to next index. */ | |
27fae429 JA |
203 | alloc_hint = 0; |
204 | if (++index >= sb->map_nr) | |
88459642 | 205 | index = 0; |
88459642 OS |
206 | } |
207 | ||
208 | return nr; | |
209 | } | |
210 | EXPORT_SYMBOL_GPL(sbitmap_get); | |
211 | ||
c05e6673 OS |
212 | int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint, |
213 | unsigned long shallow_depth) | |
214 | { | |
215 | unsigned int i, index; | |
216 | int nr = -1; | |
217 | ||
218 | index = SB_NR_TO_INDEX(sb, alloc_hint); | |
219 | ||
220 | for (i = 0; i < sb->map_nr; i++) { | |
b2dbff1b | 221 | again: |
c05e6673 OS |
222 | nr = __sbitmap_get_word(&sb->map[index].word, |
223 | min(sb->map[index].depth, shallow_depth), | |
224 | SB_NR_TO_BIT(sb, alloc_hint), true); | |
225 | if (nr != -1) { | |
226 | nr += index << sb->shift; | |
227 | break; | |
228 | } | |
229 | ||
b2dbff1b JA |
230 | if (sbitmap_deferred_clear(sb, index)) |
231 | goto again; | |
232 | ||
c05e6673 OS |
233 | /* Jump to next index. */ |
234 | index++; | |
235 | alloc_hint = index << sb->shift; | |
236 | ||
237 | if (index >= sb->map_nr) { | |
238 | index = 0; | |
239 | alloc_hint = 0; | |
240 | } | |
241 | } | |
242 | ||
243 | return nr; | |
244 | } | |
245 | EXPORT_SYMBOL_GPL(sbitmap_get_shallow); | |
246 | ||
88459642 OS |
247 | bool sbitmap_any_bit_set(const struct sbitmap *sb) |
248 | { | |
249 | unsigned int i; | |
250 | ||
251 | for (i = 0; i < sb->map_nr; i++) { | |
b2dbff1b | 252 | if (sb->map[i].word & ~sb->map[i].cleared) |
88459642 OS |
253 | return true; |
254 | } | |
255 | return false; | |
256 | } | |
257 | EXPORT_SYMBOL_GPL(sbitmap_any_bit_set); | |
258 | ||
259 | bool sbitmap_any_bit_clear(const struct sbitmap *sb) | |
260 | { | |
261 | unsigned int i; | |
262 | ||
263 | for (i = 0; i < sb->map_nr; i++) { | |
264 | const struct sbitmap_word *word = &sb->map[i]; | |
b2dbff1b | 265 | unsigned long mask = word->word & ~word->cleared; |
88459642 OS |
266 | unsigned long ret; |
267 | ||
b2dbff1b | 268 | ret = find_first_zero_bit(&mask, word->depth); |
88459642 OS |
269 | if (ret < word->depth) |
270 | return true; | |
271 | } | |
272 | return false; | |
273 | } | |
274 | EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear); | |
275 | ||
ea86ea2c | 276 | static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set) |
88459642 | 277 | { |
60658e0d | 278 | unsigned int i, weight = 0; |
88459642 OS |
279 | |
280 | for (i = 0; i < sb->map_nr; i++) { | |
281 | const struct sbitmap_word *word = &sb->map[i]; | |
282 | ||
ea86ea2c JA |
283 | if (set) |
284 | weight += bitmap_weight(&word->word, word->depth); | |
285 | else | |
286 | weight += bitmap_weight(&word->cleared, word->depth); | |
88459642 OS |
287 | } |
288 | return weight; | |
289 | } | |
ea86ea2c JA |
290 | |
291 | static unsigned int sbitmap_weight(const struct sbitmap *sb) | |
292 | { | |
293 | return __sbitmap_weight(sb, true); | |
294 | } | |
295 | ||
296 | static unsigned int sbitmap_cleared(const struct sbitmap *sb) | |
297 | { | |
298 | return __sbitmap_weight(sb, false); | |
299 | } | |
88459642 | 300 | |
24af1ccf OS |
301 | void sbitmap_show(struct sbitmap *sb, struct seq_file *m) |
302 | { | |
303 | seq_printf(m, "depth=%u\n", sb->depth); | |
ea86ea2c JA |
304 | seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb)); |
305 | seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb)); | |
24af1ccf OS |
306 | seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift); |
307 | seq_printf(m, "map_nr=%u\n", sb->map_nr); | |
308 | } | |
309 | EXPORT_SYMBOL_GPL(sbitmap_show); | |
310 | ||
311 | static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte) | |
312 | { | |
313 | if ((offset & 0xf) == 0) { | |
314 | if (offset != 0) | |
315 | seq_putc(m, '\n'); | |
316 | seq_printf(m, "%08x:", offset); | |
317 | } | |
318 | if ((offset & 0x1) == 0) | |
319 | seq_putc(m, ' '); | |
320 | seq_printf(m, "%02x", byte); | |
321 | } | |
322 | ||
323 | void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m) | |
324 | { | |
325 | u8 byte = 0; | |
326 | unsigned int byte_bits = 0; | |
327 | unsigned int offset = 0; | |
328 | int i; | |
329 | ||
330 | for (i = 0; i < sb->map_nr; i++) { | |
331 | unsigned long word = READ_ONCE(sb->map[i].word); | |
332 | unsigned int word_bits = READ_ONCE(sb->map[i].depth); | |
333 | ||
334 | while (word_bits > 0) { | |
335 | unsigned int bits = min(8 - byte_bits, word_bits); | |
336 | ||
337 | byte |= (word & (BIT(bits) - 1)) << byte_bits; | |
338 | byte_bits += bits; | |
339 | if (byte_bits == 8) { | |
340 | emit_byte(m, offset, byte); | |
341 | byte = 0; | |
342 | byte_bits = 0; | |
343 | offset++; | |
344 | } | |
345 | word >>= bits; | |
346 | word_bits -= bits; | |
347 | } | |
348 | } | |
349 | if (byte_bits) { | |
350 | emit_byte(m, offset, byte); | |
351 | offset++; | |
352 | } | |
353 | if (offset) | |
354 | seq_putc(m, '\n'); | |
355 | } | |
356 | EXPORT_SYMBOL_GPL(sbitmap_bitmap_show); | |
357 | ||
a3275539 OS |
358 | static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq, |
359 | unsigned int depth) | |
88459642 OS |
360 | { |
361 | unsigned int wake_batch; | |
a3275539 | 362 | unsigned int shallow_depth; |
88459642 OS |
363 | |
364 | /* | |
365 | * For each batch, we wake up one queue. We need to make sure that our | |
a3275539 OS |
366 | * batch size is small enough that the full depth of the bitmap, |
367 | * potentially limited by a shallow depth, is enough to wake up all of | |
368 | * the queues. | |
369 | * | |
370 | * Each full word of the bitmap has bits_per_word bits, and there might | |
371 | * be a partial word. There are depth / bits_per_word full words and | |
372 | * depth % bits_per_word bits left over. In bitwise arithmetic: | |
373 | * | |
374 | * bits_per_word = 1 << shift | |
375 | * depth / bits_per_word = depth >> shift | |
376 | * depth % bits_per_word = depth & ((1 << shift) - 1) | |
377 | * | |
378 | * Each word can be limited to sbq->min_shallow_depth bits. | |
88459642 | 379 | */ |
a3275539 OS |
380 | shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth); |
381 | depth = ((depth >> sbq->sb.shift) * shallow_depth + | |
382 | min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth)); | |
383 | wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1, | |
384 | SBQ_WAKE_BATCH); | |
88459642 OS |
385 | |
386 | return wake_batch; | |
387 | } | |
388 | ||
389 | int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth, | |
f4a644db | 390 | int shift, bool round_robin, gfp_t flags, int node) |
88459642 OS |
391 | { |
392 | int ret; | |
393 | int i; | |
394 | ||
395 | ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node); | |
396 | if (ret) | |
397 | return ret; | |
398 | ||
40aabb67 OS |
399 | sbq->alloc_hint = alloc_percpu_gfp(unsigned int, flags); |
400 | if (!sbq->alloc_hint) { | |
401 | sbitmap_free(&sbq->sb); | |
402 | return -ENOMEM; | |
403 | } | |
404 | ||
98d95416 OS |
405 | if (depth && !round_robin) { |
406 | for_each_possible_cpu(i) | |
407 | *per_cpu_ptr(sbq->alloc_hint, i) = prandom_u32() % depth; | |
408 | } | |
409 | ||
a3275539 OS |
410 | sbq->min_shallow_depth = UINT_MAX; |
411 | sbq->wake_batch = sbq_calc_wake_batch(sbq, depth); | |
88459642 | 412 | atomic_set(&sbq->wake_index, 0); |
5d2ee712 | 413 | atomic_set(&sbq->ws_active, 0); |
88459642 | 414 | |
48e28166 | 415 | sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node); |
88459642 | 416 | if (!sbq->ws) { |
40aabb67 | 417 | free_percpu(sbq->alloc_hint); |
88459642 OS |
418 | sbitmap_free(&sbq->sb); |
419 | return -ENOMEM; | |
420 | } | |
421 | ||
422 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
423 | init_waitqueue_head(&sbq->ws[i].wait); | |
424 | atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch); | |
425 | } | |
f4a644db OS |
426 | |
427 | sbq->round_robin = round_robin; | |
88459642 OS |
428 | return 0; |
429 | } | |
430 | EXPORT_SYMBOL_GPL(sbitmap_queue_init_node); | |
431 | ||
a3275539 OS |
432 | static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq, |
433 | unsigned int depth) | |
88459642 | 434 | { |
a3275539 | 435 | unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth); |
6c0ca7ae OS |
436 | int i; |
437 | ||
438 | if (sbq->wake_batch != wake_batch) { | |
439 | WRITE_ONCE(sbq->wake_batch, wake_batch); | |
440 | /* | |
e6fc4649 ML |
441 | * Pairs with the memory barrier in sbitmap_queue_wake_up() |
442 | * to ensure that the batch size is updated before the wait | |
443 | * counts. | |
6c0ca7ae OS |
444 | */ |
445 | smp_mb__before_atomic(); | |
446 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) | |
447 | atomic_set(&sbq->ws[i].wait_cnt, 1); | |
448 | } | |
a3275539 OS |
449 | } |
450 | ||
451 | void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth) | |
452 | { | |
453 | sbitmap_queue_update_wake_batch(sbq, depth); | |
88459642 OS |
454 | sbitmap_resize(&sbq->sb, depth); |
455 | } | |
456 | EXPORT_SYMBOL_GPL(sbitmap_queue_resize); | |
457 | ||
f4a644db | 458 | int __sbitmap_queue_get(struct sbitmap_queue *sbq) |
40aabb67 | 459 | { |
05fd095d | 460 | unsigned int hint, depth; |
40aabb67 OS |
461 | int nr; |
462 | ||
463 | hint = this_cpu_read(*sbq->alloc_hint); | |
05fd095d OS |
464 | depth = READ_ONCE(sbq->sb.depth); |
465 | if (unlikely(hint >= depth)) { | |
466 | hint = depth ? prandom_u32() % depth : 0; | |
467 | this_cpu_write(*sbq->alloc_hint, hint); | |
468 | } | |
f4a644db | 469 | nr = sbitmap_get(&sbq->sb, hint, sbq->round_robin); |
40aabb67 OS |
470 | |
471 | if (nr == -1) { | |
472 | /* If the map is full, a hint won't do us much good. */ | |
473 | this_cpu_write(*sbq->alloc_hint, 0); | |
f4a644db | 474 | } else if (nr == hint || unlikely(sbq->round_robin)) { |
40aabb67 OS |
475 | /* Only update the hint if we used it. */ |
476 | hint = nr + 1; | |
05fd095d | 477 | if (hint >= depth - 1) |
40aabb67 OS |
478 | hint = 0; |
479 | this_cpu_write(*sbq->alloc_hint, hint); | |
480 | } | |
481 | ||
482 | return nr; | |
483 | } | |
484 | EXPORT_SYMBOL_GPL(__sbitmap_queue_get); | |
485 | ||
c05e6673 OS |
486 | int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq, |
487 | unsigned int shallow_depth) | |
488 | { | |
489 | unsigned int hint, depth; | |
490 | int nr; | |
491 | ||
61445b56 OS |
492 | WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth); |
493 | ||
c05e6673 OS |
494 | hint = this_cpu_read(*sbq->alloc_hint); |
495 | depth = READ_ONCE(sbq->sb.depth); | |
496 | if (unlikely(hint >= depth)) { | |
497 | hint = depth ? prandom_u32() % depth : 0; | |
498 | this_cpu_write(*sbq->alloc_hint, hint); | |
499 | } | |
500 | nr = sbitmap_get_shallow(&sbq->sb, hint, shallow_depth); | |
501 | ||
502 | if (nr == -1) { | |
503 | /* If the map is full, a hint won't do us much good. */ | |
504 | this_cpu_write(*sbq->alloc_hint, 0); | |
505 | } else if (nr == hint || unlikely(sbq->round_robin)) { | |
506 | /* Only update the hint if we used it. */ | |
507 | hint = nr + 1; | |
508 | if (hint >= depth - 1) | |
509 | hint = 0; | |
510 | this_cpu_write(*sbq->alloc_hint, hint); | |
511 | } | |
512 | ||
513 | return nr; | |
514 | } | |
515 | EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow); | |
516 | ||
a3275539 OS |
517 | void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq, |
518 | unsigned int min_shallow_depth) | |
519 | { | |
520 | sbq->min_shallow_depth = min_shallow_depth; | |
521 | sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth); | |
522 | } | |
523 | EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth); | |
524 | ||
88459642 OS |
525 | static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq) |
526 | { | |
527 | int i, wake_index; | |
528 | ||
5d2ee712 JA |
529 | if (!atomic_read(&sbq->ws_active)) |
530 | return NULL; | |
531 | ||
88459642 OS |
532 | wake_index = atomic_read(&sbq->wake_index); |
533 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
534 | struct sbq_wait_state *ws = &sbq->ws[wake_index]; | |
535 | ||
536 | if (waitqueue_active(&ws->wait)) { | |
537 | int o = atomic_read(&sbq->wake_index); | |
538 | ||
539 | if (wake_index != o) | |
540 | atomic_cmpxchg(&sbq->wake_index, o, wake_index); | |
541 | return ws; | |
542 | } | |
543 | ||
544 | wake_index = sbq_index_inc(wake_index); | |
545 | } | |
546 | ||
547 | return NULL; | |
548 | } | |
549 | ||
c854ab57 | 550 | static bool __sbq_wake_up(struct sbitmap_queue *sbq) |
88459642 OS |
551 | { |
552 | struct sbq_wait_state *ws; | |
6c0ca7ae | 553 | unsigned int wake_batch; |
88459642 OS |
554 | int wait_cnt; |
555 | ||
88459642 OS |
556 | ws = sbq_wake_ptr(sbq); |
557 | if (!ws) | |
c854ab57 | 558 | return false; |
88459642 OS |
559 | |
560 | wait_cnt = atomic_dec_return(&ws->wait_cnt); | |
6c0ca7ae | 561 | if (wait_cnt <= 0) { |
c854ab57 JA |
562 | int ret; |
563 | ||
6c0ca7ae | 564 | wake_batch = READ_ONCE(sbq->wake_batch); |
c854ab57 | 565 | |
6c0ca7ae OS |
566 | /* |
567 | * Pairs with the memory barrier in sbitmap_queue_resize() to | |
568 | * ensure that we see the batch size update before the wait | |
569 | * count is reset. | |
570 | */ | |
571 | smp_mb__before_atomic(); | |
c854ab57 | 572 | |
6c0ca7ae | 573 | /* |
c854ab57 JA |
574 | * For concurrent callers of this, the one that failed the |
575 | * atomic_cmpxhcg() race should call this function again | |
576 | * to wakeup a new batch on a different 'ws'. | |
6c0ca7ae | 577 | */ |
c854ab57 JA |
578 | ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch); |
579 | if (ret == wait_cnt) { | |
580 | sbq_index_atomic_inc(&sbq->wake_index); | |
581 | wake_up_nr(&ws->wait, wake_batch); | |
582 | return false; | |
583 | } | |
584 | ||
585 | return true; | |
88459642 | 586 | } |
c854ab57 JA |
587 | |
588 | return false; | |
589 | } | |
590 | ||
e6fc4649 | 591 | void sbitmap_queue_wake_up(struct sbitmap_queue *sbq) |
c854ab57 JA |
592 | { |
593 | while (__sbq_wake_up(sbq)) | |
594 | ; | |
88459642 | 595 | } |
e6fc4649 | 596 | EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up); |
88459642 | 597 | |
40aabb67 | 598 | void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr, |
f4a644db | 599 | unsigned int cpu) |
88459642 | 600 | { |
ea86ea2c JA |
601 | sbitmap_deferred_clear_bit(&sbq->sb, nr); |
602 | ||
e6fc4649 ML |
603 | /* |
604 | * Pairs with the memory barrier in set_current_state() to ensure the | |
605 | * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker | |
606 | * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the | |
607 | * waiter. See the comment on waitqueue_active(). | |
608 | */ | |
609 | smp_mb__after_atomic(); | |
610 | sbitmap_queue_wake_up(sbq); | |
611 | ||
5c64a8df | 612 | if (likely(!sbq->round_robin && nr < sbq->sb.depth)) |
40aabb67 | 613 | *per_cpu_ptr(sbq->alloc_hint, cpu) = nr; |
88459642 OS |
614 | } |
615 | EXPORT_SYMBOL_GPL(sbitmap_queue_clear); | |
616 | ||
617 | void sbitmap_queue_wake_all(struct sbitmap_queue *sbq) | |
618 | { | |
619 | int i, wake_index; | |
620 | ||
621 | /* | |
f66227de | 622 | * Pairs with the memory barrier in set_current_state() like in |
e6fc4649 | 623 | * sbitmap_queue_wake_up(). |
88459642 OS |
624 | */ |
625 | smp_mb(); | |
626 | wake_index = atomic_read(&sbq->wake_index); | |
627 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
628 | struct sbq_wait_state *ws = &sbq->ws[wake_index]; | |
629 | ||
630 | if (waitqueue_active(&ws->wait)) | |
631 | wake_up(&ws->wait); | |
632 | ||
633 | wake_index = sbq_index_inc(wake_index); | |
634 | } | |
635 | } | |
636 | EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all); | |
24af1ccf OS |
637 | |
638 | void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m) | |
639 | { | |
640 | bool first; | |
641 | int i; | |
642 | ||
643 | sbitmap_show(&sbq->sb, m); | |
644 | ||
645 | seq_puts(m, "alloc_hint={"); | |
646 | first = true; | |
647 | for_each_possible_cpu(i) { | |
648 | if (!first) | |
649 | seq_puts(m, ", "); | |
650 | first = false; | |
651 | seq_printf(m, "%u", *per_cpu_ptr(sbq->alloc_hint, i)); | |
652 | } | |
653 | seq_puts(m, "}\n"); | |
654 | ||
655 | seq_printf(m, "wake_batch=%u\n", sbq->wake_batch); | |
656 | seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index)); | |
5d2ee712 | 657 | seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active)); |
24af1ccf OS |
658 | |
659 | seq_puts(m, "ws={\n"); | |
660 | for (i = 0; i < SBQ_WAIT_QUEUES; i++) { | |
661 | struct sbq_wait_state *ws = &sbq->ws[i]; | |
662 | ||
663 | seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n", | |
664 | atomic_read(&ws->wait_cnt), | |
665 | waitqueue_active(&ws->wait) ? "active" : "inactive"); | |
666 | } | |
667 | seq_puts(m, "}\n"); | |
668 | ||
669 | seq_printf(m, "round_robin=%d\n", sbq->round_robin); | |
a3275539 | 670 | seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth); |
24af1ccf OS |
671 | } |
672 | EXPORT_SYMBOL_GPL(sbitmap_queue_show); | |
5d2ee712 | 673 | |
9f6b7ef6 JA |
674 | void sbitmap_add_wait_queue(struct sbitmap_queue *sbq, |
675 | struct sbq_wait_state *ws, | |
676 | struct sbq_wait *sbq_wait) | |
677 | { | |
678 | if (!sbq_wait->sbq) { | |
679 | sbq_wait->sbq = sbq; | |
680 | atomic_inc(&sbq->ws_active); | |
681 | } | |
682 | add_wait_queue(&ws->wait, &sbq_wait->wait); | |
683 | } | |
684 | EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue); | |
685 | ||
686 | void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait) | |
687 | { | |
688 | list_del_init(&sbq_wait->wait.entry); | |
689 | if (sbq_wait->sbq) { | |
690 | atomic_dec(&sbq_wait->sbq->ws_active); | |
691 | sbq_wait->sbq = NULL; | |
692 | } | |
693 | } | |
694 | EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue); | |
695 | ||
5d2ee712 JA |
696 | void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq, |
697 | struct sbq_wait_state *ws, | |
698 | struct sbq_wait *sbq_wait, int state) | |
699 | { | |
9f6b7ef6 | 700 | if (!sbq_wait->sbq) { |
5d2ee712 | 701 | atomic_inc(&sbq->ws_active); |
9f6b7ef6 | 702 | sbq_wait->sbq = sbq; |
5d2ee712 JA |
703 | } |
704 | prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state); | |
705 | } | |
706 | EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait); | |
707 | ||
708 | void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws, | |
709 | struct sbq_wait *sbq_wait) | |
710 | { | |
711 | finish_wait(&ws->wait, &sbq_wait->wait); | |
9f6b7ef6 | 712 | if (sbq_wait->sbq) { |
5d2ee712 | 713 | atomic_dec(&sbq->ws_active); |
9f6b7ef6 | 714 | sbq_wait->sbq = NULL; |
5d2ee712 JA |
715 | } |
716 | } | |
717 | EXPORT_SYMBOL_GPL(sbitmap_finish_wait); |