1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2016 Facebook
4  * Copyright (C) 2013-2014 Jens Axboe
5  */
6 
7 #include <linux/sched.h>
8 #include <linux/random.h>
9 #include <linux/sbitmap.h>
10 #include <linux/seq_file.h>
11 
init_alloc_hint(struct sbitmap * sb,gfp_t flags)12 static int init_alloc_hint(struct sbitmap *sb, gfp_t flags)
13 {
14 	unsigned depth = sb->depth;
15 
16 	sb->alloc_hint = alloc_percpu_gfp(unsigned int, flags);
17 	if (!sb->alloc_hint)
18 		return -ENOMEM;
19 
20 	if (depth && !sb->round_robin) {
21 		int i;
22 
23 		for_each_possible_cpu(i)
24 			*per_cpu_ptr(sb->alloc_hint, i) = prandom_u32() % depth;
25 	}
26 	return 0;
27 }
28 
update_alloc_hint_before_get(struct sbitmap * sb,unsigned int depth)29 static inline unsigned update_alloc_hint_before_get(struct sbitmap *sb,
30 						    unsigned int depth)
31 {
32 	unsigned hint;
33 
34 	hint = this_cpu_read(*sb->alloc_hint);
35 	if (unlikely(hint >= depth)) {
36 		hint = depth ? prandom_u32() % depth : 0;
37 		this_cpu_write(*sb->alloc_hint, hint);
38 	}
39 
40 	return hint;
41 }
42 
update_alloc_hint_after_get(struct sbitmap * sb,unsigned int depth,unsigned int hint,unsigned int nr)43 static inline void update_alloc_hint_after_get(struct sbitmap *sb,
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. */
50 		this_cpu_write(*sb->alloc_hint, 0);
51 	} else if (nr == hint || unlikely(sb->round_robin)) {
52 		/* Only update the hint if we used it. */
53 		hint = nr + 1;
54 		if (hint >= depth - 1)
55 			hint = 0;
56 		this_cpu_write(*sb->alloc_hint, hint);
57 	}
58 }
59 
60 /*
61  * See if we have deferred clears that we can batch move
62  */
sbitmap_deferred_clear(struct sbitmap_word * map)63 static inline bool sbitmap_deferred_clear(struct sbitmap_word *map)
64 {
65 	unsigned long mask;
66 
67 	if (!READ_ONCE(map->cleared))
68 		return false;
69 
70 	/*
71 	 * First get a stable cleared mask, setting the old mask to 0.
72 	 */
73 	mask = xchg(&map->cleared, 0);
74 
75 	/*
76 	 * Now clear the masked bits in our free word
77 	 */
78 	atomic_long_andnot(mask, (atomic_long_t *)&map->word);
79 	BUILD_BUG_ON(sizeof(atomic_long_t) != sizeof(map->word));
80 	return true;
81 }
82 
sbitmap_init_node(struct sbitmap * sb,unsigned int depth,int shift,gfp_t flags,int node,bool round_robin,bool alloc_hint)83 int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
84 		      gfp_t flags, int node, bool round_robin,
85 		      bool alloc_hint)
86 {
87 	unsigned int bits_per_word;
88 
89 	if (shift < 0)
90 		shift = sbitmap_calculate_shift(depth);
91 
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);
99 	sb->round_robin = round_robin;
100 
101 	if (depth == 0) {
102 		sb->map = NULL;
103 		return 0;
104 	}
105 
106 	if (alloc_hint) {
107 		if (init_alloc_hint(sb, flags))
108 			return -ENOMEM;
109 	} else {
110 		sb->alloc_hint = NULL;
111 	}
112 
113 	sb->map = kvzalloc_node(sb->map_nr * sizeof(*sb->map), flags, node);
114 	if (!sb->map) {
115 		free_percpu(sb->alloc_hint);
116 		return -ENOMEM;
117 	}
118 
119 	return 0;
120 }
121 EXPORT_SYMBOL_GPL(sbitmap_init_node);
122 
sbitmap_resize(struct sbitmap * sb,unsigned int depth)123 void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
124 {
125 	unsigned int bits_per_word = 1U << sb->shift;
126 	unsigned int i;
127 
128 	for (i = 0; i < sb->map_nr; i++)
129 		sbitmap_deferred_clear(&sb->map[i]);
130 
131 	sb->depth = depth;
132 	sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
133 }
134 EXPORT_SYMBOL_GPL(sbitmap_resize);
135 
__sbitmap_get_word(unsigned long * word,unsigned long depth,unsigned int hint,bool wrap)136 static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
137 			      unsigned int hint, bool wrap)
138 {
139 	int nr;
140 
141 	/* don't wrap if starting from 0 */
142 	wrap = wrap && hint;
143 
144 	while (1) {
145 		nr = find_next_zero_bit(word, depth, hint);
146 		if (unlikely(nr >= depth)) {
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 			 */
152 			if (hint && wrap) {
153 				hint = 0;
154 				continue;
155 			}
156 			return -1;
157 		}
158 
159 		if (!test_and_set_bit_lock(nr, word))
160 			break;
161 
162 		hint = nr + 1;
163 		if (hint >= depth - 1)
164 			hint = 0;
165 	}
166 
167 	return nr;
168 }
169 
sbitmap_find_bit_in_index(struct sbitmap * sb,int index,unsigned int alloc_hint)170 static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
171 				     unsigned int alloc_hint)
172 {
173 	struct sbitmap_word *map = &sb->map[index];
174 	int nr;
175 
176 	do {
177 		nr = __sbitmap_get_word(&map->word, __map_depth(sb, index),
178 					alloc_hint, !sb->round_robin);
179 		if (nr != -1)
180 			break;
181 		if (!sbitmap_deferred_clear(map))
182 			break;
183 	} while (1);
184 
185 	return nr;
186 }
187 
__sbitmap_get(struct sbitmap * sb,unsigned int alloc_hint)188 static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
189 {
190 	unsigned int i, index;
191 	int nr = -1;
192 
193 	index = SB_NR_TO_INDEX(sb, alloc_hint);
194 
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 	 */
200 	if (sb->round_robin)
201 		alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
202 	else
203 		alloc_hint = 0;
204 
205 	for (i = 0; i < sb->map_nr; i++) {
206 		nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
207 		if (nr != -1) {
208 			nr += index << sb->shift;
209 			break;
210 		}
211 
212 		/* Jump to next index. */
213 		alloc_hint = 0;
214 		if (++index >= sb->map_nr)
215 			index = 0;
216 	}
217 
218 	return nr;
219 }
220 
sbitmap_get(struct sbitmap * sb)221 int 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 }
236 EXPORT_SYMBOL_GPL(sbitmap_get);
237 
__sbitmap_get_shallow(struct sbitmap * sb,unsigned int alloc_hint,unsigned long shallow_depth)238 static int __sbitmap_get_shallow(struct sbitmap *sb,
239 				 unsigned int alloc_hint,
240 				 unsigned long shallow_depth)
241 {
242 	unsigned int i, index;
243 	int nr = -1;
244 
245 	index = SB_NR_TO_INDEX(sb, alloc_hint);
246 
247 	for (i = 0; i < sb->map_nr; i++) {
248 again:
249 		nr = __sbitmap_get_word(&sb->map[index].word,
250 					min_t(unsigned int,
251 					      __map_depth(sb, index),
252 					      shallow_depth),
253 					SB_NR_TO_BIT(sb, alloc_hint), true);
254 		if (nr != -1) {
255 			nr += index << sb->shift;
256 			break;
257 		}
258 
259 		if (sbitmap_deferred_clear(&sb->map[index]))
260 			goto again;
261 
262 		/* Jump to next index. */
263 		index++;
264 		alloc_hint = index << sb->shift;
265 
266 		if (index >= sb->map_nr) {
267 			index = 0;
268 			alloc_hint = 0;
269 		}
270 	}
271 
272 	return nr;
273 }
274 
sbitmap_get_shallow(struct sbitmap * sb,unsigned long shallow_depth)275 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
276 {
277 	int nr;
278 	unsigned int hint, depth;
279 
280 	if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
281 		return -1;
282 
283 	depth = READ_ONCE(sb->depth);
284 	hint = update_alloc_hint_before_get(sb, depth);
285 	nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
286 	update_alloc_hint_after_get(sb, depth, hint, nr);
287 
288 	return nr;
289 }
290 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
291 
sbitmap_any_bit_set(const struct sbitmap * sb)292 bool sbitmap_any_bit_set(const struct sbitmap *sb)
293 {
294 	unsigned int i;
295 
296 	for (i = 0; i < sb->map_nr; i++) {
297 		if (sb->map[i].word & ~sb->map[i].cleared)
298 			return true;
299 	}
300 	return false;
301 }
302 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
303 
__sbitmap_weight(const struct sbitmap * sb,bool set)304 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
305 {
306 	unsigned int i, weight = 0;
307 
308 	for (i = 0; i < sb->map_nr; i++) {
309 		const struct sbitmap_word *word = &sb->map[i];
310 		unsigned int word_depth = __map_depth(sb, i);
311 
312 		if (set)
313 			weight += bitmap_weight(&word->word, word_depth);
314 		else
315 			weight += bitmap_weight(&word->cleared, word_depth);
316 	}
317 	return weight;
318 }
319 
sbitmap_cleared(const struct sbitmap * sb)320 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
321 {
322 	return __sbitmap_weight(sb, false);
323 }
324 
sbitmap_weight(const struct sbitmap * sb)325 unsigned int sbitmap_weight(const struct sbitmap *sb)
326 {
327 	return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
328 }
329 EXPORT_SYMBOL_GPL(sbitmap_weight);
330 
sbitmap_show(struct sbitmap * sb,struct seq_file * m)331 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
332 {
333 	seq_printf(m, "depth=%u\n", sb->depth);
334 	seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
335 	seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
336 	seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
337 	seq_printf(m, "map_nr=%u\n", sb->map_nr);
338 }
339 EXPORT_SYMBOL_GPL(sbitmap_show);
340 
emit_byte(struct seq_file * m,unsigned int offset,u8 byte)341 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
342 {
343 	if ((offset & 0xf) == 0) {
344 		if (offset != 0)
345 			seq_putc(m, '\n');
346 		seq_printf(m, "%08x:", offset);
347 	}
348 	if ((offset & 0x1) == 0)
349 		seq_putc(m, ' ');
350 	seq_printf(m, "%02x", byte);
351 }
352 
sbitmap_bitmap_show(struct sbitmap * sb,struct seq_file * m)353 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
354 {
355 	u8 byte = 0;
356 	unsigned int byte_bits = 0;
357 	unsigned int offset = 0;
358 	int i;
359 
360 	for (i = 0; i < sb->map_nr; i++) {
361 		unsigned long word = READ_ONCE(sb->map[i].word);
362 		unsigned long cleared = READ_ONCE(sb->map[i].cleared);
363 		unsigned int word_bits = __map_depth(sb, i);
364 
365 		word &= ~cleared;
366 
367 		while (word_bits > 0) {
368 			unsigned int bits = min(8 - byte_bits, word_bits);
369 
370 			byte |= (word & (BIT(bits) - 1)) << byte_bits;
371 			byte_bits += bits;
372 			if (byte_bits == 8) {
373 				emit_byte(m, offset, byte);
374 				byte = 0;
375 				byte_bits = 0;
376 				offset++;
377 			}
378 			word >>= bits;
379 			word_bits -= bits;
380 		}
381 	}
382 	if (byte_bits) {
383 		emit_byte(m, offset, byte);
384 		offset++;
385 	}
386 	if (offset)
387 		seq_putc(m, '\n');
388 }
389 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
390 
sbq_calc_wake_batch(struct sbitmap_queue * sbq,unsigned int depth)391 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
392 					unsigned int depth)
393 {
394 	unsigned int wake_batch;
395 	unsigned int shallow_depth;
396 
397 	/*
398 	 * For each batch, we wake up one queue. We need to make sure that our
399 	 * batch size is small enough that the full depth of the bitmap,
400 	 * potentially limited by a shallow depth, is enough to wake up all of
401 	 * the queues.
402 	 *
403 	 * Each full word of the bitmap has bits_per_word bits, and there might
404 	 * be a partial word. There are depth / bits_per_word full words and
405 	 * depth % bits_per_word bits left over. In bitwise arithmetic:
406 	 *
407 	 * bits_per_word = 1 << shift
408 	 * depth / bits_per_word = depth >> shift
409 	 * depth % bits_per_word = depth & ((1 << shift) - 1)
410 	 *
411 	 * Each word can be limited to sbq->min_shallow_depth bits.
412 	 */
413 	shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
414 	depth = ((depth >> sbq->sb.shift) * shallow_depth +
415 		 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
416 	wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
417 			     SBQ_WAKE_BATCH);
418 
419 	return wake_batch;
420 }
421 
sbitmap_queue_init_node(struct sbitmap_queue * sbq,unsigned int depth,int shift,bool round_robin,gfp_t flags,int node)422 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
423 			    int shift, bool round_robin, gfp_t flags, int node)
424 {
425 	int ret;
426 	int i;
427 
428 	ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
429 				round_robin, true);
430 	if (ret)
431 		return ret;
432 
433 	sbq->min_shallow_depth = UINT_MAX;
434 	sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
435 	atomic_set(&sbq->wake_index, 0);
436 	atomic_set(&sbq->ws_active, 0);
437 
438 	sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
439 	if (!sbq->ws) {
440 		sbitmap_free(&sbq->sb);
441 		return -ENOMEM;
442 	}
443 
444 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
445 		init_waitqueue_head(&sbq->ws[i].wait);
446 		atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
447 	}
448 
449 	return 0;
450 }
451 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
452 
__sbitmap_queue_update_wake_batch(struct sbitmap_queue * sbq,unsigned int wake_batch)453 static inline void __sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
454 					    unsigned int wake_batch)
455 {
456 	int i;
457 
458 	if (sbq->wake_batch != wake_batch) {
459 		WRITE_ONCE(sbq->wake_batch, wake_batch);
460 		/*
461 		 * Pairs with the memory barrier in sbitmap_queue_wake_up()
462 		 * to ensure that the batch size is updated before the wait
463 		 * counts.
464 		 */
465 		smp_mb();
466 		for (i = 0; i < SBQ_WAIT_QUEUES; i++)
467 			atomic_set(&sbq->ws[i].wait_cnt, 1);
468 	}
469 }
470 
sbitmap_queue_update_wake_batch(struct sbitmap_queue * sbq,unsigned int depth)471 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
472 					    unsigned int depth)
473 {
474 	unsigned int wake_batch;
475 
476 	wake_batch = sbq_calc_wake_batch(sbq, depth);
477 	__sbitmap_queue_update_wake_batch(sbq, wake_batch);
478 }
479 
sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue * sbq,unsigned int users)480 void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
481 					    unsigned int users)
482 {
483 	unsigned int wake_batch;
484 	unsigned int min_batch;
485 	unsigned int depth = (sbq->sb.depth + users - 1) / users;
486 
487 	min_batch = sbq->sb.depth >= (4 * SBQ_WAIT_QUEUES) ? 4 : 1;
488 
489 	wake_batch = clamp_val(depth / SBQ_WAIT_QUEUES,
490 			min_batch, SBQ_WAKE_BATCH);
491 	__sbitmap_queue_update_wake_batch(sbq, wake_batch);
492 }
493 EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
494 
sbitmap_queue_resize(struct sbitmap_queue * sbq,unsigned int depth)495 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
496 {
497 	sbitmap_queue_update_wake_batch(sbq, depth);
498 	sbitmap_resize(&sbq->sb, depth);
499 }
500 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
501 
__sbitmap_queue_get(struct sbitmap_queue * sbq)502 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
503 {
504 	return sbitmap_get(&sbq->sb);
505 }
506 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
507 
__sbitmap_queue_get_batch(struct sbitmap_queue * sbq,int nr_tags,unsigned int * offset)508 unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
509 					unsigned int *offset)
510 {
511 	struct sbitmap *sb = &sbq->sb;
512 	unsigned int hint, depth;
513 	unsigned long index, nr;
514 	int i;
515 
516 	if (unlikely(sb->round_robin))
517 		return 0;
518 
519 	depth = READ_ONCE(sb->depth);
520 	hint = update_alloc_hint_before_get(sb, depth);
521 
522 	index = SB_NR_TO_INDEX(sb, hint);
523 
524 	for (i = 0; i < sb->map_nr; i++) {
525 		struct sbitmap_word *map = &sb->map[index];
526 		unsigned long get_mask;
527 		unsigned int map_depth = __map_depth(sb, index);
528 
529 		sbitmap_deferred_clear(map);
530 		if (map->word == (1UL << (map_depth - 1)) - 1)
531 			goto next;
532 
533 		nr = find_first_zero_bit(&map->word, map_depth);
534 		if (nr + nr_tags <= map_depth) {
535 			atomic_long_t *ptr = (atomic_long_t *) &map->word;
536 			int map_tags = min_t(int, nr_tags, map_depth);
537 			unsigned long val, ret;
538 
539 			get_mask = ((1UL << map_tags) - 1) << nr;
540 			do {
541 				val = READ_ONCE(map->word);
542 				if ((val & ~get_mask) != val)
543 					goto next;
544 				ret = atomic_long_cmpxchg(ptr, val, get_mask | val);
545 			} while (ret != val);
546 			get_mask = (get_mask & ~ret) >> nr;
547 			if (get_mask) {
548 				*offset = nr + (index << sb->shift);
549 				update_alloc_hint_after_get(sb, depth, hint,
550 							*offset + map_tags - 1);
551 				return get_mask;
552 			}
553 		}
554 next:
555 		/* Jump to next index. */
556 		if (++index >= sb->map_nr)
557 			index = 0;
558 	}
559 
560 	return 0;
561 }
562 
sbitmap_queue_get_shallow(struct sbitmap_queue * sbq,unsigned int shallow_depth)563 int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
564 			      unsigned int shallow_depth)
565 {
566 	WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
567 
568 	return sbitmap_get_shallow(&sbq->sb, shallow_depth);
569 }
570 EXPORT_SYMBOL_GPL(sbitmap_queue_get_shallow);
571 
sbitmap_queue_min_shallow_depth(struct sbitmap_queue * sbq,unsigned int min_shallow_depth)572 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
573 				     unsigned int min_shallow_depth)
574 {
575 	sbq->min_shallow_depth = min_shallow_depth;
576 	sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
577 }
578 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
579 
sbq_wake_ptr(struct sbitmap_queue * sbq)580 static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
581 {
582 	int i, wake_index;
583 
584 	if (!atomic_read(&sbq->ws_active))
585 		return NULL;
586 
587 	wake_index = atomic_read(&sbq->wake_index);
588 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
589 		struct sbq_wait_state *ws = &sbq->ws[wake_index];
590 
591 		if (waitqueue_active(&ws->wait)) {
592 			if (wake_index != atomic_read(&sbq->wake_index))
593 				atomic_set(&sbq->wake_index, wake_index);
594 			return ws;
595 		}
596 
597 		wake_index = sbq_index_inc(wake_index);
598 	}
599 
600 	return NULL;
601 }
602 
__sbq_wake_up(struct sbitmap_queue * sbq)603 static bool __sbq_wake_up(struct sbitmap_queue *sbq)
604 {
605 	struct sbq_wait_state *ws;
606 	unsigned int wake_batch;
607 	int wait_cnt;
608 
609 	ws = sbq_wake_ptr(sbq);
610 	if (!ws)
611 		return false;
612 
613 	wait_cnt = atomic_dec_return(&ws->wait_cnt);
614 	if (wait_cnt <= 0) {
615 		int ret;
616 
617 		wake_batch = READ_ONCE(sbq->wake_batch);
618 
619 		/*
620 		 * Pairs with the memory barrier in sbitmap_queue_resize() to
621 		 * ensure that we see the batch size update before the wait
622 		 * count is reset.
623 		 */
624 		smp_mb__before_atomic();
625 
626 		/*
627 		 * For concurrent callers of this, the one that failed the
628 		 * atomic_cmpxhcg() race should call this function again
629 		 * to wakeup a new batch on a different 'ws'.
630 		 */
631 		ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch);
632 		if (ret == wait_cnt) {
633 			sbq_index_atomic_inc(&sbq->wake_index);
634 			wake_up_nr(&ws->wait, wake_batch);
635 			return false;
636 		}
637 
638 		return true;
639 	}
640 
641 	return false;
642 }
643 
sbitmap_queue_wake_up(struct sbitmap_queue * sbq)644 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
645 {
646 	while (__sbq_wake_up(sbq))
647 		;
648 }
649 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
650 
sbitmap_update_cpu_hint(struct sbitmap * sb,int cpu,int tag)651 static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
652 {
653 	if (likely(!sb->round_robin && tag < sb->depth))
654 		data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
655 }
656 
sbitmap_queue_clear_batch(struct sbitmap_queue * sbq,int offset,int * tags,int nr_tags)657 void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
658 				int *tags, int nr_tags)
659 {
660 	struct sbitmap *sb = &sbq->sb;
661 	unsigned long *addr = NULL;
662 	unsigned long mask = 0;
663 	int i;
664 
665 	smp_mb__before_atomic();
666 	for (i = 0; i < nr_tags; i++) {
667 		const int tag = tags[i] - offset;
668 		unsigned long *this_addr;
669 
670 		/* since we're clearing a batch, skip the deferred map */
671 		this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
672 		if (!addr) {
673 			addr = this_addr;
674 		} else if (addr != this_addr) {
675 			atomic_long_andnot(mask, (atomic_long_t *) addr);
676 			mask = 0;
677 			addr = this_addr;
678 		}
679 		mask |= (1UL << SB_NR_TO_BIT(sb, tag));
680 	}
681 
682 	if (mask)
683 		atomic_long_andnot(mask, (atomic_long_t *) addr);
684 
685 	smp_mb__after_atomic();
686 	sbitmap_queue_wake_up(sbq);
687 	sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
688 					tags[nr_tags - 1] - offset);
689 }
690 
sbitmap_queue_clear(struct sbitmap_queue * sbq,unsigned int nr,unsigned int cpu)691 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
692 			 unsigned int cpu)
693 {
694 	/*
695 	 * Once the clear bit is set, the bit may be allocated out.
696 	 *
697 	 * Orders READ/WRITE on the associated instance(such as request
698 	 * of blk_mq) by this bit for avoiding race with re-allocation,
699 	 * and its pair is the memory barrier implied in __sbitmap_get_word.
700 	 *
701 	 * One invariant is that the clear bit has to be zero when the bit
702 	 * is in use.
703 	 */
704 	smp_mb__before_atomic();
705 	sbitmap_deferred_clear_bit(&sbq->sb, nr);
706 
707 	/*
708 	 * Pairs with the memory barrier in set_current_state() to ensure the
709 	 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
710 	 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
711 	 * waiter. See the comment on waitqueue_active().
712 	 */
713 	smp_mb__after_atomic();
714 	sbitmap_queue_wake_up(sbq);
715 	sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
716 }
717 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
718 
sbitmap_queue_wake_all(struct sbitmap_queue * sbq)719 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
720 {
721 	int i, wake_index;
722 
723 	/*
724 	 * Pairs with the memory barrier in set_current_state() like in
725 	 * sbitmap_queue_wake_up().
726 	 */
727 	smp_mb();
728 	wake_index = atomic_read(&sbq->wake_index);
729 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
730 		struct sbq_wait_state *ws = &sbq->ws[wake_index];
731 
732 		if (waitqueue_active(&ws->wait))
733 			wake_up(&ws->wait);
734 
735 		wake_index = sbq_index_inc(wake_index);
736 	}
737 }
738 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
739 
sbitmap_queue_show(struct sbitmap_queue * sbq,struct seq_file * m)740 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
741 {
742 	bool first;
743 	int i;
744 
745 	sbitmap_show(&sbq->sb, m);
746 
747 	seq_puts(m, "alloc_hint={");
748 	first = true;
749 	for_each_possible_cpu(i) {
750 		if (!first)
751 			seq_puts(m, ", ");
752 		first = false;
753 		seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
754 	}
755 	seq_puts(m, "}\n");
756 
757 	seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
758 	seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
759 	seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
760 
761 	seq_puts(m, "ws={\n");
762 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
763 		struct sbq_wait_state *ws = &sbq->ws[i];
764 
765 		seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
766 			   atomic_read(&ws->wait_cnt),
767 			   waitqueue_active(&ws->wait) ? "active" : "inactive");
768 	}
769 	seq_puts(m, "}\n");
770 
771 	seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
772 	seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
773 }
774 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
775 
sbitmap_add_wait_queue(struct sbitmap_queue * sbq,struct sbq_wait_state * ws,struct sbq_wait * sbq_wait)776 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
777 			    struct sbq_wait_state *ws,
778 			    struct sbq_wait *sbq_wait)
779 {
780 	if (!sbq_wait->sbq) {
781 		sbq_wait->sbq = sbq;
782 		atomic_inc(&sbq->ws_active);
783 		add_wait_queue(&ws->wait, &sbq_wait->wait);
784 	}
785 }
786 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
787 
sbitmap_del_wait_queue(struct sbq_wait * sbq_wait)788 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
789 {
790 	list_del_init(&sbq_wait->wait.entry);
791 	if (sbq_wait->sbq) {
792 		atomic_dec(&sbq_wait->sbq->ws_active);
793 		sbq_wait->sbq = NULL;
794 	}
795 }
796 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
797 
sbitmap_prepare_to_wait(struct sbitmap_queue * sbq,struct sbq_wait_state * ws,struct sbq_wait * sbq_wait,int state)798 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
799 			     struct sbq_wait_state *ws,
800 			     struct sbq_wait *sbq_wait, int state)
801 {
802 	if (!sbq_wait->sbq) {
803 		atomic_inc(&sbq->ws_active);
804 		sbq_wait->sbq = sbq;
805 	}
806 	prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
807 }
808 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
809 
sbitmap_finish_wait(struct sbitmap_queue * sbq,struct sbq_wait_state * ws,struct sbq_wait * sbq_wait)810 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
811 			 struct sbq_wait *sbq_wait)
812 {
813 	finish_wait(&ws->wait, &sbq_wait->wait);
814 	if (sbq_wait->sbq) {
815 		atomic_dec(&sbq->ws_active);
816 		sbq_wait->sbq = NULL;
817 	}
818 }
819 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);
820