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 	unsigned int i;
89 
90 	if (shift < 0)
91 		shift = sbitmap_calculate_shift(depth);
92 
93 	bits_per_word = 1U << shift;
94 	if (bits_per_word > BITS_PER_LONG)
95 		return -EINVAL;
96 
97 	sb->shift = shift;
98 	sb->depth = depth;
99 	sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
100 	sb->round_robin = round_robin;
101 
102 	if (depth == 0) {
103 		sb->map = NULL;
104 		return 0;
105 	}
106 
107 	if (alloc_hint) {
108 		if (init_alloc_hint(sb, flags))
109 			return -ENOMEM;
110 	} else {
111 		sb->alloc_hint = NULL;
112 	}
113 
114 	sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node);
115 	if (!sb->map) {
116 		free_percpu(sb->alloc_hint);
117 		return -ENOMEM;
118 	}
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 	return 0;
125 }
126 EXPORT_SYMBOL_GPL(sbitmap_init_node);
127 
sbitmap_resize(struct sbitmap * sb,unsigned int depth)128 void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
129 {
130 	unsigned int bits_per_word = 1U << sb->shift;
131 	unsigned int i;
132 
133 	for (i = 0; i < sb->map_nr; i++)
134 		sbitmap_deferred_clear(&sb->map[i]);
135 
136 	sb->depth = depth;
137 	sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
138 
139 	for (i = 0; i < sb->map_nr; i++) {
140 		sb->map[i].depth = min(depth, bits_per_word);
141 		depth -= sb->map[i].depth;
142 	}
143 }
144 EXPORT_SYMBOL_GPL(sbitmap_resize);
145 
__sbitmap_get_word(unsigned long * word,unsigned long depth,unsigned int hint,bool wrap)146 static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
147 			      unsigned int hint, bool wrap)
148 {
149 	int nr;
150 
151 	/* don't wrap if starting from 0 */
152 	wrap = wrap && hint;
153 
154 	while (1) {
155 		nr = find_next_zero_bit(word, depth, hint);
156 		if (unlikely(nr >= depth)) {
157 			/*
158 			 * We started with an offset, and we didn't reset the
159 			 * offset to 0 in a failure case, so start from 0 to
160 			 * exhaust the map.
161 			 */
162 			if (hint && wrap) {
163 				hint = 0;
164 				continue;
165 			}
166 			return -1;
167 		}
168 
169 		if (!test_and_set_bit_lock(nr, word))
170 			break;
171 
172 		hint = nr + 1;
173 		if (hint >= depth - 1)
174 			hint = 0;
175 	}
176 
177 	return nr;
178 }
179 
sbitmap_find_bit_in_index(struct sbitmap * sb,int index,unsigned int alloc_hint)180 static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
181 				     unsigned int alloc_hint)
182 {
183 	struct sbitmap_word *map = &sb->map[index];
184 	int nr;
185 
186 	do {
187 		nr = __sbitmap_get_word(&map->word, map->depth, alloc_hint,
188 					!sb->round_robin);
189 		if (nr != -1)
190 			break;
191 		if (!sbitmap_deferred_clear(map))
192 			break;
193 	} while (1);
194 
195 	return nr;
196 }
197 
__sbitmap_get(struct sbitmap * sb,unsigned int alloc_hint)198 static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
199 {
200 	unsigned int i, index;
201 	int nr = -1;
202 
203 	index = SB_NR_TO_INDEX(sb, alloc_hint);
204 
205 	/*
206 	 * Unless we're doing round robin tag allocation, just use the
207 	 * alloc_hint to find the right word index. No point in looping
208 	 * twice in find_next_zero_bit() for that case.
209 	 */
210 	if (sb->round_robin)
211 		alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
212 	else
213 		alloc_hint = 0;
214 
215 	for (i = 0; i < sb->map_nr; i++) {
216 		nr = sbitmap_find_bit_in_index(sb, index, alloc_hint);
217 		if (nr != -1) {
218 			nr += index << sb->shift;
219 			break;
220 		}
221 
222 		/* Jump to next index. */
223 		alloc_hint = 0;
224 		if (++index >= sb->map_nr)
225 			index = 0;
226 	}
227 
228 	return nr;
229 }
230 
sbitmap_get(struct sbitmap * sb)231 int sbitmap_get(struct sbitmap *sb)
232 {
233 	int nr;
234 	unsigned int hint, depth;
235 
236 	if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
237 		return -1;
238 
239 	depth = READ_ONCE(sb->depth);
240 	hint = update_alloc_hint_before_get(sb, depth);
241 	nr = __sbitmap_get(sb, hint);
242 	update_alloc_hint_after_get(sb, depth, hint, nr);
243 
244 	return nr;
245 }
246 EXPORT_SYMBOL_GPL(sbitmap_get);
247 
__sbitmap_get_shallow(struct sbitmap * sb,unsigned int alloc_hint,unsigned long shallow_depth)248 static int __sbitmap_get_shallow(struct sbitmap *sb,
249 				 unsigned int alloc_hint,
250 				 unsigned long shallow_depth)
251 {
252 	unsigned int i, index;
253 	int nr = -1;
254 
255 	index = SB_NR_TO_INDEX(sb, alloc_hint);
256 
257 	for (i = 0; i < sb->map_nr; i++) {
258 again:
259 		nr = __sbitmap_get_word(&sb->map[index].word,
260 					min(sb->map[index].depth, shallow_depth),
261 					SB_NR_TO_BIT(sb, alloc_hint), true);
262 		if (nr != -1) {
263 			nr += index << sb->shift;
264 			break;
265 		}
266 
267 		if (sbitmap_deferred_clear(&sb->map[index]))
268 			goto again;
269 
270 		/* Jump to next index. */
271 		index++;
272 		alloc_hint = index << sb->shift;
273 
274 		if (index >= sb->map_nr) {
275 			index = 0;
276 			alloc_hint = 0;
277 		}
278 	}
279 
280 	return nr;
281 }
282 
sbitmap_get_shallow(struct sbitmap * sb,unsigned long shallow_depth)283 int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
284 {
285 	int nr;
286 	unsigned int hint, depth;
287 
288 	if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
289 		return -1;
290 
291 	depth = READ_ONCE(sb->depth);
292 	hint = update_alloc_hint_before_get(sb, depth);
293 	nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
294 	update_alloc_hint_after_get(sb, depth, hint, nr);
295 
296 	return nr;
297 }
298 EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
299 
sbitmap_any_bit_set(const struct sbitmap * sb)300 bool sbitmap_any_bit_set(const struct sbitmap *sb)
301 {
302 	unsigned int i;
303 
304 	for (i = 0; i < sb->map_nr; i++) {
305 		if (sb->map[i].word & ~sb->map[i].cleared)
306 			return true;
307 	}
308 	return false;
309 }
310 EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
311 
__sbitmap_weight(const struct sbitmap * sb,bool set)312 static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
313 {
314 	unsigned int i, weight = 0;
315 
316 	for (i = 0; i < sb->map_nr; i++) {
317 		const struct sbitmap_word *word = &sb->map[i];
318 
319 		if (set)
320 			weight += bitmap_weight(&word->word, word->depth);
321 		else
322 			weight += bitmap_weight(&word->cleared, word->depth);
323 	}
324 	return weight;
325 }
326 
sbitmap_cleared(const struct sbitmap * sb)327 static unsigned int sbitmap_cleared(const struct sbitmap *sb)
328 {
329 	return __sbitmap_weight(sb, false);
330 }
331 
sbitmap_weight(const struct sbitmap * sb)332 unsigned int sbitmap_weight(const struct sbitmap *sb)
333 {
334 	return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
335 }
336 EXPORT_SYMBOL_GPL(sbitmap_weight);
337 
sbitmap_show(struct sbitmap * sb,struct seq_file * m)338 void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
339 {
340 	seq_printf(m, "depth=%u\n", sb->depth);
341 	seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
342 	seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
343 	seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
344 	seq_printf(m, "map_nr=%u\n", sb->map_nr);
345 }
346 EXPORT_SYMBOL_GPL(sbitmap_show);
347 
emit_byte(struct seq_file * m,unsigned int offset,u8 byte)348 static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
349 {
350 	if ((offset & 0xf) == 0) {
351 		if (offset != 0)
352 			seq_putc(m, '\n');
353 		seq_printf(m, "%08x:", offset);
354 	}
355 	if ((offset & 0x1) == 0)
356 		seq_putc(m, ' ');
357 	seq_printf(m, "%02x", byte);
358 }
359 
sbitmap_bitmap_show(struct sbitmap * sb,struct seq_file * m)360 void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
361 {
362 	u8 byte = 0;
363 	unsigned int byte_bits = 0;
364 	unsigned int offset = 0;
365 	int i;
366 
367 	for (i = 0; i < sb->map_nr; i++) {
368 		unsigned long word = READ_ONCE(sb->map[i].word);
369 		unsigned long cleared = READ_ONCE(sb->map[i].cleared);
370 		unsigned int word_bits = READ_ONCE(sb->map[i].depth);
371 
372 		word &= ~cleared;
373 
374 		while (word_bits > 0) {
375 			unsigned int bits = min(8 - byte_bits, word_bits);
376 
377 			byte |= (word & (BIT(bits) - 1)) << byte_bits;
378 			byte_bits += bits;
379 			if (byte_bits == 8) {
380 				emit_byte(m, offset, byte);
381 				byte = 0;
382 				byte_bits = 0;
383 				offset++;
384 			}
385 			word >>= bits;
386 			word_bits -= bits;
387 		}
388 	}
389 	if (byte_bits) {
390 		emit_byte(m, offset, byte);
391 		offset++;
392 	}
393 	if (offset)
394 		seq_putc(m, '\n');
395 }
396 EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
397 
sbq_calc_wake_batch(struct sbitmap_queue * sbq,unsigned int depth)398 static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
399 					unsigned int depth)
400 {
401 	unsigned int wake_batch;
402 	unsigned int shallow_depth;
403 
404 	/*
405 	 * For each batch, we wake up one queue. We need to make sure that our
406 	 * batch size is small enough that the full depth of the bitmap,
407 	 * potentially limited by a shallow depth, is enough to wake up all of
408 	 * the queues.
409 	 *
410 	 * Each full word of the bitmap has bits_per_word bits, and there might
411 	 * be a partial word. There are depth / bits_per_word full words and
412 	 * depth % bits_per_word bits left over. In bitwise arithmetic:
413 	 *
414 	 * bits_per_word = 1 << shift
415 	 * depth / bits_per_word = depth >> shift
416 	 * depth % bits_per_word = depth & ((1 << shift) - 1)
417 	 *
418 	 * Each word can be limited to sbq->min_shallow_depth bits.
419 	 */
420 	shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
421 	depth = ((depth >> sbq->sb.shift) * shallow_depth +
422 		 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
423 	wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
424 			     SBQ_WAKE_BATCH);
425 
426 	return wake_batch;
427 }
428 
sbitmap_queue_init_node(struct sbitmap_queue * sbq,unsigned int depth,int shift,bool round_robin,gfp_t flags,int node)429 int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
430 			    int shift, bool round_robin, gfp_t flags, int node)
431 {
432 	int ret;
433 	int i;
434 
435 	ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
436 				round_robin, true);
437 	if (ret)
438 		return ret;
439 
440 	sbq->min_shallow_depth = UINT_MAX;
441 	sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
442 	atomic_set(&sbq->wake_index, 0);
443 	atomic_set(&sbq->ws_active, 0);
444 
445 	sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
446 	if (!sbq->ws) {
447 		sbitmap_free(&sbq->sb);
448 		return -ENOMEM;
449 	}
450 
451 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
452 		init_waitqueue_head(&sbq->ws[i].wait);
453 		atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
454 	}
455 
456 	return 0;
457 }
458 EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
459 
sbitmap_queue_update_wake_batch(struct sbitmap_queue * sbq,unsigned int depth)460 static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
461 					    unsigned int depth)
462 {
463 	unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth);
464 	int i;
465 
466 	if (sbq->wake_batch != wake_batch) {
467 		WRITE_ONCE(sbq->wake_batch, wake_batch);
468 		/*
469 		 * Pairs with the memory barrier in sbitmap_queue_wake_up()
470 		 * to ensure that the batch size is updated before the wait
471 		 * counts.
472 		 */
473 		smp_mb();
474 		for (i = 0; i < SBQ_WAIT_QUEUES; i++)
475 			atomic_set(&sbq->ws[i].wait_cnt, 1);
476 	}
477 }
478 
sbitmap_queue_resize(struct sbitmap_queue * sbq,unsigned int depth)479 void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
480 {
481 	sbitmap_queue_update_wake_batch(sbq, depth);
482 	sbitmap_resize(&sbq->sb, depth);
483 }
484 EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
485 
__sbitmap_queue_get(struct sbitmap_queue * sbq)486 int __sbitmap_queue_get(struct sbitmap_queue *sbq)
487 {
488 	return sbitmap_get(&sbq->sb);
489 }
490 EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
491 
__sbitmap_queue_get_batch(struct sbitmap_queue * sbq,int nr_tags,unsigned int * offset)492 unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
493 					unsigned int *offset)
494 {
495 	struct sbitmap *sb = &sbq->sb;
496 	unsigned int hint, depth;
497 	unsigned long index, nr;
498 	int i;
499 
500 	if (unlikely(sb->round_robin))
501 		return 0;
502 
503 	depth = READ_ONCE(sb->depth);
504 	hint = update_alloc_hint_before_get(sb, depth);
505 
506 	index = SB_NR_TO_INDEX(sb, hint);
507 
508 	for (i = 0; i < sb->map_nr; i++) {
509 		struct sbitmap_word *map = &sb->map[index];
510 		unsigned long get_mask;
511 
512 		sbitmap_deferred_clear(map);
513 		if (map->word == (1UL << (map->depth - 1)) - 1)
514 			continue;
515 
516 		nr = find_first_zero_bit(&map->word, map->depth);
517 		if (nr + nr_tags <= map->depth) {
518 			atomic_long_t *ptr = (atomic_long_t *) &map->word;
519 			int map_tags = min_t(int, nr_tags, map->depth);
520 			unsigned long val, ret;
521 
522 			get_mask = ((1UL << map_tags) - 1) << nr;
523 			do {
524 				val = READ_ONCE(map->word);
525 				ret = atomic_long_cmpxchg(ptr, val, get_mask | val);
526 			} while (ret != val);
527 			get_mask = (get_mask & ~ret) >> nr;
528 			if (get_mask) {
529 				*offset = nr + (index << sb->shift);
530 				update_alloc_hint_after_get(sb, depth, hint,
531 							*offset + map_tags - 1);
532 				return get_mask;
533 			}
534 		}
535 		/* Jump to next index. */
536 		if (++index >= sb->map_nr)
537 			index = 0;
538 	}
539 
540 	return 0;
541 }
542 
__sbitmap_queue_get_shallow(struct sbitmap_queue * sbq,unsigned int shallow_depth)543 int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
544 				unsigned int shallow_depth)
545 {
546 	WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
547 
548 	return sbitmap_get_shallow(&sbq->sb, shallow_depth);
549 }
550 EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow);
551 
sbitmap_queue_min_shallow_depth(struct sbitmap_queue * sbq,unsigned int min_shallow_depth)552 void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
553 				     unsigned int min_shallow_depth)
554 {
555 	sbq->min_shallow_depth = min_shallow_depth;
556 	sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
557 }
558 EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
559 
sbq_wake_ptr(struct sbitmap_queue * sbq)560 static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
561 {
562 	int i, wake_index;
563 
564 	if (!atomic_read(&sbq->ws_active))
565 		return NULL;
566 
567 	wake_index = atomic_read(&sbq->wake_index);
568 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
569 		struct sbq_wait_state *ws = &sbq->ws[wake_index];
570 
571 		if (waitqueue_active(&ws->wait)) {
572 			if (wake_index != atomic_read(&sbq->wake_index))
573 				atomic_set(&sbq->wake_index, wake_index);
574 			return ws;
575 		}
576 
577 		wake_index = sbq_index_inc(wake_index);
578 	}
579 
580 	return NULL;
581 }
582 
__sbq_wake_up(struct sbitmap_queue * sbq)583 static bool __sbq_wake_up(struct sbitmap_queue *sbq)
584 {
585 	struct sbq_wait_state *ws;
586 	unsigned int wake_batch;
587 	int wait_cnt;
588 
589 	ws = sbq_wake_ptr(sbq);
590 	if (!ws)
591 		return false;
592 
593 	wait_cnt = atomic_dec_return(&ws->wait_cnt);
594 	if (wait_cnt <= 0) {
595 		int ret;
596 
597 		wake_batch = READ_ONCE(sbq->wake_batch);
598 
599 		/*
600 		 * Pairs with the memory barrier in sbitmap_queue_resize() to
601 		 * ensure that we see the batch size update before the wait
602 		 * count is reset.
603 		 */
604 		smp_mb__before_atomic();
605 
606 		/*
607 		 * For concurrent callers of this, the one that failed the
608 		 * atomic_cmpxhcg() race should call this function again
609 		 * to wakeup a new batch on a different 'ws'.
610 		 */
611 		ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch);
612 		if (ret == wait_cnt) {
613 			sbq_index_atomic_inc(&sbq->wake_index);
614 			wake_up_nr(&ws->wait, wake_batch);
615 			return false;
616 		}
617 
618 		return true;
619 	}
620 
621 	return false;
622 }
623 
sbitmap_queue_wake_up(struct sbitmap_queue * sbq)624 void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
625 {
626 	while (__sbq_wake_up(sbq))
627 		;
628 }
629 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
630 
sbitmap_update_cpu_hint(struct sbitmap * sb,int cpu,int tag)631 static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
632 {
633 	if (likely(!sb->round_robin && tag < sb->depth))
634 		data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
635 }
636 
sbitmap_queue_clear_batch(struct sbitmap_queue * sbq,int offset,int * tags,int nr_tags)637 void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
638 				int *tags, int nr_tags)
639 {
640 	struct sbitmap *sb = &sbq->sb;
641 	unsigned long *addr = NULL;
642 	unsigned long mask = 0;
643 	int i;
644 
645 	smp_mb__before_atomic();
646 	for (i = 0; i < nr_tags; i++) {
647 		const int tag = tags[i] - offset;
648 		unsigned long *this_addr;
649 
650 		/* since we're clearing a batch, skip the deferred map */
651 		this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
652 		if (!addr) {
653 			addr = this_addr;
654 		} else if (addr != this_addr) {
655 			atomic_long_andnot(mask, (atomic_long_t *) addr);
656 			mask = 0;
657 			addr = this_addr;
658 		}
659 		mask |= (1UL << SB_NR_TO_BIT(sb, tag));
660 	}
661 
662 	if (mask)
663 		atomic_long_andnot(mask, (atomic_long_t *) addr);
664 
665 	smp_mb__after_atomic();
666 	sbitmap_queue_wake_up(sbq);
667 	sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
668 					tags[nr_tags - 1] - offset);
669 }
670 
sbitmap_queue_clear(struct sbitmap_queue * sbq,unsigned int nr,unsigned int cpu)671 void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
672 			 unsigned int cpu)
673 {
674 	/*
675 	 * Once the clear bit is set, the bit may be allocated out.
676 	 *
677 	 * Orders READ/WRITE on the associated instance(such as request
678 	 * of blk_mq) by this bit for avoiding race with re-allocation,
679 	 * and its pair is the memory barrier implied in __sbitmap_get_word.
680 	 *
681 	 * One invariant is that the clear bit has to be zero when the bit
682 	 * is in use.
683 	 */
684 	smp_mb__before_atomic();
685 	sbitmap_deferred_clear_bit(&sbq->sb, nr);
686 
687 	/*
688 	 * Pairs with the memory barrier in set_current_state() to ensure the
689 	 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
690 	 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
691 	 * waiter. See the comment on waitqueue_active().
692 	 */
693 	smp_mb__after_atomic();
694 	sbitmap_queue_wake_up(sbq);
695 	sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
696 }
697 EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
698 
sbitmap_queue_wake_all(struct sbitmap_queue * sbq)699 void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
700 {
701 	int i, wake_index;
702 
703 	/*
704 	 * Pairs with the memory barrier in set_current_state() like in
705 	 * sbitmap_queue_wake_up().
706 	 */
707 	smp_mb();
708 	wake_index = atomic_read(&sbq->wake_index);
709 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
710 		struct sbq_wait_state *ws = &sbq->ws[wake_index];
711 
712 		if (waitqueue_active(&ws->wait))
713 			wake_up(&ws->wait);
714 
715 		wake_index = sbq_index_inc(wake_index);
716 	}
717 }
718 EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
719 
sbitmap_queue_show(struct sbitmap_queue * sbq,struct seq_file * m)720 void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
721 {
722 	bool first;
723 	int i;
724 
725 	sbitmap_show(&sbq->sb, m);
726 
727 	seq_puts(m, "alloc_hint={");
728 	first = true;
729 	for_each_possible_cpu(i) {
730 		if (!first)
731 			seq_puts(m, ", ");
732 		first = false;
733 		seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
734 	}
735 	seq_puts(m, "}\n");
736 
737 	seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
738 	seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
739 	seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
740 
741 	seq_puts(m, "ws={\n");
742 	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
743 		struct sbq_wait_state *ws = &sbq->ws[i];
744 
745 		seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
746 			   atomic_read(&ws->wait_cnt),
747 			   waitqueue_active(&ws->wait) ? "active" : "inactive");
748 	}
749 	seq_puts(m, "}\n");
750 
751 	seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
752 	seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
753 }
754 EXPORT_SYMBOL_GPL(sbitmap_queue_show);
755 
sbitmap_add_wait_queue(struct sbitmap_queue * sbq,struct sbq_wait_state * ws,struct sbq_wait * sbq_wait)756 void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
757 			    struct sbq_wait_state *ws,
758 			    struct sbq_wait *sbq_wait)
759 {
760 	if (!sbq_wait->sbq) {
761 		sbq_wait->sbq = sbq;
762 		atomic_inc(&sbq->ws_active);
763 		add_wait_queue(&ws->wait, &sbq_wait->wait);
764 	}
765 }
766 EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
767 
sbitmap_del_wait_queue(struct sbq_wait * sbq_wait)768 void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
769 {
770 	list_del_init(&sbq_wait->wait.entry);
771 	if (sbq_wait->sbq) {
772 		atomic_dec(&sbq_wait->sbq->ws_active);
773 		sbq_wait->sbq = NULL;
774 	}
775 }
776 EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
777 
sbitmap_prepare_to_wait(struct sbitmap_queue * sbq,struct sbq_wait_state * ws,struct sbq_wait * sbq_wait,int state)778 void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
779 			     struct sbq_wait_state *ws,
780 			     struct sbq_wait *sbq_wait, int state)
781 {
782 	if (!sbq_wait->sbq) {
783 		atomic_inc(&sbq->ws_active);
784 		sbq_wait->sbq = sbq;
785 	}
786 	prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
787 }
788 EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
789 
sbitmap_finish_wait(struct sbitmap_queue * sbq,struct sbq_wait_state * ws,struct sbq_wait * sbq_wait)790 void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
791 			 struct sbq_wait *sbq_wait)
792 {
793 	finish_wait(&ws->wait, &sbq_wait->wait);
794 	if (sbq_wait->sbq) {
795 		atomic_dec(&sbq->ws_active);
796 		sbq_wait->sbq = NULL;
797 	}
798 }
799 EXPORT_SYMBOL_GPL(sbitmap_finish_wait);
800