Linux Audio

Check our new training course

Embedded Linux training

Mar 10-20, 2025, special US time zones
Register
Loading...
v5.4
  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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 12/*
 13 * See if we have deferred clears that we can batch move
 14 */
 15static inline bool sbitmap_deferred_clear(struct sbitmap *sb, int index)
 16{
 17	unsigned long mask, val;
 18	bool ret = false;
 19	unsigned long flags;
 20
 21	spin_lock_irqsave(&sb->map[index].swap_lock, flags);
 22
 23	if (!sb->map[index].cleared)
 24		goto out_unlock;
 25
 26	/*
 27	 * First get a stable cleared mask, setting the old mask to 0.
 28	 */
 29	mask = xchg(&sb->map[index].cleared, 0);
 30
 31	/*
 32	 * Now clear the masked bits in our free word
 33	 */
 34	do {
 35		val = sb->map[index].word;
 36	} while (cmpxchg(&sb->map[index].word, val, val & ~mask) != val);
 37
 38	ret = true;
 39out_unlock:
 40	spin_unlock_irqrestore(&sb->map[index].swap_lock, flags);
 41	return ret;
 42}
 43
 44int sbitmap_init_node(struct sbitmap *sb, unsigned int depth, int shift,
 45		      gfp_t flags, int node)
 
 46{
 47	unsigned int bits_per_word;
 48	unsigned int i;
 49
 50	if (shift < 0) {
 51		shift = ilog2(BITS_PER_LONG);
 52		/*
 53		 * If the bitmap is small, shrink the number of bits per word so
 54		 * we spread over a few cachelines, at least. If less than 4
 55		 * bits, just forget about it, it's not going to work optimally
 56		 * anyway.
 57		 */
 58		if (depth >= 4) {
 59			while ((4U << shift) > depth)
 60				shift--;
 61		}
 62	}
 63	bits_per_word = 1U << shift;
 64	if (bits_per_word > BITS_PER_LONG)
 65		return -EINVAL;
 66
 67	sb->shift = shift;
 68	sb->depth = depth;
 69	sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
 
 70
 71	if (depth == 0) {
 72		sb->map = NULL;
 73		return 0;
 74	}
 75
 76	sb->map = kcalloc_node(sb->map_nr, sizeof(*sb->map), flags, node);
 77	if (!sb->map)
 78		return -ENOMEM;
 
 
 
 79
 80	for (i = 0; i < sb->map_nr; i++) {
 81		sb->map[i].depth = min(depth, bits_per_word);
 82		depth -= sb->map[i].depth;
 83		spin_lock_init(&sb->map[i].swap_lock);
 84	}
 
 85	return 0;
 86}
 87EXPORT_SYMBOL_GPL(sbitmap_init_node);
 88
 89void sbitmap_resize(struct sbitmap *sb, unsigned int depth)
 90{
 91	unsigned int bits_per_word = 1U << sb->shift;
 92	unsigned int i;
 93
 94	for (i = 0; i < sb->map_nr; i++)
 95		sbitmap_deferred_clear(sb, i);
 96
 97	sb->depth = depth;
 98	sb->map_nr = DIV_ROUND_UP(sb->depth, bits_per_word);
 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;
103	}
104}
105EXPORT_SYMBOL_GPL(sbitmap_resize);
106
107static int __sbitmap_get_word(unsigned long *word, unsigned long depth,
108			      unsigned int hint, bool wrap)
109{
110	unsigned int orig_hint = hint;
111	int nr;
112
 
 
 
113	while (1) {
114		nr = find_next_zero_bit(word, depth, hint);
115		if (unlikely(nr >= depth)) {
116			/*
117			 * We started with an offset, and we didn't reset the
118			 * offset to 0 in a failure case, so start from 0 to
119			 * exhaust the map.
120			 */
121			if (orig_hint && hint && wrap) {
122				hint = orig_hint = 0;
123				continue;
124			}
125			return -1;
126		}
127
128		if (!test_and_set_bit_lock(nr, word))
129			break;
130
131		hint = nr + 1;
132		if (hint >= depth - 1)
133			hint = 0;
134	}
135
136	return nr;
137}
138
139static int sbitmap_find_bit_in_index(struct sbitmap *sb, int index,
140				     unsigned int alloc_hint, bool round_robin)
 
 
141{
142	int nr;
143
144	do {
145		nr = __sbitmap_get_word(&sb->map[index].word,
146					sb->map[index].depth, alloc_hint,
147					!round_robin);
148		if (nr != -1)
149			break;
150		if (!sbitmap_deferred_clear(sb, index))
151			break;
152	} while (1);
153
154	return nr;
155}
156
157int sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint, bool round_robin)
 
 
 
 
158{
159	unsigned int i, index;
160	int nr = -1;
161
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
162	index = SB_NR_TO_INDEX(sb, alloc_hint);
163
164	/*
165	 * Unless we're doing round robin tag allocation, just use the
166	 * alloc_hint to find the right word index. No point in looping
167	 * twice in find_next_zero_bit() for that case.
168	 */
169	if (round_robin)
170		alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
171	else
172		alloc_hint = 0;
173
174	for (i = 0; i < sb->map_nr; i++) {
175		nr = sbitmap_find_bit_in_index(sb, index, alloc_hint,
176						round_robin);
177		if (nr != -1) {
178			nr += index << sb->shift;
179			break;
180		}
181
182		/* Jump to next index. */
183		alloc_hint = 0;
184		if (++index >= sb->map_nr)
185			index = 0;
186	}
 
 
 
 
 
 
 
187
188	return nr;
189}
190EXPORT_SYMBOL_GPL(sbitmap_get);
191
192int sbitmap_get_shallow(struct sbitmap *sb, unsigned int alloc_hint,
193			unsigned long shallow_depth)
 
194{
195	unsigned int i, index;
196	int nr = -1;
197
198	index = SB_NR_TO_INDEX(sb, alloc_hint);
 
199
200	for (i = 0; i < sb->map_nr; i++) {
201again:
202		nr = __sbitmap_get_word(&sb->map[index].word,
203					min(sb->map[index].depth, shallow_depth),
204					SB_NR_TO_BIT(sb, alloc_hint), true);
205		if (nr != -1) {
206			nr += index << sb->shift;
207			break;
208		}
209
210		if (sbitmap_deferred_clear(sb, index))
211			goto again;
 
 
212
213		/* Jump to next index. */
214		index++;
215		alloc_hint = index << sb->shift;
216
217		if (index >= sb->map_nr) {
218			index = 0;
219			alloc_hint = 0;
220		}
221	}
222
223	return nr;
224}
225EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
226
227bool sbitmap_any_bit_set(const struct sbitmap *sb)
228{
229	unsigned int i;
230
231	for (i = 0; i < sb->map_nr; i++) {
232		if (sb->map[i].word & ~sb->map[i].cleared)
233			return true;
234	}
235	return false;
236}
237EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
238
239bool sbitmap_any_bit_clear(const struct sbitmap *sb)
240{
241	unsigned int i;
242
243	for (i = 0; i < sb->map_nr; i++) {
244		const struct sbitmap_word *word = &sb->map[i];
245		unsigned long mask = word->word & ~word->cleared;
246		unsigned long ret;
247
248		ret = find_first_zero_bit(&mask, word->depth);
249		if (ret < word->depth)
250			return true;
251	}
252	return false;
253}
254EXPORT_SYMBOL_GPL(sbitmap_any_bit_clear);
255
256static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
257{
258	unsigned int i, weight = 0;
259
260	for (i = 0; i < sb->map_nr; i++) {
261		const struct sbitmap_word *word = &sb->map[i];
 
262
263		if (set)
264			weight += bitmap_weight(&word->word, word->depth);
265		else
266			weight += bitmap_weight(&word->cleared, word->depth);
267	}
268	return weight;
269}
270
271static unsigned int sbitmap_weight(const struct sbitmap *sb)
272{
273	return __sbitmap_weight(sb, true);
274}
275
276static unsigned int sbitmap_cleared(const struct sbitmap *sb)
277{
278	return __sbitmap_weight(sb, false);
279}
 
280
281void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
282{
283	seq_printf(m, "depth=%u\n", sb->depth);
284	seq_printf(m, "busy=%u\n", sbitmap_weight(sb) - sbitmap_cleared(sb));
285	seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
286	seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
287	seq_printf(m, "map_nr=%u\n", sb->map_nr);
288}
289EXPORT_SYMBOL_GPL(sbitmap_show);
290
291static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
292{
293	if ((offset & 0xf) == 0) {
294		if (offset != 0)
295			seq_putc(m, '\n');
296		seq_printf(m, "%08x:", offset);
297	}
298	if ((offset & 0x1) == 0)
299		seq_putc(m, ' ');
300	seq_printf(m, "%02x", byte);
301}
302
303void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
304{
305	u8 byte = 0;
306	unsigned int byte_bits = 0;
307	unsigned int offset = 0;
308	int i;
309
310	for (i = 0; i < sb->map_nr; i++) {
311		unsigned long word = READ_ONCE(sb->map[i].word);
312		unsigned int word_bits = READ_ONCE(sb->map[i].depth);
 
 
 
313
314		while (word_bits > 0) {
315			unsigned int bits = min(8 - byte_bits, word_bits);
316
317			byte |= (word & (BIT(bits) - 1)) << byte_bits;
318			byte_bits += bits;
319			if (byte_bits == 8) {
320				emit_byte(m, offset, byte);
321				byte = 0;
322				byte_bits = 0;
323				offset++;
324			}
325			word >>= bits;
326			word_bits -= bits;
327		}
328	}
329	if (byte_bits) {
330		emit_byte(m, offset, byte);
331		offset++;
332	}
333	if (offset)
334		seq_putc(m, '\n');
335}
336EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
337
338static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
339					unsigned int depth)
340{
341	unsigned int wake_batch;
342	unsigned int shallow_depth;
343
344	/*
345	 * For each batch, we wake up one queue. We need to make sure that our
346	 * batch size is small enough that the full depth of the bitmap,
347	 * potentially limited by a shallow depth, is enough to wake up all of
348	 * the queues.
349	 *
350	 * Each full word of the bitmap has bits_per_word bits, and there might
351	 * be a partial word. There are depth / bits_per_word full words and
352	 * depth % bits_per_word bits left over. In bitwise arithmetic:
353	 *
354	 * bits_per_word = 1 << shift
355	 * depth / bits_per_word = depth >> shift
356	 * depth % bits_per_word = depth & ((1 << shift) - 1)
357	 *
358	 * Each word can be limited to sbq->min_shallow_depth bits.
359	 */
360	shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
361	depth = ((depth >> sbq->sb.shift) * shallow_depth +
362		 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
363	wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
364			     SBQ_WAKE_BATCH);
365
366	return wake_batch;
367}
368
369int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
370			    int shift, bool round_robin, gfp_t flags, int node)
371{
372	int ret;
373	int i;
374
375	ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node);
 
376	if (ret)
377		return ret;
378
379	sbq->alloc_hint = alloc_percpu_gfp(unsigned int, flags);
380	if (!sbq->alloc_hint) {
381		sbitmap_free(&sbq->sb);
382		return -ENOMEM;
383	}
384
385	if (depth && !round_robin) {
386		for_each_possible_cpu(i)
387			*per_cpu_ptr(sbq->alloc_hint, i) = prandom_u32() % depth;
388	}
389
390	sbq->min_shallow_depth = UINT_MAX;
391	sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
392	atomic_set(&sbq->wake_index, 0);
393	atomic_set(&sbq->ws_active, 0);
 
 
394
395	sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
396	if (!sbq->ws) {
397		free_percpu(sbq->alloc_hint);
398		sbitmap_free(&sbq->sb);
399		return -ENOMEM;
400	}
401
402	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
403		init_waitqueue_head(&sbq->ws[i].wait);
404		atomic_set(&sbq->ws[i].wait_cnt, sbq->wake_batch);
405	}
406
407	sbq->round_robin = round_robin;
408	return 0;
409}
410EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
411
412static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
413					    unsigned int depth)
414{
415	unsigned int wake_batch = sbq_calc_wake_batch(sbq, depth);
416	int i;
417
418	if (sbq->wake_batch != wake_batch) {
 
419		WRITE_ONCE(sbq->wake_batch, wake_batch);
420		/*
421		 * Pairs with the memory barrier in sbitmap_queue_wake_up()
422		 * to ensure that the batch size is updated before the wait
423		 * counts.
424		 */
425		smp_mb();
426		for (i = 0; i < SBQ_WAIT_QUEUES; i++)
427			atomic_set(&sbq->ws[i].wait_cnt, 1);
428	}
429}
430
 
 
 
 
 
 
 
 
 
 
 
 
 
431void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
432{
433	sbitmap_queue_update_wake_batch(sbq, depth);
434	sbitmap_resize(&sbq->sb, depth);
435}
436EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
437
438int __sbitmap_queue_get(struct sbitmap_queue *sbq)
439{
440	unsigned int hint, depth;
441	int nr;
442
443	hint = this_cpu_read(*sbq->alloc_hint);
444	depth = READ_ONCE(sbq->sb.depth);
445	if (unlikely(hint >= depth)) {
446		hint = depth ? prandom_u32() % depth : 0;
447		this_cpu_write(*sbq->alloc_hint, hint);
448	}
449	nr = sbitmap_get(&sbq->sb, hint, sbq->round_robin);
450
451	if (nr == -1) {
452		/* If the map is full, a hint won't do us much good. */
453		this_cpu_write(*sbq->alloc_hint, 0);
454	} else if (nr == hint || unlikely(sbq->round_robin)) {
455		/* Only update the hint if we used it. */
456		hint = nr + 1;
457		if (hint >= depth - 1)
458			hint = 0;
459		this_cpu_write(*sbq->alloc_hint, hint);
460	}
461
462	return nr;
463}
464EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
465
466int __sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
467				unsigned int shallow_depth)
468{
 
469	unsigned int hint, depth;
470	int nr;
 
471
472	WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
 
473
474	hint = this_cpu_read(*sbq->alloc_hint);
475	depth = READ_ONCE(sbq->sb.depth);
476	if (unlikely(hint >= depth)) {
477		hint = depth ? prandom_u32() % depth : 0;
478		this_cpu_write(*sbq->alloc_hint, hint);
479	}
480	nr = sbitmap_get_shallow(&sbq->sb, hint, shallow_depth);
481
482	if (nr == -1) {
483		/* If the map is full, a hint won't do us much good. */
484		this_cpu_write(*sbq->alloc_hint, 0);
485	} else if (nr == hint || unlikely(sbq->round_robin)) {
486		/* Only update the hint if we used it. */
487		hint = nr + 1;
488		if (hint >= depth - 1)
489			hint = 0;
490		this_cpu_write(*sbq->alloc_hint, hint);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
491	}
492
493	return nr;
494}
495EXPORT_SYMBOL_GPL(__sbitmap_queue_get_shallow);
 
 
 
 
 
 
 
 
496
497void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
498				     unsigned int min_shallow_depth)
499{
500	sbq->min_shallow_depth = min_shallow_depth;
501	sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
502}
503EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
504
505static struct sbq_wait_state *sbq_wake_ptr(struct sbitmap_queue *sbq)
506{
507	int i, wake_index;
508
509	if (!atomic_read(&sbq->ws_active))
510		return NULL;
511
512	wake_index = atomic_read(&sbq->wake_index);
513	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
514		struct sbq_wait_state *ws = &sbq->ws[wake_index];
515
 
 
 
 
 
 
 
 
516		if (waitqueue_active(&ws->wait)) {
517			if (wake_index != atomic_read(&sbq->wake_index))
518				atomic_set(&sbq->wake_index, wake_index);
519			return ws;
 
520		}
521
522		wake_index = sbq_index_inc(wake_index);
523	}
524
525	return NULL;
 
526}
527
528static bool __sbq_wake_up(struct sbitmap_queue *sbq)
529{
530	struct sbq_wait_state *ws;
531	unsigned int wake_batch;
532	int wait_cnt;
533
534	ws = sbq_wake_ptr(sbq);
535	if (!ws)
536		return false;
537
538	wait_cnt = atomic_dec_return(&ws->wait_cnt);
539	if (wait_cnt <= 0) {
540		int ret;
541
542		wake_batch = READ_ONCE(sbq->wake_batch);
 
 
 
 
543
544		/*
545		 * Pairs with the memory barrier in sbitmap_queue_resize() to
546		 * ensure that we see the batch size update before the wait
547		 * count is reset.
548		 */
549		smp_mb__before_atomic();
550
551		/*
552		 * For concurrent callers of this, the one that failed the
553		 * atomic_cmpxhcg() race should call this function again
554		 * to wakeup a new batch on a different 'ws'.
555		 */
556		ret = atomic_cmpxchg(&ws->wait_cnt, wait_cnt, wake_batch);
557		if (ret == wait_cnt) {
558			sbq_index_atomic_inc(&sbq->wake_index);
559			wake_up_nr(&ws->wait, wake_batch);
560			return false;
561		}
562
563		return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
564	}
565
566	return false;
567}
568
569void sbitmap_queue_wake_up(struct sbitmap_queue *sbq)
570{
571	while (__sbq_wake_up(sbq))
572		;
573}
574EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
575
576void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
577			 unsigned int cpu)
578{
579	/*
580	 * Once the clear bit is set, the bit may be allocated out.
581	 *
582	 * Orders READ/WRITE on the asssociated instance(such as request
583	 * of blk_mq) by this bit for avoiding race with re-allocation,
584	 * and its pair is the memory barrier implied in __sbitmap_get_word.
585	 *
586	 * One invariant is that the clear bit has to be zero when the bit
587	 * is in use.
588	 */
589	smp_mb__before_atomic();
590	sbitmap_deferred_clear_bit(&sbq->sb, nr);
591
592	/*
593	 * Pairs with the memory barrier in set_current_state() to ensure the
594	 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
595	 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
596	 * waiter. See the comment on waitqueue_active().
597	 */
598	smp_mb__after_atomic();
599	sbitmap_queue_wake_up(sbq);
600
601	if (likely(!sbq->round_robin && nr < sbq->sb.depth))
602		*per_cpu_ptr(sbq->alloc_hint, cpu) = nr;
603}
604EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
605
606void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
607{
608	int i, wake_index;
609
610	/*
611	 * Pairs with the memory barrier in set_current_state() like in
612	 * sbitmap_queue_wake_up().
613	 */
614	smp_mb();
615	wake_index = atomic_read(&sbq->wake_index);
616	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
617		struct sbq_wait_state *ws = &sbq->ws[wake_index];
618
619		if (waitqueue_active(&ws->wait))
620			wake_up(&ws->wait);
621
622		wake_index = sbq_index_inc(wake_index);
623	}
624}
625EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
626
627void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
628{
629	bool first;
630	int i;
631
632	sbitmap_show(&sbq->sb, m);
633
634	seq_puts(m, "alloc_hint={");
635	first = true;
636	for_each_possible_cpu(i) {
637		if (!first)
638			seq_puts(m, ", ");
639		first = false;
640		seq_printf(m, "%u", *per_cpu_ptr(sbq->alloc_hint, i));
641	}
642	seq_puts(m, "}\n");
643
644	seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
645	seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
646	seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
647
648	seq_puts(m, "ws={\n");
649	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
650		struct sbq_wait_state *ws = &sbq->ws[i];
651
652		seq_printf(m, "\t{.wait_cnt=%d, .wait=%s},\n",
653			   atomic_read(&ws->wait_cnt),
654			   waitqueue_active(&ws->wait) ? "active" : "inactive");
655	}
656	seq_puts(m, "}\n");
657
658	seq_printf(m, "round_robin=%d\n", sbq->round_robin);
659	seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
660}
661EXPORT_SYMBOL_GPL(sbitmap_queue_show);
662
663void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
664			    struct sbq_wait_state *ws,
665			    struct sbq_wait *sbq_wait)
666{
667	if (!sbq_wait->sbq) {
668		sbq_wait->sbq = sbq;
669		atomic_inc(&sbq->ws_active);
 
670	}
671	add_wait_queue(&ws->wait, &sbq_wait->wait);
672}
673EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
674
675void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
676{
677	list_del_init(&sbq_wait->wait.entry);
678	if (sbq_wait->sbq) {
679		atomic_dec(&sbq_wait->sbq->ws_active);
680		sbq_wait->sbq = NULL;
681	}
682}
683EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
684
685void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
686			     struct sbq_wait_state *ws,
687			     struct sbq_wait *sbq_wait, int state)
688{
689	if (!sbq_wait->sbq) {
690		atomic_inc(&sbq->ws_active);
691		sbq_wait->sbq = sbq;
692	}
693	prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
694}
695EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
696
697void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
698			 struct sbq_wait *sbq_wait)
699{
700	finish_wait(&ws->wait, &sbq_wait->wait);
701	if (sbq_wait->sbq) {
702		atomic_dec(&sbq->ws_active);
703		sbq_wait->sbq = NULL;
704	}
705}
706EXPORT_SYMBOL_GPL(sbitmap_finish_wait);
v6.8
  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
 12static 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) = get_random_u32_below(depth);
 25	}
 26	return 0;
 27}
 28
 29static 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 ? get_random_u32_below(depth) : 0;
 37		this_cpu_write(*sb->alloc_hint, hint);
 38	}
 39
 40	return hint;
 41}
 42
 43static 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 */
 63static 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
 83int 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}
121EXPORT_SYMBOL_GPL(sbitmap_init_node);
122
123void 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}
134EXPORT_SYMBOL_GPL(sbitmap_resize);
135
136static 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
170static int sbitmap_find_bit_in_word(struct sbitmap_word *map,
171				    unsigned int depth,
172				    unsigned int alloc_hint,
173				    bool wrap)
174{
175	int nr;
176
177	do {
178		nr = __sbitmap_get_word(&map->word, depth,
179					alloc_hint, wrap);
 
180		if (nr != -1)
181			break;
182		if (!sbitmap_deferred_clear(map))
183			break;
184	} while (1);
185
186	return nr;
187}
188
189static int sbitmap_find_bit(struct sbitmap *sb,
190			    unsigned int depth,
191			    unsigned int index,
192			    unsigned int alloc_hint,
193			    bool wrap)
194{
195	unsigned int i;
196	int nr = -1;
197
198	for (i = 0; i < sb->map_nr; i++) {
199		nr = sbitmap_find_bit_in_word(&sb->map[index],
200					      min_t(unsigned int,
201						    __map_depth(sb, index),
202						    depth),
203					      alloc_hint, wrap);
204
205		if (nr != -1) {
206			nr += index << sb->shift;
207			break;
208		}
209
210		/* Jump to next index. */
211		alloc_hint = 0;
212		if (++index >= sb->map_nr)
213			index = 0;
214	}
215
216	return nr;
217}
218
219static int __sbitmap_get(struct sbitmap *sb, unsigned int alloc_hint)
220{
221	unsigned int index;
222
223	index = SB_NR_TO_INDEX(sb, alloc_hint);
224
225	/*
226	 * Unless we're doing round robin tag allocation, just use the
227	 * alloc_hint to find the right word index. No point in looping
228	 * twice in find_next_zero_bit() for that case.
229	 */
230	if (sb->round_robin)
231		alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
232	else
233		alloc_hint = 0;
234
235	return sbitmap_find_bit(sb, UINT_MAX, index, alloc_hint,
236				!sb->round_robin);
237}
 
 
 
 
238
239int sbitmap_get(struct sbitmap *sb)
240{
241	int nr;
242	unsigned int hint, depth;
243
244	if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
245		return -1;
246
247	depth = READ_ONCE(sb->depth);
248	hint = update_alloc_hint_before_get(sb, depth);
249	nr = __sbitmap_get(sb, hint);
250	update_alloc_hint_after_get(sb, depth, hint, nr);
251
252	return nr;
253}
254EXPORT_SYMBOL_GPL(sbitmap_get);
255
256static int __sbitmap_get_shallow(struct sbitmap *sb,
257				 unsigned int alloc_hint,
258				 unsigned long shallow_depth)
259{
260	unsigned int index;
 
261
262	index = SB_NR_TO_INDEX(sb, alloc_hint);
263	alloc_hint = SB_NR_TO_BIT(sb, alloc_hint);
264
265	return sbitmap_find_bit(sb, shallow_depth, index, alloc_hint, true);
266}
 
 
 
 
 
 
 
267
268int sbitmap_get_shallow(struct sbitmap *sb, unsigned long shallow_depth)
269{
270	int nr;
271	unsigned int hint, depth;
272
273	if (WARN_ON_ONCE(unlikely(!sb->alloc_hint)))
274		return -1;
 
275
276	depth = READ_ONCE(sb->depth);
277	hint = update_alloc_hint_before_get(sb, depth);
278	nr = __sbitmap_get_shallow(sb, hint, shallow_depth);
279	update_alloc_hint_after_get(sb, depth, hint, nr);
 
280
281	return nr;
282}
283EXPORT_SYMBOL_GPL(sbitmap_get_shallow);
284
285bool sbitmap_any_bit_set(const struct sbitmap *sb)
286{
287	unsigned int i;
288
289	for (i = 0; i < sb->map_nr; i++) {
290		if (sb->map[i].word & ~sb->map[i].cleared)
291			return true;
292	}
293	return false;
294}
295EXPORT_SYMBOL_GPL(sbitmap_any_bit_set);
296
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
297static unsigned int __sbitmap_weight(const struct sbitmap *sb, bool set)
298{
299	unsigned int i, weight = 0;
300
301	for (i = 0; i < sb->map_nr; i++) {
302		const struct sbitmap_word *word = &sb->map[i];
303		unsigned int word_depth = __map_depth(sb, i);
304
305		if (set)
306			weight += bitmap_weight(&word->word, word_depth);
307		else
308			weight += bitmap_weight(&word->cleared, word_depth);
309	}
310	return weight;
311}
312
313static unsigned int sbitmap_cleared(const struct sbitmap *sb)
314{
315	return __sbitmap_weight(sb, false);
316}
317
318unsigned int sbitmap_weight(const struct sbitmap *sb)
319{
320	return __sbitmap_weight(sb, true) - sbitmap_cleared(sb);
321}
322EXPORT_SYMBOL_GPL(sbitmap_weight);
323
324void sbitmap_show(struct sbitmap *sb, struct seq_file *m)
325{
326	seq_printf(m, "depth=%u\n", sb->depth);
327	seq_printf(m, "busy=%u\n", sbitmap_weight(sb));
328	seq_printf(m, "cleared=%u\n", sbitmap_cleared(sb));
329	seq_printf(m, "bits_per_word=%u\n", 1U << sb->shift);
330	seq_printf(m, "map_nr=%u\n", sb->map_nr);
331}
332EXPORT_SYMBOL_GPL(sbitmap_show);
333
334static inline void emit_byte(struct seq_file *m, unsigned int offset, u8 byte)
335{
336	if ((offset & 0xf) == 0) {
337		if (offset != 0)
338			seq_putc(m, '\n');
339		seq_printf(m, "%08x:", offset);
340	}
341	if ((offset & 0x1) == 0)
342		seq_putc(m, ' ');
343	seq_printf(m, "%02x", byte);
344}
345
346void sbitmap_bitmap_show(struct sbitmap *sb, struct seq_file *m)
347{
348	u8 byte = 0;
349	unsigned int byte_bits = 0;
350	unsigned int offset = 0;
351	int i;
352
353	for (i = 0; i < sb->map_nr; i++) {
354		unsigned long word = READ_ONCE(sb->map[i].word);
355		unsigned long cleared = READ_ONCE(sb->map[i].cleared);
356		unsigned int word_bits = __map_depth(sb, i);
357
358		word &= ~cleared;
359
360		while (word_bits > 0) {
361			unsigned int bits = min(8 - byte_bits, word_bits);
362
363			byte |= (word & (BIT(bits) - 1)) << byte_bits;
364			byte_bits += bits;
365			if (byte_bits == 8) {
366				emit_byte(m, offset, byte);
367				byte = 0;
368				byte_bits = 0;
369				offset++;
370			}
371			word >>= bits;
372			word_bits -= bits;
373		}
374	}
375	if (byte_bits) {
376		emit_byte(m, offset, byte);
377		offset++;
378	}
379	if (offset)
380		seq_putc(m, '\n');
381}
382EXPORT_SYMBOL_GPL(sbitmap_bitmap_show);
383
384static unsigned int sbq_calc_wake_batch(struct sbitmap_queue *sbq,
385					unsigned int depth)
386{
387	unsigned int wake_batch;
388	unsigned int shallow_depth;
389
390	/*
 
 
 
 
 
391	 * Each full word of the bitmap has bits_per_word bits, and there might
392	 * be a partial word. There are depth / bits_per_word full words and
393	 * depth % bits_per_word bits left over. In bitwise arithmetic:
394	 *
395	 * bits_per_word = 1 << shift
396	 * depth / bits_per_word = depth >> shift
397	 * depth % bits_per_word = depth & ((1 << shift) - 1)
398	 *
399	 * Each word can be limited to sbq->min_shallow_depth bits.
400	 */
401	shallow_depth = min(1U << sbq->sb.shift, sbq->min_shallow_depth);
402	depth = ((depth >> sbq->sb.shift) * shallow_depth +
403		 min(depth & ((1U << sbq->sb.shift) - 1), shallow_depth));
404	wake_batch = clamp_t(unsigned int, depth / SBQ_WAIT_QUEUES, 1,
405			     SBQ_WAKE_BATCH);
406
407	return wake_batch;
408}
409
410int sbitmap_queue_init_node(struct sbitmap_queue *sbq, unsigned int depth,
411			    int shift, bool round_robin, gfp_t flags, int node)
412{
413	int ret;
414	int i;
415
416	ret = sbitmap_init_node(&sbq->sb, depth, shift, flags, node,
417				round_robin, true);
418	if (ret)
419		return ret;
420
 
 
 
 
 
 
 
 
 
 
 
421	sbq->min_shallow_depth = UINT_MAX;
422	sbq->wake_batch = sbq_calc_wake_batch(sbq, depth);
423	atomic_set(&sbq->wake_index, 0);
424	atomic_set(&sbq->ws_active, 0);
425	atomic_set(&sbq->completion_cnt, 0);
426	atomic_set(&sbq->wakeup_cnt, 0);
427
428	sbq->ws = kzalloc_node(SBQ_WAIT_QUEUES * sizeof(*sbq->ws), flags, node);
429	if (!sbq->ws) {
 
430		sbitmap_free(&sbq->sb);
431		return -ENOMEM;
432	}
433
434	for (i = 0; i < SBQ_WAIT_QUEUES; i++)
435		init_waitqueue_head(&sbq->ws[i].wait);
 
 
436
 
437	return 0;
438}
439EXPORT_SYMBOL_GPL(sbitmap_queue_init_node);
440
441static void sbitmap_queue_update_wake_batch(struct sbitmap_queue *sbq,
442					    unsigned int depth)
443{
444	unsigned int wake_batch;
 
445
446	wake_batch = sbq_calc_wake_batch(sbq, depth);
447	if (sbq->wake_batch != wake_batch)
448		WRITE_ONCE(sbq->wake_batch, wake_batch);
 
 
 
 
 
 
 
 
 
449}
450
451void sbitmap_queue_recalculate_wake_batch(struct sbitmap_queue *sbq,
452					    unsigned int users)
453{
454	unsigned int wake_batch;
455	unsigned int depth = (sbq->sb.depth + users - 1) / users;
456
457	wake_batch = clamp_val(depth / SBQ_WAIT_QUEUES,
458			1, SBQ_WAKE_BATCH);
459
460	WRITE_ONCE(sbq->wake_batch, wake_batch);
461}
462EXPORT_SYMBOL_GPL(sbitmap_queue_recalculate_wake_batch);
463
464void sbitmap_queue_resize(struct sbitmap_queue *sbq, unsigned int depth)
465{
466	sbitmap_queue_update_wake_batch(sbq, depth);
467	sbitmap_resize(&sbq->sb, depth);
468}
469EXPORT_SYMBOL_GPL(sbitmap_queue_resize);
470
471int __sbitmap_queue_get(struct sbitmap_queue *sbq)
472{
473	return sbitmap_get(&sbq->sb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
474}
475EXPORT_SYMBOL_GPL(__sbitmap_queue_get);
476
477unsigned long __sbitmap_queue_get_batch(struct sbitmap_queue *sbq, int nr_tags,
478					unsigned int *offset)
479{
480	struct sbitmap *sb = &sbq->sb;
481	unsigned int hint, depth;
482	unsigned long index, nr;
483	int i;
484
485	if (unlikely(sb->round_robin))
486		return 0;
487
488	depth = READ_ONCE(sb->depth);
489	hint = update_alloc_hint_before_get(sb, depth);
 
 
 
 
 
490
491	index = SB_NR_TO_INDEX(sb, hint);
492
493	for (i = 0; i < sb->map_nr; i++) {
494		struct sbitmap_word *map = &sb->map[index];
495		unsigned long get_mask;
496		unsigned int map_depth = __map_depth(sb, index);
497
498		sbitmap_deferred_clear(map);
499		if (map->word == (1UL << (map_depth - 1)) - 1)
500			goto next;
501
502		nr = find_first_zero_bit(&map->word, map_depth);
503		if (nr + nr_tags <= map_depth) {
504			atomic_long_t *ptr = (atomic_long_t *) &map->word;
505			unsigned long val;
506
507			get_mask = ((1UL << nr_tags) - 1) << nr;
508			val = READ_ONCE(map->word);
509			while (!atomic_long_try_cmpxchg(ptr, &val,
510							  get_mask | val))
511				;
512			get_mask = (get_mask & ~val) >> nr;
513			if (get_mask) {
514				*offset = nr + (index << sb->shift);
515				update_alloc_hint_after_get(sb, depth, hint,
516							*offset + nr_tags - 1);
517				return get_mask;
518			}
519		}
520next:
521		/* Jump to next index. */
522		if (++index >= sb->map_nr)
523			index = 0;
524	}
525
526	return 0;
527}
528
529int sbitmap_queue_get_shallow(struct sbitmap_queue *sbq,
530			      unsigned int shallow_depth)
531{
532	WARN_ON_ONCE(shallow_depth < sbq->min_shallow_depth);
533
534	return sbitmap_get_shallow(&sbq->sb, shallow_depth);
535}
536EXPORT_SYMBOL_GPL(sbitmap_queue_get_shallow);
537
538void sbitmap_queue_min_shallow_depth(struct sbitmap_queue *sbq,
539				     unsigned int min_shallow_depth)
540{
541	sbq->min_shallow_depth = min_shallow_depth;
542	sbitmap_queue_update_wake_batch(sbq, sbq->sb.depth);
543}
544EXPORT_SYMBOL_GPL(sbitmap_queue_min_shallow_depth);
545
546static void __sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
547{
548	int i, wake_index, woken;
549
550	if (!atomic_read(&sbq->ws_active))
551		return;
552
553	wake_index = atomic_read(&sbq->wake_index);
554	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
555		struct sbq_wait_state *ws = &sbq->ws[wake_index];
556
557		/*
558		 * Advance the index before checking the current queue.
559		 * It improves fairness, by ensuring the queue doesn't
560		 * need to be fully emptied before trying to wake up
561		 * from the next one.
562		 */
563		wake_index = sbq_index_inc(wake_index);
564
565		if (waitqueue_active(&ws->wait)) {
566			woken = wake_up_nr(&ws->wait, nr);
567			if (woken == nr)
568				break;
569			nr -= woken;
570		}
 
 
571	}
572
573	if (wake_index != atomic_read(&sbq->wake_index))
574		atomic_set(&sbq->wake_index, wake_index);
575}
576
577void sbitmap_queue_wake_up(struct sbitmap_queue *sbq, int nr)
578{
579	unsigned int wake_batch = READ_ONCE(sbq->wake_batch);
580	unsigned int wakeups;
 
581
582	if (!atomic_read(&sbq->ws_active))
583		return;
 
584
585	atomic_add(nr, &sbq->completion_cnt);
586	wakeups = atomic_read(&sbq->wakeup_cnt);
 
587
588	do {
589		if (atomic_read(&sbq->completion_cnt) - wakeups < wake_batch)
590			return;
591	} while (!atomic_try_cmpxchg(&sbq->wakeup_cnt,
592				     &wakeups, wakeups + wake_batch));
593
594	__sbitmap_queue_wake_up(sbq, wake_batch);
595}
596EXPORT_SYMBOL_GPL(sbitmap_queue_wake_up);
 
 
 
597
598static inline void sbitmap_update_cpu_hint(struct sbitmap *sb, int cpu, int tag)
599{
600	if (likely(!sb->round_robin && tag < sb->depth))
601		data_race(*per_cpu_ptr(sb->alloc_hint, cpu) = tag);
602}
 
 
 
 
 
 
603
604void sbitmap_queue_clear_batch(struct sbitmap_queue *sbq, int offset,
605				int *tags, int nr_tags)
606{
607	struct sbitmap *sb = &sbq->sb;
608	unsigned long *addr = NULL;
609	unsigned long mask = 0;
610	int i;
611
612	smp_mb__before_atomic();
613	for (i = 0; i < nr_tags; i++) {
614		const int tag = tags[i] - offset;
615		unsigned long *this_addr;
616
617		/* since we're clearing a batch, skip the deferred map */
618		this_addr = &sb->map[SB_NR_TO_INDEX(sb, tag)].word;
619		if (!addr) {
620			addr = this_addr;
621		} else if (addr != this_addr) {
622			atomic_long_andnot(mask, (atomic_long_t *) addr);
623			mask = 0;
624			addr = this_addr;
625		}
626		mask |= (1UL << SB_NR_TO_BIT(sb, tag));
627	}
628
629	if (mask)
630		atomic_long_andnot(mask, (atomic_long_t *) addr);
631
632	smp_mb__after_atomic();
633	sbitmap_queue_wake_up(sbq, nr_tags);
634	sbitmap_update_cpu_hint(&sbq->sb, raw_smp_processor_id(),
635					tags[nr_tags - 1] - offset);
636}
 
637
638void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
639			 unsigned int cpu)
640{
641	/*
642	 * Once the clear bit is set, the bit may be allocated out.
643	 *
644	 * Orders READ/WRITE on the associated instance(such as request
645	 * of blk_mq) by this bit for avoiding race with re-allocation,
646	 * and its pair is the memory barrier implied in __sbitmap_get_word.
647	 *
648	 * One invariant is that the clear bit has to be zero when the bit
649	 * is in use.
650	 */
651	smp_mb__before_atomic();
652	sbitmap_deferred_clear_bit(&sbq->sb, nr);
653
654	/*
655	 * Pairs with the memory barrier in set_current_state() to ensure the
656	 * proper ordering of clear_bit_unlock()/waitqueue_active() in the waker
657	 * and test_and_set_bit_lock()/prepare_to_wait()/finish_wait() in the
658	 * waiter. See the comment on waitqueue_active().
659	 */
660	smp_mb__after_atomic();
661	sbitmap_queue_wake_up(sbq, 1);
662	sbitmap_update_cpu_hint(&sbq->sb, cpu, nr);
 
 
663}
664EXPORT_SYMBOL_GPL(sbitmap_queue_clear);
665
666void sbitmap_queue_wake_all(struct sbitmap_queue *sbq)
667{
668	int i, wake_index;
669
670	/*
671	 * Pairs with the memory barrier in set_current_state() like in
672	 * sbitmap_queue_wake_up().
673	 */
674	smp_mb();
675	wake_index = atomic_read(&sbq->wake_index);
676	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
677		struct sbq_wait_state *ws = &sbq->ws[wake_index];
678
679		if (waitqueue_active(&ws->wait))
680			wake_up(&ws->wait);
681
682		wake_index = sbq_index_inc(wake_index);
683	}
684}
685EXPORT_SYMBOL_GPL(sbitmap_queue_wake_all);
686
687void sbitmap_queue_show(struct sbitmap_queue *sbq, struct seq_file *m)
688{
689	bool first;
690	int i;
691
692	sbitmap_show(&sbq->sb, m);
693
694	seq_puts(m, "alloc_hint={");
695	first = true;
696	for_each_possible_cpu(i) {
697		if (!first)
698			seq_puts(m, ", ");
699		first = false;
700		seq_printf(m, "%u", *per_cpu_ptr(sbq->sb.alloc_hint, i));
701	}
702	seq_puts(m, "}\n");
703
704	seq_printf(m, "wake_batch=%u\n", sbq->wake_batch);
705	seq_printf(m, "wake_index=%d\n", atomic_read(&sbq->wake_index));
706	seq_printf(m, "ws_active=%d\n", atomic_read(&sbq->ws_active));
707
708	seq_puts(m, "ws={\n");
709	for (i = 0; i < SBQ_WAIT_QUEUES; i++) {
710		struct sbq_wait_state *ws = &sbq->ws[i];
711		seq_printf(m, "\t{.wait=%s},\n",
 
 
712			   waitqueue_active(&ws->wait) ? "active" : "inactive");
713	}
714	seq_puts(m, "}\n");
715
716	seq_printf(m, "round_robin=%d\n", sbq->sb.round_robin);
717	seq_printf(m, "min_shallow_depth=%u\n", sbq->min_shallow_depth);
718}
719EXPORT_SYMBOL_GPL(sbitmap_queue_show);
720
721void sbitmap_add_wait_queue(struct sbitmap_queue *sbq,
722			    struct sbq_wait_state *ws,
723			    struct sbq_wait *sbq_wait)
724{
725	if (!sbq_wait->sbq) {
726		sbq_wait->sbq = sbq;
727		atomic_inc(&sbq->ws_active);
728		add_wait_queue(&ws->wait, &sbq_wait->wait);
729	}
 
730}
731EXPORT_SYMBOL_GPL(sbitmap_add_wait_queue);
732
733void sbitmap_del_wait_queue(struct sbq_wait *sbq_wait)
734{
735	list_del_init(&sbq_wait->wait.entry);
736	if (sbq_wait->sbq) {
737		atomic_dec(&sbq_wait->sbq->ws_active);
738		sbq_wait->sbq = NULL;
739	}
740}
741EXPORT_SYMBOL_GPL(sbitmap_del_wait_queue);
742
743void sbitmap_prepare_to_wait(struct sbitmap_queue *sbq,
744			     struct sbq_wait_state *ws,
745			     struct sbq_wait *sbq_wait, int state)
746{
747	if (!sbq_wait->sbq) {
748		atomic_inc(&sbq->ws_active);
749		sbq_wait->sbq = sbq;
750	}
751	prepare_to_wait_exclusive(&ws->wait, &sbq_wait->wait, state);
752}
753EXPORT_SYMBOL_GPL(sbitmap_prepare_to_wait);
754
755void sbitmap_finish_wait(struct sbitmap_queue *sbq, struct sbq_wait_state *ws,
756			 struct sbq_wait *sbq_wait)
757{
758	finish_wait(&ws->wait, &sbq_wait->wait);
759	if (sbq_wait->sbq) {
760		atomic_dec(&sbq->ws_active);
761		sbq_wait->sbq = NULL;
762	}
763}
764EXPORT_SYMBOL_GPL(sbitmap_finish_wait);