1/*
2 * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread
3 * over multiple cachelines to avoid ping-pong between multiple submitters
4 * or submitter and completer. Uses rolling wakeups to avoid falling of
5 * the scaling cliff when we run out of tags and have to start putting
6 * submitters to sleep.
7 *
8 * Uses active queue tracking to support fairer distribution of tags
9 * between multiple submitters when a shared tag map is used.
10 *
11 * Copyright (C) 2013-2014 Jens Axboe
12 */
13#include <linux/kernel.h>
14#include <linux/module.h>
15#include <linux/random.h>
16
17#include <linux/blk-mq.h>
18#include "blk.h"
19#include "blk-mq.h"
20#include "blk-mq-tag.h"
21
22static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
23{
24	int i;
25
26	for (i = 0; i < bt->map_nr; i++) {
27		struct blk_align_bitmap *bm = &bt->map[i];
28		int ret;
29
30		ret = find_first_zero_bit(&bm->word, bm->depth);
31		if (ret < bm->depth)
32			return true;
33	}
34
35	return false;
36}
37
38bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
39{
40	if (!tags)
41		return true;
42
43	return bt_has_free_tags(&tags->bitmap_tags);
44}
45
46static inline int bt_index_inc(int index)
47{
48	return (index + 1) & (BT_WAIT_QUEUES - 1);
49}
50
51static inline void bt_index_atomic_inc(atomic_t *index)
52{
53	int old = atomic_read(index);
54	int new = bt_index_inc(old);
55	atomic_cmpxchg(index, old, new);
56}
57
58/*
59 * If a previously inactive queue goes active, bump the active user count.
60 */
61bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
62{
63	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
64	    !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
65		atomic_inc(&hctx->tags->active_queues);
66
67	return true;
68}
69
70/*
71 * Wakeup all potentially sleeping on tags
72 */
73void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
74{
75	struct blk_mq_bitmap_tags *bt;
76	int i, wake_index;
77
78	bt = &tags->bitmap_tags;
79	wake_index = atomic_read(&bt->wake_index);
80	for (i = 0; i < BT_WAIT_QUEUES; i++) {
81		struct bt_wait_state *bs = &bt->bs[wake_index];
82
83		if (waitqueue_active(&bs->wait))
84			wake_up(&bs->wait);
85
86		wake_index = bt_index_inc(wake_index);
87	}
88
89	if (include_reserve) {
90		bt = &tags->breserved_tags;
91		if (waitqueue_active(&bt->bs[0].wait))
92			wake_up(&bt->bs[0].wait);
93	}
94}
95
96/*
97 * If a previously busy queue goes inactive, potential waiters could now
98 * be allowed to queue. Wake them up and check.
99 */
100void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
101{
102	struct blk_mq_tags *tags = hctx->tags;
103
104	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
105		return;
106
107	atomic_dec(&tags->active_queues);
108
109	blk_mq_tag_wakeup_all(tags, false);
110}
111
112/*
113 * For shared tag users, we track the number of currently active users
114 * and attempt to provide a fair share of the tag depth for each of them.
115 */
116static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
117				  struct blk_mq_bitmap_tags *bt)
118{
119	unsigned int depth, users;
120
121	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
122		return true;
123	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
124		return true;
125
126	/*
127	 * Don't try dividing an ant
128	 */
129	if (bt->depth == 1)
130		return true;
131
132	users = atomic_read(&hctx->tags->active_queues);
133	if (!users)
134		return true;
135
136	/*
137	 * Allow at least some tags
138	 */
139	depth = max((bt->depth + users - 1) / users, 4U);
140	return atomic_read(&hctx->nr_active) < depth;
141}
142
143static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag,
144			 bool nowrap)
145{
146	int tag, org_last_tag = last_tag;
147
148	while (1) {
149		tag = find_next_zero_bit(&bm->word, bm->depth, last_tag);
150		if (unlikely(tag >= bm->depth)) {
151			/*
152			 * We started with an offset, and we didn't reset the
153			 * offset to 0 in a failure case, so start from 0 to
154			 * exhaust the map.
155			 */
156			if (org_last_tag && last_tag && !nowrap) {
157				last_tag = org_last_tag = 0;
158				continue;
159			}
160			return -1;
161		}
162
163		if (!test_and_set_bit(tag, &bm->word))
164			break;
165
166		last_tag = tag + 1;
167		if (last_tag >= bm->depth - 1)
168			last_tag = 0;
169	}
170
171	return tag;
172}
173
174#define BT_ALLOC_RR(tags) (tags->alloc_policy == BLK_TAG_ALLOC_RR)
175
176/*
177 * Straight forward bitmap tag implementation, where each bit is a tag
178 * (cleared == free, and set == busy). The small twist is using per-cpu
179 * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
180 * contexts. This enables us to drastically limit the space searched,
181 * without dirtying an extra shared cacheline like we would if we stored
182 * the cache value inside the shared blk_mq_bitmap_tags structure. On top
183 * of that, each word of tags is in a separate cacheline. This means that
184 * multiple users will tend to stick to different cachelines, at least
185 * until the map is exhausted.
186 */
187static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
188		    unsigned int *tag_cache, struct blk_mq_tags *tags)
189{
190	unsigned int last_tag, org_last_tag;
191	int index, i, tag;
192
193	if (!hctx_may_queue(hctx, bt))
194		return -1;
195
196	last_tag = org_last_tag = *tag_cache;
197	index = TAG_TO_INDEX(bt, last_tag);
198
199	for (i = 0; i < bt->map_nr; i++) {
200		tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag),
201				    BT_ALLOC_RR(tags));
202		if (tag != -1) {
203			tag += (index << bt->bits_per_word);
204			goto done;
205		}
206
207		/*
208		 * Jump to next index, and reset the last tag to be the
209		 * first tag of that index
210		 */
211		index++;
212		last_tag = (index << bt->bits_per_word);
213
214		if (index >= bt->map_nr) {
215			index = 0;
216			last_tag = 0;
217		}
218	}
219
220	*tag_cache = 0;
221	return -1;
222
223	/*
224	 * Only update the cache from the allocation path, if we ended
225	 * up using the specific cached tag.
226	 */
227done:
228	if (tag == org_last_tag || unlikely(BT_ALLOC_RR(tags))) {
229		last_tag = tag + 1;
230		if (last_tag >= bt->depth - 1)
231			last_tag = 0;
232
233		*tag_cache = last_tag;
234	}
235
236	return tag;
237}
238
239static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
240					 struct blk_mq_hw_ctx *hctx)
241{
242	struct bt_wait_state *bs;
243	int wait_index;
244
245	if (!hctx)
246		return &bt->bs[0];
247
248	wait_index = atomic_read(&hctx->wait_index);
249	bs = &bt->bs[wait_index];
250	bt_index_atomic_inc(&hctx->wait_index);
251	return bs;
252}
253
254static int bt_get(struct blk_mq_alloc_data *data,
255		struct blk_mq_bitmap_tags *bt,
256		struct blk_mq_hw_ctx *hctx,
257		unsigned int *last_tag, struct blk_mq_tags *tags)
258{
259	struct bt_wait_state *bs;
260	DEFINE_WAIT(wait);
261	int tag;
262
263	tag = __bt_get(hctx, bt, last_tag, tags);
264	if (tag != -1)
265		return tag;
266
267	if (!(data->gfp & __GFP_WAIT))
268		return -1;
269
270	bs = bt_wait_ptr(bt, hctx);
271	do {
272		prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
273
274		tag = __bt_get(hctx, bt, last_tag, tags);
275		if (tag != -1)
276			break;
277
278		/*
279		 * We're out of tags on this hardware queue, kick any
280		 * pending IO submits before going to sleep waiting for
281		 * some to complete. Note that hctx can be NULL here for
282		 * reserved tag allocation.
283		 */
284		if (hctx)
285			blk_mq_run_hw_queue(hctx, false);
286
287		/*
288		 * Retry tag allocation after running the hardware queue,
289		 * as running the queue may also have found completions.
290		 */
291		tag = __bt_get(hctx, bt, last_tag, tags);
292		if (tag != -1)
293			break;
294
295		blk_mq_put_ctx(data->ctx);
296
297		io_schedule();
298
299		data->ctx = blk_mq_get_ctx(data->q);
300		data->hctx = data->q->mq_ops->map_queue(data->q,
301				data->ctx->cpu);
302		if (data->reserved) {
303			bt = &data->hctx->tags->breserved_tags;
304		} else {
305			last_tag = &data->ctx->last_tag;
306			hctx = data->hctx;
307			bt = &hctx->tags->bitmap_tags;
308		}
309		finish_wait(&bs->wait, &wait);
310		bs = bt_wait_ptr(bt, hctx);
311	} while (1);
312
313	finish_wait(&bs->wait, &wait);
314	return tag;
315}
316
317static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
318{
319	int tag;
320
321	tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
322			&data->ctx->last_tag, data->hctx->tags);
323	if (tag >= 0)
324		return tag + data->hctx->tags->nr_reserved_tags;
325
326	return BLK_MQ_TAG_FAIL;
327}
328
329static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
330{
331	int tag, zero = 0;
332
333	if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
334		WARN_ON_ONCE(1);
335		return BLK_MQ_TAG_FAIL;
336	}
337
338	tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero,
339		data->hctx->tags);
340	if (tag < 0)
341		return BLK_MQ_TAG_FAIL;
342
343	return tag;
344}
345
346unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
347{
348	if (!data->reserved)
349		return __blk_mq_get_tag(data);
350
351	return __blk_mq_get_reserved_tag(data);
352}
353
354static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
355{
356	int i, wake_index;
357
358	wake_index = atomic_read(&bt->wake_index);
359	for (i = 0; i < BT_WAIT_QUEUES; i++) {
360		struct bt_wait_state *bs = &bt->bs[wake_index];
361
362		if (waitqueue_active(&bs->wait)) {
363			int o = atomic_read(&bt->wake_index);
364			if (wake_index != o)
365				atomic_cmpxchg(&bt->wake_index, o, wake_index);
366
367			return bs;
368		}
369
370		wake_index = bt_index_inc(wake_index);
371	}
372
373	return NULL;
374}
375
376static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
377{
378	const int index = TAG_TO_INDEX(bt, tag);
379	struct bt_wait_state *bs;
380	int wait_cnt;
381
382	clear_bit(TAG_TO_BIT(bt, tag), &bt->map[index].word);
383
384	/* Ensure that the wait list checks occur after clear_bit(). */
385	smp_mb();
386
387	bs = bt_wake_ptr(bt);
388	if (!bs)
389		return;
390
391	wait_cnt = atomic_dec_return(&bs->wait_cnt);
392	if (unlikely(wait_cnt < 0))
393		wait_cnt = atomic_inc_return(&bs->wait_cnt);
394	if (wait_cnt == 0) {
395		atomic_add(bt->wake_cnt, &bs->wait_cnt);
396		bt_index_atomic_inc(&bt->wake_index);
397		wake_up(&bs->wait);
398	}
399}
400
401void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
402		    unsigned int *last_tag)
403{
404	struct blk_mq_tags *tags = hctx->tags;
405
406	if (tag >= tags->nr_reserved_tags) {
407		const int real_tag = tag - tags->nr_reserved_tags;
408
409		BUG_ON(real_tag >= tags->nr_tags);
410		bt_clear_tag(&tags->bitmap_tags, real_tag);
411		if (likely(tags->alloc_policy == BLK_TAG_ALLOC_FIFO))
412			*last_tag = real_tag;
413	} else {
414		BUG_ON(tag >= tags->nr_reserved_tags);
415		bt_clear_tag(&tags->breserved_tags, tag);
416	}
417}
418
419static void bt_for_each(struct blk_mq_hw_ctx *hctx,
420		struct blk_mq_bitmap_tags *bt, unsigned int off,
421		busy_iter_fn *fn, void *data, bool reserved)
422{
423	struct request *rq;
424	int bit, i;
425
426	for (i = 0; i < bt->map_nr; i++) {
427		struct blk_align_bitmap *bm = &bt->map[i];
428
429		for (bit = find_first_bit(&bm->word, bm->depth);
430		     bit < bm->depth;
431		     bit = find_next_bit(&bm->word, bm->depth, bit + 1)) {
432		     	rq = blk_mq_tag_to_rq(hctx->tags, off + bit);
433			if (rq->q == hctx->queue)
434				fn(hctx, rq, data, reserved);
435		}
436
437		off += (1 << bt->bits_per_word);
438	}
439}
440
441void blk_mq_tag_busy_iter(struct blk_mq_hw_ctx *hctx, busy_iter_fn *fn,
442		void *priv)
443{
444	struct blk_mq_tags *tags = hctx->tags;
445
446	if (tags->nr_reserved_tags)
447		bt_for_each(hctx, &tags->breserved_tags, 0, fn, priv, true);
448	bt_for_each(hctx, &tags->bitmap_tags, tags->nr_reserved_tags, fn, priv,
449			false);
450}
451EXPORT_SYMBOL(blk_mq_tag_busy_iter);
452
453static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
454{
455	unsigned int i, used;
456
457	for (i = 0, used = 0; i < bt->map_nr; i++) {
458		struct blk_align_bitmap *bm = &bt->map[i];
459
460		used += bitmap_weight(&bm->word, bm->depth);
461	}
462
463	return bt->depth - used;
464}
465
466static void bt_update_count(struct blk_mq_bitmap_tags *bt,
467			    unsigned int depth)
468{
469	unsigned int tags_per_word = 1U << bt->bits_per_word;
470	unsigned int map_depth = depth;
471
472	if (depth) {
473		int i;
474
475		for (i = 0; i < bt->map_nr; i++) {
476			bt->map[i].depth = min(map_depth, tags_per_word);
477			map_depth -= bt->map[i].depth;
478		}
479	}
480
481	bt->wake_cnt = BT_WAIT_BATCH;
482	if (bt->wake_cnt > depth / BT_WAIT_QUEUES)
483		bt->wake_cnt = max(1U, depth / BT_WAIT_QUEUES);
484
485	bt->depth = depth;
486}
487
488static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
489			int node, bool reserved)
490{
491	int i;
492
493	bt->bits_per_word = ilog2(BITS_PER_LONG);
494
495	/*
496	 * Depth can be zero for reserved tags, that's not a failure
497	 * condition.
498	 */
499	if (depth) {
500		unsigned int nr, tags_per_word;
501
502		tags_per_word = (1 << bt->bits_per_word);
503
504		/*
505		 * If the tag space is small, shrink the number of tags
506		 * per word so we spread over a few cachelines, at least.
507		 * If less than 4 tags, just forget about it, it's not
508		 * going to work optimally anyway.
509		 */
510		if (depth >= 4) {
511			while (tags_per_word * 4 > depth) {
512				bt->bits_per_word--;
513				tags_per_word = (1 << bt->bits_per_word);
514			}
515		}
516
517		nr = ALIGN(depth, tags_per_word) / tags_per_word;
518		bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
519						GFP_KERNEL, node);
520		if (!bt->map)
521			return -ENOMEM;
522
523		bt->map_nr = nr;
524	}
525
526	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
527	if (!bt->bs) {
528		kfree(bt->map);
529		bt->map = NULL;
530		return -ENOMEM;
531	}
532
533	bt_update_count(bt, depth);
534
535	for (i = 0; i < BT_WAIT_QUEUES; i++) {
536		init_waitqueue_head(&bt->bs[i].wait);
537		atomic_set(&bt->bs[i].wait_cnt, bt->wake_cnt);
538	}
539
540	return 0;
541}
542
543static void bt_free(struct blk_mq_bitmap_tags *bt)
544{
545	kfree(bt->map);
546	kfree(bt->bs);
547}
548
549static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
550						   int node, int alloc_policy)
551{
552	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
553
554	tags->alloc_policy = alloc_policy;
555
556	if (bt_alloc(&tags->bitmap_tags, depth, node, false))
557		goto enomem;
558	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
559		goto enomem;
560
561	return tags;
562enomem:
563	bt_free(&tags->bitmap_tags);
564	kfree(tags);
565	return NULL;
566}
567
568struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
569				     unsigned int reserved_tags,
570				     int node, int alloc_policy)
571{
572	struct blk_mq_tags *tags;
573
574	if (total_tags > BLK_MQ_TAG_MAX) {
575		pr_err("blk-mq: tag depth too large\n");
576		return NULL;
577	}
578
579	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
580	if (!tags)
581		return NULL;
582
583	tags->nr_tags = total_tags;
584	tags->nr_reserved_tags = reserved_tags;
585
586	return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
587}
588
589void blk_mq_free_tags(struct blk_mq_tags *tags)
590{
591	bt_free(&tags->bitmap_tags);
592	bt_free(&tags->breserved_tags);
593	kfree(tags);
594}
595
596void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
597{
598	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
599
600	*tag = prandom_u32() % depth;
601}
602
603int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
604{
605	tdepth -= tags->nr_reserved_tags;
606	if (tdepth > tags->nr_tags)
607		return -EINVAL;
608
609	/*
610	 * Don't need (or can't) update reserved tags here, they remain
611	 * static and should never need resizing.
612	 */
613	bt_update_count(&tags->bitmap_tags, tdepth);
614	blk_mq_tag_wakeup_all(tags, false);
615	return 0;
616}
617
618/**
619 * blk_mq_unique_tag() - return a tag that is unique queue-wide
620 * @rq: request for which to compute a unique tag
621 *
622 * The tag field in struct request is unique per hardware queue but not over
623 * all hardware queues. Hence this function that returns a tag with the
624 * hardware context index in the upper bits and the per hardware queue tag in
625 * the lower bits.
626 *
627 * Note: When called for a request that is queued on a non-multiqueue request
628 * queue, the hardware context index is set to zero.
629 */
630u32 blk_mq_unique_tag(struct request *rq)
631{
632	struct request_queue *q = rq->q;
633	struct blk_mq_hw_ctx *hctx;
634	int hwq = 0;
635
636	if (q->mq_ops) {
637		hctx = q->mq_ops->map_queue(q, rq->mq_ctx->cpu);
638		hwq = hctx->queue_num;
639	}
640
641	return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
642		(rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
643}
644EXPORT_SYMBOL(blk_mq_unique_tag);
645
646ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
647{
648	char *orig_page = page;
649	unsigned int free, res;
650
651	if (!tags)
652		return 0;
653
654	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
655			"bits_per_word=%u\n",
656			tags->nr_tags, tags->nr_reserved_tags,
657			tags->bitmap_tags.bits_per_word);
658
659	free = bt_unused_tags(&tags->bitmap_tags);
660	res = bt_unused_tags(&tags->breserved_tags);
661
662	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
663	page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
664
665	return page - orig_page;
666}
667