xref: /openbmc/linux/block/blk-mq-tag.c (revision 79a5a18a)
1 /*
2  * Tag allocation using scalable bitmaps. Uses active queue tracking to support
3  * fairer distribution of tags between multiple submitters when a shared tag map
4  * is used.
5  *
6  * Copyright (C) 2013-2014 Jens Axboe
7  */
8 #include <linux/kernel.h>
9 #include <linux/module.h>
10 
11 #include <linux/blk-mq.h>
12 #include "blk.h"
13 #include "blk-mq.h"
14 #include "blk-mq-tag.h"
15 
16 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
17 {
18 	if (!tags)
19 		return true;
20 
21 	return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
22 }
23 
24 /*
25  * If a previously inactive queue goes active, bump the active user count.
26  * We need to do this before try to allocate driver tag, then even if fail
27  * to get tag when first time, the other shared-tag users could reserve
28  * budget for it.
29  */
30 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
31 {
32 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
33 	    !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
34 		atomic_inc(&hctx->tags->active_queues);
35 
36 	return true;
37 }
38 
39 /*
40  * Wakeup all potentially sleeping on tags
41  */
42 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
43 {
44 	sbitmap_queue_wake_all(&tags->bitmap_tags);
45 	if (include_reserve)
46 		sbitmap_queue_wake_all(&tags->breserved_tags);
47 }
48 
49 /*
50  * If a previously busy queue goes inactive, potential waiters could now
51  * be allowed to queue. Wake them up and check.
52  */
53 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
54 {
55 	struct blk_mq_tags *tags = hctx->tags;
56 
57 	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
58 		return;
59 
60 	atomic_dec(&tags->active_queues);
61 
62 	blk_mq_tag_wakeup_all(tags, false);
63 }
64 
65 /*
66  * For shared tag users, we track the number of currently active users
67  * and attempt to provide a fair share of the tag depth for each of them.
68  */
69 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
70 				  struct sbitmap_queue *bt)
71 {
72 	unsigned int depth, users;
73 
74 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
75 		return true;
76 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
77 		return true;
78 
79 	/*
80 	 * Don't try dividing an ant
81 	 */
82 	if (bt->sb.depth == 1)
83 		return true;
84 
85 	users = atomic_read(&hctx->tags->active_queues);
86 	if (!users)
87 		return true;
88 
89 	/*
90 	 * Allow at least some tags
91 	 */
92 	depth = max((bt->sb.depth + users - 1) / users, 4U);
93 	return atomic_read(&hctx->nr_active) < depth;
94 }
95 
96 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
97 			    struct sbitmap_queue *bt)
98 {
99 	if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
100 	    !hctx_may_queue(data->hctx, bt))
101 		return -1;
102 	if (data->shallow_depth)
103 		return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
104 	else
105 		return __sbitmap_queue_get(bt);
106 }
107 
108 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
109 {
110 	struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
111 	struct sbitmap_queue *bt;
112 	struct sbq_wait_state *ws;
113 	DEFINE_WAIT(wait);
114 	unsigned int tag_offset;
115 	bool drop_ctx;
116 	int tag;
117 
118 	if (data->flags & BLK_MQ_REQ_RESERVED) {
119 		if (unlikely(!tags->nr_reserved_tags)) {
120 			WARN_ON_ONCE(1);
121 			return BLK_MQ_TAG_FAIL;
122 		}
123 		bt = &tags->breserved_tags;
124 		tag_offset = 0;
125 	} else {
126 		bt = &tags->bitmap_tags;
127 		tag_offset = tags->nr_reserved_tags;
128 	}
129 
130 	tag = __blk_mq_get_tag(data, bt);
131 	if (tag != -1)
132 		goto found_tag;
133 
134 	if (data->flags & BLK_MQ_REQ_NOWAIT)
135 		return BLK_MQ_TAG_FAIL;
136 
137 	ws = bt_wait_ptr(bt, data->hctx);
138 	drop_ctx = data->ctx == NULL;
139 	do {
140 		struct sbitmap_queue *bt_prev;
141 
142 		/*
143 		 * We're out of tags on this hardware queue, kick any
144 		 * pending IO submits before going to sleep waiting for
145 		 * some to complete.
146 		 */
147 		blk_mq_run_hw_queue(data->hctx, false);
148 
149 		/*
150 		 * Retry tag allocation after running the hardware queue,
151 		 * as running the queue may also have found completions.
152 		 */
153 		tag = __blk_mq_get_tag(data, bt);
154 		if (tag != -1)
155 			break;
156 
157 		prepare_to_wait_exclusive(&ws->wait, &wait,
158 						TASK_UNINTERRUPTIBLE);
159 
160 		tag = __blk_mq_get_tag(data, bt);
161 		if (tag != -1)
162 			break;
163 
164 		if (data->ctx)
165 			blk_mq_put_ctx(data->ctx);
166 
167 		bt_prev = bt;
168 		io_schedule();
169 
170 		data->ctx = blk_mq_get_ctx(data->q);
171 		data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
172 		tags = blk_mq_tags_from_data(data);
173 		if (data->flags & BLK_MQ_REQ_RESERVED)
174 			bt = &tags->breserved_tags;
175 		else
176 			bt = &tags->bitmap_tags;
177 
178 		finish_wait(&ws->wait, &wait);
179 
180 		/*
181 		 * If destination hw queue is changed, fake wake up on
182 		 * previous queue for compensating the wake up miss, so
183 		 * other allocations on previous queue won't be starved.
184 		 */
185 		if (bt != bt_prev)
186 			sbitmap_queue_wake_up(bt_prev);
187 
188 		ws = bt_wait_ptr(bt, data->hctx);
189 	} while (1);
190 
191 	if (drop_ctx && data->ctx)
192 		blk_mq_put_ctx(data->ctx);
193 
194 	finish_wait(&ws->wait, &wait);
195 
196 found_tag:
197 	return tag + tag_offset;
198 }
199 
200 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
201 		    struct blk_mq_ctx *ctx, unsigned int tag)
202 {
203 	if (!blk_mq_tag_is_reserved(tags, tag)) {
204 		const int real_tag = tag - tags->nr_reserved_tags;
205 
206 		BUG_ON(real_tag >= tags->nr_tags);
207 		sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
208 	} else {
209 		BUG_ON(tag >= tags->nr_reserved_tags);
210 		sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
211 	}
212 }
213 
214 struct bt_iter_data {
215 	struct blk_mq_hw_ctx *hctx;
216 	busy_iter_fn *fn;
217 	void *data;
218 	bool reserved;
219 };
220 
221 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
222 {
223 	struct bt_iter_data *iter_data = data;
224 	struct blk_mq_hw_ctx *hctx = iter_data->hctx;
225 	struct blk_mq_tags *tags = hctx->tags;
226 	bool reserved = iter_data->reserved;
227 	struct request *rq;
228 
229 	if (!reserved)
230 		bitnr += tags->nr_reserved_tags;
231 	rq = tags->rqs[bitnr];
232 
233 	/*
234 	 * We can hit rq == NULL here, because the tagging functions
235 	 * test and set the bit before assigning ->rqs[].
236 	 */
237 	if (rq && rq->q == hctx->queue)
238 		iter_data->fn(hctx, rq, iter_data->data, reserved);
239 	return true;
240 }
241 
242 /**
243  * bt_for_each - iterate over the requests associated with a hardware queue
244  * @hctx:	Hardware queue to examine.
245  * @bt:		sbitmap to examine. This is either the breserved_tags member
246  *		or the bitmap_tags member of struct blk_mq_tags.
247  * @fn:		Pointer to the function that will be called for each request
248  *		associated with @hctx that has been assigned a driver tag.
249  *		@fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
250  *		where rq is a pointer to a request.
251  * @data:	Will be passed as third argument to @fn.
252  * @reserved:	Indicates whether @bt is the breserved_tags member or the
253  *		bitmap_tags member of struct blk_mq_tags.
254  */
255 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
256 			busy_iter_fn *fn, void *data, bool reserved)
257 {
258 	struct bt_iter_data iter_data = {
259 		.hctx = hctx,
260 		.fn = fn,
261 		.data = data,
262 		.reserved = reserved,
263 	};
264 
265 	sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
266 }
267 
268 struct bt_tags_iter_data {
269 	struct blk_mq_tags *tags;
270 	busy_tag_iter_fn *fn;
271 	void *data;
272 	bool reserved;
273 };
274 
275 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
276 {
277 	struct bt_tags_iter_data *iter_data = data;
278 	struct blk_mq_tags *tags = iter_data->tags;
279 	bool reserved = iter_data->reserved;
280 	struct request *rq;
281 
282 	if (!reserved)
283 		bitnr += tags->nr_reserved_tags;
284 
285 	/*
286 	 * We can hit rq == NULL here, because the tagging functions
287 	 * test and set the bit before assining ->rqs[].
288 	 */
289 	rq = tags->rqs[bitnr];
290 	if (rq && blk_mq_request_started(rq))
291 		iter_data->fn(rq, iter_data->data, reserved);
292 
293 	return true;
294 }
295 
296 /**
297  * bt_tags_for_each - iterate over the requests in a tag map
298  * @tags:	Tag map to iterate over.
299  * @bt:		sbitmap to examine. This is either the breserved_tags member
300  *		or the bitmap_tags member of struct blk_mq_tags.
301  * @fn:		Pointer to the function that will be called for each started
302  *		request. @fn will be called as follows: @fn(rq, @data,
303  *		@reserved) where rq is a pointer to a request.
304  * @data:	Will be passed as second argument to @fn.
305  * @reserved:	Indicates whether @bt is the breserved_tags member or the
306  *		bitmap_tags member of struct blk_mq_tags.
307  */
308 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
309 			     busy_tag_iter_fn *fn, void *data, bool reserved)
310 {
311 	struct bt_tags_iter_data iter_data = {
312 		.tags = tags,
313 		.fn = fn,
314 		.data = data,
315 		.reserved = reserved,
316 	};
317 
318 	if (tags->rqs)
319 		sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
320 }
321 
322 /**
323  * blk_mq_all_tag_busy_iter - iterate over all started requests in a tag map
324  * @tags:	Tag map to iterate over.
325  * @fn:		Pointer to the function that will be called for each started
326  *		request. @fn will be called as follows: @fn(rq, @priv,
327  *		reserved) where rq is a pointer to a request. 'reserved'
328  *		indicates whether or not @rq is a reserved request.
329  * @priv:	Will be passed as second argument to @fn.
330  */
331 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
332 		busy_tag_iter_fn *fn, void *priv)
333 {
334 	if (tags->nr_reserved_tags)
335 		bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
336 	bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
337 }
338 
339 /**
340  * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
341  * @tagset:	Tag set to iterate over.
342  * @fn:		Pointer to the function that will be called for each started
343  *		request. @fn will be called as follows: @fn(rq, @priv,
344  *		reserved) where rq is a pointer to a request. 'reserved'
345  *		indicates whether or not @rq is a reserved request.
346  * @priv:	Will be passed as second argument to @fn.
347  */
348 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
349 		busy_tag_iter_fn *fn, void *priv)
350 {
351 	int i;
352 
353 	for (i = 0; i < tagset->nr_hw_queues; i++) {
354 		if (tagset->tags && tagset->tags[i])
355 			blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
356 	}
357 }
358 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
359 
360 /**
361  * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
362  * @q:		Request queue to examine.
363  * @fn:		Pointer to the function that will be called for each request
364  *		on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
365  *		reserved) where rq is a pointer to a request and hctx points
366  *		to the hardware queue associated with the request. 'reserved'
367  *		indicates whether or not @rq is a reserved request.
368  * @priv:	Will be passed as third argument to @fn.
369  *
370  * Note: if @q->tag_set is shared with other request queues then @fn will be
371  * called for all requests on all queues that share that tag set and not only
372  * for requests associated with @q.
373  */
374 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
375 		void *priv)
376 {
377 	struct blk_mq_hw_ctx *hctx;
378 	int i;
379 
380 	/*
381 	 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
382 	 * while the queue is frozen. So we can use q_usage_counter to avoid
383 	 * racing with it. __blk_mq_update_nr_hw_queues() uses
384 	 * synchronize_rcu() to ensure this function left the critical section
385 	 * below.
386 	 */
387 	if (!percpu_ref_tryget(&q->q_usage_counter))
388 		return;
389 
390 	queue_for_each_hw_ctx(q, hctx, i) {
391 		struct blk_mq_tags *tags = hctx->tags;
392 
393 		/*
394 		 * If no software queues are currently mapped to this
395 		 * hardware queue, there's nothing to check
396 		 */
397 		if (!blk_mq_hw_queue_mapped(hctx))
398 			continue;
399 
400 		if (tags->nr_reserved_tags)
401 			bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
402 		bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
403 	}
404 	blk_queue_exit(q);
405 }
406 
407 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
408 		    bool round_robin, int node)
409 {
410 	return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
411 				       node);
412 }
413 
414 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
415 						   int node, int alloc_policy)
416 {
417 	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
418 	bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
419 
420 	if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
421 		goto free_tags;
422 	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
423 		     node))
424 		goto free_bitmap_tags;
425 
426 	return tags;
427 free_bitmap_tags:
428 	sbitmap_queue_free(&tags->bitmap_tags);
429 free_tags:
430 	kfree(tags);
431 	return NULL;
432 }
433 
434 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
435 				     unsigned int reserved_tags,
436 				     int node, int alloc_policy)
437 {
438 	struct blk_mq_tags *tags;
439 
440 	if (total_tags > BLK_MQ_TAG_MAX) {
441 		pr_err("blk-mq: tag depth too large\n");
442 		return NULL;
443 	}
444 
445 	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
446 	if (!tags)
447 		return NULL;
448 
449 	tags->nr_tags = total_tags;
450 	tags->nr_reserved_tags = reserved_tags;
451 
452 	return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
453 }
454 
455 void blk_mq_free_tags(struct blk_mq_tags *tags)
456 {
457 	sbitmap_queue_free(&tags->bitmap_tags);
458 	sbitmap_queue_free(&tags->breserved_tags);
459 	kfree(tags);
460 }
461 
462 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
463 			    struct blk_mq_tags **tagsptr, unsigned int tdepth,
464 			    bool can_grow)
465 {
466 	struct blk_mq_tags *tags = *tagsptr;
467 
468 	if (tdepth <= tags->nr_reserved_tags)
469 		return -EINVAL;
470 
471 	/*
472 	 * If we are allowed to grow beyond the original size, allocate
473 	 * a new set of tags before freeing the old one.
474 	 */
475 	if (tdepth > tags->nr_tags) {
476 		struct blk_mq_tag_set *set = hctx->queue->tag_set;
477 		struct blk_mq_tags *new;
478 		bool ret;
479 
480 		if (!can_grow)
481 			return -EINVAL;
482 
483 		/*
484 		 * We need some sort of upper limit, set it high enough that
485 		 * no valid use cases should require more.
486 		 */
487 		if (tdepth > 16 * BLKDEV_MAX_RQ)
488 			return -EINVAL;
489 
490 		new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
491 				tags->nr_reserved_tags);
492 		if (!new)
493 			return -ENOMEM;
494 		ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
495 		if (ret) {
496 			blk_mq_free_rq_map(new);
497 			return -ENOMEM;
498 		}
499 
500 		blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
501 		blk_mq_free_rq_map(*tagsptr);
502 		*tagsptr = new;
503 	} else {
504 		/*
505 		 * Don't need (or can't) update reserved tags here, they
506 		 * remain static and should never need resizing.
507 		 */
508 		sbitmap_queue_resize(&tags->bitmap_tags,
509 				tdepth - tags->nr_reserved_tags);
510 	}
511 
512 	return 0;
513 }
514 
515 /**
516  * blk_mq_unique_tag() - return a tag that is unique queue-wide
517  * @rq: request for which to compute a unique tag
518  *
519  * The tag field in struct request is unique per hardware queue but not over
520  * all hardware queues. Hence this function that returns a tag with the
521  * hardware context index in the upper bits and the per hardware queue tag in
522  * the lower bits.
523  *
524  * Note: When called for a request that is queued on a non-multiqueue request
525  * queue, the hardware context index is set to zero.
526  */
527 u32 blk_mq_unique_tag(struct request *rq)
528 {
529 	struct request_queue *q = rq->q;
530 	struct blk_mq_hw_ctx *hctx;
531 	int hwq = 0;
532 
533 	if (q->mq_ops) {
534 		hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
535 		hwq = hctx->queue_num;
536 	}
537 
538 	return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
539 		(rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
540 }
541 EXPORT_SYMBOL(blk_mq_unique_tag);
542