xref: /linux/block/blk-mq-tag.c (revision 6fdcba32711044c35c0e1b094cbd8f3f0b4472c9)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Tag allocation using scalable bitmaps. Uses active queue tracking to support
4  * fairer distribution of tags between multiple submitters when a shared tag map
5  * is used.
6  *
7  * Copyright (C) 2013-2014 Jens Axboe
8  */
9 #include <linux/kernel.h>
10 #include <linux/module.h>
11 
12 #include <linux/blk-mq.h>
13 #include <linux/delay.h>
14 #include "blk.h"
15 #include "blk-mq.h"
16 #include "blk-mq-tag.h"
17 
18 /*
19  * If a previously inactive queue goes active, bump the active user count.
20  * We need to do this before try to allocate driver tag, then even if fail
21  * to get tag when first time, the other shared-tag users could reserve
22  * budget for it.
23  */
24 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
25 {
26 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
27 	    !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
28 		atomic_inc(&hctx->tags->active_queues);
29 
30 	return true;
31 }
32 
33 /*
34  * Wakeup all potentially sleeping on tags
35  */
36 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
37 {
38 	sbitmap_queue_wake_all(&tags->bitmap_tags);
39 	if (include_reserve)
40 		sbitmap_queue_wake_all(&tags->breserved_tags);
41 }
42 
43 /*
44  * If a previously busy queue goes inactive, potential waiters could now
45  * be allowed to queue. Wake them up and check.
46  */
47 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
48 {
49 	struct blk_mq_tags *tags = hctx->tags;
50 
51 	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
52 		return;
53 
54 	atomic_dec(&tags->active_queues);
55 
56 	blk_mq_tag_wakeup_all(tags, false);
57 }
58 
59 /*
60  * For shared tag users, we track the number of currently active users
61  * and attempt to provide a fair share of the tag depth for each of them.
62  */
63 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
64 				  struct sbitmap_queue *bt)
65 {
66 	unsigned int depth, users;
67 
68 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
69 		return true;
70 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
71 		return true;
72 
73 	/*
74 	 * Don't try dividing an ant
75 	 */
76 	if (bt->sb.depth == 1)
77 		return true;
78 
79 	users = atomic_read(&hctx->tags->active_queues);
80 	if (!users)
81 		return true;
82 
83 	/*
84 	 * Allow at least some tags
85 	 */
86 	depth = max((bt->sb.depth + users - 1) / users, 4U);
87 	return atomic_read(&hctx->nr_active) < depth;
88 }
89 
90 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
91 			    struct sbitmap_queue *bt)
92 {
93 	if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
94 	    !hctx_may_queue(data->hctx, bt))
95 		return -1;
96 	if (data->shallow_depth)
97 		return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
98 	else
99 		return __sbitmap_queue_get(bt);
100 }
101 
102 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
103 {
104 	struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
105 	struct sbitmap_queue *bt;
106 	struct sbq_wait_state *ws;
107 	DEFINE_SBQ_WAIT(wait);
108 	unsigned int tag_offset;
109 	int tag;
110 
111 	if (data->flags & BLK_MQ_REQ_RESERVED) {
112 		if (unlikely(!tags->nr_reserved_tags)) {
113 			WARN_ON_ONCE(1);
114 			return BLK_MQ_TAG_FAIL;
115 		}
116 		bt = &tags->breserved_tags;
117 		tag_offset = 0;
118 	} else {
119 		bt = &tags->bitmap_tags;
120 		tag_offset = tags->nr_reserved_tags;
121 	}
122 
123 	tag = __blk_mq_get_tag(data, bt);
124 	if (tag != -1)
125 		goto found_tag;
126 
127 	if (data->flags & BLK_MQ_REQ_NOWAIT)
128 		return BLK_MQ_TAG_FAIL;
129 
130 	ws = bt_wait_ptr(bt, data->hctx);
131 	do {
132 		struct sbitmap_queue *bt_prev;
133 
134 		/*
135 		 * We're out of tags on this hardware queue, kick any
136 		 * pending IO submits before going to sleep waiting for
137 		 * some to complete.
138 		 */
139 		blk_mq_run_hw_queue(data->hctx, false);
140 
141 		/*
142 		 * Retry tag allocation after running the hardware queue,
143 		 * as running the queue may also have found completions.
144 		 */
145 		tag = __blk_mq_get_tag(data, bt);
146 		if (tag != -1)
147 			break;
148 
149 		sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
150 
151 		tag = __blk_mq_get_tag(data, bt);
152 		if (tag != -1)
153 			break;
154 
155 		bt_prev = bt;
156 		io_schedule();
157 
158 		sbitmap_finish_wait(bt, ws, &wait);
159 
160 		data->ctx = blk_mq_get_ctx(data->q);
161 		data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
162 						data->ctx);
163 		tags = blk_mq_tags_from_data(data);
164 		if (data->flags & BLK_MQ_REQ_RESERVED)
165 			bt = &tags->breserved_tags;
166 		else
167 			bt = &tags->bitmap_tags;
168 
169 		/*
170 		 * If destination hw queue is changed, fake wake up on
171 		 * previous queue for compensating the wake up miss, so
172 		 * other allocations on previous queue won't be starved.
173 		 */
174 		if (bt != bt_prev)
175 			sbitmap_queue_wake_up(bt_prev);
176 
177 		ws = bt_wait_ptr(bt, data->hctx);
178 	} while (1);
179 
180 	sbitmap_finish_wait(bt, ws, &wait);
181 
182 found_tag:
183 	return tag + tag_offset;
184 }
185 
186 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
187 		    struct blk_mq_ctx *ctx, unsigned int tag)
188 {
189 	if (!blk_mq_tag_is_reserved(tags, tag)) {
190 		const int real_tag = tag - tags->nr_reserved_tags;
191 
192 		BUG_ON(real_tag >= tags->nr_tags);
193 		sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
194 	} else {
195 		BUG_ON(tag >= tags->nr_reserved_tags);
196 		sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
197 	}
198 }
199 
200 struct bt_iter_data {
201 	struct blk_mq_hw_ctx *hctx;
202 	busy_iter_fn *fn;
203 	void *data;
204 	bool reserved;
205 };
206 
207 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
208 {
209 	struct bt_iter_data *iter_data = data;
210 	struct blk_mq_hw_ctx *hctx = iter_data->hctx;
211 	struct blk_mq_tags *tags = hctx->tags;
212 	bool reserved = iter_data->reserved;
213 	struct request *rq;
214 
215 	if (!reserved)
216 		bitnr += tags->nr_reserved_tags;
217 	rq = tags->rqs[bitnr];
218 
219 	/*
220 	 * We can hit rq == NULL here, because the tagging functions
221 	 * test and set the bit before assigning ->rqs[].
222 	 */
223 	if (rq && rq->q == hctx->queue)
224 		return iter_data->fn(hctx, rq, iter_data->data, reserved);
225 	return true;
226 }
227 
228 /**
229  * bt_for_each - iterate over the requests associated with a hardware queue
230  * @hctx:	Hardware queue to examine.
231  * @bt:		sbitmap to examine. This is either the breserved_tags member
232  *		or the bitmap_tags member of struct blk_mq_tags.
233  * @fn:		Pointer to the function that will be called for each request
234  *		associated with @hctx that has been assigned a driver tag.
235  *		@fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
236  *		where rq is a pointer to a request. Return true to continue
237  *		iterating tags, false to stop.
238  * @data:	Will be passed as third argument to @fn.
239  * @reserved:	Indicates whether @bt is the breserved_tags member or the
240  *		bitmap_tags member of struct blk_mq_tags.
241  */
242 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
243 			busy_iter_fn *fn, void *data, bool reserved)
244 {
245 	struct bt_iter_data iter_data = {
246 		.hctx = hctx,
247 		.fn = fn,
248 		.data = data,
249 		.reserved = reserved,
250 	};
251 
252 	sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
253 }
254 
255 struct bt_tags_iter_data {
256 	struct blk_mq_tags *tags;
257 	busy_tag_iter_fn *fn;
258 	void *data;
259 	bool reserved;
260 };
261 
262 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
263 {
264 	struct bt_tags_iter_data *iter_data = data;
265 	struct blk_mq_tags *tags = iter_data->tags;
266 	bool reserved = iter_data->reserved;
267 	struct request *rq;
268 
269 	if (!reserved)
270 		bitnr += tags->nr_reserved_tags;
271 
272 	/*
273 	 * We can hit rq == NULL here, because the tagging functions
274 	 * test and set the bit before assining ->rqs[].
275 	 */
276 	rq = tags->rqs[bitnr];
277 	if (rq && blk_mq_request_started(rq))
278 		return iter_data->fn(rq, iter_data->data, reserved);
279 
280 	return true;
281 }
282 
283 /**
284  * bt_tags_for_each - iterate over the requests in a tag map
285  * @tags:	Tag map to iterate over.
286  * @bt:		sbitmap to examine. This is either the breserved_tags member
287  *		or the bitmap_tags member of struct blk_mq_tags.
288  * @fn:		Pointer to the function that will be called for each started
289  *		request. @fn will be called as follows: @fn(rq, @data,
290  *		@reserved) where rq is a pointer to a request. Return true
291  *		to continue iterating tags, false to stop.
292  * @data:	Will be passed as second argument to @fn.
293  * @reserved:	Indicates whether @bt is the breserved_tags member or the
294  *		bitmap_tags member of struct blk_mq_tags.
295  */
296 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
297 			     busy_tag_iter_fn *fn, void *data, bool reserved)
298 {
299 	struct bt_tags_iter_data iter_data = {
300 		.tags = tags,
301 		.fn = fn,
302 		.data = data,
303 		.reserved = reserved,
304 	};
305 
306 	if (tags->rqs)
307 		sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
308 }
309 
310 /**
311  * blk_mq_all_tag_busy_iter - iterate over all started requests in a tag map
312  * @tags:	Tag map to iterate over.
313  * @fn:		Pointer to the function that will be called for each started
314  *		request. @fn will be called as follows: @fn(rq, @priv,
315  *		reserved) where rq is a pointer to a request. 'reserved'
316  *		indicates whether or not @rq is a reserved request. Return
317  *		true to continue iterating tags, false to stop.
318  * @priv:	Will be passed as second argument to @fn.
319  */
320 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
321 		busy_tag_iter_fn *fn, void *priv)
322 {
323 	if (tags->nr_reserved_tags)
324 		bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
325 	bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
326 }
327 
328 /**
329  * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
330  * @tagset:	Tag set to iterate over.
331  * @fn:		Pointer to the function that will be called for each started
332  *		request. @fn will be called as follows: @fn(rq, @priv,
333  *		reserved) where rq is a pointer to a request. 'reserved'
334  *		indicates whether or not @rq is a reserved request. Return
335  *		true to continue iterating tags, false to stop.
336  * @priv:	Will be passed as second argument to @fn.
337  */
338 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
339 		busy_tag_iter_fn *fn, void *priv)
340 {
341 	int i;
342 
343 	for (i = 0; i < tagset->nr_hw_queues; i++) {
344 		if (tagset->tags && tagset->tags[i])
345 			blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
346 	}
347 }
348 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
349 
350 static bool blk_mq_tagset_count_completed_rqs(struct request *rq,
351 		void *data, bool reserved)
352 {
353 	unsigned *count = data;
354 
355 	if (blk_mq_request_completed(rq))
356 		(*count)++;
357 	return true;
358 }
359 
360 /**
361  * blk_mq_tagset_wait_completed_request - wait until all completed req's
362  * complete funtion is run
363  * @tagset:	Tag set to drain completed request
364  *
365  * Note: This function has to be run after all IO queues are shutdown
366  */
367 void blk_mq_tagset_wait_completed_request(struct blk_mq_tag_set *tagset)
368 {
369 	while (true) {
370 		unsigned count = 0;
371 
372 		blk_mq_tagset_busy_iter(tagset,
373 				blk_mq_tagset_count_completed_rqs, &count);
374 		if (!count)
375 			break;
376 		msleep(5);
377 	}
378 }
379 EXPORT_SYMBOL(blk_mq_tagset_wait_completed_request);
380 
381 /**
382  * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
383  * @q:		Request queue to examine.
384  * @fn:		Pointer to the function that will be called for each request
385  *		on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
386  *		reserved) where rq is a pointer to a request and hctx points
387  *		to the hardware queue associated with the request. 'reserved'
388  *		indicates whether or not @rq is a reserved request.
389  * @priv:	Will be passed as third argument to @fn.
390  *
391  * Note: if @q->tag_set is shared with other request queues then @fn will be
392  * called for all requests on all queues that share that tag set and not only
393  * for requests associated with @q.
394  */
395 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
396 		void *priv)
397 {
398 	struct blk_mq_hw_ctx *hctx;
399 	int i;
400 
401 	/*
402 	 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
403 	 * while the queue is frozen. So we can use q_usage_counter to avoid
404 	 * racing with it. __blk_mq_update_nr_hw_queues() uses
405 	 * synchronize_rcu() to ensure this function left the critical section
406 	 * below.
407 	 */
408 	if (!percpu_ref_tryget(&q->q_usage_counter))
409 		return;
410 
411 	queue_for_each_hw_ctx(q, hctx, i) {
412 		struct blk_mq_tags *tags = hctx->tags;
413 
414 		/*
415 		 * If no software queues are currently mapped to this
416 		 * hardware queue, there's nothing to check
417 		 */
418 		if (!blk_mq_hw_queue_mapped(hctx))
419 			continue;
420 
421 		if (tags->nr_reserved_tags)
422 			bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
423 		bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
424 	}
425 	blk_queue_exit(q);
426 }
427 
428 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
429 		    bool round_robin, int node)
430 {
431 	return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
432 				       node);
433 }
434 
435 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
436 						   int node, int alloc_policy)
437 {
438 	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
439 	bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
440 
441 	if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
442 		goto free_tags;
443 	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
444 		     node))
445 		goto free_bitmap_tags;
446 
447 	return tags;
448 free_bitmap_tags:
449 	sbitmap_queue_free(&tags->bitmap_tags);
450 free_tags:
451 	kfree(tags);
452 	return NULL;
453 }
454 
455 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
456 				     unsigned int reserved_tags,
457 				     int node, int alloc_policy)
458 {
459 	struct blk_mq_tags *tags;
460 
461 	if (total_tags > BLK_MQ_TAG_MAX) {
462 		pr_err("blk-mq: tag depth too large\n");
463 		return NULL;
464 	}
465 
466 	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
467 	if (!tags)
468 		return NULL;
469 
470 	tags->nr_tags = total_tags;
471 	tags->nr_reserved_tags = reserved_tags;
472 
473 	return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
474 }
475 
476 void blk_mq_free_tags(struct blk_mq_tags *tags)
477 {
478 	sbitmap_queue_free(&tags->bitmap_tags);
479 	sbitmap_queue_free(&tags->breserved_tags);
480 	kfree(tags);
481 }
482 
483 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
484 			    struct blk_mq_tags **tagsptr, unsigned int tdepth,
485 			    bool can_grow)
486 {
487 	struct blk_mq_tags *tags = *tagsptr;
488 
489 	if (tdepth <= tags->nr_reserved_tags)
490 		return -EINVAL;
491 
492 	/*
493 	 * If we are allowed to grow beyond the original size, allocate
494 	 * a new set of tags before freeing the old one.
495 	 */
496 	if (tdepth > tags->nr_tags) {
497 		struct blk_mq_tag_set *set = hctx->queue->tag_set;
498 		struct blk_mq_tags *new;
499 		bool ret;
500 
501 		if (!can_grow)
502 			return -EINVAL;
503 
504 		/*
505 		 * We need some sort of upper limit, set it high enough that
506 		 * no valid use cases should require more.
507 		 */
508 		if (tdepth > 16 * BLKDEV_MAX_RQ)
509 			return -EINVAL;
510 
511 		new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
512 				tags->nr_reserved_tags);
513 		if (!new)
514 			return -ENOMEM;
515 		ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
516 		if (ret) {
517 			blk_mq_free_rq_map(new);
518 			return -ENOMEM;
519 		}
520 
521 		blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
522 		blk_mq_free_rq_map(*tagsptr);
523 		*tagsptr = new;
524 	} else {
525 		/*
526 		 * Don't need (or can't) update reserved tags here, they
527 		 * remain static and should never need resizing.
528 		 */
529 		sbitmap_queue_resize(&tags->bitmap_tags,
530 				tdepth - tags->nr_reserved_tags);
531 	}
532 
533 	return 0;
534 }
535 
536 /**
537  * blk_mq_unique_tag() - return a tag that is unique queue-wide
538  * @rq: request for which to compute a unique tag
539  *
540  * The tag field in struct request is unique per hardware queue but not over
541  * all hardware queues. Hence this function that returns a tag with the
542  * hardware context index in the upper bits and the per hardware queue tag in
543  * the lower bits.
544  *
545  * Note: When called for a request that is queued on a non-multiqueue request
546  * queue, the hardware context index is set to zero.
547  */
548 u32 blk_mq_unique_tag(struct request *rq)
549 {
550 	return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
551 		(rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
552 }
553 EXPORT_SYMBOL(blk_mq_unique_tag);
554