xref: /openbmc/linux/block/blk-mq-tag.c (revision b830f94f)
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 "blk.h"
14 #include "blk-mq.h"
15 #include "blk-mq-tag.h"
16 
17 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
18 {
19 	if (!tags)
20 		return true;
21 
22 	return sbitmap_any_bit_clear(&tags->bitmap_tags.sb);
23 }
24 
25 /*
26  * If a previously inactive queue goes active, bump the active user count.
27  * We need to do this before try to allocate driver tag, then even if fail
28  * to get tag when first time, the other shared-tag users could reserve
29  * budget for it.
30  */
31 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
32 {
33 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
34 	    !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
35 		atomic_inc(&hctx->tags->active_queues);
36 
37 	return true;
38 }
39 
40 /*
41  * Wakeup all potentially sleeping on tags
42  */
43 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
44 {
45 	sbitmap_queue_wake_all(&tags->bitmap_tags);
46 	if (include_reserve)
47 		sbitmap_queue_wake_all(&tags->breserved_tags);
48 }
49 
50 /*
51  * If a previously busy queue goes inactive, potential waiters could now
52  * be allowed to queue. Wake them up and check.
53  */
54 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
55 {
56 	struct blk_mq_tags *tags = hctx->tags;
57 
58 	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
59 		return;
60 
61 	atomic_dec(&tags->active_queues);
62 
63 	blk_mq_tag_wakeup_all(tags, false);
64 }
65 
66 /*
67  * For shared tag users, we track the number of currently active users
68  * and attempt to provide a fair share of the tag depth for each of them.
69  */
70 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
71 				  struct sbitmap_queue *bt)
72 {
73 	unsigned int depth, users;
74 
75 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
76 		return true;
77 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
78 		return true;
79 
80 	/*
81 	 * Don't try dividing an ant
82 	 */
83 	if (bt->sb.depth == 1)
84 		return true;
85 
86 	users = atomic_read(&hctx->tags->active_queues);
87 	if (!users)
88 		return true;
89 
90 	/*
91 	 * Allow at least some tags
92 	 */
93 	depth = max((bt->sb.depth + users - 1) / users, 4U);
94 	return atomic_read(&hctx->nr_active) < depth;
95 }
96 
97 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
98 			    struct sbitmap_queue *bt)
99 {
100 	if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
101 	    !hctx_may_queue(data->hctx, bt))
102 		return -1;
103 	if (data->shallow_depth)
104 		return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
105 	else
106 		return __sbitmap_queue_get(bt);
107 }
108 
109 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
110 {
111 	struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
112 	struct sbitmap_queue *bt;
113 	struct sbq_wait_state *ws;
114 	DEFINE_SBQ_WAIT(wait);
115 	unsigned int tag_offset;
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 	do {
139 		struct sbitmap_queue *bt_prev;
140 
141 		/*
142 		 * We're out of tags on this hardware queue, kick any
143 		 * pending IO submits before going to sleep waiting for
144 		 * some to complete.
145 		 */
146 		blk_mq_run_hw_queue(data->hctx, false);
147 
148 		/*
149 		 * Retry tag allocation after running the hardware queue,
150 		 * as running the queue may also have found completions.
151 		 */
152 		tag = __blk_mq_get_tag(data, bt);
153 		if (tag != -1)
154 			break;
155 
156 		sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE);
157 
158 		tag = __blk_mq_get_tag(data, bt);
159 		if (tag != -1)
160 			break;
161 
162 		bt_prev = bt;
163 		io_schedule();
164 
165 		sbitmap_finish_wait(bt, ws, &wait);
166 
167 		data->ctx = blk_mq_get_ctx(data->q);
168 		data->hctx = blk_mq_map_queue(data->q, data->cmd_flags,
169 						data->ctx);
170 		tags = blk_mq_tags_from_data(data);
171 		if (data->flags & BLK_MQ_REQ_RESERVED)
172 			bt = &tags->breserved_tags;
173 		else
174 			bt = &tags->bitmap_tags;
175 
176 		/*
177 		 * If destination hw queue is changed, fake wake up on
178 		 * previous queue for compensating the wake up miss, so
179 		 * other allocations on previous queue won't be starved.
180 		 */
181 		if (bt != bt_prev)
182 			sbitmap_queue_wake_up(bt_prev);
183 
184 		ws = bt_wait_ptr(bt, data->hctx);
185 	} while (1);
186 
187 	sbitmap_finish_wait(bt, ws, &wait);
188 
189 found_tag:
190 	return tag + tag_offset;
191 }
192 
193 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
194 		    struct blk_mq_ctx *ctx, unsigned int tag)
195 {
196 	if (!blk_mq_tag_is_reserved(tags, tag)) {
197 		const int real_tag = tag - tags->nr_reserved_tags;
198 
199 		BUG_ON(real_tag >= tags->nr_tags);
200 		sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
201 	} else {
202 		BUG_ON(tag >= tags->nr_reserved_tags);
203 		sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
204 	}
205 }
206 
207 struct bt_iter_data {
208 	struct blk_mq_hw_ctx *hctx;
209 	busy_iter_fn *fn;
210 	void *data;
211 	bool reserved;
212 };
213 
214 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
215 {
216 	struct bt_iter_data *iter_data = data;
217 	struct blk_mq_hw_ctx *hctx = iter_data->hctx;
218 	struct blk_mq_tags *tags = hctx->tags;
219 	bool reserved = iter_data->reserved;
220 	struct request *rq;
221 
222 	if (!reserved)
223 		bitnr += tags->nr_reserved_tags;
224 	rq = tags->rqs[bitnr];
225 
226 	/*
227 	 * We can hit rq == NULL here, because the tagging functions
228 	 * test and set the bit before assigning ->rqs[].
229 	 */
230 	if (rq && rq->q == hctx->queue)
231 		return iter_data->fn(hctx, rq, iter_data->data, reserved);
232 	return true;
233 }
234 
235 /**
236  * bt_for_each - iterate over the requests associated with a hardware queue
237  * @hctx:	Hardware queue to examine.
238  * @bt:		sbitmap to examine. This is either the breserved_tags member
239  *		or the bitmap_tags member of struct blk_mq_tags.
240  * @fn:		Pointer to the function that will be called for each request
241  *		associated with @hctx that has been assigned a driver tag.
242  *		@fn will be called as follows: @fn(@hctx, rq, @data, @reserved)
243  *		where rq is a pointer to a request. Return true to continue
244  *		iterating tags, false to stop.
245  * @data:	Will be passed as third argument to @fn.
246  * @reserved:	Indicates whether @bt is the breserved_tags member or the
247  *		bitmap_tags member of struct blk_mq_tags.
248  */
249 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
250 			busy_iter_fn *fn, void *data, bool reserved)
251 {
252 	struct bt_iter_data iter_data = {
253 		.hctx = hctx,
254 		.fn = fn,
255 		.data = data,
256 		.reserved = reserved,
257 	};
258 
259 	sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
260 }
261 
262 struct bt_tags_iter_data {
263 	struct blk_mq_tags *tags;
264 	busy_tag_iter_fn *fn;
265 	void *data;
266 	bool reserved;
267 };
268 
269 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
270 {
271 	struct bt_tags_iter_data *iter_data = data;
272 	struct blk_mq_tags *tags = iter_data->tags;
273 	bool reserved = iter_data->reserved;
274 	struct request *rq;
275 
276 	if (!reserved)
277 		bitnr += tags->nr_reserved_tags;
278 
279 	/*
280 	 * We can hit rq == NULL here, because the tagging functions
281 	 * test and set the bit before assining ->rqs[].
282 	 */
283 	rq = tags->rqs[bitnr];
284 	if (rq && blk_mq_request_started(rq))
285 		return iter_data->fn(rq, iter_data->data, reserved);
286 
287 	return true;
288 }
289 
290 /**
291  * bt_tags_for_each - iterate over the requests in a tag map
292  * @tags:	Tag map to iterate over.
293  * @bt:		sbitmap to examine. This is either the breserved_tags member
294  *		or the bitmap_tags member of struct blk_mq_tags.
295  * @fn:		Pointer to the function that will be called for each started
296  *		request. @fn will be called as follows: @fn(rq, @data,
297  *		@reserved) where rq is a pointer to a request. Return true
298  *		to continue iterating tags, false to stop.
299  * @data:	Will be passed as second argument to @fn.
300  * @reserved:	Indicates whether @bt is the breserved_tags member or the
301  *		bitmap_tags member of struct blk_mq_tags.
302  */
303 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
304 			     busy_tag_iter_fn *fn, void *data, bool reserved)
305 {
306 	struct bt_tags_iter_data iter_data = {
307 		.tags = tags,
308 		.fn = fn,
309 		.data = data,
310 		.reserved = reserved,
311 	};
312 
313 	if (tags->rqs)
314 		sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
315 }
316 
317 /**
318  * blk_mq_all_tag_busy_iter - iterate over all started requests in a tag map
319  * @tags:	Tag map to iterate over.
320  * @fn:		Pointer to the function that will be called for each started
321  *		request. @fn will be called as follows: @fn(rq, @priv,
322  *		reserved) where rq is a pointer to a request. 'reserved'
323  *		indicates whether or not @rq is a reserved request. Return
324  *		true to continue iterating tags, false to stop.
325  * @priv:	Will be passed as second argument to @fn.
326  */
327 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
328 		busy_tag_iter_fn *fn, void *priv)
329 {
330 	if (tags->nr_reserved_tags)
331 		bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
332 	bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
333 }
334 
335 /**
336  * blk_mq_tagset_busy_iter - iterate over all started requests in a tag set
337  * @tagset:	Tag set to iterate over.
338  * @fn:		Pointer to the function that will be called for each started
339  *		request. @fn will be called as follows: @fn(rq, @priv,
340  *		reserved) where rq is a pointer to a request. 'reserved'
341  *		indicates whether or not @rq is a reserved request. Return
342  *		true to continue iterating tags, false to stop.
343  * @priv:	Will be passed as second argument to @fn.
344  */
345 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
346 		busy_tag_iter_fn *fn, void *priv)
347 {
348 	int i;
349 
350 	for (i = 0; i < tagset->nr_hw_queues; i++) {
351 		if (tagset->tags && tagset->tags[i])
352 			blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
353 	}
354 }
355 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
356 
357 /**
358  * blk_mq_queue_tag_busy_iter - iterate over all requests with a driver tag
359  * @q:		Request queue to examine.
360  * @fn:		Pointer to the function that will be called for each request
361  *		on @q. @fn will be called as follows: @fn(hctx, rq, @priv,
362  *		reserved) where rq is a pointer to a request and hctx points
363  *		to the hardware queue associated with the request. 'reserved'
364  *		indicates whether or not @rq is a reserved request.
365  * @priv:	Will be passed as third argument to @fn.
366  *
367  * Note: if @q->tag_set is shared with other request queues then @fn will be
368  * called for all requests on all queues that share that tag set and not only
369  * for requests associated with @q.
370  */
371 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
372 		void *priv)
373 {
374 	struct blk_mq_hw_ctx *hctx;
375 	int i;
376 
377 	/*
378 	 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
379 	 * while the queue is frozen. So we can use q_usage_counter to avoid
380 	 * racing with it. __blk_mq_update_nr_hw_queues() uses
381 	 * synchronize_rcu() to ensure this function left the critical section
382 	 * below.
383 	 */
384 	if (!percpu_ref_tryget(&q->q_usage_counter))
385 		return;
386 
387 	queue_for_each_hw_ctx(q, hctx, i) {
388 		struct blk_mq_tags *tags = hctx->tags;
389 
390 		/*
391 		 * If no software queues are currently mapped to this
392 		 * hardware queue, there's nothing to check
393 		 */
394 		if (!blk_mq_hw_queue_mapped(hctx))
395 			continue;
396 
397 		if (tags->nr_reserved_tags)
398 			bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
399 		bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
400 	}
401 	blk_queue_exit(q);
402 }
403 
404 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
405 		    bool round_robin, int node)
406 {
407 	return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
408 				       node);
409 }
410 
411 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
412 						   int node, int alloc_policy)
413 {
414 	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
415 	bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
416 
417 	if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
418 		goto free_tags;
419 	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
420 		     node))
421 		goto free_bitmap_tags;
422 
423 	return tags;
424 free_bitmap_tags:
425 	sbitmap_queue_free(&tags->bitmap_tags);
426 free_tags:
427 	kfree(tags);
428 	return NULL;
429 }
430 
431 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
432 				     unsigned int reserved_tags,
433 				     int node, int alloc_policy)
434 {
435 	struct blk_mq_tags *tags;
436 
437 	if (total_tags > BLK_MQ_TAG_MAX) {
438 		pr_err("blk-mq: tag depth too large\n");
439 		return NULL;
440 	}
441 
442 	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
443 	if (!tags)
444 		return NULL;
445 
446 	tags->nr_tags = total_tags;
447 	tags->nr_reserved_tags = reserved_tags;
448 
449 	return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
450 }
451 
452 void blk_mq_free_tags(struct blk_mq_tags *tags)
453 {
454 	sbitmap_queue_free(&tags->bitmap_tags);
455 	sbitmap_queue_free(&tags->breserved_tags);
456 	kfree(tags);
457 }
458 
459 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
460 			    struct blk_mq_tags **tagsptr, unsigned int tdepth,
461 			    bool can_grow)
462 {
463 	struct blk_mq_tags *tags = *tagsptr;
464 
465 	if (tdepth <= tags->nr_reserved_tags)
466 		return -EINVAL;
467 
468 	/*
469 	 * If we are allowed to grow beyond the original size, allocate
470 	 * a new set of tags before freeing the old one.
471 	 */
472 	if (tdepth > tags->nr_tags) {
473 		struct blk_mq_tag_set *set = hctx->queue->tag_set;
474 		struct blk_mq_tags *new;
475 		bool ret;
476 
477 		if (!can_grow)
478 			return -EINVAL;
479 
480 		/*
481 		 * We need some sort of upper limit, set it high enough that
482 		 * no valid use cases should require more.
483 		 */
484 		if (tdepth > 16 * BLKDEV_MAX_RQ)
485 			return -EINVAL;
486 
487 		new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth,
488 				tags->nr_reserved_tags);
489 		if (!new)
490 			return -ENOMEM;
491 		ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
492 		if (ret) {
493 			blk_mq_free_rq_map(new);
494 			return -ENOMEM;
495 		}
496 
497 		blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
498 		blk_mq_free_rq_map(*tagsptr);
499 		*tagsptr = new;
500 	} else {
501 		/*
502 		 * Don't need (or can't) update reserved tags here, they
503 		 * remain static and should never need resizing.
504 		 */
505 		sbitmap_queue_resize(&tags->bitmap_tags,
506 				tdepth - tags->nr_reserved_tags);
507 	}
508 
509 	return 0;
510 }
511 
512 /**
513  * blk_mq_unique_tag() - return a tag that is unique queue-wide
514  * @rq: request for which to compute a unique tag
515  *
516  * The tag field in struct request is unique per hardware queue but not over
517  * all hardware queues. Hence this function that returns a tag with the
518  * hardware context index in the upper bits and the per hardware queue tag in
519  * the lower bits.
520  *
521  * Note: When called for a request that is queued on a non-multiqueue request
522  * queue, the hardware context index is set to zero.
523  */
524 u32 blk_mq_unique_tag(struct request *rq)
525 {
526 	return (rq->mq_hctx->queue_num << BLK_MQ_UNIQUE_TAG_BITS) |
527 		(rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
528 }
529 EXPORT_SYMBOL(blk_mq_unique_tag);
530