xref: /openbmc/linux/block/blk-mq-tag.c (revision e5c86679) 
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  */
27 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
28 {
29 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
30 	    !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
31 		atomic_inc(&hctx->tags->active_queues);
32 
33 	return true;
34 }
35 
36 /*
37  * Wakeup all potentially sleeping on tags
38  */
39 void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags, bool include_reserve)
40 {
41 	sbitmap_queue_wake_all(&tags->bitmap_tags);
42 	if (include_reserve)
43 		sbitmap_queue_wake_all(&tags->breserved_tags);
44 }
45 
46 /*
47  * If a previously busy queue goes inactive, potential waiters could now
48  * be allowed to queue. Wake them up and check.
49  */
50 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
51 {
52 	struct blk_mq_tags *tags = hctx->tags;
53 
54 	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
55 		return;
56 
57 	atomic_dec(&tags->active_queues);
58 
59 	blk_mq_tag_wakeup_all(tags, false);
60 }
61 
62 /*
63  * For shared tag users, we track the number of currently active users
64  * and attempt to provide a fair share of the tag depth for each of them.
65  */
66 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
67 				  struct sbitmap_queue *bt)
68 {
69 	unsigned int depth, users;
70 
71 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
72 		return true;
73 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
74 		return true;
75 
76 	/*
77 	 * Don't try dividing an ant
78 	 */
79 	if (bt->sb.depth == 1)
80 		return true;
81 
82 	users = atomic_read(&hctx->tags->active_queues);
83 	if (!users)
84 		return true;
85 
86 	/*
87 	 * Allow at least some tags
88 	 */
89 	depth = max((bt->sb.depth + users - 1) / users, 4U);
90 	return atomic_read(&hctx->nr_active) < depth;
91 }
92 
93 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
94 			    struct sbitmap_queue *bt)
95 {
96 	if (!(data->flags & BLK_MQ_REQ_INTERNAL) &&
97 	    !hctx_may_queue(data->hctx, bt))
98 		return -1;
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_WAIT(wait);
108 	unsigned int tag_offset;
109 	bool drop_ctx;
110 	int tag;
111 
112 	if (data->flags & BLK_MQ_REQ_RESERVED) {
113 		if (unlikely(!tags->nr_reserved_tags)) {
114 			WARN_ON_ONCE(1);
115 			return BLK_MQ_TAG_FAIL;
116 		}
117 		bt = &tags->breserved_tags;
118 		tag_offset = 0;
119 	} else {
120 		bt = &tags->bitmap_tags;
121 		tag_offset = tags->nr_reserved_tags;
122 	}
123 
124 	tag = __blk_mq_get_tag(data, bt);
125 	if (tag != -1)
126 		goto found_tag;
127 
128 	if (data->flags & BLK_MQ_REQ_NOWAIT)
129 		return BLK_MQ_TAG_FAIL;
130 
131 	ws = bt_wait_ptr(bt, data->hctx);
132 	drop_ctx = data->ctx == NULL;
133 	do {
134 		prepare_to_wait(&ws->wait, &wait, TASK_UNINTERRUPTIBLE);
135 
136 		tag = __blk_mq_get_tag(data, bt);
137 		if (tag != -1)
138 			break;
139 
140 		/*
141 		 * We're out of tags on this hardware queue, kick any
142 		 * pending IO submits before going to sleep waiting for
143 		 * some to complete.
144 		 */
145 		blk_mq_run_hw_queue(data->hctx, false);
146 
147 		/*
148 		 * Retry tag allocation after running the hardware queue,
149 		 * as running the queue may also have found completions.
150 		 */
151 		tag = __blk_mq_get_tag(data, bt);
152 		if (tag != -1)
153 			break;
154 
155 		if (data->ctx)
156 			blk_mq_put_ctx(data->ctx);
157 
158 		io_schedule();
159 
160 		data->ctx = blk_mq_get_ctx(data->q);
161 		data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
162 		tags = blk_mq_tags_from_data(data);
163 		if (data->flags & BLK_MQ_REQ_RESERVED)
164 			bt = &tags->breserved_tags;
165 		else
166 			bt = &tags->bitmap_tags;
167 
168 		finish_wait(&ws->wait, &wait);
169 		ws = bt_wait_ptr(bt, data->hctx);
170 	} while (1);
171 
172 	if (drop_ctx && data->ctx)
173 		blk_mq_put_ctx(data->ctx);
174 
175 	finish_wait(&ws->wait, &wait);
176 
177 found_tag:
178 	return tag + tag_offset;
179 }
180 
181 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
182 		    struct blk_mq_ctx *ctx, unsigned int tag)
183 {
184 	if (!blk_mq_tag_is_reserved(tags, tag)) {
185 		const int real_tag = tag - tags->nr_reserved_tags;
186 
187 		BUG_ON(real_tag >= tags->nr_tags);
188 		sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
189 	} else {
190 		BUG_ON(tag >= tags->nr_reserved_tags);
191 		sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
192 	}
193 }
194 
195 struct bt_iter_data {
196 	struct blk_mq_hw_ctx *hctx;
197 	busy_iter_fn *fn;
198 	void *data;
199 	bool reserved;
200 };
201 
202 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
203 {
204 	struct bt_iter_data *iter_data = data;
205 	struct blk_mq_hw_ctx *hctx = iter_data->hctx;
206 	struct blk_mq_tags *tags = hctx->tags;
207 	bool reserved = iter_data->reserved;
208 	struct request *rq;
209 
210 	if (!reserved)
211 		bitnr += tags->nr_reserved_tags;
212 	rq = tags->rqs[bitnr];
213 
214 	if (rq->q == hctx->queue)
215 		iter_data->fn(hctx, rq, iter_data->data, reserved);
216 	return true;
217 }
218 
219 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
220 			busy_iter_fn *fn, void *data, bool reserved)
221 {
222 	struct bt_iter_data iter_data = {
223 		.hctx = hctx,
224 		.fn = fn,
225 		.data = data,
226 		.reserved = reserved,
227 	};
228 
229 	sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
230 }
231 
232 struct bt_tags_iter_data {
233 	struct blk_mq_tags *tags;
234 	busy_tag_iter_fn *fn;
235 	void *data;
236 	bool reserved;
237 };
238 
239 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
240 {
241 	struct bt_tags_iter_data *iter_data = data;
242 	struct blk_mq_tags *tags = iter_data->tags;
243 	bool reserved = iter_data->reserved;
244 	struct request *rq;
245 
246 	if (!reserved)
247 		bitnr += tags->nr_reserved_tags;
248 	rq = tags->rqs[bitnr];
249 
250 	iter_data->fn(rq, iter_data->data, reserved);
251 	return true;
252 }
253 
254 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
255 			     busy_tag_iter_fn *fn, void *data, bool reserved)
256 {
257 	struct bt_tags_iter_data iter_data = {
258 		.tags = tags,
259 		.fn = fn,
260 		.data = data,
261 		.reserved = reserved,
262 	};
263 
264 	if (tags->rqs)
265 		sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
266 }
267 
268 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
269 		busy_tag_iter_fn *fn, void *priv)
270 {
271 	if (tags->nr_reserved_tags)
272 		bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
273 	bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
274 }
275 
276 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
277 		busy_tag_iter_fn *fn, void *priv)
278 {
279 	int i;
280 
281 	for (i = 0; i < tagset->nr_hw_queues; i++) {
282 		if (tagset->tags && tagset->tags[i])
283 			blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
284 	}
285 }
286 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
287 
288 int blk_mq_reinit_tagset(struct blk_mq_tag_set *set)
289 {
290 	int i, j, ret = 0;
291 
292 	if (!set->ops->reinit_request)
293 		goto out;
294 
295 	for (i = 0; i < set->nr_hw_queues; i++) {
296 		struct blk_mq_tags *tags = set->tags[i];
297 
298 		if (!tags)
299 			continue;
300 
301 		for (j = 0; j < tags->nr_tags; j++) {
302 			if (!tags->static_rqs[j])
303 				continue;
304 
305 			ret = set->ops->reinit_request(set->driver_data,
306 						tags->static_rqs[j]);
307 			if (ret)
308 				goto out;
309 		}
310 	}
311 
312 out:
313 	return ret;
314 }
315 EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset);
316 
317 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
318 		void *priv)
319 {
320 	struct blk_mq_hw_ctx *hctx;
321 	int i;
322 
323 
324 	queue_for_each_hw_ctx(q, hctx, i) {
325 		struct blk_mq_tags *tags = hctx->tags;
326 
327 		/*
328 		 * If not software queues are currently mapped to this
329 		 * hardware queue, there's nothing to check
330 		 */
331 		if (!blk_mq_hw_queue_mapped(hctx))
332 			continue;
333 
334 		if (tags->nr_reserved_tags)
335 			bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
336 		bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
337 	}
338 
339 }
340 
341 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
342 		    bool round_robin, int node)
343 {
344 	return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
345 				       node);
346 }
347 
348 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
349 						   int node, int alloc_policy)
350 {
351 	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
352 	bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
353 
354 	if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
355 		goto free_tags;
356 	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
357 		     node))
358 		goto free_bitmap_tags;
359 
360 	return tags;
361 free_bitmap_tags:
362 	sbitmap_queue_free(&tags->bitmap_tags);
363 free_tags:
364 	kfree(tags);
365 	return NULL;
366 }
367 
368 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
369 				     unsigned int reserved_tags,
370 				     int node, int alloc_policy)
371 {
372 	struct blk_mq_tags *tags;
373 
374 	if (total_tags > BLK_MQ_TAG_MAX) {
375 		pr_err("blk-mq: tag depth too large\n");
376 		return NULL;
377 	}
378 
379 	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
380 	if (!tags)
381 		return NULL;
382 
383 	tags->nr_tags = total_tags;
384 	tags->nr_reserved_tags = reserved_tags;
385 
386 	return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
387 }
388 
389 void blk_mq_free_tags(struct blk_mq_tags *tags)
390 {
391 	sbitmap_queue_free(&tags->bitmap_tags);
392 	sbitmap_queue_free(&tags->breserved_tags);
393 	kfree(tags);
394 }
395 
396 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
397 			    struct blk_mq_tags **tagsptr, unsigned int tdepth,
398 			    bool can_grow)
399 {
400 	struct blk_mq_tags *tags = *tagsptr;
401 
402 	if (tdepth <= tags->nr_reserved_tags)
403 		return -EINVAL;
404 
405 	tdepth -= tags->nr_reserved_tags;
406 
407 	/*
408 	 * If we are allowed to grow beyond the original size, allocate
409 	 * a new set of tags before freeing the old one.
410 	 */
411 	if (tdepth > tags->nr_tags) {
412 		struct blk_mq_tag_set *set = hctx->queue->tag_set;
413 		struct blk_mq_tags *new;
414 		bool ret;
415 
416 		if (!can_grow)
417 			return -EINVAL;
418 
419 		/*
420 		 * We need some sort of upper limit, set it high enough that
421 		 * no valid use cases should require more.
422 		 */
423 		if (tdepth > 16 * BLKDEV_MAX_RQ)
424 			return -EINVAL;
425 
426 		new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth, 0);
427 		if (!new)
428 			return -ENOMEM;
429 		ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
430 		if (ret) {
431 			blk_mq_free_rq_map(new);
432 			return -ENOMEM;
433 		}
434 
435 		blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
436 		blk_mq_free_rq_map(*tagsptr);
437 		*tagsptr = new;
438 	} else {
439 		/*
440 		 * Don't need (or can't) update reserved tags here, they
441 		 * remain static and should never need resizing.
442 		 */
443 		sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
444 	}
445 
446 	return 0;
447 }
448 
449 /**
450  * blk_mq_unique_tag() - return a tag that is unique queue-wide
451  * @rq: request for which to compute a unique tag
452  *
453  * The tag field in struct request is unique per hardware queue but not over
454  * all hardware queues. Hence this function that returns a tag with the
455  * hardware context index in the upper bits and the per hardware queue tag in
456  * the lower bits.
457  *
458  * Note: When called for a request that is queued on a non-multiqueue request
459  * queue, the hardware context index is set to zero.
460  */
461 u32 blk_mq_unique_tag(struct request *rq)
462 {
463 	struct request_queue *q = rq->q;
464 	struct blk_mq_hw_ctx *hctx;
465 	int hwq = 0;
466 
467 	if (q->mq_ops) {
468 		hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
469 		hwq = hctx->queue_num;
470 	}
471 
472 	return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
473 		(rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
474 }
475 EXPORT_SYMBOL(blk_mq_unique_tag);
476