xref: /openbmc/linux/block/blk-mq-tag.c (revision f7d84fa7) 
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 	if (data->shallow_depth)
100 		return __sbitmap_queue_get_shallow(bt, data->shallow_depth);
101 	else
102 		return __sbitmap_queue_get(bt);
103 }
104 
105 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
106 {
107 	struct blk_mq_tags *tags = blk_mq_tags_from_data(data);
108 	struct sbitmap_queue *bt;
109 	struct sbq_wait_state *ws;
110 	DEFINE_WAIT(wait);
111 	unsigned int tag_offset;
112 	bool drop_ctx;
113 	int tag;
114 
115 	if (data->flags & BLK_MQ_REQ_RESERVED) {
116 		if (unlikely(!tags->nr_reserved_tags)) {
117 			WARN_ON_ONCE(1);
118 			return BLK_MQ_TAG_FAIL;
119 		}
120 		bt = &tags->breserved_tags;
121 		tag_offset = 0;
122 	} else {
123 		bt = &tags->bitmap_tags;
124 		tag_offset = tags->nr_reserved_tags;
125 	}
126 
127 	tag = __blk_mq_get_tag(data, bt);
128 	if (tag != -1)
129 		goto found_tag;
130 
131 	if (data->flags & BLK_MQ_REQ_NOWAIT)
132 		return BLK_MQ_TAG_FAIL;
133 
134 	ws = bt_wait_ptr(bt, data->hctx);
135 	drop_ctx = data->ctx == NULL;
136 	do {
137 		prepare_to_wait(&ws->wait, &wait, TASK_UNINTERRUPTIBLE);
138 
139 		tag = __blk_mq_get_tag(data, bt);
140 		if (tag != -1)
141 			break;
142 
143 		/*
144 		 * We're out of tags on this hardware queue, kick any
145 		 * pending IO submits before going to sleep waiting for
146 		 * some to complete.
147 		 */
148 		blk_mq_run_hw_queue(data->hctx, false);
149 
150 		/*
151 		 * Retry tag allocation after running the hardware queue,
152 		 * as running the queue may also have found completions.
153 		 */
154 		tag = __blk_mq_get_tag(data, bt);
155 		if (tag != -1)
156 			break;
157 
158 		if (data->ctx)
159 			blk_mq_put_ctx(data->ctx);
160 
161 		io_schedule();
162 
163 		data->ctx = blk_mq_get_ctx(data->q);
164 		data->hctx = blk_mq_map_queue(data->q, data->ctx->cpu);
165 		tags = blk_mq_tags_from_data(data);
166 		if (data->flags & BLK_MQ_REQ_RESERVED)
167 			bt = &tags->breserved_tags;
168 		else
169 			bt = &tags->bitmap_tags;
170 
171 		finish_wait(&ws->wait, &wait);
172 		ws = bt_wait_ptr(bt, data->hctx);
173 	} while (1);
174 
175 	if (drop_ctx && data->ctx)
176 		blk_mq_put_ctx(data->ctx);
177 
178 	finish_wait(&ws->wait, &wait);
179 
180 found_tag:
181 	return tag + tag_offset;
182 }
183 
184 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, struct blk_mq_tags *tags,
185 		    struct blk_mq_ctx *ctx, unsigned int tag)
186 {
187 	if (!blk_mq_tag_is_reserved(tags, tag)) {
188 		const int real_tag = tag - tags->nr_reserved_tags;
189 
190 		BUG_ON(real_tag >= tags->nr_tags);
191 		sbitmap_queue_clear(&tags->bitmap_tags, real_tag, ctx->cpu);
192 	} else {
193 		BUG_ON(tag >= tags->nr_reserved_tags);
194 		sbitmap_queue_clear(&tags->breserved_tags, tag, ctx->cpu);
195 	}
196 }
197 
198 struct bt_iter_data {
199 	struct blk_mq_hw_ctx *hctx;
200 	busy_iter_fn *fn;
201 	void *data;
202 	bool reserved;
203 };
204 
205 static bool bt_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
206 {
207 	struct bt_iter_data *iter_data = data;
208 	struct blk_mq_hw_ctx *hctx = iter_data->hctx;
209 	struct blk_mq_tags *tags = hctx->tags;
210 	bool reserved = iter_data->reserved;
211 	struct request *rq;
212 
213 	if (!reserved)
214 		bitnr += tags->nr_reserved_tags;
215 	rq = tags->rqs[bitnr];
216 
217 	if (rq->q == hctx->queue)
218 		iter_data->fn(hctx, rq, iter_data->data, reserved);
219 	return true;
220 }
221 
222 static void bt_for_each(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt,
223 			busy_iter_fn *fn, void *data, bool reserved)
224 {
225 	struct bt_iter_data iter_data = {
226 		.hctx = hctx,
227 		.fn = fn,
228 		.data = data,
229 		.reserved = reserved,
230 	};
231 
232 	sbitmap_for_each_set(&bt->sb, bt_iter, &iter_data);
233 }
234 
235 struct bt_tags_iter_data {
236 	struct blk_mq_tags *tags;
237 	busy_tag_iter_fn *fn;
238 	void *data;
239 	bool reserved;
240 };
241 
242 static bool bt_tags_iter(struct sbitmap *bitmap, unsigned int bitnr, void *data)
243 {
244 	struct bt_tags_iter_data *iter_data = data;
245 	struct blk_mq_tags *tags = iter_data->tags;
246 	bool reserved = iter_data->reserved;
247 	struct request *rq;
248 
249 	if (!reserved)
250 		bitnr += tags->nr_reserved_tags;
251 	rq = tags->rqs[bitnr];
252 
253 	iter_data->fn(rq, iter_data->data, reserved);
254 	return true;
255 }
256 
257 static void bt_tags_for_each(struct blk_mq_tags *tags, struct sbitmap_queue *bt,
258 			     busy_tag_iter_fn *fn, void *data, bool reserved)
259 {
260 	struct bt_tags_iter_data iter_data = {
261 		.tags = tags,
262 		.fn = fn,
263 		.data = data,
264 		.reserved = reserved,
265 	};
266 
267 	if (tags->rqs)
268 		sbitmap_for_each_set(&bt->sb, bt_tags_iter, &iter_data);
269 }
270 
271 static void blk_mq_all_tag_busy_iter(struct blk_mq_tags *tags,
272 		busy_tag_iter_fn *fn, void *priv)
273 {
274 	if (tags->nr_reserved_tags)
275 		bt_tags_for_each(tags, &tags->breserved_tags, fn, priv, true);
276 	bt_tags_for_each(tags, &tags->bitmap_tags, fn, priv, false);
277 }
278 
279 void blk_mq_tagset_busy_iter(struct blk_mq_tag_set *tagset,
280 		busy_tag_iter_fn *fn, void *priv)
281 {
282 	int i;
283 
284 	for (i = 0; i < tagset->nr_hw_queues; i++) {
285 		if (tagset->tags && tagset->tags[i])
286 			blk_mq_all_tag_busy_iter(tagset->tags[i], fn, priv);
287 	}
288 }
289 EXPORT_SYMBOL(blk_mq_tagset_busy_iter);
290 
291 int blk_mq_reinit_tagset(struct blk_mq_tag_set *set)
292 {
293 	int i, j, ret = 0;
294 
295 	if (!set->ops->reinit_request)
296 		goto out;
297 
298 	for (i = 0; i < set->nr_hw_queues; i++) {
299 		struct blk_mq_tags *tags = set->tags[i];
300 
301 		if (!tags)
302 			continue;
303 
304 		for (j = 0; j < tags->nr_tags; j++) {
305 			if (!tags->static_rqs[j])
306 				continue;
307 
308 			ret = set->ops->reinit_request(set->driver_data,
309 						tags->static_rqs[j]);
310 			if (ret)
311 				goto out;
312 		}
313 	}
314 
315 out:
316 	return ret;
317 }
318 EXPORT_SYMBOL_GPL(blk_mq_reinit_tagset);
319 
320 void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_iter_fn *fn,
321 		void *priv)
322 {
323 	struct blk_mq_hw_ctx *hctx;
324 	int i;
325 
326 
327 	queue_for_each_hw_ctx(q, hctx, i) {
328 		struct blk_mq_tags *tags = hctx->tags;
329 
330 		/*
331 		 * If not software queues are currently mapped to this
332 		 * hardware queue, there's nothing to check
333 		 */
334 		if (!blk_mq_hw_queue_mapped(hctx))
335 			continue;
336 
337 		if (tags->nr_reserved_tags)
338 			bt_for_each(hctx, &tags->breserved_tags, fn, priv, true);
339 		bt_for_each(hctx, &tags->bitmap_tags, fn, priv, false);
340 	}
341 
342 }
343 
344 static int bt_alloc(struct sbitmap_queue *bt, unsigned int depth,
345 		    bool round_robin, int node)
346 {
347 	return sbitmap_queue_init_node(bt, depth, -1, round_robin, GFP_KERNEL,
348 				       node);
349 }
350 
351 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
352 						   int node, int alloc_policy)
353 {
354 	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
355 	bool round_robin = alloc_policy == BLK_TAG_ALLOC_RR;
356 
357 	if (bt_alloc(&tags->bitmap_tags, depth, round_robin, node))
358 		goto free_tags;
359 	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, round_robin,
360 		     node))
361 		goto free_bitmap_tags;
362 
363 	return tags;
364 free_bitmap_tags:
365 	sbitmap_queue_free(&tags->bitmap_tags);
366 free_tags:
367 	kfree(tags);
368 	return NULL;
369 }
370 
371 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
372 				     unsigned int reserved_tags,
373 				     int node, int alloc_policy)
374 {
375 	struct blk_mq_tags *tags;
376 
377 	if (total_tags > BLK_MQ_TAG_MAX) {
378 		pr_err("blk-mq: tag depth too large\n");
379 		return NULL;
380 	}
381 
382 	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
383 	if (!tags)
384 		return NULL;
385 
386 	tags->nr_tags = total_tags;
387 	tags->nr_reserved_tags = reserved_tags;
388 
389 	return blk_mq_init_bitmap_tags(tags, node, alloc_policy);
390 }
391 
392 void blk_mq_free_tags(struct blk_mq_tags *tags)
393 {
394 	sbitmap_queue_free(&tags->bitmap_tags);
395 	sbitmap_queue_free(&tags->breserved_tags);
396 	kfree(tags);
397 }
398 
399 int blk_mq_tag_update_depth(struct blk_mq_hw_ctx *hctx,
400 			    struct blk_mq_tags **tagsptr, unsigned int tdepth,
401 			    bool can_grow)
402 {
403 	struct blk_mq_tags *tags = *tagsptr;
404 
405 	if (tdepth <= tags->nr_reserved_tags)
406 		return -EINVAL;
407 
408 	tdepth -= tags->nr_reserved_tags;
409 
410 	/*
411 	 * If we are allowed to grow beyond the original size, allocate
412 	 * a new set of tags before freeing the old one.
413 	 */
414 	if (tdepth > tags->nr_tags) {
415 		struct blk_mq_tag_set *set = hctx->queue->tag_set;
416 		struct blk_mq_tags *new;
417 		bool ret;
418 
419 		if (!can_grow)
420 			return -EINVAL;
421 
422 		/*
423 		 * We need some sort of upper limit, set it high enough that
424 		 * no valid use cases should require more.
425 		 */
426 		if (tdepth > 16 * BLKDEV_MAX_RQ)
427 			return -EINVAL;
428 
429 		new = blk_mq_alloc_rq_map(set, hctx->queue_num, tdepth, 0);
430 		if (!new)
431 			return -ENOMEM;
432 		ret = blk_mq_alloc_rqs(set, new, hctx->queue_num, tdepth);
433 		if (ret) {
434 			blk_mq_free_rq_map(new);
435 			return -ENOMEM;
436 		}
437 
438 		blk_mq_free_rqs(set, *tagsptr, hctx->queue_num);
439 		blk_mq_free_rq_map(*tagsptr);
440 		*tagsptr = new;
441 	} else {
442 		/*
443 		 * Don't need (or can't) update reserved tags here, they
444 		 * remain static and should never need resizing.
445 		 */
446 		sbitmap_queue_resize(&tags->bitmap_tags, tdepth);
447 	}
448 
449 	return 0;
450 }
451 
452 /**
453  * blk_mq_unique_tag() - return a tag that is unique queue-wide
454  * @rq: request for which to compute a unique tag
455  *
456  * The tag field in struct request is unique per hardware queue but not over
457  * all hardware queues. Hence this function that returns a tag with the
458  * hardware context index in the upper bits and the per hardware queue tag in
459  * the lower bits.
460  *
461  * Note: When called for a request that is queued on a non-multiqueue request
462  * queue, the hardware context index is set to zero.
463  */
464 u32 blk_mq_unique_tag(struct request *rq)
465 {
466 	struct request_queue *q = rq->q;
467 	struct blk_mq_hw_ctx *hctx;
468 	int hwq = 0;
469 
470 	if (q->mq_ops) {
471 		hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
472 		hwq = hctx->queue_num;
473 	}
474 
475 	return (hwq << BLK_MQ_UNIQUE_TAG_BITS) |
476 		(rq->tag & BLK_MQ_UNIQUE_TAG_MASK);
477 }
478 EXPORT_SYMBOL(blk_mq_unique_tag);
479