xref: /openbmc/linux/block/mq-deadline.c (revision 2f164822)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  MQ Deadline i/o scheduler - adaptation of the legacy deadline scheduler,
4  *  for the blk-mq scheduling framework
5  *
6  *  Copyright (C) 2016 Jens Axboe <axboe@kernel.dk>
7  */
8 #include <linux/kernel.h>
9 #include <linux/fs.h>
10 #include <linux/blkdev.h>
11 #include <linux/blk-mq.h>
12 #include <linux/bio.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/init.h>
16 #include <linux/compiler.h>
17 #include <linux/rbtree.h>
18 #include <linux/sbitmap.h>
19 
20 #include <trace/events/block.h>
21 
22 #include "elevator.h"
23 #include "blk.h"
24 #include "blk-mq.h"
25 #include "blk-mq-debugfs.h"
26 #include "blk-mq-tag.h"
27 #include "blk-mq-sched.h"
28 
29 /*
30  * See Documentation/block/deadline-iosched.rst
31  */
32 static const int read_expire = HZ / 2;  /* max time before a read is submitted. */
33 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
34 /*
35  * Time after which to dispatch lower priority requests even if higher
36  * priority requests are pending.
37  */
38 static const int prio_aging_expire = 10 * HZ;
39 static const int writes_starved = 2;    /* max times reads can starve a write */
40 static const int fifo_batch = 16;       /* # of sequential requests treated as one
41 				     by the above parameters. For throughput. */
42 
43 enum dd_data_dir {
44 	DD_READ		= READ,
45 	DD_WRITE	= WRITE,
46 };
47 
48 enum { DD_DIR_COUNT = 2 };
49 
50 enum dd_prio {
51 	DD_RT_PRIO	= 0,
52 	DD_BE_PRIO	= 1,
53 	DD_IDLE_PRIO	= 2,
54 	DD_PRIO_MAX	= 2,
55 };
56 
57 enum { DD_PRIO_COUNT = 3 };
58 
59 /*
60  * I/O statistics per I/O priority. It is fine if these counters overflow.
61  * What matters is that these counters are at least as wide as
62  * log2(max_outstanding_requests).
63  */
64 struct io_stats_per_prio {
65 	uint32_t inserted;
66 	uint32_t merged;
67 	uint32_t dispatched;
68 	atomic_t completed;
69 };
70 
71 /*
72  * Deadline scheduler data per I/O priority (enum dd_prio). Requests are
73  * present on both sort_list[] and fifo_list[].
74  */
75 struct dd_per_prio {
76 	struct list_head dispatch;
77 	struct rb_root sort_list[DD_DIR_COUNT];
78 	struct list_head fifo_list[DD_DIR_COUNT];
79 	/* Next request in FIFO order. Read, write or both are NULL. */
80 	struct request *next_rq[DD_DIR_COUNT];
81 	struct io_stats_per_prio stats;
82 };
83 
84 struct deadline_data {
85 	/*
86 	 * run time data
87 	 */
88 
89 	struct dd_per_prio per_prio[DD_PRIO_COUNT];
90 
91 	/* Data direction of latest dispatched request. */
92 	enum dd_data_dir last_dir;
93 	unsigned int batching;		/* number of sequential requests made */
94 	unsigned int starved;		/* times reads have starved writes */
95 
96 	/*
97 	 * settings that change how the i/o scheduler behaves
98 	 */
99 	int fifo_expire[DD_DIR_COUNT];
100 	int fifo_batch;
101 	int writes_starved;
102 	int front_merges;
103 	u32 async_depth;
104 	int prio_aging_expire;
105 
106 	spinlock_t lock;
107 	spinlock_t zone_lock;
108 };
109 
110 /* Maps an I/O priority class to a deadline scheduler priority. */
111 static const enum dd_prio ioprio_class_to_prio[] = {
112 	[IOPRIO_CLASS_NONE]	= DD_BE_PRIO,
113 	[IOPRIO_CLASS_RT]	= DD_RT_PRIO,
114 	[IOPRIO_CLASS_BE]	= DD_BE_PRIO,
115 	[IOPRIO_CLASS_IDLE]	= DD_IDLE_PRIO,
116 };
117 
118 static inline struct rb_root *
119 deadline_rb_root(struct dd_per_prio *per_prio, struct request *rq)
120 {
121 	return &per_prio->sort_list[rq_data_dir(rq)];
122 }
123 
124 /*
125  * Returns the I/O priority class (IOPRIO_CLASS_*) that has been assigned to a
126  * request.
127  */
128 static u8 dd_rq_ioclass(struct request *rq)
129 {
130 	return IOPRIO_PRIO_CLASS(req_get_ioprio(rq));
131 }
132 
133 /*
134  * get the request before `rq' in sector-sorted order
135  */
136 static inline struct request *
137 deadline_earlier_request(struct request *rq)
138 {
139 	struct rb_node *node = rb_prev(&rq->rb_node);
140 
141 	if (node)
142 		return rb_entry_rq(node);
143 
144 	return NULL;
145 }
146 
147 /*
148  * get the request after `rq' in sector-sorted order
149  */
150 static inline struct request *
151 deadline_latter_request(struct request *rq)
152 {
153 	struct rb_node *node = rb_next(&rq->rb_node);
154 
155 	if (node)
156 		return rb_entry_rq(node);
157 
158 	return NULL;
159 }
160 
161 static void
162 deadline_add_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
163 {
164 	struct rb_root *root = deadline_rb_root(per_prio, rq);
165 
166 	elv_rb_add(root, rq);
167 }
168 
169 static inline void
170 deadline_del_rq_rb(struct dd_per_prio *per_prio, struct request *rq)
171 {
172 	const enum dd_data_dir data_dir = rq_data_dir(rq);
173 
174 	if (per_prio->next_rq[data_dir] == rq)
175 		per_prio->next_rq[data_dir] = deadline_latter_request(rq);
176 
177 	elv_rb_del(deadline_rb_root(per_prio, rq), rq);
178 }
179 
180 /*
181  * remove rq from rbtree and fifo.
182  */
183 static void deadline_remove_request(struct request_queue *q,
184 				    struct dd_per_prio *per_prio,
185 				    struct request *rq)
186 {
187 	list_del_init(&rq->queuelist);
188 
189 	/*
190 	 * We might not be on the rbtree, if we are doing an insert merge
191 	 */
192 	if (!RB_EMPTY_NODE(&rq->rb_node))
193 		deadline_del_rq_rb(per_prio, rq);
194 
195 	elv_rqhash_del(q, rq);
196 	if (q->last_merge == rq)
197 		q->last_merge = NULL;
198 }
199 
200 static void dd_request_merged(struct request_queue *q, struct request *req,
201 			      enum elv_merge type)
202 {
203 	struct deadline_data *dd = q->elevator->elevator_data;
204 	const u8 ioprio_class = dd_rq_ioclass(req);
205 	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
206 	struct dd_per_prio *per_prio = &dd->per_prio[prio];
207 
208 	/*
209 	 * if the merge was a front merge, we need to reposition request
210 	 */
211 	if (type == ELEVATOR_FRONT_MERGE) {
212 		elv_rb_del(deadline_rb_root(per_prio, req), req);
213 		deadline_add_rq_rb(per_prio, req);
214 	}
215 }
216 
217 /*
218  * Callback function that is invoked after @next has been merged into @req.
219  */
220 static void dd_merged_requests(struct request_queue *q, struct request *req,
221 			       struct request *next)
222 {
223 	struct deadline_data *dd = q->elevator->elevator_data;
224 	const u8 ioprio_class = dd_rq_ioclass(next);
225 	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
226 
227 	lockdep_assert_held(&dd->lock);
228 
229 	dd->per_prio[prio].stats.merged++;
230 
231 	/*
232 	 * if next expires before rq, assign its expire time to rq
233 	 * and move into next position (next will be deleted) in fifo
234 	 */
235 	if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
236 		if (time_before((unsigned long)next->fifo_time,
237 				(unsigned long)req->fifo_time)) {
238 			list_move(&req->queuelist, &next->queuelist);
239 			req->fifo_time = next->fifo_time;
240 		}
241 	}
242 
243 	/*
244 	 * kill knowledge of next, this one is a goner
245 	 */
246 	deadline_remove_request(q, &dd->per_prio[prio], next);
247 }
248 
249 /*
250  * move an entry to dispatch queue
251  */
252 static void
253 deadline_move_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
254 		      struct request *rq)
255 {
256 	const enum dd_data_dir data_dir = rq_data_dir(rq);
257 
258 	per_prio->next_rq[data_dir] = deadline_latter_request(rq);
259 
260 	/*
261 	 * take it off the sort and fifo list
262 	 */
263 	deadline_remove_request(rq->q, per_prio, rq);
264 }
265 
266 /* Number of requests queued for a given priority level. */
267 static u32 dd_queued(struct deadline_data *dd, enum dd_prio prio)
268 {
269 	const struct io_stats_per_prio *stats = &dd->per_prio[prio].stats;
270 
271 	lockdep_assert_held(&dd->lock);
272 
273 	return stats->inserted - atomic_read(&stats->completed);
274 }
275 
276 /*
277  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
278  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
279  */
280 static inline int deadline_check_fifo(struct dd_per_prio *per_prio,
281 				      enum dd_data_dir data_dir)
282 {
283 	struct request *rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
284 
285 	/*
286 	 * rq is expired!
287 	 */
288 	if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
289 		return 1;
290 
291 	return 0;
292 }
293 
294 /*
295  * Check if rq has a sequential request preceding it.
296  */
297 static bool deadline_is_seq_write(struct deadline_data *dd, struct request *rq)
298 {
299 	struct request *prev = deadline_earlier_request(rq);
300 
301 	if (!prev)
302 		return false;
303 
304 	return blk_rq_pos(prev) + blk_rq_sectors(prev) == blk_rq_pos(rq);
305 }
306 
307 /*
308  * Skip all write requests that are sequential from @rq, even if we cross
309  * a zone boundary.
310  */
311 static struct request *deadline_skip_seq_writes(struct deadline_data *dd,
312 						struct request *rq)
313 {
314 	sector_t pos = blk_rq_pos(rq);
315 	sector_t skipped_sectors = 0;
316 
317 	while (rq) {
318 		if (blk_rq_pos(rq) != pos + skipped_sectors)
319 			break;
320 		skipped_sectors += blk_rq_sectors(rq);
321 		rq = deadline_latter_request(rq);
322 	}
323 
324 	return rq;
325 }
326 
327 /*
328  * For the specified data direction, return the next request to
329  * dispatch using arrival ordered lists.
330  */
331 static struct request *
332 deadline_fifo_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
333 		      enum dd_data_dir data_dir)
334 {
335 	struct request *rq;
336 	unsigned long flags;
337 
338 	if (list_empty(&per_prio->fifo_list[data_dir]))
339 		return NULL;
340 
341 	rq = rq_entry_fifo(per_prio->fifo_list[data_dir].next);
342 	if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
343 		return rq;
344 
345 	/*
346 	 * Look for a write request that can be dispatched, that is one with
347 	 * an unlocked target zone. For some HDDs, breaking a sequential
348 	 * write stream can lead to lower throughput, so make sure to preserve
349 	 * sequential write streams, even if that stream crosses into the next
350 	 * zones and these zones are unlocked.
351 	 */
352 	spin_lock_irqsave(&dd->zone_lock, flags);
353 	list_for_each_entry(rq, &per_prio->fifo_list[DD_WRITE], queuelist) {
354 		if (blk_req_can_dispatch_to_zone(rq) &&
355 		    (blk_queue_nonrot(rq->q) ||
356 		     !deadline_is_seq_write(dd, rq)))
357 			goto out;
358 	}
359 	rq = NULL;
360 out:
361 	spin_unlock_irqrestore(&dd->zone_lock, flags);
362 
363 	return rq;
364 }
365 
366 /*
367  * For the specified data direction, return the next request to
368  * dispatch using sector position sorted lists.
369  */
370 static struct request *
371 deadline_next_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
372 		      enum dd_data_dir data_dir)
373 {
374 	struct request *rq;
375 	unsigned long flags;
376 
377 	rq = per_prio->next_rq[data_dir];
378 	if (!rq)
379 		return NULL;
380 
381 	if (data_dir == DD_READ || !blk_queue_is_zoned(rq->q))
382 		return rq;
383 
384 	/*
385 	 * Look for a write request that can be dispatched, that is one with
386 	 * an unlocked target zone. For some HDDs, breaking a sequential
387 	 * write stream can lead to lower throughput, so make sure to preserve
388 	 * sequential write streams, even if that stream crosses into the next
389 	 * zones and these zones are unlocked.
390 	 */
391 	spin_lock_irqsave(&dd->zone_lock, flags);
392 	while (rq) {
393 		if (blk_req_can_dispatch_to_zone(rq))
394 			break;
395 		if (blk_queue_nonrot(rq->q))
396 			rq = deadline_latter_request(rq);
397 		else
398 			rq = deadline_skip_seq_writes(dd, rq);
399 	}
400 	spin_unlock_irqrestore(&dd->zone_lock, flags);
401 
402 	return rq;
403 }
404 
405 /*
406  * Returns true if and only if @rq started after @latest_start where
407  * @latest_start is in jiffies.
408  */
409 static bool started_after(struct deadline_data *dd, struct request *rq,
410 			  unsigned long latest_start)
411 {
412 	unsigned long start_time = (unsigned long)rq->fifo_time;
413 
414 	start_time -= dd->fifo_expire[rq_data_dir(rq)];
415 
416 	return time_after(start_time, latest_start);
417 }
418 
419 /*
420  * deadline_dispatch_requests selects the best request according to
421  * read/write expire, fifo_batch, etc and with a start time <= @latest_start.
422  */
423 static struct request *__dd_dispatch_request(struct deadline_data *dd,
424 					     struct dd_per_prio *per_prio,
425 					     unsigned long latest_start)
426 {
427 	struct request *rq, *next_rq;
428 	enum dd_data_dir data_dir;
429 	enum dd_prio prio;
430 	u8 ioprio_class;
431 
432 	lockdep_assert_held(&dd->lock);
433 
434 	if (!list_empty(&per_prio->dispatch)) {
435 		rq = list_first_entry(&per_prio->dispatch, struct request,
436 				      queuelist);
437 		if (started_after(dd, rq, latest_start))
438 			return NULL;
439 		list_del_init(&rq->queuelist);
440 		goto done;
441 	}
442 
443 	/*
444 	 * batches are currently reads XOR writes
445 	 */
446 	rq = deadline_next_request(dd, per_prio, dd->last_dir);
447 	if (rq && dd->batching < dd->fifo_batch)
448 		/* we have a next request are still entitled to batch */
449 		goto dispatch_request;
450 
451 	/*
452 	 * at this point we are not running a batch. select the appropriate
453 	 * data direction (read / write)
454 	 */
455 
456 	if (!list_empty(&per_prio->fifo_list[DD_READ])) {
457 		BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_READ]));
458 
459 		if (deadline_fifo_request(dd, per_prio, DD_WRITE) &&
460 		    (dd->starved++ >= dd->writes_starved))
461 			goto dispatch_writes;
462 
463 		data_dir = DD_READ;
464 
465 		goto dispatch_find_request;
466 	}
467 
468 	/*
469 	 * there are either no reads or writes have been starved
470 	 */
471 
472 	if (!list_empty(&per_prio->fifo_list[DD_WRITE])) {
473 dispatch_writes:
474 		BUG_ON(RB_EMPTY_ROOT(&per_prio->sort_list[DD_WRITE]));
475 
476 		dd->starved = 0;
477 
478 		data_dir = DD_WRITE;
479 
480 		goto dispatch_find_request;
481 	}
482 
483 	return NULL;
484 
485 dispatch_find_request:
486 	/*
487 	 * we are not running a batch, find best request for selected data_dir
488 	 */
489 	next_rq = deadline_next_request(dd, per_prio, data_dir);
490 	if (deadline_check_fifo(per_prio, data_dir) || !next_rq) {
491 		/*
492 		 * A deadline has expired, the last request was in the other
493 		 * direction, or we have run out of higher-sectored requests.
494 		 * Start again from the request with the earliest expiry time.
495 		 */
496 		rq = deadline_fifo_request(dd, per_prio, data_dir);
497 	} else {
498 		/*
499 		 * The last req was the same dir and we have a next request in
500 		 * sort order. No expired requests so continue on from here.
501 		 */
502 		rq = next_rq;
503 	}
504 
505 	/*
506 	 * For a zoned block device, if we only have writes queued and none of
507 	 * them can be dispatched, rq will be NULL.
508 	 */
509 	if (!rq)
510 		return NULL;
511 
512 	dd->last_dir = data_dir;
513 	dd->batching = 0;
514 
515 dispatch_request:
516 	if (started_after(dd, rq, latest_start))
517 		return NULL;
518 
519 	/*
520 	 * rq is the selected appropriate request.
521 	 */
522 	dd->batching++;
523 	deadline_move_request(dd, per_prio, rq);
524 done:
525 	ioprio_class = dd_rq_ioclass(rq);
526 	prio = ioprio_class_to_prio[ioprio_class];
527 	dd->per_prio[prio].stats.dispatched++;
528 	/*
529 	 * If the request needs its target zone locked, do it.
530 	 */
531 	blk_req_zone_write_lock(rq);
532 	rq->rq_flags |= RQF_STARTED;
533 	return rq;
534 }
535 
536 /*
537  * Check whether there are any requests with priority other than DD_RT_PRIO
538  * that were inserted more than prio_aging_expire jiffies ago.
539  */
540 static struct request *dd_dispatch_prio_aged_requests(struct deadline_data *dd,
541 						      unsigned long now)
542 {
543 	struct request *rq;
544 	enum dd_prio prio;
545 	int prio_cnt;
546 
547 	lockdep_assert_held(&dd->lock);
548 
549 	prio_cnt = !!dd_queued(dd, DD_RT_PRIO) + !!dd_queued(dd, DD_BE_PRIO) +
550 		   !!dd_queued(dd, DD_IDLE_PRIO);
551 	if (prio_cnt < 2)
552 		return NULL;
553 
554 	for (prio = DD_BE_PRIO; prio <= DD_PRIO_MAX; prio++) {
555 		rq = __dd_dispatch_request(dd, &dd->per_prio[prio],
556 					   now - dd->prio_aging_expire);
557 		if (rq)
558 			return rq;
559 	}
560 
561 	return NULL;
562 }
563 
564 /*
565  * Called from blk_mq_run_hw_queue() -> __blk_mq_sched_dispatch_requests().
566  *
567  * One confusing aspect here is that we get called for a specific
568  * hardware queue, but we may return a request that is for a
569  * different hardware queue. This is because mq-deadline has shared
570  * state for all hardware queues, in terms of sorting, FIFOs, etc.
571  */
572 static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
573 {
574 	struct deadline_data *dd = hctx->queue->elevator->elevator_data;
575 	const unsigned long now = jiffies;
576 	struct request *rq;
577 	enum dd_prio prio;
578 
579 	spin_lock(&dd->lock);
580 	rq = dd_dispatch_prio_aged_requests(dd, now);
581 	if (rq)
582 		goto unlock;
583 
584 	/*
585 	 * Next, dispatch requests in priority order. Ignore lower priority
586 	 * requests if any higher priority requests are pending.
587 	 */
588 	for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
589 		rq = __dd_dispatch_request(dd, &dd->per_prio[prio], now);
590 		if (rq || dd_queued(dd, prio))
591 			break;
592 	}
593 
594 unlock:
595 	spin_unlock(&dd->lock);
596 
597 	return rq;
598 }
599 
600 /*
601  * Called by __blk_mq_alloc_request(). The shallow_depth value set by this
602  * function is used by __blk_mq_get_tag().
603  */
604 static void dd_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
605 {
606 	struct deadline_data *dd = data->q->elevator->elevator_data;
607 
608 	/* Do not throttle synchronous reads. */
609 	if (op_is_sync(opf) && !op_is_write(opf))
610 		return;
611 
612 	/*
613 	 * Throttle asynchronous requests and writes such that these requests
614 	 * do not block the allocation of synchronous requests.
615 	 */
616 	data->shallow_depth = dd->async_depth;
617 }
618 
619 /* Called by blk_mq_update_nr_requests(). */
620 static void dd_depth_updated(struct blk_mq_hw_ctx *hctx)
621 {
622 	struct request_queue *q = hctx->queue;
623 	struct deadline_data *dd = q->elevator->elevator_data;
624 	struct blk_mq_tags *tags = hctx->sched_tags;
625 
626 	dd->async_depth = max(1UL, 3 * q->nr_requests / 4);
627 
628 	sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, dd->async_depth);
629 }
630 
631 /* Called by blk_mq_init_hctx() and blk_mq_init_sched(). */
632 static int dd_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
633 {
634 	dd_depth_updated(hctx);
635 	return 0;
636 }
637 
638 static void dd_exit_sched(struct elevator_queue *e)
639 {
640 	struct deadline_data *dd = e->elevator_data;
641 	enum dd_prio prio;
642 
643 	for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
644 		struct dd_per_prio *per_prio = &dd->per_prio[prio];
645 		const struct io_stats_per_prio *stats = &per_prio->stats;
646 		uint32_t queued;
647 
648 		WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_READ]));
649 		WARN_ON_ONCE(!list_empty(&per_prio->fifo_list[DD_WRITE]));
650 
651 		spin_lock(&dd->lock);
652 		queued = dd_queued(dd, prio);
653 		spin_unlock(&dd->lock);
654 
655 		WARN_ONCE(queued != 0,
656 			  "statistics for priority %d: i %u m %u d %u c %u\n",
657 			  prio, stats->inserted, stats->merged,
658 			  stats->dispatched, atomic_read(&stats->completed));
659 	}
660 
661 	kfree(dd);
662 }
663 
664 /*
665  * initialize elevator private data (deadline_data).
666  */
667 static int dd_init_sched(struct request_queue *q, struct elevator_type *e)
668 {
669 	struct deadline_data *dd;
670 	struct elevator_queue *eq;
671 	enum dd_prio prio;
672 	int ret = -ENOMEM;
673 
674 	eq = elevator_alloc(q, e);
675 	if (!eq)
676 		return ret;
677 
678 	dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
679 	if (!dd)
680 		goto put_eq;
681 
682 	eq->elevator_data = dd;
683 
684 	for (prio = 0; prio <= DD_PRIO_MAX; prio++) {
685 		struct dd_per_prio *per_prio = &dd->per_prio[prio];
686 
687 		INIT_LIST_HEAD(&per_prio->dispatch);
688 		INIT_LIST_HEAD(&per_prio->fifo_list[DD_READ]);
689 		INIT_LIST_HEAD(&per_prio->fifo_list[DD_WRITE]);
690 		per_prio->sort_list[DD_READ] = RB_ROOT;
691 		per_prio->sort_list[DD_WRITE] = RB_ROOT;
692 	}
693 	dd->fifo_expire[DD_READ] = read_expire;
694 	dd->fifo_expire[DD_WRITE] = write_expire;
695 	dd->writes_starved = writes_starved;
696 	dd->front_merges = 1;
697 	dd->last_dir = DD_WRITE;
698 	dd->fifo_batch = fifo_batch;
699 	dd->prio_aging_expire = prio_aging_expire;
700 	spin_lock_init(&dd->lock);
701 	spin_lock_init(&dd->zone_lock);
702 
703 	/* We dispatch from request queue wide instead of hw queue */
704 	blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
705 
706 	q->elevator = eq;
707 	return 0;
708 
709 put_eq:
710 	kobject_put(&eq->kobj);
711 	return ret;
712 }
713 
714 /*
715  * Try to merge @bio into an existing request. If @bio has been merged into
716  * an existing request, store the pointer to that request into *@rq.
717  */
718 static int dd_request_merge(struct request_queue *q, struct request **rq,
719 			    struct bio *bio)
720 {
721 	struct deadline_data *dd = q->elevator->elevator_data;
722 	const u8 ioprio_class = IOPRIO_PRIO_CLASS(bio->bi_ioprio);
723 	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
724 	struct dd_per_prio *per_prio = &dd->per_prio[prio];
725 	sector_t sector = bio_end_sector(bio);
726 	struct request *__rq;
727 
728 	if (!dd->front_merges)
729 		return ELEVATOR_NO_MERGE;
730 
731 	__rq = elv_rb_find(&per_prio->sort_list[bio_data_dir(bio)], sector);
732 	if (__rq) {
733 		BUG_ON(sector != blk_rq_pos(__rq));
734 
735 		if (elv_bio_merge_ok(__rq, bio)) {
736 			*rq = __rq;
737 			if (blk_discard_mergable(__rq))
738 				return ELEVATOR_DISCARD_MERGE;
739 			return ELEVATOR_FRONT_MERGE;
740 		}
741 	}
742 
743 	return ELEVATOR_NO_MERGE;
744 }
745 
746 /*
747  * Attempt to merge a bio into an existing request. This function is called
748  * before @bio is associated with a request.
749  */
750 static bool dd_bio_merge(struct request_queue *q, struct bio *bio,
751 		unsigned int nr_segs)
752 {
753 	struct deadline_data *dd = q->elevator->elevator_data;
754 	struct request *free = NULL;
755 	bool ret;
756 
757 	spin_lock(&dd->lock);
758 	ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
759 	spin_unlock(&dd->lock);
760 
761 	if (free)
762 		blk_mq_free_request(free);
763 
764 	return ret;
765 }
766 
767 /*
768  * add rq to rbtree and fifo
769  */
770 static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
771 			      bool at_head)
772 {
773 	struct request_queue *q = hctx->queue;
774 	struct deadline_data *dd = q->elevator->elevator_data;
775 	const enum dd_data_dir data_dir = rq_data_dir(rq);
776 	u16 ioprio = req_get_ioprio(rq);
777 	u8 ioprio_class = IOPRIO_PRIO_CLASS(ioprio);
778 	struct dd_per_prio *per_prio;
779 	enum dd_prio prio;
780 	LIST_HEAD(free);
781 
782 	lockdep_assert_held(&dd->lock);
783 
784 	/*
785 	 * This may be a requeue of a write request that has locked its
786 	 * target zone. If it is the case, this releases the zone lock.
787 	 */
788 	blk_req_zone_write_unlock(rq);
789 
790 	prio = ioprio_class_to_prio[ioprio_class];
791 	per_prio = &dd->per_prio[prio];
792 	if (!rq->elv.priv[0]) {
793 		per_prio->stats.inserted++;
794 		rq->elv.priv[0] = (void *)(uintptr_t)1;
795 	}
796 
797 	if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
798 		blk_mq_free_requests(&free);
799 		return;
800 	}
801 
802 	trace_block_rq_insert(rq);
803 
804 	if (at_head) {
805 		list_add(&rq->queuelist, &per_prio->dispatch);
806 		rq->fifo_time = jiffies;
807 	} else {
808 		deadline_add_rq_rb(per_prio, rq);
809 
810 		if (rq_mergeable(rq)) {
811 			elv_rqhash_add(q, rq);
812 			if (!q->last_merge)
813 				q->last_merge = rq;
814 		}
815 
816 		/*
817 		 * set expire time and add to fifo list
818 		 */
819 		rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
820 		list_add_tail(&rq->queuelist, &per_prio->fifo_list[data_dir]);
821 	}
822 }
823 
824 /*
825  * Called from blk_mq_sched_insert_request() or blk_mq_sched_insert_requests().
826  */
827 static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
828 			       struct list_head *list, bool at_head)
829 {
830 	struct request_queue *q = hctx->queue;
831 	struct deadline_data *dd = q->elevator->elevator_data;
832 
833 	spin_lock(&dd->lock);
834 	while (!list_empty(list)) {
835 		struct request *rq;
836 
837 		rq = list_first_entry(list, struct request, queuelist);
838 		list_del_init(&rq->queuelist);
839 		dd_insert_request(hctx, rq, at_head);
840 	}
841 	spin_unlock(&dd->lock);
842 }
843 
844 /* Callback from inside blk_mq_rq_ctx_init(). */
845 static void dd_prepare_request(struct request *rq)
846 {
847 	rq->elv.priv[0] = NULL;
848 }
849 
850 static bool dd_has_write_work(struct blk_mq_hw_ctx *hctx)
851 {
852 	struct deadline_data *dd = hctx->queue->elevator->elevator_data;
853 	enum dd_prio p;
854 
855 	for (p = 0; p <= DD_PRIO_MAX; p++)
856 		if (!list_empty_careful(&dd->per_prio[p].fifo_list[DD_WRITE]))
857 			return true;
858 
859 	return false;
860 }
861 
862 /*
863  * Callback from inside blk_mq_free_request().
864  *
865  * For zoned block devices, write unlock the target zone of
866  * completed write requests. Do this while holding the zone lock
867  * spinlock so that the zone is never unlocked while deadline_fifo_request()
868  * or deadline_next_request() are executing. This function is called for
869  * all requests, whether or not these requests complete successfully.
870  *
871  * For a zoned block device, __dd_dispatch_request() may have stopped
872  * dispatching requests if all the queued requests are write requests directed
873  * at zones that are already locked due to on-going write requests. To ensure
874  * write request dispatch progress in this case, mark the queue as needing a
875  * restart to ensure that the queue is run again after completion of the
876  * request and zones being unlocked.
877  */
878 static void dd_finish_request(struct request *rq)
879 {
880 	struct request_queue *q = rq->q;
881 	struct deadline_data *dd = q->elevator->elevator_data;
882 	const u8 ioprio_class = dd_rq_ioclass(rq);
883 	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
884 	struct dd_per_prio *per_prio = &dd->per_prio[prio];
885 
886 	/*
887 	 * The block layer core may call dd_finish_request() without having
888 	 * called dd_insert_requests(). Skip requests that bypassed I/O
889 	 * scheduling. See also blk_mq_request_bypass_insert().
890 	 */
891 	if (!rq->elv.priv[0])
892 		return;
893 
894 	atomic_inc(&per_prio->stats.completed);
895 
896 	if (blk_queue_is_zoned(q)) {
897 		unsigned long flags;
898 
899 		spin_lock_irqsave(&dd->zone_lock, flags);
900 		blk_req_zone_write_unlock(rq);
901 		spin_unlock_irqrestore(&dd->zone_lock, flags);
902 
903 		if (dd_has_write_work(rq->mq_hctx))
904 			blk_mq_sched_mark_restart_hctx(rq->mq_hctx);
905 	}
906 }
907 
908 static bool dd_has_work_for_prio(struct dd_per_prio *per_prio)
909 {
910 	return !list_empty_careful(&per_prio->dispatch) ||
911 		!list_empty_careful(&per_prio->fifo_list[DD_READ]) ||
912 		!list_empty_careful(&per_prio->fifo_list[DD_WRITE]);
913 }
914 
915 static bool dd_has_work(struct blk_mq_hw_ctx *hctx)
916 {
917 	struct deadline_data *dd = hctx->queue->elevator->elevator_data;
918 	enum dd_prio prio;
919 
920 	for (prio = 0; prio <= DD_PRIO_MAX; prio++)
921 		if (dd_has_work_for_prio(&dd->per_prio[prio]))
922 			return true;
923 
924 	return false;
925 }
926 
927 /*
928  * sysfs parts below
929  */
930 #define SHOW_INT(__FUNC, __VAR)						\
931 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
932 {									\
933 	struct deadline_data *dd = e->elevator_data;			\
934 									\
935 	return sysfs_emit(page, "%d\n", __VAR);				\
936 }
937 #define SHOW_JIFFIES(__FUNC, __VAR) SHOW_INT(__FUNC, jiffies_to_msecs(__VAR))
938 SHOW_JIFFIES(deadline_read_expire_show, dd->fifo_expire[DD_READ]);
939 SHOW_JIFFIES(deadline_write_expire_show, dd->fifo_expire[DD_WRITE]);
940 SHOW_JIFFIES(deadline_prio_aging_expire_show, dd->prio_aging_expire);
941 SHOW_INT(deadline_writes_starved_show, dd->writes_starved);
942 SHOW_INT(deadline_front_merges_show, dd->front_merges);
943 SHOW_INT(deadline_async_depth_show, dd->async_depth);
944 SHOW_INT(deadline_fifo_batch_show, dd->fifo_batch);
945 #undef SHOW_INT
946 #undef SHOW_JIFFIES
947 
948 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
949 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)	\
950 {									\
951 	struct deadline_data *dd = e->elevator_data;			\
952 	int __data, __ret;						\
953 									\
954 	__ret = kstrtoint(page, 0, &__data);				\
955 	if (__ret < 0)							\
956 		return __ret;						\
957 	if (__data < (MIN))						\
958 		__data = (MIN);						\
959 	else if (__data > (MAX))					\
960 		__data = (MAX);						\
961 	*(__PTR) = __CONV(__data);					\
962 	return count;							\
963 }
964 #define STORE_INT(__FUNC, __PTR, MIN, MAX)				\
965 	STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, )
966 #define STORE_JIFFIES(__FUNC, __PTR, MIN, MAX)				\
967 	STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, msecs_to_jiffies)
968 STORE_JIFFIES(deadline_read_expire_store, &dd->fifo_expire[DD_READ], 0, INT_MAX);
969 STORE_JIFFIES(deadline_write_expire_store, &dd->fifo_expire[DD_WRITE], 0, INT_MAX);
970 STORE_JIFFIES(deadline_prio_aging_expire_store, &dd->prio_aging_expire, 0, INT_MAX);
971 STORE_INT(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX);
972 STORE_INT(deadline_front_merges_store, &dd->front_merges, 0, 1);
973 STORE_INT(deadline_async_depth_store, &dd->async_depth, 1, INT_MAX);
974 STORE_INT(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX);
975 #undef STORE_FUNCTION
976 #undef STORE_INT
977 #undef STORE_JIFFIES
978 
979 #define DD_ATTR(name) \
980 	__ATTR(name, 0644, deadline_##name##_show, deadline_##name##_store)
981 
982 static struct elv_fs_entry deadline_attrs[] = {
983 	DD_ATTR(read_expire),
984 	DD_ATTR(write_expire),
985 	DD_ATTR(writes_starved),
986 	DD_ATTR(front_merges),
987 	DD_ATTR(async_depth),
988 	DD_ATTR(fifo_batch),
989 	DD_ATTR(prio_aging_expire),
990 	__ATTR_NULL
991 };
992 
993 #ifdef CONFIG_BLK_DEBUG_FS
994 #define DEADLINE_DEBUGFS_DDIR_ATTRS(prio, data_dir, name)		\
995 static void *deadline_##name##_fifo_start(struct seq_file *m,		\
996 					  loff_t *pos)			\
997 	__acquires(&dd->lock)						\
998 {									\
999 	struct request_queue *q = m->private;				\
1000 	struct deadline_data *dd = q->elevator->elevator_data;		\
1001 	struct dd_per_prio *per_prio = &dd->per_prio[prio];		\
1002 									\
1003 	spin_lock(&dd->lock);						\
1004 	return seq_list_start(&per_prio->fifo_list[data_dir], *pos);	\
1005 }									\
1006 									\
1007 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v,	\
1008 					 loff_t *pos)			\
1009 {									\
1010 	struct request_queue *q = m->private;				\
1011 	struct deadline_data *dd = q->elevator->elevator_data;		\
1012 	struct dd_per_prio *per_prio = &dd->per_prio[prio];		\
1013 									\
1014 	return seq_list_next(v, &per_prio->fifo_list[data_dir], pos);	\
1015 }									\
1016 									\
1017 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v)	\
1018 	__releases(&dd->lock)						\
1019 {									\
1020 	struct request_queue *q = m->private;				\
1021 	struct deadline_data *dd = q->elevator->elevator_data;		\
1022 									\
1023 	spin_unlock(&dd->lock);						\
1024 }									\
1025 									\
1026 static const struct seq_operations deadline_##name##_fifo_seq_ops = {	\
1027 	.start	= deadline_##name##_fifo_start,				\
1028 	.next	= deadline_##name##_fifo_next,				\
1029 	.stop	= deadline_##name##_fifo_stop,				\
1030 	.show	= blk_mq_debugfs_rq_show,				\
1031 };									\
1032 									\
1033 static int deadline_##name##_next_rq_show(void *data,			\
1034 					  struct seq_file *m)		\
1035 {									\
1036 	struct request_queue *q = data;					\
1037 	struct deadline_data *dd = q->elevator->elevator_data;		\
1038 	struct dd_per_prio *per_prio = &dd->per_prio[prio];		\
1039 	struct request *rq = per_prio->next_rq[data_dir];		\
1040 									\
1041 	if (rq)								\
1042 		__blk_mq_debugfs_rq_show(m, rq);			\
1043 	return 0;							\
1044 }
1045 
1046 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_READ, read0);
1047 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO, DD_WRITE, write0);
1048 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_READ, read1);
1049 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO, DD_WRITE, write1);
1050 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_READ, read2);
1051 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO, DD_WRITE, write2);
1052 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
1053 
1054 static int deadline_batching_show(void *data, struct seq_file *m)
1055 {
1056 	struct request_queue *q = data;
1057 	struct deadline_data *dd = q->elevator->elevator_data;
1058 
1059 	seq_printf(m, "%u\n", dd->batching);
1060 	return 0;
1061 }
1062 
1063 static int deadline_starved_show(void *data, struct seq_file *m)
1064 {
1065 	struct request_queue *q = data;
1066 	struct deadline_data *dd = q->elevator->elevator_data;
1067 
1068 	seq_printf(m, "%u\n", dd->starved);
1069 	return 0;
1070 }
1071 
1072 static int dd_async_depth_show(void *data, struct seq_file *m)
1073 {
1074 	struct request_queue *q = data;
1075 	struct deadline_data *dd = q->elevator->elevator_data;
1076 
1077 	seq_printf(m, "%u\n", dd->async_depth);
1078 	return 0;
1079 }
1080 
1081 static int dd_queued_show(void *data, struct seq_file *m)
1082 {
1083 	struct request_queue *q = data;
1084 	struct deadline_data *dd = q->elevator->elevator_data;
1085 	u32 rt, be, idle;
1086 
1087 	spin_lock(&dd->lock);
1088 	rt = dd_queued(dd, DD_RT_PRIO);
1089 	be = dd_queued(dd, DD_BE_PRIO);
1090 	idle = dd_queued(dd, DD_IDLE_PRIO);
1091 	spin_unlock(&dd->lock);
1092 
1093 	seq_printf(m, "%u %u %u\n", rt, be, idle);
1094 
1095 	return 0;
1096 }
1097 
1098 /* Number of requests owned by the block driver for a given priority. */
1099 static u32 dd_owned_by_driver(struct deadline_data *dd, enum dd_prio prio)
1100 {
1101 	const struct io_stats_per_prio *stats = &dd->per_prio[prio].stats;
1102 
1103 	lockdep_assert_held(&dd->lock);
1104 
1105 	return stats->dispatched + stats->merged -
1106 		atomic_read(&stats->completed);
1107 }
1108 
1109 static int dd_owned_by_driver_show(void *data, struct seq_file *m)
1110 {
1111 	struct request_queue *q = data;
1112 	struct deadline_data *dd = q->elevator->elevator_data;
1113 	u32 rt, be, idle;
1114 
1115 	spin_lock(&dd->lock);
1116 	rt = dd_owned_by_driver(dd, DD_RT_PRIO);
1117 	be = dd_owned_by_driver(dd, DD_BE_PRIO);
1118 	idle = dd_owned_by_driver(dd, DD_IDLE_PRIO);
1119 	spin_unlock(&dd->lock);
1120 
1121 	seq_printf(m, "%u %u %u\n", rt, be, idle);
1122 
1123 	return 0;
1124 }
1125 
1126 #define DEADLINE_DISPATCH_ATTR(prio)					\
1127 static void *deadline_dispatch##prio##_start(struct seq_file *m,	\
1128 					     loff_t *pos)		\
1129 	__acquires(&dd->lock)						\
1130 {									\
1131 	struct request_queue *q = m->private;				\
1132 	struct deadline_data *dd = q->elevator->elevator_data;		\
1133 	struct dd_per_prio *per_prio = &dd->per_prio[prio];		\
1134 									\
1135 	spin_lock(&dd->lock);						\
1136 	return seq_list_start(&per_prio->dispatch, *pos);		\
1137 }									\
1138 									\
1139 static void *deadline_dispatch##prio##_next(struct seq_file *m,		\
1140 					    void *v, loff_t *pos)	\
1141 {									\
1142 	struct request_queue *q = m->private;				\
1143 	struct deadline_data *dd = q->elevator->elevator_data;		\
1144 	struct dd_per_prio *per_prio = &dd->per_prio[prio];		\
1145 									\
1146 	return seq_list_next(v, &per_prio->dispatch, pos);		\
1147 }									\
1148 									\
1149 static void deadline_dispatch##prio##_stop(struct seq_file *m, void *v)	\
1150 	__releases(&dd->lock)						\
1151 {									\
1152 	struct request_queue *q = m->private;				\
1153 	struct deadline_data *dd = q->elevator->elevator_data;		\
1154 									\
1155 	spin_unlock(&dd->lock);						\
1156 }									\
1157 									\
1158 static const struct seq_operations deadline_dispatch##prio##_seq_ops = { \
1159 	.start	= deadline_dispatch##prio##_start,			\
1160 	.next	= deadline_dispatch##prio##_next,			\
1161 	.stop	= deadline_dispatch##prio##_stop,			\
1162 	.show	= blk_mq_debugfs_rq_show,				\
1163 }
1164 
1165 DEADLINE_DISPATCH_ATTR(0);
1166 DEADLINE_DISPATCH_ATTR(1);
1167 DEADLINE_DISPATCH_ATTR(2);
1168 #undef DEADLINE_DISPATCH_ATTR
1169 
1170 #define DEADLINE_QUEUE_DDIR_ATTRS(name)					\
1171 	{#name "_fifo_list", 0400,					\
1172 			.seq_ops = &deadline_##name##_fifo_seq_ops}
1173 #define DEADLINE_NEXT_RQ_ATTR(name)					\
1174 	{#name "_next_rq", 0400, deadline_##name##_next_rq_show}
1175 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
1176 	DEADLINE_QUEUE_DDIR_ATTRS(read0),
1177 	DEADLINE_QUEUE_DDIR_ATTRS(write0),
1178 	DEADLINE_QUEUE_DDIR_ATTRS(read1),
1179 	DEADLINE_QUEUE_DDIR_ATTRS(write1),
1180 	DEADLINE_QUEUE_DDIR_ATTRS(read2),
1181 	DEADLINE_QUEUE_DDIR_ATTRS(write2),
1182 	DEADLINE_NEXT_RQ_ATTR(read0),
1183 	DEADLINE_NEXT_RQ_ATTR(write0),
1184 	DEADLINE_NEXT_RQ_ATTR(read1),
1185 	DEADLINE_NEXT_RQ_ATTR(write1),
1186 	DEADLINE_NEXT_RQ_ATTR(read2),
1187 	DEADLINE_NEXT_RQ_ATTR(write2),
1188 	{"batching", 0400, deadline_batching_show},
1189 	{"starved", 0400, deadline_starved_show},
1190 	{"async_depth", 0400, dd_async_depth_show},
1191 	{"dispatch0", 0400, .seq_ops = &deadline_dispatch0_seq_ops},
1192 	{"dispatch1", 0400, .seq_ops = &deadline_dispatch1_seq_ops},
1193 	{"dispatch2", 0400, .seq_ops = &deadline_dispatch2_seq_ops},
1194 	{"owned_by_driver", 0400, dd_owned_by_driver_show},
1195 	{"queued", 0400, dd_queued_show},
1196 	{},
1197 };
1198 #undef DEADLINE_QUEUE_DDIR_ATTRS
1199 #endif
1200 
1201 static struct elevator_type mq_deadline = {
1202 	.ops = {
1203 		.depth_updated		= dd_depth_updated,
1204 		.limit_depth		= dd_limit_depth,
1205 		.insert_requests	= dd_insert_requests,
1206 		.dispatch_request	= dd_dispatch_request,
1207 		.prepare_request	= dd_prepare_request,
1208 		.finish_request		= dd_finish_request,
1209 		.next_request		= elv_rb_latter_request,
1210 		.former_request		= elv_rb_former_request,
1211 		.bio_merge		= dd_bio_merge,
1212 		.request_merge		= dd_request_merge,
1213 		.requests_merged	= dd_merged_requests,
1214 		.request_merged		= dd_request_merged,
1215 		.has_work		= dd_has_work,
1216 		.init_sched		= dd_init_sched,
1217 		.exit_sched		= dd_exit_sched,
1218 		.init_hctx		= dd_init_hctx,
1219 	},
1220 
1221 #ifdef CONFIG_BLK_DEBUG_FS
1222 	.queue_debugfs_attrs = deadline_queue_debugfs_attrs,
1223 #endif
1224 	.elevator_attrs = deadline_attrs,
1225 	.elevator_name = "mq-deadline",
1226 	.elevator_alias = "deadline",
1227 	.elevator_features = ELEVATOR_F_ZBD_SEQ_WRITE,
1228 	.elevator_owner = THIS_MODULE,
1229 };
1230 MODULE_ALIAS("mq-deadline-iosched");
1231 
1232 static int __init deadline_init(void)
1233 {
1234 	return elv_register(&mq_deadline);
1235 }
1236 
1237 static void __exit deadline_exit(void)
1238 {
1239 	elv_unregister(&mq_deadline);
1240 }
1241 
1242 module_init(deadline_init);
1243 module_exit(deadline_exit);
1244 
1245 MODULE_AUTHOR("Jens Axboe, Damien Le Moal and Bart Van Assche");
1246 MODULE_LICENSE("GPL");
1247 MODULE_DESCRIPTION("MQ deadline IO scheduler");
1248