xref: /openbmc/linux/block/mq-deadline.c (revision fcc8487d)
1 /*
2  *  MQ Deadline i/o scheduler - adaptation of the legacy deadline scheduler,
3  *  for the blk-mq scheduling framework
4  *
5  *  Copyright (C) 2016 Jens Axboe <axboe@kernel.dk>
6  */
7 #include <linux/kernel.h>
8 #include <linux/fs.h>
9 #include <linux/blkdev.h>
10 #include <linux/blk-mq.h>
11 #include <linux/elevator.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 "blk.h"
21 #include "blk-mq.h"
22 #include "blk-mq-debugfs.h"
23 #include "blk-mq-tag.h"
24 #include "blk-mq-sched.h"
25 
26 /*
27  * See Documentation/block/deadline-iosched.txt
28  */
29 static const int read_expire = HZ / 2;  /* max time before a read is submitted. */
30 static const int write_expire = 5 * HZ; /* ditto for writes, these limits are SOFT! */
31 static const int writes_starved = 2;    /* max times reads can starve a write */
32 static const int fifo_batch = 16;       /* # of sequential requests treated as one
33 				     by the above parameters. For throughput. */
34 
35 struct deadline_data {
36 	/*
37 	 * run time data
38 	 */
39 
40 	/*
41 	 * requests (deadline_rq s) are present on both sort_list and fifo_list
42 	 */
43 	struct rb_root sort_list[2];
44 	struct list_head fifo_list[2];
45 
46 	/*
47 	 * next in sort order. read, write or both are NULL
48 	 */
49 	struct request *next_rq[2];
50 	unsigned int batching;		/* number of sequential requests made */
51 	unsigned int starved;		/* times reads have starved writes */
52 
53 	/*
54 	 * settings that change how the i/o scheduler behaves
55 	 */
56 	int fifo_expire[2];
57 	int fifo_batch;
58 	int writes_starved;
59 	int front_merges;
60 
61 	spinlock_t lock;
62 	struct list_head dispatch;
63 };
64 
65 static inline struct rb_root *
66 deadline_rb_root(struct deadline_data *dd, struct request *rq)
67 {
68 	return &dd->sort_list[rq_data_dir(rq)];
69 }
70 
71 /*
72  * get the request after `rq' in sector-sorted order
73  */
74 static inline struct request *
75 deadline_latter_request(struct request *rq)
76 {
77 	struct rb_node *node = rb_next(&rq->rb_node);
78 
79 	if (node)
80 		return rb_entry_rq(node);
81 
82 	return NULL;
83 }
84 
85 static void
86 deadline_add_rq_rb(struct deadline_data *dd, struct request *rq)
87 {
88 	struct rb_root *root = deadline_rb_root(dd, rq);
89 
90 	elv_rb_add(root, rq);
91 }
92 
93 static inline void
94 deadline_del_rq_rb(struct deadline_data *dd, struct request *rq)
95 {
96 	const int data_dir = rq_data_dir(rq);
97 
98 	if (dd->next_rq[data_dir] == rq)
99 		dd->next_rq[data_dir] = deadline_latter_request(rq);
100 
101 	elv_rb_del(deadline_rb_root(dd, rq), rq);
102 }
103 
104 /*
105  * remove rq from rbtree and fifo.
106  */
107 static void deadline_remove_request(struct request_queue *q, struct request *rq)
108 {
109 	struct deadline_data *dd = q->elevator->elevator_data;
110 
111 	list_del_init(&rq->queuelist);
112 
113 	/*
114 	 * We might not be on the rbtree, if we are doing an insert merge
115 	 */
116 	if (!RB_EMPTY_NODE(&rq->rb_node))
117 		deadline_del_rq_rb(dd, rq);
118 
119 	elv_rqhash_del(q, rq);
120 	if (q->last_merge == rq)
121 		q->last_merge = NULL;
122 }
123 
124 static void dd_request_merged(struct request_queue *q, struct request *req,
125 			      enum elv_merge type)
126 {
127 	struct deadline_data *dd = q->elevator->elevator_data;
128 
129 	/*
130 	 * if the merge was a front merge, we need to reposition request
131 	 */
132 	if (type == ELEVATOR_FRONT_MERGE) {
133 		elv_rb_del(deadline_rb_root(dd, req), req);
134 		deadline_add_rq_rb(dd, req);
135 	}
136 }
137 
138 static void dd_merged_requests(struct request_queue *q, struct request *req,
139 			       struct request *next)
140 {
141 	/*
142 	 * if next expires before rq, assign its expire time to rq
143 	 * and move into next position (next will be deleted) in fifo
144 	 */
145 	if (!list_empty(&req->queuelist) && !list_empty(&next->queuelist)) {
146 		if (time_before((unsigned long)next->fifo_time,
147 				(unsigned long)req->fifo_time)) {
148 			list_move(&req->queuelist, &next->queuelist);
149 			req->fifo_time = next->fifo_time;
150 		}
151 	}
152 
153 	/*
154 	 * kill knowledge of next, this one is a goner
155 	 */
156 	deadline_remove_request(q, next);
157 }
158 
159 /*
160  * move an entry to dispatch queue
161  */
162 static void
163 deadline_move_request(struct deadline_data *dd, struct request *rq)
164 {
165 	const int data_dir = rq_data_dir(rq);
166 
167 	dd->next_rq[READ] = NULL;
168 	dd->next_rq[WRITE] = NULL;
169 	dd->next_rq[data_dir] = deadline_latter_request(rq);
170 
171 	/*
172 	 * take it off the sort and fifo list
173 	 */
174 	deadline_remove_request(rq->q, rq);
175 }
176 
177 /*
178  * deadline_check_fifo returns 0 if there are no expired requests on the fifo,
179  * 1 otherwise. Requires !list_empty(&dd->fifo_list[data_dir])
180  */
181 static inline int deadline_check_fifo(struct deadline_data *dd, int ddir)
182 {
183 	struct request *rq = rq_entry_fifo(dd->fifo_list[ddir].next);
184 
185 	/*
186 	 * rq is expired!
187 	 */
188 	if (time_after_eq(jiffies, (unsigned long)rq->fifo_time))
189 		return 1;
190 
191 	return 0;
192 }
193 
194 /*
195  * deadline_dispatch_requests selects the best request according to
196  * read/write expire, fifo_batch, etc
197  */
198 static struct request *__dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
199 {
200 	struct deadline_data *dd = hctx->queue->elevator->elevator_data;
201 	struct request *rq;
202 	bool reads, writes;
203 	int data_dir;
204 
205 	if (!list_empty(&dd->dispatch)) {
206 		rq = list_first_entry(&dd->dispatch, struct request, queuelist);
207 		list_del_init(&rq->queuelist);
208 		goto done;
209 	}
210 
211 	reads = !list_empty(&dd->fifo_list[READ]);
212 	writes = !list_empty(&dd->fifo_list[WRITE]);
213 
214 	/*
215 	 * batches are currently reads XOR writes
216 	 */
217 	if (dd->next_rq[WRITE])
218 		rq = dd->next_rq[WRITE];
219 	else
220 		rq = dd->next_rq[READ];
221 
222 	if (rq && dd->batching < dd->fifo_batch)
223 		/* we have a next request are still entitled to batch */
224 		goto dispatch_request;
225 
226 	/*
227 	 * at this point we are not running a batch. select the appropriate
228 	 * data direction (read / write)
229 	 */
230 
231 	if (reads) {
232 		BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[READ]));
233 
234 		if (writes && (dd->starved++ >= dd->writes_starved))
235 			goto dispatch_writes;
236 
237 		data_dir = READ;
238 
239 		goto dispatch_find_request;
240 	}
241 
242 	/*
243 	 * there are either no reads or writes have been starved
244 	 */
245 
246 	if (writes) {
247 dispatch_writes:
248 		BUG_ON(RB_EMPTY_ROOT(&dd->sort_list[WRITE]));
249 
250 		dd->starved = 0;
251 
252 		data_dir = WRITE;
253 
254 		goto dispatch_find_request;
255 	}
256 
257 	return NULL;
258 
259 dispatch_find_request:
260 	/*
261 	 * we are not running a batch, find best request for selected data_dir
262 	 */
263 	if (deadline_check_fifo(dd, data_dir) || !dd->next_rq[data_dir]) {
264 		/*
265 		 * A deadline has expired, the last request was in the other
266 		 * direction, or we have run out of higher-sectored requests.
267 		 * Start again from the request with the earliest expiry time.
268 		 */
269 		rq = rq_entry_fifo(dd->fifo_list[data_dir].next);
270 	} else {
271 		/*
272 		 * The last req was the same dir and we have a next request in
273 		 * sort order. No expired requests so continue on from here.
274 		 */
275 		rq = dd->next_rq[data_dir];
276 	}
277 
278 	dd->batching = 0;
279 
280 dispatch_request:
281 	/*
282 	 * rq is the selected appropriate request.
283 	 */
284 	dd->batching++;
285 	deadline_move_request(dd, rq);
286 done:
287 	rq->rq_flags |= RQF_STARTED;
288 	return rq;
289 }
290 
291 static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
292 {
293 	struct deadline_data *dd = hctx->queue->elevator->elevator_data;
294 	struct request *rq;
295 
296 	spin_lock(&dd->lock);
297 	rq = __dd_dispatch_request(hctx);
298 	spin_unlock(&dd->lock);
299 
300 	return rq;
301 }
302 
303 static void dd_exit_queue(struct elevator_queue *e)
304 {
305 	struct deadline_data *dd = e->elevator_data;
306 
307 	BUG_ON(!list_empty(&dd->fifo_list[READ]));
308 	BUG_ON(!list_empty(&dd->fifo_list[WRITE]));
309 
310 	kfree(dd);
311 }
312 
313 /*
314  * initialize elevator private data (deadline_data).
315  */
316 static int dd_init_queue(struct request_queue *q, struct elevator_type *e)
317 {
318 	struct deadline_data *dd;
319 	struct elevator_queue *eq;
320 
321 	eq = elevator_alloc(q, e);
322 	if (!eq)
323 		return -ENOMEM;
324 
325 	dd = kzalloc_node(sizeof(*dd), GFP_KERNEL, q->node);
326 	if (!dd) {
327 		kobject_put(&eq->kobj);
328 		return -ENOMEM;
329 	}
330 	eq->elevator_data = dd;
331 
332 	INIT_LIST_HEAD(&dd->fifo_list[READ]);
333 	INIT_LIST_HEAD(&dd->fifo_list[WRITE]);
334 	dd->sort_list[READ] = RB_ROOT;
335 	dd->sort_list[WRITE] = RB_ROOT;
336 	dd->fifo_expire[READ] = read_expire;
337 	dd->fifo_expire[WRITE] = write_expire;
338 	dd->writes_starved = writes_starved;
339 	dd->front_merges = 1;
340 	dd->fifo_batch = fifo_batch;
341 	spin_lock_init(&dd->lock);
342 	INIT_LIST_HEAD(&dd->dispatch);
343 
344 	q->elevator = eq;
345 	return 0;
346 }
347 
348 static int dd_request_merge(struct request_queue *q, struct request **rq,
349 			    struct bio *bio)
350 {
351 	struct deadline_data *dd = q->elevator->elevator_data;
352 	sector_t sector = bio_end_sector(bio);
353 	struct request *__rq;
354 
355 	if (!dd->front_merges)
356 		return ELEVATOR_NO_MERGE;
357 
358 	__rq = elv_rb_find(&dd->sort_list[bio_data_dir(bio)], sector);
359 	if (__rq) {
360 		BUG_ON(sector != blk_rq_pos(__rq));
361 
362 		if (elv_bio_merge_ok(__rq, bio)) {
363 			*rq = __rq;
364 			return ELEVATOR_FRONT_MERGE;
365 		}
366 	}
367 
368 	return ELEVATOR_NO_MERGE;
369 }
370 
371 static bool dd_bio_merge(struct blk_mq_hw_ctx *hctx, struct bio *bio)
372 {
373 	struct request_queue *q = hctx->queue;
374 	struct deadline_data *dd = q->elevator->elevator_data;
375 	struct request *free = NULL;
376 	bool ret;
377 
378 	spin_lock(&dd->lock);
379 	ret = blk_mq_sched_try_merge(q, bio, &free);
380 	spin_unlock(&dd->lock);
381 
382 	if (free)
383 		blk_mq_free_request(free);
384 
385 	return ret;
386 }
387 
388 /*
389  * add rq to rbtree and fifo
390  */
391 static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
392 			      bool at_head)
393 {
394 	struct request_queue *q = hctx->queue;
395 	struct deadline_data *dd = q->elevator->elevator_data;
396 	const int data_dir = rq_data_dir(rq);
397 
398 	if (blk_mq_sched_try_insert_merge(q, rq))
399 		return;
400 
401 	blk_mq_sched_request_inserted(rq);
402 
403 	if (at_head || blk_rq_is_passthrough(rq)) {
404 		if (at_head)
405 			list_add(&rq->queuelist, &dd->dispatch);
406 		else
407 			list_add_tail(&rq->queuelist, &dd->dispatch);
408 	} else {
409 		deadline_add_rq_rb(dd, rq);
410 
411 		if (rq_mergeable(rq)) {
412 			elv_rqhash_add(q, rq);
413 			if (!q->last_merge)
414 				q->last_merge = rq;
415 		}
416 
417 		/*
418 		 * set expire time and add to fifo list
419 		 */
420 		rq->fifo_time = jiffies + dd->fifo_expire[data_dir];
421 		list_add_tail(&rq->queuelist, &dd->fifo_list[data_dir]);
422 	}
423 }
424 
425 static void dd_insert_requests(struct blk_mq_hw_ctx *hctx,
426 			       struct list_head *list, bool at_head)
427 {
428 	struct request_queue *q = hctx->queue;
429 	struct deadline_data *dd = q->elevator->elevator_data;
430 
431 	spin_lock(&dd->lock);
432 	while (!list_empty(list)) {
433 		struct request *rq;
434 
435 		rq = list_first_entry(list, struct request, queuelist);
436 		list_del_init(&rq->queuelist);
437 		dd_insert_request(hctx, rq, at_head);
438 	}
439 	spin_unlock(&dd->lock);
440 }
441 
442 static bool dd_has_work(struct blk_mq_hw_ctx *hctx)
443 {
444 	struct deadline_data *dd = hctx->queue->elevator->elevator_data;
445 
446 	return !list_empty_careful(&dd->dispatch) ||
447 		!list_empty_careful(&dd->fifo_list[0]) ||
448 		!list_empty_careful(&dd->fifo_list[1]);
449 }
450 
451 /*
452  * sysfs parts below
453  */
454 static ssize_t
455 deadline_var_show(int var, char *page)
456 {
457 	return sprintf(page, "%d\n", var);
458 }
459 
460 static ssize_t
461 deadline_var_store(int *var, const char *page, size_t count)
462 {
463 	char *p = (char *) page;
464 
465 	*var = simple_strtol(p, &p, 10);
466 	return count;
467 }
468 
469 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)				\
470 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
471 {									\
472 	struct deadline_data *dd = e->elevator_data;			\
473 	int __data = __VAR;						\
474 	if (__CONV)							\
475 		__data = jiffies_to_msecs(__data);			\
476 	return deadline_var_show(__data, (page));			\
477 }
478 SHOW_FUNCTION(deadline_read_expire_show, dd->fifo_expire[READ], 1);
479 SHOW_FUNCTION(deadline_write_expire_show, dd->fifo_expire[WRITE], 1);
480 SHOW_FUNCTION(deadline_writes_starved_show, dd->writes_starved, 0);
481 SHOW_FUNCTION(deadline_front_merges_show, dd->front_merges, 0);
482 SHOW_FUNCTION(deadline_fifo_batch_show, dd->fifo_batch, 0);
483 #undef SHOW_FUNCTION
484 
485 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
486 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)	\
487 {									\
488 	struct deadline_data *dd = e->elevator_data;			\
489 	int __data;							\
490 	int ret = deadline_var_store(&__data, (page), count);		\
491 	if (__data < (MIN))						\
492 		__data = (MIN);						\
493 	else if (__data > (MAX))					\
494 		__data = (MAX);						\
495 	if (__CONV)							\
496 		*(__PTR) = msecs_to_jiffies(__data);			\
497 	else								\
498 		*(__PTR) = __data;					\
499 	return ret;							\
500 }
501 STORE_FUNCTION(deadline_read_expire_store, &dd->fifo_expire[READ], 0, INT_MAX, 1);
502 STORE_FUNCTION(deadline_write_expire_store, &dd->fifo_expire[WRITE], 0, INT_MAX, 1);
503 STORE_FUNCTION(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX, 0);
504 STORE_FUNCTION(deadline_front_merges_store, &dd->front_merges, 0, 1, 0);
505 STORE_FUNCTION(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX, 0);
506 #undef STORE_FUNCTION
507 
508 #define DD_ATTR(name) \
509 	__ATTR(name, S_IRUGO|S_IWUSR, deadline_##name##_show, \
510 				      deadline_##name##_store)
511 
512 static struct elv_fs_entry deadline_attrs[] = {
513 	DD_ATTR(read_expire),
514 	DD_ATTR(write_expire),
515 	DD_ATTR(writes_starved),
516 	DD_ATTR(front_merges),
517 	DD_ATTR(fifo_batch),
518 	__ATTR_NULL
519 };
520 
521 #ifdef CONFIG_BLK_DEBUG_FS
522 #define DEADLINE_DEBUGFS_DDIR_ATTRS(ddir, name)				\
523 static void *deadline_##name##_fifo_start(struct seq_file *m,		\
524 					  loff_t *pos)			\
525 	__acquires(&dd->lock)						\
526 {									\
527 	struct request_queue *q = m->private;				\
528 	struct deadline_data *dd = q->elevator->elevator_data;		\
529 									\
530 	spin_lock(&dd->lock);						\
531 	return seq_list_start(&dd->fifo_list[ddir], *pos);		\
532 }									\
533 									\
534 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v,	\
535 					 loff_t *pos)			\
536 {									\
537 	struct request_queue *q = m->private;				\
538 	struct deadline_data *dd = q->elevator->elevator_data;		\
539 									\
540 	return seq_list_next(v, &dd->fifo_list[ddir], pos);		\
541 }									\
542 									\
543 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v)	\
544 	__releases(&dd->lock)						\
545 {									\
546 	struct request_queue *q = m->private;				\
547 	struct deadline_data *dd = q->elevator->elevator_data;		\
548 									\
549 	spin_unlock(&dd->lock);						\
550 }									\
551 									\
552 static const struct seq_operations deadline_##name##_fifo_seq_ops = {	\
553 	.start	= deadline_##name##_fifo_start,				\
554 	.next	= deadline_##name##_fifo_next,				\
555 	.stop	= deadline_##name##_fifo_stop,				\
556 	.show	= blk_mq_debugfs_rq_show,				\
557 };									\
558 									\
559 static int deadline_##name##_next_rq_show(void *data,			\
560 					  struct seq_file *m)		\
561 {									\
562 	struct request_queue *q = data;					\
563 	struct deadline_data *dd = q->elevator->elevator_data;		\
564 	struct request *rq = dd->next_rq[ddir];				\
565 									\
566 	if (rq)								\
567 		__blk_mq_debugfs_rq_show(m, rq);			\
568 	return 0;							\
569 }
570 DEADLINE_DEBUGFS_DDIR_ATTRS(READ, read)
571 DEADLINE_DEBUGFS_DDIR_ATTRS(WRITE, write)
572 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
573 
574 static int deadline_batching_show(void *data, struct seq_file *m)
575 {
576 	struct request_queue *q = data;
577 	struct deadline_data *dd = q->elevator->elevator_data;
578 
579 	seq_printf(m, "%u\n", dd->batching);
580 	return 0;
581 }
582 
583 static int deadline_starved_show(void *data, struct seq_file *m)
584 {
585 	struct request_queue *q = data;
586 	struct deadline_data *dd = q->elevator->elevator_data;
587 
588 	seq_printf(m, "%u\n", dd->starved);
589 	return 0;
590 }
591 
592 static void *deadline_dispatch_start(struct seq_file *m, loff_t *pos)
593 	__acquires(&dd->lock)
594 {
595 	struct request_queue *q = m->private;
596 	struct deadline_data *dd = q->elevator->elevator_data;
597 
598 	spin_lock(&dd->lock);
599 	return seq_list_start(&dd->dispatch, *pos);
600 }
601 
602 static void *deadline_dispatch_next(struct seq_file *m, void *v, loff_t *pos)
603 {
604 	struct request_queue *q = m->private;
605 	struct deadline_data *dd = q->elevator->elevator_data;
606 
607 	return seq_list_next(v, &dd->dispatch, pos);
608 }
609 
610 static void deadline_dispatch_stop(struct seq_file *m, void *v)
611 	__releases(&dd->lock)
612 {
613 	struct request_queue *q = m->private;
614 	struct deadline_data *dd = q->elevator->elevator_data;
615 
616 	spin_unlock(&dd->lock);
617 }
618 
619 static const struct seq_operations deadline_dispatch_seq_ops = {
620 	.start	= deadline_dispatch_start,
621 	.next	= deadline_dispatch_next,
622 	.stop	= deadline_dispatch_stop,
623 	.show	= blk_mq_debugfs_rq_show,
624 };
625 
626 #define DEADLINE_QUEUE_DDIR_ATTRS(name)						\
627 	{#name "_fifo_list", 0400, .seq_ops = &deadline_##name##_fifo_seq_ops},	\
628 	{#name "_next_rq", 0400, deadline_##name##_next_rq_show}
629 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
630 	DEADLINE_QUEUE_DDIR_ATTRS(read),
631 	DEADLINE_QUEUE_DDIR_ATTRS(write),
632 	{"batching", 0400, deadline_batching_show},
633 	{"starved", 0400, deadline_starved_show},
634 	{"dispatch", 0400, .seq_ops = &deadline_dispatch_seq_ops},
635 	{},
636 };
637 #undef DEADLINE_QUEUE_DDIR_ATTRS
638 #endif
639 
640 static struct elevator_type mq_deadline = {
641 	.ops.mq = {
642 		.insert_requests	= dd_insert_requests,
643 		.dispatch_request	= dd_dispatch_request,
644 		.next_request		= elv_rb_latter_request,
645 		.former_request		= elv_rb_former_request,
646 		.bio_merge		= dd_bio_merge,
647 		.request_merge		= dd_request_merge,
648 		.requests_merged	= dd_merged_requests,
649 		.request_merged		= dd_request_merged,
650 		.has_work		= dd_has_work,
651 		.init_sched		= dd_init_queue,
652 		.exit_sched		= dd_exit_queue,
653 	},
654 
655 	.uses_mq	= true,
656 #ifdef CONFIG_BLK_DEBUG_FS
657 	.queue_debugfs_attrs = deadline_queue_debugfs_attrs,
658 #endif
659 	.elevator_attrs = deadline_attrs,
660 	.elevator_name = "mq-deadline",
661 	.elevator_owner = THIS_MODULE,
662 };
663 
664 static int __init deadline_init(void)
665 {
666 	return elv_register(&mq_deadline);
667 }
668 
669 static void __exit deadline_exit(void)
670 {
671 	elv_unregister(&mq_deadline);
672 }
673 
674 module_init(deadline_init);
675 module_exit(deadline_exit);
676 
677 MODULE_AUTHOR("Jens Axboe");
678 MODULE_LICENSE("GPL");
679 MODULE_DESCRIPTION("MQ deadline IO scheduler");
680