reset: Create subdirectory for StarFive drivers
[platform/kernel/linux-starfive.git] / block / mq-deadline.c
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");