quota: Push dqio_sem down to ->read_dqblk()
[sfrench/cifs-2.6.git] / block / mq-deadline.c
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");