1 // SPDX-License-Identifier: GPL-2.0
3 * MQ Deadline i/o scheduler - adaptation of the legacy deadline scheduler,
4 * for the blk-mq scheduling framework
6 * Copyright (C) 2016 Jens Axboe <axboe@kernel.dk>
8 #include <linux/kernel.h>
10 #include <linux/blkdev.h>
11 #include <linux/bio.h>
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/init.h>
15 #include <linux/compiler.h>
16 #include <linux/rbtree.h>
17 #include <linux/sbitmap.h>
19 #include <trace/events/block.h>
24 #include "blk-mq-debugfs.h"
25 #include "blk-mq-sched.h"
28 * See Documentation/block/deadline-iosched.rst
30 static const int read_expire
= HZ
/ 2; /* max time before a read is submitted. */
31 static const int write_expire
= 5 * HZ
; /* ditto for writes, these limits are SOFT! */
33 * Time after which to dispatch lower priority requests even if higher
34 * priority requests are pending.
36 static const int prio_aging_expire
= 10 * HZ
;
37 static const int writes_starved
= 2; /* max times reads can starve a write */
38 static const int fifo_batch
= 16; /* # of sequential requests treated as one
39 by the above parameters. For throughput. */
46 enum { DD_DIR_COUNT
= 2 };
55 enum { DD_PRIO_COUNT
= 3 };
58 * I/O statistics per I/O priority. It is fine if these counters overflow.
59 * What matters is that these counters are at least as wide as
60 * log2(max_outstanding_requests).
62 struct io_stats_per_prio
{
70 * Deadline scheduler data per I/O priority (enum dd_prio). Requests are
71 * present on both sort_list[] and fifo_list[].
74 struct list_head dispatch
;
75 struct rb_root sort_list
[DD_DIR_COUNT
];
76 struct list_head fifo_list
[DD_DIR_COUNT
];
77 /* Position of the most recently dispatched request. */
78 sector_t latest_pos
[DD_DIR_COUNT
];
79 struct io_stats_per_prio stats
;
82 struct deadline_data
{
87 struct dd_per_prio per_prio
[DD_PRIO_COUNT
];
89 /* Data direction of latest dispatched request. */
90 enum dd_data_dir last_dir
;
91 unsigned int batching
; /* number of sequential requests made */
92 unsigned int starved
; /* times reads have starved writes */
95 * settings that change how the i/o scheduler behaves
97 int fifo_expire
[DD_DIR_COUNT
];
102 int prio_aging_expire
;
105 spinlock_t zone_lock
;
108 /* Maps an I/O priority class to a deadline scheduler priority. */
109 static const enum dd_prio ioprio_class_to_prio
[] = {
110 [IOPRIO_CLASS_NONE
] = DD_BE_PRIO
,
111 [IOPRIO_CLASS_RT
] = DD_RT_PRIO
,
112 [IOPRIO_CLASS_BE
] = DD_BE_PRIO
,
113 [IOPRIO_CLASS_IDLE
] = DD_IDLE_PRIO
,
116 static inline struct rb_root
*
117 deadline_rb_root(struct dd_per_prio
*per_prio
, struct request
*rq
)
119 return &per_prio
->sort_list
[rq_data_dir(rq
)];
123 * Returns the I/O priority class (IOPRIO_CLASS_*) that has been assigned to a
126 static u8
dd_rq_ioclass(struct request
*rq
)
128 return IOPRIO_PRIO_CLASS(req_get_ioprio(rq
));
132 * get the request before `rq' in sector-sorted order
134 static inline struct request
*
135 deadline_earlier_request(struct request
*rq
)
137 struct rb_node
*node
= rb_prev(&rq
->rb_node
);
140 return rb_entry_rq(node
);
146 * get the request after `rq' in sector-sorted order
148 static inline struct request
*
149 deadline_latter_request(struct request
*rq
)
151 struct rb_node
*node
= rb_next(&rq
->rb_node
);
154 return rb_entry_rq(node
);
160 * Return the first request for which blk_rq_pos() >= @pos. For zoned devices,
161 * return the first request after the start of the zone containing @pos.
163 static inline struct request
*deadline_from_pos(struct dd_per_prio
*per_prio
,
164 enum dd_data_dir data_dir
, sector_t pos
)
166 struct rb_node
*node
= per_prio
->sort_list
[data_dir
].rb_node
;
167 struct request
*rq
, *res
= NULL
;
172 rq
= rb_entry_rq(node
);
174 * A zoned write may have been requeued with a starting position that
175 * is below that of the most recently dispatched request. Hence, for
176 * zoned writes, start searching from the start of a zone.
178 if (blk_rq_is_seq_zoned_write(rq
))
179 pos
= round_down(pos
, rq
->q
->limits
.chunk_sectors
);
182 rq
= rb_entry_rq(node
);
183 if (blk_rq_pos(rq
) >= pos
) {
185 node
= node
->rb_left
;
187 node
= node
->rb_right
;
194 deadline_add_rq_rb(struct dd_per_prio
*per_prio
, struct request
*rq
)
196 struct rb_root
*root
= deadline_rb_root(per_prio
, rq
);
198 elv_rb_add(root
, rq
);
202 deadline_del_rq_rb(struct dd_per_prio
*per_prio
, struct request
*rq
)
204 elv_rb_del(deadline_rb_root(per_prio
, rq
), rq
);
208 * remove rq from rbtree and fifo.
210 static void deadline_remove_request(struct request_queue
*q
,
211 struct dd_per_prio
*per_prio
,
214 list_del_init(&rq
->queuelist
);
217 * We might not be on the rbtree, if we are doing an insert merge
219 if (!RB_EMPTY_NODE(&rq
->rb_node
))
220 deadline_del_rq_rb(per_prio
, rq
);
222 elv_rqhash_del(q
, rq
);
223 if (q
->last_merge
== rq
)
224 q
->last_merge
= NULL
;
227 static void dd_request_merged(struct request_queue
*q
, struct request
*req
,
230 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
231 const u8 ioprio_class
= dd_rq_ioclass(req
);
232 const enum dd_prio prio
= ioprio_class_to_prio
[ioprio_class
];
233 struct dd_per_prio
*per_prio
= &dd
->per_prio
[prio
];
236 * if the merge was a front merge, we need to reposition request
238 if (type
== ELEVATOR_FRONT_MERGE
) {
239 elv_rb_del(deadline_rb_root(per_prio
, req
), req
);
240 deadline_add_rq_rb(per_prio
, req
);
245 * Callback function that is invoked after @next has been merged into @req.
247 static void dd_merged_requests(struct request_queue
*q
, struct request
*req
,
248 struct request
*next
)
250 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
251 const u8 ioprio_class
= dd_rq_ioclass(next
);
252 const enum dd_prio prio
= ioprio_class_to_prio
[ioprio_class
];
254 lockdep_assert_held(&dd
->lock
);
256 dd
->per_prio
[prio
].stats
.merged
++;
259 * if next expires before rq, assign its expire time to rq
260 * and move into next position (next will be deleted) in fifo
262 if (!list_empty(&req
->queuelist
) && !list_empty(&next
->queuelist
)) {
263 if (time_before((unsigned long)next
->fifo_time
,
264 (unsigned long)req
->fifo_time
)) {
265 list_move(&req
->queuelist
, &next
->queuelist
);
266 req
->fifo_time
= next
->fifo_time
;
271 * kill knowledge of next, this one is a goner
273 deadline_remove_request(q
, &dd
->per_prio
[prio
], next
);
277 * move an entry to dispatch queue
280 deadline_move_request(struct deadline_data
*dd
, struct dd_per_prio
*per_prio
,
284 * take it off the sort and fifo list
286 deadline_remove_request(rq
->q
, per_prio
, rq
);
289 /* Number of requests queued for a given priority level. */
290 static u32
dd_queued(struct deadline_data
*dd
, enum dd_prio prio
)
292 const struct io_stats_per_prio
*stats
= &dd
->per_prio
[prio
].stats
;
294 lockdep_assert_held(&dd
->lock
);
296 return stats
->inserted
- atomic_read(&stats
->completed
);
300 * deadline_check_fifo returns true if and only if there are expired requests
301 * in the FIFO list. Requires !list_empty(&dd->fifo_list[data_dir]).
303 static inline bool deadline_check_fifo(struct dd_per_prio
*per_prio
,
304 enum dd_data_dir data_dir
)
306 struct request
*rq
= rq_entry_fifo(per_prio
->fifo_list
[data_dir
].next
);
308 return time_is_before_eq_jiffies((unsigned long)rq
->fifo_time
);
312 * Check if rq has a sequential request preceding it.
314 static bool deadline_is_seq_write(struct deadline_data
*dd
, struct request
*rq
)
316 struct request
*prev
= deadline_earlier_request(rq
);
321 return blk_rq_pos(prev
) + blk_rq_sectors(prev
) == blk_rq_pos(rq
);
325 * Skip all write requests that are sequential from @rq, even if we cross
328 static struct request
*deadline_skip_seq_writes(struct deadline_data
*dd
,
331 sector_t pos
= blk_rq_pos(rq
);
334 pos
+= blk_rq_sectors(rq
);
335 rq
= deadline_latter_request(rq
);
336 } while (rq
&& blk_rq_pos(rq
) == pos
);
342 * For the specified data direction, return the next request to
343 * dispatch using arrival ordered lists.
345 static struct request
*
346 deadline_fifo_request(struct deadline_data
*dd
, struct dd_per_prio
*per_prio
,
347 enum dd_data_dir data_dir
)
349 struct request
*rq
, *rb_rq
, *next
;
352 if (list_empty(&per_prio
->fifo_list
[data_dir
]))
355 rq
= rq_entry_fifo(per_prio
->fifo_list
[data_dir
].next
);
356 if (data_dir
== DD_READ
|| !blk_queue_is_zoned(rq
->q
))
360 * Look for a write request that can be dispatched, that is one with
361 * an unlocked target zone. For some HDDs, breaking a sequential
362 * write stream can lead to lower throughput, so make sure to preserve
363 * sequential write streams, even if that stream crosses into the next
364 * zones and these zones are unlocked.
366 spin_lock_irqsave(&dd
->zone_lock
, flags
);
367 list_for_each_entry_safe(rq
, next
, &per_prio
->fifo_list
[DD_WRITE
],
369 /* Check whether a prior request exists for the same zone. */
370 rb_rq
= deadline_from_pos(per_prio
, data_dir
, blk_rq_pos(rq
));
371 if (rb_rq
&& blk_rq_pos(rb_rq
) < blk_rq_pos(rq
))
373 if (blk_req_can_dispatch_to_zone(rq
) &&
374 (blk_queue_nonrot(rq
->q
) ||
375 !deadline_is_seq_write(dd
, rq
)))
380 spin_unlock_irqrestore(&dd
->zone_lock
, flags
);
386 * For the specified data direction, return the next request to
387 * dispatch using sector position sorted lists.
389 static struct request
*
390 deadline_next_request(struct deadline_data
*dd
, struct dd_per_prio
*per_prio
,
391 enum dd_data_dir data_dir
)
396 rq
= deadline_from_pos(per_prio
, data_dir
,
397 per_prio
->latest_pos
[data_dir
]);
401 if (data_dir
== DD_READ
|| !blk_queue_is_zoned(rq
->q
))
405 * Look for a write request that can be dispatched, that is one with
406 * an unlocked target zone. For some HDDs, breaking a sequential
407 * write stream can lead to lower throughput, so make sure to preserve
408 * sequential write streams, even if that stream crosses into the next
409 * zones and these zones are unlocked.
411 spin_lock_irqsave(&dd
->zone_lock
, flags
);
413 if (blk_req_can_dispatch_to_zone(rq
))
415 if (blk_queue_nonrot(rq
->q
))
416 rq
= deadline_latter_request(rq
);
418 rq
= deadline_skip_seq_writes(dd
, rq
);
420 spin_unlock_irqrestore(&dd
->zone_lock
, flags
);
426 * Returns true if and only if @rq started after @latest_start where
427 * @latest_start is in jiffies.
429 static bool started_after(struct deadline_data
*dd
, struct request
*rq
,
430 unsigned long latest_start
)
432 unsigned long start_time
= (unsigned long)rq
->fifo_time
;
434 start_time
-= dd
->fifo_expire
[rq_data_dir(rq
)];
436 return time_after(start_time
, latest_start
);
440 * deadline_dispatch_requests selects the best request according to
441 * read/write expire, fifo_batch, etc and with a start time <= @latest_start.
443 static struct request
*__dd_dispatch_request(struct deadline_data
*dd
,
444 struct dd_per_prio
*per_prio
,
445 unsigned long latest_start
)
447 struct request
*rq
, *next_rq
;
448 enum dd_data_dir data_dir
;
452 lockdep_assert_held(&dd
->lock
);
454 if (!list_empty(&per_prio
->dispatch
)) {
455 rq
= list_first_entry(&per_prio
->dispatch
, struct request
,
457 if (started_after(dd
, rq
, latest_start
))
459 list_del_init(&rq
->queuelist
);
460 data_dir
= rq_data_dir(rq
);
465 * batches are currently reads XOR writes
467 rq
= deadline_next_request(dd
, per_prio
, dd
->last_dir
);
468 if (rq
&& dd
->batching
< dd
->fifo_batch
) {
469 /* we have a next request and are still entitled to batch */
470 data_dir
= rq_data_dir(rq
);
471 goto dispatch_request
;
475 * at this point we are not running a batch. select the appropriate
476 * data direction (read / write)
479 if (!list_empty(&per_prio
->fifo_list
[DD_READ
])) {
480 BUG_ON(RB_EMPTY_ROOT(&per_prio
->sort_list
[DD_READ
]));
482 if (deadline_fifo_request(dd
, per_prio
, DD_WRITE
) &&
483 (dd
->starved
++ >= dd
->writes_starved
))
484 goto dispatch_writes
;
488 goto dispatch_find_request
;
492 * there are either no reads or writes have been starved
495 if (!list_empty(&per_prio
->fifo_list
[DD_WRITE
])) {
497 BUG_ON(RB_EMPTY_ROOT(&per_prio
->sort_list
[DD_WRITE
]));
503 goto dispatch_find_request
;
508 dispatch_find_request
:
510 * we are not running a batch, find best request for selected data_dir
512 next_rq
= deadline_next_request(dd
, per_prio
, data_dir
);
513 if (deadline_check_fifo(per_prio
, data_dir
) || !next_rq
) {
515 * A deadline has expired, the last request was in the other
516 * direction, or we have run out of higher-sectored requests.
517 * Start again from the request with the earliest expiry time.
519 rq
= deadline_fifo_request(dd
, per_prio
, data_dir
);
522 * The last req was the same dir and we have a next request in
523 * sort order. No expired requests so continue on from here.
529 * For a zoned block device, if we only have writes queued and none of
530 * them can be dispatched, rq will be NULL.
535 dd
->last_dir
= data_dir
;
539 if (started_after(dd
, rq
, latest_start
))
543 * rq is the selected appropriate request.
546 deadline_move_request(dd
, per_prio
, rq
);
548 ioprio_class
= dd_rq_ioclass(rq
);
549 prio
= ioprio_class_to_prio
[ioprio_class
];
550 dd
->per_prio
[prio
].latest_pos
[data_dir
] = blk_rq_pos(rq
);
551 dd
->per_prio
[prio
].stats
.dispatched
++;
553 * If the request needs its target zone locked, do it.
555 blk_req_zone_write_lock(rq
);
556 rq
->rq_flags
|= RQF_STARTED
;
561 * Check whether there are any requests with priority other than DD_RT_PRIO
562 * that were inserted more than prio_aging_expire jiffies ago.
564 static struct request
*dd_dispatch_prio_aged_requests(struct deadline_data
*dd
,
571 lockdep_assert_held(&dd
->lock
);
573 prio_cnt
= !!dd_queued(dd
, DD_RT_PRIO
) + !!dd_queued(dd
, DD_BE_PRIO
) +
574 !!dd_queued(dd
, DD_IDLE_PRIO
);
578 for (prio
= DD_BE_PRIO
; prio
<= DD_PRIO_MAX
; prio
++) {
579 rq
= __dd_dispatch_request(dd
, &dd
->per_prio
[prio
],
580 now
- dd
->prio_aging_expire
);
589 * Called from blk_mq_run_hw_queue() -> __blk_mq_sched_dispatch_requests().
591 * One confusing aspect here is that we get called for a specific
592 * hardware queue, but we may return a request that is for a
593 * different hardware queue. This is because mq-deadline has shared
594 * state for all hardware queues, in terms of sorting, FIFOs, etc.
596 static struct request
*dd_dispatch_request(struct blk_mq_hw_ctx
*hctx
)
598 struct deadline_data
*dd
= hctx
->queue
->elevator
->elevator_data
;
599 const unsigned long now
= jiffies
;
603 spin_lock(&dd
->lock
);
604 rq
= dd_dispatch_prio_aged_requests(dd
, now
);
609 * Next, dispatch requests in priority order. Ignore lower priority
610 * requests if any higher priority requests are pending.
612 for (prio
= 0; prio
<= DD_PRIO_MAX
; prio
++) {
613 rq
= __dd_dispatch_request(dd
, &dd
->per_prio
[prio
], now
);
614 if (rq
|| dd_queued(dd
, prio
))
619 spin_unlock(&dd
->lock
);
625 * Called by __blk_mq_alloc_request(). The shallow_depth value set by this
626 * function is used by __blk_mq_get_tag().
628 static void dd_limit_depth(blk_opf_t opf
, struct blk_mq_alloc_data
*data
)
630 struct deadline_data
*dd
= data
->q
->elevator
->elevator_data
;
632 /* Do not throttle synchronous reads. */
633 if (op_is_sync(opf
) && !op_is_write(opf
))
637 * Throttle asynchronous requests and writes such that these requests
638 * do not block the allocation of synchronous requests.
640 data
->shallow_depth
= dd
->async_depth
;
643 /* Called by blk_mq_update_nr_requests(). */
644 static void dd_depth_updated(struct blk_mq_hw_ctx
*hctx
)
646 struct request_queue
*q
= hctx
->queue
;
647 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
648 struct blk_mq_tags
*tags
= hctx
->sched_tags
;
649 unsigned int shift
= tags
->bitmap_tags
.sb
.shift
;
651 dd
->async_depth
= max(1U, 3 * (1U << shift
) / 4);
653 sbitmap_queue_min_shallow_depth(&tags
->bitmap_tags
, dd
->async_depth
);
656 /* Called by blk_mq_init_hctx() and blk_mq_init_sched(). */
657 static int dd_init_hctx(struct blk_mq_hw_ctx
*hctx
, unsigned int hctx_idx
)
659 dd_depth_updated(hctx
);
663 static void dd_exit_sched(struct elevator_queue
*e
)
665 struct deadline_data
*dd
= e
->elevator_data
;
668 for (prio
= 0; prio
<= DD_PRIO_MAX
; prio
++) {
669 struct dd_per_prio
*per_prio
= &dd
->per_prio
[prio
];
670 const struct io_stats_per_prio
*stats
= &per_prio
->stats
;
673 WARN_ON_ONCE(!list_empty(&per_prio
->fifo_list
[DD_READ
]));
674 WARN_ON_ONCE(!list_empty(&per_prio
->fifo_list
[DD_WRITE
]));
676 spin_lock(&dd
->lock
);
677 queued
= dd_queued(dd
, prio
);
678 spin_unlock(&dd
->lock
);
680 WARN_ONCE(queued
!= 0,
681 "statistics for priority %d: i %u m %u d %u c %u\n",
682 prio
, stats
->inserted
, stats
->merged
,
683 stats
->dispatched
, atomic_read(&stats
->completed
));
690 * initialize elevator private data (deadline_data).
692 static int dd_init_sched(struct request_queue
*q
, struct elevator_type
*e
)
694 struct deadline_data
*dd
;
695 struct elevator_queue
*eq
;
699 eq
= elevator_alloc(q
, e
);
703 dd
= kzalloc_node(sizeof(*dd
), GFP_KERNEL
, q
->node
);
707 eq
->elevator_data
= dd
;
709 for (prio
= 0; prio
<= DD_PRIO_MAX
; prio
++) {
710 struct dd_per_prio
*per_prio
= &dd
->per_prio
[prio
];
712 INIT_LIST_HEAD(&per_prio
->dispatch
);
713 INIT_LIST_HEAD(&per_prio
->fifo_list
[DD_READ
]);
714 INIT_LIST_HEAD(&per_prio
->fifo_list
[DD_WRITE
]);
715 per_prio
->sort_list
[DD_READ
] = RB_ROOT
;
716 per_prio
->sort_list
[DD_WRITE
] = RB_ROOT
;
718 dd
->fifo_expire
[DD_READ
] = read_expire
;
719 dd
->fifo_expire
[DD_WRITE
] = write_expire
;
720 dd
->writes_starved
= writes_starved
;
721 dd
->front_merges
= 1;
722 dd
->last_dir
= DD_WRITE
;
723 dd
->fifo_batch
= fifo_batch
;
724 dd
->prio_aging_expire
= prio_aging_expire
;
725 spin_lock_init(&dd
->lock
);
726 spin_lock_init(&dd
->zone_lock
);
728 /* We dispatch from request queue wide instead of hw queue */
729 blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED
, q
);
735 kobject_put(&eq
->kobj
);
740 * Try to merge @bio into an existing request. If @bio has been merged into
741 * an existing request, store the pointer to that request into *@rq.
743 static int dd_request_merge(struct request_queue
*q
, struct request
**rq
,
746 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
747 const u8 ioprio_class
= IOPRIO_PRIO_CLASS(bio
->bi_ioprio
);
748 const enum dd_prio prio
= ioprio_class_to_prio
[ioprio_class
];
749 struct dd_per_prio
*per_prio
= &dd
->per_prio
[prio
];
750 sector_t sector
= bio_end_sector(bio
);
751 struct request
*__rq
;
753 if (!dd
->front_merges
)
754 return ELEVATOR_NO_MERGE
;
756 __rq
= elv_rb_find(&per_prio
->sort_list
[bio_data_dir(bio
)], sector
);
758 BUG_ON(sector
!= blk_rq_pos(__rq
));
760 if (elv_bio_merge_ok(__rq
, bio
)) {
762 if (blk_discard_mergable(__rq
))
763 return ELEVATOR_DISCARD_MERGE
;
764 return ELEVATOR_FRONT_MERGE
;
768 return ELEVATOR_NO_MERGE
;
772 * Attempt to merge a bio into an existing request. This function is called
773 * before @bio is associated with a request.
775 static bool dd_bio_merge(struct request_queue
*q
, struct bio
*bio
,
776 unsigned int nr_segs
)
778 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
779 struct request
*free
= NULL
;
782 spin_lock(&dd
->lock
);
783 ret
= blk_mq_sched_try_merge(q
, bio
, nr_segs
, &free
);
784 spin_unlock(&dd
->lock
);
787 blk_mq_free_request(free
);
793 * add rq to rbtree and fifo
795 static void dd_insert_request(struct blk_mq_hw_ctx
*hctx
, struct request
*rq
,
796 blk_insert_t flags
, struct list_head
*free
)
798 struct request_queue
*q
= hctx
->queue
;
799 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
800 const enum dd_data_dir data_dir
= rq_data_dir(rq
);
801 u16 ioprio
= req_get_ioprio(rq
);
802 u8 ioprio_class
= IOPRIO_PRIO_CLASS(ioprio
);
803 struct dd_per_prio
*per_prio
;
806 lockdep_assert_held(&dd
->lock
);
809 * This may be a requeue of a write request that has locked its
810 * target zone. If it is the case, this releases the zone lock.
812 blk_req_zone_write_unlock(rq
);
814 prio
= ioprio_class_to_prio
[ioprio_class
];
815 per_prio
= &dd
->per_prio
[prio
];
816 if (!rq
->elv
.priv
[0]) {
817 per_prio
->stats
.inserted
++;
818 rq
->elv
.priv
[0] = (void *)(uintptr_t)1;
821 if (blk_mq_sched_try_insert_merge(q
, rq
, free
))
824 trace_block_rq_insert(rq
);
826 if (flags
& BLK_MQ_INSERT_AT_HEAD
) {
827 list_add(&rq
->queuelist
, &per_prio
->dispatch
);
828 rq
->fifo_time
= jiffies
;
830 struct list_head
*insert_before
;
832 deadline_add_rq_rb(per_prio
, rq
);
834 if (rq_mergeable(rq
)) {
835 elv_rqhash_add(q
, rq
);
841 * set expire time and add to fifo list
843 rq
->fifo_time
= jiffies
+ dd
->fifo_expire
[data_dir
];
844 insert_before
= &per_prio
->fifo_list
[data_dir
];
845 #ifdef CONFIG_BLK_DEV_ZONED
847 * Insert zoned writes such that requests are sorted by
850 if (blk_rq_is_seq_zoned_write(rq
)) {
851 struct request
*rq2
= deadline_latter_request(rq
);
853 if (rq2
&& blk_rq_zone_no(rq2
) == blk_rq_zone_no(rq
))
854 insert_before
= &rq2
->queuelist
;
857 list_add_tail(&rq
->queuelist
, insert_before
);
862 * Called from blk_mq_insert_request() or blk_mq_dispatch_plug_list().
864 static void dd_insert_requests(struct blk_mq_hw_ctx
*hctx
,
865 struct list_head
*list
,
868 struct request_queue
*q
= hctx
->queue
;
869 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
872 spin_lock(&dd
->lock
);
873 while (!list_empty(list
)) {
876 rq
= list_first_entry(list
, struct request
, queuelist
);
877 list_del_init(&rq
->queuelist
);
878 dd_insert_request(hctx
, rq
, flags
, &free
);
880 spin_unlock(&dd
->lock
);
882 blk_mq_free_requests(&free
);
885 /* Callback from inside blk_mq_rq_ctx_init(). */
886 static void dd_prepare_request(struct request
*rq
)
888 rq
->elv
.priv
[0] = NULL
;
891 static bool dd_has_write_work(struct blk_mq_hw_ctx
*hctx
)
893 struct deadline_data
*dd
= hctx
->queue
->elevator
->elevator_data
;
896 for (p
= 0; p
<= DD_PRIO_MAX
; p
++)
897 if (!list_empty_careful(&dd
->per_prio
[p
].fifo_list
[DD_WRITE
]))
904 * Callback from inside blk_mq_free_request().
906 * For zoned block devices, write unlock the target zone of
907 * completed write requests. Do this while holding the zone lock
908 * spinlock so that the zone is never unlocked while deadline_fifo_request()
909 * or deadline_next_request() are executing. This function is called for
910 * all requests, whether or not these requests complete successfully.
912 * For a zoned block device, __dd_dispatch_request() may have stopped
913 * dispatching requests if all the queued requests are write requests directed
914 * at zones that are already locked due to on-going write requests. To ensure
915 * write request dispatch progress in this case, mark the queue as needing a
916 * restart to ensure that the queue is run again after completion of the
917 * request and zones being unlocked.
919 static void dd_finish_request(struct request
*rq
)
921 struct request_queue
*q
= rq
->q
;
922 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
923 const u8 ioprio_class
= dd_rq_ioclass(rq
);
924 const enum dd_prio prio
= ioprio_class_to_prio
[ioprio_class
];
925 struct dd_per_prio
*per_prio
= &dd
->per_prio
[prio
];
928 * The block layer core may call dd_finish_request() without having
929 * called dd_insert_requests(). Skip requests that bypassed I/O
930 * scheduling. See also blk_mq_request_bypass_insert().
932 if (!rq
->elv
.priv
[0])
935 atomic_inc(&per_prio
->stats
.completed
);
937 if (blk_queue_is_zoned(q
)) {
940 spin_lock_irqsave(&dd
->zone_lock
, flags
);
941 blk_req_zone_write_unlock(rq
);
942 spin_unlock_irqrestore(&dd
->zone_lock
, flags
);
944 if (dd_has_write_work(rq
->mq_hctx
))
945 blk_mq_sched_mark_restart_hctx(rq
->mq_hctx
);
949 static bool dd_has_work_for_prio(struct dd_per_prio
*per_prio
)
951 return !list_empty_careful(&per_prio
->dispatch
) ||
952 !list_empty_careful(&per_prio
->fifo_list
[DD_READ
]) ||
953 !list_empty_careful(&per_prio
->fifo_list
[DD_WRITE
]);
956 static bool dd_has_work(struct blk_mq_hw_ctx
*hctx
)
958 struct deadline_data
*dd
= hctx
->queue
->elevator
->elevator_data
;
961 for (prio
= 0; prio
<= DD_PRIO_MAX
; prio
++)
962 if (dd_has_work_for_prio(&dd
->per_prio
[prio
]))
971 #define SHOW_INT(__FUNC, __VAR) \
972 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
974 struct deadline_data *dd = e->elevator_data; \
976 return sysfs_emit(page, "%d\n", __VAR); \
978 #define SHOW_JIFFIES(__FUNC, __VAR) SHOW_INT(__FUNC, jiffies_to_msecs(__VAR))
979 SHOW_JIFFIES(deadline_read_expire_show
, dd
->fifo_expire
[DD_READ
]);
980 SHOW_JIFFIES(deadline_write_expire_show
, dd
->fifo_expire
[DD_WRITE
]);
981 SHOW_JIFFIES(deadline_prio_aging_expire_show
, dd
->prio_aging_expire
);
982 SHOW_INT(deadline_writes_starved_show
, dd
->writes_starved
);
983 SHOW_INT(deadline_front_merges_show
, dd
->front_merges
);
984 SHOW_INT(deadline_async_depth_show
, dd
->async_depth
);
985 SHOW_INT(deadline_fifo_batch_show
, dd
->fifo_batch
);
989 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
990 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
992 struct deadline_data *dd = e->elevator_data; \
995 __ret = kstrtoint(page, 0, &__data); \
998 if (__data < (MIN)) \
1000 else if (__data > (MAX)) \
1002 *(__PTR) = __CONV(__data); \
1005 #define STORE_INT(__FUNC, __PTR, MIN, MAX) \
1006 STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, )
1007 #define STORE_JIFFIES(__FUNC, __PTR, MIN, MAX) \
1008 STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, msecs_to_jiffies)
1009 STORE_JIFFIES(deadline_read_expire_store
, &dd
->fifo_expire
[DD_READ
], 0, INT_MAX
);
1010 STORE_JIFFIES(deadline_write_expire_store
, &dd
->fifo_expire
[DD_WRITE
], 0, INT_MAX
);
1011 STORE_JIFFIES(deadline_prio_aging_expire_store
, &dd
->prio_aging_expire
, 0, INT_MAX
);
1012 STORE_INT(deadline_writes_starved_store
, &dd
->writes_starved
, INT_MIN
, INT_MAX
);
1013 STORE_INT(deadline_front_merges_store
, &dd
->front_merges
, 0, 1);
1014 STORE_INT(deadline_async_depth_store
, &dd
->async_depth
, 1, INT_MAX
);
1015 STORE_INT(deadline_fifo_batch_store
, &dd
->fifo_batch
, 0, INT_MAX
);
1016 #undef STORE_FUNCTION
1018 #undef STORE_JIFFIES
1020 #define DD_ATTR(name) \
1021 __ATTR(name, 0644, deadline_##name##_show, deadline_##name##_store)
1023 static struct elv_fs_entry deadline_attrs
[] = {
1024 DD_ATTR(read_expire
),
1025 DD_ATTR(write_expire
),
1026 DD_ATTR(writes_starved
),
1027 DD_ATTR(front_merges
),
1028 DD_ATTR(async_depth
),
1029 DD_ATTR(fifo_batch
),
1030 DD_ATTR(prio_aging_expire
),
1034 #ifdef CONFIG_BLK_DEBUG_FS
1035 #define DEADLINE_DEBUGFS_DDIR_ATTRS(prio, data_dir, name) \
1036 static void *deadline_##name##_fifo_start(struct seq_file *m, \
1038 __acquires(&dd->lock) \
1040 struct request_queue *q = m->private; \
1041 struct deadline_data *dd = q->elevator->elevator_data; \
1042 struct dd_per_prio *per_prio = &dd->per_prio[prio]; \
1044 spin_lock(&dd->lock); \
1045 return seq_list_start(&per_prio->fifo_list[data_dir], *pos); \
1048 static void *deadline_##name##_fifo_next(struct seq_file *m, void *v, \
1051 struct request_queue *q = m->private; \
1052 struct deadline_data *dd = q->elevator->elevator_data; \
1053 struct dd_per_prio *per_prio = &dd->per_prio[prio]; \
1055 return seq_list_next(v, &per_prio->fifo_list[data_dir], pos); \
1058 static void deadline_##name##_fifo_stop(struct seq_file *m, void *v) \
1059 __releases(&dd->lock) \
1061 struct request_queue *q = m->private; \
1062 struct deadline_data *dd = q->elevator->elevator_data; \
1064 spin_unlock(&dd->lock); \
1067 static const struct seq_operations deadline_##name##_fifo_seq_ops = { \
1068 .start = deadline_##name##_fifo_start, \
1069 .next = deadline_##name##_fifo_next, \
1070 .stop = deadline_##name##_fifo_stop, \
1071 .show = blk_mq_debugfs_rq_show, \
1074 static int deadline_##name##_next_rq_show(void *data, \
1075 struct seq_file *m) \
1077 struct request_queue *q = data; \
1078 struct deadline_data *dd = q->elevator->elevator_data; \
1079 struct dd_per_prio *per_prio = &dd->per_prio[prio]; \
1080 struct request *rq; \
1082 rq = deadline_from_pos(per_prio, data_dir, \
1083 per_prio->latest_pos[data_dir]); \
1085 __blk_mq_debugfs_rq_show(m, rq); \
1089 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO
, DD_READ
, read0
);
1090 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_RT_PRIO
, DD_WRITE
, write0
);
1091 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO
, DD_READ
, read1
);
1092 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_BE_PRIO
, DD_WRITE
, write1
);
1093 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO
, DD_READ
, read2
);
1094 DEADLINE_DEBUGFS_DDIR_ATTRS(DD_IDLE_PRIO
, DD_WRITE
, write2
);
1095 #undef DEADLINE_DEBUGFS_DDIR_ATTRS
1097 static int deadline_batching_show(void *data
, struct seq_file
*m
)
1099 struct request_queue
*q
= data
;
1100 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
1102 seq_printf(m
, "%u\n", dd
->batching
);
1106 static int deadline_starved_show(void *data
, struct seq_file
*m
)
1108 struct request_queue
*q
= data
;
1109 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
1111 seq_printf(m
, "%u\n", dd
->starved
);
1115 static int dd_async_depth_show(void *data
, struct seq_file
*m
)
1117 struct request_queue
*q
= data
;
1118 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
1120 seq_printf(m
, "%u\n", dd
->async_depth
);
1124 static int dd_queued_show(void *data
, struct seq_file
*m
)
1126 struct request_queue
*q
= data
;
1127 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
1130 spin_lock(&dd
->lock
);
1131 rt
= dd_queued(dd
, DD_RT_PRIO
);
1132 be
= dd_queued(dd
, DD_BE_PRIO
);
1133 idle
= dd_queued(dd
, DD_IDLE_PRIO
);
1134 spin_unlock(&dd
->lock
);
1136 seq_printf(m
, "%u %u %u\n", rt
, be
, idle
);
1141 /* Number of requests owned by the block driver for a given priority. */
1142 static u32
dd_owned_by_driver(struct deadline_data
*dd
, enum dd_prio prio
)
1144 const struct io_stats_per_prio
*stats
= &dd
->per_prio
[prio
].stats
;
1146 lockdep_assert_held(&dd
->lock
);
1148 return stats
->dispatched
+ stats
->merged
-
1149 atomic_read(&stats
->completed
);
1152 static int dd_owned_by_driver_show(void *data
, struct seq_file
*m
)
1154 struct request_queue
*q
= data
;
1155 struct deadline_data
*dd
= q
->elevator
->elevator_data
;
1158 spin_lock(&dd
->lock
);
1159 rt
= dd_owned_by_driver(dd
, DD_RT_PRIO
);
1160 be
= dd_owned_by_driver(dd
, DD_BE_PRIO
);
1161 idle
= dd_owned_by_driver(dd
, DD_IDLE_PRIO
);
1162 spin_unlock(&dd
->lock
);
1164 seq_printf(m
, "%u %u %u\n", rt
, be
, idle
);
1169 #define DEADLINE_DISPATCH_ATTR(prio) \
1170 static void *deadline_dispatch##prio##_start(struct seq_file *m, \
1172 __acquires(&dd->lock) \
1174 struct request_queue *q = m->private; \
1175 struct deadline_data *dd = q->elevator->elevator_data; \
1176 struct dd_per_prio *per_prio = &dd->per_prio[prio]; \
1178 spin_lock(&dd->lock); \
1179 return seq_list_start(&per_prio->dispatch, *pos); \
1182 static void *deadline_dispatch##prio##_next(struct seq_file *m, \
1183 void *v, loff_t *pos) \
1185 struct request_queue *q = m->private; \
1186 struct deadline_data *dd = q->elevator->elevator_data; \
1187 struct dd_per_prio *per_prio = &dd->per_prio[prio]; \
1189 return seq_list_next(v, &per_prio->dispatch, pos); \
1192 static void deadline_dispatch##prio##_stop(struct seq_file *m, void *v) \
1193 __releases(&dd->lock) \
1195 struct request_queue *q = m->private; \
1196 struct deadline_data *dd = q->elevator->elevator_data; \
1198 spin_unlock(&dd->lock); \
1201 static const struct seq_operations deadline_dispatch##prio##_seq_ops = { \
1202 .start = deadline_dispatch##prio##_start, \
1203 .next = deadline_dispatch##prio##_next, \
1204 .stop = deadline_dispatch##prio##_stop, \
1205 .show = blk_mq_debugfs_rq_show, \
1208 DEADLINE_DISPATCH_ATTR(0);
1209 DEADLINE_DISPATCH_ATTR(1);
1210 DEADLINE_DISPATCH_ATTR(2);
1211 #undef DEADLINE_DISPATCH_ATTR
1213 #define DEADLINE_QUEUE_DDIR_ATTRS(name) \
1214 {#name "_fifo_list", 0400, \
1215 .seq_ops = &deadline_##name##_fifo_seq_ops}
1216 #define DEADLINE_NEXT_RQ_ATTR(name) \
1217 {#name "_next_rq", 0400, deadline_##name##_next_rq_show}
1218 static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs
[] = {
1219 DEADLINE_QUEUE_DDIR_ATTRS(read0
),
1220 DEADLINE_QUEUE_DDIR_ATTRS(write0
),
1221 DEADLINE_QUEUE_DDIR_ATTRS(read1
),
1222 DEADLINE_QUEUE_DDIR_ATTRS(write1
),
1223 DEADLINE_QUEUE_DDIR_ATTRS(read2
),
1224 DEADLINE_QUEUE_DDIR_ATTRS(write2
),
1225 DEADLINE_NEXT_RQ_ATTR(read0
),
1226 DEADLINE_NEXT_RQ_ATTR(write0
),
1227 DEADLINE_NEXT_RQ_ATTR(read1
),
1228 DEADLINE_NEXT_RQ_ATTR(write1
),
1229 DEADLINE_NEXT_RQ_ATTR(read2
),
1230 DEADLINE_NEXT_RQ_ATTR(write2
),
1231 {"batching", 0400, deadline_batching_show
},
1232 {"starved", 0400, deadline_starved_show
},
1233 {"async_depth", 0400, dd_async_depth_show
},
1234 {"dispatch0", 0400, .seq_ops
= &deadline_dispatch0_seq_ops
},
1235 {"dispatch1", 0400, .seq_ops
= &deadline_dispatch1_seq_ops
},
1236 {"dispatch2", 0400, .seq_ops
= &deadline_dispatch2_seq_ops
},
1237 {"owned_by_driver", 0400, dd_owned_by_driver_show
},
1238 {"queued", 0400, dd_queued_show
},
1241 #undef DEADLINE_QUEUE_DDIR_ATTRS
1244 static struct elevator_type mq_deadline
= {
1246 .depth_updated
= dd_depth_updated
,
1247 .limit_depth
= dd_limit_depth
,
1248 .insert_requests
= dd_insert_requests
,
1249 .dispatch_request
= dd_dispatch_request
,
1250 .prepare_request
= dd_prepare_request
,
1251 .finish_request
= dd_finish_request
,
1252 .next_request
= elv_rb_latter_request
,
1253 .former_request
= elv_rb_former_request
,
1254 .bio_merge
= dd_bio_merge
,
1255 .request_merge
= dd_request_merge
,
1256 .requests_merged
= dd_merged_requests
,
1257 .request_merged
= dd_request_merged
,
1258 .has_work
= dd_has_work
,
1259 .init_sched
= dd_init_sched
,
1260 .exit_sched
= dd_exit_sched
,
1261 .init_hctx
= dd_init_hctx
,
1264 #ifdef CONFIG_BLK_DEBUG_FS
1265 .queue_debugfs_attrs
= deadline_queue_debugfs_attrs
,
1267 .elevator_attrs
= deadline_attrs
,
1268 .elevator_name
= "mq-deadline",
1269 .elevator_alias
= "deadline",
1270 .elevator_features
= ELEVATOR_F_ZBD_SEQ_WRITE
,
1271 .elevator_owner
= THIS_MODULE
,
1273 MODULE_ALIAS("mq-deadline-iosched");
1275 static int __init
deadline_init(void)
1277 return elv_register(&mq_deadline
);
1280 static void __exit
deadline_exit(void)
1282 elv_unregister(&mq_deadline
);
1285 module_init(deadline_init
);
1286 module_exit(deadline_exit
);
1288 MODULE_AUTHOR("Jens Axboe, Damien Le Moal and Bart Van Assche");
1289 MODULE_LICENSE("GPL");
1290 MODULE_DESCRIPTION("MQ deadline IO scheduler");