2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
20 /* max queue in one round of service */
21 static const int cfq_quantum = 4;
22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
23 /* maximum backwards seek, in KiB */
24 static const int cfq_back_max = 16 * 1024;
25 /* penalty of a backwards seek */
26 static const int cfq_back_penalty = 2;
27 static const int cfq_slice_sync = HZ / 10;
28 static int cfq_slice_async = HZ / 25;
29 static const int cfq_slice_async_rq = 2;
30 static int cfq_slice_idle = HZ / 125;
31 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
32 static const int cfq_hist_divisor = 4;
35 * offset from end of service tree
37 #define CFQ_IDLE_DELAY (HZ / 5)
40 * below this threshold, we consider thinktime immediate
42 #define CFQ_MIN_TT (2)
45 * Allow merged cfqqs to perform this amount of seeky I/O before
46 * deciding to break the queues up again.
48 #define CFQQ_COOP_TOUT (HZ)
50 #define CFQ_SLICE_SCALE (5)
51 #define CFQ_HW_QUEUE_MIN (5)
54 ((struct cfq_io_context *) (rq)->elevator_private)
55 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
57 static struct kmem_cache *cfq_pool;
58 static struct kmem_cache *cfq_ioc_pool;
60 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
61 static struct completion *ioc_gone;
62 static DEFINE_SPINLOCK(ioc_gone_lock);
64 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68 #define sample_valid(samples) ((samples) > 80)
71 * Most of our rbtree usage is for sorting with min extraction, so
72 * if we cache the leftmost node we don't have to walk down the tree
73 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
74 * move this into the elevator for the rq sorting as well.
81 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, }
84 * Per process-grouping structure
89 /* various state flags, see below */
92 struct cfq_data *cfqd;
93 /* service_tree member */
94 struct rb_node rb_node;
95 /* service_tree key */
97 /* prio tree member */
98 struct rb_node p_node;
99 /* prio tree root we belong to, if any */
100 struct rb_root *p_root;
101 /* sorted list of pending requests */
102 struct rb_root sort_list;
103 /* if fifo isn't expired, next request to serve */
104 struct request *next_rq;
105 /* requests queued in sort_list */
107 /* currently allocated requests */
109 /* fifo list of requests in sort_list */
110 struct list_head fifo;
112 unsigned long slice_end;
114 unsigned int slice_dispatch;
116 /* pending metadata requests */
118 /* number of requests that are on the dispatch list or inside driver */
121 /* io prio of this group */
122 unsigned short ioprio, org_ioprio;
123 unsigned short ioprio_class, org_ioprio_class;
125 unsigned int seek_samples;
128 sector_t last_request_pos;
129 unsigned long seeky_start;
133 struct cfq_rb_root *service_tree;
134 struct cfq_queue *new_cfqq;
135 struct cfq_group *cfqg;
139 * First index in the service_trees.
140 * IDLE is handled separately, so it has negative index
149 * Second index in the service_trees.
153 SYNC_NOIDLE_WORKLOAD = 1,
157 /* This is per cgroup per device grouping structure */
160 * rr lists of queues with requests, onle rr for each priority class.
161 * Counts are embedded in the cfq_rb_root
163 struct cfq_rb_root service_trees[2][3];
164 struct cfq_rb_root service_tree_idle;
168 * Per block device queue structure
171 struct request_queue *queue;
172 struct cfq_group root_group;
175 * The priority currently being served
177 enum wl_prio_t serving_prio;
178 enum wl_type_t serving_type;
179 unsigned long workload_expires;
180 struct cfq_group *serving_group;
181 bool noidle_tree_requires_idle;
184 * Each priority tree is sorted by next_request position. These
185 * trees are used when determining if two or more queues are
186 * interleaving requests (see cfq_close_cooperator).
188 struct rb_root prio_trees[CFQ_PRIO_LISTS];
190 unsigned int busy_queues;
191 unsigned int busy_queues_avg[2];
197 * queue-depth detection
203 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
204 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
207 int hw_tag_est_depth;
208 unsigned int hw_tag_samples;
211 * idle window management
213 struct timer_list idle_slice_timer;
214 struct work_struct unplug_work;
216 struct cfq_queue *active_queue;
217 struct cfq_io_context *active_cic;
220 * async queue for each priority case
222 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
223 struct cfq_queue *async_idle_cfqq;
225 sector_t last_position;
228 * tunables, see top of file
230 unsigned int cfq_quantum;
231 unsigned int cfq_fifo_expire[2];
232 unsigned int cfq_back_penalty;
233 unsigned int cfq_back_max;
234 unsigned int cfq_slice[2];
235 unsigned int cfq_slice_async_rq;
236 unsigned int cfq_slice_idle;
237 unsigned int cfq_latency;
239 struct list_head cic_list;
242 * Fallback dummy cfqq for extreme OOM conditions
244 struct cfq_queue oom_cfqq;
246 unsigned long last_end_sync_rq;
249 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
252 struct cfq_data *cfqd)
254 if (prio == IDLE_WORKLOAD)
255 return &cfqg->service_tree_idle;
257 return &cfqg->service_trees[prio][type];
260 enum cfqq_state_flags {
261 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
262 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
263 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
264 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
265 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
266 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
267 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
268 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
269 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
270 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
271 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
274 #define CFQ_CFQQ_FNS(name) \
275 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
277 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
279 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
281 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
283 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
285 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
289 CFQ_CFQQ_FNS(wait_request);
290 CFQ_CFQQ_FNS(must_dispatch);
291 CFQ_CFQQ_FNS(must_alloc_slice);
292 CFQ_CFQQ_FNS(fifo_expire);
293 CFQ_CFQQ_FNS(idle_window);
294 CFQ_CFQQ_FNS(prio_changed);
295 CFQ_CFQQ_FNS(slice_new);
301 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
302 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
303 #define cfq_log(cfqd, fmt, args...) \
304 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
306 /* Traverses through cfq group service trees */
307 #define for_each_cfqg_st(cfqg, i, j, st) \
308 for (i = 0; i <= IDLE_WORKLOAD; i++) \
309 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
310 : &cfqg->service_tree_idle; \
311 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
312 (i == IDLE_WORKLOAD && j == 0); \
313 j++, st = i < IDLE_WORKLOAD ? \
314 &cfqg->service_trees[i][j]: NULL) \
317 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
319 if (cfq_class_idle(cfqq))
320 return IDLE_WORKLOAD;
321 if (cfq_class_rt(cfqq))
327 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
329 if (!cfq_cfqq_sync(cfqq))
330 return ASYNC_WORKLOAD;
331 if (!cfq_cfqq_idle_window(cfqq))
332 return SYNC_NOIDLE_WORKLOAD;
333 return SYNC_WORKLOAD;
336 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
338 struct cfq_group *cfqg = &cfqd->root_group;
340 if (wl == IDLE_WORKLOAD)
341 return cfqg->service_tree_idle.count;
343 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
344 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
345 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
348 static void cfq_dispatch_insert(struct request_queue *, struct request *);
349 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
350 struct io_context *, gfp_t);
351 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
352 struct io_context *);
354 static inline int rq_in_driver(struct cfq_data *cfqd)
356 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
359 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
362 return cic->cfqq[is_sync];
365 static inline void cic_set_cfqq(struct cfq_io_context *cic,
366 struct cfq_queue *cfqq, bool is_sync)
368 cic->cfqq[is_sync] = cfqq;
372 * We regard a request as SYNC, if it's either a read or has the SYNC bit
373 * set (in which case it could also be direct WRITE).
375 static inline bool cfq_bio_sync(struct bio *bio)
377 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
381 * scheduler run of queue, if there are requests pending and no one in the
382 * driver that will restart queueing
384 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
386 if (cfqd->busy_queues) {
387 cfq_log(cfqd, "schedule dispatch");
388 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
392 static int cfq_queue_empty(struct request_queue *q)
394 struct cfq_data *cfqd = q->elevator->elevator_data;
396 return !cfqd->rq_queued;
400 * Scale schedule slice based on io priority. Use the sync time slice only
401 * if a queue is marked sync and has sync io queued. A sync queue with async
402 * io only, should not get full sync slice length.
404 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
407 const int base_slice = cfqd->cfq_slice[sync];
409 WARN_ON(prio >= IOPRIO_BE_NR);
411 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
415 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
417 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
421 * get averaged number of queues of RT/BE priority.
422 * average is updated, with a formula that gives more weight to higher numbers,
423 * to quickly follows sudden increases and decrease slowly
426 static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt)
428 unsigned min_q, max_q;
429 unsigned mult = cfq_hist_divisor - 1;
430 unsigned round = cfq_hist_divisor / 2;
431 unsigned busy = cfq_busy_queues_wl(rt, cfqd);
433 min_q = min(cfqd->busy_queues_avg[rt], busy);
434 max_q = max(cfqd->busy_queues_avg[rt], busy);
435 cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
437 return cfqd->busy_queues_avg[rt];
441 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
443 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
444 if (cfqd->cfq_latency) {
445 /* interested queues (we consider only the ones with the same
447 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
448 unsigned sync_slice = cfqd->cfq_slice[1];
449 unsigned expect_latency = sync_slice * iq;
450 if (expect_latency > cfq_target_latency) {
451 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
452 /* scale low_slice according to IO priority
453 * and sync vs async */
455 min(slice, base_low_slice * slice / sync_slice);
456 /* the adapted slice value is scaled to fit all iqs
457 * into the target latency */
458 slice = max(slice * cfq_target_latency / expect_latency,
462 cfqq->slice_end = jiffies + slice;
463 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
467 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
468 * isn't valid until the first request from the dispatch is activated
469 * and the slice time set.
471 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
473 if (cfq_cfqq_slice_new(cfqq))
475 if (time_before(jiffies, cfqq->slice_end))
482 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
483 * We choose the request that is closest to the head right now. Distance
484 * behind the head is penalized and only allowed to a certain extent.
486 static struct request *
487 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
489 sector_t s1, s2, d1 = 0, d2 = 0;
490 unsigned long back_max;
491 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
492 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
493 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
495 if (rq1 == NULL || rq1 == rq2)
500 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
502 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
504 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
506 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
509 s1 = blk_rq_pos(rq1);
510 s2 = blk_rq_pos(rq2);
513 * by definition, 1KiB is 2 sectors
515 back_max = cfqd->cfq_back_max * 2;
518 * Strict one way elevator _except_ in the case where we allow
519 * short backward seeks which are biased as twice the cost of a
520 * similar forward seek.
524 else if (s1 + back_max >= last)
525 d1 = (last - s1) * cfqd->cfq_back_penalty;
527 wrap |= CFQ_RQ1_WRAP;
531 else if (s2 + back_max >= last)
532 d2 = (last - s2) * cfqd->cfq_back_penalty;
534 wrap |= CFQ_RQ2_WRAP;
536 /* Found required data */
539 * By doing switch() on the bit mask "wrap" we avoid having to
540 * check two variables for all permutations: --> faster!
543 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
559 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
562 * Since both rqs are wrapped,
563 * start with the one that's further behind head
564 * (--> only *one* back seek required),
565 * since back seek takes more time than forward.
575 * The below is leftmost cache rbtree addon
577 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
579 /* Service tree is empty */
584 root->left = rb_first(&root->rb);
587 return rb_entry(root->left, struct cfq_queue, rb_node);
592 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
598 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
602 rb_erase_init(n, &root->rb);
607 * would be nice to take fifo expire time into account as well
609 static struct request *
610 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
611 struct request *last)
613 struct rb_node *rbnext = rb_next(&last->rb_node);
614 struct rb_node *rbprev = rb_prev(&last->rb_node);
615 struct request *next = NULL, *prev = NULL;
617 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
620 prev = rb_entry_rq(rbprev);
623 next = rb_entry_rq(rbnext);
625 rbnext = rb_first(&cfqq->sort_list);
626 if (rbnext && rbnext != &last->rb_node)
627 next = rb_entry_rq(rbnext);
630 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
633 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
634 struct cfq_queue *cfqq)
637 * just an approximation, should be ok.
639 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
640 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
644 * The cfqd->service_trees holds all pending cfq_queue's that have
645 * requests waiting to be processed. It is sorted in the order that
646 * we will service the queues.
648 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
651 struct rb_node **p, *parent;
652 struct cfq_queue *__cfqq;
653 unsigned long rb_key;
654 struct cfq_rb_root *service_tree;
657 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
658 cfqq_type(cfqq), cfqd);
659 if (cfq_class_idle(cfqq)) {
660 rb_key = CFQ_IDLE_DELAY;
661 parent = rb_last(&service_tree->rb);
662 if (parent && parent != &cfqq->rb_node) {
663 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
664 rb_key += __cfqq->rb_key;
667 } else if (!add_front) {
669 * Get our rb key offset. Subtract any residual slice
670 * value carried from last service. A negative resid
671 * count indicates slice overrun, and this should position
672 * the next service time further away in the tree.
674 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
675 rb_key -= cfqq->slice_resid;
676 cfqq->slice_resid = 0;
679 __cfqq = cfq_rb_first(service_tree);
680 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
683 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
685 * same position, nothing more to do
687 if (rb_key == cfqq->rb_key &&
688 cfqq->service_tree == service_tree)
691 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
692 cfqq->service_tree = NULL;
697 cfqq->service_tree = service_tree;
698 p = &service_tree->rb.rb_node;
703 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
706 * sort by key, that represents service time.
708 if (time_before(rb_key, __cfqq->rb_key))
719 service_tree->left = &cfqq->rb_node;
721 cfqq->rb_key = rb_key;
722 rb_link_node(&cfqq->rb_node, parent, p);
723 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
724 service_tree->count++;
727 static struct cfq_queue *
728 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
729 sector_t sector, struct rb_node **ret_parent,
730 struct rb_node ***rb_link)
732 struct rb_node **p, *parent;
733 struct cfq_queue *cfqq = NULL;
741 cfqq = rb_entry(parent, struct cfq_queue, p_node);
744 * Sort strictly based on sector. Smallest to the left,
745 * largest to the right.
747 if (sector > blk_rq_pos(cfqq->next_rq))
749 else if (sector < blk_rq_pos(cfqq->next_rq))
757 *ret_parent = parent;
763 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
765 struct rb_node **p, *parent;
766 struct cfq_queue *__cfqq;
769 rb_erase(&cfqq->p_node, cfqq->p_root);
773 if (cfq_class_idle(cfqq))
778 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
779 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
780 blk_rq_pos(cfqq->next_rq), &parent, &p);
782 rb_link_node(&cfqq->p_node, parent, p);
783 rb_insert_color(&cfqq->p_node, cfqq->p_root);
789 * Update cfqq's position in the service tree.
791 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
794 * Resorting requires the cfqq to be on the RR list already.
796 if (cfq_cfqq_on_rr(cfqq)) {
797 cfq_service_tree_add(cfqd, cfqq, 0);
798 cfq_prio_tree_add(cfqd, cfqq);
803 * add to busy list of queues for service, trying to be fair in ordering
804 * the pending list according to last request service
806 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
808 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
809 BUG_ON(cfq_cfqq_on_rr(cfqq));
810 cfq_mark_cfqq_on_rr(cfqq);
813 cfq_resort_rr_list(cfqd, cfqq);
817 * Called when the cfqq no longer has requests pending, remove it from
820 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
822 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
823 BUG_ON(!cfq_cfqq_on_rr(cfqq));
824 cfq_clear_cfqq_on_rr(cfqq);
826 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
827 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
828 cfqq->service_tree = NULL;
831 rb_erase(&cfqq->p_node, cfqq->p_root);
835 BUG_ON(!cfqd->busy_queues);
840 * rb tree support functions
842 static void cfq_del_rq_rb(struct request *rq)
844 struct cfq_queue *cfqq = RQ_CFQQ(rq);
845 const int sync = rq_is_sync(rq);
847 BUG_ON(!cfqq->queued[sync]);
848 cfqq->queued[sync]--;
850 elv_rb_del(&cfqq->sort_list, rq);
852 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
854 * Queue will be deleted from service tree when we actually
855 * expire it later. Right now just remove it from prio tree
859 rb_erase(&cfqq->p_node, cfqq->p_root);
865 static void cfq_add_rq_rb(struct request *rq)
867 struct cfq_queue *cfqq = RQ_CFQQ(rq);
868 struct cfq_data *cfqd = cfqq->cfqd;
869 struct request *__alias, *prev;
871 cfqq->queued[rq_is_sync(rq)]++;
874 * looks a little odd, but the first insert might return an alias.
875 * if that happens, put the alias on the dispatch list
877 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
878 cfq_dispatch_insert(cfqd->queue, __alias);
880 if (!cfq_cfqq_on_rr(cfqq))
881 cfq_add_cfqq_rr(cfqd, cfqq);
884 * check if this request is a better next-serve candidate
886 prev = cfqq->next_rq;
887 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
890 * adjust priority tree position, if ->next_rq changes
892 if (prev != cfqq->next_rq)
893 cfq_prio_tree_add(cfqd, cfqq);
895 BUG_ON(!cfqq->next_rq);
898 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
900 elv_rb_del(&cfqq->sort_list, rq);
901 cfqq->queued[rq_is_sync(rq)]--;
905 static struct request *
906 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
908 struct task_struct *tsk = current;
909 struct cfq_io_context *cic;
910 struct cfq_queue *cfqq;
912 cic = cfq_cic_lookup(cfqd, tsk->io_context);
916 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
918 sector_t sector = bio->bi_sector + bio_sectors(bio);
920 return elv_rb_find(&cfqq->sort_list, sector);
926 static void cfq_activate_request(struct request_queue *q, struct request *rq)
928 struct cfq_data *cfqd = q->elevator->elevator_data;
930 cfqd->rq_in_driver[rq_is_sync(rq)]++;
931 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
934 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
937 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
939 struct cfq_data *cfqd = q->elevator->elevator_data;
940 const int sync = rq_is_sync(rq);
942 WARN_ON(!cfqd->rq_in_driver[sync]);
943 cfqd->rq_in_driver[sync]--;
944 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
948 static void cfq_remove_request(struct request *rq)
950 struct cfq_queue *cfqq = RQ_CFQQ(rq);
952 if (cfqq->next_rq == rq)
953 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
955 list_del_init(&rq->queuelist);
958 cfqq->cfqd->rq_queued--;
959 if (rq_is_meta(rq)) {
960 WARN_ON(!cfqq->meta_pending);
961 cfqq->meta_pending--;
965 static int cfq_merge(struct request_queue *q, struct request **req,
968 struct cfq_data *cfqd = q->elevator->elevator_data;
969 struct request *__rq;
971 __rq = cfq_find_rq_fmerge(cfqd, bio);
972 if (__rq && elv_rq_merge_ok(__rq, bio)) {
974 return ELEVATOR_FRONT_MERGE;
977 return ELEVATOR_NO_MERGE;
980 static void cfq_merged_request(struct request_queue *q, struct request *req,
983 if (type == ELEVATOR_FRONT_MERGE) {
984 struct cfq_queue *cfqq = RQ_CFQQ(req);
986 cfq_reposition_rq_rb(cfqq, req);
991 cfq_merged_requests(struct request_queue *q, struct request *rq,
992 struct request *next)
994 struct cfq_queue *cfqq = RQ_CFQQ(rq);
996 * reposition in fifo if next is older than rq
998 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
999 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1000 list_move(&rq->queuelist, &next->queuelist);
1001 rq_set_fifo_time(rq, rq_fifo_time(next));
1004 if (cfqq->next_rq == next)
1006 cfq_remove_request(next);
1009 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1012 struct cfq_data *cfqd = q->elevator->elevator_data;
1013 struct cfq_io_context *cic;
1014 struct cfq_queue *cfqq;
1017 * Disallow merge of a sync bio into an async request.
1019 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1023 * Lookup the cfqq that this bio will be queued with. Allow
1024 * merge only if rq is queued there.
1026 cic = cfq_cic_lookup(cfqd, current->io_context);
1030 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1031 return cfqq == RQ_CFQQ(rq);
1034 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1035 struct cfq_queue *cfqq)
1038 cfq_log_cfqq(cfqd, cfqq, "set_active");
1039 cfqq->slice_end = 0;
1040 cfqq->slice_dispatch = 0;
1042 cfq_clear_cfqq_wait_request(cfqq);
1043 cfq_clear_cfqq_must_dispatch(cfqq);
1044 cfq_clear_cfqq_must_alloc_slice(cfqq);
1045 cfq_clear_cfqq_fifo_expire(cfqq);
1046 cfq_mark_cfqq_slice_new(cfqq);
1048 del_timer(&cfqd->idle_slice_timer);
1051 cfqd->active_queue = cfqq;
1055 * current cfqq expired its slice (or was too idle), select new one
1058 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1061 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1063 if (cfq_cfqq_wait_request(cfqq))
1064 del_timer(&cfqd->idle_slice_timer);
1066 cfq_clear_cfqq_wait_request(cfqq);
1069 * store what was left of this slice, if the queue idled/timed out
1071 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1072 cfqq->slice_resid = cfqq->slice_end - jiffies;
1073 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1076 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1077 cfq_del_cfqq_rr(cfqd, cfqq);
1079 cfq_resort_rr_list(cfqd, cfqq);
1081 if (cfqq == cfqd->active_queue)
1082 cfqd->active_queue = NULL;
1084 if (cfqd->active_cic) {
1085 put_io_context(cfqd->active_cic->ioc);
1086 cfqd->active_cic = NULL;
1090 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1092 struct cfq_queue *cfqq = cfqd->active_queue;
1095 __cfq_slice_expired(cfqd, cfqq, timed_out);
1099 * Get next queue for service. Unless we have a queue preemption,
1100 * we'll simply select the first cfqq in the service tree.
1102 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1104 struct cfq_rb_root *service_tree =
1105 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1106 cfqd->serving_type, cfqd);
1108 if (!cfqd->rq_queued)
1111 if (RB_EMPTY_ROOT(&service_tree->rb))
1113 return cfq_rb_first(service_tree);
1116 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1118 struct cfq_group *cfqg = &cfqd->root_group;
1119 struct cfq_queue *cfqq;
1121 struct cfq_rb_root *st;
1123 if (!cfqd->rq_queued)
1126 for_each_cfqg_st(cfqg, i, j, st)
1127 if ((cfqq = cfq_rb_first(st)) != NULL)
1133 * Get and set a new active queue for service.
1135 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1136 struct cfq_queue *cfqq)
1139 cfqq = cfq_get_next_queue(cfqd);
1141 __cfq_set_active_queue(cfqd, cfqq);
1145 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1148 if (blk_rq_pos(rq) >= cfqd->last_position)
1149 return blk_rq_pos(rq) - cfqd->last_position;
1151 return cfqd->last_position - blk_rq_pos(rq);
1154 #define CFQQ_SEEK_THR 8 * 1024
1155 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1157 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1160 sector_t sdist = cfqq->seek_mean;
1162 if (!sample_valid(cfqq->seek_samples))
1163 sdist = CFQQ_SEEK_THR;
1165 return cfq_dist_from_last(cfqd, rq) <= sdist;
1168 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1169 struct cfq_queue *cur_cfqq)
1171 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1172 struct rb_node *parent, *node;
1173 struct cfq_queue *__cfqq;
1174 sector_t sector = cfqd->last_position;
1176 if (RB_EMPTY_ROOT(root))
1180 * First, if we find a request starting at the end of the last
1181 * request, choose it.
1183 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1188 * If the exact sector wasn't found, the parent of the NULL leaf
1189 * will contain the closest sector.
1191 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1192 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1195 if (blk_rq_pos(__cfqq->next_rq) < sector)
1196 node = rb_next(&__cfqq->p_node);
1198 node = rb_prev(&__cfqq->p_node);
1202 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1203 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1211 * cur_cfqq - passed in so that we don't decide that the current queue is
1212 * closely cooperating with itself.
1214 * So, basically we're assuming that that cur_cfqq has dispatched at least
1215 * one request, and that cfqd->last_position reflects a position on the disk
1216 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1219 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1220 struct cfq_queue *cur_cfqq)
1222 struct cfq_queue *cfqq;
1224 if (!cfq_cfqq_sync(cur_cfqq))
1226 if (CFQQ_SEEKY(cur_cfqq))
1230 * We should notice if some of the queues are cooperating, eg
1231 * working closely on the same area of the disk. In that case,
1232 * we can group them together and don't waste time idling.
1234 cfqq = cfqq_close(cfqd, cur_cfqq);
1239 * It only makes sense to merge sync queues.
1241 if (!cfq_cfqq_sync(cfqq))
1243 if (CFQQ_SEEKY(cfqq))
1247 * Do not merge queues of different priority classes
1249 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1256 * Determine whether we should enforce idle window for this queue.
1259 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1261 enum wl_prio_t prio = cfqq_prio(cfqq);
1262 struct cfq_rb_root *service_tree = cfqq->service_tree;
1264 BUG_ON(!service_tree);
1265 BUG_ON(!service_tree->count);
1267 /* We never do for idle class queues. */
1268 if (prio == IDLE_WORKLOAD)
1271 /* We do for queues that were marked with idle window flag. */
1272 if (cfq_cfqq_idle_window(cfqq))
1276 * Otherwise, we do only if they are the last ones
1277 * in their service tree.
1279 return service_tree->count == 1;
1282 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1284 struct cfq_queue *cfqq = cfqd->active_queue;
1285 struct cfq_io_context *cic;
1289 * SSD device without seek penalty, disable idling. But only do so
1290 * for devices that support queuing, otherwise we still have a problem
1291 * with sync vs async workloads.
1293 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1296 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1297 WARN_ON(cfq_cfqq_slice_new(cfqq));
1300 * idle is disabled, either manually or by past process history
1302 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1306 * still active requests from this queue, don't idle
1308 if (cfqq->dispatched)
1312 * task has exited, don't wait
1314 cic = cfqd->active_cic;
1315 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1319 * If our average think time is larger than the remaining time
1320 * slice, then don't idle. This avoids overrunning the allotted
1323 if (sample_valid(cic->ttime_samples) &&
1324 (cfqq->slice_end - jiffies < cic->ttime_mean))
1327 cfq_mark_cfqq_wait_request(cfqq);
1329 sl = cfqd->cfq_slice_idle;
1331 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1332 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1336 * Move request from internal lists to the request queue dispatch list.
1338 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1340 struct cfq_data *cfqd = q->elevator->elevator_data;
1341 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1343 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1345 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1346 cfq_remove_request(rq);
1348 elv_dispatch_sort(q, rq);
1350 if (cfq_cfqq_sync(cfqq))
1351 cfqd->sync_flight++;
1355 * return expired entry, or NULL to just start from scratch in rbtree
1357 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1359 struct request *rq = NULL;
1361 if (cfq_cfqq_fifo_expire(cfqq))
1364 cfq_mark_cfqq_fifo_expire(cfqq);
1366 if (list_empty(&cfqq->fifo))
1369 rq = rq_entry_fifo(cfqq->fifo.next);
1370 if (time_before(jiffies, rq_fifo_time(rq)))
1373 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1378 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1380 const int base_rq = cfqd->cfq_slice_async_rq;
1382 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1384 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1388 * Must be called with the queue_lock held.
1390 static int cfqq_process_refs(struct cfq_queue *cfqq)
1392 int process_refs, io_refs;
1394 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1395 process_refs = atomic_read(&cfqq->ref) - io_refs;
1396 BUG_ON(process_refs < 0);
1397 return process_refs;
1400 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1402 int process_refs, new_process_refs;
1403 struct cfq_queue *__cfqq;
1405 /* Avoid a circular list and skip interim queue merges */
1406 while ((__cfqq = new_cfqq->new_cfqq)) {
1412 process_refs = cfqq_process_refs(cfqq);
1414 * If the process for the cfqq has gone away, there is no
1415 * sense in merging the queues.
1417 if (process_refs == 0)
1421 * Merge in the direction of the lesser amount of work.
1423 new_process_refs = cfqq_process_refs(new_cfqq);
1424 if (new_process_refs >= process_refs) {
1425 cfqq->new_cfqq = new_cfqq;
1426 atomic_add(process_refs, &new_cfqq->ref);
1428 new_cfqq->new_cfqq = cfqq;
1429 atomic_add(new_process_refs, &cfqq->ref);
1433 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1434 struct cfq_group *cfqg, enum wl_prio_t prio,
1437 struct cfq_queue *queue;
1439 bool key_valid = false;
1440 unsigned long lowest_key = 0;
1441 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1445 * When priorities switched, we prefer starting
1446 * from SYNC_NOIDLE (first choice), or just SYNC
1449 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1451 cur_best = SYNC_WORKLOAD;
1452 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1455 return ASYNC_WORKLOAD;
1458 for (i = 0; i < 3; ++i) {
1459 /* otherwise, select the one with lowest rb_key */
1460 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1462 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1463 lowest_key = queue->rb_key;
1472 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1474 enum wl_prio_t previous_prio = cfqd->serving_prio;
1478 struct cfq_rb_root *st;
1480 /* Choose next priority. RT > BE > IDLE */
1481 if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1482 cfqd->serving_prio = RT_WORKLOAD;
1483 else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1484 cfqd->serving_prio = BE_WORKLOAD;
1486 cfqd->serving_prio = IDLE_WORKLOAD;
1487 cfqd->workload_expires = jiffies + 1;
1492 * For RT and BE, we have to choose also the type
1493 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1496 prio_changed = (cfqd->serving_prio != previous_prio);
1497 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1502 * If priority didn't change, check workload expiration,
1503 * and that we still have other queues ready
1505 if (!prio_changed && count &&
1506 !time_after(jiffies, cfqd->workload_expires))
1509 /* otherwise select new workload type */
1510 cfqd->serving_type =
1511 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
1512 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1517 * the workload slice is computed as a fraction of target latency
1518 * proportional to the number of queues in that workload, over
1519 * all the queues in the same priority class
1521 slice = cfq_target_latency * count /
1522 max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio],
1523 cfq_busy_queues_wl(cfqd->serving_prio, cfqd));
1525 if (cfqd->serving_type == ASYNC_WORKLOAD)
1526 /* async workload slice is scaled down according to
1527 * the sync/async slice ratio. */
1528 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1530 /* sync workload slice is at least 2 * cfq_slice_idle */
1531 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1533 slice = max_t(unsigned, slice, CFQ_MIN_TT);
1534 cfqd->workload_expires = jiffies + slice;
1535 cfqd->noidle_tree_requires_idle = false;
1538 static void cfq_choose_cfqg(struct cfq_data *cfqd)
1540 cfqd->serving_group = &cfqd->root_group;
1541 choose_service_tree(cfqd, &cfqd->root_group);
1545 * Select a queue for service. If we have a current active queue,
1546 * check whether to continue servicing it, or retrieve and set a new one.
1548 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1550 struct cfq_queue *cfqq, *new_cfqq = NULL;
1552 cfqq = cfqd->active_queue;
1556 if (!cfqd->rq_queued)
1559 * The active queue has run out of time, expire it and select new.
1561 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1565 * The active queue has requests and isn't expired, allow it to
1568 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1572 * If another queue has a request waiting within our mean seek
1573 * distance, let it run. The expire code will check for close
1574 * cooperators and put the close queue at the front of the service
1575 * tree. If possible, merge the expiring queue with the new cfqq.
1577 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1579 if (!cfqq->new_cfqq)
1580 cfq_setup_merge(cfqq, new_cfqq);
1585 * No requests pending. If the active queue still has requests in
1586 * flight or is idling for a new request, allow either of these
1587 * conditions to happen (or time out) before selecting a new queue.
1589 if (timer_pending(&cfqd->idle_slice_timer) ||
1590 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1596 cfq_slice_expired(cfqd, 0);
1599 * Current queue expired. Check if we have to switch to a new
1603 cfq_choose_cfqg(cfqd);
1605 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1610 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1614 while (cfqq->next_rq) {
1615 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1619 BUG_ON(!list_empty(&cfqq->fifo));
1621 /* By default cfqq is not expired if it is empty. Do it explicitly */
1622 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
1627 * Drain our current requests. Used for barriers and when switching
1628 * io schedulers on-the-fly.
1630 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1632 struct cfq_queue *cfqq;
1635 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
1636 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1638 cfq_slice_expired(cfqd, 0);
1639 BUG_ON(cfqd->busy_queues);
1641 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1645 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1647 unsigned int max_dispatch;
1650 * Drain async requests before we start sync IO
1652 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1656 * If this is an async queue and we have sync IO in flight, let it wait
1658 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1661 max_dispatch = cfqd->cfq_quantum;
1662 if (cfq_class_idle(cfqq))
1666 * Does this cfqq already have too much IO in flight?
1668 if (cfqq->dispatched >= max_dispatch) {
1670 * idle queue must always only have a single IO in flight
1672 if (cfq_class_idle(cfqq))
1676 * We have other queues, don't allow more IO from this one
1678 if (cfqd->busy_queues > 1)
1682 * Sole queue user, no limit
1688 * Async queues must wait a bit before being allowed dispatch.
1689 * We also ramp up the dispatch depth gradually for async IO,
1690 * based on the last sync IO we serviced
1692 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1693 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1696 depth = last_sync / cfqd->cfq_slice[1];
1697 if (!depth && !cfqq->dispatched)
1699 if (depth < max_dispatch)
1700 max_dispatch = depth;
1704 * If we're below the current max, allow a dispatch
1706 return cfqq->dispatched < max_dispatch;
1710 * Dispatch a request from cfqq, moving them to the request queue
1713 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1717 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1719 if (!cfq_may_dispatch(cfqd, cfqq))
1723 * follow expired path, else get first next available
1725 rq = cfq_check_fifo(cfqq);
1730 * insert request into driver dispatch list
1732 cfq_dispatch_insert(cfqd->queue, rq);
1734 if (!cfqd->active_cic) {
1735 struct cfq_io_context *cic = RQ_CIC(rq);
1737 atomic_long_inc(&cic->ioc->refcount);
1738 cfqd->active_cic = cic;
1745 * Find the cfqq that we need to service and move a request from that to the
1748 static int cfq_dispatch_requests(struct request_queue *q, int force)
1750 struct cfq_data *cfqd = q->elevator->elevator_data;
1751 struct cfq_queue *cfqq;
1753 if (!cfqd->busy_queues)
1756 if (unlikely(force))
1757 return cfq_forced_dispatch(cfqd);
1759 cfqq = cfq_select_queue(cfqd);
1764 * Dispatch a request from this cfqq, if it is allowed
1766 if (!cfq_dispatch_request(cfqd, cfqq))
1769 cfqq->slice_dispatch++;
1770 cfq_clear_cfqq_must_dispatch(cfqq);
1773 * expire an async queue immediately if it has used up its slice. idle
1774 * queue always expire after 1 dispatch round.
1776 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1777 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1778 cfq_class_idle(cfqq))) {
1779 cfqq->slice_end = jiffies + 1;
1780 cfq_slice_expired(cfqd, 0);
1783 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1788 * task holds one reference to the queue, dropped when task exits. each rq
1789 * in-flight on this queue also holds a reference, dropped when rq is freed.
1791 * queue lock must be held here.
1793 static void cfq_put_queue(struct cfq_queue *cfqq)
1795 struct cfq_data *cfqd = cfqq->cfqd;
1797 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1799 if (!atomic_dec_and_test(&cfqq->ref))
1802 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1803 BUG_ON(rb_first(&cfqq->sort_list));
1804 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1806 if (unlikely(cfqd->active_queue == cfqq)) {
1807 __cfq_slice_expired(cfqd, cfqq, 0);
1808 cfq_schedule_dispatch(cfqd);
1811 BUG_ON(cfq_cfqq_on_rr(cfqq));
1812 kmem_cache_free(cfq_pool, cfqq);
1816 * Must always be called with the rcu_read_lock() held
1819 __call_for_each_cic(struct io_context *ioc,
1820 void (*func)(struct io_context *, struct cfq_io_context *))
1822 struct cfq_io_context *cic;
1823 struct hlist_node *n;
1825 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1830 * Call func for each cic attached to this ioc.
1833 call_for_each_cic(struct io_context *ioc,
1834 void (*func)(struct io_context *, struct cfq_io_context *))
1837 __call_for_each_cic(ioc, func);
1841 static void cfq_cic_free_rcu(struct rcu_head *head)
1843 struct cfq_io_context *cic;
1845 cic = container_of(head, struct cfq_io_context, rcu_head);
1847 kmem_cache_free(cfq_ioc_pool, cic);
1848 elv_ioc_count_dec(cfq_ioc_count);
1852 * CFQ scheduler is exiting, grab exit lock and check
1853 * the pending io context count. If it hits zero,
1854 * complete ioc_gone and set it back to NULL
1856 spin_lock(&ioc_gone_lock);
1857 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1861 spin_unlock(&ioc_gone_lock);
1865 static void cfq_cic_free(struct cfq_io_context *cic)
1867 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1870 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1872 unsigned long flags;
1874 BUG_ON(!cic->dead_key);
1876 spin_lock_irqsave(&ioc->lock, flags);
1877 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1878 hlist_del_rcu(&cic->cic_list);
1879 spin_unlock_irqrestore(&ioc->lock, flags);
1885 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1886 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1887 * and ->trim() which is called with the task lock held
1889 static void cfq_free_io_context(struct io_context *ioc)
1892 * ioc->refcount is zero here, or we are called from elv_unregister(),
1893 * so no more cic's are allowed to be linked into this ioc. So it
1894 * should be ok to iterate over the known list, we will see all cic's
1895 * since no new ones are added.
1897 __call_for_each_cic(ioc, cic_free_func);
1900 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1902 struct cfq_queue *__cfqq, *next;
1904 if (unlikely(cfqq == cfqd->active_queue)) {
1905 __cfq_slice_expired(cfqd, cfqq, 0);
1906 cfq_schedule_dispatch(cfqd);
1910 * If this queue was scheduled to merge with another queue, be
1911 * sure to drop the reference taken on that queue (and others in
1912 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1914 __cfqq = cfqq->new_cfqq;
1916 if (__cfqq == cfqq) {
1917 WARN(1, "cfqq->new_cfqq loop detected\n");
1920 next = __cfqq->new_cfqq;
1921 cfq_put_queue(__cfqq);
1925 cfq_put_queue(cfqq);
1928 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1929 struct cfq_io_context *cic)
1931 struct io_context *ioc = cic->ioc;
1933 list_del_init(&cic->queue_list);
1936 * Make sure key == NULL is seen for dead queues
1939 cic->dead_key = (unsigned long) cic->key;
1942 if (ioc->ioc_data == cic)
1943 rcu_assign_pointer(ioc->ioc_data, NULL);
1945 if (cic->cfqq[BLK_RW_ASYNC]) {
1946 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1947 cic->cfqq[BLK_RW_ASYNC] = NULL;
1950 if (cic->cfqq[BLK_RW_SYNC]) {
1951 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1952 cic->cfqq[BLK_RW_SYNC] = NULL;
1956 static void cfq_exit_single_io_context(struct io_context *ioc,
1957 struct cfq_io_context *cic)
1959 struct cfq_data *cfqd = cic->key;
1962 struct request_queue *q = cfqd->queue;
1963 unsigned long flags;
1965 spin_lock_irqsave(q->queue_lock, flags);
1968 * Ensure we get a fresh copy of the ->key to prevent
1969 * race between exiting task and queue
1971 smp_read_barrier_depends();
1973 __cfq_exit_single_io_context(cfqd, cic);
1975 spin_unlock_irqrestore(q->queue_lock, flags);
1980 * The process that ioc belongs to has exited, we need to clean up
1981 * and put the internal structures we have that belongs to that process.
1983 static void cfq_exit_io_context(struct io_context *ioc)
1985 call_for_each_cic(ioc, cfq_exit_single_io_context);
1988 static struct cfq_io_context *
1989 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1991 struct cfq_io_context *cic;
1993 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1996 cic->last_end_request = jiffies;
1997 INIT_LIST_HEAD(&cic->queue_list);
1998 INIT_HLIST_NODE(&cic->cic_list);
1999 cic->dtor = cfq_free_io_context;
2000 cic->exit = cfq_exit_io_context;
2001 elv_ioc_count_inc(cfq_ioc_count);
2007 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2009 struct task_struct *tsk = current;
2012 if (!cfq_cfqq_prio_changed(cfqq))
2015 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2016 switch (ioprio_class) {
2018 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2019 case IOPRIO_CLASS_NONE:
2021 * no prio set, inherit CPU scheduling settings
2023 cfqq->ioprio = task_nice_ioprio(tsk);
2024 cfqq->ioprio_class = task_nice_ioclass(tsk);
2026 case IOPRIO_CLASS_RT:
2027 cfqq->ioprio = task_ioprio(ioc);
2028 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2030 case IOPRIO_CLASS_BE:
2031 cfqq->ioprio = task_ioprio(ioc);
2032 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2034 case IOPRIO_CLASS_IDLE:
2035 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2037 cfq_clear_cfqq_idle_window(cfqq);
2042 * keep track of original prio settings in case we have to temporarily
2043 * elevate the priority of this queue
2045 cfqq->org_ioprio = cfqq->ioprio;
2046 cfqq->org_ioprio_class = cfqq->ioprio_class;
2047 cfq_clear_cfqq_prio_changed(cfqq);
2050 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2052 struct cfq_data *cfqd = cic->key;
2053 struct cfq_queue *cfqq;
2054 unsigned long flags;
2056 if (unlikely(!cfqd))
2059 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2061 cfqq = cic->cfqq[BLK_RW_ASYNC];
2063 struct cfq_queue *new_cfqq;
2064 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2067 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2068 cfq_put_queue(cfqq);
2072 cfqq = cic->cfqq[BLK_RW_SYNC];
2074 cfq_mark_cfqq_prio_changed(cfqq);
2076 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2079 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2081 call_for_each_cic(ioc, changed_ioprio);
2082 ioc->ioprio_changed = 0;
2085 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2086 pid_t pid, bool is_sync)
2088 RB_CLEAR_NODE(&cfqq->rb_node);
2089 RB_CLEAR_NODE(&cfqq->p_node);
2090 INIT_LIST_HEAD(&cfqq->fifo);
2092 atomic_set(&cfqq->ref, 0);
2095 cfq_mark_cfqq_prio_changed(cfqq);
2098 if (!cfq_class_idle(cfqq))
2099 cfq_mark_cfqq_idle_window(cfqq);
2100 cfq_mark_cfqq_sync(cfqq);
2105 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
2110 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
2112 return &cfqd->root_group;
2115 static struct cfq_queue *
2116 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2117 struct io_context *ioc, gfp_t gfp_mask)
2119 struct cfq_queue *cfqq, *new_cfqq = NULL;
2120 struct cfq_io_context *cic;
2121 struct cfq_group *cfqg;
2124 cfqg = cfq_get_cfqg(cfqd, 1);
2125 cic = cfq_cic_lookup(cfqd, ioc);
2126 /* cic always exists here */
2127 cfqq = cic_to_cfqq(cic, is_sync);
2130 * Always try a new alloc if we fell back to the OOM cfqq
2131 * originally, since it should just be a temporary situation.
2133 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2138 } else if (gfp_mask & __GFP_WAIT) {
2139 spin_unlock_irq(cfqd->queue->queue_lock);
2140 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2141 gfp_mask | __GFP_ZERO,
2143 spin_lock_irq(cfqd->queue->queue_lock);
2147 cfqq = kmem_cache_alloc_node(cfq_pool,
2148 gfp_mask | __GFP_ZERO,
2153 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2154 cfq_init_prio_data(cfqq, ioc);
2155 cfq_link_cfqq_cfqg(cfqq, cfqg);
2156 cfq_log_cfqq(cfqd, cfqq, "alloced");
2158 cfqq = &cfqd->oom_cfqq;
2162 kmem_cache_free(cfq_pool, new_cfqq);
2167 static struct cfq_queue **
2168 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2170 switch (ioprio_class) {
2171 case IOPRIO_CLASS_RT:
2172 return &cfqd->async_cfqq[0][ioprio];
2173 case IOPRIO_CLASS_BE:
2174 return &cfqd->async_cfqq[1][ioprio];
2175 case IOPRIO_CLASS_IDLE:
2176 return &cfqd->async_idle_cfqq;
2182 static struct cfq_queue *
2183 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2186 const int ioprio = task_ioprio(ioc);
2187 const int ioprio_class = task_ioprio_class(ioc);
2188 struct cfq_queue **async_cfqq = NULL;
2189 struct cfq_queue *cfqq = NULL;
2192 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2197 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2200 * pin the queue now that it's allocated, scheduler exit will prune it
2202 if (!is_sync && !(*async_cfqq)) {
2203 atomic_inc(&cfqq->ref);
2207 atomic_inc(&cfqq->ref);
2212 * We drop cfq io contexts lazily, so we may find a dead one.
2215 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2216 struct cfq_io_context *cic)
2218 unsigned long flags;
2220 WARN_ON(!list_empty(&cic->queue_list));
2222 spin_lock_irqsave(&ioc->lock, flags);
2224 BUG_ON(ioc->ioc_data == cic);
2226 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2227 hlist_del_rcu(&cic->cic_list);
2228 spin_unlock_irqrestore(&ioc->lock, flags);
2233 static struct cfq_io_context *
2234 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2236 struct cfq_io_context *cic;
2237 unsigned long flags;
2246 * we maintain a last-hit cache, to avoid browsing over the tree
2248 cic = rcu_dereference(ioc->ioc_data);
2249 if (cic && cic->key == cfqd) {
2255 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2259 /* ->key must be copied to avoid race with cfq_exit_queue() */
2262 cfq_drop_dead_cic(cfqd, ioc, cic);
2267 spin_lock_irqsave(&ioc->lock, flags);
2268 rcu_assign_pointer(ioc->ioc_data, cic);
2269 spin_unlock_irqrestore(&ioc->lock, flags);
2277 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2278 * the process specific cfq io context when entered from the block layer.
2279 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2281 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2282 struct cfq_io_context *cic, gfp_t gfp_mask)
2284 unsigned long flags;
2287 ret = radix_tree_preload(gfp_mask);
2292 spin_lock_irqsave(&ioc->lock, flags);
2293 ret = radix_tree_insert(&ioc->radix_root,
2294 (unsigned long) cfqd, cic);
2296 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2297 spin_unlock_irqrestore(&ioc->lock, flags);
2299 radix_tree_preload_end();
2302 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2303 list_add(&cic->queue_list, &cfqd->cic_list);
2304 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2309 printk(KERN_ERR "cfq: cic link failed!\n");
2315 * Setup general io context and cfq io context. There can be several cfq
2316 * io contexts per general io context, if this process is doing io to more
2317 * than one device managed by cfq.
2319 static struct cfq_io_context *
2320 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2322 struct io_context *ioc = NULL;
2323 struct cfq_io_context *cic;
2325 might_sleep_if(gfp_mask & __GFP_WAIT);
2327 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2331 cic = cfq_cic_lookup(cfqd, ioc);
2335 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2339 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2343 smp_read_barrier_depends();
2344 if (unlikely(ioc->ioprio_changed))
2345 cfq_ioc_set_ioprio(ioc);
2351 put_io_context(ioc);
2356 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2358 unsigned long elapsed = jiffies - cic->last_end_request;
2359 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2361 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2362 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2363 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2367 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2373 if (!cfqq->last_request_pos)
2375 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2376 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2378 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2381 * Don't allow the seek distance to get too large from the
2382 * odd fragment, pagein, etc
2384 if (cfqq->seek_samples <= 60) /* second&third seek */
2385 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2387 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2389 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2390 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2391 total = cfqq->seek_total + (cfqq->seek_samples/2);
2392 do_div(total, cfqq->seek_samples);
2393 cfqq->seek_mean = (sector_t)total;
2396 * If this cfqq is shared between multiple processes, check to
2397 * make sure that those processes are still issuing I/Os within
2398 * the mean seek distance. If not, it may be time to break the
2399 * queues apart again.
2401 if (cfq_cfqq_coop(cfqq)) {
2402 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2403 cfqq->seeky_start = jiffies;
2404 else if (!CFQQ_SEEKY(cfqq))
2405 cfqq->seeky_start = 0;
2410 * Disable idle window if the process thinks too long or seeks so much that
2414 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2415 struct cfq_io_context *cic)
2417 int old_idle, enable_idle;
2420 * Don't idle for async or idle io prio class
2422 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2425 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2427 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2428 cfq_mark_cfqq_deep(cfqq);
2430 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2431 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
2432 && CFQQ_SEEKY(cfqq)))
2434 else if (sample_valid(cic->ttime_samples)) {
2435 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2441 if (old_idle != enable_idle) {
2442 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2444 cfq_mark_cfqq_idle_window(cfqq);
2446 cfq_clear_cfqq_idle_window(cfqq);
2451 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2452 * no or if we aren't sure, a 1 will cause a preempt.
2455 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2458 struct cfq_queue *cfqq;
2460 cfqq = cfqd->active_queue;
2464 if (cfq_slice_used(cfqq))
2467 if (cfq_class_idle(new_cfqq))
2470 if (cfq_class_idle(cfqq))
2473 /* Allow preemption only if we are idling on sync-noidle tree */
2474 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
2475 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
2476 new_cfqq->service_tree->count == 2 &&
2477 RB_EMPTY_ROOT(&cfqq->sort_list))
2481 * if the new request is sync, but the currently running queue is
2482 * not, let the sync request have priority.
2484 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2488 * So both queues are sync. Let the new request get disk time if
2489 * it's a metadata request and the current queue is doing regular IO.
2491 if (rq_is_meta(rq) && !cfqq->meta_pending)
2495 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2497 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2500 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2504 * if this request is as-good as one we would expect from the
2505 * current cfqq, let it preempt
2507 if (cfq_rq_close(cfqd, cfqq, rq))
2514 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2515 * let it have half of its nominal slice.
2517 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2519 cfq_log_cfqq(cfqd, cfqq, "preempt");
2520 cfq_slice_expired(cfqd, 1);
2523 * Put the new queue at the front of the of the current list,
2524 * so we know that it will be selected next.
2526 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2528 cfq_service_tree_add(cfqd, cfqq, 1);
2530 cfqq->slice_end = 0;
2531 cfq_mark_cfqq_slice_new(cfqq);
2535 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2536 * something we should do about it
2539 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2542 struct cfq_io_context *cic = RQ_CIC(rq);
2546 cfqq->meta_pending++;
2548 cfq_update_io_thinktime(cfqd, cic);
2549 cfq_update_io_seektime(cfqd, cfqq, rq);
2550 cfq_update_idle_window(cfqd, cfqq, cic);
2552 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2554 if (cfqq == cfqd->active_queue) {
2556 * Remember that we saw a request from this process, but
2557 * don't start queuing just yet. Otherwise we risk seeing lots
2558 * of tiny requests, because we disrupt the normal plugging
2559 * and merging. If the request is already larger than a single
2560 * page, let it rip immediately. For that case we assume that
2561 * merging is already done. Ditto for a busy system that
2562 * has other work pending, don't risk delaying until the
2563 * idle timer unplug to continue working.
2565 if (cfq_cfqq_wait_request(cfqq)) {
2566 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2567 cfqd->busy_queues > 1) {
2568 del_timer(&cfqd->idle_slice_timer);
2569 __blk_run_queue(cfqd->queue);
2571 cfq_mark_cfqq_must_dispatch(cfqq);
2573 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2575 * not the active queue - expire current slice if it is
2576 * idle and has expired it's mean thinktime or this new queue
2577 * has some old slice time left and is of higher priority or
2578 * this new queue is RT and the current one is BE
2580 cfq_preempt_queue(cfqd, cfqq);
2581 __blk_run_queue(cfqd->queue);
2585 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2587 struct cfq_data *cfqd = q->elevator->elevator_data;
2588 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2590 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2591 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2593 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2594 list_add_tail(&rq->queuelist, &cfqq->fifo);
2597 cfq_rq_enqueued(cfqd, cfqq, rq);
2601 * Update hw_tag based on peak queue depth over 50 samples under
2604 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2606 struct cfq_queue *cfqq = cfqd->active_queue;
2608 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
2609 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
2611 if (cfqd->hw_tag == 1)
2614 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2615 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2619 * If active queue hasn't enough requests and can idle, cfq might not
2620 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2623 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2624 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2625 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2628 if (cfqd->hw_tag_samples++ < 50)
2631 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
2637 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2639 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2640 struct cfq_data *cfqd = cfqq->cfqd;
2641 const int sync = rq_is_sync(rq);
2645 cfq_log_cfqq(cfqd, cfqq, "complete");
2647 cfq_update_hw_tag(cfqd);
2649 WARN_ON(!cfqd->rq_in_driver[sync]);
2650 WARN_ON(!cfqq->dispatched);
2651 cfqd->rq_in_driver[sync]--;
2654 if (cfq_cfqq_sync(cfqq))
2655 cfqd->sync_flight--;
2658 RQ_CIC(rq)->last_end_request = now;
2659 cfqd->last_end_sync_rq = now;
2663 * If this is the active queue, check if it needs to be expired,
2664 * or if we want to idle in case it has no pending requests.
2666 if (cfqd->active_queue == cfqq) {
2667 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2669 if (cfq_cfqq_slice_new(cfqq)) {
2670 cfq_set_prio_slice(cfqd, cfqq);
2671 cfq_clear_cfqq_slice_new(cfqq);
2674 * Idling is not enabled on:
2676 * - idle-priority queues
2678 * - queues with still some requests queued
2679 * - when there is a close cooperator
2681 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2682 cfq_slice_expired(cfqd, 1);
2683 else if (sync && cfqq_empty &&
2684 !cfq_close_cooperator(cfqd, cfqq)) {
2685 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
2687 * Idling is enabled for SYNC_WORKLOAD.
2688 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
2689 * only if we processed at least one !rq_noidle request
2691 if (cfqd->serving_type == SYNC_WORKLOAD
2692 || cfqd->noidle_tree_requires_idle)
2693 cfq_arm_slice_timer(cfqd);
2697 if (!rq_in_driver(cfqd))
2698 cfq_schedule_dispatch(cfqd);
2702 * we temporarily boost lower priority queues if they are holding fs exclusive
2703 * resources. they are boosted to normal prio (CLASS_BE/4)
2705 static void cfq_prio_boost(struct cfq_queue *cfqq)
2707 if (has_fs_excl()) {
2709 * boost idle prio on transactions that would lock out other
2710 * users of the filesystem
2712 if (cfq_class_idle(cfqq))
2713 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2714 if (cfqq->ioprio > IOPRIO_NORM)
2715 cfqq->ioprio = IOPRIO_NORM;
2718 * unboost the queue (if needed)
2720 cfqq->ioprio_class = cfqq->org_ioprio_class;
2721 cfqq->ioprio = cfqq->org_ioprio;
2725 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2727 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2728 cfq_mark_cfqq_must_alloc_slice(cfqq);
2729 return ELV_MQUEUE_MUST;
2732 return ELV_MQUEUE_MAY;
2735 static int cfq_may_queue(struct request_queue *q, int rw)
2737 struct cfq_data *cfqd = q->elevator->elevator_data;
2738 struct task_struct *tsk = current;
2739 struct cfq_io_context *cic;
2740 struct cfq_queue *cfqq;
2743 * don't force setup of a queue from here, as a call to may_queue
2744 * does not necessarily imply that a request actually will be queued.
2745 * so just lookup a possibly existing queue, or return 'may queue'
2748 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2750 return ELV_MQUEUE_MAY;
2752 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2754 cfq_init_prio_data(cfqq, cic->ioc);
2755 cfq_prio_boost(cfqq);
2757 return __cfq_may_queue(cfqq);
2760 return ELV_MQUEUE_MAY;
2764 * queue lock held here
2766 static void cfq_put_request(struct request *rq)
2768 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2771 const int rw = rq_data_dir(rq);
2773 BUG_ON(!cfqq->allocated[rw]);
2774 cfqq->allocated[rw]--;
2776 put_io_context(RQ_CIC(rq)->ioc);
2778 rq->elevator_private = NULL;
2779 rq->elevator_private2 = NULL;
2781 cfq_put_queue(cfqq);
2785 static struct cfq_queue *
2786 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2787 struct cfq_queue *cfqq)
2789 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2790 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2791 cfq_mark_cfqq_coop(cfqq->new_cfqq);
2792 cfq_put_queue(cfqq);
2793 return cic_to_cfqq(cic, 1);
2796 static int should_split_cfqq(struct cfq_queue *cfqq)
2798 if (cfqq->seeky_start &&
2799 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2805 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2806 * was the last process referring to said cfqq.
2808 static struct cfq_queue *
2809 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2811 if (cfqq_process_refs(cfqq) == 1) {
2812 cfqq->seeky_start = 0;
2813 cfqq->pid = current->pid;
2814 cfq_clear_cfqq_coop(cfqq);
2818 cic_set_cfqq(cic, NULL, 1);
2819 cfq_put_queue(cfqq);
2823 * Allocate cfq data structures associated with this request.
2826 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2828 struct cfq_data *cfqd = q->elevator->elevator_data;
2829 struct cfq_io_context *cic;
2830 const int rw = rq_data_dir(rq);
2831 const bool is_sync = rq_is_sync(rq);
2832 struct cfq_queue *cfqq;
2833 unsigned long flags;
2835 might_sleep_if(gfp_mask & __GFP_WAIT);
2837 cic = cfq_get_io_context(cfqd, gfp_mask);
2839 spin_lock_irqsave(q->queue_lock, flags);
2845 cfqq = cic_to_cfqq(cic, is_sync);
2846 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2847 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2848 cic_set_cfqq(cic, cfqq, is_sync);
2851 * If the queue was seeky for too long, break it apart.
2853 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2854 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2855 cfqq = split_cfqq(cic, cfqq);
2861 * Check to see if this queue is scheduled to merge with
2862 * another, closely cooperating queue. The merging of
2863 * queues happens here as it must be done in process context.
2864 * The reference on new_cfqq was taken in merge_cfqqs.
2867 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2870 cfqq->allocated[rw]++;
2871 atomic_inc(&cfqq->ref);
2873 spin_unlock_irqrestore(q->queue_lock, flags);
2875 rq->elevator_private = cic;
2876 rq->elevator_private2 = cfqq;
2881 put_io_context(cic->ioc);
2883 cfq_schedule_dispatch(cfqd);
2884 spin_unlock_irqrestore(q->queue_lock, flags);
2885 cfq_log(cfqd, "set_request fail");
2889 static void cfq_kick_queue(struct work_struct *work)
2891 struct cfq_data *cfqd =
2892 container_of(work, struct cfq_data, unplug_work);
2893 struct request_queue *q = cfqd->queue;
2895 spin_lock_irq(q->queue_lock);
2896 __blk_run_queue(cfqd->queue);
2897 spin_unlock_irq(q->queue_lock);
2901 * Timer running if the active_queue is currently idling inside its time slice
2903 static void cfq_idle_slice_timer(unsigned long data)
2905 struct cfq_data *cfqd = (struct cfq_data *) data;
2906 struct cfq_queue *cfqq;
2907 unsigned long flags;
2910 cfq_log(cfqd, "idle timer fired");
2912 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2914 cfqq = cfqd->active_queue;
2919 * We saw a request before the queue expired, let it through
2921 if (cfq_cfqq_must_dispatch(cfqq))
2927 if (cfq_slice_used(cfqq))
2931 * only expire and reinvoke request handler, if there are
2932 * other queues with pending requests
2934 if (!cfqd->busy_queues)
2938 * not expired and it has a request pending, let it dispatch
2940 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2944 * Queue depth flag is reset only when the idle didn't succeed
2946 cfq_clear_cfqq_deep(cfqq);
2949 cfq_slice_expired(cfqd, timed_out);
2951 cfq_schedule_dispatch(cfqd);
2953 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2956 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2958 del_timer_sync(&cfqd->idle_slice_timer);
2959 cancel_work_sync(&cfqd->unplug_work);
2962 static void cfq_put_async_queues(struct cfq_data *cfqd)
2966 for (i = 0; i < IOPRIO_BE_NR; i++) {
2967 if (cfqd->async_cfqq[0][i])
2968 cfq_put_queue(cfqd->async_cfqq[0][i]);
2969 if (cfqd->async_cfqq[1][i])
2970 cfq_put_queue(cfqd->async_cfqq[1][i]);
2973 if (cfqd->async_idle_cfqq)
2974 cfq_put_queue(cfqd->async_idle_cfqq);
2977 static void cfq_exit_queue(struct elevator_queue *e)
2979 struct cfq_data *cfqd = e->elevator_data;
2980 struct request_queue *q = cfqd->queue;
2982 cfq_shutdown_timer_wq(cfqd);
2984 spin_lock_irq(q->queue_lock);
2986 if (cfqd->active_queue)
2987 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2989 while (!list_empty(&cfqd->cic_list)) {
2990 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2991 struct cfq_io_context,
2994 __cfq_exit_single_io_context(cfqd, cic);
2997 cfq_put_async_queues(cfqd);
2999 spin_unlock_irq(q->queue_lock);
3001 cfq_shutdown_timer_wq(cfqd);
3006 static void *cfq_init_queue(struct request_queue *q)
3008 struct cfq_data *cfqd;
3010 struct cfq_group *cfqg;
3011 struct cfq_rb_root *st;
3013 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3017 /* Init root group */
3018 cfqg = &cfqd->root_group;
3019 for_each_cfqg_st(cfqg, i, j, st)
3023 * Not strictly needed (since RB_ROOT just clears the node and we
3024 * zeroed cfqd on alloc), but better be safe in case someone decides
3025 * to add magic to the rb code
3027 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3028 cfqd->prio_trees[i] = RB_ROOT;
3031 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3032 * Grab a permanent reference to it, so that the normal code flow
3033 * will not attempt to free it.
3035 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3036 atomic_inc(&cfqd->oom_cfqq.ref);
3037 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3039 INIT_LIST_HEAD(&cfqd->cic_list);
3043 init_timer(&cfqd->idle_slice_timer);
3044 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3045 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3047 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3049 cfqd->cfq_quantum = cfq_quantum;
3050 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3051 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3052 cfqd->cfq_back_max = cfq_back_max;
3053 cfqd->cfq_back_penalty = cfq_back_penalty;
3054 cfqd->cfq_slice[0] = cfq_slice_async;
3055 cfqd->cfq_slice[1] = cfq_slice_sync;
3056 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3057 cfqd->cfq_slice_idle = cfq_slice_idle;
3058 cfqd->cfq_latency = 1;
3060 cfqd->last_end_sync_rq = jiffies;
3064 static void cfq_slab_kill(void)
3067 * Caller already ensured that pending RCU callbacks are completed,
3068 * so we should have no busy allocations at this point.
3071 kmem_cache_destroy(cfq_pool);
3073 kmem_cache_destroy(cfq_ioc_pool);
3076 static int __init cfq_slab_setup(void)
3078 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3082 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3093 * sysfs parts below -->
3096 cfq_var_show(unsigned int var, char *page)
3098 return sprintf(page, "%d\n", var);
3102 cfq_var_store(unsigned int *var, const char *page, size_t count)
3104 char *p = (char *) page;
3106 *var = simple_strtoul(p, &p, 10);
3110 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3111 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3113 struct cfq_data *cfqd = e->elevator_data; \
3114 unsigned int __data = __VAR; \
3116 __data = jiffies_to_msecs(__data); \
3117 return cfq_var_show(__data, (page)); \
3119 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3120 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3121 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3122 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3123 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3124 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3125 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3126 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3127 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3128 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3129 #undef SHOW_FUNCTION
3131 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3132 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3134 struct cfq_data *cfqd = e->elevator_data; \
3135 unsigned int __data; \
3136 int ret = cfq_var_store(&__data, (page), count); \
3137 if (__data < (MIN)) \
3139 else if (__data > (MAX)) \
3142 *(__PTR) = msecs_to_jiffies(__data); \
3144 *(__PTR) = __data; \
3147 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3148 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3150 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3152 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3153 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3155 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3156 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3157 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3158 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3160 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3161 #undef STORE_FUNCTION
3163 #define CFQ_ATTR(name) \
3164 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3166 static struct elv_fs_entry cfq_attrs[] = {
3168 CFQ_ATTR(fifo_expire_sync),
3169 CFQ_ATTR(fifo_expire_async),
3170 CFQ_ATTR(back_seek_max),
3171 CFQ_ATTR(back_seek_penalty),
3172 CFQ_ATTR(slice_sync),
3173 CFQ_ATTR(slice_async),
3174 CFQ_ATTR(slice_async_rq),
3175 CFQ_ATTR(slice_idle),
3176 CFQ_ATTR(low_latency),
3180 static struct elevator_type iosched_cfq = {
3182 .elevator_merge_fn = cfq_merge,
3183 .elevator_merged_fn = cfq_merged_request,
3184 .elevator_merge_req_fn = cfq_merged_requests,
3185 .elevator_allow_merge_fn = cfq_allow_merge,
3186 .elevator_dispatch_fn = cfq_dispatch_requests,
3187 .elevator_add_req_fn = cfq_insert_request,
3188 .elevator_activate_req_fn = cfq_activate_request,
3189 .elevator_deactivate_req_fn = cfq_deactivate_request,
3190 .elevator_queue_empty_fn = cfq_queue_empty,
3191 .elevator_completed_req_fn = cfq_completed_request,
3192 .elevator_former_req_fn = elv_rb_former_request,
3193 .elevator_latter_req_fn = elv_rb_latter_request,
3194 .elevator_set_req_fn = cfq_set_request,
3195 .elevator_put_req_fn = cfq_put_request,
3196 .elevator_may_queue_fn = cfq_may_queue,
3197 .elevator_init_fn = cfq_init_queue,
3198 .elevator_exit_fn = cfq_exit_queue,
3199 .trim = cfq_free_io_context,
3201 .elevator_attrs = cfq_attrs,
3202 .elevator_name = "cfq",
3203 .elevator_owner = THIS_MODULE,
3206 static int __init cfq_init(void)
3209 * could be 0 on HZ < 1000 setups
3211 if (!cfq_slice_async)
3212 cfq_slice_async = 1;
3213 if (!cfq_slice_idle)
3216 if (cfq_slab_setup())
3219 elv_register(&iosched_cfq);
3224 static void __exit cfq_exit(void)
3226 DECLARE_COMPLETION_ONSTACK(all_gone);
3227 elv_unregister(&iosched_cfq);
3228 ioc_gone = &all_gone;
3229 /* ioc_gone's update must be visible before reading ioc_count */
3233 * this also protects us from entering cfq_slab_kill() with
3234 * pending RCU callbacks
3236 if (elv_ioc_count_read(cfq_ioc_count))
3237 wait_for_completion(&all_gone);
3241 module_init(cfq_init);
3242 module_exit(cfq_exit);
3244 MODULE_AUTHOR("Jens Axboe");
3245 MODULE_LICENSE("GPL");
3246 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");