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/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
61 static struct kmem_cache *cfq_pool;
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
68 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
71 unsigned long last_end_request;
73 unsigned long ttime_total;
74 unsigned long ttime_samples;
75 unsigned long ttime_mean;
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
88 unsigned total_weight;
90 struct cfq_ttime ttime;
92 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
93 .ttime = {.last_end_request = jiffies,},}
96 * Per process-grouping structure
101 /* various state flags, see below */
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
119 /* currently allocated requests */
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
126 unsigned int allocated_slice;
127 unsigned int slice_dispatch;
128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
130 unsigned long slice_end;
133 /* pending priority requests */
135 /* number of requests that are on the dispatch list or inside driver */
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
140 unsigned short ioprio_class;
145 sector_t last_request_pos;
147 struct cfq_rb_root *service_tree;
148 struct cfq_queue *new_cfqq;
149 struct cfq_group *cfqg;
150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
155 * First index in the service_trees.
156 * IDLE is handled separately, so it has negative index
166 * Second index in the service_trees.
170 SYNC_NOIDLE_WORKLOAD = 1,
174 /* This is per cgroup per device grouping structure */
176 /* group service_tree member */
177 struct rb_node rb_node;
179 /* group service_tree key */
182 unsigned int new_weight;
185 /* number of cfqq currently on this group */
189 * Per group busy queues average. Useful for workload slice calc. We
190 * create the array for each prio class but at run time it is used
191 * only for RT and BE class and slot for IDLE class remains unused.
192 * This is primarily done to avoid confusion and a gcc warning.
194 unsigned int busy_queues_avg[CFQ_PRIO_NR];
196 * rr lists of queues with requests. We maintain service trees for
197 * RT and BE classes. These trees are subdivided in subclasses
198 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199 * class there is no subclassification and all the cfq queues go on
200 * a single tree service_tree_idle.
201 * Counts are embedded in the cfq_rb_root
203 struct cfq_rb_root service_trees[2][3];
204 struct cfq_rb_root service_tree_idle;
206 unsigned long saved_workload_slice;
207 enum wl_type_t saved_workload;
208 enum wl_prio_t saved_serving_prio;
209 struct blkio_group blkg;
210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
211 struct hlist_node cfqd_node;
214 /* number of requests that are on the dispatch list or inside driver */
216 struct cfq_ttime ttime;
220 struct io_cq icq; /* must be the first member */
221 struct cfq_queue *cfqq[2];
222 struct cfq_ttime ttime;
226 * Per block device queue structure
229 struct request_queue *queue;
230 /* Root service tree for cfq_groups */
231 struct cfq_rb_root grp_service_tree;
232 struct cfq_group *root_group;
235 * The priority currently being served
237 enum wl_prio_t serving_prio;
238 enum wl_type_t serving_type;
239 unsigned long workload_expires;
240 struct cfq_group *serving_group;
243 * Each priority tree is sorted by next_request position. These
244 * trees are used when determining if two or more queues are
245 * interleaving requests (see cfq_close_cooperator).
247 struct rb_root prio_trees[CFQ_PRIO_LISTS];
249 unsigned int busy_queues;
250 unsigned int busy_sync_queues;
256 * queue-depth detection
262 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
263 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
266 int hw_tag_est_depth;
267 unsigned int hw_tag_samples;
270 * idle window management
272 struct timer_list idle_slice_timer;
273 struct work_struct unplug_work;
275 struct cfq_queue *active_queue;
276 struct cfq_io_cq *active_cic;
279 * async queue for each priority case
281 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
282 struct cfq_queue *async_idle_cfqq;
284 sector_t last_position;
287 * tunables, see top of file
289 unsigned int cfq_quantum;
290 unsigned int cfq_fifo_expire[2];
291 unsigned int cfq_back_penalty;
292 unsigned int cfq_back_max;
293 unsigned int cfq_slice[2];
294 unsigned int cfq_slice_async_rq;
295 unsigned int cfq_slice_idle;
296 unsigned int cfq_group_idle;
297 unsigned int cfq_latency;
300 * Fallback dummy cfqq for extreme OOM conditions
302 struct cfq_queue oom_cfqq;
304 unsigned long last_delayed_sync;
306 /* List of cfq groups being managed on this device*/
307 struct hlist_head cfqg_list;
309 /* Number of groups which are on blkcg->blkg_list */
310 unsigned int nr_blkcg_linked_grps;
313 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
315 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
322 if (prio == IDLE_WORKLOAD)
323 return &cfqg->service_tree_idle;
325 return &cfqg->service_trees[prio][type];
328 enum cfqq_state_flags {
329 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
330 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
331 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
332 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
333 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
334 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
335 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
336 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
337 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
338 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
339 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
340 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
341 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
344 #define CFQ_CFQQ_FNS(name) \
345 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
347 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
349 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
351 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
353 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
355 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
359 CFQ_CFQQ_FNS(wait_request);
360 CFQ_CFQQ_FNS(must_dispatch);
361 CFQ_CFQQ_FNS(must_alloc_slice);
362 CFQ_CFQQ_FNS(fifo_expire);
363 CFQ_CFQQ_FNS(idle_window);
364 CFQ_CFQQ_FNS(prio_changed);
365 CFQ_CFQQ_FNS(slice_new);
368 CFQ_CFQQ_FNS(split_coop);
370 CFQ_CFQQ_FNS(wait_busy);
373 #ifdef CONFIG_CFQ_GROUP_IOSCHED
374 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
376 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
377 blkg_path(&(cfqq)->cfqg->blkg), ##args)
379 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
381 blkg_path(&(cfqg)->blkg), ##args) \
384 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
385 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
386 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
388 #define cfq_log(cfqd, fmt, args...) \
389 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
391 /* Traverses through cfq group service trees */
392 #define for_each_cfqg_st(cfqg, i, j, st) \
393 for (i = 0; i <= IDLE_WORKLOAD; i++) \
394 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
395 : &cfqg->service_tree_idle; \
396 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
397 (i == IDLE_WORKLOAD && j == 0); \
398 j++, st = i < IDLE_WORKLOAD ? \
399 &cfqg->service_trees[i][j]: NULL) \
401 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
402 struct cfq_ttime *ttime, bool group_idle)
405 if (!sample_valid(ttime->ttime_samples))
408 slice = cfqd->cfq_group_idle;
410 slice = cfqd->cfq_slice_idle;
411 return ttime->ttime_mean > slice;
414 static inline bool iops_mode(struct cfq_data *cfqd)
417 * If we are not idling on queues and it is a NCQ drive, parallel
418 * execution of requests is on and measuring time is not possible
419 * in most of the cases until and unless we drive shallower queue
420 * depths and that becomes a performance bottleneck. In such cases
421 * switch to start providing fairness in terms of number of IOs.
423 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
429 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
431 if (cfq_class_idle(cfqq))
432 return IDLE_WORKLOAD;
433 if (cfq_class_rt(cfqq))
439 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
441 if (!cfq_cfqq_sync(cfqq))
442 return ASYNC_WORKLOAD;
443 if (!cfq_cfqq_idle_window(cfqq))
444 return SYNC_NOIDLE_WORKLOAD;
445 return SYNC_WORKLOAD;
448 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
449 struct cfq_data *cfqd,
450 struct cfq_group *cfqg)
452 if (wl == IDLE_WORKLOAD)
453 return cfqg->service_tree_idle.count;
455 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
456 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
457 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
460 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
461 struct cfq_group *cfqg)
463 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
464 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
467 static void cfq_dispatch_insert(struct request_queue *, struct request *);
468 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
469 struct io_context *, gfp_t);
471 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
473 /* cic->icq is the first member, %NULL will convert to %NULL */
474 return container_of(icq, struct cfq_io_cq, icq);
477 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
478 struct io_context *ioc)
481 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
485 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
487 return cic->cfqq[is_sync];
490 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
493 cic->cfqq[is_sync] = cfqq;
496 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
498 return cic->icq.q->elevator->elevator_data;
502 * We regard a request as SYNC, if it's either a read or has the SYNC bit
503 * set (in which case it could also be direct WRITE).
505 static inline bool cfq_bio_sync(struct bio *bio)
507 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
511 * scheduler run of queue, if there are requests pending and no one in the
512 * driver that will restart queueing
514 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
516 if (cfqd->busy_queues) {
517 cfq_log(cfqd, "schedule dispatch");
518 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
523 * Scale schedule slice based on io priority. Use the sync time slice only
524 * if a queue is marked sync and has sync io queued. A sync queue with async
525 * io only, should not get full sync slice length.
527 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
530 const int base_slice = cfqd->cfq_slice[sync];
532 WARN_ON(prio >= IOPRIO_BE_NR);
534 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
538 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
540 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
543 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
545 u64 d = delta << CFQ_SERVICE_SHIFT;
547 d = d * BLKIO_WEIGHT_DEFAULT;
548 do_div(d, cfqg->weight);
552 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
554 s64 delta = (s64)(vdisktime - min_vdisktime);
556 min_vdisktime = vdisktime;
558 return min_vdisktime;
561 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
563 s64 delta = (s64)(vdisktime - min_vdisktime);
565 min_vdisktime = vdisktime;
567 return min_vdisktime;
570 static void update_min_vdisktime(struct cfq_rb_root *st)
572 struct cfq_group *cfqg;
575 cfqg = rb_entry_cfqg(st->left);
576 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
582 * get averaged number of queues of RT/BE priority.
583 * average is updated, with a formula that gives more weight to higher numbers,
584 * to quickly follows sudden increases and decrease slowly
587 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
588 struct cfq_group *cfqg, bool rt)
590 unsigned min_q, max_q;
591 unsigned mult = cfq_hist_divisor - 1;
592 unsigned round = cfq_hist_divisor / 2;
593 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
595 min_q = min(cfqg->busy_queues_avg[rt], busy);
596 max_q = max(cfqg->busy_queues_avg[rt], busy);
597 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
599 return cfqg->busy_queues_avg[rt];
602 static inline unsigned
603 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
605 struct cfq_rb_root *st = &cfqd->grp_service_tree;
607 return cfq_target_latency * cfqg->weight / st->total_weight;
610 static inline unsigned
611 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
613 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
614 if (cfqd->cfq_latency) {
616 * interested queues (we consider only the ones with the same
617 * priority class in the cfq group)
619 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
621 unsigned sync_slice = cfqd->cfq_slice[1];
622 unsigned expect_latency = sync_slice * iq;
623 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
625 if (expect_latency > group_slice) {
626 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
627 /* scale low_slice according to IO priority
628 * and sync vs async */
630 min(slice, base_low_slice * slice / sync_slice);
631 /* the adapted slice value is scaled to fit all iqs
632 * into the target latency */
633 slice = max(slice * group_slice / expect_latency,
641 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
643 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
645 cfqq->slice_start = jiffies;
646 cfqq->slice_end = jiffies + slice;
647 cfqq->allocated_slice = slice;
648 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
652 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
653 * isn't valid until the first request from the dispatch is activated
654 * and the slice time set.
656 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
658 if (cfq_cfqq_slice_new(cfqq))
660 if (time_before(jiffies, cfqq->slice_end))
667 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
668 * We choose the request that is closest to the head right now. Distance
669 * behind the head is penalized and only allowed to a certain extent.
671 static struct request *
672 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
674 sector_t s1, s2, d1 = 0, d2 = 0;
675 unsigned long back_max;
676 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
677 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
678 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
680 if (rq1 == NULL || rq1 == rq2)
685 if (rq_is_sync(rq1) != rq_is_sync(rq2))
686 return rq_is_sync(rq1) ? rq1 : rq2;
688 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
689 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
691 s1 = blk_rq_pos(rq1);
692 s2 = blk_rq_pos(rq2);
695 * by definition, 1KiB is 2 sectors
697 back_max = cfqd->cfq_back_max * 2;
700 * Strict one way elevator _except_ in the case where we allow
701 * short backward seeks which are biased as twice the cost of a
702 * similar forward seek.
706 else if (s1 + back_max >= last)
707 d1 = (last - s1) * cfqd->cfq_back_penalty;
709 wrap |= CFQ_RQ1_WRAP;
713 else if (s2 + back_max >= last)
714 d2 = (last - s2) * cfqd->cfq_back_penalty;
716 wrap |= CFQ_RQ2_WRAP;
718 /* Found required data */
721 * By doing switch() on the bit mask "wrap" we avoid having to
722 * check two variables for all permutations: --> faster!
725 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
741 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
744 * Since both rqs are wrapped,
745 * start with the one that's further behind head
746 * (--> only *one* back seek required),
747 * since back seek takes more time than forward.
757 * The below is leftmost cache rbtree addon
759 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
761 /* Service tree is empty */
766 root->left = rb_first(&root->rb);
769 return rb_entry(root->left, struct cfq_queue, rb_node);
774 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
777 root->left = rb_first(&root->rb);
780 return rb_entry_cfqg(root->left);
785 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
791 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
795 rb_erase_init(n, &root->rb);
800 * would be nice to take fifo expire time into account as well
802 static struct request *
803 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
804 struct request *last)
806 struct rb_node *rbnext = rb_next(&last->rb_node);
807 struct rb_node *rbprev = rb_prev(&last->rb_node);
808 struct request *next = NULL, *prev = NULL;
810 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
813 prev = rb_entry_rq(rbprev);
816 next = rb_entry_rq(rbnext);
818 rbnext = rb_first(&cfqq->sort_list);
819 if (rbnext && rbnext != &last->rb_node)
820 next = rb_entry_rq(rbnext);
823 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
826 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
827 struct cfq_queue *cfqq)
830 * just an approximation, should be ok.
832 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
833 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
837 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
839 return cfqg->vdisktime - st->min_vdisktime;
843 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
845 struct rb_node **node = &st->rb.rb_node;
846 struct rb_node *parent = NULL;
847 struct cfq_group *__cfqg;
848 s64 key = cfqg_key(st, cfqg);
851 while (*node != NULL) {
853 __cfqg = rb_entry_cfqg(parent);
855 if (key < cfqg_key(st, __cfqg))
856 node = &parent->rb_left;
858 node = &parent->rb_right;
864 st->left = &cfqg->rb_node;
866 rb_link_node(&cfqg->rb_node, parent, node);
867 rb_insert_color(&cfqg->rb_node, &st->rb);
871 cfq_update_group_weight(struct cfq_group *cfqg)
873 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
874 if (cfqg->needs_update) {
875 cfqg->weight = cfqg->new_weight;
876 cfqg->needs_update = false;
881 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
883 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
885 cfq_update_group_weight(cfqg);
886 __cfq_group_service_tree_add(st, cfqg);
887 st->total_weight += cfqg->weight;
891 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
893 struct cfq_rb_root *st = &cfqd->grp_service_tree;
894 struct cfq_group *__cfqg;
898 if (!RB_EMPTY_NODE(&cfqg->rb_node))
902 * Currently put the group at the end. Later implement something
903 * so that groups get lesser vtime based on their weights, so that
904 * if group does not loose all if it was not continuously backlogged.
906 n = rb_last(&st->rb);
908 __cfqg = rb_entry_cfqg(n);
909 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
911 cfqg->vdisktime = st->min_vdisktime;
912 cfq_group_service_tree_add(st, cfqg);
916 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
918 st->total_weight -= cfqg->weight;
919 if (!RB_EMPTY_NODE(&cfqg->rb_node))
920 cfq_rb_erase(&cfqg->rb_node, st);
924 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
926 struct cfq_rb_root *st = &cfqd->grp_service_tree;
928 BUG_ON(cfqg->nr_cfqq < 1);
931 /* If there are other cfq queues under this group, don't delete it */
935 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
936 cfq_group_service_tree_del(st, cfqg);
937 cfqg->saved_workload_slice = 0;
938 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
941 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
942 unsigned int *unaccounted_time)
944 unsigned int slice_used;
947 * Queue got expired before even a single request completed or
948 * got expired immediately after first request completion.
950 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
952 * Also charge the seek time incurred to the group, otherwise
953 * if there are mutiple queues in the group, each can dispatch
954 * a single request on seeky media and cause lots of seek time
955 * and group will never know it.
957 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
960 slice_used = jiffies - cfqq->slice_start;
961 if (slice_used > cfqq->allocated_slice) {
962 *unaccounted_time = slice_used - cfqq->allocated_slice;
963 slice_used = cfqq->allocated_slice;
965 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
966 *unaccounted_time += cfqq->slice_start -
967 cfqq->dispatch_start;
973 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
974 struct cfq_queue *cfqq)
976 struct cfq_rb_root *st = &cfqd->grp_service_tree;
977 unsigned int used_sl, charge, unaccounted_sl = 0;
978 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
979 - cfqg->service_tree_idle.count;
982 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
985 charge = cfqq->slice_dispatch;
986 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
987 charge = cfqq->allocated_slice;
989 /* Can't update vdisktime while group is on service tree */
990 cfq_group_service_tree_del(st, cfqg);
991 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
992 /* If a new weight was requested, update now, off tree */
993 cfq_group_service_tree_add(st, cfqg);
995 /* This group is being expired. Save the context */
996 if (time_after(cfqd->workload_expires, jiffies)) {
997 cfqg->saved_workload_slice = cfqd->workload_expires
999 cfqg->saved_workload = cfqd->serving_type;
1000 cfqg->saved_serving_prio = cfqd->serving_prio;
1002 cfqg->saved_workload_slice = 0;
1004 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1006 cfq_log_cfqq(cfqq->cfqd, cfqq,
1007 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1008 used_sl, cfqq->slice_dispatch, charge,
1009 iops_mode(cfqd), cfqq->nr_sectors);
1010 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1012 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1016 * cfq_init_cfqg_base - initialize base part of a cfq_group
1017 * @cfqg: cfq_group to initialize
1019 * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1020 * is enabled or not.
1022 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1024 struct cfq_rb_root *st;
1027 for_each_cfqg_st(cfqg, i, j, st)
1029 RB_CLEAR_NODE(&cfqg->rb_node);
1031 cfqg->ttime.last_end_request = jiffies;
1034 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1035 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1038 return container_of(blkg, struct cfq_group, blkg);
1042 static void cfq_update_blkio_group_weight(struct request_queue *q,
1043 struct blkio_group *blkg,
1044 unsigned int weight)
1046 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1047 cfqg->new_weight = weight;
1048 cfqg->needs_update = true;
1051 static void cfq_link_blkio_group(struct request_queue *q,
1052 struct blkio_group *blkg)
1054 struct cfq_data *cfqd = q->elevator->elevator_data;
1055 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1057 cfqd->nr_blkcg_linked_grps++;
1059 /* Add group on cfqd list */
1060 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1063 static struct blkio_group *cfq_alloc_blkio_group(struct request_queue *q,
1064 struct blkio_cgroup *blkcg)
1066 struct cfq_group *cfqg;
1068 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, q->node);
1072 cfq_init_cfqg_base(cfqg);
1073 cfqg->weight = blkcg->weight;
1076 * Take the initial reference that will be released on destroy
1077 * This can be thought of a joint reference by cgroup and
1078 * elevator which will be dropped by either elevator exit
1079 * or cgroup deletion path depending on who is exiting first.
1087 * Search for the cfq group current task belongs to. request_queue lock must
1090 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1091 struct blkio_cgroup *blkcg)
1093 struct request_queue *q = cfqd->queue;
1094 struct cfq_group *cfqg = NULL;
1096 /* avoid lookup for the common case where there's no blkio cgroup */
1097 if (blkcg == &blkio_root_cgroup) {
1098 cfqg = cfqd->root_group;
1100 struct blkio_group *blkg;
1102 blkg = blkg_lookup_create(blkcg, q, BLKIO_POLICY_PROP, false);
1104 cfqg = cfqg_of_blkg(blkg);
1110 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1116 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1118 /* Currently, all async queues are mapped to root group */
1119 if (!cfq_cfqq_sync(cfqq))
1120 cfqg = cfqq->cfqd->root_group;
1123 /* cfqq reference on cfqg */
1127 static void cfq_put_cfqg(struct cfq_group *cfqg)
1129 struct cfq_rb_root *st;
1132 BUG_ON(cfqg->ref <= 0);
1136 for_each_cfqg_st(cfqg, i, j, st)
1137 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1138 free_percpu(cfqg->blkg.stats_cpu);
1142 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1144 /* Something wrong if we are trying to remove same group twice */
1145 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1147 hlist_del_init(&cfqg->cfqd_node);
1149 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1150 cfqd->nr_blkcg_linked_grps--;
1153 * Put the reference taken at the time of creation so that when all
1154 * queues are gone, group can be destroyed.
1159 static bool cfq_release_cfq_groups(struct cfq_data *cfqd)
1161 struct hlist_node *pos, *n;
1162 struct cfq_group *cfqg;
1165 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1167 * If cgroup removal path got to blk_group first and removed
1168 * it from cgroup list, then it will take care of destroying
1171 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1172 cfq_destroy_cfqg(cfqd, cfqg);
1180 * Blk cgroup controller notification saying that blkio_group object is being
1181 * delinked as associated cgroup object is going away. That also means that
1182 * no new IO will come in this group. So get rid of this group as soon as
1183 * any pending IO in the group is finished.
1185 * This function is called under rcu_read_lock(). key is the rcu protected
1186 * pointer. That means @q is a valid request_queue pointer as long as we
1187 * are rcu read lock.
1189 * @q was fetched from blkio_group under blkio_cgroup->lock. That means
1190 * it should not be NULL as even if elevator was exiting, cgroup deltion
1191 * path got to it first.
1193 static void cfq_unlink_blkio_group(struct request_queue *q,
1194 struct blkio_group *blkg)
1196 struct cfq_data *cfqd = q->elevator->elevator_data;
1197 unsigned long flags;
1199 spin_lock_irqsave(q->queue_lock, flags);
1200 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1201 spin_unlock_irqrestore(q->queue_lock, flags);
1204 static struct elevator_type iosched_cfq;
1206 static bool cfq_clear_queue(struct request_queue *q)
1208 lockdep_assert_held(q->queue_lock);
1210 /* shoot down blkgs iff the current elevator is cfq */
1211 if (!q->elevator || q->elevator->type != &iosched_cfq)
1214 return cfq_release_cfq_groups(q->elevator->elevator_data);
1217 #else /* GROUP_IOSCHED */
1218 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1219 struct blkio_cgroup *blkcg)
1221 return cfqd->root_group;
1224 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1230 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1234 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1235 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1237 #endif /* GROUP_IOSCHED */
1240 * The cfqd->service_trees holds all pending cfq_queue's that have
1241 * requests waiting to be processed. It is sorted in the order that
1242 * we will service the queues.
1244 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1247 struct rb_node **p, *parent;
1248 struct cfq_queue *__cfqq;
1249 unsigned long rb_key;
1250 struct cfq_rb_root *service_tree;
1254 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1256 if (cfq_class_idle(cfqq)) {
1257 rb_key = CFQ_IDLE_DELAY;
1258 parent = rb_last(&service_tree->rb);
1259 if (parent && parent != &cfqq->rb_node) {
1260 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1261 rb_key += __cfqq->rb_key;
1264 } else if (!add_front) {
1266 * Get our rb key offset. Subtract any residual slice
1267 * value carried from last service. A negative resid
1268 * count indicates slice overrun, and this should position
1269 * the next service time further away in the tree.
1271 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1272 rb_key -= cfqq->slice_resid;
1273 cfqq->slice_resid = 0;
1276 __cfqq = cfq_rb_first(service_tree);
1277 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1280 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1283 * same position, nothing more to do
1285 if (rb_key == cfqq->rb_key &&
1286 cfqq->service_tree == service_tree)
1289 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1290 cfqq->service_tree = NULL;
1295 cfqq->service_tree = service_tree;
1296 p = &service_tree->rb.rb_node;
1301 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1304 * sort by key, that represents service time.
1306 if (time_before(rb_key, __cfqq->rb_key))
1309 n = &(*p)->rb_right;
1317 service_tree->left = &cfqq->rb_node;
1319 cfqq->rb_key = rb_key;
1320 rb_link_node(&cfqq->rb_node, parent, p);
1321 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1322 service_tree->count++;
1323 if (add_front || !new_cfqq)
1325 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1328 static struct cfq_queue *
1329 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1330 sector_t sector, struct rb_node **ret_parent,
1331 struct rb_node ***rb_link)
1333 struct rb_node **p, *parent;
1334 struct cfq_queue *cfqq = NULL;
1342 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1345 * Sort strictly based on sector. Smallest to the left,
1346 * largest to the right.
1348 if (sector > blk_rq_pos(cfqq->next_rq))
1349 n = &(*p)->rb_right;
1350 else if (sector < blk_rq_pos(cfqq->next_rq))
1358 *ret_parent = parent;
1364 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1366 struct rb_node **p, *parent;
1367 struct cfq_queue *__cfqq;
1370 rb_erase(&cfqq->p_node, cfqq->p_root);
1371 cfqq->p_root = NULL;
1374 if (cfq_class_idle(cfqq))
1379 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1380 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1381 blk_rq_pos(cfqq->next_rq), &parent, &p);
1383 rb_link_node(&cfqq->p_node, parent, p);
1384 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1386 cfqq->p_root = NULL;
1390 * Update cfqq's position in the service tree.
1392 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1395 * Resorting requires the cfqq to be on the RR list already.
1397 if (cfq_cfqq_on_rr(cfqq)) {
1398 cfq_service_tree_add(cfqd, cfqq, 0);
1399 cfq_prio_tree_add(cfqd, cfqq);
1404 * add to busy list of queues for service, trying to be fair in ordering
1405 * the pending list according to last request service
1407 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1409 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1410 BUG_ON(cfq_cfqq_on_rr(cfqq));
1411 cfq_mark_cfqq_on_rr(cfqq);
1412 cfqd->busy_queues++;
1413 if (cfq_cfqq_sync(cfqq))
1414 cfqd->busy_sync_queues++;
1416 cfq_resort_rr_list(cfqd, cfqq);
1420 * Called when the cfqq no longer has requests pending, remove it from
1423 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1425 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1426 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1427 cfq_clear_cfqq_on_rr(cfqq);
1429 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1430 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1431 cfqq->service_tree = NULL;
1434 rb_erase(&cfqq->p_node, cfqq->p_root);
1435 cfqq->p_root = NULL;
1438 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1439 BUG_ON(!cfqd->busy_queues);
1440 cfqd->busy_queues--;
1441 if (cfq_cfqq_sync(cfqq))
1442 cfqd->busy_sync_queues--;
1446 * rb tree support functions
1448 static void cfq_del_rq_rb(struct request *rq)
1450 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1451 const int sync = rq_is_sync(rq);
1453 BUG_ON(!cfqq->queued[sync]);
1454 cfqq->queued[sync]--;
1456 elv_rb_del(&cfqq->sort_list, rq);
1458 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1460 * Queue will be deleted from service tree when we actually
1461 * expire it later. Right now just remove it from prio tree
1465 rb_erase(&cfqq->p_node, cfqq->p_root);
1466 cfqq->p_root = NULL;
1471 static void cfq_add_rq_rb(struct request *rq)
1473 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1474 struct cfq_data *cfqd = cfqq->cfqd;
1475 struct request *prev;
1477 cfqq->queued[rq_is_sync(rq)]++;
1479 elv_rb_add(&cfqq->sort_list, rq);
1481 if (!cfq_cfqq_on_rr(cfqq))
1482 cfq_add_cfqq_rr(cfqd, cfqq);
1485 * check if this request is a better next-serve candidate
1487 prev = cfqq->next_rq;
1488 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1491 * adjust priority tree position, if ->next_rq changes
1493 if (prev != cfqq->next_rq)
1494 cfq_prio_tree_add(cfqd, cfqq);
1496 BUG_ON(!cfqq->next_rq);
1499 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1501 elv_rb_del(&cfqq->sort_list, rq);
1502 cfqq->queued[rq_is_sync(rq)]--;
1503 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1504 rq_data_dir(rq), rq_is_sync(rq));
1506 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1507 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1511 static struct request *
1512 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1514 struct task_struct *tsk = current;
1515 struct cfq_io_cq *cic;
1516 struct cfq_queue *cfqq;
1518 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1522 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1524 sector_t sector = bio->bi_sector + bio_sectors(bio);
1526 return elv_rb_find(&cfqq->sort_list, sector);
1532 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1534 struct cfq_data *cfqd = q->elevator->elevator_data;
1536 cfqd->rq_in_driver++;
1537 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1538 cfqd->rq_in_driver);
1540 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1543 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1545 struct cfq_data *cfqd = q->elevator->elevator_data;
1547 WARN_ON(!cfqd->rq_in_driver);
1548 cfqd->rq_in_driver--;
1549 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1550 cfqd->rq_in_driver);
1553 static void cfq_remove_request(struct request *rq)
1555 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1557 if (cfqq->next_rq == rq)
1558 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1560 list_del_init(&rq->queuelist);
1563 cfqq->cfqd->rq_queued--;
1564 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1565 rq_data_dir(rq), rq_is_sync(rq));
1566 if (rq->cmd_flags & REQ_PRIO) {
1567 WARN_ON(!cfqq->prio_pending);
1568 cfqq->prio_pending--;
1572 static int cfq_merge(struct request_queue *q, struct request **req,
1575 struct cfq_data *cfqd = q->elevator->elevator_data;
1576 struct request *__rq;
1578 __rq = cfq_find_rq_fmerge(cfqd, bio);
1579 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1581 return ELEVATOR_FRONT_MERGE;
1584 return ELEVATOR_NO_MERGE;
1587 static void cfq_merged_request(struct request_queue *q, struct request *req,
1590 if (type == ELEVATOR_FRONT_MERGE) {
1591 struct cfq_queue *cfqq = RQ_CFQQ(req);
1593 cfq_reposition_rq_rb(cfqq, req);
1597 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1600 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1601 bio_data_dir(bio), cfq_bio_sync(bio));
1605 cfq_merged_requests(struct request_queue *q, struct request *rq,
1606 struct request *next)
1608 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1609 struct cfq_data *cfqd = q->elevator->elevator_data;
1612 * reposition in fifo if next is older than rq
1614 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1615 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1616 list_move(&rq->queuelist, &next->queuelist);
1617 rq_set_fifo_time(rq, rq_fifo_time(next));
1620 if (cfqq->next_rq == next)
1622 cfq_remove_request(next);
1623 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1624 rq_data_dir(next), rq_is_sync(next));
1626 cfqq = RQ_CFQQ(next);
1628 * all requests of this queue are merged to other queues, delete it
1629 * from the service tree. If it's the active_queue,
1630 * cfq_dispatch_requests() will choose to expire it or do idle
1632 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1633 cfqq != cfqd->active_queue)
1634 cfq_del_cfqq_rr(cfqd, cfqq);
1637 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1640 struct cfq_data *cfqd = q->elevator->elevator_data;
1641 struct cfq_io_cq *cic;
1642 struct cfq_queue *cfqq;
1645 * Disallow merge of a sync bio into an async request.
1647 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1651 * Lookup the cfqq that this bio will be queued with and allow
1652 * merge only if rq is queued there.
1654 cic = cfq_cic_lookup(cfqd, current->io_context);
1658 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1659 return cfqq == RQ_CFQQ(rq);
1662 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1664 del_timer(&cfqd->idle_slice_timer);
1665 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1668 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1669 struct cfq_queue *cfqq)
1672 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1673 cfqd->serving_prio, cfqd->serving_type);
1674 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1675 cfqq->slice_start = 0;
1676 cfqq->dispatch_start = jiffies;
1677 cfqq->allocated_slice = 0;
1678 cfqq->slice_end = 0;
1679 cfqq->slice_dispatch = 0;
1680 cfqq->nr_sectors = 0;
1682 cfq_clear_cfqq_wait_request(cfqq);
1683 cfq_clear_cfqq_must_dispatch(cfqq);
1684 cfq_clear_cfqq_must_alloc_slice(cfqq);
1685 cfq_clear_cfqq_fifo_expire(cfqq);
1686 cfq_mark_cfqq_slice_new(cfqq);
1688 cfq_del_timer(cfqd, cfqq);
1691 cfqd->active_queue = cfqq;
1695 * current cfqq expired its slice (or was too idle), select new one
1698 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1701 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1703 if (cfq_cfqq_wait_request(cfqq))
1704 cfq_del_timer(cfqd, cfqq);
1706 cfq_clear_cfqq_wait_request(cfqq);
1707 cfq_clear_cfqq_wait_busy(cfqq);
1710 * If this cfqq is shared between multiple processes, check to
1711 * make sure that those processes are still issuing I/Os within
1712 * the mean seek distance. If not, it may be time to break the
1713 * queues apart again.
1715 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1716 cfq_mark_cfqq_split_coop(cfqq);
1719 * store what was left of this slice, if the queue idled/timed out
1722 if (cfq_cfqq_slice_new(cfqq))
1723 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1725 cfqq->slice_resid = cfqq->slice_end - jiffies;
1726 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1729 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1731 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1732 cfq_del_cfqq_rr(cfqd, cfqq);
1734 cfq_resort_rr_list(cfqd, cfqq);
1736 if (cfqq == cfqd->active_queue)
1737 cfqd->active_queue = NULL;
1739 if (cfqd->active_cic) {
1740 put_io_context(cfqd->active_cic->icq.ioc);
1741 cfqd->active_cic = NULL;
1745 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1747 struct cfq_queue *cfqq = cfqd->active_queue;
1750 __cfq_slice_expired(cfqd, cfqq, timed_out);
1754 * Get next queue for service. Unless we have a queue preemption,
1755 * we'll simply select the first cfqq in the service tree.
1757 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1759 struct cfq_rb_root *service_tree =
1760 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1761 cfqd->serving_type);
1763 if (!cfqd->rq_queued)
1766 /* There is nothing to dispatch */
1769 if (RB_EMPTY_ROOT(&service_tree->rb))
1771 return cfq_rb_first(service_tree);
1774 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1776 struct cfq_group *cfqg;
1777 struct cfq_queue *cfqq;
1779 struct cfq_rb_root *st;
1781 if (!cfqd->rq_queued)
1784 cfqg = cfq_get_next_cfqg(cfqd);
1788 for_each_cfqg_st(cfqg, i, j, st)
1789 if ((cfqq = cfq_rb_first(st)) != NULL)
1795 * Get and set a new active queue for service.
1797 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1798 struct cfq_queue *cfqq)
1801 cfqq = cfq_get_next_queue(cfqd);
1803 __cfq_set_active_queue(cfqd, cfqq);
1807 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1810 if (blk_rq_pos(rq) >= cfqd->last_position)
1811 return blk_rq_pos(rq) - cfqd->last_position;
1813 return cfqd->last_position - blk_rq_pos(rq);
1816 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1819 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1822 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1823 struct cfq_queue *cur_cfqq)
1825 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1826 struct rb_node *parent, *node;
1827 struct cfq_queue *__cfqq;
1828 sector_t sector = cfqd->last_position;
1830 if (RB_EMPTY_ROOT(root))
1834 * First, if we find a request starting at the end of the last
1835 * request, choose it.
1837 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1842 * If the exact sector wasn't found, the parent of the NULL leaf
1843 * will contain the closest sector.
1845 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1846 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1849 if (blk_rq_pos(__cfqq->next_rq) < sector)
1850 node = rb_next(&__cfqq->p_node);
1852 node = rb_prev(&__cfqq->p_node);
1856 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1857 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1865 * cur_cfqq - passed in so that we don't decide that the current queue is
1866 * closely cooperating with itself.
1868 * So, basically we're assuming that that cur_cfqq has dispatched at least
1869 * one request, and that cfqd->last_position reflects a position on the disk
1870 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1873 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1874 struct cfq_queue *cur_cfqq)
1876 struct cfq_queue *cfqq;
1878 if (cfq_class_idle(cur_cfqq))
1880 if (!cfq_cfqq_sync(cur_cfqq))
1882 if (CFQQ_SEEKY(cur_cfqq))
1886 * Don't search priority tree if it's the only queue in the group.
1888 if (cur_cfqq->cfqg->nr_cfqq == 1)
1892 * We should notice if some of the queues are cooperating, eg
1893 * working closely on the same area of the disk. In that case,
1894 * we can group them together and don't waste time idling.
1896 cfqq = cfqq_close(cfqd, cur_cfqq);
1900 /* If new queue belongs to different cfq_group, don't choose it */
1901 if (cur_cfqq->cfqg != cfqq->cfqg)
1905 * It only makes sense to merge sync queues.
1907 if (!cfq_cfqq_sync(cfqq))
1909 if (CFQQ_SEEKY(cfqq))
1913 * Do not merge queues of different priority classes
1915 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1922 * Determine whether we should enforce idle window for this queue.
1925 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1927 enum wl_prio_t prio = cfqq_prio(cfqq);
1928 struct cfq_rb_root *service_tree = cfqq->service_tree;
1930 BUG_ON(!service_tree);
1931 BUG_ON(!service_tree->count);
1933 if (!cfqd->cfq_slice_idle)
1936 /* We never do for idle class queues. */
1937 if (prio == IDLE_WORKLOAD)
1940 /* We do for queues that were marked with idle window flag. */
1941 if (cfq_cfqq_idle_window(cfqq) &&
1942 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1946 * Otherwise, we do only if they are the last ones
1947 * in their service tree.
1949 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1950 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1952 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1953 service_tree->count);
1957 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1959 struct cfq_queue *cfqq = cfqd->active_queue;
1960 struct cfq_io_cq *cic;
1961 unsigned long sl, group_idle = 0;
1964 * SSD device without seek penalty, disable idling. But only do so
1965 * for devices that support queuing, otherwise we still have a problem
1966 * with sync vs async workloads.
1968 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1971 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1972 WARN_ON(cfq_cfqq_slice_new(cfqq));
1975 * idle is disabled, either manually or by past process history
1977 if (!cfq_should_idle(cfqd, cfqq)) {
1978 /* no queue idling. Check for group idling */
1979 if (cfqd->cfq_group_idle)
1980 group_idle = cfqd->cfq_group_idle;
1986 * still active requests from this queue, don't idle
1988 if (cfqq->dispatched)
1992 * task has exited, don't wait
1994 cic = cfqd->active_cic;
1995 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
1999 * If our average think time is larger than the remaining time
2000 * slice, then don't idle. This avoids overrunning the allotted
2003 if (sample_valid(cic->ttime.ttime_samples) &&
2004 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2005 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2006 cic->ttime.ttime_mean);
2010 /* There are other queues in the group, don't do group idle */
2011 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2014 cfq_mark_cfqq_wait_request(cfqq);
2017 sl = cfqd->cfq_group_idle;
2019 sl = cfqd->cfq_slice_idle;
2021 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2022 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2023 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2024 group_idle ? 1 : 0);
2028 * Move request from internal lists to the request queue dispatch list.
2030 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2032 struct cfq_data *cfqd = q->elevator->elevator_data;
2033 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2035 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2037 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2038 cfq_remove_request(rq);
2040 (RQ_CFQG(rq))->dispatched++;
2041 elv_dispatch_sort(q, rq);
2043 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2044 cfqq->nr_sectors += blk_rq_sectors(rq);
2045 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2046 rq_data_dir(rq), rq_is_sync(rq));
2050 * return expired entry, or NULL to just start from scratch in rbtree
2052 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2054 struct request *rq = NULL;
2056 if (cfq_cfqq_fifo_expire(cfqq))
2059 cfq_mark_cfqq_fifo_expire(cfqq);
2061 if (list_empty(&cfqq->fifo))
2064 rq = rq_entry_fifo(cfqq->fifo.next);
2065 if (time_before(jiffies, rq_fifo_time(rq)))
2068 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2073 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2075 const int base_rq = cfqd->cfq_slice_async_rq;
2077 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2079 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2083 * Must be called with the queue_lock held.
2085 static int cfqq_process_refs(struct cfq_queue *cfqq)
2087 int process_refs, io_refs;
2089 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2090 process_refs = cfqq->ref - io_refs;
2091 BUG_ON(process_refs < 0);
2092 return process_refs;
2095 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2097 int process_refs, new_process_refs;
2098 struct cfq_queue *__cfqq;
2101 * If there are no process references on the new_cfqq, then it is
2102 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2103 * chain may have dropped their last reference (not just their
2104 * last process reference).
2106 if (!cfqq_process_refs(new_cfqq))
2109 /* Avoid a circular list and skip interim queue merges */
2110 while ((__cfqq = new_cfqq->new_cfqq)) {
2116 process_refs = cfqq_process_refs(cfqq);
2117 new_process_refs = cfqq_process_refs(new_cfqq);
2119 * If the process for the cfqq has gone away, there is no
2120 * sense in merging the queues.
2122 if (process_refs == 0 || new_process_refs == 0)
2126 * Merge in the direction of the lesser amount of work.
2128 if (new_process_refs >= process_refs) {
2129 cfqq->new_cfqq = new_cfqq;
2130 new_cfqq->ref += process_refs;
2132 new_cfqq->new_cfqq = cfqq;
2133 cfqq->ref += new_process_refs;
2137 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2138 struct cfq_group *cfqg, enum wl_prio_t prio)
2140 struct cfq_queue *queue;
2142 bool key_valid = false;
2143 unsigned long lowest_key = 0;
2144 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2146 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2147 /* select the one with lowest rb_key */
2148 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2150 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2151 lowest_key = queue->rb_key;
2160 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2164 struct cfq_rb_root *st;
2165 unsigned group_slice;
2166 enum wl_prio_t original_prio = cfqd->serving_prio;
2168 /* Choose next priority. RT > BE > IDLE */
2169 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2170 cfqd->serving_prio = RT_WORKLOAD;
2171 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2172 cfqd->serving_prio = BE_WORKLOAD;
2174 cfqd->serving_prio = IDLE_WORKLOAD;
2175 cfqd->workload_expires = jiffies + 1;
2179 if (original_prio != cfqd->serving_prio)
2183 * For RT and BE, we have to choose also the type
2184 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2187 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2191 * check workload expiration, and that we still have other queues ready
2193 if (count && !time_after(jiffies, cfqd->workload_expires))
2197 /* otherwise select new workload type */
2198 cfqd->serving_type =
2199 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2200 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2204 * the workload slice is computed as a fraction of target latency
2205 * proportional to the number of queues in that workload, over
2206 * all the queues in the same priority class
2208 group_slice = cfq_group_slice(cfqd, cfqg);
2210 slice = group_slice * count /
2211 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2212 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2214 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2218 * Async queues are currently system wide. Just taking
2219 * proportion of queues with-in same group will lead to higher
2220 * async ratio system wide as generally root group is going
2221 * to have higher weight. A more accurate thing would be to
2222 * calculate system wide asnc/sync ratio.
2224 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2225 tmp = tmp/cfqd->busy_queues;
2226 slice = min_t(unsigned, slice, tmp);
2228 /* async workload slice is scaled down according to
2229 * the sync/async slice ratio. */
2230 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2232 /* sync workload slice is at least 2 * cfq_slice_idle */
2233 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2235 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2236 cfq_log(cfqd, "workload slice:%d", slice);
2237 cfqd->workload_expires = jiffies + slice;
2240 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2242 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2243 struct cfq_group *cfqg;
2245 if (RB_EMPTY_ROOT(&st->rb))
2247 cfqg = cfq_rb_first_group(st);
2248 update_min_vdisktime(st);
2252 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2254 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2256 cfqd->serving_group = cfqg;
2258 /* Restore the workload type data */
2259 if (cfqg->saved_workload_slice) {
2260 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2261 cfqd->serving_type = cfqg->saved_workload;
2262 cfqd->serving_prio = cfqg->saved_serving_prio;
2264 cfqd->workload_expires = jiffies - 1;
2266 choose_service_tree(cfqd, cfqg);
2270 * Select a queue for service. If we have a current active queue,
2271 * check whether to continue servicing it, or retrieve and set a new one.
2273 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2275 struct cfq_queue *cfqq, *new_cfqq = NULL;
2277 cfqq = cfqd->active_queue;
2281 if (!cfqd->rq_queued)
2285 * We were waiting for group to get backlogged. Expire the queue
2287 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2291 * The active queue has run out of time, expire it and select new.
2293 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2295 * If slice had not expired at the completion of last request
2296 * we might not have turned on wait_busy flag. Don't expire
2297 * the queue yet. Allow the group to get backlogged.
2299 * The very fact that we have used the slice, that means we
2300 * have been idling all along on this queue and it should be
2301 * ok to wait for this request to complete.
2303 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2304 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2308 goto check_group_idle;
2312 * The active queue has requests and isn't expired, allow it to
2315 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2319 * If another queue has a request waiting within our mean seek
2320 * distance, let it run. The expire code will check for close
2321 * cooperators and put the close queue at the front of the service
2322 * tree. If possible, merge the expiring queue with the new cfqq.
2324 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2326 if (!cfqq->new_cfqq)
2327 cfq_setup_merge(cfqq, new_cfqq);
2332 * No requests pending. If the active queue still has requests in
2333 * flight or is idling for a new request, allow either of these
2334 * conditions to happen (or time out) before selecting a new queue.
2336 if (timer_pending(&cfqd->idle_slice_timer)) {
2342 * This is a deep seek queue, but the device is much faster than
2343 * the queue can deliver, don't idle
2345 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2346 (cfq_cfqq_slice_new(cfqq) ||
2347 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2348 cfq_clear_cfqq_deep(cfqq);
2349 cfq_clear_cfqq_idle_window(cfqq);
2352 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2358 * If group idle is enabled and there are requests dispatched from
2359 * this group, wait for requests to complete.
2362 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2363 cfqq->cfqg->dispatched &&
2364 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2370 cfq_slice_expired(cfqd, 0);
2373 * Current queue expired. Check if we have to switch to a new
2377 cfq_choose_cfqg(cfqd);
2379 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2384 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2388 while (cfqq->next_rq) {
2389 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2393 BUG_ON(!list_empty(&cfqq->fifo));
2395 /* By default cfqq is not expired if it is empty. Do it explicitly */
2396 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2401 * Drain our current requests. Used for barriers and when switching
2402 * io schedulers on-the-fly.
2404 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2406 struct cfq_queue *cfqq;
2409 /* Expire the timeslice of the current active queue first */
2410 cfq_slice_expired(cfqd, 0);
2411 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2412 __cfq_set_active_queue(cfqd, cfqq);
2413 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2416 BUG_ON(cfqd->busy_queues);
2418 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2422 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2423 struct cfq_queue *cfqq)
2425 /* the queue hasn't finished any request, can't estimate */
2426 if (cfq_cfqq_slice_new(cfqq))
2428 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2435 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2437 unsigned int max_dispatch;
2440 * Drain async requests before we start sync IO
2442 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2446 * If this is an async queue and we have sync IO in flight, let it wait
2448 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2451 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2452 if (cfq_class_idle(cfqq))
2456 * Does this cfqq already have too much IO in flight?
2458 if (cfqq->dispatched >= max_dispatch) {
2459 bool promote_sync = false;
2461 * idle queue must always only have a single IO in flight
2463 if (cfq_class_idle(cfqq))
2467 * If there is only one sync queue
2468 * we can ignore async queue here and give the sync
2469 * queue no dispatch limit. The reason is a sync queue can
2470 * preempt async queue, limiting the sync queue doesn't make
2471 * sense. This is useful for aiostress test.
2473 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2474 promote_sync = true;
2477 * We have other queues, don't allow more IO from this one
2479 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2484 * Sole queue user, no limit
2486 if (cfqd->busy_queues == 1 || promote_sync)
2490 * Normally we start throttling cfqq when cfq_quantum/2
2491 * requests have been dispatched. But we can drive
2492 * deeper queue depths at the beginning of slice
2493 * subjected to upper limit of cfq_quantum.
2495 max_dispatch = cfqd->cfq_quantum;
2499 * Async queues must wait a bit before being allowed dispatch.
2500 * We also ramp up the dispatch depth gradually for async IO,
2501 * based on the last sync IO we serviced
2503 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2504 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2507 depth = last_sync / cfqd->cfq_slice[1];
2508 if (!depth && !cfqq->dispatched)
2510 if (depth < max_dispatch)
2511 max_dispatch = depth;
2515 * If we're below the current max, allow a dispatch
2517 return cfqq->dispatched < max_dispatch;
2521 * Dispatch a request from cfqq, moving them to the request queue
2524 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2528 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2530 if (!cfq_may_dispatch(cfqd, cfqq))
2534 * follow expired path, else get first next available
2536 rq = cfq_check_fifo(cfqq);
2541 * insert request into driver dispatch list
2543 cfq_dispatch_insert(cfqd->queue, rq);
2545 if (!cfqd->active_cic) {
2546 struct cfq_io_cq *cic = RQ_CIC(rq);
2548 atomic_long_inc(&cic->icq.ioc->refcount);
2549 cfqd->active_cic = cic;
2556 * Find the cfqq that we need to service and move a request from that to the
2559 static int cfq_dispatch_requests(struct request_queue *q, int force)
2561 struct cfq_data *cfqd = q->elevator->elevator_data;
2562 struct cfq_queue *cfqq;
2564 if (!cfqd->busy_queues)
2567 if (unlikely(force))
2568 return cfq_forced_dispatch(cfqd);
2570 cfqq = cfq_select_queue(cfqd);
2575 * Dispatch a request from this cfqq, if it is allowed
2577 if (!cfq_dispatch_request(cfqd, cfqq))
2580 cfqq->slice_dispatch++;
2581 cfq_clear_cfqq_must_dispatch(cfqq);
2584 * expire an async queue immediately if it has used up its slice. idle
2585 * queue always expire after 1 dispatch round.
2587 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2588 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2589 cfq_class_idle(cfqq))) {
2590 cfqq->slice_end = jiffies + 1;
2591 cfq_slice_expired(cfqd, 0);
2594 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2599 * task holds one reference to the queue, dropped when task exits. each rq
2600 * in-flight on this queue also holds a reference, dropped when rq is freed.
2602 * Each cfq queue took a reference on the parent group. Drop it now.
2603 * queue lock must be held here.
2605 static void cfq_put_queue(struct cfq_queue *cfqq)
2607 struct cfq_data *cfqd = cfqq->cfqd;
2608 struct cfq_group *cfqg;
2610 BUG_ON(cfqq->ref <= 0);
2616 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2617 BUG_ON(rb_first(&cfqq->sort_list));
2618 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2621 if (unlikely(cfqd->active_queue == cfqq)) {
2622 __cfq_slice_expired(cfqd, cfqq, 0);
2623 cfq_schedule_dispatch(cfqd);
2626 BUG_ON(cfq_cfqq_on_rr(cfqq));
2627 kmem_cache_free(cfq_pool, cfqq);
2631 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2633 struct cfq_queue *__cfqq, *next;
2636 * If this queue was scheduled to merge with another queue, be
2637 * sure to drop the reference taken on that queue (and others in
2638 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2640 __cfqq = cfqq->new_cfqq;
2642 if (__cfqq == cfqq) {
2643 WARN(1, "cfqq->new_cfqq loop detected\n");
2646 next = __cfqq->new_cfqq;
2647 cfq_put_queue(__cfqq);
2652 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2654 if (unlikely(cfqq == cfqd->active_queue)) {
2655 __cfq_slice_expired(cfqd, cfqq, 0);
2656 cfq_schedule_dispatch(cfqd);
2659 cfq_put_cooperator(cfqq);
2661 cfq_put_queue(cfqq);
2664 static void cfq_init_icq(struct io_cq *icq)
2666 struct cfq_io_cq *cic = icq_to_cic(icq);
2668 cic->ttime.last_end_request = jiffies;
2671 static void cfq_exit_icq(struct io_cq *icq)
2673 struct cfq_io_cq *cic = icq_to_cic(icq);
2674 struct cfq_data *cfqd = cic_to_cfqd(cic);
2676 if (cic->cfqq[BLK_RW_ASYNC]) {
2677 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2678 cic->cfqq[BLK_RW_ASYNC] = NULL;
2681 if (cic->cfqq[BLK_RW_SYNC]) {
2682 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2683 cic->cfqq[BLK_RW_SYNC] = NULL;
2687 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2689 struct task_struct *tsk = current;
2692 if (!cfq_cfqq_prio_changed(cfqq))
2695 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2696 switch (ioprio_class) {
2698 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2699 case IOPRIO_CLASS_NONE:
2701 * no prio set, inherit CPU scheduling settings
2703 cfqq->ioprio = task_nice_ioprio(tsk);
2704 cfqq->ioprio_class = task_nice_ioclass(tsk);
2706 case IOPRIO_CLASS_RT:
2707 cfqq->ioprio = task_ioprio(ioc);
2708 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2710 case IOPRIO_CLASS_BE:
2711 cfqq->ioprio = task_ioprio(ioc);
2712 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2714 case IOPRIO_CLASS_IDLE:
2715 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2717 cfq_clear_cfqq_idle_window(cfqq);
2722 * keep track of original prio settings in case we have to temporarily
2723 * elevate the priority of this queue
2725 cfqq->org_ioprio = cfqq->ioprio;
2726 cfq_clear_cfqq_prio_changed(cfqq);
2729 static void changed_ioprio(struct cfq_io_cq *cic)
2731 struct cfq_data *cfqd = cic_to_cfqd(cic);
2732 struct cfq_queue *cfqq;
2734 if (unlikely(!cfqd))
2737 cfqq = cic->cfqq[BLK_RW_ASYNC];
2739 struct cfq_queue *new_cfqq;
2740 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2743 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2744 cfq_put_queue(cfqq);
2748 cfqq = cic->cfqq[BLK_RW_SYNC];
2750 cfq_mark_cfqq_prio_changed(cfqq);
2753 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2754 pid_t pid, bool is_sync)
2756 RB_CLEAR_NODE(&cfqq->rb_node);
2757 RB_CLEAR_NODE(&cfqq->p_node);
2758 INIT_LIST_HEAD(&cfqq->fifo);
2763 cfq_mark_cfqq_prio_changed(cfqq);
2766 if (!cfq_class_idle(cfqq))
2767 cfq_mark_cfqq_idle_window(cfqq);
2768 cfq_mark_cfqq_sync(cfqq);
2773 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2774 static void changed_cgroup(struct cfq_io_cq *cic)
2776 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2777 struct cfq_data *cfqd = cic_to_cfqd(cic);
2778 struct request_queue *q;
2780 if (unlikely(!cfqd))
2787 * Drop reference to sync queue. A new sync queue will be
2788 * assigned in new group upon arrival of a fresh request.
2790 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2791 cic_set_cfqq(cic, NULL, 1);
2792 cfq_put_queue(sync_cfqq);
2795 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2797 static struct cfq_queue *
2798 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2799 struct io_context *ioc, gfp_t gfp_mask)
2801 struct blkio_cgroup *blkcg;
2802 struct cfq_queue *cfqq, *new_cfqq = NULL;
2803 struct cfq_io_cq *cic;
2804 struct cfq_group *cfqg;
2809 blkcg = task_blkio_cgroup(current);
2811 cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
2813 cic = cfq_cic_lookup(cfqd, ioc);
2814 /* cic always exists here */
2815 cfqq = cic_to_cfqq(cic, is_sync);
2818 * Always try a new alloc if we fell back to the OOM cfqq
2819 * originally, since it should just be a temporary situation.
2821 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2826 } else if (gfp_mask & __GFP_WAIT) {
2828 spin_unlock_irq(cfqd->queue->queue_lock);
2829 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2830 gfp_mask | __GFP_ZERO,
2832 spin_lock_irq(cfqd->queue->queue_lock);
2836 cfqq = kmem_cache_alloc_node(cfq_pool,
2837 gfp_mask | __GFP_ZERO,
2842 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2843 cfq_init_prio_data(cfqq, ioc);
2844 cfq_link_cfqq_cfqg(cfqq, cfqg);
2845 cfq_log_cfqq(cfqd, cfqq, "alloced");
2847 cfqq = &cfqd->oom_cfqq;
2851 kmem_cache_free(cfq_pool, new_cfqq);
2857 static struct cfq_queue **
2858 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2860 switch (ioprio_class) {
2861 case IOPRIO_CLASS_RT:
2862 return &cfqd->async_cfqq[0][ioprio];
2863 case IOPRIO_CLASS_BE:
2864 return &cfqd->async_cfqq[1][ioprio];
2865 case IOPRIO_CLASS_IDLE:
2866 return &cfqd->async_idle_cfqq;
2872 static struct cfq_queue *
2873 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2876 const int ioprio = task_ioprio(ioc);
2877 const int ioprio_class = task_ioprio_class(ioc);
2878 struct cfq_queue **async_cfqq = NULL;
2879 struct cfq_queue *cfqq = NULL;
2882 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2887 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2890 * pin the queue now that it's allocated, scheduler exit will prune it
2892 if (!is_sync && !(*async_cfqq)) {
2902 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
2904 unsigned long elapsed = jiffies - ttime->last_end_request;
2905 elapsed = min(elapsed, 2UL * slice_idle);
2907 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2908 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2909 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2913 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2914 struct cfq_io_cq *cic)
2916 if (cfq_cfqq_sync(cfqq)) {
2917 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
2918 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2919 cfqd->cfq_slice_idle);
2921 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2922 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2927 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2931 sector_t n_sec = blk_rq_sectors(rq);
2932 if (cfqq->last_request_pos) {
2933 if (cfqq->last_request_pos < blk_rq_pos(rq))
2934 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2936 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2939 cfqq->seek_history <<= 1;
2940 if (blk_queue_nonrot(cfqd->queue))
2941 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2943 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2947 * Disable idle window if the process thinks too long or seeks so much that
2951 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2952 struct cfq_io_cq *cic)
2954 int old_idle, enable_idle;
2957 * Don't idle for async or idle io prio class
2959 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2962 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2964 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2965 cfq_mark_cfqq_deep(cfqq);
2967 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
2969 else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
2970 !cfqd->cfq_slice_idle ||
2971 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
2973 else if (sample_valid(cic->ttime.ttime_samples)) {
2974 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
2980 if (old_idle != enable_idle) {
2981 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2983 cfq_mark_cfqq_idle_window(cfqq);
2985 cfq_clear_cfqq_idle_window(cfqq);
2990 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2991 * no or if we aren't sure, a 1 will cause a preempt.
2994 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2997 struct cfq_queue *cfqq;
2999 cfqq = cfqd->active_queue;
3003 if (cfq_class_idle(new_cfqq))
3006 if (cfq_class_idle(cfqq))
3010 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3012 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3016 * if the new request is sync, but the currently running queue is
3017 * not, let the sync request have priority.
3019 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3022 if (new_cfqq->cfqg != cfqq->cfqg)
3025 if (cfq_slice_used(cfqq))
3028 /* Allow preemption only if we are idling on sync-noidle tree */
3029 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3030 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3031 new_cfqq->service_tree->count == 2 &&
3032 RB_EMPTY_ROOT(&cfqq->sort_list))
3036 * So both queues are sync. Let the new request get disk time if
3037 * it's a metadata request and the current queue is doing regular IO.
3039 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3043 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3045 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3048 /* An idle queue should not be idle now for some reason */
3049 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3052 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3056 * if this request is as-good as one we would expect from the
3057 * current cfqq, let it preempt
3059 if (cfq_rq_close(cfqd, cfqq, rq))
3066 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3067 * let it have half of its nominal slice.
3069 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3071 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3073 cfq_log_cfqq(cfqd, cfqq, "preempt");
3074 cfq_slice_expired(cfqd, 1);
3077 * workload type is changed, don't save slice, otherwise preempt
3080 if (old_type != cfqq_type(cfqq))
3081 cfqq->cfqg->saved_workload_slice = 0;
3084 * Put the new queue at the front of the of the current list,
3085 * so we know that it will be selected next.
3087 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3089 cfq_service_tree_add(cfqd, cfqq, 1);
3091 cfqq->slice_end = 0;
3092 cfq_mark_cfqq_slice_new(cfqq);
3096 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3097 * something we should do about it
3100 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3103 struct cfq_io_cq *cic = RQ_CIC(rq);
3106 if (rq->cmd_flags & REQ_PRIO)
3107 cfqq->prio_pending++;
3109 cfq_update_io_thinktime(cfqd, cfqq, cic);
3110 cfq_update_io_seektime(cfqd, cfqq, rq);
3111 cfq_update_idle_window(cfqd, cfqq, cic);
3113 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3115 if (cfqq == cfqd->active_queue) {
3117 * Remember that we saw a request from this process, but
3118 * don't start queuing just yet. Otherwise we risk seeing lots
3119 * of tiny requests, because we disrupt the normal plugging
3120 * and merging. If the request is already larger than a single
3121 * page, let it rip immediately. For that case we assume that
3122 * merging is already done. Ditto for a busy system that
3123 * has other work pending, don't risk delaying until the
3124 * idle timer unplug to continue working.
3126 if (cfq_cfqq_wait_request(cfqq)) {
3127 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3128 cfqd->busy_queues > 1) {
3129 cfq_del_timer(cfqd, cfqq);
3130 cfq_clear_cfqq_wait_request(cfqq);
3131 __blk_run_queue(cfqd->queue);
3133 cfq_blkiocg_update_idle_time_stats(
3135 cfq_mark_cfqq_must_dispatch(cfqq);
3138 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3140 * not the active queue - expire current slice if it is
3141 * idle and has expired it's mean thinktime or this new queue
3142 * has some old slice time left and is of higher priority or
3143 * this new queue is RT and the current one is BE
3145 cfq_preempt_queue(cfqd, cfqq);
3146 __blk_run_queue(cfqd->queue);
3150 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3152 struct cfq_data *cfqd = q->elevator->elevator_data;
3153 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3155 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3156 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3158 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3159 list_add_tail(&rq->queuelist, &cfqq->fifo);
3161 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3162 &cfqd->serving_group->blkg, rq_data_dir(rq),
3164 cfq_rq_enqueued(cfqd, cfqq, rq);
3168 * Update hw_tag based on peak queue depth over 50 samples under
3171 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3173 struct cfq_queue *cfqq = cfqd->active_queue;
3175 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3176 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3178 if (cfqd->hw_tag == 1)
3181 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3182 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3186 * If active queue hasn't enough requests and can idle, cfq might not
3187 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3190 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3191 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3192 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3195 if (cfqd->hw_tag_samples++ < 50)
3198 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3204 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3206 struct cfq_io_cq *cic = cfqd->active_cic;
3208 /* If the queue already has requests, don't wait */
3209 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3212 /* If there are other queues in the group, don't wait */
3213 if (cfqq->cfqg->nr_cfqq > 1)
3216 /* the only queue in the group, but think time is big */
3217 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3220 if (cfq_slice_used(cfqq))
3223 /* if slice left is less than think time, wait busy */
3224 if (cic && sample_valid(cic->ttime.ttime_samples)
3225 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3229 * If think times is less than a jiffy than ttime_mean=0 and above
3230 * will not be true. It might happen that slice has not expired yet
3231 * but will expire soon (4-5 ns) during select_queue(). To cover the
3232 * case where think time is less than a jiffy, mark the queue wait
3233 * busy if only 1 jiffy is left in the slice.
3235 if (cfqq->slice_end - jiffies == 1)
3241 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3243 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3244 struct cfq_data *cfqd = cfqq->cfqd;
3245 const int sync = rq_is_sync(rq);
3249 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3250 !!(rq->cmd_flags & REQ_NOIDLE));
3252 cfq_update_hw_tag(cfqd);
3254 WARN_ON(!cfqd->rq_in_driver);
3255 WARN_ON(!cfqq->dispatched);
3256 cfqd->rq_in_driver--;
3258 (RQ_CFQG(rq))->dispatched--;
3259 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3260 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3261 rq_data_dir(rq), rq_is_sync(rq));
3263 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3266 struct cfq_rb_root *service_tree;
3268 RQ_CIC(rq)->ttime.last_end_request = now;
3270 if (cfq_cfqq_on_rr(cfqq))
3271 service_tree = cfqq->service_tree;
3273 service_tree = service_tree_for(cfqq->cfqg,
3274 cfqq_prio(cfqq), cfqq_type(cfqq));
3275 service_tree->ttime.last_end_request = now;
3276 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3277 cfqd->last_delayed_sync = now;
3280 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3281 cfqq->cfqg->ttime.last_end_request = now;
3285 * If this is the active queue, check if it needs to be expired,
3286 * or if we want to idle in case it has no pending requests.
3288 if (cfqd->active_queue == cfqq) {
3289 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3291 if (cfq_cfqq_slice_new(cfqq)) {
3292 cfq_set_prio_slice(cfqd, cfqq);
3293 cfq_clear_cfqq_slice_new(cfqq);
3297 * Should we wait for next request to come in before we expire
3300 if (cfq_should_wait_busy(cfqd, cfqq)) {
3301 unsigned long extend_sl = cfqd->cfq_slice_idle;
3302 if (!cfqd->cfq_slice_idle)
3303 extend_sl = cfqd->cfq_group_idle;
3304 cfqq->slice_end = jiffies + extend_sl;
3305 cfq_mark_cfqq_wait_busy(cfqq);
3306 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3310 * Idling is not enabled on:
3312 * - idle-priority queues
3314 * - queues with still some requests queued
3315 * - when there is a close cooperator
3317 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3318 cfq_slice_expired(cfqd, 1);
3319 else if (sync && cfqq_empty &&
3320 !cfq_close_cooperator(cfqd, cfqq)) {
3321 cfq_arm_slice_timer(cfqd);
3325 if (!cfqd->rq_in_driver)
3326 cfq_schedule_dispatch(cfqd);
3329 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3331 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3332 cfq_mark_cfqq_must_alloc_slice(cfqq);
3333 return ELV_MQUEUE_MUST;
3336 return ELV_MQUEUE_MAY;
3339 static int cfq_may_queue(struct request_queue *q, int rw)
3341 struct cfq_data *cfqd = q->elevator->elevator_data;
3342 struct task_struct *tsk = current;
3343 struct cfq_io_cq *cic;
3344 struct cfq_queue *cfqq;
3347 * don't force setup of a queue from here, as a call to may_queue
3348 * does not necessarily imply that a request actually will be queued.
3349 * so just lookup a possibly existing queue, or return 'may queue'
3352 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3354 return ELV_MQUEUE_MAY;
3356 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3358 cfq_init_prio_data(cfqq, cic->icq.ioc);
3360 return __cfq_may_queue(cfqq);
3363 return ELV_MQUEUE_MAY;
3367 * queue lock held here
3369 static void cfq_put_request(struct request *rq)
3371 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3374 const int rw = rq_data_dir(rq);
3376 BUG_ON(!cfqq->allocated[rw]);
3377 cfqq->allocated[rw]--;
3379 /* Put down rq reference on cfqg */
3380 cfq_put_cfqg(RQ_CFQG(rq));
3381 rq->elv.priv[0] = NULL;
3382 rq->elv.priv[1] = NULL;
3384 cfq_put_queue(cfqq);
3388 static struct cfq_queue *
3389 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3390 struct cfq_queue *cfqq)
3392 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3393 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3394 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3395 cfq_put_queue(cfqq);
3396 return cic_to_cfqq(cic, 1);
3400 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3401 * was the last process referring to said cfqq.
3403 static struct cfq_queue *
3404 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3406 if (cfqq_process_refs(cfqq) == 1) {
3407 cfqq->pid = current->pid;
3408 cfq_clear_cfqq_coop(cfqq);
3409 cfq_clear_cfqq_split_coop(cfqq);
3413 cic_set_cfqq(cic, NULL, 1);
3415 cfq_put_cooperator(cfqq);
3417 cfq_put_queue(cfqq);
3421 * Allocate cfq data structures associated with this request.
3424 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3426 struct cfq_data *cfqd = q->elevator->elevator_data;
3427 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3428 const int rw = rq_data_dir(rq);
3429 const bool is_sync = rq_is_sync(rq);
3430 struct cfq_queue *cfqq;
3431 unsigned int changed;
3433 might_sleep_if(gfp_mask & __GFP_WAIT);
3435 spin_lock_irq(q->queue_lock);
3437 /* handle changed notifications */
3438 changed = icq_get_changed(&cic->icq);
3439 if (unlikely(changed & ICQ_IOPRIO_CHANGED))
3440 changed_ioprio(cic);
3441 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3442 if (unlikely(changed & ICQ_CGROUP_CHANGED))
3443 changed_cgroup(cic);
3447 cfqq = cic_to_cfqq(cic, is_sync);
3448 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3449 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3450 cic_set_cfqq(cic, cfqq, is_sync);
3453 * If the queue was seeky for too long, break it apart.
3455 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3456 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3457 cfqq = split_cfqq(cic, cfqq);
3463 * Check to see if this queue is scheduled to merge with
3464 * another, closely cooperating queue. The merging of
3465 * queues happens here as it must be done in process context.
3466 * The reference on new_cfqq was taken in merge_cfqqs.
3469 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3472 cfqq->allocated[rw]++;
3475 rq->elv.priv[0] = cfqq;
3476 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
3477 spin_unlock_irq(q->queue_lock);
3481 static void cfq_kick_queue(struct work_struct *work)
3483 struct cfq_data *cfqd =
3484 container_of(work, struct cfq_data, unplug_work);
3485 struct request_queue *q = cfqd->queue;
3487 spin_lock_irq(q->queue_lock);
3488 __blk_run_queue(cfqd->queue);
3489 spin_unlock_irq(q->queue_lock);
3493 * Timer running if the active_queue is currently idling inside its time slice
3495 static void cfq_idle_slice_timer(unsigned long data)
3497 struct cfq_data *cfqd = (struct cfq_data *) data;
3498 struct cfq_queue *cfqq;
3499 unsigned long flags;
3502 cfq_log(cfqd, "idle timer fired");
3504 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3506 cfqq = cfqd->active_queue;
3511 * We saw a request before the queue expired, let it through
3513 if (cfq_cfqq_must_dispatch(cfqq))
3519 if (cfq_slice_used(cfqq))
3523 * only expire and reinvoke request handler, if there are
3524 * other queues with pending requests
3526 if (!cfqd->busy_queues)
3530 * not expired and it has a request pending, let it dispatch
3532 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3536 * Queue depth flag is reset only when the idle didn't succeed
3538 cfq_clear_cfqq_deep(cfqq);
3541 cfq_slice_expired(cfqd, timed_out);
3543 cfq_schedule_dispatch(cfqd);
3545 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3548 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3550 del_timer_sync(&cfqd->idle_slice_timer);
3551 cancel_work_sync(&cfqd->unplug_work);
3554 static void cfq_put_async_queues(struct cfq_data *cfqd)
3558 for (i = 0; i < IOPRIO_BE_NR; i++) {
3559 if (cfqd->async_cfqq[0][i])
3560 cfq_put_queue(cfqd->async_cfqq[0][i]);
3561 if (cfqd->async_cfqq[1][i])
3562 cfq_put_queue(cfqd->async_cfqq[1][i]);
3565 if (cfqd->async_idle_cfqq)
3566 cfq_put_queue(cfqd->async_idle_cfqq);
3569 static void cfq_exit_queue(struct elevator_queue *e)
3571 struct cfq_data *cfqd = e->elevator_data;
3572 struct request_queue *q = cfqd->queue;
3575 cfq_shutdown_timer_wq(cfqd);
3577 spin_lock_irq(q->queue_lock);
3579 if (cfqd->active_queue)
3580 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3582 cfq_put_async_queues(cfqd);
3583 cfq_release_cfq_groups(cfqd);
3586 * If there are groups which we could not unlink from blkcg list,
3587 * wait for a rcu period for them to be freed.
3589 if (cfqd->nr_blkcg_linked_grps)
3592 spin_unlock_irq(q->queue_lock);
3594 cfq_shutdown_timer_wq(cfqd);
3597 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3598 * Do this wait only if there are other unlinked groups out
3599 * there. This can happen if cgroup deletion path claimed the
3600 * responsibility of cleaning up a group before queue cleanup code
3603 * Do not call synchronize_rcu() unconditionally as there are drivers
3604 * which create/delete request queue hundreds of times during scan/boot
3605 * and synchronize_rcu() can take significant time and slow down boot.
3610 #ifndef CONFIG_CFQ_GROUP_IOSCHED
3611 kfree(cfqd->root_group);
3616 static int cfq_init_queue(struct request_queue *q)
3618 struct cfq_data *cfqd;
3619 struct blkio_group *blkg __maybe_unused;
3622 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3627 q->elevator->elevator_data = cfqd;
3629 /* Init root service tree */
3630 cfqd->grp_service_tree = CFQ_RB_ROOT;
3632 /* Init root group and prefer root group over other groups by default */
3633 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3635 spin_lock_irq(q->queue_lock);
3637 blkg = blkg_lookup_create(&blkio_root_cgroup, q, BLKIO_POLICY_PROP,
3640 cfqd->root_group = cfqg_of_blkg(blkg);
3642 spin_unlock_irq(q->queue_lock);
3645 cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
3646 GFP_KERNEL, cfqd->queue->node);
3647 if (cfqd->root_group)
3648 cfq_init_cfqg_base(cfqd->root_group);
3650 if (!cfqd->root_group) {
3655 cfqd->root_group->weight = 2*BLKIO_WEIGHT_DEFAULT;
3658 * Not strictly needed (since RB_ROOT just clears the node and we
3659 * zeroed cfqd on alloc), but better be safe in case someone decides
3660 * to add magic to the rb code
3662 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3663 cfqd->prio_trees[i] = RB_ROOT;
3666 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3667 * Grab a permanent reference to it, so that the normal code flow
3668 * will not attempt to free it. oom_cfqq is linked to root_group
3669 * but shouldn't hold a reference as it'll never be unlinked. Lose
3670 * the reference from linking right away.
3672 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3673 cfqd->oom_cfqq.ref++;
3674 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
3675 cfq_put_cfqg(cfqd->root_group);
3677 init_timer(&cfqd->idle_slice_timer);
3678 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3679 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3681 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3683 cfqd->cfq_quantum = cfq_quantum;
3684 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3685 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3686 cfqd->cfq_back_max = cfq_back_max;
3687 cfqd->cfq_back_penalty = cfq_back_penalty;
3688 cfqd->cfq_slice[0] = cfq_slice_async;
3689 cfqd->cfq_slice[1] = cfq_slice_sync;
3690 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3691 cfqd->cfq_slice_idle = cfq_slice_idle;
3692 cfqd->cfq_group_idle = cfq_group_idle;
3693 cfqd->cfq_latency = 1;
3696 * we optimistically start assuming sync ops weren't delayed in last
3697 * second, in order to have larger depth for async operations.
3699 cfqd->last_delayed_sync = jiffies - HZ;
3704 * sysfs parts below -->
3707 cfq_var_show(unsigned int var, char *page)
3709 return sprintf(page, "%d\n", var);
3713 cfq_var_store(unsigned int *var, const char *page, size_t count)
3715 char *p = (char *) page;
3717 *var = simple_strtoul(p, &p, 10);
3721 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3722 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3724 struct cfq_data *cfqd = e->elevator_data; \
3725 unsigned int __data = __VAR; \
3727 __data = jiffies_to_msecs(__data); \
3728 return cfq_var_show(__data, (page)); \
3730 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3731 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3732 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3733 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3734 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3735 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3736 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3737 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3738 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3739 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3740 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3741 #undef SHOW_FUNCTION
3743 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3744 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3746 struct cfq_data *cfqd = e->elevator_data; \
3747 unsigned int __data; \
3748 int ret = cfq_var_store(&__data, (page), count); \
3749 if (__data < (MIN)) \
3751 else if (__data > (MAX)) \
3754 *(__PTR) = msecs_to_jiffies(__data); \
3756 *(__PTR) = __data; \
3759 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3760 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3762 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3764 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3765 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3767 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3768 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
3769 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3770 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3771 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3773 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3774 #undef STORE_FUNCTION
3776 #define CFQ_ATTR(name) \
3777 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3779 static struct elv_fs_entry cfq_attrs[] = {
3781 CFQ_ATTR(fifo_expire_sync),
3782 CFQ_ATTR(fifo_expire_async),
3783 CFQ_ATTR(back_seek_max),
3784 CFQ_ATTR(back_seek_penalty),
3785 CFQ_ATTR(slice_sync),
3786 CFQ_ATTR(slice_async),
3787 CFQ_ATTR(slice_async_rq),
3788 CFQ_ATTR(slice_idle),
3789 CFQ_ATTR(group_idle),
3790 CFQ_ATTR(low_latency),
3794 static struct elevator_type iosched_cfq = {
3796 .elevator_merge_fn = cfq_merge,
3797 .elevator_merged_fn = cfq_merged_request,
3798 .elevator_merge_req_fn = cfq_merged_requests,
3799 .elevator_allow_merge_fn = cfq_allow_merge,
3800 .elevator_bio_merged_fn = cfq_bio_merged,
3801 .elevator_dispatch_fn = cfq_dispatch_requests,
3802 .elevator_add_req_fn = cfq_insert_request,
3803 .elevator_activate_req_fn = cfq_activate_request,
3804 .elevator_deactivate_req_fn = cfq_deactivate_request,
3805 .elevator_completed_req_fn = cfq_completed_request,
3806 .elevator_former_req_fn = elv_rb_former_request,
3807 .elevator_latter_req_fn = elv_rb_latter_request,
3808 .elevator_init_icq_fn = cfq_init_icq,
3809 .elevator_exit_icq_fn = cfq_exit_icq,
3810 .elevator_set_req_fn = cfq_set_request,
3811 .elevator_put_req_fn = cfq_put_request,
3812 .elevator_may_queue_fn = cfq_may_queue,
3813 .elevator_init_fn = cfq_init_queue,
3814 .elevator_exit_fn = cfq_exit_queue,
3816 .icq_size = sizeof(struct cfq_io_cq),
3817 .icq_align = __alignof__(struct cfq_io_cq),
3818 .elevator_attrs = cfq_attrs,
3819 .elevator_name = "cfq",
3820 .elevator_owner = THIS_MODULE,
3823 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3824 static struct blkio_policy_type blkio_policy_cfq = {
3826 .blkio_alloc_group_fn = cfq_alloc_blkio_group,
3827 .blkio_link_group_fn = cfq_link_blkio_group,
3828 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3829 .blkio_clear_queue_fn = cfq_clear_queue,
3830 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3832 .plid = BLKIO_POLICY_PROP,
3836 static int __init cfq_init(void)
3841 * could be 0 on HZ < 1000 setups
3843 if (!cfq_slice_async)
3844 cfq_slice_async = 1;
3845 if (!cfq_slice_idle)
3848 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3849 if (!cfq_group_idle)
3854 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3858 ret = elv_register(&iosched_cfq);
3860 kmem_cache_destroy(cfq_pool);
3864 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3865 blkio_policy_register(&blkio_policy_cfq);
3870 static void __exit cfq_exit(void)
3872 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3873 blkio_policy_unregister(&blkio_policy_cfq);
3875 elv_unregister(&iosched_cfq);
3876 kmem_cache_destroy(cfq_pool);
3879 module_init(cfq_init);
3880 module_exit(cfq_exit);
3882 MODULE_AUTHOR("Jens Axboe");
3883 MODULE_LICENSE("GPL");
3884 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");