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>
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
38 * offset from end of service tree
40 #define CFQ_IDLE_DELAY (HZ / 5)
43 * below this threshold, we consider thinktime immediate
45 #define CFQ_MIN_TT (2)
47 #define CFQ_SLICE_SCALE (5)
48 #define CFQ_HW_QUEUE_MIN (5)
49 #define CFQ_SERVICE_SHIFT 12
51 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
71 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
75 #define sample_valid(samples) ((samples) > 80)
76 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
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;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
95 * Per process-grouping structure
100 /* various state flags, see below */
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
118 /* currently allocated requests */
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
125 unsigned int allocated_slice;
126 unsigned int slice_dispatch;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
129 unsigned long slice_end;
132 /* pending metadata requests */
134 /* number of requests that are on the dispatch list or inside driver */
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
144 sector_t last_request_pos;
146 struct cfq_rb_root *service_tree;
147 struct cfq_queue *new_cfqq;
148 struct cfq_group *cfqg;
149 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors;
154 * First index in the service_trees.
155 * IDLE is handled separately, so it has negative index
165 * Second index in the service_trees.
169 SYNC_NOIDLE_WORKLOAD = 1,
173 /* This is per cgroup per device grouping structure */
175 /* group service_tree member */
176 struct rb_node rb_node;
178 /* group service_tree key */
181 unsigned int new_weight;
184 /* number of cfqq currently on this group */
188 * Per group busy queus average. Useful for workload slice calc. We
189 * create the array for each prio class but at run time it is used
190 * only for RT and BE class and slot for IDLE class remains unused.
191 * This is primarily done to avoid confusion and a gcc warning.
193 unsigned int busy_queues_avg[CFQ_PRIO_NR];
195 * rr lists of queues with requests. We maintain service trees for
196 * RT and BE classes. These trees are subdivided in subclasses
197 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198 * class there is no subclassification and all the cfq queues go on
199 * a single tree service_tree_idle.
200 * Counts are embedded in the cfq_rb_root
202 struct cfq_rb_root service_trees[2][3];
203 struct cfq_rb_root service_tree_idle;
205 unsigned long saved_workload_slice;
206 enum wl_type_t saved_workload;
207 enum wl_prio_t saved_serving_prio;
208 struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210 struct hlist_node cfqd_node;
213 /* number of requests that are on the dispatch list or inside driver */
218 * Per block device queue structure
221 struct request_queue *queue;
222 /* Root service tree for cfq_groups */
223 struct cfq_rb_root grp_service_tree;
224 struct cfq_group root_group;
227 * The priority currently being served
229 enum wl_prio_t serving_prio;
230 enum wl_type_t serving_type;
231 unsigned long workload_expires;
232 struct cfq_group *serving_group;
235 * Each priority tree is sorted by next_request position. These
236 * trees are used when determining if two or more queues are
237 * interleaving requests (see cfq_close_cooperator).
239 struct rb_root prio_trees[CFQ_PRIO_LISTS];
241 unsigned int busy_queues;
242 unsigned int busy_sync_queues;
248 * queue-depth detection
254 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
258 int hw_tag_est_depth;
259 unsigned int hw_tag_samples;
262 * idle window management
264 struct timer_list idle_slice_timer;
265 struct work_struct unplug_work;
267 struct cfq_queue *active_queue;
268 struct cfq_io_context *active_cic;
271 * async queue for each priority case
273 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274 struct cfq_queue *async_idle_cfqq;
276 sector_t last_position;
279 * tunables, see top of file
281 unsigned int cfq_quantum;
282 unsigned int cfq_fifo_expire[2];
283 unsigned int cfq_back_penalty;
284 unsigned int cfq_back_max;
285 unsigned int cfq_slice[2];
286 unsigned int cfq_slice_async_rq;
287 unsigned int cfq_slice_idle;
288 unsigned int cfq_group_idle;
289 unsigned int cfq_latency;
291 unsigned int cic_index;
292 struct list_head cic_list;
295 * Fallback dummy cfqq for extreme OOM conditions
297 struct cfq_queue oom_cfqq;
299 unsigned long last_delayed_sync;
301 /* List of cfq groups being managed on this device*/
302 struct hlist_head cfqg_list;
304 /* Number of groups which are on blkcg->blkg_list */
305 unsigned int nr_blkcg_linked_grps;
308 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
310 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
317 if (prio == IDLE_WORKLOAD)
318 return &cfqg->service_tree_idle;
320 return &cfqg->service_trees[prio][type];
323 enum cfqq_state_flags {
324 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
325 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
326 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
327 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
328 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
329 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
330 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
331 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
332 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
333 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
334 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
335 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
336 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
339 #define CFQ_CFQQ_FNS(name) \
340 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
342 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
344 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
346 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
348 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
350 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
354 CFQ_CFQQ_FNS(wait_request);
355 CFQ_CFQQ_FNS(must_dispatch);
356 CFQ_CFQQ_FNS(must_alloc_slice);
357 CFQ_CFQQ_FNS(fifo_expire);
358 CFQ_CFQQ_FNS(idle_window);
359 CFQ_CFQQ_FNS(prio_changed);
360 CFQ_CFQQ_FNS(slice_new);
363 CFQ_CFQQ_FNS(split_coop);
365 CFQ_CFQQ_FNS(wait_busy);
368 #ifdef CONFIG_CFQ_GROUP_IOSCHED
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
371 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
372 blkg_path(&(cfqq)->cfqg->blkg), ##args);
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
376 blkg_path(&(cfqg)->blkg), ##args); \
379 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
383 #define cfq_log(cfqd, fmt, args...) \
384 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
386 /* Traverses through cfq group service trees */
387 #define for_each_cfqg_st(cfqg, i, j, st) \
388 for (i = 0; i <= IDLE_WORKLOAD; i++) \
389 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
390 : &cfqg->service_tree_idle; \
391 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
392 (i == IDLE_WORKLOAD && j == 0); \
393 j++, st = i < IDLE_WORKLOAD ? \
394 &cfqg->service_trees[i][j]: NULL) \
397 static inline bool iops_mode(struct cfq_data *cfqd)
400 * If we are not idling on queues and it is a NCQ drive, parallel
401 * execution of requests is on and measuring time is not possible
402 * in most of the cases until and unless we drive shallower queue
403 * depths and that becomes a performance bottleneck. In such cases
404 * switch to start providing fairness in terms of number of IOs.
406 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
412 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
414 if (cfq_class_idle(cfqq))
415 return IDLE_WORKLOAD;
416 if (cfq_class_rt(cfqq))
422 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
424 if (!cfq_cfqq_sync(cfqq))
425 return ASYNC_WORKLOAD;
426 if (!cfq_cfqq_idle_window(cfqq))
427 return SYNC_NOIDLE_WORKLOAD;
428 return SYNC_WORKLOAD;
431 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
432 struct cfq_data *cfqd,
433 struct cfq_group *cfqg)
435 if (wl == IDLE_WORKLOAD)
436 return cfqg->service_tree_idle.count;
438 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
439 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
440 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
443 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
444 struct cfq_group *cfqg)
446 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
447 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
450 static void cfq_dispatch_insert(struct request_queue *, struct request *);
451 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
452 struct io_context *, gfp_t);
453 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
454 struct io_context *);
456 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
459 return cic->cfqq[is_sync];
462 static inline void cic_set_cfqq(struct cfq_io_context *cic,
463 struct cfq_queue *cfqq, bool is_sync)
465 cic->cfqq[is_sync] = cfqq;
468 #define CIC_DEAD_KEY 1ul
469 #define CIC_DEAD_INDEX_SHIFT 1
471 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
473 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
476 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
478 struct cfq_data *cfqd = cic->key;
480 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
487 * We regard a request as SYNC, if it's either a read or has the SYNC bit
488 * set (in which case it could also be direct WRITE).
490 static inline bool cfq_bio_sync(struct bio *bio)
492 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
496 * scheduler run of queue, if there are requests pending and no one in the
497 * driver that will restart queueing
499 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
501 if (cfqd->busy_queues) {
502 cfq_log(cfqd, "schedule dispatch");
503 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
508 * Scale schedule slice based on io priority. Use the sync time slice only
509 * if a queue is marked sync and has sync io queued. A sync queue with async
510 * io only, should not get full sync slice length.
512 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
515 const int base_slice = cfqd->cfq_slice[sync];
517 WARN_ON(prio >= IOPRIO_BE_NR);
519 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
523 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
525 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
528 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
530 u64 d = delta << CFQ_SERVICE_SHIFT;
532 d = d * BLKIO_WEIGHT_DEFAULT;
533 do_div(d, cfqg->weight);
537 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
539 s64 delta = (s64)(vdisktime - min_vdisktime);
541 min_vdisktime = vdisktime;
543 return min_vdisktime;
546 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
548 s64 delta = (s64)(vdisktime - min_vdisktime);
550 min_vdisktime = vdisktime;
552 return min_vdisktime;
555 static void update_min_vdisktime(struct cfq_rb_root *st)
557 struct cfq_group *cfqg;
560 cfqg = rb_entry_cfqg(st->left);
561 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
567 * get averaged number of queues of RT/BE priority.
568 * average is updated, with a formula that gives more weight to higher numbers,
569 * to quickly follows sudden increases and decrease slowly
572 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
573 struct cfq_group *cfqg, bool rt)
575 unsigned min_q, max_q;
576 unsigned mult = cfq_hist_divisor - 1;
577 unsigned round = cfq_hist_divisor / 2;
578 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
580 min_q = min(cfqg->busy_queues_avg[rt], busy);
581 max_q = max(cfqg->busy_queues_avg[rt], busy);
582 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
584 return cfqg->busy_queues_avg[rt];
587 static inline unsigned
588 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
590 struct cfq_rb_root *st = &cfqd->grp_service_tree;
592 return cfq_target_latency * cfqg->weight / st->total_weight;
595 static inline unsigned
596 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
598 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
599 if (cfqd->cfq_latency) {
601 * interested queues (we consider only the ones with the same
602 * priority class in the cfq group)
604 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
606 unsigned sync_slice = cfqd->cfq_slice[1];
607 unsigned expect_latency = sync_slice * iq;
608 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
610 if (expect_latency > group_slice) {
611 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
612 /* scale low_slice according to IO priority
613 * and sync vs async */
615 min(slice, base_low_slice * slice / sync_slice);
616 /* the adapted slice value is scaled to fit all iqs
617 * into the target latency */
618 slice = max(slice * group_slice / expect_latency,
626 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
628 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
630 cfqq->slice_start = jiffies;
631 cfqq->slice_end = jiffies + slice;
632 cfqq->allocated_slice = slice;
633 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
637 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
638 * isn't valid until the first request from the dispatch is activated
639 * and the slice time set.
641 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
643 if (cfq_cfqq_slice_new(cfqq))
645 if (time_before(jiffies, cfqq->slice_end))
652 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
653 * We choose the request that is closest to the head right now. Distance
654 * behind the head is penalized and only allowed to a certain extent.
656 static struct request *
657 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
659 sector_t s1, s2, d1 = 0, d2 = 0;
660 unsigned long back_max;
661 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
662 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
663 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
665 if (rq1 == NULL || rq1 == rq2)
670 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
672 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
674 if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
676 else if ((rq2->cmd_flags & REQ_META) &&
677 !(rq1->cmd_flags & REQ_META))
680 s1 = blk_rq_pos(rq1);
681 s2 = blk_rq_pos(rq2);
684 * by definition, 1KiB is 2 sectors
686 back_max = cfqd->cfq_back_max * 2;
689 * Strict one way elevator _except_ in the case where we allow
690 * short backward seeks which are biased as twice the cost of a
691 * similar forward seek.
695 else if (s1 + back_max >= last)
696 d1 = (last - s1) * cfqd->cfq_back_penalty;
698 wrap |= CFQ_RQ1_WRAP;
702 else if (s2 + back_max >= last)
703 d2 = (last - s2) * cfqd->cfq_back_penalty;
705 wrap |= CFQ_RQ2_WRAP;
707 /* Found required data */
710 * By doing switch() on the bit mask "wrap" we avoid having to
711 * check two variables for all permutations: --> faster!
714 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
730 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
733 * Since both rqs are wrapped,
734 * start with the one that's further behind head
735 * (--> only *one* back seek required),
736 * since back seek takes more time than forward.
746 * The below is leftmost cache rbtree addon
748 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
750 /* Service tree is empty */
755 root->left = rb_first(&root->rb);
758 return rb_entry(root->left, struct cfq_queue, rb_node);
763 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
766 root->left = rb_first(&root->rb);
769 return rb_entry_cfqg(root->left);
774 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
780 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
784 rb_erase_init(n, &root->rb);
789 * would be nice to take fifo expire time into account as well
791 static struct request *
792 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
793 struct request *last)
795 struct rb_node *rbnext = rb_next(&last->rb_node);
796 struct rb_node *rbprev = rb_prev(&last->rb_node);
797 struct request *next = NULL, *prev = NULL;
799 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
802 prev = rb_entry_rq(rbprev);
805 next = rb_entry_rq(rbnext);
807 rbnext = rb_first(&cfqq->sort_list);
808 if (rbnext && rbnext != &last->rb_node)
809 next = rb_entry_rq(rbnext);
812 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
815 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
816 struct cfq_queue *cfqq)
819 * just an approximation, should be ok.
821 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
822 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
826 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
828 return cfqg->vdisktime - st->min_vdisktime;
832 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
834 struct rb_node **node = &st->rb.rb_node;
835 struct rb_node *parent = NULL;
836 struct cfq_group *__cfqg;
837 s64 key = cfqg_key(st, cfqg);
840 while (*node != NULL) {
842 __cfqg = rb_entry_cfqg(parent);
844 if (key < cfqg_key(st, __cfqg))
845 node = &parent->rb_left;
847 node = &parent->rb_right;
853 st->left = &cfqg->rb_node;
855 rb_link_node(&cfqg->rb_node, parent, node);
856 rb_insert_color(&cfqg->rb_node, &st->rb);
860 cfq_update_group_weight(struct cfq_group *cfqg)
862 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
863 if (cfqg->needs_update) {
864 cfqg->weight = cfqg->new_weight;
865 cfqg->needs_update = false;
870 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
872 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
874 cfq_update_group_weight(cfqg);
875 __cfq_group_service_tree_add(st, cfqg);
876 st->total_weight += cfqg->weight;
880 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
882 struct cfq_rb_root *st = &cfqd->grp_service_tree;
883 struct cfq_group *__cfqg;
887 if (!RB_EMPTY_NODE(&cfqg->rb_node))
891 * Currently put the group at the end. Later implement something
892 * so that groups get lesser vtime based on their weights, so that
893 * if group does not loose all if it was not continuously backlogged.
895 n = rb_last(&st->rb);
897 __cfqg = rb_entry_cfqg(n);
898 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
900 cfqg->vdisktime = st->min_vdisktime;
901 cfq_group_service_tree_add(st, cfqg);
905 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
907 st->total_weight -= cfqg->weight;
908 if (!RB_EMPTY_NODE(&cfqg->rb_node))
909 cfq_rb_erase(&cfqg->rb_node, st);
913 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
915 struct cfq_rb_root *st = &cfqd->grp_service_tree;
917 BUG_ON(cfqg->nr_cfqq < 1);
920 /* If there are other cfq queues under this group, don't delete it */
924 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
925 cfq_group_service_tree_del(st, cfqg);
926 cfqg->saved_workload_slice = 0;
927 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
930 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
931 unsigned int *unaccounted_time)
933 unsigned int slice_used;
936 * Queue got expired before even a single request completed or
937 * got expired immediately after first request completion.
939 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
941 * Also charge the seek time incurred to the group, otherwise
942 * if there are mutiple queues in the group, each can dispatch
943 * a single request on seeky media and cause lots of seek time
944 * and group will never know it.
946 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
949 slice_used = jiffies - cfqq->slice_start;
950 if (slice_used > cfqq->allocated_slice) {
951 *unaccounted_time = slice_used - cfqq->allocated_slice;
952 slice_used = cfqq->allocated_slice;
954 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
955 *unaccounted_time += cfqq->slice_start -
956 cfqq->dispatch_start;
962 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
963 struct cfq_queue *cfqq)
965 struct cfq_rb_root *st = &cfqd->grp_service_tree;
966 unsigned int used_sl, charge, unaccounted_sl = 0;
967 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
968 - cfqg->service_tree_idle.count;
971 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
974 charge = cfqq->slice_dispatch;
975 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
976 charge = cfqq->allocated_slice;
978 /* Can't update vdisktime while group is on service tree */
979 cfq_group_service_tree_del(st, cfqg);
980 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
981 /* If a new weight was requested, update now, off tree */
982 cfq_group_service_tree_add(st, cfqg);
984 /* This group is being expired. Save the context */
985 if (time_after(cfqd->workload_expires, jiffies)) {
986 cfqg->saved_workload_slice = cfqd->workload_expires
988 cfqg->saved_workload = cfqd->serving_type;
989 cfqg->saved_serving_prio = cfqd->serving_prio;
991 cfqg->saved_workload_slice = 0;
993 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
995 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
996 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
997 iops_mode(cfqd), cfqq->nr_sectors);
998 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1000 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1003 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1004 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1007 return container_of(blkg, struct cfq_group, blkg);
1011 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1012 unsigned int weight)
1014 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1015 cfqg->new_weight = weight;
1016 cfqg->needs_update = true;
1019 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1020 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1022 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1023 unsigned int major, minor;
1026 * Add group onto cgroup list. It might happen that bdi->dev is
1027 * not initialized yet. Initialize this new group without major
1028 * and minor info and this info will be filled in once a new thread
1032 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1033 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1034 (void *)cfqd, MKDEV(major, minor));
1036 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1039 cfqd->nr_blkcg_linked_grps++;
1040 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1042 /* Add group on cfqd list */
1043 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1047 * Should be called from sleepable context. No request queue lock as per
1048 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1049 * from sleepable context.
1051 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1053 struct cfq_group *cfqg = NULL;
1055 struct cfq_rb_root *st;
1057 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1061 for_each_cfqg_st(cfqg, i, j, st)
1063 RB_CLEAR_NODE(&cfqg->rb_node);
1066 * Take the initial reference that will be released on destroy
1067 * This can be thought of a joint reference by cgroup and
1068 * elevator which will be dropped by either elevator exit
1069 * or cgroup deletion path depending on who is exiting first.
1075 static struct cfq_group *
1076 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1078 struct cfq_group *cfqg = NULL;
1080 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1081 unsigned int major, minor;
1084 * This is the common case when there are no blkio cgroups.
1085 * Avoid lookup in this case
1087 if (blkcg == &blkio_root_cgroup)
1088 cfqg = &cfqd->root_group;
1090 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1092 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1093 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1094 cfqg->blkg.dev = MKDEV(major, minor);
1101 * Search for the cfq group current task belongs to. request_queue lock must
1104 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1106 struct blkio_cgroup *blkcg;
1107 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1108 struct request_queue *q = cfqd->queue;
1111 blkcg = task_blkio_cgroup(current);
1112 cfqg = cfq_find_cfqg(cfqd, blkcg);
1119 * Need to allocate a group. Allocation of group also needs allocation
1120 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1121 * we need to drop rcu lock and queue_lock before we call alloc.
1123 * Not taking any queue reference here and assuming that queue is
1124 * around by the time we return. CFQ queue allocation code does
1125 * the same. It might be racy though.
1129 spin_unlock_irq(q->queue_lock);
1131 cfqg = cfq_alloc_cfqg(cfqd);
1133 spin_lock_irq(q->queue_lock);
1136 blkcg = task_blkio_cgroup(current);
1139 * If some other thread already allocated the group while we were
1140 * not holding queue lock, free up the group
1142 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1151 cfqg = &cfqd->root_group;
1153 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1158 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1164 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1166 /* Currently, all async queues are mapped to root group */
1167 if (!cfq_cfqq_sync(cfqq))
1168 cfqg = &cfqq->cfqd->root_group;
1171 /* cfqq reference on cfqg */
1175 static void cfq_put_cfqg(struct cfq_group *cfqg)
1177 struct cfq_rb_root *st;
1180 BUG_ON(cfqg->ref <= 0);
1184 for_each_cfqg_st(cfqg, i, j, st)
1185 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1189 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1191 /* Something wrong if we are trying to remove same group twice */
1192 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1194 hlist_del_init(&cfqg->cfqd_node);
1197 * Put the reference taken at the time of creation so that when all
1198 * queues are gone, group can be destroyed.
1203 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1205 struct hlist_node *pos, *n;
1206 struct cfq_group *cfqg;
1208 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1210 * If cgroup removal path got to blk_group first and removed
1211 * it from cgroup list, then it will take care of destroying
1214 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1215 cfq_destroy_cfqg(cfqd, cfqg);
1220 * Blk cgroup controller notification saying that blkio_group object is being
1221 * delinked as associated cgroup object is going away. That also means that
1222 * no new IO will come in this group. So get rid of this group as soon as
1223 * any pending IO in the group is finished.
1225 * This function is called under rcu_read_lock(). key is the rcu protected
1226 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1229 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1230 * it should not be NULL as even if elevator was exiting, cgroup deltion
1231 * path got to it first.
1233 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1235 unsigned long flags;
1236 struct cfq_data *cfqd = key;
1238 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1239 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1240 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1243 #else /* GROUP_IOSCHED */
1244 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1246 return &cfqd->root_group;
1249 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1255 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1259 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1260 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1262 #endif /* GROUP_IOSCHED */
1265 * The cfqd->service_trees holds all pending cfq_queue's that have
1266 * requests waiting to be processed. It is sorted in the order that
1267 * we will service the queues.
1269 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1272 struct rb_node **p, *parent;
1273 struct cfq_queue *__cfqq;
1274 unsigned long rb_key;
1275 struct cfq_rb_root *service_tree;
1278 int group_changed = 0;
1280 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1282 if (cfq_class_idle(cfqq)) {
1283 rb_key = CFQ_IDLE_DELAY;
1284 parent = rb_last(&service_tree->rb);
1285 if (parent && parent != &cfqq->rb_node) {
1286 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1287 rb_key += __cfqq->rb_key;
1290 } else if (!add_front) {
1292 * Get our rb key offset. Subtract any residual slice
1293 * value carried from last service. A negative resid
1294 * count indicates slice overrun, and this should position
1295 * the next service time further away in the tree.
1297 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1298 rb_key -= cfqq->slice_resid;
1299 cfqq->slice_resid = 0;
1302 __cfqq = cfq_rb_first(service_tree);
1303 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1306 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1309 * same position, nothing more to do
1311 if (rb_key == cfqq->rb_key &&
1312 cfqq->service_tree == service_tree)
1315 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1316 cfqq->service_tree = NULL;
1321 cfqq->service_tree = service_tree;
1322 p = &service_tree->rb.rb_node;
1327 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1330 * sort by key, that represents service time.
1332 if (time_before(rb_key, __cfqq->rb_key))
1335 n = &(*p)->rb_right;
1343 service_tree->left = &cfqq->rb_node;
1345 cfqq->rb_key = rb_key;
1346 rb_link_node(&cfqq->rb_node, parent, p);
1347 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1348 service_tree->count++;
1349 if ((add_front || !new_cfqq) && !group_changed)
1351 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1354 static struct cfq_queue *
1355 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1356 sector_t sector, struct rb_node **ret_parent,
1357 struct rb_node ***rb_link)
1359 struct rb_node **p, *parent;
1360 struct cfq_queue *cfqq = NULL;
1368 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1371 * Sort strictly based on sector. Smallest to the left,
1372 * largest to the right.
1374 if (sector > blk_rq_pos(cfqq->next_rq))
1375 n = &(*p)->rb_right;
1376 else if (sector < blk_rq_pos(cfqq->next_rq))
1384 *ret_parent = parent;
1390 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1392 struct rb_node **p, *parent;
1393 struct cfq_queue *__cfqq;
1396 rb_erase(&cfqq->p_node, cfqq->p_root);
1397 cfqq->p_root = NULL;
1400 if (cfq_class_idle(cfqq))
1405 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1406 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1407 blk_rq_pos(cfqq->next_rq), &parent, &p);
1409 rb_link_node(&cfqq->p_node, parent, p);
1410 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1412 cfqq->p_root = NULL;
1416 * Update cfqq's position in the service tree.
1418 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1421 * Resorting requires the cfqq to be on the RR list already.
1423 if (cfq_cfqq_on_rr(cfqq)) {
1424 cfq_service_tree_add(cfqd, cfqq, 0);
1425 cfq_prio_tree_add(cfqd, cfqq);
1430 * add to busy list of queues for service, trying to be fair in ordering
1431 * the pending list according to last request service
1433 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1435 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1436 BUG_ON(cfq_cfqq_on_rr(cfqq));
1437 cfq_mark_cfqq_on_rr(cfqq);
1438 cfqd->busy_queues++;
1439 if (cfq_cfqq_sync(cfqq))
1440 cfqd->busy_sync_queues++;
1442 cfq_resort_rr_list(cfqd, cfqq);
1446 * Called when the cfqq no longer has requests pending, remove it from
1449 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1451 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1452 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1453 cfq_clear_cfqq_on_rr(cfqq);
1455 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1456 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1457 cfqq->service_tree = NULL;
1460 rb_erase(&cfqq->p_node, cfqq->p_root);
1461 cfqq->p_root = NULL;
1464 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1465 BUG_ON(!cfqd->busy_queues);
1466 cfqd->busy_queues--;
1467 if (cfq_cfqq_sync(cfqq))
1468 cfqd->busy_sync_queues--;
1472 * rb tree support functions
1474 static void cfq_del_rq_rb(struct request *rq)
1476 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1477 const int sync = rq_is_sync(rq);
1479 BUG_ON(!cfqq->queued[sync]);
1480 cfqq->queued[sync]--;
1482 elv_rb_del(&cfqq->sort_list, rq);
1484 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1486 * Queue will be deleted from service tree when we actually
1487 * expire it later. Right now just remove it from prio tree
1491 rb_erase(&cfqq->p_node, cfqq->p_root);
1492 cfqq->p_root = NULL;
1497 static void cfq_add_rq_rb(struct request *rq)
1499 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1500 struct cfq_data *cfqd = cfqq->cfqd;
1501 struct request *__alias, *prev;
1503 cfqq->queued[rq_is_sync(rq)]++;
1506 * looks a little odd, but the first insert might return an alias.
1507 * if that happens, put the alias on the dispatch list
1509 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1510 cfq_dispatch_insert(cfqd->queue, __alias);
1512 if (!cfq_cfqq_on_rr(cfqq))
1513 cfq_add_cfqq_rr(cfqd, cfqq);
1516 * check if this request is a better next-serve candidate
1518 prev = cfqq->next_rq;
1519 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1522 * adjust priority tree position, if ->next_rq changes
1524 if (prev != cfqq->next_rq)
1525 cfq_prio_tree_add(cfqd, cfqq);
1527 BUG_ON(!cfqq->next_rq);
1530 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1532 elv_rb_del(&cfqq->sort_list, rq);
1533 cfqq->queued[rq_is_sync(rq)]--;
1534 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1535 rq_data_dir(rq), rq_is_sync(rq));
1537 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1538 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1542 static struct request *
1543 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1545 struct task_struct *tsk = current;
1546 struct cfq_io_context *cic;
1547 struct cfq_queue *cfqq;
1549 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1553 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1555 sector_t sector = bio->bi_sector + bio_sectors(bio);
1557 return elv_rb_find(&cfqq->sort_list, sector);
1563 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1565 struct cfq_data *cfqd = q->elevator->elevator_data;
1567 cfqd->rq_in_driver++;
1568 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1569 cfqd->rq_in_driver);
1571 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1574 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1576 struct cfq_data *cfqd = q->elevator->elevator_data;
1578 WARN_ON(!cfqd->rq_in_driver);
1579 cfqd->rq_in_driver--;
1580 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1581 cfqd->rq_in_driver);
1584 static void cfq_remove_request(struct request *rq)
1586 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1588 if (cfqq->next_rq == rq)
1589 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1591 list_del_init(&rq->queuelist);
1594 cfqq->cfqd->rq_queued--;
1595 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1596 rq_data_dir(rq), rq_is_sync(rq));
1597 if (rq->cmd_flags & REQ_META) {
1598 WARN_ON(!cfqq->meta_pending);
1599 cfqq->meta_pending--;
1603 static int cfq_merge(struct request_queue *q, struct request **req,
1606 struct cfq_data *cfqd = q->elevator->elevator_data;
1607 struct request *__rq;
1609 __rq = cfq_find_rq_fmerge(cfqd, bio);
1610 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1612 return ELEVATOR_FRONT_MERGE;
1615 return ELEVATOR_NO_MERGE;
1618 static void cfq_merged_request(struct request_queue *q, struct request *req,
1621 if (type == ELEVATOR_FRONT_MERGE) {
1622 struct cfq_queue *cfqq = RQ_CFQQ(req);
1624 cfq_reposition_rq_rb(cfqq, req);
1628 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1631 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1632 bio_data_dir(bio), cfq_bio_sync(bio));
1636 cfq_merged_requests(struct request_queue *q, struct request *rq,
1637 struct request *next)
1639 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1641 * reposition in fifo if next is older than rq
1643 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1644 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1645 list_move(&rq->queuelist, &next->queuelist);
1646 rq_set_fifo_time(rq, rq_fifo_time(next));
1649 if (cfqq->next_rq == next)
1651 cfq_remove_request(next);
1652 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1653 rq_data_dir(next), rq_is_sync(next));
1656 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1659 struct cfq_data *cfqd = q->elevator->elevator_data;
1660 struct cfq_io_context *cic;
1661 struct cfq_queue *cfqq;
1664 * Disallow merge of a sync bio into an async request.
1666 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1670 * Lookup the cfqq that this bio will be queued with. Allow
1671 * merge only if rq is queued there.
1673 cic = cfq_cic_lookup(cfqd, current->io_context);
1677 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1678 return cfqq == RQ_CFQQ(rq);
1681 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1683 del_timer(&cfqd->idle_slice_timer);
1684 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1687 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1688 struct cfq_queue *cfqq)
1691 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1692 cfqd->serving_prio, cfqd->serving_type);
1693 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1694 cfqq->slice_start = 0;
1695 cfqq->dispatch_start = jiffies;
1696 cfqq->allocated_slice = 0;
1697 cfqq->slice_end = 0;
1698 cfqq->slice_dispatch = 0;
1699 cfqq->nr_sectors = 0;
1701 cfq_clear_cfqq_wait_request(cfqq);
1702 cfq_clear_cfqq_must_dispatch(cfqq);
1703 cfq_clear_cfqq_must_alloc_slice(cfqq);
1704 cfq_clear_cfqq_fifo_expire(cfqq);
1705 cfq_mark_cfqq_slice_new(cfqq);
1707 cfq_del_timer(cfqd, cfqq);
1710 cfqd->active_queue = cfqq;
1714 * current cfqq expired its slice (or was too idle), select new one
1717 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1720 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1722 if (cfq_cfqq_wait_request(cfqq))
1723 cfq_del_timer(cfqd, cfqq);
1725 cfq_clear_cfqq_wait_request(cfqq);
1726 cfq_clear_cfqq_wait_busy(cfqq);
1729 * If this cfqq is shared between multiple processes, check to
1730 * make sure that those processes are still issuing I/Os within
1731 * the mean seek distance. If not, it may be time to break the
1732 * queues apart again.
1734 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1735 cfq_mark_cfqq_split_coop(cfqq);
1738 * store what was left of this slice, if the queue idled/timed out
1741 if (cfq_cfqq_slice_new(cfqq))
1742 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1744 cfqq->slice_resid = cfqq->slice_end - jiffies;
1745 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1748 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1750 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1751 cfq_del_cfqq_rr(cfqd, cfqq);
1753 cfq_resort_rr_list(cfqd, cfqq);
1755 if (cfqq == cfqd->active_queue)
1756 cfqd->active_queue = NULL;
1758 if (cfqd->active_cic) {
1759 put_io_context(cfqd->active_cic->ioc);
1760 cfqd->active_cic = NULL;
1764 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1766 struct cfq_queue *cfqq = cfqd->active_queue;
1769 __cfq_slice_expired(cfqd, cfqq, timed_out);
1773 * Get next queue for service. Unless we have a queue preemption,
1774 * we'll simply select the first cfqq in the service tree.
1776 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1778 struct cfq_rb_root *service_tree =
1779 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1780 cfqd->serving_type);
1782 if (!cfqd->rq_queued)
1785 /* There is nothing to dispatch */
1788 if (RB_EMPTY_ROOT(&service_tree->rb))
1790 return cfq_rb_first(service_tree);
1793 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1795 struct cfq_group *cfqg;
1796 struct cfq_queue *cfqq;
1798 struct cfq_rb_root *st;
1800 if (!cfqd->rq_queued)
1803 cfqg = cfq_get_next_cfqg(cfqd);
1807 for_each_cfqg_st(cfqg, i, j, st)
1808 if ((cfqq = cfq_rb_first(st)) != NULL)
1814 * Get and set a new active queue for service.
1816 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1817 struct cfq_queue *cfqq)
1820 cfqq = cfq_get_next_queue(cfqd);
1822 __cfq_set_active_queue(cfqd, cfqq);
1826 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1829 if (blk_rq_pos(rq) >= cfqd->last_position)
1830 return blk_rq_pos(rq) - cfqd->last_position;
1832 return cfqd->last_position - blk_rq_pos(rq);
1835 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1838 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1841 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1842 struct cfq_queue *cur_cfqq)
1844 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1845 struct rb_node *parent, *node;
1846 struct cfq_queue *__cfqq;
1847 sector_t sector = cfqd->last_position;
1849 if (RB_EMPTY_ROOT(root))
1853 * First, if we find a request starting at the end of the last
1854 * request, choose it.
1856 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1861 * If the exact sector wasn't found, the parent of the NULL leaf
1862 * will contain the closest sector.
1864 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1865 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1868 if (blk_rq_pos(__cfqq->next_rq) < sector)
1869 node = rb_next(&__cfqq->p_node);
1871 node = rb_prev(&__cfqq->p_node);
1875 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1876 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1884 * cur_cfqq - passed in so that we don't decide that the current queue is
1885 * closely cooperating with itself.
1887 * So, basically we're assuming that that cur_cfqq has dispatched at least
1888 * one request, and that cfqd->last_position reflects a position on the disk
1889 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1892 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1893 struct cfq_queue *cur_cfqq)
1895 struct cfq_queue *cfqq;
1897 if (cfq_class_idle(cur_cfqq))
1899 if (!cfq_cfqq_sync(cur_cfqq))
1901 if (CFQQ_SEEKY(cur_cfqq))
1905 * Don't search priority tree if it's the only queue in the group.
1907 if (cur_cfqq->cfqg->nr_cfqq == 1)
1911 * We should notice if some of the queues are cooperating, eg
1912 * working closely on the same area of the disk. In that case,
1913 * we can group them together and don't waste time idling.
1915 cfqq = cfqq_close(cfqd, cur_cfqq);
1919 /* If new queue belongs to different cfq_group, don't choose it */
1920 if (cur_cfqq->cfqg != cfqq->cfqg)
1924 * It only makes sense to merge sync queues.
1926 if (!cfq_cfqq_sync(cfqq))
1928 if (CFQQ_SEEKY(cfqq))
1932 * Do not merge queues of different priority classes
1934 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1941 * Determine whether we should enforce idle window for this queue.
1944 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1946 enum wl_prio_t prio = cfqq_prio(cfqq);
1947 struct cfq_rb_root *service_tree = cfqq->service_tree;
1949 BUG_ON(!service_tree);
1950 BUG_ON(!service_tree->count);
1952 if (!cfqd->cfq_slice_idle)
1955 /* We never do for idle class queues. */
1956 if (prio == IDLE_WORKLOAD)
1959 /* We do for queues that were marked with idle window flag. */
1960 if (cfq_cfqq_idle_window(cfqq) &&
1961 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1965 * Otherwise, we do only if they are the last ones
1966 * in their service tree.
1968 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1970 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1971 service_tree->count);
1975 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1977 struct cfq_queue *cfqq = cfqd->active_queue;
1978 struct cfq_io_context *cic;
1979 unsigned long sl, group_idle = 0;
1982 * SSD device without seek penalty, disable idling. But only do so
1983 * for devices that support queuing, otherwise we still have a problem
1984 * with sync vs async workloads.
1986 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1989 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1990 WARN_ON(cfq_cfqq_slice_new(cfqq));
1993 * idle is disabled, either manually or by past process history
1995 if (!cfq_should_idle(cfqd, cfqq)) {
1996 /* no queue idling. Check for group idling */
1997 if (cfqd->cfq_group_idle)
1998 group_idle = cfqd->cfq_group_idle;
2004 * still active requests from this queue, don't idle
2006 if (cfqq->dispatched)
2010 * task has exited, don't wait
2012 cic = cfqd->active_cic;
2013 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2017 * If our average think time is larger than the remaining time
2018 * slice, then don't idle. This avoids overrunning the allotted
2021 if (sample_valid(cic->ttime_samples) &&
2022 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
2023 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
2028 /* There are other queues in the group, don't do group idle */
2029 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2032 cfq_mark_cfqq_wait_request(cfqq);
2035 sl = cfqd->cfq_group_idle;
2037 sl = cfqd->cfq_slice_idle;
2039 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2040 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2041 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2042 group_idle ? 1 : 0);
2046 * Move request from internal lists to the request queue dispatch list.
2048 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2050 struct cfq_data *cfqd = q->elevator->elevator_data;
2051 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2053 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2055 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2056 cfq_remove_request(rq);
2058 (RQ_CFQG(rq))->dispatched++;
2059 elv_dispatch_sort(q, rq);
2061 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2062 cfqq->nr_sectors += blk_rq_sectors(rq);
2063 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2064 rq_data_dir(rq), rq_is_sync(rq));
2068 * return expired entry, or NULL to just start from scratch in rbtree
2070 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2072 struct request *rq = NULL;
2074 if (cfq_cfqq_fifo_expire(cfqq))
2077 cfq_mark_cfqq_fifo_expire(cfqq);
2079 if (list_empty(&cfqq->fifo))
2082 rq = rq_entry_fifo(cfqq->fifo.next);
2083 if (time_before(jiffies, rq_fifo_time(rq)))
2086 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2091 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2093 const int base_rq = cfqd->cfq_slice_async_rq;
2095 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2097 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
2101 * Must be called with the queue_lock held.
2103 static int cfqq_process_refs(struct cfq_queue *cfqq)
2105 int process_refs, io_refs;
2107 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2108 process_refs = cfqq->ref - io_refs;
2109 BUG_ON(process_refs < 0);
2110 return process_refs;
2113 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2115 int process_refs, new_process_refs;
2116 struct cfq_queue *__cfqq;
2119 * If there are no process references on the new_cfqq, then it is
2120 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2121 * chain may have dropped their last reference (not just their
2122 * last process reference).
2124 if (!cfqq_process_refs(new_cfqq))
2127 /* Avoid a circular list and skip interim queue merges */
2128 while ((__cfqq = new_cfqq->new_cfqq)) {
2134 process_refs = cfqq_process_refs(cfqq);
2135 new_process_refs = cfqq_process_refs(new_cfqq);
2137 * If the process for the cfqq has gone away, there is no
2138 * sense in merging the queues.
2140 if (process_refs == 0 || new_process_refs == 0)
2144 * Merge in the direction of the lesser amount of work.
2146 if (new_process_refs >= process_refs) {
2147 cfqq->new_cfqq = new_cfqq;
2148 new_cfqq->ref += process_refs;
2150 new_cfqq->new_cfqq = cfqq;
2151 cfqq->ref += new_process_refs;
2155 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2156 struct cfq_group *cfqg, enum wl_prio_t prio)
2158 struct cfq_queue *queue;
2160 bool key_valid = false;
2161 unsigned long lowest_key = 0;
2162 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2164 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2165 /* select the one with lowest rb_key */
2166 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2168 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2169 lowest_key = queue->rb_key;
2178 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2182 struct cfq_rb_root *st;
2183 unsigned group_slice;
2184 enum wl_prio_t original_prio = cfqd->serving_prio;
2186 /* Choose next priority. RT > BE > IDLE */
2187 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2188 cfqd->serving_prio = RT_WORKLOAD;
2189 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2190 cfqd->serving_prio = BE_WORKLOAD;
2192 cfqd->serving_prio = IDLE_WORKLOAD;
2193 cfqd->workload_expires = jiffies + 1;
2197 if (original_prio != cfqd->serving_prio)
2201 * For RT and BE, we have to choose also the type
2202 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2205 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2209 * check workload expiration, and that we still have other queues ready
2211 if (count && !time_after(jiffies, cfqd->workload_expires))
2215 /* otherwise select new workload type */
2216 cfqd->serving_type =
2217 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2218 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2222 * the workload slice is computed as a fraction of target latency
2223 * proportional to the number of queues in that workload, over
2224 * all the queues in the same priority class
2226 group_slice = cfq_group_slice(cfqd, cfqg);
2228 slice = group_slice * count /
2229 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2230 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2232 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2236 * Async queues are currently system wide. Just taking
2237 * proportion of queues with-in same group will lead to higher
2238 * async ratio system wide as generally root group is going
2239 * to have higher weight. A more accurate thing would be to
2240 * calculate system wide asnc/sync ratio.
2242 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2243 tmp = tmp/cfqd->busy_queues;
2244 slice = min_t(unsigned, slice, tmp);
2246 /* async workload slice is scaled down according to
2247 * the sync/async slice ratio. */
2248 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2250 /* sync workload slice is at least 2 * cfq_slice_idle */
2251 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2253 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2254 cfq_log(cfqd, "workload slice:%d", slice);
2255 cfqd->workload_expires = jiffies + slice;
2258 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2260 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2261 struct cfq_group *cfqg;
2263 if (RB_EMPTY_ROOT(&st->rb))
2265 cfqg = cfq_rb_first_group(st);
2266 update_min_vdisktime(st);
2270 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2272 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2274 cfqd->serving_group = cfqg;
2276 /* Restore the workload type data */
2277 if (cfqg->saved_workload_slice) {
2278 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2279 cfqd->serving_type = cfqg->saved_workload;
2280 cfqd->serving_prio = cfqg->saved_serving_prio;
2282 cfqd->workload_expires = jiffies - 1;
2284 choose_service_tree(cfqd, cfqg);
2288 * Select a queue for service. If we have a current active queue,
2289 * check whether to continue servicing it, or retrieve and set a new one.
2291 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2293 struct cfq_queue *cfqq, *new_cfqq = NULL;
2295 cfqq = cfqd->active_queue;
2299 if (!cfqd->rq_queued)
2303 * We were waiting for group to get backlogged. Expire the queue
2305 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2309 * The active queue has run out of time, expire it and select new.
2311 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2313 * If slice had not expired at the completion of last request
2314 * we might not have turned on wait_busy flag. Don't expire
2315 * the queue yet. Allow the group to get backlogged.
2317 * The very fact that we have used the slice, that means we
2318 * have been idling all along on this queue and it should be
2319 * ok to wait for this request to complete.
2321 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2322 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2326 goto check_group_idle;
2330 * The active queue has requests and isn't expired, allow it to
2333 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2337 * If another queue has a request waiting within our mean seek
2338 * distance, let it run. The expire code will check for close
2339 * cooperators and put the close queue at the front of the service
2340 * tree. If possible, merge the expiring queue with the new cfqq.
2342 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2344 if (!cfqq->new_cfqq)
2345 cfq_setup_merge(cfqq, new_cfqq);
2350 * No requests pending. If the active queue still has requests in
2351 * flight or is idling for a new request, allow either of these
2352 * conditions to happen (or time out) before selecting a new queue.
2354 if (timer_pending(&cfqd->idle_slice_timer)) {
2360 * This is a deep seek queue, but the device is much faster than
2361 * the queue can deliver, don't idle
2363 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2364 (cfq_cfqq_slice_new(cfqq) ||
2365 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2366 cfq_clear_cfqq_deep(cfqq);
2367 cfq_clear_cfqq_idle_window(cfqq);
2370 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2376 * If group idle is enabled and there are requests dispatched from
2377 * this group, wait for requests to complete.
2380 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2381 && cfqq->cfqg->dispatched) {
2387 cfq_slice_expired(cfqd, 0);
2390 * Current queue expired. Check if we have to switch to a new
2394 cfq_choose_cfqg(cfqd);
2396 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2401 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2405 while (cfqq->next_rq) {
2406 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2410 BUG_ON(!list_empty(&cfqq->fifo));
2412 /* By default cfqq is not expired if it is empty. Do it explicitly */
2413 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2418 * Drain our current requests. Used for barriers and when switching
2419 * io schedulers on-the-fly.
2421 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2423 struct cfq_queue *cfqq;
2426 /* Expire the timeslice of the current active queue first */
2427 cfq_slice_expired(cfqd, 0);
2428 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2429 __cfq_set_active_queue(cfqd, cfqq);
2430 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2433 BUG_ON(cfqd->busy_queues);
2435 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2439 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2440 struct cfq_queue *cfqq)
2442 /* the queue hasn't finished any request, can't estimate */
2443 if (cfq_cfqq_slice_new(cfqq))
2445 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2452 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2454 unsigned int max_dispatch;
2457 * Drain async requests before we start sync IO
2459 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2463 * If this is an async queue and we have sync IO in flight, let it wait
2465 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2468 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2469 if (cfq_class_idle(cfqq))
2473 * Does this cfqq already have too much IO in flight?
2475 if (cfqq->dispatched >= max_dispatch) {
2476 bool promote_sync = false;
2478 * idle queue must always only have a single IO in flight
2480 if (cfq_class_idle(cfqq))
2484 * If there is only one sync queue
2485 * we can ignore async queue here and give the sync
2486 * queue no dispatch limit. The reason is a sync queue can
2487 * preempt async queue, limiting the sync queue doesn't make
2488 * sense. This is useful for aiostress test.
2490 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2491 promote_sync = true;
2494 * We have other queues, don't allow more IO from this one
2496 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2501 * Sole queue user, no limit
2503 if (cfqd->busy_queues == 1 || promote_sync)
2507 * Normally we start throttling cfqq when cfq_quantum/2
2508 * requests have been dispatched. But we can drive
2509 * deeper queue depths at the beginning of slice
2510 * subjected to upper limit of cfq_quantum.
2512 max_dispatch = cfqd->cfq_quantum;
2516 * Async queues must wait a bit before being allowed dispatch.
2517 * We also ramp up the dispatch depth gradually for async IO,
2518 * based on the last sync IO we serviced
2520 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2521 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2524 depth = last_sync / cfqd->cfq_slice[1];
2525 if (!depth && !cfqq->dispatched)
2527 if (depth < max_dispatch)
2528 max_dispatch = depth;
2532 * If we're below the current max, allow a dispatch
2534 return cfqq->dispatched < max_dispatch;
2538 * Dispatch a request from cfqq, moving them to the request queue
2541 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2545 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2547 if (!cfq_may_dispatch(cfqd, cfqq))
2551 * follow expired path, else get first next available
2553 rq = cfq_check_fifo(cfqq);
2558 * insert request into driver dispatch list
2560 cfq_dispatch_insert(cfqd->queue, rq);
2562 if (!cfqd->active_cic) {
2563 struct cfq_io_context *cic = RQ_CIC(rq);
2565 atomic_long_inc(&cic->ioc->refcount);
2566 cfqd->active_cic = cic;
2573 * Find the cfqq that we need to service and move a request from that to the
2576 static int cfq_dispatch_requests(struct request_queue *q, int force)
2578 struct cfq_data *cfqd = q->elevator->elevator_data;
2579 struct cfq_queue *cfqq;
2581 if (!cfqd->busy_queues)
2584 if (unlikely(force))
2585 return cfq_forced_dispatch(cfqd);
2587 cfqq = cfq_select_queue(cfqd);
2592 * Dispatch a request from this cfqq, if it is allowed
2594 if (!cfq_dispatch_request(cfqd, cfqq))
2597 cfqq->slice_dispatch++;
2598 cfq_clear_cfqq_must_dispatch(cfqq);
2601 * expire an async queue immediately if it has used up its slice. idle
2602 * queue always expire after 1 dispatch round.
2604 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2605 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2606 cfq_class_idle(cfqq))) {
2607 cfqq->slice_end = jiffies + 1;
2608 cfq_slice_expired(cfqd, 0);
2611 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2616 * task holds one reference to the queue, dropped when task exits. each rq
2617 * in-flight on this queue also holds a reference, dropped when rq is freed.
2619 * Each cfq queue took a reference on the parent group. Drop it now.
2620 * queue lock must be held here.
2622 static void cfq_put_queue(struct cfq_queue *cfqq)
2624 struct cfq_data *cfqd = cfqq->cfqd;
2625 struct cfq_group *cfqg;
2627 BUG_ON(cfqq->ref <= 0);
2633 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2634 BUG_ON(rb_first(&cfqq->sort_list));
2635 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2638 if (unlikely(cfqd->active_queue == cfqq)) {
2639 __cfq_slice_expired(cfqd, cfqq, 0);
2640 cfq_schedule_dispatch(cfqd);
2643 BUG_ON(cfq_cfqq_on_rr(cfqq));
2644 kmem_cache_free(cfq_pool, cfqq);
2649 * Call func for each cic attached to this ioc.
2652 call_for_each_cic(struct io_context *ioc,
2653 void (*func)(struct io_context *, struct cfq_io_context *))
2655 struct cfq_io_context *cic;
2656 struct hlist_node *n;
2660 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2666 static void cfq_cic_free_rcu(struct rcu_head *head)
2668 struct cfq_io_context *cic;
2670 cic = container_of(head, struct cfq_io_context, rcu_head);
2672 kmem_cache_free(cfq_ioc_pool, cic);
2673 elv_ioc_count_dec(cfq_ioc_count);
2677 * CFQ scheduler is exiting, grab exit lock and check
2678 * the pending io context count. If it hits zero,
2679 * complete ioc_gone and set it back to NULL
2681 spin_lock(&ioc_gone_lock);
2682 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2686 spin_unlock(&ioc_gone_lock);
2690 static void cfq_cic_free(struct cfq_io_context *cic)
2692 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2695 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2697 unsigned long flags;
2698 unsigned long dead_key = (unsigned long) cic->key;
2700 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2702 spin_lock_irqsave(&ioc->lock, flags);
2703 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2704 hlist_del_rcu(&cic->cic_list);
2705 spin_unlock_irqrestore(&ioc->lock, flags);
2711 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2712 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2713 * and ->trim() which is called with the task lock held
2715 static void cfq_free_io_context(struct io_context *ioc)
2718 * ioc->refcount is zero here, or we are called from elv_unregister(),
2719 * so no more cic's are allowed to be linked into this ioc. So it
2720 * should be ok to iterate over the known list, we will see all cic's
2721 * since no new ones are added.
2723 call_for_each_cic(ioc, cic_free_func);
2726 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2728 struct cfq_queue *__cfqq, *next;
2731 * If this queue was scheduled to merge with another queue, be
2732 * sure to drop the reference taken on that queue (and others in
2733 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2735 __cfqq = cfqq->new_cfqq;
2737 if (__cfqq == cfqq) {
2738 WARN(1, "cfqq->new_cfqq loop detected\n");
2741 next = __cfqq->new_cfqq;
2742 cfq_put_queue(__cfqq);
2747 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2749 if (unlikely(cfqq == cfqd->active_queue)) {
2750 __cfq_slice_expired(cfqd, cfqq, 0);
2751 cfq_schedule_dispatch(cfqd);
2754 cfq_put_cooperator(cfqq);
2756 cfq_put_queue(cfqq);
2759 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2760 struct cfq_io_context *cic)
2762 struct io_context *ioc = cic->ioc;
2764 list_del_init(&cic->queue_list);
2767 * Make sure dead mark is seen for dead queues
2770 cic->key = cfqd_dead_key(cfqd);
2772 if (ioc->ioc_data == cic)
2773 rcu_assign_pointer(ioc->ioc_data, NULL);
2775 if (cic->cfqq[BLK_RW_ASYNC]) {
2776 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2777 cic->cfqq[BLK_RW_ASYNC] = NULL;
2780 if (cic->cfqq[BLK_RW_SYNC]) {
2781 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2782 cic->cfqq[BLK_RW_SYNC] = NULL;
2786 static void cfq_exit_single_io_context(struct io_context *ioc,
2787 struct cfq_io_context *cic)
2789 struct cfq_data *cfqd = cic_to_cfqd(cic);
2792 struct request_queue *q = cfqd->queue;
2793 unsigned long flags;
2795 spin_lock_irqsave(q->queue_lock, flags);
2798 * Ensure we get a fresh copy of the ->key to prevent
2799 * race between exiting task and queue
2801 smp_read_barrier_depends();
2802 if (cic->key == cfqd)
2803 __cfq_exit_single_io_context(cfqd, cic);
2805 spin_unlock_irqrestore(q->queue_lock, flags);
2810 * The process that ioc belongs to has exited, we need to clean up
2811 * and put the internal structures we have that belongs to that process.
2813 static void cfq_exit_io_context(struct io_context *ioc)
2815 call_for_each_cic(ioc, cfq_exit_single_io_context);
2818 static struct cfq_io_context *
2819 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2821 struct cfq_io_context *cic;
2823 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2826 cic->last_end_request = jiffies;
2827 INIT_LIST_HEAD(&cic->queue_list);
2828 INIT_HLIST_NODE(&cic->cic_list);
2829 cic->dtor = cfq_free_io_context;
2830 cic->exit = cfq_exit_io_context;
2831 elv_ioc_count_inc(cfq_ioc_count);
2837 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2839 struct task_struct *tsk = current;
2842 if (!cfq_cfqq_prio_changed(cfqq))
2845 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2846 switch (ioprio_class) {
2848 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2849 case IOPRIO_CLASS_NONE:
2851 * no prio set, inherit CPU scheduling settings
2853 cfqq->ioprio = task_nice_ioprio(tsk);
2854 cfqq->ioprio_class = task_nice_ioclass(tsk);
2856 case IOPRIO_CLASS_RT:
2857 cfqq->ioprio = task_ioprio(ioc);
2858 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2860 case IOPRIO_CLASS_BE:
2861 cfqq->ioprio = task_ioprio(ioc);
2862 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2864 case IOPRIO_CLASS_IDLE:
2865 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2867 cfq_clear_cfqq_idle_window(cfqq);
2872 * keep track of original prio settings in case we have to temporarily
2873 * elevate the priority of this queue
2875 cfqq->org_ioprio = cfqq->ioprio;
2876 cfqq->org_ioprio_class = cfqq->ioprio_class;
2877 cfq_clear_cfqq_prio_changed(cfqq);
2880 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2882 struct cfq_data *cfqd = cic_to_cfqd(cic);
2883 struct cfq_queue *cfqq;
2884 unsigned long flags;
2886 if (unlikely(!cfqd))
2889 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2891 cfqq = cic->cfqq[BLK_RW_ASYNC];
2893 struct cfq_queue *new_cfqq;
2894 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2897 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2898 cfq_put_queue(cfqq);
2902 cfqq = cic->cfqq[BLK_RW_SYNC];
2904 cfq_mark_cfqq_prio_changed(cfqq);
2906 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2909 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2911 call_for_each_cic(ioc, changed_ioprio);
2912 ioc->ioprio_changed = 0;
2915 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2916 pid_t pid, bool is_sync)
2918 RB_CLEAR_NODE(&cfqq->rb_node);
2919 RB_CLEAR_NODE(&cfqq->p_node);
2920 INIT_LIST_HEAD(&cfqq->fifo);
2925 cfq_mark_cfqq_prio_changed(cfqq);
2928 if (!cfq_class_idle(cfqq))
2929 cfq_mark_cfqq_idle_window(cfqq);
2930 cfq_mark_cfqq_sync(cfqq);
2935 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2936 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2938 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2939 struct cfq_data *cfqd = cic_to_cfqd(cic);
2940 unsigned long flags;
2941 struct request_queue *q;
2943 if (unlikely(!cfqd))
2948 spin_lock_irqsave(q->queue_lock, flags);
2952 * Drop reference to sync queue. A new sync queue will be
2953 * assigned in new group upon arrival of a fresh request.
2955 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2956 cic_set_cfqq(cic, NULL, 1);
2957 cfq_put_queue(sync_cfqq);
2960 spin_unlock_irqrestore(q->queue_lock, flags);
2963 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2965 call_for_each_cic(ioc, changed_cgroup);
2966 ioc->cgroup_changed = 0;
2968 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2970 static struct cfq_queue *
2971 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2972 struct io_context *ioc, gfp_t gfp_mask)
2974 struct cfq_queue *cfqq, *new_cfqq = NULL;
2975 struct cfq_io_context *cic;
2976 struct cfq_group *cfqg;
2979 cfqg = cfq_get_cfqg(cfqd);
2980 cic = cfq_cic_lookup(cfqd, ioc);
2981 /* cic always exists here */
2982 cfqq = cic_to_cfqq(cic, is_sync);
2985 * Always try a new alloc if we fell back to the OOM cfqq
2986 * originally, since it should just be a temporary situation.
2988 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2993 } else if (gfp_mask & __GFP_WAIT) {
2994 spin_unlock_irq(cfqd->queue->queue_lock);
2995 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2996 gfp_mask | __GFP_ZERO,
2998 spin_lock_irq(cfqd->queue->queue_lock);
3002 cfqq = kmem_cache_alloc_node(cfq_pool,
3003 gfp_mask | __GFP_ZERO,
3008 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3009 cfq_init_prio_data(cfqq, ioc);
3010 cfq_link_cfqq_cfqg(cfqq, cfqg);
3011 cfq_log_cfqq(cfqd, cfqq, "alloced");
3013 cfqq = &cfqd->oom_cfqq;
3017 kmem_cache_free(cfq_pool, new_cfqq);
3022 static struct cfq_queue **
3023 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3025 switch (ioprio_class) {
3026 case IOPRIO_CLASS_RT:
3027 return &cfqd->async_cfqq[0][ioprio];
3028 case IOPRIO_CLASS_BE:
3029 return &cfqd->async_cfqq[1][ioprio];
3030 case IOPRIO_CLASS_IDLE:
3031 return &cfqd->async_idle_cfqq;
3037 static struct cfq_queue *
3038 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3041 const int ioprio = task_ioprio(ioc);
3042 const int ioprio_class = task_ioprio_class(ioc);
3043 struct cfq_queue **async_cfqq = NULL;
3044 struct cfq_queue *cfqq = NULL;
3047 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3052 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3055 * pin the queue now that it's allocated, scheduler exit will prune it
3057 if (!is_sync && !(*async_cfqq)) {
3067 * We drop cfq io contexts lazily, so we may find a dead one.
3070 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3071 struct cfq_io_context *cic)
3073 unsigned long flags;
3075 WARN_ON(!list_empty(&cic->queue_list));
3076 BUG_ON(cic->key != cfqd_dead_key(cfqd));
3078 spin_lock_irqsave(&ioc->lock, flags);
3080 BUG_ON(ioc->ioc_data == cic);
3082 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3083 hlist_del_rcu(&cic->cic_list);
3084 spin_unlock_irqrestore(&ioc->lock, flags);
3089 static struct cfq_io_context *
3090 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3092 struct cfq_io_context *cic;
3093 unsigned long flags;
3101 * we maintain a last-hit cache, to avoid browsing over the tree
3103 cic = rcu_dereference(ioc->ioc_data);
3104 if (cic && cic->key == cfqd) {
3110 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3114 if (unlikely(cic->key != cfqd)) {
3115 cfq_drop_dead_cic(cfqd, ioc, cic);
3120 spin_lock_irqsave(&ioc->lock, flags);
3121 rcu_assign_pointer(ioc->ioc_data, cic);
3122 spin_unlock_irqrestore(&ioc->lock, flags);
3130 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3131 * the process specific cfq io context when entered from the block layer.
3132 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3134 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3135 struct cfq_io_context *cic, gfp_t gfp_mask)
3137 unsigned long flags;
3140 ret = radix_tree_preload(gfp_mask);
3145 spin_lock_irqsave(&ioc->lock, flags);
3146 ret = radix_tree_insert(&ioc->radix_root,
3147 cfqd->cic_index, cic);
3149 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3150 spin_unlock_irqrestore(&ioc->lock, flags);
3152 radix_tree_preload_end();
3155 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3156 list_add(&cic->queue_list, &cfqd->cic_list);
3157 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3162 printk(KERN_ERR "cfq: cic link failed!\n");
3168 * Setup general io context and cfq io context. There can be several cfq
3169 * io contexts per general io context, if this process is doing io to more
3170 * than one device managed by cfq.
3172 static struct cfq_io_context *
3173 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3175 struct io_context *ioc = NULL;
3176 struct cfq_io_context *cic;
3178 might_sleep_if(gfp_mask & __GFP_WAIT);
3180 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3184 cic = cfq_cic_lookup(cfqd, ioc);
3188 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3192 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3196 smp_read_barrier_depends();
3197 if (unlikely(ioc->ioprio_changed))
3198 cfq_ioc_set_ioprio(ioc);
3200 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3201 if (unlikely(ioc->cgroup_changed))
3202 cfq_ioc_set_cgroup(ioc);
3208 put_io_context(ioc);
3213 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3215 unsigned long elapsed = jiffies - cic->last_end_request;
3216 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3218 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3219 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3220 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3224 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3228 sector_t n_sec = blk_rq_sectors(rq);
3229 if (cfqq->last_request_pos) {
3230 if (cfqq->last_request_pos < blk_rq_pos(rq))
3231 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3233 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3236 cfqq->seek_history <<= 1;
3237 if (blk_queue_nonrot(cfqd->queue))
3238 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3240 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3244 * Disable idle window if the process thinks too long or seeks so much that
3248 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3249 struct cfq_io_context *cic)
3251 int old_idle, enable_idle;
3254 * Don't idle for async or idle io prio class
3256 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3259 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3261 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3262 cfq_mark_cfqq_deep(cfqq);
3264 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3266 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3267 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3269 else if (sample_valid(cic->ttime_samples)) {
3270 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3276 if (old_idle != enable_idle) {
3277 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3279 cfq_mark_cfqq_idle_window(cfqq);
3281 cfq_clear_cfqq_idle_window(cfqq);
3286 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3287 * no or if we aren't sure, a 1 will cause a preempt.
3290 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3293 struct cfq_queue *cfqq;
3295 cfqq = cfqd->active_queue;
3299 if (cfq_class_idle(new_cfqq))
3302 if (cfq_class_idle(cfqq))
3306 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3308 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3312 * if the new request is sync, but the currently running queue is
3313 * not, let the sync request have priority.
3315 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3318 if (new_cfqq->cfqg != cfqq->cfqg)
3321 if (cfq_slice_used(cfqq))
3324 /* Allow preemption only if we are idling on sync-noidle tree */
3325 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3326 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3327 new_cfqq->service_tree->count == 2 &&
3328 RB_EMPTY_ROOT(&cfqq->sort_list))
3332 * So both queues are sync. Let the new request get disk time if
3333 * it's a metadata request and the current queue is doing regular IO.
3335 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3339 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3341 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3344 /* An idle queue should not be idle now for some reason */
3345 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3348 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3352 * if this request is as-good as one we would expect from the
3353 * current cfqq, let it preempt
3355 if (cfq_rq_close(cfqd, cfqq, rq))
3362 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3363 * let it have half of its nominal slice.
3365 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3367 struct cfq_queue *old_cfqq = cfqd->active_queue;
3369 cfq_log_cfqq(cfqd, cfqq, "preempt");
3370 cfq_slice_expired(cfqd, 1);
3373 * workload type is changed, don't save slice, otherwise preempt
3376 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3377 cfqq->cfqg->saved_workload_slice = 0;
3380 * Put the new queue at the front of the of the current list,
3381 * so we know that it will be selected next.
3383 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3385 cfq_service_tree_add(cfqd, cfqq, 1);
3387 cfqq->slice_end = 0;
3388 cfq_mark_cfqq_slice_new(cfqq);
3392 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3393 * something we should do about it
3396 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3399 struct cfq_io_context *cic = RQ_CIC(rq);
3402 if (rq->cmd_flags & REQ_META)
3403 cfqq->meta_pending++;
3405 cfq_update_io_thinktime(cfqd, cic);
3406 cfq_update_io_seektime(cfqd, cfqq, rq);
3407 cfq_update_idle_window(cfqd, cfqq, cic);
3409 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3411 if (cfqq == cfqd->active_queue) {
3413 * Remember that we saw a request from this process, but
3414 * don't start queuing just yet. Otherwise we risk seeing lots
3415 * of tiny requests, because we disrupt the normal plugging
3416 * and merging. If the request is already larger than a single
3417 * page, let it rip immediately. For that case we assume that
3418 * merging is already done. Ditto for a busy system that
3419 * has other work pending, don't risk delaying until the
3420 * idle timer unplug to continue working.
3422 if (cfq_cfqq_wait_request(cfqq)) {
3423 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3424 cfqd->busy_queues > 1) {
3425 cfq_del_timer(cfqd, cfqq);
3426 cfq_clear_cfqq_wait_request(cfqq);
3427 __blk_run_queue(cfqd->queue);
3429 cfq_blkiocg_update_idle_time_stats(
3431 cfq_mark_cfqq_must_dispatch(cfqq);
3434 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3436 * not the active queue - expire current slice if it is
3437 * idle and has expired it's mean thinktime or this new queue
3438 * has some old slice time left and is of higher priority or
3439 * this new queue is RT and the current one is BE
3441 cfq_preempt_queue(cfqd, cfqq);
3442 __blk_run_queue(cfqd->queue);
3446 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3448 struct cfq_data *cfqd = q->elevator->elevator_data;
3449 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3451 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3452 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3454 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3455 list_add_tail(&rq->queuelist, &cfqq->fifo);
3457 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3458 &cfqd->serving_group->blkg, rq_data_dir(rq),
3460 cfq_rq_enqueued(cfqd, cfqq, rq);
3464 * Update hw_tag based on peak queue depth over 50 samples under
3467 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3469 struct cfq_queue *cfqq = cfqd->active_queue;
3471 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3472 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3474 if (cfqd->hw_tag == 1)
3477 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3478 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3482 * If active queue hasn't enough requests and can idle, cfq might not
3483 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3486 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3487 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3488 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3491 if (cfqd->hw_tag_samples++ < 50)
3494 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3500 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3502 struct cfq_io_context *cic = cfqd->active_cic;
3504 /* If the queue already has requests, don't wait */
3505 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3508 /* If there are other queues in the group, don't wait */
3509 if (cfqq->cfqg->nr_cfqq > 1)
3512 if (cfq_slice_used(cfqq))
3515 /* if slice left is less than think time, wait busy */
3516 if (cic && sample_valid(cic->ttime_samples)
3517 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3521 * If think times is less than a jiffy than ttime_mean=0 and above
3522 * will not be true. It might happen that slice has not expired yet
3523 * but will expire soon (4-5 ns) during select_queue(). To cover the
3524 * case where think time is less than a jiffy, mark the queue wait
3525 * busy if only 1 jiffy is left in the slice.
3527 if (cfqq->slice_end - jiffies == 1)
3533 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3535 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3536 struct cfq_data *cfqd = cfqq->cfqd;
3537 const int sync = rq_is_sync(rq);
3541 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3542 !!(rq->cmd_flags & REQ_NOIDLE));
3544 cfq_update_hw_tag(cfqd);
3546 WARN_ON(!cfqd->rq_in_driver);
3547 WARN_ON(!cfqq->dispatched);
3548 cfqd->rq_in_driver--;
3550 (RQ_CFQG(rq))->dispatched--;
3551 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3552 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3553 rq_data_dir(rq), rq_is_sync(rq));
3555 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3558 RQ_CIC(rq)->last_end_request = now;
3559 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3560 cfqd->last_delayed_sync = now;
3564 * If this is the active queue, check if it needs to be expired,
3565 * or if we want to idle in case it has no pending requests.
3567 if (cfqd->active_queue == cfqq) {
3568 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3570 if (cfq_cfqq_slice_new(cfqq)) {
3571 cfq_set_prio_slice(cfqd, cfqq);
3572 cfq_clear_cfqq_slice_new(cfqq);
3576 * Should we wait for next request to come in before we expire
3579 if (cfq_should_wait_busy(cfqd, cfqq)) {
3580 unsigned long extend_sl = cfqd->cfq_slice_idle;
3581 if (!cfqd->cfq_slice_idle)
3582 extend_sl = cfqd->cfq_group_idle;
3583 cfqq->slice_end = jiffies + extend_sl;
3584 cfq_mark_cfqq_wait_busy(cfqq);
3585 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3589 * Idling is not enabled on:
3591 * - idle-priority queues
3593 * - queues with still some requests queued
3594 * - when there is a close cooperator
3596 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3597 cfq_slice_expired(cfqd, 1);
3598 else if (sync && cfqq_empty &&
3599 !cfq_close_cooperator(cfqd, cfqq)) {
3600 cfq_arm_slice_timer(cfqd);
3604 if (!cfqd->rq_in_driver)
3605 cfq_schedule_dispatch(cfqd);
3609 * we temporarily boost lower priority queues if they are holding fs exclusive
3610 * resources. they are boosted to normal prio (CLASS_BE/4)
3612 static void cfq_prio_boost(struct cfq_queue *cfqq)
3614 if (has_fs_excl()) {
3616 * boost idle prio on transactions that would lock out other
3617 * users of the filesystem
3619 if (cfq_class_idle(cfqq))
3620 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3621 if (cfqq->ioprio > IOPRIO_NORM)
3622 cfqq->ioprio = IOPRIO_NORM;
3625 * unboost the queue (if needed)
3627 cfqq->ioprio_class = cfqq->org_ioprio_class;
3628 cfqq->ioprio = cfqq->org_ioprio;
3632 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3634 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3635 cfq_mark_cfqq_must_alloc_slice(cfqq);
3636 return ELV_MQUEUE_MUST;
3639 return ELV_MQUEUE_MAY;
3642 static int cfq_may_queue(struct request_queue *q, int rw)
3644 struct cfq_data *cfqd = q->elevator->elevator_data;
3645 struct task_struct *tsk = current;
3646 struct cfq_io_context *cic;
3647 struct cfq_queue *cfqq;
3650 * don't force setup of a queue from here, as a call to may_queue
3651 * does not necessarily imply that a request actually will be queued.
3652 * so just lookup a possibly existing queue, or return 'may queue'
3655 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3657 return ELV_MQUEUE_MAY;
3659 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3661 cfq_init_prio_data(cfqq, cic->ioc);
3662 cfq_prio_boost(cfqq);
3664 return __cfq_may_queue(cfqq);
3667 return ELV_MQUEUE_MAY;
3671 * queue lock held here
3673 static void cfq_put_request(struct request *rq)
3675 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3678 const int rw = rq_data_dir(rq);
3680 BUG_ON(!cfqq->allocated[rw]);
3681 cfqq->allocated[rw]--;
3683 put_io_context(RQ_CIC(rq)->ioc);
3685 rq->elevator_private[0] = NULL;
3686 rq->elevator_private[1] = NULL;
3688 /* Put down rq reference on cfqg */
3689 cfq_put_cfqg(RQ_CFQG(rq));
3690 rq->elevator_private[2] = NULL;
3692 cfq_put_queue(cfqq);
3696 static struct cfq_queue *
3697 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3698 struct cfq_queue *cfqq)
3700 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3701 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3702 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3703 cfq_put_queue(cfqq);
3704 return cic_to_cfqq(cic, 1);
3708 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3709 * was the last process referring to said cfqq.
3711 static struct cfq_queue *
3712 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3714 if (cfqq_process_refs(cfqq) == 1) {
3715 cfqq->pid = current->pid;
3716 cfq_clear_cfqq_coop(cfqq);
3717 cfq_clear_cfqq_split_coop(cfqq);
3721 cic_set_cfqq(cic, NULL, 1);
3723 cfq_put_cooperator(cfqq);
3725 cfq_put_queue(cfqq);
3729 * Allocate cfq data structures associated with this request.
3732 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3734 struct cfq_data *cfqd = q->elevator->elevator_data;
3735 struct cfq_io_context *cic;
3736 const int rw = rq_data_dir(rq);
3737 const bool is_sync = rq_is_sync(rq);
3738 struct cfq_queue *cfqq;
3739 unsigned long flags;
3741 might_sleep_if(gfp_mask & __GFP_WAIT);
3743 cic = cfq_get_io_context(cfqd, gfp_mask);
3745 spin_lock_irqsave(q->queue_lock, flags);
3751 cfqq = cic_to_cfqq(cic, is_sync);
3752 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3753 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3754 cic_set_cfqq(cic, cfqq, is_sync);
3757 * If the queue was seeky for too long, break it apart.
3759 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3760 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3761 cfqq = split_cfqq(cic, cfqq);
3767 * Check to see if this queue is scheduled to merge with
3768 * another, closely cooperating queue. The merging of
3769 * queues happens here as it must be done in process context.
3770 * The reference on new_cfqq was taken in merge_cfqqs.
3773 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3776 cfqq->allocated[rw]++;
3779 rq->elevator_private[0] = cic;
3780 rq->elevator_private[1] = cfqq;
3781 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3782 spin_unlock_irqrestore(q->queue_lock, flags);
3787 put_io_context(cic->ioc);
3789 cfq_schedule_dispatch(cfqd);
3790 spin_unlock_irqrestore(q->queue_lock, flags);
3791 cfq_log(cfqd, "set_request fail");
3795 static void cfq_kick_queue(struct work_struct *work)
3797 struct cfq_data *cfqd =
3798 container_of(work, struct cfq_data, unplug_work);
3799 struct request_queue *q = cfqd->queue;
3801 spin_lock_irq(q->queue_lock);
3802 __blk_run_queue(cfqd->queue);
3803 spin_unlock_irq(q->queue_lock);
3807 * Timer running if the active_queue is currently idling inside its time slice
3809 static void cfq_idle_slice_timer(unsigned long data)
3811 struct cfq_data *cfqd = (struct cfq_data *) data;
3812 struct cfq_queue *cfqq;
3813 unsigned long flags;
3816 cfq_log(cfqd, "idle timer fired");
3818 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3820 cfqq = cfqd->active_queue;
3825 * We saw a request before the queue expired, let it through
3827 if (cfq_cfqq_must_dispatch(cfqq))
3833 if (cfq_slice_used(cfqq))
3837 * only expire and reinvoke request handler, if there are
3838 * other queues with pending requests
3840 if (!cfqd->busy_queues)
3844 * not expired and it has a request pending, let it dispatch
3846 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3850 * Queue depth flag is reset only when the idle didn't succeed
3852 cfq_clear_cfqq_deep(cfqq);
3855 cfq_slice_expired(cfqd, timed_out);
3857 cfq_schedule_dispatch(cfqd);
3859 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3862 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3864 del_timer_sync(&cfqd->idle_slice_timer);
3865 cancel_work_sync(&cfqd->unplug_work);
3868 static void cfq_put_async_queues(struct cfq_data *cfqd)
3872 for (i = 0; i < IOPRIO_BE_NR; i++) {
3873 if (cfqd->async_cfqq[0][i])
3874 cfq_put_queue(cfqd->async_cfqq[0][i]);
3875 if (cfqd->async_cfqq[1][i])
3876 cfq_put_queue(cfqd->async_cfqq[1][i]);
3879 if (cfqd->async_idle_cfqq)
3880 cfq_put_queue(cfqd->async_idle_cfqq);
3883 static void cfq_exit_queue(struct elevator_queue *e)
3885 struct cfq_data *cfqd = e->elevator_data;
3886 struct request_queue *q = cfqd->queue;
3889 cfq_shutdown_timer_wq(cfqd);
3891 spin_lock_irq(q->queue_lock);
3893 if (cfqd->active_queue)
3894 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3896 while (!list_empty(&cfqd->cic_list)) {
3897 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3898 struct cfq_io_context,
3901 __cfq_exit_single_io_context(cfqd, cic);
3904 cfq_put_async_queues(cfqd);
3905 cfq_release_cfq_groups(cfqd);
3908 * If there are groups which we could not unlink from blkcg list,
3909 * wait for a rcu period for them to be freed.
3911 if (cfqd->nr_blkcg_linked_grps)
3914 spin_unlock_irq(q->queue_lock);
3916 cfq_shutdown_timer_wq(cfqd);
3918 spin_lock(&cic_index_lock);
3919 ida_remove(&cic_index_ida, cfqd->cic_index);
3920 spin_unlock(&cic_index_lock);
3923 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3924 * Do this wait only if there are other unlinked groups out
3925 * there. This can happen if cgroup deletion path claimed the
3926 * responsibility of cleaning up a group before queue cleanup code
3929 * Do not call synchronize_rcu() unconditionally as there are drivers
3930 * which create/delete request queue hundreds of times during scan/boot
3931 * and synchronize_rcu() can take significant time and slow down boot.
3938 static int cfq_alloc_cic_index(void)
3943 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3946 spin_lock(&cic_index_lock);
3947 error = ida_get_new(&cic_index_ida, &index);
3948 spin_unlock(&cic_index_lock);
3949 if (error && error != -EAGAIN)
3956 static void *cfq_init_queue(struct request_queue *q)
3958 struct cfq_data *cfqd;
3960 struct cfq_group *cfqg;
3961 struct cfq_rb_root *st;
3963 i = cfq_alloc_cic_index();
3967 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3972 * Don't need take queue_lock in the routine, since we are
3973 * initializing the ioscheduler, and nobody is using cfqd
3975 cfqd->cic_index = i;
3977 /* Init root service tree */
3978 cfqd->grp_service_tree = CFQ_RB_ROOT;
3980 /* Init root group */
3981 cfqg = &cfqd->root_group;
3982 for_each_cfqg_st(cfqg, i, j, st)
3984 RB_CLEAR_NODE(&cfqg->rb_node);
3986 /* Give preference to root group over other groups */
3987 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3989 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3991 * Set root group reference to 2. One reference will be dropped when
3992 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3993 * Other reference will remain there as we don't want to delete this
3994 * group as it is statically allocated and gets destroyed when
3995 * throtl_data goes away.
3999 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4002 cfqd->nr_blkcg_linked_grps++;
4004 /* Add group on cfqd->cfqg_list */
4005 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4008 * Not strictly needed (since RB_ROOT just clears the node and we
4009 * zeroed cfqd on alloc), but better be safe in case someone decides
4010 * to add magic to the rb code
4012 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4013 cfqd->prio_trees[i] = RB_ROOT;
4016 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4017 * Grab a permanent reference to it, so that the normal code flow
4018 * will not attempt to free it.
4020 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4021 cfqd->oom_cfqq.ref++;
4022 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4024 INIT_LIST_HEAD(&cfqd->cic_list);
4028 init_timer(&cfqd->idle_slice_timer);
4029 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4030 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4032 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4034 cfqd->cfq_quantum = cfq_quantum;
4035 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4036 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4037 cfqd->cfq_back_max = cfq_back_max;
4038 cfqd->cfq_back_penalty = cfq_back_penalty;
4039 cfqd->cfq_slice[0] = cfq_slice_async;
4040 cfqd->cfq_slice[1] = cfq_slice_sync;
4041 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4042 cfqd->cfq_slice_idle = cfq_slice_idle;
4043 cfqd->cfq_group_idle = cfq_group_idle;
4044 cfqd->cfq_latency = 1;
4047 * we optimistically start assuming sync ops weren't delayed in last
4048 * second, in order to have larger depth for async operations.
4050 cfqd->last_delayed_sync = jiffies - HZ;
4054 static void cfq_slab_kill(void)
4057 * Caller already ensured that pending RCU callbacks are completed,
4058 * so we should have no busy allocations at this point.
4061 kmem_cache_destroy(cfq_pool);
4063 kmem_cache_destroy(cfq_ioc_pool);
4066 static int __init cfq_slab_setup(void)
4068 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4072 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4083 * sysfs parts below -->
4086 cfq_var_show(unsigned int var, char *page)
4088 return sprintf(page, "%d\n", var);
4092 cfq_var_store(unsigned int *var, const char *page, size_t count)
4094 char *p = (char *) page;
4096 *var = simple_strtoul(p, &p, 10);
4100 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4101 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4103 struct cfq_data *cfqd = e->elevator_data; \
4104 unsigned int __data = __VAR; \
4106 __data = jiffies_to_msecs(__data); \
4107 return cfq_var_show(__data, (page)); \
4109 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4110 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4111 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4112 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4113 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4114 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4115 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4116 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4117 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4118 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4119 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4120 #undef SHOW_FUNCTION
4122 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4123 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4125 struct cfq_data *cfqd = e->elevator_data; \
4126 unsigned int __data; \
4127 int ret = cfq_var_store(&__data, (page), count); \
4128 if (__data < (MIN)) \
4130 else if (__data > (MAX)) \
4133 *(__PTR) = msecs_to_jiffies(__data); \
4135 *(__PTR) = __data; \
4138 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4139 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4141 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4143 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4144 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4146 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4147 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4148 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4149 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4150 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4152 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4153 #undef STORE_FUNCTION
4155 #define CFQ_ATTR(name) \
4156 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4158 static struct elv_fs_entry cfq_attrs[] = {
4160 CFQ_ATTR(fifo_expire_sync),
4161 CFQ_ATTR(fifo_expire_async),
4162 CFQ_ATTR(back_seek_max),
4163 CFQ_ATTR(back_seek_penalty),
4164 CFQ_ATTR(slice_sync),
4165 CFQ_ATTR(slice_async),
4166 CFQ_ATTR(slice_async_rq),
4167 CFQ_ATTR(slice_idle),
4168 CFQ_ATTR(group_idle),
4169 CFQ_ATTR(low_latency),
4173 static struct elevator_type iosched_cfq = {
4175 .elevator_merge_fn = cfq_merge,
4176 .elevator_merged_fn = cfq_merged_request,
4177 .elevator_merge_req_fn = cfq_merged_requests,
4178 .elevator_allow_merge_fn = cfq_allow_merge,
4179 .elevator_bio_merged_fn = cfq_bio_merged,
4180 .elevator_dispatch_fn = cfq_dispatch_requests,
4181 .elevator_add_req_fn = cfq_insert_request,
4182 .elevator_activate_req_fn = cfq_activate_request,
4183 .elevator_deactivate_req_fn = cfq_deactivate_request,
4184 .elevator_completed_req_fn = cfq_completed_request,
4185 .elevator_former_req_fn = elv_rb_former_request,
4186 .elevator_latter_req_fn = elv_rb_latter_request,
4187 .elevator_set_req_fn = cfq_set_request,
4188 .elevator_put_req_fn = cfq_put_request,
4189 .elevator_may_queue_fn = cfq_may_queue,
4190 .elevator_init_fn = cfq_init_queue,
4191 .elevator_exit_fn = cfq_exit_queue,
4192 .trim = cfq_free_io_context,
4194 .elevator_attrs = cfq_attrs,
4195 .elevator_name = "cfq",
4196 .elevator_owner = THIS_MODULE,
4199 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4200 static struct blkio_policy_type blkio_policy_cfq = {
4202 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4203 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4205 .plid = BLKIO_POLICY_PROP,
4208 static struct blkio_policy_type blkio_policy_cfq;
4211 static int __init cfq_init(void)
4214 * could be 0 on HZ < 1000 setups
4216 if (!cfq_slice_async)
4217 cfq_slice_async = 1;
4218 if (!cfq_slice_idle)
4221 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4222 if (!cfq_group_idle)
4227 if (cfq_slab_setup())
4230 elv_register(&iosched_cfq);
4231 blkio_policy_register(&blkio_policy_cfq);
4236 static void __exit cfq_exit(void)
4238 DECLARE_COMPLETION_ONSTACK(all_gone);
4239 blkio_policy_unregister(&blkio_policy_cfq);
4240 elv_unregister(&iosched_cfq);
4241 ioc_gone = &all_gone;
4242 /* ioc_gone's update must be visible before reading ioc_count */
4246 * this also protects us from entering cfq_slab_kill() with
4247 * pending RCU callbacks
4249 if (elv_ioc_count_read(cfq_ioc_count))
4250 wait_for_completion(&all_gone);
4251 ida_destroy(&cic_index_ida);
4255 module_init(cfq_init);
4256 module_exit(cfq_exit);
4258 MODULE_AUTHOR("Jens Axboe");
4259 MODULE_LICENSE("GPL");
4260 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");