2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
16 #include "blk-cgroup.h"
21 /* max queue in one round of service */
22 static const int cfq_quantum = 8;
23 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
24 /* maximum backwards seek, in KiB */
25 static const int cfq_back_max = 16 * 1024;
26 /* penalty of a backwards seek */
27 static const int cfq_back_penalty = 2;
28 static const int cfq_slice_sync = HZ / 10;
29 static int cfq_slice_async = HZ / 25;
30 static const int cfq_slice_async_rq = 2;
31 static int cfq_slice_idle = HZ / 125;
32 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
33 static const int cfq_hist_divisor = 4;
36 * offset from end of service tree
38 #define CFQ_IDLE_DELAY (HZ / 5)
41 * below this threshold, we consider thinktime immediate
43 #define CFQ_MIN_TT (2)
45 #define CFQ_SLICE_SCALE (5)
46 #define CFQ_HW_QUEUE_MIN (5)
47 #define CFQ_SERVICE_SHIFT 12
49 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
50 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
51 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
52 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
55 ((struct cfq_io_context *) (rq)->elevator_private)
56 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
58 static struct kmem_cache *cfq_pool;
59 static struct kmem_cache *cfq_ioc_pool;
61 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
62 static struct completion *ioc_gone;
63 static DEFINE_SPINLOCK(ioc_gone_lock);
65 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
66 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
67 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
69 #define sample_valid(samples) ((samples) > 80)
70 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
73 * Most of our rbtree usage is for sorting with min extraction, so
74 * if we cache the leftmost node we don't have to walk down the tree
75 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
76 * move this into the elevator for the rq sorting as well.
82 unsigned total_weight;
84 struct rb_node *active;
86 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
87 .count = 0, .min_vdisktime = 0, }
90 * Per process-grouping structure
95 /* various state flags, see below */
98 struct cfq_data *cfqd;
99 /* service_tree member */
100 struct rb_node rb_node;
101 /* service_tree key */
102 unsigned long rb_key;
103 /* prio tree member */
104 struct rb_node p_node;
105 /* prio tree root we belong to, if any */
106 struct rb_root *p_root;
107 /* sorted list of pending requests */
108 struct rb_root sort_list;
109 /* if fifo isn't expired, next request to serve */
110 struct request *next_rq;
111 /* requests queued in sort_list */
113 /* currently allocated requests */
115 /* fifo list of requests in sort_list */
116 struct list_head fifo;
118 /* time when queue got scheduled in to dispatch first request. */
119 unsigned long dispatch_start;
120 unsigned int allocated_slice;
121 unsigned int slice_dispatch;
122 /* time when first request from queue completed and slice started. */
123 unsigned long slice_start;
124 unsigned long slice_end;
127 /* pending metadata requests */
129 /* number of requests that are on the dispatch list or inside driver */
132 /* io prio of this group */
133 unsigned short ioprio, org_ioprio;
134 unsigned short ioprio_class, org_ioprio_class;
139 sector_t last_request_pos;
141 struct cfq_rb_root *service_tree;
142 struct cfq_queue *new_cfqq;
143 struct cfq_group *cfqg;
144 struct cfq_group *orig_cfqg;
148 * First index in the service_trees.
149 * IDLE is handled separately, so it has negative index
158 * Second index in the service_trees.
162 SYNC_NOIDLE_WORKLOAD = 1,
166 /* This is per cgroup per device grouping structure */
168 /* group service_tree member */
169 struct rb_node rb_node;
171 /* group service_tree key */
176 /* number of cfqq currently on this group */
179 /* Per group busy queus average. Useful for workload slice calc. */
180 unsigned int busy_queues_avg[2];
182 * rr lists of queues with requests, onle rr for each priority class.
183 * Counts are embedded in the cfq_rb_root
185 struct cfq_rb_root service_trees[2][3];
186 struct cfq_rb_root service_tree_idle;
188 unsigned long saved_workload_slice;
189 enum wl_type_t saved_workload;
190 enum wl_prio_t saved_serving_prio;
191 struct blkio_group blkg;
192 #ifdef CONFIG_CFQ_GROUP_IOSCHED
193 struct hlist_node cfqd_node;
199 * Per block device queue structure
202 struct request_queue *queue;
203 /* Root service tree for cfq_groups */
204 struct cfq_rb_root grp_service_tree;
205 struct cfq_group root_group;
208 * The priority currently being served
210 enum wl_prio_t serving_prio;
211 enum wl_type_t serving_type;
212 unsigned long workload_expires;
213 struct cfq_group *serving_group;
214 bool noidle_tree_requires_idle;
217 * Each priority tree is sorted by next_request position. These
218 * trees are used when determining if two or more queues are
219 * interleaving requests (see cfq_close_cooperator).
221 struct rb_root prio_trees[CFQ_PRIO_LISTS];
223 unsigned int busy_queues;
229 * queue-depth detection
235 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
236 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
239 int hw_tag_est_depth;
240 unsigned int hw_tag_samples;
243 * idle window management
245 struct timer_list idle_slice_timer;
246 struct work_struct unplug_work;
248 struct cfq_queue *active_queue;
249 struct cfq_io_context *active_cic;
252 * async queue for each priority case
254 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
255 struct cfq_queue *async_idle_cfqq;
257 sector_t last_position;
260 * tunables, see top of file
262 unsigned int cfq_quantum;
263 unsigned int cfq_fifo_expire[2];
264 unsigned int cfq_back_penalty;
265 unsigned int cfq_back_max;
266 unsigned int cfq_slice[2];
267 unsigned int cfq_slice_async_rq;
268 unsigned int cfq_slice_idle;
269 unsigned int cfq_latency;
270 unsigned int cfq_group_isolation;
272 struct list_head cic_list;
275 * Fallback dummy cfqq for extreme OOM conditions
277 struct cfq_queue oom_cfqq;
279 unsigned long last_delayed_sync;
281 /* List of cfq groups being managed on this device*/
282 struct hlist_head cfqg_list;
286 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
288 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
295 if (prio == IDLE_WORKLOAD)
296 return &cfqg->service_tree_idle;
298 return &cfqg->service_trees[prio][type];
301 enum cfqq_state_flags {
302 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
303 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
304 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
305 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
306 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
307 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
308 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
309 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
310 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
311 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
312 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
313 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
314 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
317 #define CFQ_CFQQ_FNS(name) \
318 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
320 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
322 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
324 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
326 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
328 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
332 CFQ_CFQQ_FNS(wait_request);
333 CFQ_CFQQ_FNS(must_dispatch);
334 CFQ_CFQQ_FNS(must_alloc_slice);
335 CFQ_CFQQ_FNS(fifo_expire);
336 CFQ_CFQQ_FNS(idle_window);
337 CFQ_CFQQ_FNS(prio_changed);
338 CFQ_CFQQ_FNS(slice_new);
341 CFQ_CFQQ_FNS(split_coop);
343 CFQ_CFQQ_FNS(wait_busy);
346 #ifdef CONFIG_DEBUG_CFQ_IOSCHED
347 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
348 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
349 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
350 blkg_path(&(cfqq)->cfqg->blkg), ##args);
352 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
353 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
354 blkg_path(&(cfqg)->blkg), ##args); \
357 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
358 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
359 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
361 #define cfq_log(cfqd, fmt, args...) \
362 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
364 /* Traverses through cfq group service trees */
365 #define for_each_cfqg_st(cfqg, i, j, st) \
366 for (i = 0; i <= IDLE_WORKLOAD; i++) \
367 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
368 : &cfqg->service_tree_idle; \
369 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
370 (i == IDLE_WORKLOAD && j == 0); \
371 j++, st = i < IDLE_WORKLOAD ? \
372 &cfqg->service_trees[i][j]: NULL) \
375 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
377 if (cfq_class_idle(cfqq))
378 return IDLE_WORKLOAD;
379 if (cfq_class_rt(cfqq))
385 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
387 if (!cfq_cfqq_sync(cfqq))
388 return ASYNC_WORKLOAD;
389 if (!cfq_cfqq_idle_window(cfqq))
390 return SYNC_NOIDLE_WORKLOAD;
391 return SYNC_WORKLOAD;
394 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
395 struct cfq_data *cfqd,
396 struct cfq_group *cfqg)
398 if (wl == IDLE_WORKLOAD)
399 return cfqg->service_tree_idle.count;
401 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
402 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
403 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
406 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
407 struct cfq_group *cfqg)
409 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
410 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
413 static void cfq_dispatch_insert(struct request_queue *, struct request *);
414 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
415 struct io_context *, gfp_t);
416 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
417 struct io_context *);
419 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
422 return cic->cfqq[is_sync];
425 static inline void cic_set_cfqq(struct cfq_io_context *cic,
426 struct cfq_queue *cfqq, bool is_sync)
428 cic->cfqq[is_sync] = cfqq;
432 * We regard a request as SYNC, if it's either a read or has the SYNC bit
433 * set (in which case it could also be direct WRITE).
435 static inline bool cfq_bio_sync(struct bio *bio)
437 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
441 * scheduler run of queue, if there are requests pending and no one in the
442 * driver that will restart queueing
444 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
446 if (cfqd->busy_queues) {
447 cfq_log(cfqd, "schedule dispatch");
448 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
452 static int cfq_queue_empty(struct request_queue *q)
454 struct cfq_data *cfqd = q->elevator->elevator_data;
456 return !cfqd->rq_queued;
460 * Scale schedule slice based on io priority. Use the sync time slice only
461 * if a queue is marked sync and has sync io queued. A sync queue with async
462 * io only, should not get full sync slice length.
464 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
467 const int base_slice = cfqd->cfq_slice[sync];
469 WARN_ON(prio >= IOPRIO_BE_NR);
471 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
475 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
477 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
480 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
482 u64 d = delta << CFQ_SERVICE_SHIFT;
484 d = d * BLKIO_WEIGHT_DEFAULT;
485 do_div(d, cfqg->weight);
489 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
491 s64 delta = (s64)(vdisktime - min_vdisktime);
493 min_vdisktime = vdisktime;
495 return min_vdisktime;
498 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
500 s64 delta = (s64)(vdisktime - min_vdisktime);
502 min_vdisktime = vdisktime;
504 return min_vdisktime;
507 static void update_min_vdisktime(struct cfq_rb_root *st)
509 u64 vdisktime = st->min_vdisktime;
510 struct cfq_group *cfqg;
513 cfqg = rb_entry_cfqg(st->active);
514 vdisktime = cfqg->vdisktime;
518 cfqg = rb_entry_cfqg(st->left);
519 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
522 st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
526 * get averaged number of queues of RT/BE priority.
527 * average is updated, with a formula that gives more weight to higher numbers,
528 * to quickly follows sudden increases and decrease slowly
531 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
532 struct cfq_group *cfqg, bool rt)
534 unsigned min_q, max_q;
535 unsigned mult = cfq_hist_divisor - 1;
536 unsigned round = cfq_hist_divisor / 2;
537 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
539 min_q = min(cfqg->busy_queues_avg[rt], busy);
540 max_q = max(cfqg->busy_queues_avg[rt], busy);
541 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
543 return cfqg->busy_queues_avg[rt];
546 static inline unsigned
547 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
549 struct cfq_rb_root *st = &cfqd->grp_service_tree;
551 return cfq_target_latency * cfqg->weight / st->total_weight;
555 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
557 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
558 if (cfqd->cfq_latency) {
560 * interested queues (we consider only the ones with the same
561 * priority class in the cfq group)
563 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
565 unsigned sync_slice = cfqd->cfq_slice[1];
566 unsigned expect_latency = sync_slice * iq;
567 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
569 if (expect_latency > group_slice) {
570 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
571 /* scale low_slice according to IO priority
572 * and sync vs async */
574 min(slice, base_low_slice * slice / sync_slice);
575 /* the adapted slice value is scaled to fit all iqs
576 * into the target latency */
577 slice = max(slice * group_slice / expect_latency,
581 cfqq->slice_start = jiffies;
582 cfqq->slice_end = jiffies + slice;
583 cfqq->allocated_slice = slice;
584 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
588 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
589 * isn't valid until the first request from the dispatch is activated
590 * and the slice time set.
592 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
594 if (cfq_cfqq_slice_new(cfqq))
596 if (time_before(jiffies, cfqq->slice_end))
603 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
604 * We choose the request that is closest to the head right now. Distance
605 * behind the head is penalized and only allowed to a certain extent.
607 static struct request *
608 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
610 sector_t s1, s2, d1 = 0, d2 = 0;
611 unsigned long back_max;
612 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
613 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
614 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
616 if (rq1 == NULL || rq1 == rq2)
621 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
623 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
625 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
627 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
630 s1 = blk_rq_pos(rq1);
631 s2 = blk_rq_pos(rq2);
634 * by definition, 1KiB is 2 sectors
636 back_max = cfqd->cfq_back_max * 2;
639 * Strict one way elevator _except_ in the case where we allow
640 * short backward seeks which are biased as twice the cost of a
641 * similar forward seek.
645 else if (s1 + back_max >= last)
646 d1 = (last - s1) * cfqd->cfq_back_penalty;
648 wrap |= CFQ_RQ1_WRAP;
652 else if (s2 + back_max >= last)
653 d2 = (last - s2) * cfqd->cfq_back_penalty;
655 wrap |= CFQ_RQ2_WRAP;
657 /* Found required data */
660 * By doing switch() on the bit mask "wrap" we avoid having to
661 * check two variables for all permutations: --> faster!
664 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
680 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
683 * Since both rqs are wrapped,
684 * start with the one that's further behind head
685 * (--> only *one* back seek required),
686 * since back seek takes more time than forward.
696 * The below is leftmost cache rbtree addon
698 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
700 /* Service tree is empty */
705 root->left = rb_first(&root->rb);
708 return rb_entry(root->left, struct cfq_queue, rb_node);
713 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
716 root->left = rb_first(&root->rb);
719 return rb_entry_cfqg(root->left);
724 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
730 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
734 rb_erase_init(n, &root->rb);
739 * would be nice to take fifo expire time into account as well
741 static struct request *
742 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
743 struct request *last)
745 struct rb_node *rbnext = rb_next(&last->rb_node);
746 struct rb_node *rbprev = rb_prev(&last->rb_node);
747 struct request *next = NULL, *prev = NULL;
749 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
752 prev = rb_entry_rq(rbprev);
755 next = rb_entry_rq(rbnext);
757 rbnext = rb_first(&cfqq->sort_list);
758 if (rbnext && rbnext != &last->rb_node)
759 next = rb_entry_rq(rbnext);
762 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
765 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
766 struct cfq_queue *cfqq)
769 * just an approximation, should be ok.
771 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
772 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
776 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
778 return cfqg->vdisktime - st->min_vdisktime;
782 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
784 struct rb_node **node = &st->rb.rb_node;
785 struct rb_node *parent = NULL;
786 struct cfq_group *__cfqg;
787 s64 key = cfqg_key(st, cfqg);
790 while (*node != NULL) {
792 __cfqg = rb_entry_cfqg(parent);
794 if (key < cfqg_key(st, __cfqg))
795 node = &parent->rb_left;
797 node = &parent->rb_right;
803 st->left = &cfqg->rb_node;
805 rb_link_node(&cfqg->rb_node, parent, node);
806 rb_insert_color(&cfqg->rb_node, &st->rb);
810 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
812 struct cfq_rb_root *st = &cfqd->grp_service_tree;
813 struct cfq_group *__cfqg;
821 * Currently put the group at the end. Later implement something
822 * so that groups get lesser vtime based on their weights, so that
823 * if group does not loose all if it was not continously backlogged.
825 n = rb_last(&st->rb);
827 __cfqg = rb_entry_cfqg(n);
828 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
830 cfqg->vdisktime = st->min_vdisktime;
832 __cfq_group_service_tree_add(st, cfqg);
834 st->total_weight += cfqg->weight;
838 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
840 struct cfq_rb_root *st = &cfqd->grp_service_tree;
842 if (st->active == &cfqg->rb_node)
845 BUG_ON(cfqg->nr_cfqq < 1);
848 /* If there are other cfq queues under this group, don't delete it */
852 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
854 st->total_weight -= cfqg->weight;
855 if (!RB_EMPTY_NODE(&cfqg->rb_node))
856 cfq_rb_erase(&cfqg->rb_node, st);
857 cfqg->saved_workload_slice = 0;
858 blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
861 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
863 unsigned int slice_used;
866 * Queue got expired before even a single request completed or
867 * got expired immediately after first request completion.
869 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
871 * Also charge the seek time incurred to the group, otherwise
872 * if there are mutiple queues in the group, each can dispatch
873 * a single request on seeky media and cause lots of seek time
874 * and group will never know it.
876 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
879 slice_used = jiffies - cfqq->slice_start;
880 if (slice_used > cfqq->allocated_slice)
881 slice_used = cfqq->allocated_slice;
884 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
888 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
889 struct cfq_queue *cfqq)
891 struct cfq_rb_root *st = &cfqd->grp_service_tree;
892 unsigned int used_sl, charge_sl;
893 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
894 - cfqg->service_tree_idle.count;
897 used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
899 if (!cfq_cfqq_sync(cfqq) && !nr_sync)
900 charge_sl = cfqq->allocated_slice;
902 /* Can't update vdisktime while group is on service tree */
903 cfq_rb_erase(&cfqg->rb_node, st);
904 cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
905 __cfq_group_service_tree_add(st, cfqg);
907 /* This group is being expired. Save the context */
908 if (time_after(cfqd->workload_expires, jiffies)) {
909 cfqg->saved_workload_slice = cfqd->workload_expires
911 cfqg->saved_workload = cfqd->serving_type;
912 cfqg->saved_serving_prio = cfqd->serving_prio;
914 cfqg->saved_workload_slice = 0;
916 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
918 blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
921 #ifdef CONFIG_CFQ_GROUP_IOSCHED
922 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
925 return container_of(blkg, struct cfq_group, blkg);
930 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
932 cfqg_of_blkg(blkg)->weight = weight;
935 static struct cfq_group *
936 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
938 struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
939 struct cfq_group *cfqg = NULL;
942 struct cfq_rb_root *st;
943 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
944 unsigned int major, minor;
946 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
950 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
954 cfqg->weight = blkcg->weight;
955 for_each_cfqg_st(cfqg, i, j, st)
957 RB_CLEAR_NODE(&cfqg->rb_node);
958 blkio_group_init(&cfqg->blkg);
961 * Take the initial reference that will be released on destroy
962 * This can be thought of a joint reference by cgroup and
963 * elevator which will be dropped by either elevator exit
964 * or cgroup deletion path depending on who is exiting first.
966 atomic_set(&cfqg->ref, 1);
968 /* Add group onto cgroup list */
969 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
970 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
971 MKDEV(major, minor));
973 /* Add group on cfqd list */
974 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
981 * Search for the cfq group current task belongs to. If create = 1, then also
982 * create the cfq group if it does not exist. request_queue lock must be held.
984 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
986 struct cgroup *cgroup;
987 struct cfq_group *cfqg = NULL;
990 cgroup = task_cgroup(current, blkio_subsys_id);
991 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
993 cfqg = &cfqd->root_group;
998 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1000 /* Currently, all async queues are mapped to root group */
1001 if (!cfq_cfqq_sync(cfqq))
1002 cfqg = &cfqq->cfqd->root_group;
1005 /* cfqq reference on cfqg */
1006 atomic_inc(&cfqq->cfqg->ref);
1009 static void cfq_put_cfqg(struct cfq_group *cfqg)
1011 struct cfq_rb_root *st;
1014 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1015 if (!atomic_dec_and_test(&cfqg->ref))
1017 for_each_cfqg_st(cfqg, i, j, st)
1018 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1022 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1024 /* Something wrong if we are trying to remove same group twice */
1025 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1027 hlist_del_init(&cfqg->cfqd_node);
1030 * Put the reference taken at the time of creation so that when all
1031 * queues are gone, group can be destroyed.
1036 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1038 struct hlist_node *pos, *n;
1039 struct cfq_group *cfqg;
1041 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1043 * If cgroup removal path got to blk_group first and removed
1044 * it from cgroup list, then it will take care of destroying
1047 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1048 cfq_destroy_cfqg(cfqd, cfqg);
1053 * Blk cgroup controller notification saying that blkio_group object is being
1054 * delinked as associated cgroup object is going away. That also means that
1055 * no new IO will come in this group. So get rid of this group as soon as
1056 * any pending IO in the group is finished.
1058 * This function is called under rcu_read_lock(). key is the rcu protected
1059 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1062 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1063 * it should not be NULL as even if elevator was exiting, cgroup deltion
1064 * path got to it first.
1066 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1068 unsigned long flags;
1069 struct cfq_data *cfqd = key;
1071 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1072 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1073 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1076 #else /* GROUP_IOSCHED */
1077 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1079 return &cfqd->root_group;
1082 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1086 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1087 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1089 #endif /* GROUP_IOSCHED */
1092 * The cfqd->service_trees holds all pending cfq_queue's that have
1093 * requests waiting to be processed. It is sorted in the order that
1094 * we will service the queues.
1096 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1099 struct rb_node **p, *parent;
1100 struct cfq_queue *__cfqq;
1101 unsigned long rb_key;
1102 struct cfq_rb_root *service_tree;
1105 int group_changed = 0;
1107 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1108 if (!cfqd->cfq_group_isolation
1109 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1110 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1111 /* Move this cfq to root group */
1112 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1113 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1114 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1115 cfqq->orig_cfqg = cfqq->cfqg;
1116 cfqq->cfqg = &cfqd->root_group;
1117 atomic_inc(&cfqd->root_group.ref);
1119 } else if (!cfqd->cfq_group_isolation
1120 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1121 /* cfqq is sequential now needs to go to its original group */
1122 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1123 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1124 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1125 cfq_put_cfqg(cfqq->cfqg);
1126 cfqq->cfqg = cfqq->orig_cfqg;
1127 cfqq->orig_cfqg = NULL;
1129 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1133 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1135 if (cfq_class_idle(cfqq)) {
1136 rb_key = CFQ_IDLE_DELAY;
1137 parent = rb_last(&service_tree->rb);
1138 if (parent && parent != &cfqq->rb_node) {
1139 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1140 rb_key += __cfqq->rb_key;
1143 } else if (!add_front) {
1145 * Get our rb key offset. Subtract any residual slice
1146 * value carried from last service. A negative resid
1147 * count indicates slice overrun, and this should position
1148 * the next service time further away in the tree.
1150 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1151 rb_key -= cfqq->slice_resid;
1152 cfqq->slice_resid = 0;
1155 __cfqq = cfq_rb_first(service_tree);
1156 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1159 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1162 * same position, nothing more to do
1164 if (rb_key == cfqq->rb_key &&
1165 cfqq->service_tree == service_tree)
1168 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1169 cfqq->service_tree = NULL;
1174 cfqq->service_tree = service_tree;
1175 p = &service_tree->rb.rb_node;
1180 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1183 * sort by key, that represents service time.
1185 if (time_before(rb_key, __cfqq->rb_key))
1188 n = &(*p)->rb_right;
1196 service_tree->left = &cfqq->rb_node;
1198 cfqq->rb_key = rb_key;
1199 rb_link_node(&cfqq->rb_node, parent, p);
1200 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1201 service_tree->count++;
1202 if ((add_front || !new_cfqq) && !group_changed)
1204 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1207 static struct cfq_queue *
1208 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1209 sector_t sector, struct rb_node **ret_parent,
1210 struct rb_node ***rb_link)
1212 struct rb_node **p, *parent;
1213 struct cfq_queue *cfqq = NULL;
1221 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1224 * Sort strictly based on sector. Smallest to the left,
1225 * largest to the right.
1227 if (sector > blk_rq_pos(cfqq->next_rq))
1228 n = &(*p)->rb_right;
1229 else if (sector < blk_rq_pos(cfqq->next_rq))
1237 *ret_parent = parent;
1243 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1245 struct rb_node **p, *parent;
1246 struct cfq_queue *__cfqq;
1249 rb_erase(&cfqq->p_node, cfqq->p_root);
1250 cfqq->p_root = NULL;
1253 if (cfq_class_idle(cfqq))
1258 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1259 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1260 blk_rq_pos(cfqq->next_rq), &parent, &p);
1262 rb_link_node(&cfqq->p_node, parent, p);
1263 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1265 cfqq->p_root = NULL;
1269 * Update cfqq's position in the service tree.
1271 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1274 * Resorting requires the cfqq to be on the RR list already.
1276 if (cfq_cfqq_on_rr(cfqq)) {
1277 cfq_service_tree_add(cfqd, cfqq, 0);
1278 cfq_prio_tree_add(cfqd, cfqq);
1283 * add to busy list of queues for service, trying to be fair in ordering
1284 * the pending list according to last request service
1286 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1288 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1289 BUG_ON(cfq_cfqq_on_rr(cfqq));
1290 cfq_mark_cfqq_on_rr(cfqq);
1291 cfqd->busy_queues++;
1293 cfq_resort_rr_list(cfqd, cfqq);
1297 * Called when the cfqq no longer has requests pending, remove it from
1300 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1302 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1303 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1304 cfq_clear_cfqq_on_rr(cfqq);
1306 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1307 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1308 cfqq->service_tree = NULL;
1311 rb_erase(&cfqq->p_node, cfqq->p_root);
1312 cfqq->p_root = NULL;
1315 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1316 BUG_ON(!cfqd->busy_queues);
1317 cfqd->busy_queues--;
1321 * rb tree support functions
1323 static void cfq_del_rq_rb(struct request *rq)
1325 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1326 const int sync = rq_is_sync(rq);
1328 BUG_ON(!cfqq->queued[sync]);
1329 cfqq->queued[sync]--;
1331 elv_rb_del(&cfqq->sort_list, rq);
1333 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1335 * Queue will be deleted from service tree when we actually
1336 * expire it later. Right now just remove it from prio tree
1340 rb_erase(&cfqq->p_node, cfqq->p_root);
1341 cfqq->p_root = NULL;
1346 static void cfq_add_rq_rb(struct request *rq)
1348 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1349 struct cfq_data *cfqd = cfqq->cfqd;
1350 struct request *__alias, *prev;
1352 cfqq->queued[rq_is_sync(rq)]++;
1355 * looks a little odd, but the first insert might return an alias.
1356 * if that happens, put the alias on the dispatch list
1358 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1359 cfq_dispatch_insert(cfqd->queue, __alias);
1361 if (!cfq_cfqq_on_rr(cfqq))
1362 cfq_add_cfqq_rr(cfqd, cfqq);
1365 * check if this request is a better next-serve candidate
1367 prev = cfqq->next_rq;
1368 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1371 * adjust priority tree position, if ->next_rq changes
1373 if (prev != cfqq->next_rq)
1374 cfq_prio_tree_add(cfqd, cfqq);
1376 BUG_ON(!cfqq->next_rq);
1379 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1381 elv_rb_del(&cfqq->sort_list, rq);
1382 cfqq->queued[rq_is_sync(rq)]--;
1386 static struct request *
1387 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1389 struct task_struct *tsk = current;
1390 struct cfq_io_context *cic;
1391 struct cfq_queue *cfqq;
1393 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1397 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1399 sector_t sector = bio->bi_sector + bio_sectors(bio);
1401 return elv_rb_find(&cfqq->sort_list, sector);
1407 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1409 struct cfq_data *cfqd = q->elevator->elevator_data;
1411 cfqd->rq_in_driver++;
1412 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1413 cfqd->rq_in_driver);
1415 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1418 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1420 struct cfq_data *cfqd = q->elevator->elevator_data;
1422 WARN_ON(!cfqd->rq_in_driver);
1423 cfqd->rq_in_driver--;
1424 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1425 cfqd->rq_in_driver);
1428 static void cfq_remove_request(struct request *rq)
1430 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1432 if (cfqq->next_rq == rq)
1433 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1435 list_del_init(&rq->queuelist);
1438 cfqq->cfqd->rq_queued--;
1439 if (rq_is_meta(rq)) {
1440 WARN_ON(!cfqq->meta_pending);
1441 cfqq->meta_pending--;
1445 static int cfq_merge(struct request_queue *q, struct request **req,
1448 struct cfq_data *cfqd = q->elevator->elevator_data;
1449 struct request *__rq;
1451 __rq = cfq_find_rq_fmerge(cfqd, bio);
1452 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1454 return ELEVATOR_FRONT_MERGE;
1457 return ELEVATOR_NO_MERGE;
1460 static void cfq_merged_request(struct request_queue *q, struct request *req,
1463 if (type == ELEVATOR_FRONT_MERGE) {
1464 struct cfq_queue *cfqq = RQ_CFQQ(req);
1466 cfq_reposition_rq_rb(cfqq, req);
1470 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1473 struct cfq_queue *cfqq = RQ_CFQQ(req);
1474 blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, bio_data_dir(bio),
1479 cfq_merged_requests(struct request_queue *q, struct request *rq,
1480 struct request *next)
1482 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1484 * reposition in fifo if next is older than rq
1486 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1487 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1488 list_move(&rq->queuelist, &next->queuelist);
1489 rq_set_fifo_time(rq, rq_fifo_time(next));
1492 if (cfqq->next_rq == next)
1494 cfq_remove_request(next);
1495 blkiocg_update_io_merged_stats(&cfqq->cfqg->blkg, rq_data_dir(next),
1499 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1502 struct cfq_data *cfqd = q->elevator->elevator_data;
1503 struct cfq_io_context *cic;
1504 struct cfq_queue *cfqq;
1507 * Disallow merge of a sync bio into an async request.
1509 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1513 * Lookup the cfqq that this bio will be queued with. Allow
1514 * merge only if rq is queued there.
1516 cic = cfq_cic_lookup(cfqd, current->io_context);
1520 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1521 return cfqq == RQ_CFQQ(rq);
1524 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1525 struct cfq_queue *cfqq)
1528 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1529 cfqd->serving_prio, cfqd->serving_type);
1530 cfqq->slice_start = 0;
1531 cfqq->dispatch_start = jiffies;
1532 cfqq->allocated_slice = 0;
1533 cfqq->slice_end = 0;
1534 cfqq->slice_dispatch = 0;
1536 cfq_clear_cfqq_wait_request(cfqq);
1537 cfq_clear_cfqq_must_dispatch(cfqq);
1538 cfq_clear_cfqq_must_alloc_slice(cfqq);
1539 cfq_clear_cfqq_fifo_expire(cfqq);
1540 cfq_mark_cfqq_slice_new(cfqq);
1542 del_timer(&cfqd->idle_slice_timer);
1545 cfqd->active_queue = cfqq;
1549 * current cfqq expired its slice (or was too idle), select new one
1552 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1555 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1557 if (cfq_cfqq_wait_request(cfqq))
1558 del_timer(&cfqd->idle_slice_timer);
1560 cfq_clear_cfqq_wait_request(cfqq);
1561 cfq_clear_cfqq_wait_busy(cfqq);
1564 * If this cfqq is shared between multiple processes, check to
1565 * make sure that those processes are still issuing I/Os within
1566 * the mean seek distance. If not, it may be time to break the
1567 * queues apart again.
1569 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1570 cfq_mark_cfqq_split_coop(cfqq);
1573 * store what was left of this slice, if the queue idled/timed out
1575 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1576 cfqq->slice_resid = cfqq->slice_end - jiffies;
1577 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1580 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1582 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1583 cfq_del_cfqq_rr(cfqd, cfqq);
1585 cfq_resort_rr_list(cfqd, cfqq);
1587 if (cfqq == cfqd->active_queue)
1588 cfqd->active_queue = NULL;
1590 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1591 cfqd->grp_service_tree.active = NULL;
1593 if (cfqd->active_cic) {
1594 put_io_context(cfqd->active_cic->ioc);
1595 cfqd->active_cic = NULL;
1599 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1601 struct cfq_queue *cfqq = cfqd->active_queue;
1604 __cfq_slice_expired(cfqd, cfqq, timed_out);
1608 * Get next queue for service. Unless we have a queue preemption,
1609 * we'll simply select the first cfqq in the service tree.
1611 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1613 struct cfq_rb_root *service_tree =
1614 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1615 cfqd->serving_type);
1617 if (!cfqd->rq_queued)
1620 /* There is nothing to dispatch */
1623 if (RB_EMPTY_ROOT(&service_tree->rb))
1625 return cfq_rb_first(service_tree);
1628 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1630 struct cfq_group *cfqg;
1631 struct cfq_queue *cfqq;
1633 struct cfq_rb_root *st;
1635 if (!cfqd->rq_queued)
1638 cfqg = cfq_get_next_cfqg(cfqd);
1642 for_each_cfqg_st(cfqg, i, j, st)
1643 if ((cfqq = cfq_rb_first(st)) != NULL)
1649 * Get and set a new active queue for service.
1651 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1652 struct cfq_queue *cfqq)
1655 cfqq = cfq_get_next_queue(cfqd);
1657 __cfq_set_active_queue(cfqd, cfqq);
1661 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1664 if (blk_rq_pos(rq) >= cfqd->last_position)
1665 return blk_rq_pos(rq) - cfqd->last_position;
1667 return cfqd->last_position - blk_rq_pos(rq);
1670 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1673 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1676 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1677 struct cfq_queue *cur_cfqq)
1679 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1680 struct rb_node *parent, *node;
1681 struct cfq_queue *__cfqq;
1682 sector_t sector = cfqd->last_position;
1684 if (RB_EMPTY_ROOT(root))
1688 * First, if we find a request starting at the end of the last
1689 * request, choose it.
1691 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1696 * If the exact sector wasn't found, the parent of the NULL leaf
1697 * will contain the closest sector.
1699 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1700 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1703 if (blk_rq_pos(__cfqq->next_rq) < sector)
1704 node = rb_next(&__cfqq->p_node);
1706 node = rb_prev(&__cfqq->p_node);
1710 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1711 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1719 * cur_cfqq - passed in so that we don't decide that the current queue is
1720 * closely cooperating with itself.
1722 * So, basically we're assuming that that cur_cfqq has dispatched at least
1723 * one request, and that cfqd->last_position reflects a position on the disk
1724 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1727 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1728 struct cfq_queue *cur_cfqq)
1730 struct cfq_queue *cfqq;
1732 if (cfq_class_idle(cur_cfqq))
1734 if (!cfq_cfqq_sync(cur_cfqq))
1736 if (CFQQ_SEEKY(cur_cfqq))
1740 * Don't search priority tree if it's the only queue in the group.
1742 if (cur_cfqq->cfqg->nr_cfqq == 1)
1746 * We should notice if some of the queues are cooperating, eg
1747 * working closely on the same area of the disk. In that case,
1748 * we can group them together and don't waste time idling.
1750 cfqq = cfqq_close(cfqd, cur_cfqq);
1754 /* If new queue belongs to different cfq_group, don't choose it */
1755 if (cur_cfqq->cfqg != cfqq->cfqg)
1759 * It only makes sense to merge sync queues.
1761 if (!cfq_cfqq_sync(cfqq))
1763 if (CFQQ_SEEKY(cfqq))
1767 * Do not merge queues of different priority classes
1769 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1776 * Determine whether we should enforce idle window for this queue.
1779 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1781 enum wl_prio_t prio = cfqq_prio(cfqq);
1782 struct cfq_rb_root *service_tree = cfqq->service_tree;
1784 BUG_ON(!service_tree);
1785 BUG_ON(!service_tree->count);
1787 /* We never do for idle class queues. */
1788 if (prio == IDLE_WORKLOAD)
1791 /* We do for queues that were marked with idle window flag. */
1792 if (cfq_cfqq_idle_window(cfqq) &&
1793 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1797 * Otherwise, we do only if they are the last ones
1798 * in their service tree.
1800 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1802 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1803 service_tree->count);
1807 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1809 struct cfq_queue *cfqq = cfqd->active_queue;
1810 struct cfq_io_context *cic;
1814 * SSD device without seek penalty, disable idling. But only do so
1815 * for devices that support queuing, otherwise we still have a problem
1816 * with sync vs async workloads.
1818 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1821 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1822 WARN_ON(cfq_cfqq_slice_new(cfqq));
1825 * idle is disabled, either manually or by past process history
1827 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1831 * still active requests from this queue, don't idle
1833 if (cfqq->dispatched)
1837 * task has exited, don't wait
1839 cic = cfqd->active_cic;
1840 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1844 * If our average think time is larger than the remaining time
1845 * slice, then don't idle. This avoids overrunning the allotted
1848 if (sample_valid(cic->ttime_samples) &&
1849 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1850 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1855 cfq_mark_cfqq_wait_request(cfqq);
1857 sl = cfqd->cfq_slice_idle;
1859 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1860 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1864 * Move request from internal lists to the request queue dispatch list.
1866 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1868 struct cfq_data *cfqd = q->elevator->elevator_data;
1869 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1871 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1873 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1874 cfq_remove_request(rq);
1876 elv_dispatch_sort(q, rq);
1878 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1879 blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1880 rq_data_dir(rq), rq_is_sync(rq));
1884 * return expired entry, or NULL to just start from scratch in rbtree
1886 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1888 struct request *rq = NULL;
1890 if (cfq_cfqq_fifo_expire(cfqq))
1893 cfq_mark_cfqq_fifo_expire(cfqq);
1895 if (list_empty(&cfqq->fifo))
1898 rq = rq_entry_fifo(cfqq->fifo.next);
1899 if (time_before(jiffies, rq_fifo_time(rq)))
1902 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1907 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1909 const int base_rq = cfqd->cfq_slice_async_rq;
1911 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1913 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1917 * Must be called with the queue_lock held.
1919 static int cfqq_process_refs(struct cfq_queue *cfqq)
1921 int process_refs, io_refs;
1923 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1924 process_refs = atomic_read(&cfqq->ref) - io_refs;
1925 BUG_ON(process_refs < 0);
1926 return process_refs;
1929 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1931 int process_refs, new_process_refs;
1932 struct cfq_queue *__cfqq;
1934 /* Avoid a circular list and skip interim queue merges */
1935 while ((__cfqq = new_cfqq->new_cfqq)) {
1941 process_refs = cfqq_process_refs(cfqq);
1943 * If the process for the cfqq has gone away, there is no
1944 * sense in merging the queues.
1946 if (process_refs == 0)
1950 * Merge in the direction of the lesser amount of work.
1952 new_process_refs = cfqq_process_refs(new_cfqq);
1953 if (new_process_refs >= process_refs) {
1954 cfqq->new_cfqq = new_cfqq;
1955 atomic_add(process_refs, &new_cfqq->ref);
1957 new_cfqq->new_cfqq = cfqq;
1958 atomic_add(new_process_refs, &cfqq->ref);
1962 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1963 struct cfq_group *cfqg, enum wl_prio_t prio)
1965 struct cfq_queue *queue;
1967 bool key_valid = false;
1968 unsigned long lowest_key = 0;
1969 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1971 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1972 /* select the one with lowest rb_key */
1973 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1975 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1976 lowest_key = queue->rb_key;
1985 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1989 struct cfq_rb_root *st;
1990 unsigned group_slice;
1993 cfqd->serving_prio = IDLE_WORKLOAD;
1994 cfqd->workload_expires = jiffies + 1;
1998 /* Choose next priority. RT > BE > IDLE */
1999 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2000 cfqd->serving_prio = RT_WORKLOAD;
2001 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2002 cfqd->serving_prio = BE_WORKLOAD;
2004 cfqd->serving_prio = IDLE_WORKLOAD;
2005 cfqd->workload_expires = jiffies + 1;
2010 * For RT and BE, we have to choose also the type
2011 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2014 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2018 * check workload expiration, and that we still have other queues ready
2020 if (count && !time_after(jiffies, cfqd->workload_expires))
2023 /* otherwise select new workload type */
2024 cfqd->serving_type =
2025 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2026 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2030 * the workload slice is computed as a fraction of target latency
2031 * proportional to the number of queues in that workload, over
2032 * all the queues in the same priority class
2034 group_slice = cfq_group_slice(cfqd, cfqg);
2036 slice = group_slice * count /
2037 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2038 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2040 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2044 * Async queues are currently system wide. Just taking
2045 * proportion of queues with-in same group will lead to higher
2046 * async ratio system wide as generally root group is going
2047 * to have higher weight. A more accurate thing would be to
2048 * calculate system wide asnc/sync ratio.
2050 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2051 tmp = tmp/cfqd->busy_queues;
2052 slice = min_t(unsigned, slice, tmp);
2054 /* async workload slice is scaled down according to
2055 * the sync/async slice ratio. */
2056 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2058 /* sync workload slice is at least 2 * cfq_slice_idle */
2059 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2061 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2062 cfq_log(cfqd, "workload slice:%d", slice);
2063 cfqd->workload_expires = jiffies + slice;
2064 cfqd->noidle_tree_requires_idle = false;
2067 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2069 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2070 struct cfq_group *cfqg;
2072 if (RB_EMPTY_ROOT(&st->rb))
2074 cfqg = cfq_rb_first_group(st);
2075 st->active = &cfqg->rb_node;
2076 update_min_vdisktime(st);
2080 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2082 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2084 cfqd->serving_group = cfqg;
2086 /* Restore the workload type data */
2087 if (cfqg->saved_workload_slice) {
2088 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2089 cfqd->serving_type = cfqg->saved_workload;
2090 cfqd->serving_prio = cfqg->saved_serving_prio;
2092 cfqd->workload_expires = jiffies - 1;
2094 choose_service_tree(cfqd, cfqg);
2098 * Select a queue for service. If we have a current active queue,
2099 * check whether to continue servicing it, or retrieve and set a new one.
2101 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2103 struct cfq_queue *cfqq, *new_cfqq = NULL;
2105 cfqq = cfqd->active_queue;
2109 if (!cfqd->rq_queued)
2113 * We were waiting for group to get backlogged. Expire the queue
2115 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2119 * The active queue has run out of time, expire it and select new.
2121 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2123 * If slice had not expired at the completion of last request
2124 * we might not have turned on wait_busy flag. Don't expire
2125 * the queue yet. Allow the group to get backlogged.
2127 * The very fact that we have used the slice, that means we
2128 * have been idling all along on this queue and it should be
2129 * ok to wait for this request to complete.
2131 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2132 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2140 * The active queue has requests and isn't expired, allow it to
2143 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2147 * If another queue has a request waiting within our mean seek
2148 * distance, let it run. The expire code will check for close
2149 * cooperators and put the close queue at the front of the service
2150 * tree. If possible, merge the expiring queue with the new cfqq.
2152 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2154 if (!cfqq->new_cfqq)
2155 cfq_setup_merge(cfqq, new_cfqq);
2160 * No requests pending. If the active queue still has requests in
2161 * flight or is idling for a new request, allow either of these
2162 * conditions to happen (or time out) before selecting a new queue.
2164 if (timer_pending(&cfqd->idle_slice_timer) ||
2165 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2171 cfq_slice_expired(cfqd, 0);
2174 * Current queue expired. Check if we have to switch to a new
2178 cfq_choose_cfqg(cfqd);
2180 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2185 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2189 while (cfqq->next_rq) {
2190 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2194 BUG_ON(!list_empty(&cfqq->fifo));
2196 /* By default cfqq is not expired if it is empty. Do it explicitly */
2197 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2202 * Drain our current requests. Used for barriers and when switching
2203 * io schedulers on-the-fly.
2205 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2207 struct cfq_queue *cfqq;
2210 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2211 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2213 cfq_slice_expired(cfqd, 0);
2214 BUG_ON(cfqd->busy_queues);
2216 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2220 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2221 struct cfq_queue *cfqq)
2223 /* the queue hasn't finished any request, can't estimate */
2224 if (cfq_cfqq_slice_new(cfqq))
2226 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2233 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2235 unsigned int max_dispatch;
2238 * Drain async requests before we start sync IO
2240 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2244 * If this is an async queue and we have sync IO in flight, let it wait
2246 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2249 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2250 if (cfq_class_idle(cfqq))
2254 * Does this cfqq already have too much IO in flight?
2256 if (cfqq->dispatched >= max_dispatch) {
2258 * idle queue must always only have a single IO in flight
2260 if (cfq_class_idle(cfqq))
2264 * We have other queues, don't allow more IO from this one
2266 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2270 * Sole queue user, no limit
2272 if (cfqd->busy_queues == 1)
2276 * Normally we start throttling cfqq when cfq_quantum/2
2277 * requests have been dispatched. But we can drive
2278 * deeper queue depths at the beginning of slice
2279 * subjected to upper limit of cfq_quantum.
2281 max_dispatch = cfqd->cfq_quantum;
2285 * Async queues must wait a bit before being allowed dispatch.
2286 * We also ramp up the dispatch depth gradually for async IO,
2287 * based on the last sync IO we serviced
2289 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2290 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2293 depth = last_sync / cfqd->cfq_slice[1];
2294 if (!depth && !cfqq->dispatched)
2296 if (depth < max_dispatch)
2297 max_dispatch = depth;
2301 * If we're below the current max, allow a dispatch
2303 return cfqq->dispatched < max_dispatch;
2307 * Dispatch a request from cfqq, moving them to the request queue
2310 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2314 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2316 if (!cfq_may_dispatch(cfqd, cfqq))
2320 * follow expired path, else get first next available
2322 rq = cfq_check_fifo(cfqq);
2327 * insert request into driver dispatch list
2329 cfq_dispatch_insert(cfqd->queue, rq);
2331 if (!cfqd->active_cic) {
2332 struct cfq_io_context *cic = RQ_CIC(rq);
2334 atomic_long_inc(&cic->ioc->refcount);
2335 cfqd->active_cic = cic;
2342 * Find the cfqq that we need to service and move a request from that to the
2345 static int cfq_dispatch_requests(struct request_queue *q, int force)
2347 struct cfq_data *cfqd = q->elevator->elevator_data;
2348 struct cfq_queue *cfqq;
2350 if (!cfqd->busy_queues)
2353 if (unlikely(force))
2354 return cfq_forced_dispatch(cfqd);
2356 cfqq = cfq_select_queue(cfqd);
2361 * Dispatch a request from this cfqq, if it is allowed
2363 if (!cfq_dispatch_request(cfqd, cfqq))
2366 cfqq->slice_dispatch++;
2367 cfq_clear_cfqq_must_dispatch(cfqq);
2370 * expire an async queue immediately if it has used up its slice. idle
2371 * queue always expire after 1 dispatch round.
2373 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2374 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2375 cfq_class_idle(cfqq))) {
2376 cfqq->slice_end = jiffies + 1;
2377 cfq_slice_expired(cfqd, 0);
2380 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2385 * task holds one reference to the queue, dropped when task exits. each rq
2386 * in-flight on this queue also holds a reference, dropped when rq is freed.
2388 * Each cfq queue took a reference on the parent group. Drop it now.
2389 * queue lock must be held here.
2391 static void cfq_put_queue(struct cfq_queue *cfqq)
2393 struct cfq_data *cfqd = cfqq->cfqd;
2394 struct cfq_group *cfqg, *orig_cfqg;
2396 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2398 if (!atomic_dec_and_test(&cfqq->ref))
2401 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2402 BUG_ON(rb_first(&cfqq->sort_list));
2403 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2405 orig_cfqg = cfqq->orig_cfqg;
2407 if (unlikely(cfqd->active_queue == cfqq)) {
2408 __cfq_slice_expired(cfqd, cfqq, 0);
2409 cfq_schedule_dispatch(cfqd);
2412 BUG_ON(cfq_cfqq_on_rr(cfqq));
2413 kmem_cache_free(cfq_pool, cfqq);
2416 cfq_put_cfqg(orig_cfqg);
2420 * Must always be called with the rcu_read_lock() held
2423 __call_for_each_cic(struct io_context *ioc,
2424 void (*func)(struct io_context *, struct cfq_io_context *))
2426 struct cfq_io_context *cic;
2427 struct hlist_node *n;
2429 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2434 * Call func for each cic attached to this ioc.
2437 call_for_each_cic(struct io_context *ioc,
2438 void (*func)(struct io_context *, struct cfq_io_context *))
2441 __call_for_each_cic(ioc, func);
2445 static void cfq_cic_free_rcu(struct rcu_head *head)
2447 struct cfq_io_context *cic;
2449 cic = container_of(head, struct cfq_io_context, rcu_head);
2451 kmem_cache_free(cfq_ioc_pool, cic);
2452 elv_ioc_count_dec(cfq_ioc_count);
2456 * CFQ scheduler is exiting, grab exit lock and check
2457 * the pending io context count. If it hits zero,
2458 * complete ioc_gone and set it back to NULL
2460 spin_lock(&ioc_gone_lock);
2461 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2465 spin_unlock(&ioc_gone_lock);
2469 static void cfq_cic_free(struct cfq_io_context *cic)
2471 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2474 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2476 unsigned long flags;
2478 BUG_ON(!cic->dead_key);
2480 spin_lock_irqsave(&ioc->lock, flags);
2481 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2482 hlist_del_rcu(&cic->cic_list);
2483 spin_unlock_irqrestore(&ioc->lock, flags);
2489 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2490 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2491 * and ->trim() which is called with the task lock held
2493 static void cfq_free_io_context(struct io_context *ioc)
2496 * ioc->refcount is zero here, or we are called from elv_unregister(),
2497 * so no more cic's are allowed to be linked into this ioc. So it
2498 * should be ok to iterate over the known list, we will see all cic's
2499 * since no new ones are added.
2501 __call_for_each_cic(ioc, cic_free_func);
2504 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2506 struct cfq_queue *__cfqq, *next;
2508 if (unlikely(cfqq == cfqd->active_queue)) {
2509 __cfq_slice_expired(cfqd, cfqq, 0);
2510 cfq_schedule_dispatch(cfqd);
2514 * If this queue was scheduled to merge with another queue, be
2515 * sure to drop the reference taken on that queue (and others in
2516 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2518 __cfqq = cfqq->new_cfqq;
2520 if (__cfqq == cfqq) {
2521 WARN(1, "cfqq->new_cfqq loop detected\n");
2524 next = __cfqq->new_cfqq;
2525 cfq_put_queue(__cfqq);
2529 cfq_put_queue(cfqq);
2532 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2533 struct cfq_io_context *cic)
2535 struct io_context *ioc = cic->ioc;
2537 list_del_init(&cic->queue_list);
2540 * Make sure key == NULL is seen for dead queues
2543 cic->dead_key = (unsigned long) cic->key;
2546 if (ioc->ioc_data == cic)
2547 rcu_assign_pointer(ioc->ioc_data, NULL);
2549 if (cic->cfqq[BLK_RW_ASYNC]) {
2550 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2551 cic->cfqq[BLK_RW_ASYNC] = NULL;
2554 if (cic->cfqq[BLK_RW_SYNC]) {
2555 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2556 cic->cfqq[BLK_RW_SYNC] = NULL;
2560 static void cfq_exit_single_io_context(struct io_context *ioc,
2561 struct cfq_io_context *cic)
2563 struct cfq_data *cfqd = cic->key;
2566 struct request_queue *q = cfqd->queue;
2567 unsigned long flags;
2569 spin_lock_irqsave(q->queue_lock, flags);
2572 * Ensure we get a fresh copy of the ->key to prevent
2573 * race between exiting task and queue
2575 smp_read_barrier_depends();
2577 __cfq_exit_single_io_context(cfqd, cic);
2579 spin_unlock_irqrestore(q->queue_lock, flags);
2584 * The process that ioc belongs to has exited, we need to clean up
2585 * and put the internal structures we have that belongs to that process.
2587 static void cfq_exit_io_context(struct io_context *ioc)
2589 call_for_each_cic(ioc, cfq_exit_single_io_context);
2592 static struct cfq_io_context *
2593 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2595 struct cfq_io_context *cic;
2597 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2600 cic->last_end_request = jiffies;
2601 INIT_LIST_HEAD(&cic->queue_list);
2602 INIT_HLIST_NODE(&cic->cic_list);
2603 cic->dtor = cfq_free_io_context;
2604 cic->exit = cfq_exit_io_context;
2605 elv_ioc_count_inc(cfq_ioc_count);
2611 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2613 struct task_struct *tsk = current;
2616 if (!cfq_cfqq_prio_changed(cfqq))
2619 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2620 switch (ioprio_class) {
2622 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2623 case IOPRIO_CLASS_NONE:
2625 * no prio set, inherit CPU scheduling settings
2627 cfqq->ioprio = task_nice_ioprio(tsk);
2628 cfqq->ioprio_class = task_nice_ioclass(tsk);
2630 case IOPRIO_CLASS_RT:
2631 cfqq->ioprio = task_ioprio(ioc);
2632 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2634 case IOPRIO_CLASS_BE:
2635 cfqq->ioprio = task_ioprio(ioc);
2636 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2638 case IOPRIO_CLASS_IDLE:
2639 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2641 cfq_clear_cfqq_idle_window(cfqq);
2646 * keep track of original prio settings in case we have to temporarily
2647 * elevate the priority of this queue
2649 cfqq->org_ioprio = cfqq->ioprio;
2650 cfqq->org_ioprio_class = cfqq->ioprio_class;
2651 cfq_clear_cfqq_prio_changed(cfqq);
2654 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2656 struct cfq_data *cfqd = cic->key;
2657 struct cfq_queue *cfqq;
2658 unsigned long flags;
2660 if (unlikely(!cfqd))
2663 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2665 cfqq = cic->cfqq[BLK_RW_ASYNC];
2667 struct cfq_queue *new_cfqq;
2668 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2671 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2672 cfq_put_queue(cfqq);
2676 cfqq = cic->cfqq[BLK_RW_SYNC];
2678 cfq_mark_cfqq_prio_changed(cfqq);
2680 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2683 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2685 call_for_each_cic(ioc, changed_ioprio);
2686 ioc->ioprio_changed = 0;
2689 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2690 pid_t pid, bool is_sync)
2692 RB_CLEAR_NODE(&cfqq->rb_node);
2693 RB_CLEAR_NODE(&cfqq->p_node);
2694 INIT_LIST_HEAD(&cfqq->fifo);
2696 atomic_set(&cfqq->ref, 0);
2699 cfq_mark_cfqq_prio_changed(cfqq);
2702 if (!cfq_class_idle(cfqq))
2703 cfq_mark_cfqq_idle_window(cfqq);
2704 cfq_mark_cfqq_sync(cfqq);
2709 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2710 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2712 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2713 struct cfq_data *cfqd = cic->key;
2714 unsigned long flags;
2715 struct request_queue *q;
2717 if (unlikely(!cfqd))
2722 spin_lock_irqsave(q->queue_lock, flags);
2726 * Drop reference to sync queue. A new sync queue will be
2727 * assigned in new group upon arrival of a fresh request.
2729 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2730 cic_set_cfqq(cic, NULL, 1);
2731 cfq_put_queue(sync_cfqq);
2734 spin_unlock_irqrestore(q->queue_lock, flags);
2737 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2739 call_for_each_cic(ioc, changed_cgroup);
2740 ioc->cgroup_changed = 0;
2742 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2744 static struct cfq_queue *
2745 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2746 struct io_context *ioc, gfp_t gfp_mask)
2748 struct cfq_queue *cfqq, *new_cfqq = NULL;
2749 struct cfq_io_context *cic;
2750 struct cfq_group *cfqg;
2753 cfqg = cfq_get_cfqg(cfqd, 1);
2754 cic = cfq_cic_lookup(cfqd, ioc);
2755 /* cic always exists here */
2756 cfqq = cic_to_cfqq(cic, is_sync);
2759 * Always try a new alloc if we fell back to the OOM cfqq
2760 * originally, since it should just be a temporary situation.
2762 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2767 } else if (gfp_mask & __GFP_WAIT) {
2768 spin_unlock_irq(cfqd->queue->queue_lock);
2769 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2770 gfp_mask | __GFP_ZERO,
2772 spin_lock_irq(cfqd->queue->queue_lock);
2776 cfqq = kmem_cache_alloc_node(cfq_pool,
2777 gfp_mask | __GFP_ZERO,
2782 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2783 cfq_init_prio_data(cfqq, ioc);
2784 cfq_link_cfqq_cfqg(cfqq, cfqg);
2785 cfq_log_cfqq(cfqd, cfqq, "alloced");
2787 cfqq = &cfqd->oom_cfqq;
2791 kmem_cache_free(cfq_pool, new_cfqq);
2796 static struct cfq_queue **
2797 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2799 switch (ioprio_class) {
2800 case IOPRIO_CLASS_RT:
2801 return &cfqd->async_cfqq[0][ioprio];
2802 case IOPRIO_CLASS_BE:
2803 return &cfqd->async_cfqq[1][ioprio];
2804 case IOPRIO_CLASS_IDLE:
2805 return &cfqd->async_idle_cfqq;
2811 static struct cfq_queue *
2812 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2815 const int ioprio = task_ioprio(ioc);
2816 const int ioprio_class = task_ioprio_class(ioc);
2817 struct cfq_queue **async_cfqq = NULL;
2818 struct cfq_queue *cfqq = NULL;
2821 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2826 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2829 * pin the queue now that it's allocated, scheduler exit will prune it
2831 if (!is_sync && !(*async_cfqq)) {
2832 atomic_inc(&cfqq->ref);
2836 atomic_inc(&cfqq->ref);
2841 * We drop cfq io contexts lazily, so we may find a dead one.
2844 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2845 struct cfq_io_context *cic)
2847 unsigned long flags;
2849 WARN_ON(!list_empty(&cic->queue_list));
2851 spin_lock_irqsave(&ioc->lock, flags);
2853 BUG_ON(ioc->ioc_data == cic);
2855 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2856 hlist_del_rcu(&cic->cic_list);
2857 spin_unlock_irqrestore(&ioc->lock, flags);
2862 static struct cfq_io_context *
2863 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2865 struct cfq_io_context *cic;
2866 unsigned long flags;
2875 * we maintain a last-hit cache, to avoid browsing over the tree
2877 cic = rcu_dereference(ioc->ioc_data);
2878 if (cic && cic->key == cfqd) {
2884 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2888 /* ->key must be copied to avoid race with cfq_exit_queue() */
2891 cfq_drop_dead_cic(cfqd, ioc, cic);
2896 spin_lock_irqsave(&ioc->lock, flags);
2897 rcu_assign_pointer(ioc->ioc_data, cic);
2898 spin_unlock_irqrestore(&ioc->lock, flags);
2906 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2907 * the process specific cfq io context when entered from the block layer.
2908 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2910 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2911 struct cfq_io_context *cic, gfp_t gfp_mask)
2913 unsigned long flags;
2916 ret = radix_tree_preload(gfp_mask);
2921 spin_lock_irqsave(&ioc->lock, flags);
2922 ret = radix_tree_insert(&ioc->radix_root,
2923 (unsigned long) cfqd, cic);
2925 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2926 spin_unlock_irqrestore(&ioc->lock, flags);
2928 radix_tree_preload_end();
2931 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2932 list_add(&cic->queue_list, &cfqd->cic_list);
2933 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2938 printk(KERN_ERR "cfq: cic link failed!\n");
2944 * Setup general io context and cfq io context. There can be several cfq
2945 * io contexts per general io context, if this process is doing io to more
2946 * than one device managed by cfq.
2948 static struct cfq_io_context *
2949 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2951 struct io_context *ioc = NULL;
2952 struct cfq_io_context *cic;
2954 might_sleep_if(gfp_mask & __GFP_WAIT);
2956 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2960 cic = cfq_cic_lookup(cfqd, ioc);
2964 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2968 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2972 smp_read_barrier_depends();
2973 if (unlikely(ioc->ioprio_changed))
2974 cfq_ioc_set_ioprio(ioc);
2976 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2977 if (unlikely(ioc->cgroup_changed))
2978 cfq_ioc_set_cgroup(ioc);
2984 put_io_context(ioc);
2989 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2991 unsigned long elapsed = jiffies - cic->last_end_request;
2992 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2994 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2995 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2996 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3000 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3004 sector_t n_sec = blk_rq_sectors(rq);
3005 if (cfqq->last_request_pos) {
3006 if (cfqq->last_request_pos < blk_rq_pos(rq))
3007 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3009 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3012 cfqq->seek_history <<= 1;
3013 if (blk_queue_nonrot(cfqd->queue))
3014 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3016 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3020 * Disable idle window if the process thinks too long or seeks so much that
3024 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3025 struct cfq_io_context *cic)
3027 int old_idle, enable_idle;
3030 * Don't idle for async or idle io prio class
3032 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3035 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3037 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3038 cfq_mark_cfqq_deep(cfqq);
3040 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3041 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3043 else if (sample_valid(cic->ttime_samples)) {
3044 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3050 if (old_idle != enable_idle) {
3051 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3053 cfq_mark_cfqq_idle_window(cfqq);
3055 cfq_clear_cfqq_idle_window(cfqq);
3060 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3061 * no or if we aren't sure, a 1 will cause a preempt.
3064 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3067 struct cfq_queue *cfqq;
3069 cfqq = cfqd->active_queue;
3073 if (cfq_class_idle(new_cfqq))
3076 if (cfq_class_idle(cfqq))
3080 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3082 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3086 * if the new request is sync, but the currently running queue is
3087 * not, let the sync request have priority.
3089 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3092 if (new_cfqq->cfqg != cfqq->cfqg)
3095 if (cfq_slice_used(cfqq))
3098 /* Allow preemption only if we are idling on sync-noidle tree */
3099 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3100 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3101 new_cfqq->service_tree->count == 2 &&
3102 RB_EMPTY_ROOT(&cfqq->sort_list))
3106 * So both queues are sync. Let the new request get disk time if
3107 * it's a metadata request and the current queue is doing regular IO.
3109 if (rq_is_meta(rq) && !cfqq->meta_pending)
3113 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3115 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3118 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3122 * if this request is as-good as one we would expect from the
3123 * current cfqq, let it preempt
3125 if (cfq_rq_close(cfqd, cfqq, rq))
3132 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3133 * let it have half of its nominal slice.
3135 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3137 cfq_log_cfqq(cfqd, cfqq, "preempt");
3138 cfq_slice_expired(cfqd, 1);
3141 * Put the new queue at the front of the of the current list,
3142 * so we know that it will be selected next.
3144 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3146 cfq_service_tree_add(cfqd, cfqq, 1);
3148 cfqq->slice_end = 0;
3149 cfq_mark_cfqq_slice_new(cfqq);
3153 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3154 * something we should do about it
3157 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3160 struct cfq_io_context *cic = RQ_CIC(rq);
3164 cfqq->meta_pending++;
3166 cfq_update_io_thinktime(cfqd, cic);
3167 cfq_update_io_seektime(cfqd, cfqq, rq);
3168 cfq_update_idle_window(cfqd, cfqq, cic);
3170 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3172 if (cfqq == cfqd->active_queue) {
3174 * Remember that we saw a request from this process, but
3175 * don't start queuing just yet. Otherwise we risk seeing lots
3176 * of tiny requests, because we disrupt the normal plugging
3177 * and merging. If the request is already larger than a single
3178 * page, let it rip immediately. For that case we assume that
3179 * merging is already done. Ditto for a busy system that
3180 * has other work pending, don't risk delaying until the
3181 * idle timer unplug to continue working.
3183 if (cfq_cfqq_wait_request(cfqq)) {
3184 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3185 cfqd->busy_queues > 1) {
3186 del_timer(&cfqd->idle_slice_timer);
3187 cfq_clear_cfqq_wait_request(cfqq);
3188 __blk_run_queue(cfqd->queue);
3190 cfq_mark_cfqq_must_dispatch(cfqq);
3192 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3194 * not the active queue - expire current slice if it is
3195 * idle and has expired it's mean thinktime or this new queue
3196 * has some old slice time left and is of higher priority or
3197 * this new queue is RT and the current one is BE
3199 cfq_preempt_queue(cfqd, cfqq);
3200 __blk_run_queue(cfqd->queue);
3204 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3206 struct cfq_data *cfqd = q->elevator->elevator_data;
3207 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3209 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3210 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3212 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3213 list_add_tail(&rq->queuelist, &cfqq->fifo);
3216 cfq_rq_enqueued(cfqd, cfqq, rq);
3220 * Update hw_tag based on peak queue depth over 50 samples under
3223 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3225 struct cfq_queue *cfqq = cfqd->active_queue;
3227 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3228 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3230 if (cfqd->hw_tag == 1)
3233 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3234 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3238 * If active queue hasn't enough requests and can idle, cfq might not
3239 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3242 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3243 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3244 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3247 if (cfqd->hw_tag_samples++ < 50)
3250 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3256 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3258 struct cfq_io_context *cic = cfqd->active_cic;
3260 /* If there are other queues in the group, don't wait */
3261 if (cfqq->cfqg->nr_cfqq > 1)
3264 if (cfq_slice_used(cfqq))
3267 /* if slice left is less than think time, wait busy */
3268 if (cic && sample_valid(cic->ttime_samples)
3269 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3273 * If think times is less than a jiffy than ttime_mean=0 and above
3274 * will not be true. It might happen that slice has not expired yet
3275 * but will expire soon (4-5 ns) during select_queue(). To cover the
3276 * case where think time is less than a jiffy, mark the queue wait
3277 * busy if only 1 jiffy is left in the slice.
3279 if (cfqq->slice_end - jiffies == 1)
3285 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3287 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3288 struct cfq_data *cfqd = cfqq->cfqd;
3289 const int sync = rq_is_sync(rq);
3293 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3295 cfq_update_hw_tag(cfqd);
3297 WARN_ON(!cfqd->rq_in_driver);
3298 WARN_ON(!cfqq->dispatched);
3299 cfqd->rq_in_driver--;
3301 blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3302 rq_io_start_time_ns(rq), rq_data_dir(rq),
3305 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3308 RQ_CIC(rq)->last_end_request = now;
3309 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3310 cfqd->last_delayed_sync = now;
3314 * If this is the active queue, check if it needs to be expired,
3315 * or if we want to idle in case it has no pending requests.
3317 if (cfqd->active_queue == cfqq) {
3318 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3320 if (cfq_cfqq_slice_new(cfqq)) {
3321 cfq_set_prio_slice(cfqd, cfqq);
3322 cfq_clear_cfqq_slice_new(cfqq);
3326 * Should we wait for next request to come in before we expire
3329 if (cfq_should_wait_busy(cfqd, cfqq)) {
3330 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3331 cfq_mark_cfqq_wait_busy(cfqq);
3332 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3336 * Idling is not enabled on:
3338 * - idle-priority queues
3340 * - queues with still some requests queued
3341 * - when there is a close cooperator
3343 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3344 cfq_slice_expired(cfqd, 1);
3345 else if (sync && cfqq_empty &&
3346 !cfq_close_cooperator(cfqd, cfqq)) {
3347 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3349 * Idling is enabled for SYNC_WORKLOAD.
3350 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3351 * only if we processed at least one !rq_noidle request
3353 if (cfqd->serving_type == SYNC_WORKLOAD
3354 || cfqd->noidle_tree_requires_idle
3355 || cfqq->cfqg->nr_cfqq == 1)
3356 cfq_arm_slice_timer(cfqd);
3360 if (!cfqd->rq_in_driver)
3361 cfq_schedule_dispatch(cfqd);
3365 * we temporarily boost lower priority queues if they are holding fs exclusive
3366 * resources. they are boosted to normal prio (CLASS_BE/4)
3368 static void cfq_prio_boost(struct cfq_queue *cfqq)
3370 if (has_fs_excl()) {
3372 * boost idle prio on transactions that would lock out other
3373 * users of the filesystem
3375 if (cfq_class_idle(cfqq))
3376 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3377 if (cfqq->ioprio > IOPRIO_NORM)
3378 cfqq->ioprio = IOPRIO_NORM;
3381 * unboost the queue (if needed)
3383 cfqq->ioprio_class = cfqq->org_ioprio_class;
3384 cfqq->ioprio = cfqq->org_ioprio;
3388 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3390 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3391 cfq_mark_cfqq_must_alloc_slice(cfqq);
3392 return ELV_MQUEUE_MUST;
3395 return ELV_MQUEUE_MAY;
3398 static int cfq_may_queue(struct request_queue *q, int rw)
3400 struct cfq_data *cfqd = q->elevator->elevator_data;
3401 struct task_struct *tsk = current;
3402 struct cfq_io_context *cic;
3403 struct cfq_queue *cfqq;
3406 * don't force setup of a queue from here, as a call to may_queue
3407 * does not necessarily imply that a request actually will be queued.
3408 * so just lookup a possibly existing queue, or return 'may queue'
3411 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3413 return ELV_MQUEUE_MAY;
3415 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3417 cfq_init_prio_data(cfqq, cic->ioc);
3418 cfq_prio_boost(cfqq);
3420 return __cfq_may_queue(cfqq);
3423 return ELV_MQUEUE_MAY;
3427 * queue lock held here
3429 static void cfq_put_request(struct request *rq)
3431 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3434 const int rw = rq_data_dir(rq);
3436 BUG_ON(!cfqq->allocated[rw]);
3437 cfqq->allocated[rw]--;
3439 put_io_context(RQ_CIC(rq)->ioc);
3441 rq->elevator_private = NULL;
3442 rq->elevator_private2 = NULL;
3444 cfq_put_queue(cfqq);
3448 static struct cfq_queue *
3449 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3450 struct cfq_queue *cfqq)
3452 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3453 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3454 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3455 cfq_put_queue(cfqq);
3456 return cic_to_cfqq(cic, 1);
3460 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3461 * was the last process referring to said cfqq.
3463 static struct cfq_queue *
3464 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3466 if (cfqq_process_refs(cfqq) == 1) {
3467 cfqq->pid = current->pid;
3468 cfq_clear_cfqq_coop(cfqq);
3469 cfq_clear_cfqq_split_coop(cfqq);
3473 cic_set_cfqq(cic, NULL, 1);
3474 cfq_put_queue(cfqq);
3478 * Allocate cfq data structures associated with this request.
3481 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3483 struct cfq_data *cfqd = q->elevator->elevator_data;
3484 struct cfq_io_context *cic;
3485 const int rw = rq_data_dir(rq);
3486 const bool is_sync = rq_is_sync(rq);
3487 struct cfq_queue *cfqq;
3488 unsigned long flags;
3490 might_sleep_if(gfp_mask & __GFP_WAIT);
3492 cic = cfq_get_io_context(cfqd, gfp_mask);
3494 spin_lock_irqsave(q->queue_lock, flags);
3500 cfqq = cic_to_cfqq(cic, is_sync);
3501 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3502 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3503 cic_set_cfqq(cic, cfqq, is_sync);
3506 * If the queue was seeky for too long, break it apart.
3508 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3509 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3510 cfqq = split_cfqq(cic, cfqq);
3516 * Check to see if this queue is scheduled to merge with
3517 * another, closely cooperating queue. The merging of
3518 * queues happens here as it must be done in process context.
3519 * The reference on new_cfqq was taken in merge_cfqqs.
3522 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3525 cfqq->allocated[rw]++;
3526 atomic_inc(&cfqq->ref);
3528 spin_unlock_irqrestore(q->queue_lock, flags);
3530 rq->elevator_private = cic;
3531 rq->elevator_private2 = cfqq;
3536 put_io_context(cic->ioc);
3538 cfq_schedule_dispatch(cfqd);
3539 spin_unlock_irqrestore(q->queue_lock, flags);
3540 cfq_log(cfqd, "set_request fail");
3544 static void cfq_kick_queue(struct work_struct *work)
3546 struct cfq_data *cfqd =
3547 container_of(work, struct cfq_data, unplug_work);
3548 struct request_queue *q = cfqd->queue;
3550 spin_lock_irq(q->queue_lock);
3551 __blk_run_queue(cfqd->queue);
3552 spin_unlock_irq(q->queue_lock);
3556 * Timer running if the active_queue is currently idling inside its time slice
3558 static void cfq_idle_slice_timer(unsigned long data)
3560 struct cfq_data *cfqd = (struct cfq_data *) data;
3561 struct cfq_queue *cfqq;
3562 unsigned long flags;
3565 cfq_log(cfqd, "idle timer fired");
3567 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3569 cfqq = cfqd->active_queue;
3574 * We saw a request before the queue expired, let it through
3576 if (cfq_cfqq_must_dispatch(cfqq))
3582 if (cfq_slice_used(cfqq))
3586 * only expire and reinvoke request handler, if there are
3587 * other queues with pending requests
3589 if (!cfqd->busy_queues)
3593 * not expired and it has a request pending, let it dispatch
3595 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3599 * Queue depth flag is reset only when the idle didn't succeed
3601 cfq_clear_cfqq_deep(cfqq);
3604 cfq_slice_expired(cfqd, timed_out);
3606 cfq_schedule_dispatch(cfqd);
3608 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3611 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3613 del_timer_sync(&cfqd->idle_slice_timer);
3614 cancel_work_sync(&cfqd->unplug_work);
3617 static void cfq_put_async_queues(struct cfq_data *cfqd)
3621 for (i = 0; i < IOPRIO_BE_NR; i++) {
3622 if (cfqd->async_cfqq[0][i])
3623 cfq_put_queue(cfqd->async_cfqq[0][i]);
3624 if (cfqd->async_cfqq[1][i])
3625 cfq_put_queue(cfqd->async_cfqq[1][i]);
3628 if (cfqd->async_idle_cfqq)
3629 cfq_put_queue(cfqd->async_idle_cfqq);
3632 static void cfq_cfqd_free(struct rcu_head *head)
3634 kfree(container_of(head, struct cfq_data, rcu));
3637 static void cfq_exit_queue(struct elevator_queue *e)
3639 struct cfq_data *cfqd = e->elevator_data;
3640 struct request_queue *q = cfqd->queue;
3642 cfq_shutdown_timer_wq(cfqd);
3644 spin_lock_irq(q->queue_lock);
3646 if (cfqd->active_queue)
3647 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3649 while (!list_empty(&cfqd->cic_list)) {
3650 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3651 struct cfq_io_context,
3654 __cfq_exit_single_io_context(cfqd, cic);
3657 cfq_put_async_queues(cfqd);
3658 cfq_release_cfq_groups(cfqd);
3659 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3661 spin_unlock_irq(q->queue_lock);
3663 cfq_shutdown_timer_wq(cfqd);
3665 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3666 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3669 static void *cfq_init_queue(struct request_queue *q)
3671 struct cfq_data *cfqd;
3673 struct cfq_group *cfqg;
3674 struct cfq_rb_root *st;
3676 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3680 /* Init root service tree */
3681 cfqd->grp_service_tree = CFQ_RB_ROOT;
3683 /* Init root group */
3684 cfqg = &cfqd->root_group;
3685 for_each_cfqg_st(cfqg, i, j, st)
3687 RB_CLEAR_NODE(&cfqg->rb_node);
3689 /* Give preference to root group over other groups */
3690 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3692 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3694 * Take a reference to root group which we never drop. This is just
3695 * to make sure that cfq_put_cfqg() does not try to kfree root group
3697 atomic_set(&cfqg->ref, 1);
3698 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3702 * Not strictly needed (since RB_ROOT just clears the node and we
3703 * zeroed cfqd on alloc), but better be safe in case someone decides
3704 * to add magic to the rb code
3706 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3707 cfqd->prio_trees[i] = RB_ROOT;
3710 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3711 * Grab a permanent reference to it, so that the normal code flow
3712 * will not attempt to free it.
3714 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3715 atomic_inc(&cfqd->oom_cfqq.ref);
3716 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3718 INIT_LIST_HEAD(&cfqd->cic_list);
3722 init_timer(&cfqd->idle_slice_timer);
3723 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3724 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3726 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3728 cfqd->cfq_quantum = cfq_quantum;
3729 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3730 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3731 cfqd->cfq_back_max = cfq_back_max;
3732 cfqd->cfq_back_penalty = cfq_back_penalty;
3733 cfqd->cfq_slice[0] = cfq_slice_async;
3734 cfqd->cfq_slice[1] = cfq_slice_sync;
3735 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3736 cfqd->cfq_slice_idle = cfq_slice_idle;
3737 cfqd->cfq_latency = 1;
3738 cfqd->cfq_group_isolation = 0;
3741 * we optimistically start assuming sync ops weren't delayed in last
3742 * second, in order to have larger depth for async operations.
3744 cfqd->last_delayed_sync = jiffies - HZ;
3745 INIT_RCU_HEAD(&cfqd->rcu);
3749 static void cfq_slab_kill(void)
3752 * Caller already ensured that pending RCU callbacks are completed,
3753 * so we should have no busy allocations at this point.
3756 kmem_cache_destroy(cfq_pool);
3758 kmem_cache_destroy(cfq_ioc_pool);
3761 static int __init cfq_slab_setup(void)
3763 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3767 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3778 * sysfs parts below -->
3781 cfq_var_show(unsigned int var, char *page)
3783 return sprintf(page, "%d\n", var);
3787 cfq_var_store(unsigned int *var, const char *page, size_t count)
3789 char *p = (char *) page;
3791 *var = simple_strtoul(p, &p, 10);
3795 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3796 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3798 struct cfq_data *cfqd = e->elevator_data; \
3799 unsigned int __data = __VAR; \
3801 __data = jiffies_to_msecs(__data); \
3802 return cfq_var_show(__data, (page)); \
3804 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3805 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3806 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3807 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3808 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3809 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3810 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3811 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3812 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3813 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3814 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3815 #undef SHOW_FUNCTION
3817 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3818 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3820 struct cfq_data *cfqd = e->elevator_data; \
3821 unsigned int __data; \
3822 int ret = cfq_var_store(&__data, (page), count); \
3823 if (__data < (MIN)) \
3825 else if (__data > (MAX)) \
3828 *(__PTR) = msecs_to_jiffies(__data); \
3830 *(__PTR) = __data; \
3833 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3834 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3836 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3838 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3839 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3841 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3842 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3843 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3844 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3846 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3847 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3848 #undef STORE_FUNCTION
3850 #define CFQ_ATTR(name) \
3851 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3853 static struct elv_fs_entry cfq_attrs[] = {
3855 CFQ_ATTR(fifo_expire_sync),
3856 CFQ_ATTR(fifo_expire_async),
3857 CFQ_ATTR(back_seek_max),
3858 CFQ_ATTR(back_seek_penalty),
3859 CFQ_ATTR(slice_sync),
3860 CFQ_ATTR(slice_async),
3861 CFQ_ATTR(slice_async_rq),
3862 CFQ_ATTR(slice_idle),
3863 CFQ_ATTR(low_latency),
3864 CFQ_ATTR(group_isolation),
3868 static struct elevator_type iosched_cfq = {
3870 .elevator_merge_fn = cfq_merge,
3871 .elevator_merged_fn = cfq_merged_request,
3872 .elevator_merge_req_fn = cfq_merged_requests,
3873 .elevator_allow_merge_fn = cfq_allow_merge,
3874 .elevator_bio_merged_fn = cfq_bio_merged,
3875 .elevator_dispatch_fn = cfq_dispatch_requests,
3876 .elevator_add_req_fn = cfq_insert_request,
3877 .elevator_activate_req_fn = cfq_activate_request,
3878 .elevator_deactivate_req_fn = cfq_deactivate_request,
3879 .elevator_queue_empty_fn = cfq_queue_empty,
3880 .elevator_completed_req_fn = cfq_completed_request,
3881 .elevator_former_req_fn = elv_rb_former_request,
3882 .elevator_latter_req_fn = elv_rb_latter_request,
3883 .elevator_set_req_fn = cfq_set_request,
3884 .elevator_put_req_fn = cfq_put_request,
3885 .elevator_may_queue_fn = cfq_may_queue,
3886 .elevator_init_fn = cfq_init_queue,
3887 .elevator_exit_fn = cfq_exit_queue,
3888 .trim = cfq_free_io_context,
3890 .elevator_attrs = cfq_attrs,
3891 .elevator_name = "cfq",
3892 .elevator_owner = THIS_MODULE,
3895 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3896 static struct blkio_policy_type blkio_policy_cfq = {
3898 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3899 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3903 static struct blkio_policy_type blkio_policy_cfq;
3906 static int __init cfq_init(void)
3909 * could be 0 on HZ < 1000 setups
3911 if (!cfq_slice_async)
3912 cfq_slice_async = 1;
3913 if (!cfq_slice_idle)
3916 if (cfq_slab_setup())
3919 elv_register(&iosched_cfq);
3920 blkio_policy_register(&blkio_policy_cfq);
3925 static void __exit cfq_exit(void)
3927 DECLARE_COMPLETION_ONSTACK(all_gone);
3928 blkio_policy_unregister(&blkio_policy_cfq);
3929 elv_unregister(&iosched_cfq);
3930 ioc_gone = &all_gone;
3931 /* ioc_gone's update must be visible before reading ioc_count */
3935 * this also protects us from entering cfq_slab_kill() with
3936 * pending RCU callbacks
3938 if (elv_ioc_count_read(cfq_ioc_count))
3939 wait_for_completion(&all_gone);
3943 module_init(cfq_init);
3944 module_exit(cfq_exit);
3946 MODULE_AUTHOR("Jens Axboe");
3947 MODULE_LICENSE("GPL");
3948 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");