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_blkio_group_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_blkio_group_stats(&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);
960 * Take the initial reference that will be released on destroy
961 * This can be thought of a joint reference by cgroup and
962 * elevator which will be dropped by either elevator exit
963 * or cgroup deletion path depending on who is exiting first.
965 atomic_set(&cfqg->ref, 1);
967 /* Add group onto cgroup list */
968 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
969 blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
970 MKDEV(major, minor));
972 /* Add group on cfqd list */
973 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
980 * Search for the cfq group current task belongs to. If create = 1, then also
981 * create the cfq group if it does not exist. request_queue lock must be held.
983 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
985 struct cgroup *cgroup;
986 struct cfq_group *cfqg = NULL;
989 cgroup = task_cgroup(current, blkio_subsys_id);
990 cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
992 cfqg = &cfqd->root_group;
997 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
999 /* Currently, all async queues are mapped to root group */
1000 if (!cfq_cfqq_sync(cfqq))
1001 cfqg = &cfqq->cfqd->root_group;
1004 /* cfqq reference on cfqg */
1005 atomic_inc(&cfqq->cfqg->ref);
1008 static void cfq_put_cfqg(struct cfq_group *cfqg)
1010 struct cfq_rb_root *st;
1013 BUG_ON(atomic_read(&cfqg->ref) <= 0);
1014 if (!atomic_dec_and_test(&cfqg->ref))
1016 for_each_cfqg_st(cfqg, i, j, st)
1017 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1021 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1023 /* Something wrong if we are trying to remove same group twice */
1024 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1026 hlist_del_init(&cfqg->cfqd_node);
1029 * Put the reference taken at the time of creation so that when all
1030 * queues are gone, group can be destroyed.
1035 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1037 struct hlist_node *pos, *n;
1038 struct cfq_group *cfqg;
1040 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1042 * If cgroup removal path got to blk_group first and removed
1043 * it from cgroup list, then it will take care of destroying
1046 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1047 cfq_destroy_cfqg(cfqd, cfqg);
1052 * Blk cgroup controller notification saying that blkio_group object is being
1053 * delinked as associated cgroup object is going away. That also means that
1054 * no new IO will come in this group. So get rid of this group as soon as
1055 * any pending IO in the group is finished.
1057 * This function is called under rcu_read_lock(). key is the rcu protected
1058 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1061 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1062 * it should not be NULL as even if elevator was exiting, cgroup deltion
1063 * path got to it first.
1065 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1067 unsigned long flags;
1068 struct cfq_data *cfqd = key;
1070 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1071 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1072 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1075 #else /* GROUP_IOSCHED */
1076 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1078 return &cfqd->root_group;
1081 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1085 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1086 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1088 #endif /* GROUP_IOSCHED */
1091 * The cfqd->service_trees holds all pending cfq_queue's that have
1092 * requests waiting to be processed. It is sorted in the order that
1093 * we will service the queues.
1095 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1098 struct rb_node **p, *parent;
1099 struct cfq_queue *__cfqq;
1100 unsigned long rb_key;
1101 struct cfq_rb_root *service_tree;
1104 int group_changed = 0;
1106 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1107 if (!cfqd->cfq_group_isolation
1108 && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1109 && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1110 /* Move this cfq to root group */
1111 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1112 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1113 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1114 cfqq->orig_cfqg = cfqq->cfqg;
1115 cfqq->cfqg = &cfqd->root_group;
1116 atomic_inc(&cfqd->root_group.ref);
1118 } else if (!cfqd->cfq_group_isolation
1119 && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1120 /* cfqq is sequential now needs to go to its original group */
1121 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1122 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1123 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1124 cfq_put_cfqg(cfqq->cfqg);
1125 cfqq->cfqg = cfqq->orig_cfqg;
1126 cfqq->orig_cfqg = NULL;
1128 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1132 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1134 if (cfq_class_idle(cfqq)) {
1135 rb_key = CFQ_IDLE_DELAY;
1136 parent = rb_last(&service_tree->rb);
1137 if (parent && parent != &cfqq->rb_node) {
1138 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1139 rb_key += __cfqq->rb_key;
1142 } else if (!add_front) {
1144 * Get our rb key offset. Subtract any residual slice
1145 * value carried from last service. A negative resid
1146 * count indicates slice overrun, and this should position
1147 * the next service time further away in the tree.
1149 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1150 rb_key -= cfqq->slice_resid;
1151 cfqq->slice_resid = 0;
1154 __cfqq = cfq_rb_first(service_tree);
1155 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1158 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1161 * same position, nothing more to do
1163 if (rb_key == cfqq->rb_key &&
1164 cfqq->service_tree == service_tree)
1167 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1168 cfqq->service_tree = NULL;
1173 cfqq->service_tree = service_tree;
1174 p = &service_tree->rb.rb_node;
1179 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1182 * sort by key, that represents service time.
1184 if (time_before(rb_key, __cfqq->rb_key))
1187 n = &(*p)->rb_right;
1195 service_tree->left = &cfqq->rb_node;
1197 cfqq->rb_key = rb_key;
1198 rb_link_node(&cfqq->rb_node, parent, p);
1199 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1200 service_tree->count++;
1201 if ((add_front || !new_cfqq) && !group_changed)
1203 cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1206 static struct cfq_queue *
1207 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1208 sector_t sector, struct rb_node **ret_parent,
1209 struct rb_node ***rb_link)
1211 struct rb_node **p, *parent;
1212 struct cfq_queue *cfqq = NULL;
1220 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1223 * Sort strictly based on sector. Smallest to the left,
1224 * largest to the right.
1226 if (sector > blk_rq_pos(cfqq->next_rq))
1227 n = &(*p)->rb_right;
1228 else if (sector < blk_rq_pos(cfqq->next_rq))
1236 *ret_parent = parent;
1242 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1244 struct rb_node **p, *parent;
1245 struct cfq_queue *__cfqq;
1248 rb_erase(&cfqq->p_node, cfqq->p_root);
1249 cfqq->p_root = NULL;
1252 if (cfq_class_idle(cfqq))
1257 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1258 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1259 blk_rq_pos(cfqq->next_rq), &parent, &p);
1261 rb_link_node(&cfqq->p_node, parent, p);
1262 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1264 cfqq->p_root = NULL;
1268 * Update cfqq's position in the service tree.
1270 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1273 * Resorting requires the cfqq to be on the RR list already.
1275 if (cfq_cfqq_on_rr(cfqq)) {
1276 cfq_service_tree_add(cfqd, cfqq, 0);
1277 cfq_prio_tree_add(cfqd, cfqq);
1282 * add to busy list of queues for service, trying to be fair in ordering
1283 * the pending list according to last request service
1285 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1287 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1288 BUG_ON(cfq_cfqq_on_rr(cfqq));
1289 cfq_mark_cfqq_on_rr(cfqq);
1290 cfqd->busy_queues++;
1292 cfq_resort_rr_list(cfqd, cfqq);
1296 * Called when the cfqq no longer has requests pending, remove it from
1299 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1301 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1302 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1303 cfq_clear_cfqq_on_rr(cfqq);
1305 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1306 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1307 cfqq->service_tree = NULL;
1310 rb_erase(&cfqq->p_node, cfqq->p_root);
1311 cfqq->p_root = NULL;
1314 cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1315 BUG_ON(!cfqd->busy_queues);
1316 cfqd->busy_queues--;
1320 * rb tree support functions
1322 static void cfq_del_rq_rb(struct request *rq)
1324 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1325 const int sync = rq_is_sync(rq);
1327 BUG_ON(!cfqq->queued[sync]);
1328 cfqq->queued[sync]--;
1330 elv_rb_del(&cfqq->sort_list, rq);
1332 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1334 * Queue will be deleted from service tree when we actually
1335 * expire it later. Right now just remove it from prio tree
1339 rb_erase(&cfqq->p_node, cfqq->p_root);
1340 cfqq->p_root = NULL;
1345 static void cfq_add_rq_rb(struct request *rq)
1347 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1348 struct cfq_data *cfqd = cfqq->cfqd;
1349 struct request *__alias, *prev;
1351 cfqq->queued[rq_is_sync(rq)]++;
1354 * looks a little odd, but the first insert might return an alias.
1355 * if that happens, put the alias on the dispatch list
1357 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1358 cfq_dispatch_insert(cfqd->queue, __alias);
1360 if (!cfq_cfqq_on_rr(cfqq))
1361 cfq_add_cfqq_rr(cfqd, cfqq);
1364 * check if this request is a better next-serve candidate
1366 prev = cfqq->next_rq;
1367 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1370 * adjust priority tree position, if ->next_rq changes
1372 if (prev != cfqq->next_rq)
1373 cfq_prio_tree_add(cfqd, cfqq);
1375 BUG_ON(!cfqq->next_rq);
1378 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1380 elv_rb_del(&cfqq->sort_list, rq);
1381 cfqq->queued[rq_is_sync(rq)]--;
1385 static struct request *
1386 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1388 struct task_struct *tsk = current;
1389 struct cfq_io_context *cic;
1390 struct cfq_queue *cfqq;
1392 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1396 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1398 sector_t sector = bio->bi_sector + bio_sectors(bio);
1400 return elv_rb_find(&cfqq->sort_list, sector);
1406 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1408 struct cfq_data *cfqd = q->elevator->elevator_data;
1410 cfqd->rq_in_driver++;
1411 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1412 cfqd->rq_in_driver);
1414 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1417 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1419 struct cfq_data *cfqd = q->elevator->elevator_data;
1421 WARN_ON(!cfqd->rq_in_driver);
1422 cfqd->rq_in_driver--;
1423 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1424 cfqd->rq_in_driver);
1427 static void cfq_remove_request(struct request *rq)
1429 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1431 if (cfqq->next_rq == rq)
1432 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1434 list_del_init(&rq->queuelist);
1437 cfqq->cfqd->rq_queued--;
1438 if (rq_is_meta(rq)) {
1439 WARN_ON(!cfqq->meta_pending);
1440 cfqq->meta_pending--;
1444 static int cfq_merge(struct request_queue *q, struct request **req,
1447 struct cfq_data *cfqd = q->elevator->elevator_data;
1448 struct request *__rq;
1450 __rq = cfq_find_rq_fmerge(cfqd, bio);
1451 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1453 return ELEVATOR_FRONT_MERGE;
1456 return ELEVATOR_NO_MERGE;
1459 static void cfq_merged_request(struct request_queue *q, struct request *req,
1462 if (type == ELEVATOR_FRONT_MERGE) {
1463 struct cfq_queue *cfqq = RQ_CFQQ(req);
1465 cfq_reposition_rq_rb(cfqq, req);
1470 cfq_merged_requests(struct request_queue *q, struct request *rq,
1471 struct request *next)
1473 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1475 * reposition in fifo if next is older than rq
1477 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1478 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1479 list_move(&rq->queuelist, &next->queuelist);
1480 rq_set_fifo_time(rq, rq_fifo_time(next));
1483 if (cfqq->next_rq == next)
1485 cfq_remove_request(next);
1488 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1491 struct cfq_data *cfqd = q->elevator->elevator_data;
1492 struct cfq_io_context *cic;
1493 struct cfq_queue *cfqq;
1496 * Disallow merge of a sync bio into an async request.
1498 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1502 * Lookup the cfqq that this bio will be queued with. Allow
1503 * merge only if rq is queued there.
1505 cic = cfq_cic_lookup(cfqd, current->io_context);
1509 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1510 return cfqq == RQ_CFQQ(rq);
1513 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1514 struct cfq_queue *cfqq)
1517 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1518 cfqd->serving_prio, cfqd->serving_type);
1519 cfqq->slice_start = 0;
1520 cfqq->dispatch_start = jiffies;
1521 cfqq->allocated_slice = 0;
1522 cfqq->slice_end = 0;
1523 cfqq->slice_dispatch = 0;
1525 cfq_clear_cfqq_wait_request(cfqq);
1526 cfq_clear_cfqq_must_dispatch(cfqq);
1527 cfq_clear_cfqq_must_alloc_slice(cfqq);
1528 cfq_clear_cfqq_fifo_expire(cfqq);
1529 cfq_mark_cfqq_slice_new(cfqq);
1531 del_timer(&cfqd->idle_slice_timer);
1534 cfqd->active_queue = cfqq;
1538 * current cfqq expired its slice (or was too idle), select new one
1541 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1544 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1546 if (cfq_cfqq_wait_request(cfqq))
1547 del_timer(&cfqd->idle_slice_timer);
1549 cfq_clear_cfqq_wait_request(cfqq);
1550 cfq_clear_cfqq_wait_busy(cfqq);
1553 * If this cfqq is shared between multiple processes, check to
1554 * make sure that those processes are still issuing I/Os within
1555 * the mean seek distance. If not, it may be time to break the
1556 * queues apart again.
1558 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1559 cfq_mark_cfqq_split_coop(cfqq);
1562 * store what was left of this slice, if the queue idled/timed out
1564 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1565 cfqq->slice_resid = cfqq->slice_end - jiffies;
1566 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1569 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1571 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1572 cfq_del_cfqq_rr(cfqd, cfqq);
1574 cfq_resort_rr_list(cfqd, cfqq);
1576 if (cfqq == cfqd->active_queue)
1577 cfqd->active_queue = NULL;
1579 if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1580 cfqd->grp_service_tree.active = NULL;
1582 if (cfqd->active_cic) {
1583 put_io_context(cfqd->active_cic->ioc);
1584 cfqd->active_cic = NULL;
1588 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1590 struct cfq_queue *cfqq = cfqd->active_queue;
1593 __cfq_slice_expired(cfqd, cfqq, timed_out);
1597 * Get next queue for service. Unless we have a queue preemption,
1598 * we'll simply select the first cfqq in the service tree.
1600 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1602 struct cfq_rb_root *service_tree =
1603 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1604 cfqd->serving_type);
1606 if (!cfqd->rq_queued)
1609 /* There is nothing to dispatch */
1612 if (RB_EMPTY_ROOT(&service_tree->rb))
1614 return cfq_rb_first(service_tree);
1617 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1619 struct cfq_group *cfqg;
1620 struct cfq_queue *cfqq;
1622 struct cfq_rb_root *st;
1624 if (!cfqd->rq_queued)
1627 cfqg = cfq_get_next_cfqg(cfqd);
1631 for_each_cfqg_st(cfqg, i, j, st)
1632 if ((cfqq = cfq_rb_first(st)) != NULL)
1638 * Get and set a new active queue for service.
1640 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1641 struct cfq_queue *cfqq)
1644 cfqq = cfq_get_next_queue(cfqd);
1646 __cfq_set_active_queue(cfqd, cfqq);
1650 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1653 if (blk_rq_pos(rq) >= cfqd->last_position)
1654 return blk_rq_pos(rq) - cfqd->last_position;
1656 return cfqd->last_position - blk_rq_pos(rq);
1659 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1662 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1665 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1666 struct cfq_queue *cur_cfqq)
1668 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1669 struct rb_node *parent, *node;
1670 struct cfq_queue *__cfqq;
1671 sector_t sector = cfqd->last_position;
1673 if (RB_EMPTY_ROOT(root))
1677 * First, if we find a request starting at the end of the last
1678 * request, choose it.
1680 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1685 * If the exact sector wasn't found, the parent of the NULL leaf
1686 * will contain the closest sector.
1688 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1689 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1692 if (blk_rq_pos(__cfqq->next_rq) < sector)
1693 node = rb_next(&__cfqq->p_node);
1695 node = rb_prev(&__cfqq->p_node);
1699 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1700 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1708 * cur_cfqq - passed in so that we don't decide that the current queue is
1709 * closely cooperating with itself.
1711 * So, basically we're assuming that that cur_cfqq has dispatched at least
1712 * one request, and that cfqd->last_position reflects a position on the disk
1713 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1716 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1717 struct cfq_queue *cur_cfqq)
1719 struct cfq_queue *cfqq;
1721 if (cfq_class_idle(cur_cfqq))
1723 if (!cfq_cfqq_sync(cur_cfqq))
1725 if (CFQQ_SEEKY(cur_cfqq))
1729 * Don't search priority tree if it's the only queue in the group.
1731 if (cur_cfqq->cfqg->nr_cfqq == 1)
1735 * We should notice if some of the queues are cooperating, eg
1736 * working closely on the same area of the disk. In that case,
1737 * we can group them together and don't waste time idling.
1739 cfqq = cfqq_close(cfqd, cur_cfqq);
1743 /* If new queue belongs to different cfq_group, don't choose it */
1744 if (cur_cfqq->cfqg != cfqq->cfqg)
1748 * It only makes sense to merge sync queues.
1750 if (!cfq_cfqq_sync(cfqq))
1752 if (CFQQ_SEEKY(cfqq))
1756 * Do not merge queues of different priority classes
1758 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1765 * Determine whether we should enforce idle window for this queue.
1768 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1770 enum wl_prio_t prio = cfqq_prio(cfqq);
1771 struct cfq_rb_root *service_tree = cfqq->service_tree;
1773 BUG_ON(!service_tree);
1774 BUG_ON(!service_tree->count);
1776 /* We never do for idle class queues. */
1777 if (prio == IDLE_WORKLOAD)
1780 /* We do for queues that were marked with idle window flag. */
1781 if (cfq_cfqq_idle_window(cfqq) &&
1782 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1786 * Otherwise, we do only if they are the last ones
1787 * in their service tree.
1789 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1791 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1792 service_tree->count);
1796 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1798 struct cfq_queue *cfqq = cfqd->active_queue;
1799 struct cfq_io_context *cic;
1803 * SSD device without seek penalty, disable idling. But only do so
1804 * for devices that support queuing, otherwise we still have a problem
1805 * with sync vs async workloads.
1807 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1810 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1811 WARN_ON(cfq_cfqq_slice_new(cfqq));
1814 * idle is disabled, either manually or by past process history
1816 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1820 * still active requests from this queue, don't idle
1822 if (cfqq->dispatched)
1826 * task has exited, don't wait
1828 cic = cfqd->active_cic;
1829 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1833 * If our average think time is larger than the remaining time
1834 * slice, then don't idle. This avoids overrunning the allotted
1837 if (sample_valid(cic->ttime_samples) &&
1838 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1839 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1844 cfq_mark_cfqq_wait_request(cfqq);
1846 sl = cfqd->cfq_slice_idle;
1848 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1849 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1853 * Move request from internal lists to the request queue dispatch list.
1855 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1857 struct cfq_data *cfqd = q->elevator->elevator_data;
1858 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1860 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1862 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1863 cfq_remove_request(rq);
1865 elv_dispatch_sort(q, rq);
1867 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1871 * return expired entry, or NULL to just start from scratch in rbtree
1873 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1875 struct request *rq = NULL;
1877 if (cfq_cfqq_fifo_expire(cfqq))
1880 cfq_mark_cfqq_fifo_expire(cfqq);
1882 if (list_empty(&cfqq->fifo))
1885 rq = rq_entry_fifo(cfqq->fifo.next);
1886 if (time_before(jiffies, rq_fifo_time(rq)))
1889 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1894 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1896 const int base_rq = cfqd->cfq_slice_async_rq;
1898 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1900 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1904 * Must be called with the queue_lock held.
1906 static int cfqq_process_refs(struct cfq_queue *cfqq)
1908 int process_refs, io_refs;
1910 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1911 process_refs = atomic_read(&cfqq->ref) - io_refs;
1912 BUG_ON(process_refs < 0);
1913 return process_refs;
1916 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1918 int process_refs, new_process_refs;
1919 struct cfq_queue *__cfqq;
1921 /* Avoid a circular list and skip interim queue merges */
1922 while ((__cfqq = new_cfqq->new_cfqq)) {
1928 process_refs = cfqq_process_refs(cfqq);
1930 * If the process for the cfqq has gone away, there is no
1931 * sense in merging the queues.
1933 if (process_refs == 0)
1937 * Merge in the direction of the lesser amount of work.
1939 new_process_refs = cfqq_process_refs(new_cfqq);
1940 if (new_process_refs >= process_refs) {
1941 cfqq->new_cfqq = new_cfqq;
1942 atomic_add(process_refs, &new_cfqq->ref);
1944 new_cfqq->new_cfqq = cfqq;
1945 atomic_add(new_process_refs, &cfqq->ref);
1949 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1950 struct cfq_group *cfqg, enum wl_prio_t prio)
1952 struct cfq_queue *queue;
1954 bool key_valid = false;
1955 unsigned long lowest_key = 0;
1956 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1958 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
1959 /* select the one with lowest rb_key */
1960 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
1962 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1963 lowest_key = queue->rb_key;
1972 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1976 struct cfq_rb_root *st;
1977 unsigned group_slice;
1980 cfqd->serving_prio = IDLE_WORKLOAD;
1981 cfqd->workload_expires = jiffies + 1;
1985 /* Choose next priority. RT > BE > IDLE */
1986 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
1987 cfqd->serving_prio = RT_WORKLOAD;
1988 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
1989 cfqd->serving_prio = BE_WORKLOAD;
1991 cfqd->serving_prio = IDLE_WORKLOAD;
1992 cfqd->workload_expires = jiffies + 1;
1997 * For RT and BE, we have to choose also the type
1998 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2001 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2005 * check workload expiration, and that we still have other queues ready
2007 if (count && !time_after(jiffies, cfqd->workload_expires))
2010 /* otherwise select new workload type */
2011 cfqd->serving_type =
2012 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2013 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2017 * the workload slice is computed as a fraction of target latency
2018 * proportional to the number of queues in that workload, over
2019 * all the queues in the same priority class
2021 group_slice = cfq_group_slice(cfqd, cfqg);
2023 slice = group_slice * count /
2024 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2025 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2027 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2031 * Async queues are currently system wide. Just taking
2032 * proportion of queues with-in same group will lead to higher
2033 * async ratio system wide as generally root group is going
2034 * to have higher weight. A more accurate thing would be to
2035 * calculate system wide asnc/sync ratio.
2037 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2038 tmp = tmp/cfqd->busy_queues;
2039 slice = min_t(unsigned, slice, tmp);
2041 /* async workload slice is scaled down according to
2042 * the sync/async slice ratio. */
2043 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2045 /* sync workload slice is at least 2 * cfq_slice_idle */
2046 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2048 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2049 cfq_log(cfqd, "workload slice:%d", slice);
2050 cfqd->workload_expires = jiffies + slice;
2051 cfqd->noidle_tree_requires_idle = false;
2054 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2056 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2057 struct cfq_group *cfqg;
2059 if (RB_EMPTY_ROOT(&st->rb))
2061 cfqg = cfq_rb_first_group(st);
2062 st->active = &cfqg->rb_node;
2063 update_min_vdisktime(st);
2067 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2069 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2071 cfqd->serving_group = cfqg;
2073 /* Restore the workload type data */
2074 if (cfqg->saved_workload_slice) {
2075 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2076 cfqd->serving_type = cfqg->saved_workload;
2077 cfqd->serving_prio = cfqg->saved_serving_prio;
2079 cfqd->workload_expires = jiffies - 1;
2081 choose_service_tree(cfqd, cfqg);
2085 * Select a queue for service. If we have a current active queue,
2086 * check whether to continue servicing it, or retrieve and set a new one.
2088 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2090 struct cfq_queue *cfqq, *new_cfqq = NULL;
2092 cfqq = cfqd->active_queue;
2096 if (!cfqd->rq_queued)
2100 * We were waiting for group to get backlogged. Expire the queue
2102 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2106 * The active queue has run out of time, expire it and select new.
2108 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2110 * If slice had not expired at the completion of last request
2111 * we might not have turned on wait_busy flag. Don't expire
2112 * the queue yet. Allow the group to get backlogged.
2114 * The very fact that we have used the slice, that means we
2115 * have been idling all along on this queue and it should be
2116 * ok to wait for this request to complete.
2118 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2119 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2127 * The active queue has requests and isn't expired, allow it to
2130 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2134 * If another queue has a request waiting within our mean seek
2135 * distance, let it run. The expire code will check for close
2136 * cooperators and put the close queue at the front of the service
2137 * tree. If possible, merge the expiring queue with the new cfqq.
2139 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2141 if (!cfqq->new_cfqq)
2142 cfq_setup_merge(cfqq, new_cfqq);
2147 * No requests pending. If the active queue still has requests in
2148 * flight or is idling for a new request, allow either of these
2149 * conditions to happen (or time out) before selecting a new queue.
2151 if (timer_pending(&cfqd->idle_slice_timer) ||
2152 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2158 cfq_slice_expired(cfqd, 0);
2161 * Current queue expired. Check if we have to switch to a new
2165 cfq_choose_cfqg(cfqd);
2167 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2172 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2176 while (cfqq->next_rq) {
2177 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2181 BUG_ON(!list_empty(&cfqq->fifo));
2183 /* By default cfqq is not expired if it is empty. Do it explicitly */
2184 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2189 * Drain our current requests. Used for barriers and when switching
2190 * io schedulers on-the-fly.
2192 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2194 struct cfq_queue *cfqq;
2197 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL)
2198 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2200 cfq_slice_expired(cfqd, 0);
2201 BUG_ON(cfqd->busy_queues);
2203 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2207 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2208 struct cfq_queue *cfqq)
2210 /* the queue hasn't finished any request, can't estimate */
2211 if (cfq_cfqq_slice_new(cfqq))
2213 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2220 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2222 unsigned int max_dispatch;
2225 * Drain async requests before we start sync IO
2227 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2231 * If this is an async queue and we have sync IO in flight, let it wait
2233 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2236 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2237 if (cfq_class_idle(cfqq))
2241 * Does this cfqq already have too much IO in flight?
2243 if (cfqq->dispatched >= max_dispatch) {
2245 * idle queue must always only have a single IO in flight
2247 if (cfq_class_idle(cfqq))
2251 * We have other queues, don't allow more IO from this one
2253 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2257 * Sole queue user, no limit
2259 if (cfqd->busy_queues == 1)
2263 * Normally we start throttling cfqq when cfq_quantum/2
2264 * requests have been dispatched. But we can drive
2265 * deeper queue depths at the beginning of slice
2266 * subjected to upper limit of cfq_quantum.
2268 max_dispatch = cfqd->cfq_quantum;
2272 * Async queues must wait a bit before being allowed dispatch.
2273 * We also ramp up the dispatch depth gradually for async IO,
2274 * based on the last sync IO we serviced
2276 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2277 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2280 depth = last_sync / cfqd->cfq_slice[1];
2281 if (!depth && !cfqq->dispatched)
2283 if (depth < max_dispatch)
2284 max_dispatch = depth;
2288 * If we're below the current max, allow a dispatch
2290 return cfqq->dispatched < max_dispatch;
2294 * Dispatch a request from cfqq, moving them to the request queue
2297 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2301 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2303 if (!cfq_may_dispatch(cfqd, cfqq))
2307 * follow expired path, else get first next available
2309 rq = cfq_check_fifo(cfqq);
2314 * insert request into driver dispatch list
2316 cfq_dispatch_insert(cfqd->queue, rq);
2318 if (!cfqd->active_cic) {
2319 struct cfq_io_context *cic = RQ_CIC(rq);
2321 atomic_long_inc(&cic->ioc->refcount);
2322 cfqd->active_cic = cic;
2329 * Find the cfqq that we need to service and move a request from that to the
2332 static int cfq_dispatch_requests(struct request_queue *q, int force)
2334 struct cfq_data *cfqd = q->elevator->elevator_data;
2335 struct cfq_queue *cfqq;
2337 if (!cfqd->busy_queues)
2340 if (unlikely(force))
2341 return cfq_forced_dispatch(cfqd);
2343 cfqq = cfq_select_queue(cfqd);
2348 * Dispatch a request from this cfqq, if it is allowed
2350 if (!cfq_dispatch_request(cfqd, cfqq))
2353 cfqq->slice_dispatch++;
2354 cfq_clear_cfqq_must_dispatch(cfqq);
2357 * expire an async queue immediately if it has used up its slice. idle
2358 * queue always expire after 1 dispatch round.
2360 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2361 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2362 cfq_class_idle(cfqq))) {
2363 cfqq->slice_end = jiffies + 1;
2364 cfq_slice_expired(cfqd, 0);
2367 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2372 * task holds one reference to the queue, dropped when task exits. each rq
2373 * in-flight on this queue also holds a reference, dropped when rq is freed.
2375 * Each cfq queue took a reference on the parent group. Drop it now.
2376 * queue lock must be held here.
2378 static void cfq_put_queue(struct cfq_queue *cfqq)
2380 struct cfq_data *cfqd = cfqq->cfqd;
2381 struct cfq_group *cfqg, *orig_cfqg;
2383 BUG_ON(atomic_read(&cfqq->ref) <= 0);
2385 if (!atomic_dec_and_test(&cfqq->ref))
2388 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2389 BUG_ON(rb_first(&cfqq->sort_list));
2390 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2392 orig_cfqg = cfqq->orig_cfqg;
2394 if (unlikely(cfqd->active_queue == cfqq)) {
2395 __cfq_slice_expired(cfqd, cfqq, 0);
2396 cfq_schedule_dispatch(cfqd);
2399 BUG_ON(cfq_cfqq_on_rr(cfqq));
2400 kmem_cache_free(cfq_pool, cfqq);
2403 cfq_put_cfqg(orig_cfqg);
2407 * Must always be called with the rcu_read_lock() held
2410 __call_for_each_cic(struct io_context *ioc,
2411 void (*func)(struct io_context *, struct cfq_io_context *))
2413 struct cfq_io_context *cic;
2414 struct hlist_node *n;
2416 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2421 * Call func for each cic attached to this ioc.
2424 call_for_each_cic(struct io_context *ioc,
2425 void (*func)(struct io_context *, struct cfq_io_context *))
2428 __call_for_each_cic(ioc, func);
2432 static void cfq_cic_free_rcu(struct rcu_head *head)
2434 struct cfq_io_context *cic;
2436 cic = container_of(head, struct cfq_io_context, rcu_head);
2438 kmem_cache_free(cfq_ioc_pool, cic);
2439 elv_ioc_count_dec(cfq_ioc_count);
2443 * CFQ scheduler is exiting, grab exit lock and check
2444 * the pending io context count. If it hits zero,
2445 * complete ioc_gone and set it back to NULL
2447 spin_lock(&ioc_gone_lock);
2448 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2452 spin_unlock(&ioc_gone_lock);
2456 static void cfq_cic_free(struct cfq_io_context *cic)
2458 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2461 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2463 unsigned long flags;
2465 BUG_ON(!cic->dead_key);
2467 spin_lock_irqsave(&ioc->lock, flags);
2468 radix_tree_delete(&ioc->radix_root, cic->dead_key);
2469 hlist_del_rcu(&cic->cic_list);
2470 spin_unlock_irqrestore(&ioc->lock, flags);
2476 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2477 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2478 * and ->trim() which is called with the task lock held
2480 static void cfq_free_io_context(struct io_context *ioc)
2483 * ioc->refcount is zero here, or we are called from elv_unregister(),
2484 * so no more cic's are allowed to be linked into this ioc. So it
2485 * should be ok to iterate over the known list, we will see all cic's
2486 * since no new ones are added.
2488 __call_for_each_cic(ioc, cic_free_func);
2491 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2493 struct cfq_queue *__cfqq, *next;
2495 if (unlikely(cfqq == cfqd->active_queue)) {
2496 __cfq_slice_expired(cfqd, cfqq, 0);
2497 cfq_schedule_dispatch(cfqd);
2501 * If this queue was scheduled to merge with another queue, be
2502 * sure to drop the reference taken on that queue (and others in
2503 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2505 __cfqq = cfqq->new_cfqq;
2507 if (__cfqq == cfqq) {
2508 WARN(1, "cfqq->new_cfqq loop detected\n");
2511 next = __cfqq->new_cfqq;
2512 cfq_put_queue(__cfqq);
2516 cfq_put_queue(cfqq);
2519 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2520 struct cfq_io_context *cic)
2522 struct io_context *ioc = cic->ioc;
2524 list_del_init(&cic->queue_list);
2527 * Make sure key == NULL is seen for dead queues
2530 cic->dead_key = (unsigned long) cic->key;
2533 if (ioc->ioc_data == cic)
2534 rcu_assign_pointer(ioc->ioc_data, NULL);
2536 if (cic->cfqq[BLK_RW_ASYNC]) {
2537 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2538 cic->cfqq[BLK_RW_ASYNC] = NULL;
2541 if (cic->cfqq[BLK_RW_SYNC]) {
2542 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2543 cic->cfqq[BLK_RW_SYNC] = NULL;
2547 static void cfq_exit_single_io_context(struct io_context *ioc,
2548 struct cfq_io_context *cic)
2550 struct cfq_data *cfqd = cic->key;
2553 struct request_queue *q = cfqd->queue;
2554 unsigned long flags;
2556 spin_lock_irqsave(q->queue_lock, flags);
2559 * Ensure we get a fresh copy of the ->key to prevent
2560 * race between exiting task and queue
2562 smp_read_barrier_depends();
2564 __cfq_exit_single_io_context(cfqd, cic);
2566 spin_unlock_irqrestore(q->queue_lock, flags);
2571 * The process that ioc belongs to has exited, we need to clean up
2572 * and put the internal structures we have that belongs to that process.
2574 static void cfq_exit_io_context(struct io_context *ioc)
2576 call_for_each_cic(ioc, cfq_exit_single_io_context);
2579 static struct cfq_io_context *
2580 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2582 struct cfq_io_context *cic;
2584 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2587 cic->last_end_request = jiffies;
2588 INIT_LIST_HEAD(&cic->queue_list);
2589 INIT_HLIST_NODE(&cic->cic_list);
2590 cic->dtor = cfq_free_io_context;
2591 cic->exit = cfq_exit_io_context;
2592 elv_ioc_count_inc(cfq_ioc_count);
2598 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2600 struct task_struct *tsk = current;
2603 if (!cfq_cfqq_prio_changed(cfqq))
2606 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2607 switch (ioprio_class) {
2609 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2610 case IOPRIO_CLASS_NONE:
2612 * no prio set, inherit CPU scheduling settings
2614 cfqq->ioprio = task_nice_ioprio(tsk);
2615 cfqq->ioprio_class = task_nice_ioclass(tsk);
2617 case IOPRIO_CLASS_RT:
2618 cfqq->ioprio = task_ioprio(ioc);
2619 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2621 case IOPRIO_CLASS_BE:
2622 cfqq->ioprio = task_ioprio(ioc);
2623 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2625 case IOPRIO_CLASS_IDLE:
2626 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2628 cfq_clear_cfqq_idle_window(cfqq);
2633 * keep track of original prio settings in case we have to temporarily
2634 * elevate the priority of this queue
2636 cfqq->org_ioprio = cfqq->ioprio;
2637 cfqq->org_ioprio_class = cfqq->ioprio_class;
2638 cfq_clear_cfqq_prio_changed(cfqq);
2641 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2643 struct cfq_data *cfqd = cic->key;
2644 struct cfq_queue *cfqq;
2645 unsigned long flags;
2647 if (unlikely(!cfqd))
2650 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2652 cfqq = cic->cfqq[BLK_RW_ASYNC];
2654 struct cfq_queue *new_cfqq;
2655 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2658 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2659 cfq_put_queue(cfqq);
2663 cfqq = cic->cfqq[BLK_RW_SYNC];
2665 cfq_mark_cfqq_prio_changed(cfqq);
2667 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2670 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2672 call_for_each_cic(ioc, changed_ioprio);
2673 ioc->ioprio_changed = 0;
2676 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2677 pid_t pid, bool is_sync)
2679 RB_CLEAR_NODE(&cfqq->rb_node);
2680 RB_CLEAR_NODE(&cfqq->p_node);
2681 INIT_LIST_HEAD(&cfqq->fifo);
2683 atomic_set(&cfqq->ref, 0);
2686 cfq_mark_cfqq_prio_changed(cfqq);
2689 if (!cfq_class_idle(cfqq))
2690 cfq_mark_cfqq_idle_window(cfqq);
2691 cfq_mark_cfqq_sync(cfqq);
2696 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2697 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2699 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2700 struct cfq_data *cfqd = cic->key;
2701 unsigned long flags;
2702 struct request_queue *q;
2704 if (unlikely(!cfqd))
2709 spin_lock_irqsave(q->queue_lock, flags);
2713 * Drop reference to sync queue. A new sync queue will be
2714 * assigned in new group upon arrival of a fresh request.
2716 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2717 cic_set_cfqq(cic, NULL, 1);
2718 cfq_put_queue(sync_cfqq);
2721 spin_unlock_irqrestore(q->queue_lock, flags);
2724 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2726 call_for_each_cic(ioc, changed_cgroup);
2727 ioc->cgroup_changed = 0;
2729 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2731 static struct cfq_queue *
2732 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2733 struct io_context *ioc, gfp_t gfp_mask)
2735 struct cfq_queue *cfqq, *new_cfqq = NULL;
2736 struct cfq_io_context *cic;
2737 struct cfq_group *cfqg;
2740 cfqg = cfq_get_cfqg(cfqd, 1);
2741 cic = cfq_cic_lookup(cfqd, ioc);
2742 /* cic always exists here */
2743 cfqq = cic_to_cfqq(cic, is_sync);
2746 * Always try a new alloc if we fell back to the OOM cfqq
2747 * originally, since it should just be a temporary situation.
2749 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2754 } else if (gfp_mask & __GFP_WAIT) {
2755 spin_unlock_irq(cfqd->queue->queue_lock);
2756 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2757 gfp_mask | __GFP_ZERO,
2759 spin_lock_irq(cfqd->queue->queue_lock);
2763 cfqq = kmem_cache_alloc_node(cfq_pool,
2764 gfp_mask | __GFP_ZERO,
2769 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2770 cfq_init_prio_data(cfqq, ioc);
2771 cfq_link_cfqq_cfqg(cfqq, cfqg);
2772 cfq_log_cfqq(cfqd, cfqq, "alloced");
2774 cfqq = &cfqd->oom_cfqq;
2778 kmem_cache_free(cfq_pool, new_cfqq);
2783 static struct cfq_queue **
2784 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2786 switch (ioprio_class) {
2787 case IOPRIO_CLASS_RT:
2788 return &cfqd->async_cfqq[0][ioprio];
2789 case IOPRIO_CLASS_BE:
2790 return &cfqd->async_cfqq[1][ioprio];
2791 case IOPRIO_CLASS_IDLE:
2792 return &cfqd->async_idle_cfqq;
2798 static struct cfq_queue *
2799 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2802 const int ioprio = task_ioprio(ioc);
2803 const int ioprio_class = task_ioprio_class(ioc);
2804 struct cfq_queue **async_cfqq = NULL;
2805 struct cfq_queue *cfqq = NULL;
2808 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2813 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2816 * pin the queue now that it's allocated, scheduler exit will prune it
2818 if (!is_sync && !(*async_cfqq)) {
2819 atomic_inc(&cfqq->ref);
2823 atomic_inc(&cfqq->ref);
2828 * We drop cfq io contexts lazily, so we may find a dead one.
2831 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2832 struct cfq_io_context *cic)
2834 unsigned long flags;
2836 WARN_ON(!list_empty(&cic->queue_list));
2838 spin_lock_irqsave(&ioc->lock, flags);
2840 BUG_ON(ioc->ioc_data == cic);
2842 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2843 hlist_del_rcu(&cic->cic_list);
2844 spin_unlock_irqrestore(&ioc->lock, flags);
2849 static struct cfq_io_context *
2850 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2852 struct cfq_io_context *cic;
2853 unsigned long flags;
2862 * we maintain a last-hit cache, to avoid browsing over the tree
2864 cic = rcu_dereference(ioc->ioc_data);
2865 if (cic && cic->key == cfqd) {
2871 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2875 /* ->key must be copied to avoid race with cfq_exit_queue() */
2878 cfq_drop_dead_cic(cfqd, ioc, cic);
2883 spin_lock_irqsave(&ioc->lock, flags);
2884 rcu_assign_pointer(ioc->ioc_data, cic);
2885 spin_unlock_irqrestore(&ioc->lock, flags);
2893 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2894 * the process specific cfq io context when entered from the block layer.
2895 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2897 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2898 struct cfq_io_context *cic, gfp_t gfp_mask)
2900 unsigned long flags;
2903 ret = radix_tree_preload(gfp_mask);
2908 spin_lock_irqsave(&ioc->lock, flags);
2909 ret = radix_tree_insert(&ioc->radix_root,
2910 (unsigned long) cfqd, cic);
2912 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2913 spin_unlock_irqrestore(&ioc->lock, flags);
2915 radix_tree_preload_end();
2918 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2919 list_add(&cic->queue_list, &cfqd->cic_list);
2920 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2925 printk(KERN_ERR "cfq: cic link failed!\n");
2931 * Setup general io context and cfq io context. There can be several cfq
2932 * io contexts per general io context, if this process is doing io to more
2933 * than one device managed by cfq.
2935 static struct cfq_io_context *
2936 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2938 struct io_context *ioc = NULL;
2939 struct cfq_io_context *cic;
2941 might_sleep_if(gfp_mask & __GFP_WAIT);
2943 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2947 cic = cfq_cic_lookup(cfqd, ioc);
2951 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2955 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2959 smp_read_barrier_depends();
2960 if (unlikely(ioc->ioprio_changed))
2961 cfq_ioc_set_ioprio(ioc);
2963 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2964 if (unlikely(ioc->cgroup_changed))
2965 cfq_ioc_set_cgroup(ioc);
2971 put_io_context(ioc);
2976 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2978 unsigned long elapsed = jiffies - cic->last_end_request;
2979 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2981 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2982 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2983 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2987 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2991 sector_t n_sec = blk_rq_sectors(rq);
2992 if (cfqq->last_request_pos) {
2993 if (cfqq->last_request_pos < blk_rq_pos(rq))
2994 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2996 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2999 cfqq->seek_history <<= 1;
3000 if (blk_queue_nonrot(cfqd->queue))
3001 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3003 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3007 * Disable idle window if the process thinks too long or seeks so much that
3011 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3012 struct cfq_io_context *cic)
3014 int old_idle, enable_idle;
3017 * Don't idle for async or idle io prio class
3019 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3022 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3024 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3025 cfq_mark_cfqq_deep(cfqq);
3027 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3028 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3030 else if (sample_valid(cic->ttime_samples)) {
3031 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3037 if (old_idle != enable_idle) {
3038 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3040 cfq_mark_cfqq_idle_window(cfqq);
3042 cfq_clear_cfqq_idle_window(cfqq);
3047 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3048 * no or if we aren't sure, a 1 will cause a preempt.
3051 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3054 struct cfq_queue *cfqq;
3056 cfqq = cfqd->active_queue;
3060 if (cfq_class_idle(new_cfqq))
3063 if (cfq_class_idle(cfqq))
3067 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3069 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3073 * if the new request is sync, but the currently running queue is
3074 * not, let the sync request have priority.
3076 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3079 if (new_cfqq->cfqg != cfqq->cfqg)
3082 if (cfq_slice_used(cfqq))
3085 /* Allow preemption only if we are idling on sync-noidle tree */
3086 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3087 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3088 new_cfqq->service_tree->count == 2 &&
3089 RB_EMPTY_ROOT(&cfqq->sort_list))
3093 * So both queues are sync. Let the new request get disk time if
3094 * it's a metadata request and the current queue is doing regular IO.
3096 if (rq_is_meta(rq) && !cfqq->meta_pending)
3100 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3102 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3105 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3109 * if this request is as-good as one we would expect from the
3110 * current cfqq, let it preempt
3112 if (cfq_rq_close(cfqd, cfqq, rq))
3119 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3120 * let it have half of its nominal slice.
3122 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3124 cfq_log_cfqq(cfqd, cfqq, "preempt");
3125 cfq_slice_expired(cfqd, 1);
3128 * Put the new queue at the front of the of the current list,
3129 * so we know that it will be selected next.
3131 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3133 cfq_service_tree_add(cfqd, cfqq, 1);
3135 cfqq->slice_end = 0;
3136 cfq_mark_cfqq_slice_new(cfqq);
3140 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3141 * something we should do about it
3144 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3147 struct cfq_io_context *cic = RQ_CIC(rq);
3151 cfqq->meta_pending++;
3153 cfq_update_io_thinktime(cfqd, cic);
3154 cfq_update_io_seektime(cfqd, cfqq, rq);
3155 cfq_update_idle_window(cfqd, cfqq, cic);
3157 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3159 if (cfqq == cfqd->active_queue) {
3161 * Remember that we saw a request from this process, but
3162 * don't start queuing just yet. Otherwise we risk seeing lots
3163 * of tiny requests, because we disrupt the normal plugging
3164 * and merging. If the request is already larger than a single
3165 * page, let it rip immediately. For that case we assume that
3166 * merging is already done. Ditto for a busy system that
3167 * has other work pending, don't risk delaying until the
3168 * idle timer unplug to continue working.
3170 if (cfq_cfqq_wait_request(cfqq)) {
3171 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3172 cfqd->busy_queues > 1) {
3173 del_timer(&cfqd->idle_slice_timer);
3174 cfq_clear_cfqq_wait_request(cfqq);
3175 __blk_run_queue(cfqd->queue);
3177 cfq_mark_cfqq_must_dispatch(cfqq);
3179 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3181 * not the active queue - expire current slice if it is
3182 * idle and has expired it's mean thinktime or this new queue
3183 * has some old slice time left and is of higher priority or
3184 * this new queue is RT and the current one is BE
3186 cfq_preempt_queue(cfqd, cfqq);
3187 __blk_run_queue(cfqd->queue);
3191 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3193 struct cfq_data *cfqd = q->elevator->elevator_data;
3194 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3196 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3197 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3199 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3200 list_add_tail(&rq->queuelist, &cfqq->fifo);
3203 cfq_rq_enqueued(cfqd, cfqq, rq);
3207 * Update hw_tag based on peak queue depth over 50 samples under
3210 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3212 struct cfq_queue *cfqq = cfqd->active_queue;
3214 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3215 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3217 if (cfqd->hw_tag == 1)
3220 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3221 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3225 * If active queue hasn't enough requests and can idle, cfq might not
3226 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3229 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3230 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3231 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3234 if (cfqd->hw_tag_samples++ < 50)
3237 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3243 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3245 struct cfq_io_context *cic = cfqd->active_cic;
3247 /* If there are other queues in the group, don't wait */
3248 if (cfqq->cfqg->nr_cfqq > 1)
3251 if (cfq_slice_used(cfqq))
3254 /* if slice left is less than think time, wait busy */
3255 if (cic && sample_valid(cic->ttime_samples)
3256 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3260 * If think times is less than a jiffy than ttime_mean=0 and above
3261 * will not be true. It might happen that slice has not expired yet
3262 * but will expire soon (4-5 ns) during select_queue(). To cover the
3263 * case where think time is less than a jiffy, mark the queue wait
3264 * busy if only 1 jiffy is left in the slice.
3266 if (cfqq->slice_end - jiffies == 1)
3272 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3274 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3275 struct cfq_data *cfqd = cfqq->cfqd;
3276 const int sync = rq_is_sync(rq);
3280 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3282 cfq_update_hw_tag(cfqd);
3284 WARN_ON(!cfqd->rq_in_driver);
3285 WARN_ON(!cfqq->dispatched);
3286 cfqd->rq_in_driver--;
3289 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3292 RQ_CIC(rq)->last_end_request = now;
3293 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3294 cfqd->last_delayed_sync = now;
3298 * If this is the active queue, check if it needs to be expired,
3299 * or if we want to idle in case it has no pending requests.
3301 if (cfqd->active_queue == cfqq) {
3302 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3304 if (cfq_cfqq_slice_new(cfqq)) {
3305 cfq_set_prio_slice(cfqd, cfqq);
3306 cfq_clear_cfqq_slice_new(cfqq);
3310 * Should we wait for next request to come in before we expire
3313 if (cfq_should_wait_busy(cfqd, cfqq)) {
3314 cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3315 cfq_mark_cfqq_wait_busy(cfqq);
3316 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3320 * Idling is not enabled on:
3322 * - idle-priority queues
3324 * - queues with still some requests queued
3325 * - when there is a close cooperator
3327 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3328 cfq_slice_expired(cfqd, 1);
3329 else if (sync && cfqq_empty &&
3330 !cfq_close_cooperator(cfqd, cfqq)) {
3331 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3333 * Idling is enabled for SYNC_WORKLOAD.
3334 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3335 * only if we processed at least one !rq_noidle request
3337 if (cfqd->serving_type == SYNC_WORKLOAD
3338 || cfqd->noidle_tree_requires_idle
3339 || cfqq->cfqg->nr_cfqq == 1)
3340 cfq_arm_slice_timer(cfqd);
3344 if (!cfqd->rq_in_driver)
3345 cfq_schedule_dispatch(cfqd);
3349 * we temporarily boost lower priority queues if they are holding fs exclusive
3350 * resources. they are boosted to normal prio (CLASS_BE/4)
3352 static void cfq_prio_boost(struct cfq_queue *cfqq)
3354 if (has_fs_excl()) {
3356 * boost idle prio on transactions that would lock out other
3357 * users of the filesystem
3359 if (cfq_class_idle(cfqq))
3360 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3361 if (cfqq->ioprio > IOPRIO_NORM)
3362 cfqq->ioprio = IOPRIO_NORM;
3365 * unboost the queue (if needed)
3367 cfqq->ioprio_class = cfqq->org_ioprio_class;
3368 cfqq->ioprio = cfqq->org_ioprio;
3372 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3374 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3375 cfq_mark_cfqq_must_alloc_slice(cfqq);
3376 return ELV_MQUEUE_MUST;
3379 return ELV_MQUEUE_MAY;
3382 static int cfq_may_queue(struct request_queue *q, int rw)
3384 struct cfq_data *cfqd = q->elevator->elevator_data;
3385 struct task_struct *tsk = current;
3386 struct cfq_io_context *cic;
3387 struct cfq_queue *cfqq;
3390 * don't force setup of a queue from here, as a call to may_queue
3391 * does not necessarily imply that a request actually will be queued.
3392 * so just lookup a possibly existing queue, or return 'may queue'
3395 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3397 return ELV_MQUEUE_MAY;
3399 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3401 cfq_init_prio_data(cfqq, cic->ioc);
3402 cfq_prio_boost(cfqq);
3404 return __cfq_may_queue(cfqq);
3407 return ELV_MQUEUE_MAY;
3411 * queue lock held here
3413 static void cfq_put_request(struct request *rq)
3415 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3418 const int rw = rq_data_dir(rq);
3420 BUG_ON(!cfqq->allocated[rw]);
3421 cfqq->allocated[rw]--;
3423 put_io_context(RQ_CIC(rq)->ioc);
3425 rq->elevator_private = NULL;
3426 rq->elevator_private2 = NULL;
3428 cfq_put_queue(cfqq);
3432 static struct cfq_queue *
3433 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3434 struct cfq_queue *cfqq)
3436 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3437 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3438 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3439 cfq_put_queue(cfqq);
3440 return cic_to_cfqq(cic, 1);
3444 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3445 * was the last process referring to said cfqq.
3447 static struct cfq_queue *
3448 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3450 if (cfqq_process_refs(cfqq) == 1) {
3451 cfqq->pid = current->pid;
3452 cfq_clear_cfqq_coop(cfqq);
3453 cfq_clear_cfqq_split_coop(cfqq);
3457 cic_set_cfqq(cic, NULL, 1);
3458 cfq_put_queue(cfqq);
3462 * Allocate cfq data structures associated with this request.
3465 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3467 struct cfq_data *cfqd = q->elevator->elevator_data;
3468 struct cfq_io_context *cic;
3469 const int rw = rq_data_dir(rq);
3470 const bool is_sync = rq_is_sync(rq);
3471 struct cfq_queue *cfqq;
3472 unsigned long flags;
3474 might_sleep_if(gfp_mask & __GFP_WAIT);
3476 cic = cfq_get_io_context(cfqd, gfp_mask);
3478 spin_lock_irqsave(q->queue_lock, flags);
3484 cfqq = cic_to_cfqq(cic, is_sync);
3485 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3486 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3487 cic_set_cfqq(cic, cfqq, is_sync);
3490 * If the queue was seeky for too long, break it apart.
3492 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3493 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3494 cfqq = split_cfqq(cic, cfqq);
3500 * Check to see if this queue is scheduled to merge with
3501 * another, closely cooperating queue. The merging of
3502 * queues happens here as it must be done in process context.
3503 * The reference on new_cfqq was taken in merge_cfqqs.
3506 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3509 cfqq->allocated[rw]++;
3510 atomic_inc(&cfqq->ref);
3512 spin_unlock_irqrestore(q->queue_lock, flags);
3514 rq->elevator_private = cic;
3515 rq->elevator_private2 = cfqq;
3520 put_io_context(cic->ioc);
3522 cfq_schedule_dispatch(cfqd);
3523 spin_unlock_irqrestore(q->queue_lock, flags);
3524 cfq_log(cfqd, "set_request fail");
3528 static void cfq_kick_queue(struct work_struct *work)
3530 struct cfq_data *cfqd =
3531 container_of(work, struct cfq_data, unplug_work);
3532 struct request_queue *q = cfqd->queue;
3534 spin_lock_irq(q->queue_lock);
3535 __blk_run_queue(cfqd->queue);
3536 spin_unlock_irq(q->queue_lock);
3540 * Timer running if the active_queue is currently idling inside its time slice
3542 static void cfq_idle_slice_timer(unsigned long data)
3544 struct cfq_data *cfqd = (struct cfq_data *) data;
3545 struct cfq_queue *cfqq;
3546 unsigned long flags;
3549 cfq_log(cfqd, "idle timer fired");
3551 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3553 cfqq = cfqd->active_queue;
3558 * We saw a request before the queue expired, let it through
3560 if (cfq_cfqq_must_dispatch(cfqq))
3566 if (cfq_slice_used(cfqq))
3570 * only expire and reinvoke request handler, if there are
3571 * other queues with pending requests
3573 if (!cfqd->busy_queues)
3577 * not expired and it has a request pending, let it dispatch
3579 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3583 * Queue depth flag is reset only when the idle didn't succeed
3585 cfq_clear_cfqq_deep(cfqq);
3588 cfq_slice_expired(cfqd, timed_out);
3590 cfq_schedule_dispatch(cfqd);
3592 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3595 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3597 del_timer_sync(&cfqd->idle_slice_timer);
3598 cancel_work_sync(&cfqd->unplug_work);
3601 static void cfq_put_async_queues(struct cfq_data *cfqd)
3605 for (i = 0; i < IOPRIO_BE_NR; i++) {
3606 if (cfqd->async_cfqq[0][i])
3607 cfq_put_queue(cfqd->async_cfqq[0][i]);
3608 if (cfqd->async_cfqq[1][i])
3609 cfq_put_queue(cfqd->async_cfqq[1][i]);
3612 if (cfqd->async_idle_cfqq)
3613 cfq_put_queue(cfqd->async_idle_cfqq);
3616 static void cfq_cfqd_free(struct rcu_head *head)
3618 kfree(container_of(head, struct cfq_data, rcu));
3621 static void cfq_exit_queue(struct elevator_queue *e)
3623 struct cfq_data *cfqd = e->elevator_data;
3624 struct request_queue *q = cfqd->queue;
3626 cfq_shutdown_timer_wq(cfqd);
3628 spin_lock_irq(q->queue_lock);
3630 if (cfqd->active_queue)
3631 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3633 while (!list_empty(&cfqd->cic_list)) {
3634 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3635 struct cfq_io_context,
3638 __cfq_exit_single_io_context(cfqd, cic);
3641 cfq_put_async_queues(cfqd);
3642 cfq_release_cfq_groups(cfqd);
3643 blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3645 spin_unlock_irq(q->queue_lock);
3647 cfq_shutdown_timer_wq(cfqd);
3649 /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3650 call_rcu(&cfqd->rcu, cfq_cfqd_free);
3653 static void *cfq_init_queue(struct request_queue *q)
3655 struct cfq_data *cfqd;
3657 struct cfq_group *cfqg;
3658 struct cfq_rb_root *st;
3660 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3664 /* Init root service tree */
3665 cfqd->grp_service_tree = CFQ_RB_ROOT;
3667 /* Init root group */
3668 cfqg = &cfqd->root_group;
3669 for_each_cfqg_st(cfqg, i, j, st)
3671 RB_CLEAR_NODE(&cfqg->rb_node);
3673 /* Give preference to root group over other groups */
3674 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3676 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3678 * Take a reference to root group which we never drop. This is just
3679 * to make sure that cfq_put_cfqg() does not try to kfree root group
3681 atomic_set(&cfqg->ref, 1);
3682 blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3686 * Not strictly needed (since RB_ROOT just clears the node and we
3687 * zeroed cfqd on alloc), but better be safe in case someone decides
3688 * to add magic to the rb code
3690 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3691 cfqd->prio_trees[i] = RB_ROOT;
3694 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3695 * Grab a permanent reference to it, so that the normal code flow
3696 * will not attempt to free it.
3698 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3699 atomic_inc(&cfqd->oom_cfqq.ref);
3700 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3702 INIT_LIST_HEAD(&cfqd->cic_list);
3706 init_timer(&cfqd->idle_slice_timer);
3707 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3708 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3710 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3712 cfqd->cfq_quantum = cfq_quantum;
3713 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3714 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3715 cfqd->cfq_back_max = cfq_back_max;
3716 cfqd->cfq_back_penalty = cfq_back_penalty;
3717 cfqd->cfq_slice[0] = cfq_slice_async;
3718 cfqd->cfq_slice[1] = cfq_slice_sync;
3719 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3720 cfqd->cfq_slice_idle = cfq_slice_idle;
3721 cfqd->cfq_latency = 1;
3722 cfqd->cfq_group_isolation = 0;
3725 * we optimistically start assuming sync ops weren't delayed in last
3726 * second, in order to have larger depth for async operations.
3728 cfqd->last_delayed_sync = jiffies - HZ;
3729 INIT_RCU_HEAD(&cfqd->rcu);
3733 static void cfq_slab_kill(void)
3736 * Caller already ensured that pending RCU callbacks are completed,
3737 * so we should have no busy allocations at this point.
3740 kmem_cache_destroy(cfq_pool);
3742 kmem_cache_destroy(cfq_ioc_pool);
3745 static int __init cfq_slab_setup(void)
3747 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3751 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3762 * sysfs parts below -->
3765 cfq_var_show(unsigned int var, char *page)
3767 return sprintf(page, "%d\n", var);
3771 cfq_var_store(unsigned int *var, const char *page, size_t count)
3773 char *p = (char *) page;
3775 *var = simple_strtoul(p, &p, 10);
3779 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3780 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3782 struct cfq_data *cfqd = e->elevator_data; \
3783 unsigned int __data = __VAR; \
3785 __data = jiffies_to_msecs(__data); \
3786 return cfq_var_show(__data, (page)); \
3788 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3789 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3790 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3791 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3792 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3793 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3794 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3795 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3796 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3797 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3798 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3799 #undef SHOW_FUNCTION
3801 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3802 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3804 struct cfq_data *cfqd = e->elevator_data; \
3805 unsigned int __data; \
3806 int ret = cfq_var_store(&__data, (page), count); \
3807 if (__data < (MIN)) \
3809 else if (__data > (MAX)) \
3812 *(__PTR) = msecs_to_jiffies(__data); \
3814 *(__PTR) = __data; \
3817 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3818 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3820 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3822 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3823 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3825 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3826 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3827 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3828 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3830 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3831 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3832 #undef STORE_FUNCTION
3834 #define CFQ_ATTR(name) \
3835 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3837 static struct elv_fs_entry cfq_attrs[] = {
3839 CFQ_ATTR(fifo_expire_sync),
3840 CFQ_ATTR(fifo_expire_async),
3841 CFQ_ATTR(back_seek_max),
3842 CFQ_ATTR(back_seek_penalty),
3843 CFQ_ATTR(slice_sync),
3844 CFQ_ATTR(slice_async),
3845 CFQ_ATTR(slice_async_rq),
3846 CFQ_ATTR(slice_idle),
3847 CFQ_ATTR(low_latency),
3848 CFQ_ATTR(group_isolation),
3852 static struct elevator_type iosched_cfq = {
3854 .elevator_merge_fn = cfq_merge,
3855 .elevator_merged_fn = cfq_merged_request,
3856 .elevator_merge_req_fn = cfq_merged_requests,
3857 .elevator_allow_merge_fn = cfq_allow_merge,
3858 .elevator_dispatch_fn = cfq_dispatch_requests,
3859 .elevator_add_req_fn = cfq_insert_request,
3860 .elevator_activate_req_fn = cfq_activate_request,
3861 .elevator_deactivate_req_fn = cfq_deactivate_request,
3862 .elevator_queue_empty_fn = cfq_queue_empty,
3863 .elevator_completed_req_fn = cfq_completed_request,
3864 .elevator_former_req_fn = elv_rb_former_request,
3865 .elevator_latter_req_fn = elv_rb_latter_request,
3866 .elevator_set_req_fn = cfq_set_request,
3867 .elevator_put_req_fn = cfq_put_request,
3868 .elevator_may_queue_fn = cfq_may_queue,
3869 .elevator_init_fn = cfq_init_queue,
3870 .elevator_exit_fn = cfq_exit_queue,
3871 .trim = cfq_free_io_context,
3873 .elevator_attrs = cfq_attrs,
3874 .elevator_name = "cfq",
3875 .elevator_owner = THIS_MODULE,
3878 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3879 static struct blkio_policy_type blkio_policy_cfq = {
3881 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3882 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3886 static struct blkio_policy_type blkio_policy_cfq;
3889 static int __init cfq_init(void)
3892 * could be 0 on HZ < 1000 setups
3894 if (!cfq_slice_async)
3895 cfq_slice_async = 1;
3896 if (!cfq_slice_idle)
3899 if (cfq_slab_setup())
3902 elv_register(&iosched_cfq);
3903 blkio_policy_register(&blkio_policy_cfq);
3908 static void __exit cfq_exit(void)
3910 DECLARE_COMPLETION_ONSTACK(all_gone);
3911 blkio_policy_unregister(&blkio_policy_cfq);
3912 elv_unregister(&iosched_cfq);
3913 ioc_gone = &all_gone;
3914 /* ioc_gone's update must be visible before reading ioc_count */
3918 * this also protects us from entering cfq_slab_kill() with
3919 * pending RCU callbacks
3921 if (elv_ioc_count_read(cfq_ioc_count))
3922 wait_for_completion(&all_gone);
3926 module_init(cfq_init);
3927 module_exit(cfq_exit);
3929 MODULE_AUTHOR("Jens Axboe");
3930 MODULE_LICENSE("GPL");
3931 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");