2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
38 * offset from end of service tree
40 #define CFQ_IDLE_DELAY (HZ / 5)
43 * below this threshold, we consider thinktime immediate
45 #define CFQ_MIN_TT (2)
47 #define CFQ_SLICE_SCALE (5)
48 #define CFQ_HW_QUEUE_MIN (5)
49 #define CFQ_SERVICE_SHIFT 12
51 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
71 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
75 #define sample_valid(samples) ((samples) > 80)
76 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
88 unsigned total_weight;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
95 * Per process-grouping structure
100 /* various state flags, see below */
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
118 /* currently allocated requests */
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
125 unsigned int allocated_slice;
126 unsigned int slice_dispatch;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
129 unsigned long slice_end;
132 /* pending metadata requests */
134 /* number of requests that are on the dispatch list or inside driver */
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
144 sector_t last_request_pos;
146 struct cfq_rb_root *service_tree;
147 struct cfq_queue *new_cfqq;
148 struct cfq_group *cfqg;
149 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors;
154 * First index in the service_trees.
155 * IDLE is handled separately, so it has negative index
165 * Second index in the service_trees.
169 SYNC_NOIDLE_WORKLOAD = 1,
173 /* This is per cgroup per device grouping structure */
175 /* group service_tree member */
176 struct rb_node rb_node;
178 /* group service_tree key */
181 unsigned int new_weight;
184 /* number of cfqq currently on this group */
188 * Per group busy queus average. Useful for workload slice calc. We
189 * create the array for each prio class but at run time it is used
190 * only for RT and BE class and slot for IDLE class remains unused.
191 * This is primarily done to avoid confusion and a gcc warning.
193 unsigned int busy_queues_avg[CFQ_PRIO_NR];
195 * rr lists of queues with requests. We maintain service trees for
196 * RT and BE classes. These trees are subdivided in subclasses
197 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198 * class there is no subclassification and all the cfq queues go on
199 * a single tree service_tree_idle.
200 * Counts are embedded in the cfq_rb_root
202 struct cfq_rb_root service_trees[2][3];
203 struct cfq_rb_root service_tree_idle;
205 unsigned long saved_workload_slice;
206 enum wl_type_t saved_workload;
207 enum wl_prio_t saved_serving_prio;
208 struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210 struct hlist_node cfqd_node;
213 /* number of requests that are on the dispatch list or inside driver */
218 * Per block device queue structure
221 struct request_queue *queue;
222 /* Root service tree for cfq_groups */
223 struct cfq_rb_root grp_service_tree;
224 struct cfq_group root_group;
227 * The priority currently being served
229 enum wl_prio_t serving_prio;
230 enum wl_type_t serving_type;
231 unsigned long workload_expires;
232 struct cfq_group *serving_group;
235 * Each priority tree is sorted by next_request position. These
236 * trees are used when determining if two or more queues are
237 * interleaving requests (see cfq_close_cooperator).
239 struct rb_root prio_trees[CFQ_PRIO_LISTS];
241 unsigned int busy_queues;
242 unsigned int busy_sync_queues;
248 * queue-depth detection
254 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
258 int hw_tag_est_depth;
259 unsigned int hw_tag_samples;
262 * idle window management
264 struct timer_list idle_slice_timer;
265 struct work_struct unplug_work;
267 struct cfq_queue *active_queue;
268 struct cfq_io_context *active_cic;
271 * async queue for each priority case
273 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274 struct cfq_queue *async_idle_cfqq;
276 sector_t last_position;
279 * tunables, see top of file
281 unsigned int cfq_quantum;
282 unsigned int cfq_fifo_expire[2];
283 unsigned int cfq_back_penalty;
284 unsigned int cfq_back_max;
285 unsigned int cfq_slice[2];
286 unsigned int cfq_slice_async_rq;
287 unsigned int cfq_slice_idle;
288 unsigned int cfq_group_idle;
289 unsigned int cfq_latency;
291 unsigned int cic_index;
292 struct list_head cic_list;
295 * Fallback dummy cfqq for extreme OOM conditions
297 struct cfq_queue oom_cfqq;
299 unsigned long last_delayed_sync;
301 /* List of cfq groups being managed on this device*/
302 struct hlist_head cfqg_list;
304 /* Number of groups which are on blkcg->blkg_list */
305 unsigned int nr_blkcg_linked_grps;
308 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
310 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
317 if (prio == IDLE_WORKLOAD)
318 return &cfqg->service_tree_idle;
320 return &cfqg->service_trees[prio][type];
323 enum cfqq_state_flags {
324 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
325 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
326 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
327 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
328 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
329 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
330 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
331 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
332 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
333 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
334 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
335 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
336 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
339 #define CFQ_CFQQ_FNS(name) \
340 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
342 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
344 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
346 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
348 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
350 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
354 CFQ_CFQQ_FNS(wait_request);
355 CFQ_CFQQ_FNS(must_dispatch);
356 CFQ_CFQQ_FNS(must_alloc_slice);
357 CFQ_CFQQ_FNS(fifo_expire);
358 CFQ_CFQQ_FNS(idle_window);
359 CFQ_CFQQ_FNS(prio_changed);
360 CFQ_CFQQ_FNS(slice_new);
363 CFQ_CFQQ_FNS(split_coop);
365 CFQ_CFQQ_FNS(wait_busy);
368 #ifdef CONFIG_CFQ_GROUP_IOSCHED
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
371 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
372 blkg_path(&(cfqq)->cfqg->blkg), ##args);
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
376 blkg_path(&(cfqg)->blkg), ##args); \
379 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0);
383 #define cfq_log(cfqd, fmt, args...) \
384 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
386 /* Traverses through cfq group service trees */
387 #define for_each_cfqg_st(cfqg, i, j, st) \
388 for (i = 0; i <= IDLE_WORKLOAD; i++) \
389 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
390 : &cfqg->service_tree_idle; \
391 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
392 (i == IDLE_WORKLOAD && j == 0); \
393 j++, st = i < IDLE_WORKLOAD ? \
394 &cfqg->service_trees[i][j]: NULL) \
397 static inline bool iops_mode(struct cfq_data *cfqd)
400 * If we are not idling on queues and it is a NCQ drive, parallel
401 * execution of requests is on and measuring time is not possible
402 * in most of the cases until and unless we drive shallower queue
403 * depths and that becomes a performance bottleneck. In such cases
404 * switch to start providing fairness in terms of number of IOs.
406 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
412 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
414 if (cfq_class_idle(cfqq))
415 return IDLE_WORKLOAD;
416 if (cfq_class_rt(cfqq))
422 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
424 if (!cfq_cfqq_sync(cfqq))
425 return ASYNC_WORKLOAD;
426 if (!cfq_cfqq_idle_window(cfqq))
427 return SYNC_NOIDLE_WORKLOAD;
428 return SYNC_WORKLOAD;
431 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
432 struct cfq_data *cfqd,
433 struct cfq_group *cfqg)
435 if (wl == IDLE_WORKLOAD)
436 return cfqg->service_tree_idle.count;
438 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
439 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
440 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
443 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
444 struct cfq_group *cfqg)
446 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
447 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
450 static void cfq_dispatch_insert(struct request_queue *, struct request *);
451 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
452 struct io_context *, gfp_t);
453 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
454 struct io_context *);
456 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
459 return cic->cfqq[is_sync];
462 static inline void cic_set_cfqq(struct cfq_io_context *cic,
463 struct cfq_queue *cfqq, bool is_sync)
465 cic->cfqq[is_sync] = cfqq;
468 #define CIC_DEAD_KEY 1ul
469 #define CIC_DEAD_INDEX_SHIFT 1
471 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
473 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
476 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
478 struct cfq_data *cfqd = cic->key;
480 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
487 * We regard a request as SYNC, if it's either a read or has the SYNC bit
488 * set (in which case it could also be direct WRITE).
490 static inline bool cfq_bio_sync(struct bio *bio)
492 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
496 * scheduler run of queue, if there are requests pending and no one in the
497 * driver that will restart queueing
499 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
501 if (cfqd->busy_queues) {
502 cfq_log(cfqd, "schedule dispatch");
503 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
508 * Scale schedule slice based on io priority. Use the sync time slice only
509 * if a queue is marked sync and has sync io queued. A sync queue with async
510 * io only, should not get full sync slice length.
512 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
515 const int base_slice = cfqd->cfq_slice[sync];
517 WARN_ON(prio >= IOPRIO_BE_NR);
519 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
523 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
525 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
528 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
530 u64 d = delta << CFQ_SERVICE_SHIFT;
532 d = d * BLKIO_WEIGHT_DEFAULT;
533 do_div(d, cfqg->weight);
537 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
539 s64 delta = (s64)(vdisktime - min_vdisktime);
541 min_vdisktime = vdisktime;
543 return min_vdisktime;
546 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
548 s64 delta = (s64)(vdisktime - min_vdisktime);
550 min_vdisktime = vdisktime;
552 return min_vdisktime;
555 static void update_min_vdisktime(struct cfq_rb_root *st)
557 struct cfq_group *cfqg;
560 cfqg = rb_entry_cfqg(st->left);
561 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
567 * get averaged number of queues of RT/BE priority.
568 * average is updated, with a formula that gives more weight to higher numbers,
569 * to quickly follows sudden increases and decrease slowly
572 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
573 struct cfq_group *cfqg, bool rt)
575 unsigned min_q, max_q;
576 unsigned mult = cfq_hist_divisor - 1;
577 unsigned round = cfq_hist_divisor / 2;
578 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
580 min_q = min(cfqg->busy_queues_avg[rt], busy);
581 max_q = max(cfqg->busy_queues_avg[rt], busy);
582 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
584 return cfqg->busy_queues_avg[rt];
587 static inline unsigned
588 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
590 struct cfq_rb_root *st = &cfqd->grp_service_tree;
592 return cfq_target_latency * cfqg->weight / st->total_weight;
595 static inline unsigned
596 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
598 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
599 if (cfqd->cfq_latency) {
601 * interested queues (we consider only the ones with the same
602 * priority class in the cfq group)
604 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
606 unsigned sync_slice = cfqd->cfq_slice[1];
607 unsigned expect_latency = sync_slice * iq;
608 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
610 if (expect_latency > group_slice) {
611 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
612 /* scale low_slice according to IO priority
613 * and sync vs async */
615 min(slice, base_low_slice * slice / sync_slice);
616 /* the adapted slice value is scaled to fit all iqs
617 * into the target latency */
618 slice = max(slice * group_slice / expect_latency,
626 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
628 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
630 cfqq->slice_start = jiffies;
631 cfqq->slice_end = jiffies + slice;
632 cfqq->allocated_slice = slice;
633 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
637 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
638 * isn't valid until the first request from the dispatch is activated
639 * and the slice time set.
641 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
643 if (cfq_cfqq_slice_new(cfqq))
645 if (time_before(jiffies, cfqq->slice_end))
652 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
653 * We choose the request that is closest to the head right now. Distance
654 * behind the head is penalized and only allowed to a certain extent.
656 static struct request *
657 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
659 sector_t s1, s2, d1 = 0, d2 = 0;
660 unsigned long back_max;
661 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
662 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
663 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
665 if (rq1 == NULL || rq1 == rq2)
670 if (rq_is_sync(rq1) != rq_is_sync(rq2))
671 return rq_is_sync(rq1) ? rq1 : rq2;
673 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META)
674 return rq1->cmd_flags & REQ_META ? rq1 : rq2;
676 s1 = blk_rq_pos(rq1);
677 s2 = blk_rq_pos(rq2);
680 * by definition, 1KiB is 2 sectors
682 back_max = cfqd->cfq_back_max * 2;
685 * Strict one way elevator _except_ in the case where we allow
686 * short backward seeks which are biased as twice the cost of a
687 * similar forward seek.
691 else if (s1 + back_max >= last)
692 d1 = (last - s1) * cfqd->cfq_back_penalty;
694 wrap |= CFQ_RQ1_WRAP;
698 else if (s2 + back_max >= last)
699 d2 = (last - s2) * cfqd->cfq_back_penalty;
701 wrap |= CFQ_RQ2_WRAP;
703 /* Found required data */
706 * By doing switch() on the bit mask "wrap" we avoid having to
707 * check two variables for all permutations: --> faster!
710 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
726 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
729 * Since both rqs are wrapped,
730 * start with the one that's further behind head
731 * (--> only *one* back seek required),
732 * since back seek takes more time than forward.
742 * The below is leftmost cache rbtree addon
744 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
746 /* Service tree is empty */
751 root->left = rb_first(&root->rb);
754 return rb_entry(root->left, struct cfq_queue, rb_node);
759 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
762 root->left = rb_first(&root->rb);
765 return rb_entry_cfqg(root->left);
770 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
776 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
780 rb_erase_init(n, &root->rb);
785 * would be nice to take fifo expire time into account as well
787 static struct request *
788 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789 struct request *last)
791 struct rb_node *rbnext = rb_next(&last->rb_node);
792 struct rb_node *rbprev = rb_prev(&last->rb_node);
793 struct request *next = NULL, *prev = NULL;
795 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
798 prev = rb_entry_rq(rbprev);
801 next = rb_entry_rq(rbnext);
803 rbnext = rb_first(&cfqq->sort_list);
804 if (rbnext && rbnext != &last->rb_node)
805 next = rb_entry_rq(rbnext);
808 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
811 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812 struct cfq_queue *cfqq)
815 * just an approximation, should be ok.
817 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
818 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
822 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
824 return cfqg->vdisktime - st->min_vdisktime;
828 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
830 struct rb_node **node = &st->rb.rb_node;
831 struct rb_node *parent = NULL;
832 struct cfq_group *__cfqg;
833 s64 key = cfqg_key(st, cfqg);
836 while (*node != NULL) {
838 __cfqg = rb_entry_cfqg(parent);
840 if (key < cfqg_key(st, __cfqg))
841 node = &parent->rb_left;
843 node = &parent->rb_right;
849 st->left = &cfqg->rb_node;
851 rb_link_node(&cfqg->rb_node, parent, node);
852 rb_insert_color(&cfqg->rb_node, &st->rb);
856 cfq_update_group_weight(struct cfq_group *cfqg)
858 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
859 if (cfqg->needs_update) {
860 cfqg->weight = cfqg->new_weight;
861 cfqg->needs_update = false;
866 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
868 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
870 cfq_update_group_weight(cfqg);
871 __cfq_group_service_tree_add(st, cfqg);
872 st->total_weight += cfqg->weight;
876 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
878 struct cfq_rb_root *st = &cfqd->grp_service_tree;
879 struct cfq_group *__cfqg;
883 if (!RB_EMPTY_NODE(&cfqg->rb_node))
887 * Currently put the group at the end. Later implement something
888 * so that groups get lesser vtime based on their weights, so that
889 * if group does not loose all if it was not continuously backlogged.
891 n = rb_last(&st->rb);
893 __cfqg = rb_entry_cfqg(n);
894 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
896 cfqg->vdisktime = st->min_vdisktime;
897 cfq_group_service_tree_add(st, cfqg);
901 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
903 st->total_weight -= cfqg->weight;
904 if (!RB_EMPTY_NODE(&cfqg->rb_node))
905 cfq_rb_erase(&cfqg->rb_node, st);
909 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
911 struct cfq_rb_root *st = &cfqd->grp_service_tree;
913 BUG_ON(cfqg->nr_cfqq < 1);
916 /* If there are other cfq queues under this group, don't delete it */
920 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
921 cfq_group_service_tree_del(st, cfqg);
922 cfqg->saved_workload_slice = 0;
923 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
926 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
927 unsigned int *unaccounted_time)
929 unsigned int slice_used;
932 * Queue got expired before even a single request completed or
933 * got expired immediately after first request completion.
935 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
937 * Also charge the seek time incurred to the group, otherwise
938 * if there are mutiple queues in the group, each can dispatch
939 * a single request on seeky media and cause lots of seek time
940 * and group will never know it.
942 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
945 slice_used = jiffies - cfqq->slice_start;
946 if (slice_used > cfqq->allocated_slice) {
947 *unaccounted_time = slice_used - cfqq->allocated_slice;
948 slice_used = cfqq->allocated_slice;
950 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
951 *unaccounted_time += cfqq->slice_start -
952 cfqq->dispatch_start;
958 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
959 struct cfq_queue *cfqq)
961 struct cfq_rb_root *st = &cfqd->grp_service_tree;
962 unsigned int used_sl, charge, unaccounted_sl = 0;
963 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
964 - cfqg->service_tree_idle.count;
967 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
970 charge = cfqq->slice_dispatch;
971 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
972 charge = cfqq->allocated_slice;
974 /* Can't update vdisktime while group is on service tree */
975 cfq_group_service_tree_del(st, cfqg);
976 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
977 /* If a new weight was requested, update now, off tree */
978 cfq_group_service_tree_add(st, cfqg);
980 /* This group is being expired. Save the context */
981 if (time_after(cfqd->workload_expires, jiffies)) {
982 cfqg->saved_workload_slice = cfqd->workload_expires
984 cfqg->saved_workload = cfqd->serving_type;
985 cfqg->saved_serving_prio = cfqd->serving_prio;
987 cfqg->saved_workload_slice = 0;
989 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
991 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
992 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
993 iops_mode(cfqd), cfqq->nr_sectors);
994 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
996 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
999 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1000 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1003 return container_of(blkg, struct cfq_group, blkg);
1007 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1008 unsigned int weight)
1010 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1011 cfqg->new_weight = weight;
1012 cfqg->needs_update = true;
1015 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1016 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1018 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1019 unsigned int major, minor;
1022 * Add group onto cgroup list. It might happen that bdi->dev is
1023 * not initialized yet. Initialize this new group without major
1024 * and minor info and this info will be filled in once a new thread
1028 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1029 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1030 (void *)cfqd, MKDEV(major, minor));
1032 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1035 cfqd->nr_blkcg_linked_grps++;
1036 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1038 /* Add group on cfqd list */
1039 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1043 * Should be called from sleepable context. No request queue lock as per
1044 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1045 * from sleepable context.
1047 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1049 struct cfq_group *cfqg = NULL;
1051 struct cfq_rb_root *st;
1053 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1057 for_each_cfqg_st(cfqg, i, j, st)
1059 RB_CLEAR_NODE(&cfqg->rb_node);
1062 * Take the initial reference that will be released on destroy
1063 * This can be thought of a joint reference by cgroup and
1064 * elevator which will be dropped by either elevator exit
1065 * or cgroup deletion path depending on who is exiting first.
1069 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1078 static struct cfq_group *
1079 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1081 struct cfq_group *cfqg = NULL;
1083 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1084 unsigned int major, minor;
1087 * This is the common case when there are no blkio cgroups.
1088 * Avoid lookup in this case
1090 if (blkcg == &blkio_root_cgroup)
1091 cfqg = &cfqd->root_group;
1093 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1095 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1096 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1097 cfqg->blkg.dev = MKDEV(major, minor);
1104 * Search for the cfq group current task belongs to. request_queue lock must
1107 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1109 struct blkio_cgroup *blkcg;
1110 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1111 struct request_queue *q = cfqd->queue;
1114 blkcg = task_blkio_cgroup(current);
1115 cfqg = cfq_find_cfqg(cfqd, blkcg);
1122 * Need to allocate a group. Allocation of group also needs allocation
1123 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1124 * we need to drop rcu lock and queue_lock before we call alloc.
1126 * Not taking any queue reference here and assuming that queue is
1127 * around by the time we return. CFQ queue allocation code does
1128 * the same. It might be racy though.
1132 spin_unlock_irq(q->queue_lock);
1134 cfqg = cfq_alloc_cfqg(cfqd);
1136 spin_lock_irq(q->queue_lock);
1139 blkcg = task_blkio_cgroup(current);
1142 * If some other thread already allocated the group while we were
1143 * not holding queue lock, free up the group
1145 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1154 cfqg = &cfqd->root_group;
1156 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1161 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1167 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1169 /* Currently, all async queues are mapped to root group */
1170 if (!cfq_cfqq_sync(cfqq))
1171 cfqg = &cfqq->cfqd->root_group;
1174 /* cfqq reference on cfqg */
1178 static void cfq_put_cfqg(struct cfq_group *cfqg)
1180 struct cfq_rb_root *st;
1183 BUG_ON(cfqg->ref <= 0);
1187 for_each_cfqg_st(cfqg, i, j, st)
1188 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1189 free_percpu(cfqg->blkg.stats_cpu);
1193 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1195 /* Something wrong if we are trying to remove same group twice */
1196 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1198 hlist_del_init(&cfqg->cfqd_node);
1201 * Put the reference taken at the time of creation so that when all
1202 * queues are gone, group can be destroyed.
1207 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1209 struct hlist_node *pos, *n;
1210 struct cfq_group *cfqg;
1212 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1214 * If cgroup removal path got to blk_group first and removed
1215 * it from cgroup list, then it will take care of destroying
1218 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1219 cfq_destroy_cfqg(cfqd, cfqg);
1224 * Blk cgroup controller notification saying that blkio_group object is being
1225 * delinked as associated cgroup object is going away. That also means that
1226 * no new IO will come in this group. So get rid of this group as soon as
1227 * any pending IO in the group is finished.
1229 * This function is called under rcu_read_lock(). key is the rcu protected
1230 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1233 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1234 * it should not be NULL as even if elevator was exiting, cgroup deltion
1235 * path got to it first.
1237 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1239 unsigned long flags;
1240 struct cfq_data *cfqd = key;
1242 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1243 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1244 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1247 #else /* GROUP_IOSCHED */
1248 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1250 return &cfqd->root_group;
1253 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1259 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1263 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1264 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1266 #endif /* GROUP_IOSCHED */
1269 * The cfqd->service_trees holds all pending cfq_queue's that have
1270 * requests waiting to be processed. It is sorted in the order that
1271 * we will service the queues.
1273 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1276 struct rb_node **p, *parent;
1277 struct cfq_queue *__cfqq;
1278 unsigned long rb_key;
1279 struct cfq_rb_root *service_tree;
1283 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1285 if (cfq_class_idle(cfqq)) {
1286 rb_key = CFQ_IDLE_DELAY;
1287 parent = rb_last(&service_tree->rb);
1288 if (parent && parent != &cfqq->rb_node) {
1289 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1290 rb_key += __cfqq->rb_key;
1293 } else if (!add_front) {
1295 * Get our rb key offset. Subtract any residual slice
1296 * value carried from last service. A negative resid
1297 * count indicates slice overrun, and this should position
1298 * the next service time further away in the tree.
1300 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1301 rb_key -= cfqq->slice_resid;
1302 cfqq->slice_resid = 0;
1305 __cfqq = cfq_rb_first(service_tree);
1306 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1309 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1312 * same position, nothing more to do
1314 if (rb_key == cfqq->rb_key &&
1315 cfqq->service_tree == service_tree)
1318 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1319 cfqq->service_tree = NULL;
1324 cfqq->service_tree = service_tree;
1325 p = &service_tree->rb.rb_node;
1330 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1333 * sort by key, that represents service time.
1335 if (time_before(rb_key, __cfqq->rb_key))
1338 n = &(*p)->rb_right;
1346 service_tree->left = &cfqq->rb_node;
1348 cfqq->rb_key = rb_key;
1349 rb_link_node(&cfqq->rb_node, parent, p);
1350 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1351 service_tree->count++;
1352 if (add_front || !new_cfqq)
1354 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1357 static struct cfq_queue *
1358 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1359 sector_t sector, struct rb_node **ret_parent,
1360 struct rb_node ***rb_link)
1362 struct rb_node **p, *parent;
1363 struct cfq_queue *cfqq = NULL;
1371 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1374 * Sort strictly based on sector. Smallest to the left,
1375 * largest to the right.
1377 if (sector > blk_rq_pos(cfqq->next_rq))
1378 n = &(*p)->rb_right;
1379 else if (sector < blk_rq_pos(cfqq->next_rq))
1387 *ret_parent = parent;
1393 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1395 struct rb_node **p, *parent;
1396 struct cfq_queue *__cfqq;
1399 rb_erase(&cfqq->p_node, cfqq->p_root);
1400 cfqq->p_root = NULL;
1403 if (cfq_class_idle(cfqq))
1408 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1409 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1410 blk_rq_pos(cfqq->next_rq), &parent, &p);
1412 rb_link_node(&cfqq->p_node, parent, p);
1413 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1415 cfqq->p_root = NULL;
1419 * Update cfqq's position in the service tree.
1421 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1424 * Resorting requires the cfqq to be on the RR list already.
1426 if (cfq_cfqq_on_rr(cfqq)) {
1427 cfq_service_tree_add(cfqd, cfqq, 0);
1428 cfq_prio_tree_add(cfqd, cfqq);
1433 * add to busy list of queues for service, trying to be fair in ordering
1434 * the pending list according to last request service
1436 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1438 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1439 BUG_ON(cfq_cfqq_on_rr(cfqq));
1440 cfq_mark_cfqq_on_rr(cfqq);
1441 cfqd->busy_queues++;
1442 if (cfq_cfqq_sync(cfqq))
1443 cfqd->busy_sync_queues++;
1445 cfq_resort_rr_list(cfqd, cfqq);
1449 * Called when the cfqq no longer has requests pending, remove it from
1452 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1454 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1455 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1456 cfq_clear_cfqq_on_rr(cfqq);
1458 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1459 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1460 cfqq->service_tree = NULL;
1463 rb_erase(&cfqq->p_node, cfqq->p_root);
1464 cfqq->p_root = NULL;
1467 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1468 BUG_ON(!cfqd->busy_queues);
1469 cfqd->busy_queues--;
1470 if (cfq_cfqq_sync(cfqq))
1471 cfqd->busy_sync_queues--;
1475 * rb tree support functions
1477 static void cfq_del_rq_rb(struct request *rq)
1479 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1480 const int sync = rq_is_sync(rq);
1482 BUG_ON(!cfqq->queued[sync]);
1483 cfqq->queued[sync]--;
1485 elv_rb_del(&cfqq->sort_list, rq);
1487 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1489 * Queue will be deleted from service tree when we actually
1490 * expire it later. Right now just remove it from prio tree
1494 rb_erase(&cfqq->p_node, cfqq->p_root);
1495 cfqq->p_root = NULL;
1500 static void cfq_add_rq_rb(struct request *rq)
1502 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1503 struct cfq_data *cfqd = cfqq->cfqd;
1504 struct request *prev;
1506 cfqq->queued[rq_is_sync(rq)]++;
1508 elv_rb_add(&cfqq->sort_list, rq);
1510 if (!cfq_cfqq_on_rr(cfqq))
1511 cfq_add_cfqq_rr(cfqd, cfqq);
1514 * check if this request is a better next-serve candidate
1516 prev = cfqq->next_rq;
1517 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1520 * adjust priority tree position, if ->next_rq changes
1522 if (prev != cfqq->next_rq)
1523 cfq_prio_tree_add(cfqd, cfqq);
1525 BUG_ON(!cfqq->next_rq);
1528 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1530 elv_rb_del(&cfqq->sort_list, rq);
1531 cfqq->queued[rq_is_sync(rq)]--;
1532 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1533 rq_data_dir(rq), rq_is_sync(rq));
1535 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1536 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1540 static struct request *
1541 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1543 struct task_struct *tsk = current;
1544 struct cfq_io_context *cic;
1545 struct cfq_queue *cfqq;
1547 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1551 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1553 sector_t sector = bio->bi_sector + bio_sectors(bio);
1555 return elv_rb_find(&cfqq->sort_list, sector);
1561 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1563 struct cfq_data *cfqd = q->elevator->elevator_data;
1565 cfqd->rq_in_driver++;
1566 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1567 cfqd->rq_in_driver);
1569 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1572 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1574 struct cfq_data *cfqd = q->elevator->elevator_data;
1576 WARN_ON(!cfqd->rq_in_driver);
1577 cfqd->rq_in_driver--;
1578 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1579 cfqd->rq_in_driver);
1582 static void cfq_remove_request(struct request *rq)
1584 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1586 if (cfqq->next_rq == rq)
1587 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1589 list_del_init(&rq->queuelist);
1592 cfqq->cfqd->rq_queued--;
1593 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1594 rq_data_dir(rq), rq_is_sync(rq));
1595 if (rq->cmd_flags & REQ_META) {
1596 WARN_ON(!cfqq->meta_pending);
1597 cfqq->meta_pending--;
1601 static int cfq_merge(struct request_queue *q, struct request **req,
1604 struct cfq_data *cfqd = q->elevator->elevator_data;
1605 struct request *__rq;
1607 __rq = cfq_find_rq_fmerge(cfqd, bio);
1608 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1610 return ELEVATOR_FRONT_MERGE;
1613 return ELEVATOR_NO_MERGE;
1616 static void cfq_merged_request(struct request_queue *q, struct request *req,
1619 if (type == ELEVATOR_FRONT_MERGE) {
1620 struct cfq_queue *cfqq = RQ_CFQQ(req);
1622 cfq_reposition_rq_rb(cfqq, req);
1626 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1629 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1630 bio_data_dir(bio), cfq_bio_sync(bio));
1634 cfq_merged_requests(struct request_queue *q, struct request *rq,
1635 struct request *next)
1637 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1639 * reposition in fifo if next is older than rq
1641 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1642 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1643 list_move(&rq->queuelist, &next->queuelist);
1644 rq_set_fifo_time(rq, rq_fifo_time(next));
1647 if (cfqq->next_rq == next)
1649 cfq_remove_request(next);
1650 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1651 rq_data_dir(next), rq_is_sync(next));
1654 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1657 struct cfq_data *cfqd = q->elevator->elevator_data;
1658 struct cfq_io_context *cic;
1659 struct cfq_queue *cfqq;
1662 * Disallow merge of a sync bio into an async request.
1664 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1668 * Lookup the cfqq that this bio will be queued with. Allow
1669 * merge only if rq is queued there.
1671 cic = cfq_cic_lookup(cfqd, current->io_context);
1675 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1676 return cfqq == RQ_CFQQ(rq);
1679 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1681 del_timer(&cfqd->idle_slice_timer);
1682 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1685 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1686 struct cfq_queue *cfqq)
1689 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1690 cfqd->serving_prio, cfqd->serving_type);
1691 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1692 cfqq->slice_start = 0;
1693 cfqq->dispatch_start = jiffies;
1694 cfqq->allocated_slice = 0;
1695 cfqq->slice_end = 0;
1696 cfqq->slice_dispatch = 0;
1697 cfqq->nr_sectors = 0;
1699 cfq_clear_cfqq_wait_request(cfqq);
1700 cfq_clear_cfqq_must_dispatch(cfqq);
1701 cfq_clear_cfqq_must_alloc_slice(cfqq);
1702 cfq_clear_cfqq_fifo_expire(cfqq);
1703 cfq_mark_cfqq_slice_new(cfqq);
1705 cfq_del_timer(cfqd, cfqq);
1708 cfqd->active_queue = cfqq;
1712 * current cfqq expired its slice (or was too idle), select new one
1715 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1718 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1720 if (cfq_cfqq_wait_request(cfqq))
1721 cfq_del_timer(cfqd, cfqq);
1723 cfq_clear_cfqq_wait_request(cfqq);
1724 cfq_clear_cfqq_wait_busy(cfqq);
1727 * If this cfqq is shared between multiple processes, check to
1728 * make sure that those processes are still issuing I/Os within
1729 * the mean seek distance. If not, it may be time to break the
1730 * queues apart again.
1732 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1733 cfq_mark_cfqq_split_coop(cfqq);
1736 * store what was left of this slice, if the queue idled/timed out
1739 if (cfq_cfqq_slice_new(cfqq))
1740 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1742 cfqq->slice_resid = cfqq->slice_end - jiffies;
1743 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1746 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1748 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1749 cfq_del_cfqq_rr(cfqd, cfqq);
1751 cfq_resort_rr_list(cfqd, cfqq);
1753 if (cfqq == cfqd->active_queue)
1754 cfqd->active_queue = NULL;
1756 if (cfqd->active_cic) {
1757 put_io_context(cfqd->active_cic->ioc);
1758 cfqd->active_cic = NULL;
1762 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1764 struct cfq_queue *cfqq = cfqd->active_queue;
1767 __cfq_slice_expired(cfqd, cfqq, timed_out);
1771 * Get next queue for service. Unless we have a queue preemption,
1772 * we'll simply select the first cfqq in the service tree.
1774 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1776 struct cfq_rb_root *service_tree =
1777 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1778 cfqd->serving_type);
1780 if (!cfqd->rq_queued)
1783 /* There is nothing to dispatch */
1786 if (RB_EMPTY_ROOT(&service_tree->rb))
1788 return cfq_rb_first(service_tree);
1791 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1793 struct cfq_group *cfqg;
1794 struct cfq_queue *cfqq;
1796 struct cfq_rb_root *st;
1798 if (!cfqd->rq_queued)
1801 cfqg = cfq_get_next_cfqg(cfqd);
1805 for_each_cfqg_st(cfqg, i, j, st)
1806 if ((cfqq = cfq_rb_first(st)) != NULL)
1812 * Get and set a new active queue for service.
1814 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1815 struct cfq_queue *cfqq)
1818 cfqq = cfq_get_next_queue(cfqd);
1820 __cfq_set_active_queue(cfqd, cfqq);
1824 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1827 if (blk_rq_pos(rq) >= cfqd->last_position)
1828 return blk_rq_pos(rq) - cfqd->last_position;
1830 return cfqd->last_position - blk_rq_pos(rq);
1833 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1836 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1839 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1840 struct cfq_queue *cur_cfqq)
1842 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1843 struct rb_node *parent, *node;
1844 struct cfq_queue *__cfqq;
1845 sector_t sector = cfqd->last_position;
1847 if (RB_EMPTY_ROOT(root))
1851 * First, if we find a request starting at the end of the last
1852 * request, choose it.
1854 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1859 * If the exact sector wasn't found, the parent of the NULL leaf
1860 * will contain the closest sector.
1862 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1863 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1866 if (blk_rq_pos(__cfqq->next_rq) < sector)
1867 node = rb_next(&__cfqq->p_node);
1869 node = rb_prev(&__cfqq->p_node);
1873 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1874 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1882 * cur_cfqq - passed in so that we don't decide that the current queue is
1883 * closely cooperating with itself.
1885 * So, basically we're assuming that that cur_cfqq has dispatched at least
1886 * one request, and that cfqd->last_position reflects a position on the disk
1887 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1890 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1891 struct cfq_queue *cur_cfqq)
1893 struct cfq_queue *cfqq;
1895 if (cfq_class_idle(cur_cfqq))
1897 if (!cfq_cfqq_sync(cur_cfqq))
1899 if (CFQQ_SEEKY(cur_cfqq))
1903 * Don't search priority tree if it's the only queue in the group.
1905 if (cur_cfqq->cfqg->nr_cfqq == 1)
1909 * We should notice if some of the queues are cooperating, eg
1910 * working closely on the same area of the disk. In that case,
1911 * we can group them together and don't waste time idling.
1913 cfqq = cfqq_close(cfqd, cur_cfqq);
1917 /* If new queue belongs to different cfq_group, don't choose it */
1918 if (cur_cfqq->cfqg != cfqq->cfqg)
1922 * It only makes sense to merge sync queues.
1924 if (!cfq_cfqq_sync(cfqq))
1926 if (CFQQ_SEEKY(cfqq))
1930 * Do not merge queues of different priority classes
1932 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1939 * Determine whether we should enforce idle window for this queue.
1942 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1944 enum wl_prio_t prio = cfqq_prio(cfqq);
1945 struct cfq_rb_root *service_tree = cfqq->service_tree;
1947 BUG_ON(!service_tree);
1948 BUG_ON(!service_tree->count);
1950 if (!cfqd->cfq_slice_idle)
1953 /* We never do for idle class queues. */
1954 if (prio == IDLE_WORKLOAD)
1957 /* We do for queues that were marked with idle window flag. */
1958 if (cfq_cfqq_idle_window(cfqq) &&
1959 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1963 * Otherwise, we do only if they are the last ones
1964 * in their service tree.
1966 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1968 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1969 service_tree->count);
1973 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1975 struct cfq_queue *cfqq = cfqd->active_queue;
1976 struct cfq_io_context *cic;
1977 unsigned long sl, group_idle = 0;
1980 * SSD device without seek penalty, disable idling. But only do so
1981 * for devices that support queuing, otherwise we still have a problem
1982 * with sync vs async workloads.
1984 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1987 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1988 WARN_ON(cfq_cfqq_slice_new(cfqq));
1991 * idle is disabled, either manually or by past process history
1993 if (!cfq_should_idle(cfqd, cfqq)) {
1994 /* no queue idling. Check for group idling */
1995 if (cfqd->cfq_group_idle)
1996 group_idle = cfqd->cfq_group_idle;
2002 * still active requests from this queue, don't idle
2004 if (cfqq->dispatched)
2008 * task has exited, don't wait
2010 cic = cfqd->active_cic;
2011 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2015 * If our average think time is larger than the remaining time
2016 * slice, then don't idle. This avoids overrunning the allotted
2019 if (sample_valid(cic->ttime_samples) &&
2020 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
2021 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
2026 /* There are other queues in the group, don't do group idle */
2027 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2030 cfq_mark_cfqq_wait_request(cfqq);
2033 sl = cfqd->cfq_group_idle;
2035 sl = cfqd->cfq_slice_idle;
2037 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2038 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2039 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2040 group_idle ? 1 : 0);
2044 * Move request from internal lists to the request queue dispatch list.
2046 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2048 struct cfq_data *cfqd = q->elevator->elevator_data;
2049 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2051 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2053 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2054 cfq_remove_request(rq);
2056 (RQ_CFQG(rq))->dispatched++;
2057 elv_dispatch_sort(q, rq);
2059 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2060 cfqq->nr_sectors += blk_rq_sectors(rq);
2061 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2062 rq_data_dir(rq), rq_is_sync(rq));
2066 * return expired entry, or NULL to just start from scratch in rbtree
2068 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2070 struct request *rq = NULL;
2072 if (cfq_cfqq_fifo_expire(cfqq))
2075 cfq_mark_cfqq_fifo_expire(cfqq);
2077 if (list_empty(&cfqq->fifo))
2080 rq = rq_entry_fifo(cfqq->fifo.next);
2081 if (time_before(jiffies, rq_fifo_time(rq)))
2084 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2089 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2091 const int base_rq = cfqd->cfq_slice_async_rq;
2093 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2095 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2099 * Must be called with the queue_lock held.
2101 static int cfqq_process_refs(struct cfq_queue *cfqq)
2103 int process_refs, io_refs;
2105 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2106 process_refs = cfqq->ref - io_refs;
2107 BUG_ON(process_refs < 0);
2108 return process_refs;
2111 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2113 int process_refs, new_process_refs;
2114 struct cfq_queue *__cfqq;
2117 * If there are no process references on the new_cfqq, then it is
2118 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2119 * chain may have dropped their last reference (not just their
2120 * last process reference).
2122 if (!cfqq_process_refs(new_cfqq))
2125 /* Avoid a circular list and skip interim queue merges */
2126 while ((__cfqq = new_cfqq->new_cfqq)) {
2132 process_refs = cfqq_process_refs(cfqq);
2133 new_process_refs = cfqq_process_refs(new_cfqq);
2135 * If the process for the cfqq has gone away, there is no
2136 * sense in merging the queues.
2138 if (process_refs == 0 || new_process_refs == 0)
2142 * Merge in the direction of the lesser amount of work.
2144 if (new_process_refs >= process_refs) {
2145 cfqq->new_cfqq = new_cfqq;
2146 new_cfqq->ref += process_refs;
2148 new_cfqq->new_cfqq = cfqq;
2149 cfqq->ref += new_process_refs;
2153 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2154 struct cfq_group *cfqg, enum wl_prio_t prio)
2156 struct cfq_queue *queue;
2158 bool key_valid = false;
2159 unsigned long lowest_key = 0;
2160 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2162 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2163 /* select the one with lowest rb_key */
2164 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2166 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2167 lowest_key = queue->rb_key;
2176 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2180 struct cfq_rb_root *st;
2181 unsigned group_slice;
2182 enum wl_prio_t original_prio = cfqd->serving_prio;
2184 /* Choose next priority. RT > BE > IDLE */
2185 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2186 cfqd->serving_prio = RT_WORKLOAD;
2187 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2188 cfqd->serving_prio = BE_WORKLOAD;
2190 cfqd->serving_prio = IDLE_WORKLOAD;
2191 cfqd->workload_expires = jiffies + 1;
2195 if (original_prio != cfqd->serving_prio)
2199 * For RT and BE, we have to choose also the type
2200 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2203 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2207 * check workload expiration, and that we still have other queues ready
2209 if (count && !time_after(jiffies, cfqd->workload_expires))
2213 /* otherwise select new workload type */
2214 cfqd->serving_type =
2215 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2216 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2220 * the workload slice is computed as a fraction of target latency
2221 * proportional to the number of queues in that workload, over
2222 * all the queues in the same priority class
2224 group_slice = cfq_group_slice(cfqd, cfqg);
2226 slice = group_slice * count /
2227 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2228 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2230 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2234 * Async queues are currently system wide. Just taking
2235 * proportion of queues with-in same group will lead to higher
2236 * async ratio system wide as generally root group is going
2237 * to have higher weight. A more accurate thing would be to
2238 * calculate system wide asnc/sync ratio.
2240 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2241 tmp = tmp/cfqd->busy_queues;
2242 slice = min_t(unsigned, slice, tmp);
2244 /* async workload slice is scaled down according to
2245 * the sync/async slice ratio. */
2246 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2248 /* sync workload slice is at least 2 * cfq_slice_idle */
2249 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2251 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2252 cfq_log(cfqd, "workload slice:%d", slice);
2253 cfqd->workload_expires = jiffies + slice;
2256 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2258 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2259 struct cfq_group *cfqg;
2261 if (RB_EMPTY_ROOT(&st->rb))
2263 cfqg = cfq_rb_first_group(st);
2264 update_min_vdisktime(st);
2268 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2270 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2272 cfqd->serving_group = cfqg;
2274 /* Restore the workload type data */
2275 if (cfqg->saved_workload_slice) {
2276 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2277 cfqd->serving_type = cfqg->saved_workload;
2278 cfqd->serving_prio = cfqg->saved_serving_prio;
2280 cfqd->workload_expires = jiffies - 1;
2282 choose_service_tree(cfqd, cfqg);
2286 * Select a queue for service. If we have a current active queue,
2287 * check whether to continue servicing it, or retrieve and set a new one.
2289 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2291 struct cfq_queue *cfqq, *new_cfqq = NULL;
2293 cfqq = cfqd->active_queue;
2297 if (!cfqd->rq_queued)
2301 * We were waiting for group to get backlogged. Expire the queue
2303 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2307 * The active queue has run out of time, expire it and select new.
2309 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2311 * If slice had not expired at the completion of last request
2312 * we might not have turned on wait_busy flag. Don't expire
2313 * the queue yet. Allow the group to get backlogged.
2315 * The very fact that we have used the slice, that means we
2316 * have been idling all along on this queue and it should be
2317 * ok to wait for this request to complete.
2319 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2320 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2324 goto check_group_idle;
2328 * The active queue has requests and isn't expired, allow it to
2331 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2335 * If another queue has a request waiting within our mean seek
2336 * distance, let it run. The expire code will check for close
2337 * cooperators and put the close queue at the front of the service
2338 * tree. If possible, merge the expiring queue with the new cfqq.
2340 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2342 if (!cfqq->new_cfqq)
2343 cfq_setup_merge(cfqq, new_cfqq);
2348 * No requests pending. If the active queue still has requests in
2349 * flight or is idling for a new request, allow either of these
2350 * conditions to happen (or time out) before selecting a new queue.
2352 if (timer_pending(&cfqd->idle_slice_timer)) {
2358 * This is a deep seek queue, but the device is much faster than
2359 * the queue can deliver, don't idle
2361 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2362 (cfq_cfqq_slice_new(cfqq) ||
2363 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2364 cfq_clear_cfqq_deep(cfqq);
2365 cfq_clear_cfqq_idle_window(cfqq);
2368 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2374 * If group idle is enabled and there are requests dispatched from
2375 * this group, wait for requests to complete.
2378 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2379 && cfqq->cfqg->dispatched) {
2385 cfq_slice_expired(cfqd, 0);
2388 * Current queue expired. Check if we have to switch to a new
2392 cfq_choose_cfqg(cfqd);
2394 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2399 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2403 while (cfqq->next_rq) {
2404 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2408 BUG_ON(!list_empty(&cfqq->fifo));
2410 /* By default cfqq is not expired if it is empty. Do it explicitly */
2411 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2416 * Drain our current requests. Used for barriers and when switching
2417 * io schedulers on-the-fly.
2419 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2421 struct cfq_queue *cfqq;
2424 /* Expire the timeslice of the current active queue first */
2425 cfq_slice_expired(cfqd, 0);
2426 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2427 __cfq_set_active_queue(cfqd, cfqq);
2428 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2431 BUG_ON(cfqd->busy_queues);
2433 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2437 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2438 struct cfq_queue *cfqq)
2440 /* the queue hasn't finished any request, can't estimate */
2441 if (cfq_cfqq_slice_new(cfqq))
2443 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2450 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2452 unsigned int max_dispatch;
2455 * Drain async requests before we start sync IO
2457 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2461 * If this is an async queue and we have sync IO in flight, let it wait
2463 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2466 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2467 if (cfq_class_idle(cfqq))
2471 * Does this cfqq already have too much IO in flight?
2473 if (cfqq->dispatched >= max_dispatch) {
2474 bool promote_sync = false;
2476 * idle queue must always only have a single IO in flight
2478 if (cfq_class_idle(cfqq))
2482 * If there is only one sync queue
2483 * we can ignore async queue here and give the sync
2484 * queue no dispatch limit. The reason is a sync queue can
2485 * preempt async queue, limiting the sync queue doesn't make
2486 * sense. This is useful for aiostress test.
2488 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2489 promote_sync = true;
2492 * We have other queues, don't allow more IO from this one
2494 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2499 * Sole queue user, no limit
2501 if (cfqd->busy_queues == 1 || promote_sync)
2505 * Normally we start throttling cfqq when cfq_quantum/2
2506 * requests have been dispatched. But we can drive
2507 * deeper queue depths at the beginning of slice
2508 * subjected to upper limit of cfq_quantum.
2510 max_dispatch = cfqd->cfq_quantum;
2514 * Async queues must wait a bit before being allowed dispatch.
2515 * We also ramp up the dispatch depth gradually for async IO,
2516 * based on the last sync IO we serviced
2518 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2519 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2522 depth = last_sync / cfqd->cfq_slice[1];
2523 if (!depth && !cfqq->dispatched)
2525 if (depth < max_dispatch)
2526 max_dispatch = depth;
2530 * If we're below the current max, allow a dispatch
2532 return cfqq->dispatched < max_dispatch;
2536 * Dispatch a request from cfqq, moving them to the request queue
2539 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2543 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2545 if (!cfq_may_dispatch(cfqd, cfqq))
2549 * follow expired path, else get first next available
2551 rq = cfq_check_fifo(cfqq);
2556 * insert request into driver dispatch list
2558 cfq_dispatch_insert(cfqd->queue, rq);
2560 if (!cfqd->active_cic) {
2561 struct cfq_io_context *cic = RQ_CIC(rq);
2563 atomic_long_inc(&cic->ioc->refcount);
2564 cfqd->active_cic = cic;
2571 * Find the cfqq that we need to service and move a request from that to the
2574 static int cfq_dispatch_requests(struct request_queue *q, int force)
2576 struct cfq_data *cfqd = q->elevator->elevator_data;
2577 struct cfq_queue *cfqq;
2579 if (!cfqd->busy_queues)
2582 if (unlikely(force))
2583 return cfq_forced_dispatch(cfqd);
2585 cfqq = cfq_select_queue(cfqd);
2590 * Dispatch a request from this cfqq, if it is allowed
2592 if (!cfq_dispatch_request(cfqd, cfqq))
2595 cfqq->slice_dispatch++;
2596 cfq_clear_cfqq_must_dispatch(cfqq);
2599 * expire an async queue immediately if it has used up its slice. idle
2600 * queue always expire after 1 dispatch round.
2602 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2603 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2604 cfq_class_idle(cfqq))) {
2605 cfqq->slice_end = jiffies + 1;
2606 cfq_slice_expired(cfqd, 0);
2609 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2614 * task holds one reference to the queue, dropped when task exits. each rq
2615 * in-flight on this queue also holds a reference, dropped when rq is freed.
2617 * Each cfq queue took a reference on the parent group. Drop it now.
2618 * queue lock must be held here.
2620 static void cfq_put_queue(struct cfq_queue *cfqq)
2622 struct cfq_data *cfqd = cfqq->cfqd;
2623 struct cfq_group *cfqg;
2625 BUG_ON(cfqq->ref <= 0);
2631 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2632 BUG_ON(rb_first(&cfqq->sort_list));
2633 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2636 if (unlikely(cfqd->active_queue == cfqq)) {
2637 __cfq_slice_expired(cfqd, cfqq, 0);
2638 cfq_schedule_dispatch(cfqd);
2641 BUG_ON(cfq_cfqq_on_rr(cfqq));
2642 kmem_cache_free(cfq_pool, cfqq);
2647 * Call func for each cic attached to this ioc.
2650 call_for_each_cic(struct io_context *ioc,
2651 void (*func)(struct io_context *, struct cfq_io_context *))
2653 struct cfq_io_context *cic;
2654 struct hlist_node *n;
2658 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2664 static void cfq_cic_free_rcu(struct rcu_head *head)
2666 struct cfq_io_context *cic;
2668 cic = container_of(head, struct cfq_io_context, rcu_head);
2670 kmem_cache_free(cfq_ioc_pool, cic);
2671 elv_ioc_count_dec(cfq_ioc_count);
2675 * CFQ scheduler is exiting, grab exit lock and check
2676 * the pending io context count. If it hits zero,
2677 * complete ioc_gone and set it back to NULL
2679 spin_lock(&ioc_gone_lock);
2680 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2684 spin_unlock(&ioc_gone_lock);
2688 static void cfq_cic_free(struct cfq_io_context *cic)
2690 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2693 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2695 unsigned long flags;
2696 unsigned long dead_key = (unsigned long) cic->key;
2698 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2700 spin_lock_irqsave(&ioc->lock, flags);
2701 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2702 hlist_del_rcu(&cic->cic_list);
2703 spin_unlock_irqrestore(&ioc->lock, flags);
2709 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2710 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2711 * and ->trim() which is called with the task lock held
2713 static void cfq_free_io_context(struct io_context *ioc)
2716 * ioc->refcount is zero here, or we are called from elv_unregister(),
2717 * so no more cic's are allowed to be linked into this ioc. So it
2718 * should be ok to iterate over the known list, we will see all cic's
2719 * since no new ones are added.
2721 call_for_each_cic(ioc, cic_free_func);
2724 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2726 struct cfq_queue *__cfqq, *next;
2729 * If this queue was scheduled to merge with another queue, be
2730 * sure to drop the reference taken on that queue (and others in
2731 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2733 __cfqq = cfqq->new_cfqq;
2735 if (__cfqq == cfqq) {
2736 WARN(1, "cfqq->new_cfqq loop detected\n");
2739 next = __cfqq->new_cfqq;
2740 cfq_put_queue(__cfqq);
2745 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2747 if (unlikely(cfqq == cfqd->active_queue)) {
2748 __cfq_slice_expired(cfqd, cfqq, 0);
2749 cfq_schedule_dispatch(cfqd);
2752 cfq_put_cooperator(cfqq);
2754 cfq_put_queue(cfqq);
2757 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2758 struct cfq_io_context *cic)
2760 struct io_context *ioc = cic->ioc;
2762 list_del_init(&cic->queue_list);
2765 * Make sure dead mark is seen for dead queues
2768 cic->key = cfqd_dead_key(cfqd);
2770 if (ioc->ioc_data == cic)
2771 rcu_assign_pointer(ioc->ioc_data, NULL);
2773 if (cic->cfqq[BLK_RW_ASYNC]) {
2774 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2775 cic->cfqq[BLK_RW_ASYNC] = NULL;
2778 if (cic->cfqq[BLK_RW_SYNC]) {
2779 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2780 cic->cfqq[BLK_RW_SYNC] = NULL;
2784 static void cfq_exit_single_io_context(struct io_context *ioc,
2785 struct cfq_io_context *cic)
2787 struct cfq_data *cfqd = cic_to_cfqd(cic);
2790 struct request_queue *q = cfqd->queue;
2791 unsigned long flags;
2793 spin_lock_irqsave(q->queue_lock, flags);
2796 * Ensure we get a fresh copy of the ->key to prevent
2797 * race between exiting task and queue
2799 smp_read_barrier_depends();
2800 if (cic->key == cfqd)
2801 __cfq_exit_single_io_context(cfqd, cic);
2803 spin_unlock_irqrestore(q->queue_lock, flags);
2808 * The process that ioc belongs to has exited, we need to clean up
2809 * and put the internal structures we have that belongs to that process.
2811 static void cfq_exit_io_context(struct io_context *ioc)
2813 call_for_each_cic(ioc, cfq_exit_single_io_context);
2816 static struct cfq_io_context *
2817 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2819 struct cfq_io_context *cic;
2821 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2824 cic->last_end_request = jiffies;
2825 INIT_LIST_HEAD(&cic->queue_list);
2826 INIT_HLIST_NODE(&cic->cic_list);
2827 cic->dtor = cfq_free_io_context;
2828 cic->exit = cfq_exit_io_context;
2829 elv_ioc_count_inc(cfq_ioc_count);
2835 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2837 struct task_struct *tsk = current;
2840 if (!cfq_cfqq_prio_changed(cfqq))
2843 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2844 switch (ioprio_class) {
2846 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2847 case IOPRIO_CLASS_NONE:
2849 * no prio set, inherit CPU scheduling settings
2851 cfqq->ioprio = task_nice_ioprio(tsk);
2852 cfqq->ioprio_class = task_nice_ioclass(tsk);
2854 case IOPRIO_CLASS_RT:
2855 cfqq->ioprio = task_ioprio(ioc);
2856 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2858 case IOPRIO_CLASS_BE:
2859 cfqq->ioprio = task_ioprio(ioc);
2860 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2862 case IOPRIO_CLASS_IDLE:
2863 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2865 cfq_clear_cfqq_idle_window(cfqq);
2870 * keep track of original prio settings in case we have to temporarily
2871 * elevate the priority of this queue
2873 cfqq->org_ioprio = cfqq->ioprio;
2874 cfqq->org_ioprio_class = cfqq->ioprio_class;
2875 cfq_clear_cfqq_prio_changed(cfqq);
2878 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2880 struct cfq_data *cfqd = cic_to_cfqd(cic);
2881 struct cfq_queue *cfqq;
2882 unsigned long flags;
2884 if (unlikely(!cfqd))
2887 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2889 cfqq = cic->cfqq[BLK_RW_ASYNC];
2891 struct cfq_queue *new_cfqq;
2892 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2895 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2896 cfq_put_queue(cfqq);
2900 cfqq = cic->cfqq[BLK_RW_SYNC];
2902 cfq_mark_cfqq_prio_changed(cfqq);
2904 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2907 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2909 call_for_each_cic(ioc, changed_ioprio);
2910 ioc->ioprio_changed = 0;
2913 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2914 pid_t pid, bool is_sync)
2916 RB_CLEAR_NODE(&cfqq->rb_node);
2917 RB_CLEAR_NODE(&cfqq->p_node);
2918 INIT_LIST_HEAD(&cfqq->fifo);
2923 cfq_mark_cfqq_prio_changed(cfqq);
2926 if (!cfq_class_idle(cfqq))
2927 cfq_mark_cfqq_idle_window(cfqq);
2928 cfq_mark_cfqq_sync(cfqq);
2933 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2934 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2936 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2937 struct cfq_data *cfqd = cic_to_cfqd(cic);
2938 unsigned long flags;
2939 struct request_queue *q;
2941 if (unlikely(!cfqd))
2946 spin_lock_irqsave(q->queue_lock, flags);
2950 * Drop reference to sync queue. A new sync queue will be
2951 * assigned in new group upon arrival of a fresh request.
2953 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2954 cic_set_cfqq(cic, NULL, 1);
2955 cfq_put_queue(sync_cfqq);
2958 spin_unlock_irqrestore(q->queue_lock, flags);
2961 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2963 call_for_each_cic(ioc, changed_cgroup);
2964 ioc->cgroup_changed = 0;
2966 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2968 static struct cfq_queue *
2969 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2970 struct io_context *ioc, gfp_t gfp_mask)
2972 struct cfq_queue *cfqq, *new_cfqq = NULL;
2973 struct cfq_io_context *cic;
2974 struct cfq_group *cfqg;
2977 cfqg = cfq_get_cfqg(cfqd);
2978 cic = cfq_cic_lookup(cfqd, ioc);
2979 /* cic always exists here */
2980 cfqq = cic_to_cfqq(cic, is_sync);
2983 * Always try a new alloc if we fell back to the OOM cfqq
2984 * originally, since it should just be a temporary situation.
2986 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2991 } else if (gfp_mask & __GFP_WAIT) {
2992 spin_unlock_irq(cfqd->queue->queue_lock);
2993 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2994 gfp_mask | __GFP_ZERO,
2996 spin_lock_irq(cfqd->queue->queue_lock);
3000 cfqq = kmem_cache_alloc_node(cfq_pool,
3001 gfp_mask | __GFP_ZERO,
3006 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3007 cfq_init_prio_data(cfqq, ioc);
3008 cfq_link_cfqq_cfqg(cfqq, cfqg);
3009 cfq_log_cfqq(cfqd, cfqq, "alloced");
3011 cfqq = &cfqd->oom_cfqq;
3015 kmem_cache_free(cfq_pool, new_cfqq);
3020 static struct cfq_queue **
3021 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3023 switch (ioprio_class) {
3024 case IOPRIO_CLASS_RT:
3025 return &cfqd->async_cfqq[0][ioprio];
3026 case IOPRIO_CLASS_BE:
3027 return &cfqd->async_cfqq[1][ioprio];
3028 case IOPRIO_CLASS_IDLE:
3029 return &cfqd->async_idle_cfqq;
3035 static struct cfq_queue *
3036 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3039 const int ioprio = task_ioprio(ioc);
3040 const int ioprio_class = task_ioprio_class(ioc);
3041 struct cfq_queue **async_cfqq = NULL;
3042 struct cfq_queue *cfqq = NULL;
3045 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3050 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3053 * pin the queue now that it's allocated, scheduler exit will prune it
3055 if (!is_sync && !(*async_cfqq)) {
3065 * We drop cfq io contexts lazily, so we may find a dead one.
3068 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3069 struct cfq_io_context *cic)
3071 unsigned long flags;
3073 WARN_ON(!list_empty(&cic->queue_list));
3074 BUG_ON(cic->key != cfqd_dead_key(cfqd));
3076 spin_lock_irqsave(&ioc->lock, flags);
3078 BUG_ON(ioc->ioc_data == cic);
3080 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3081 hlist_del_rcu(&cic->cic_list);
3082 spin_unlock_irqrestore(&ioc->lock, flags);
3087 static struct cfq_io_context *
3088 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3090 struct cfq_io_context *cic;
3091 unsigned long flags;
3099 * we maintain a last-hit cache, to avoid browsing over the tree
3101 cic = rcu_dereference(ioc->ioc_data);
3102 if (cic && cic->key == cfqd) {
3108 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3112 if (unlikely(cic->key != cfqd)) {
3113 cfq_drop_dead_cic(cfqd, ioc, cic);
3118 spin_lock_irqsave(&ioc->lock, flags);
3119 rcu_assign_pointer(ioc->ioc_data, cic);
3120 spin_unlock_irqrestore(&ioc->lock, flags);
3128 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3129 * the process specific cfq io context when entered from the block layer.
3130 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3132 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3133 struct cfq_io_context *cic, gfp_t gfp_mask)
3135 unsigned long flags;
3138 ret = radix_tree_preload(gfp_mask);
3143 spin_lock_irqsave(&ioc->lock, flags);
3144 ret = radix_tree_insert(&ioc->radix_root,
3145 cfqd->cic_index, cic);
3147 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3148 spin_unlock_irqrestore(&ioc->lock, flags);
3150 radix_tree_preload_end();
3153 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3154 list_add(&cic->queue_list, &cfqd->cic_list);
3155 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3160 printk(KERN_ERR "cfq: cic link failed!\n");
3166 * Setup general io context and cfq io context. There can be several cfq
3167 * io contexts per general io context, if this process is doing io to more
3168 * than one device managed by cfq.
3170 static struct cfq_io_context *
3171 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3173 struct io_context *ioc = NULL;
3174 struct cfq_io_context *cic;
3176 might_sleep_if(gfp_mask & __GFP_WAIT);
3178 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3182 cic = cfq_cic_lookup(cfqd, ioc);
3186 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3190 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3194 smp_read_barrier_depends();
3195 if (unlikely(ioc->ioprio_changed))
3196 cfq_ioc_set_ioprio(ioc);
3198 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3199 if (unlikely(ioc->cgroup_changed))
3200 cfq_ioc_set_cgroup(ioc);
3206 put_io_context(ioc);
3211 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3213 unsigned long elapsed = jiffies - cic->last_end_request;
3214 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3216 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3217 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3218 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3222 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3226 sector_t n_sec = blk_rq_sectors(rq);
3227 if (cfqq->last_request_pos) {
3228 if (cfqq->last_request_pos < blk_rq_pos(rq))
3229 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3231 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3234 cfqq->seek_history <<= 1;
3235 if (blk_queue_nonrot(cfqd->queue))
3236 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3238 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3242 * Disable idle window if the process thinks too long or seeks so much that
3246 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3247 struct cfq_io_context *cic)
3249 int old_idle, enable_idle;
3252 * Don't idle for async or idle io prio class
3254 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3257 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3259 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3260 cfq_mark_cfqq_deep(cfqq);
3262 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3264 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3265 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3267 else if (sample_valid(cic->ttime_samples)) {
3268 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3274 if (old_idle != enable_idle) {
3275 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3277 cfq_mark_cfqq_idle_window(cfqq);
3279 cfq_clear_cfqq_idle_window(cfqq);
3284 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3285 * no or if we aren't sure, a 1 will cause a preempt.
3288 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3291 struct cfq_queue *cfqq;
3293 cfqq = cfqd->active_queue;
3297 if (cfq_class_idle(new_cfqq))
3300 if (cfq_class_idle(cfqq))
3304 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3306 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3310 * if the new request is sync, but the currently running queue is
3311 * not, let the sync request have priority.
3313 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3316 if (new_cfqq->cfqg != cfqq->cfqg)
3319 if (cfq_slice_used(cfqq))
3322 /* Allow preemption only if we are idling on sync-noidle tree */
3323 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3324 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3325 new_cfqq->service_tree->count == 2 &&
3326 RB_EMPTY_ROOT(&cfqq->sort_list))
3330 * So both queues are sync. Let the new request get disk time if
3331 * it's a metadata request and the current queue is doing regular IO.
3333 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3337 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3339 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3342 /* An idle queue should not be idle now for some reason */
3343 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3346 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3350 * if this request is as-good as one we would expect from the
3351 * current cfqq, let it preempt
3353 if (cfq_rq_close(cfqd, cfqq, rq))
3360 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3361 * let it have half of its nominal slice.
3363 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3365 struct cfq_queue *old_cfqq = cfqd->active_queue;
3367 cfq_log_cfqq(cfqd, cfqq, "preempt");
3368 cfq_slice_expired(cfqd, 1);
3371 * workload type is changed, don't save slice, otherwise preempt
3374 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3375 cfqq->cfqg->saved_workload_slice = 0;
3378 * Put the new queue at the front of the of the current list,
3379 * so we know that it will be selected next.
3381 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3383 cfq_service_tree_add(cfqd, cfqq, 1);
3385 cfqq->slice_end = 0;
3386 cfq_mark_cfqq_slice_new(cfqq);
3390 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3391 * something we should do about it
3394 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3397 struct cfq_io_context *cic = RQ_CIC(rq);
3400 if (rq->cmd_flags & REQ_META)
3401 cfqq->meta_pending++;
3403 cfq_update_io_thinktime(cfqd, cic);
3404 cfq_update_io_seektime(cfqd, cfqq, rq);
3405 cfq_update_idle_window(cfqd, cfqq, cic);
3407 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3409 if (cfqq == cfqd->active_queue) {
3411 * Remember that we saw a request from this process, but
3412 * don't start queuing just yet. Otherwise we risk seeing lots
3413 * of tiny requests, because we disrupt the normal plugging
3414 * and merging. If the request is already larger than a single
3415 * page, let it rip immediately. For that case we assume that
3416 * merging is already done. Ditto for a busy system that
3417 * has other work pending, don't risk delaying until the
3418 * idle timer unplug to continue working.
3420 if (cfq_cfqq_wait_request(cfqq)) {
3421 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3422 cfqd->busy_queues > 1) {
3423 cfq_del_timer(cfqd, cfqq);
3424 cfq_clear_cfqq_wait_request(cfqq);
3425 __blk_run_queue(cfqd->queue);
3427 cfq_blkiocg_update_idle_time_stats(
3429 cfq_mark_cfqq_must_dispatch(cfqq);
3432 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3434 * not the active queue - expire current slice if it is
3435 * idle and has expired it's mean thinktime or this new queue
3436 * has some old slice time left and is of higher priority or
3437 * this new queue is RT and the current one is BE
3439 cfq_preempt_queue(cfqd, cfqq);
3440 __blk_run_queue(cfqd->queue);
3444 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3446 struct cfq_data *cfqd = q->elevator->elevator_data;
3447 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3449 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3450 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3452 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3453 list_add_tail(&rq->queuelist, &cfqq->fifo);
3455 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3456 &cfqd->serving_group->blkg, rq_data_dir(rq),
3458 cfq_rq_enqueued(cfqd, cfqq, rq);
3462 * Update hw_tag based on peak queue depth over 50 samples under
3465 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3467 struct cfq_queue *cfqq = cfqd->active_queue;
3469 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3470 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3472 if (cfqd->hw_tag == 1)
3475 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3476 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3480 * If active queue hasn't enough requests and can idle, cfq might not
3481 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3484 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3485 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3486 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3489 if (cfqd->hw_tag_samples++ < 50)
3492 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3498 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3500 struct cfq_io_context *cic = cfqd->active_cic;
3502 /* If the queue already has requests, don't wait */
3503 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3506 /* If there are other queues in the group, don't wait */
3507 if (cfqq->cfqg->nr_cfqq > 1)
3510 if (cfq_slice_used(cfqq))
3513 /* if slice left is less than think time, wait busy */
3514 if (cic && sample_valid(cic->ttime_samples)
3515 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3519 * If think times is less than a jiffy than ttime_mean=0 and above
3520 * will not be true. It might happen that slice has not expired yet
3521 * but will expire soon (4-5 ns) during select_queue(). To cover the
3522 * case where think time is less than a jiffy, mark the queue wait
3523 * busy if only 1 jiffy is left in the slice.
3525 if (cfqq->slice_end - jiffies == 1)
3531 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3533 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3534 struct cfq_data *cfqd = cfqq->cfqd;
3535 const int sync = rq_is_sync(rq);
3539 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3540 !!(rq->cmd_flags & REQ_NOIDLE));
3542 cfq_update_hw_tag(cfqd);
3544 WARN_ON(!cfqd->rq_in_driver);
3545 WARN_ON(!cfqq->dispatched);
3546 cfqd->rq_in_driver--;
3548 (RQ_CFQG(rq))->dispatched--;
3549 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3550 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3551 rq_data_dir(rq), rq_is_sync(rq));
3553 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3556 RQ_CIC(rq)->last_end_request = now;
3557 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3558 cfqd->last_delayed_sync = now;
3562 * If this is the active queue, check if it needs to be expired,
3563 * or if we want to idle in case it has no pending requests.
3565 if (cfqd->active_queue == cfqq) {
3566 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3568 if (cfq_cfqq_slice_new(cfqq)) {
3569 cfq_set_prio_slice(cfqd, cfqq);
3570 cfq_clear_cfqq_slice_new(cfqq);
3574 * Should we wait for next request to come in before we expire
3577 if (cfq_should_wait_busy(cfqd, cfqq)) {
3578 unsigned long extend_sl = cfqd->cfq_slice_idle;
3579 if (!cfqd->cfq_slice_idle)
3580 extend_sl = cfqd->cfq_group_idle;
3581 cfqq->slice_end = jiffies + extend_sl;
3582 cfq_mark_cfqq_wait_busy(cfqq);
3583 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3587 * Idling is not enabled on:
3589 * - idle-priority queues
3591 * - queues with still some requests queued
3592 * - when there is a close cooperator
3594 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3595 cfq_slice_expired(cfqd, 1);
3596 else if (sync && cfqq_empty &&
3597 !cfq_close_cooperator(cfqd, cfqq)) {
3598 cfq_arm_slice_timer(cfqd);
3602 if (!cfqd->rq_in_driver)
3603 cfq_schedule_dispatch(cfqd);
3607 * we temporarily boost lower priority queues if they are holding fs exclusive
3608 * resources. they are boosted to normal prio (CLASS_BE/4)
3610 static void cfq_prio_boost(struct cfq_queue *cfqq)
3612 if (has_fs_excl()) {
3614 * boost idle prio on transactions that would lock out other
3615 * users of the filesystem
3617 if (cfq_class_idle(cfqq))
3618 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3619 if (cfqq->ioprio > IOPRIO_NORM)
3620 cfqq->ioprio = IOPRIO_NORM;
3623 * unboost the queue (if needed)
3625 cfqq->ioprio_class = cfqq->org_ioprio_class;
3626 cfqq->ioprio = cfqq->org_ioprio;
3630 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3632 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3633 cfq_mark_cfqq_must_alloc_slice(cfqq);
3634 return ELV_MQUEUE_MUST;
3637 return ELV_MQUEUE_MAY;
3640 static int cfq_may_queue(struct request_queue *q, int rw)
3642 struct cfq_data *cfqd = q->elevator->elevator_data;
3643 struct task_struct *tsk = current;
3644 struct cfq_io_context *cic;
3645 struct cfq_queue *cfqq;
3648 * don't force setup of a queue from here, as a call to may_queue
3649 * does not necessarily imply that a request actually will be queued.
3650 * so just lookup a possibly existing queue, or return 'may queue'
3653 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3655 return ELV_MQUEUE_MAY;
3657 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3659 cfq_init_prio_data(cfqq, cic->ioc);
3660 cfq_prio_boost(cfqq);
3662 return __cfq_may_queue(cfqq);
3665 return ELV_MQUEUE_MAY;
3669 * queue lock held here
3671 static void cfq_put_request(struct request *rq)
3673 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3676 const int rw = rq_data_dir(rq);
3678 BUG_ON(!cfqq->allocated[rw]);
3679 cfqq->allocated[rw]--;
3681 put_io_context(RQ_CIC(rq)->ioc);
3683 rq->elevator_private[0] = NULL;
3684 rq->elevator_private[1] = NULL;
3686 /* Put down rq reference on cfqg */
3687 cfq_put_cfqg(RQ_CFQG(rq));
3688 rq->elevator_private[2] = NULL;
3690 cfq_put_queue(cfqq);
3694 static struct cfq_queue *
3695 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3696 struct cfq_queue *cfqq)
3698 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3699 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3700 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3701 cfq_put_queue(cfqq);
3702 return cic_to_cfqq(cic, 1);
3706 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3707 * was the last process referring to said cfqq.
3709 static struct cfq_queue *
3710 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3712 if (cfqq_process_refs(cfqq) == 1) {
3713 cfqq->pid = current->pid;
3714 cfq_clear_cfqq_coop(cfqq);
3715 cfq_clear_cfqq_split_coop(cfqq);
3719 cic_set_cfqq(cic, NULL, 1);
3721 cfq_put_cooperator(cfqq);
3723 cfq_put_queue(cfqq);
3727 * Allocate cfq data structures associated with this request.
3730 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3732 struct cfq_data *cfqd = q->elevator->elevator_data;
3733 struct cfq_io_context *cic;
3734 const int rw = rq_data_dir(rq);
3735 const bool is_sync = rq_is_sync(rq);
3736 struct cfq_queue *cfqq;
3737 unsigned long flags;
3739 might_sleep_if(gfp_mask & __GFP_WAIT);
3741 cic = cfq_get_io_context(cfqd, gfp_mask);
3743 spin_lock_irqsave(q->queue_lock, flags);
3749 cfqq = cic_to_cfqq(cic, is_sync);
3750 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3751 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3752 cic_set_cfqq(cic, cfqq, is_sync);
3755 * If the queue was seeky for too long, break it apart.
3757 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3758 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3759 cfqq = split_cfqq(cic, cfqq);
3765 * Check to see if this queue is scheduled to merge with
3766 * another, closely cooperating queue. The merging of
3767 * queues happens here as it must be done in process context.
3768 * The reference on new_cfqq was taken in merge_cfqqs.
3771 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3774 cfqq->allocated[rw]++;
3777 rq->elevator_private[0] = cic;
3778 rq->elevator_private[1] = cfqq;
3779 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3780 spin_unlock_irqrestore(q->queue_lock, flags);
3785 put_io_context(cic->ioc);
3787 cfq_schedule_dispatch(cfqd);
3788 spin_unlock_irqrestore(q->queue_lock, flags);
3789 cfq_log(cfqd, "set_request fail");
3793 static void cfq_kick_queue(struct work_struct *work)
3795 struct cfq_data *cfqd =
3796 container_of(work, struct cfq_data, unplug_work);
3797 struct request_queue *q = cfqd->queue;
3799 spin_lock_irq(q->queue_lock);
3800 __blk_run_queue(cfqd->queue);
3801 spin_unlock_irq(q->queue_lock);
3805 * Timer running if the active_queue is currently idling inside its time slice
3807 static void cfq_idle_slice_timer(unsigned long data)
3809 struct cfq_data *cfqd = (struct cfq_data *) data;
3810 struct cfq_queue *cfqq;
3811 unsigned long flags;
3814 cfq_log(cfqd, "idle timer fired");
3816 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3818 cfqq = cfqd->active_queue;
3823 * We saw a request before the queue expired, let it through
3825 if (cfq_cfqq_must_dispatch(cfqq))
3831 if (cfq_slice_used(cfqq))
3835 * only expire and reinvoke request handler, if there are
3836 * other queues with pending requests
3838 if (!cfqd->busy_queues)
3842 * not expired and it has a request pending, let it dispatch
3844 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3848 * Queue depth flag is reset only when the idle didn't succeed
3850 cfq_clear_cfqq_deep(cfqq);
3853 cfq_slice_expired(cfqd, timed_out);
3855 cfq_schedule_dispatch(cfqd);
3857 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3860 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3862 del_timer_sync(&cfqd->idle_slice_timer);
3863 cancel_work_sync(&cfqd->unplug_work);
3866 static void cfq_put_async_queues(struct cfq_data *cfqd)
3870 for (i = 0; i < IOPRIO_BE_NR; i++) {
3871 if (cfqd->async_cfqq[0][i])
3872 cfq_put_queue(cfqd->async_cfqq[0][i]);
3873 if (cfqd->async_cfqq[1][i])
3874 cfq_put_queue(cfqd->async_cfqq[1][i]);
3877 if (cfqd->async_idle_cfqq)
3878 cfq_put_queue(cfqd->async_idle_cfqq);
3881 static void cfq_exit_queue(struct elevator_queue *e)
3883 struct cfq_data *cfqd = e->elevator_data;
3884 struct request_queue *q = cfqd->queue;
3887 cfq_shutdown_timer_wq(cfqd);
3889 spin_lock_irq(q->queue_lock);
3891 if (cfqd->active_queue)
3892 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3894 while (!list_empty(&cfqd->cic_list)) {
3895 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3896 struct cfq_io_context,
3899 __cfq_exit_single_io_context(cfqd, cic);
3902 cfq_put_async_queues(cfqd);
3903 cfq_release_cfq_groups(cfqd);
3906 * If there are groups which we could not unlink from blkcg list,
3907 * wait for a rcu period for them to be freed.
3909 if (cfqd->nr_blkcg_linked_grps)
3912 spin_unlock_irq(q->queue_lock);
3914 cfq_shutdown_timer_wq(cfqd);
3916 spin_lock(&cic_index_lock);
3917 ida_remove(&cic_index_ida, cfqd->cic_index);
3918 spin_unlock(&cic_index_lock);
3921 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3922 * Do this wait only if there are other unlinked groups out
3923 * there. This can happen if cgroup deletion path claimed the
3924 * responsibility of cleaning up a group before queue cleanup code
3927 * Do not call synchronize_rcu() unconditionally as there are drivers
3928 * which create/delete request queue hundreds of times during scan/boot
3929 * and synchronize_rcu() can take significant time and slow down boot.
3934 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3935 /* Free up per cpu stats for root group */
3936 free_percpu(cfqd->root_group.blkg.stats_cpu);
3941 static int cfq_alloc_cic_index(void)
3946 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3949 spin_lock(&cic_index_lock);
3950 error = ida_get_new(&cic_index_ida, &index);
3951 spin_unlock(&cic_index_lock);
3952 if (error && error != -EAGAIN)
3959 static void *cfq_init_queue(struct request_queue *q)
3961 struct cfq_data *cfqd;
3963 struct cfq_group *cfqg;
3964 struct cfq_rb_root *st;
3966 i = cfq_alloc_cic_index();
3970 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3972 spin_lock(&cic_index_lock);
3973 ida_remove(&cic_index_ida, i);
3974 spin_unlock(&cic_index_lock);
3979 * Don't need take queue_lock in the routine, since we are
3980 * initializing the ioscheduler, and nobody is using cfqd
3982 cfqd->cic_index = i;
3984 /* Init root service tree */
3985 cfqd->grp_service_tree = CFQ_RB_ROOT;
3987 /* Init root group */
3988 cfqg = &cfqd->root_group;
3989 for_each_cfqg_st(cfqg, i, j, st)
3991 RB_CLEAR_NODE(&cfqg->rb_node);
3993 /* Give preference to root group over other groups */
3994 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3996 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3998 * Set root group reference to 2. One reference will be dropped when
3999 * all groups on cfqd->cfqg_list are being deleted during queue exit.
4000 * Other reference will remain there as we don't want to delete this
4001 * group as it is statically allocated and gets destroyed when
4002 * throtl_data goes away.
4006 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4014 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4017 cfqd->nr_blkcg_linked_grps++;
4019 /* Add group on cfqd->cfqg_list */
4020 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4023 * Not strictly needed (since RB_ROOT just clears the node and we
4024 * zeroed cfqd on alloc), but better be safe in case someone decides
4025 * to add magic to the rb code
4027 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4028 cfqd->prio_trees[i] = RB_ROOT;
4031 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4032 * Grab a permanent reference to it, so that the normal code flow
4033 * will not attempt to free it.
4035 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4036 cfqd->oom_cfqq.ref++;
4037 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4039 INIT_LIST_HEAD(&cfqd->cic_list);
4043 init_timer(&cfqd->idle_slice_timer);
4044 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4045 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4047 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4049 cfqd->cfq_quantum = cfq_quantum;
4050 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4051 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4052 cfqd->cfq_back_max = cfq_back_max;
4053 cfqd->cfq_back_penalty = cfq_back_penalty;
4054 cfqd->cfq_slice[0] = cfq_slice_async;
4055 cfqd->cfq_slice[1] = cfq_slice_sync;
4056 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4057 cfqd->cfq_slice_idle = cfq_slice_idle;
4058 cfqd->cfq_group_idle = cfq_group_idle;
4059 cfqd->cfq_latency = 1;
4062 * we optimistically start assuming sync ops weren't delayed in last
4063 * second, in order to have larger depth for async operations.
4065 cfqd->last_delayed_sync = jiffies - HZ;
4069 static void cfq_slab_kill(void)
4072 * Caller already ensured that pending RCU callbacks are completed,
4073 * so we should have no busy allocations at this point.
4076 kmem_cache_destroy(cfq_pool);
4078 kmem_cache_destroy(cfq_ioc_pool);
4081 static int __init cfq_slab_setup(void)
4083 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4087 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4098 * sysfs parts below -->
4101 cfq_var_show(unsigned int var, char *page)
4103 return sprintf(page, "%d\n", var);
4107 cfq_var_store(unsigned int *var, const char *page, size_t count)
4109 char *p = (char *) page;
4111 *var = simple_strtoul(p, &p, 10);
4115 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4116 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4118 struct cfq_data *cfqd = e->elevator_data; \
4119 unsigned int __data = __VAR; \
4121 __data = jiffies_to_msecs(__data); \
4122 return cfq_var_show(__data, (page)); \
4124 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4125 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4126 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4127 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4128 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4129 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4130 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4131 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4132 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4133 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4134 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4135 #undef SHOW_FUNCTION
4137 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4138 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4140 struct cfq_data *cfqd = e->elevator_data; \
4141 unsigned int __data; \
4142 int ret = cfq_var_store(&__data, (page), count); \
4143 if (__data < (MIN)) \
4145 else if (__data > (MAX)) \
4148 *(__PTR) = msecs_to_jiffies(__data); \
4150 *(__PTR) = __data; \
4153 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4154 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4156 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4158 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4159 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4161 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4162 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4163 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4164 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4165 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4167 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4168 #undef STORE_FUNCTION
4170 #define CFQ_ATTR(name) \
4171 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4173 static struct elv_fs_entry cfq_attrs[] = {
4175 CFQ_ATTR(fifo_expire_sync),
4176 CFQ_ATTR(fifo_expire_async),
4177 CFQ_ATTR(back_seek_max),
4178 CFQ_ATTR(back_seek_penalty),
4179 CFQ_ATTR(slice_sync),
4180 CFQ_ATTR(slice_async),
4181 CFQ_ATTR(slice_async_rq),
4182 CFQ_ATTR(slice_idle),
4183 CFQ_ATTR(group_idle),
4184 CFQ_ATTR(low_latency),
4188 static struct elevator_type iosched_cfq = {
4190 .elevator_merge_fn = cfq_merge,
4191 .elevator_merged_fn = cfq_merged_request,
4192 .elevator_merge_req_fn = cfq_merged_requests,
4193 .elevator_allow_merge_fn = cfq_allow_merge,
4194 .elevator_bio_merged_fn = cfq_bio_merged,
4195 .elevator_dispatch_fn = cfq_dispatch_requests,
4196 .elevator_add_req_fn = cfq_insert_request,
4197 .elevator_activate_req_fn = cfq_activate_request,
4198 .elevator_deactivate_req_fn = cfq_deactivate_request,
4199 .elevator_completed_req_fn = cfq_completed_request,
4200 .elevator_former_req_fn = elv_rb_former_request,
4201 .elevator_latter_req_fn = elv_rb_latter_request,
4202 .elevator_set_req_fn = cfq_set_request,
4203 .elevator_put_req_fn = cfq_put_request,
4204 .elevator_may_queue_fn = cfq_may_queue,
4205 .elevator_init_fn = cfq_init_queue,
4206 .elevator_exit_fn = cfq_exit_queue,
4207 .trim = cfq_free_io_context,
4209 .elevator_attrs = cfq_attrs,
4210 .elevator_name = "cfq",
4211 .elevator_owner = THIS_MODULE,
4214 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4215 static struct blkio_policy_type blkio_policy_cfq = {
4217 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4218 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4220 .plid = BLKIO_POLICY_PROP,
4223 static struct blkio_policy_type blkio_policy_cfq;
4226 static int __init cfq_init(void)
4229 * could be 0 on HZ < 1000 setups
4231 if (!cfq_slice_async)
4232 cfq_slice_async = 1;
4233 if (!cfq_slice_idle)
4236 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4237 if (!cfq_group_idle)
4242 if (cfq_slab_setup())
4245 elv_register(&iosched_cfq);
4246 blkio_policy_register(&blkio_policy_cfq);
4251 static void __exit cfq_exit(void)
4253 DECLARE_COMPLETION_ONSTACK(all_gone);
4254 blkio_policy_unregister(&blkio_policy_cfq);
4255 elv_unregister(&iosched_cfq);
4256 ioc_gone = &all_gone;
4257 /* ioc_gone's update must be visible before reading ioc_count */
4261 * this also protects us from entering cfq_slab_kill() with
4262 * pending RCU callbacks
4264 if (elv_ioc_count_read(cfq_ioc_count))
4265 wait_for_completion(&all_gone);
4266 ida_destroy(&cic_index_ida);
4270 module_init(cfq_init);
4271 module_exit(cfq_exit);
4273 MODULE_AUTHOR("Jens Axboe");
4274 MODULE_LICENSE("GPL");
4275 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");