2 * CFQ, or complete fairness queueing, disk scheduler.
4 * Based on ideas from a previously unfinished io
5 * scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
7 * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
9 #include <linux/module.h>
10 #include <linux/blkdev.h>
11 #include <linux/elevator.h>
12 #include <linux/jiffies.h>
13 #include <linux/rbtree.h>
14 #include <linux/ioprio.h>
15 #include <linux/blktrace_api.h>
20 /* max queue in one round of service */
21 static const int cfq_quantum = 4;
22 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
23 /* maximum backwards seek, in KiB */
24 static const int cfq_back_max = 16 * 1024;
25 /* penalty of a backwards seek */
26 static const int cfq_back_penalty = 2;
27 static const int cfq_slice_sync = HZ / 10;
28 static int cfq_slice_async = HZ / 25;
29 static const int cfq_slice_async_rq = 2;
30 static int cfq_slice_idle = HZ / 125;
31 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
32 static const int cfq_hist_divisor = 4;
35 * offset from end of service tree
37 #define CFQ_IDLE_DELAY (HZ / 5)
40 * below this threshold, we consider thinktime immediate
42 #define CFQ_MIN_TT (2)
45 * Allow merged cfqqs to perform this amount of seeky I/O before
46 * deciding to break the queues up again.
48 #define CFQQ_COOP_TOUT (HZ)
50 #define CFQ_SLICE_SCALE (5)
51 #define CFQ_HW_QUEUE_MIN (5)
54 ((struct cfq_io_context *) (rq)->elevator_private)
55 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
57 static struct kmem_cache *cfq_pool;
58 static struct kmem_cache *cfq_ioc_pool;
60 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
61 static struct completion *ioc_gone;
62 static DEFINE_SPINLOCK(ioc_gone_lock);
64 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
68 #define sample_valid(samples) ((samples) > 80)
71 * Most of our rbtree usage is for sorting with min extraction, so
72 * if we cache the leftmost node we don't have to walk down the tree
73 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
74 * move this into the elevator for the rq sorting as well.
81 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, }
84 * Per process-grouping structure
89 /* various state flags, see below */
92 struct cfq_data *cfqd;
93 /* service_tree member */
94 struct rb_node rb_node;
95 /* service_tree key */
97 /* prio tree member */
98 struct rb_node p_node;
99 /* prio tree root we belong to, if any */
100 struct rb_root *p_root;
101 /* sorted list of pending requests */
102 struct rb_root sort_list;
103 /* if fifo isn't expired, next request to serve */
104 struct request *next_rq;
105 /* requests queued in sort_list */
107 /* currently allocated requests */
109 /* fifo list of requests in sort_list */
110 struct list_head fifo;
112 unsigned long slice_end;
114 unsigned int slice_dispatch;
116 /* pending metadata requests */
118 /* number of requests that are on the dispatch list or inside driver */
121 /* io prio of this group */
122 unsigned short ioprio, org_ioprio;
123 unsigned short ioprio_class, org_ioprio_class;
125 unsigned int seek_samples;
128 sector_t last_request_pos;
129 unsigned long seeky_start;
133 struct cfq_rb_root *service_tree;
134 struct cfq_queue *new_cfqq;
135 struct cfq_group *cfqg;
139 * First index in the service_trees.
140 * IDLE is handled separately, so it has negative index
149 * Second index in the service_trees.
153 SYNC_NOIDLE_WORKLOAD = 1,
157 /* This is per cgroup per device grouping structure */
160 * rr lists of queues with requests, onle rr for each priority class.
161 * Counts are embedded in the cfq_rb_root
163 struct cfq_rb_root service_trees[2][3];
164 struct cfq_rb_root service_tree_idle;
168 * Per block device queue structure
171 struct request_queue *queue;
172 struct cfq_group root_group;
175 * The priority currently being served
177 enum wl_prio_t serving_prio;
178 enum wl_type_t serving_type;
179 unsigned long workload_expires;
180 struct cfq_group *serving_group;
181 bool noidle_tree_requires_idle;
184 * Each priority tree is sorted by next_request position. These
185 * trees are used when determining if two or more queues are
186 * interleaving requests (see cfq_close_cooperator).
188 struct rb_root prio_trees[CFQ_PRIO_LISTS];
190 unsigned int busy_queues;
191 unsigned int busy_queues_avg[2];
197 * queue-depth detection
203 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
204 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
207 int hw_tag_est_depth;
208 unsigned int hw_tag_samples;
211 * idle window management
213 struct timer_list idle_slice_timer;
214 struct work_struct unplug_work;
216 struct cfq_queue *active_queue;
217 struct cfq_io_context *active_cic;
220 * async queue for each priority case
222 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
223 struct cfq_queue *async_idle_cfqq;
225 sector_t last_position;
228 * tunables, see top of file
230 unsigned int cfq_quantum;
231 unsigned int cfq_fifo_expire[2];
232 unsigned int cfq_back_penalty;
233 unsigned int cfq_back_max;
234 unsigned int cfq_slice[2];
235 unsigned int cfq_slice_async_rq;
236 unsigned int cfq_slice_idle;
237 unsigned int cfq_latency;
239 struct list_head cic_list;
242 * Fallback dummy cfqq for extreme OOM conditions
244 struct cfq_queue oom_cfqq;
246 unsigned long last_end_sync_rq;
249 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
252 struct cfq_data *cfqd)
254 if (prio == IDLE_WORKLOAD)
255 return &cfqg->service_tree_idle;
257 return &cfqg->service_trees[prio][type];
260 enum cfqq_state_flags {
261 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
262 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
263 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
264 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
265 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
266 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
267 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
268 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
269 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
270 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
271 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
274 #define CFQ_CFQQ_FNS(name) \
275 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
277 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
279 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
281 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
283 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
285 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
289 CFQ_CFQQ_FNS(wait_request);
290 CFQ_CFQQ_FNS(must_dispatch);
291 CFQ_CFQQ_FNS(must_alloc_slice);
292 CFQ_CFQQ_FNS(fifo_expire);
293 CFQ_CFQQ_FNS(idle_window);
294 CFQ_CFQQ_FNS(prio_changed);
295 CFQ_CFQQ_FNS(slice_new);
301 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
302 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
303 #define cfq_log(cfqd, fmt, args...) \
304 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
306 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
308 if (cfq_class_idle(cfqq))
309 return IDLE_WORKLOAD;
310 if (cfq_class_rt(cfqq))
316 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
318 if (!cfq_cfqq_sync(cfqq))
319 return ASYNC_WORKLOAD;
320 if (!cfq_cfqq_idle_window(cfqq))
321 return SYNC_NOIDLE_WORKLOAD;
322 return SYNC_WORKLOAD;
325 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
327 struct cfq_group *cfqg = &cfqd->root_group;
329 if (wl == IDLE_WORKLOAD)
330 return cfqg->service_tree_idle.count;
332 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
333 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
334 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
337 static void cfq_dispatch_insert(struct request_queue *, struct request *);
338 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
339 struct io_context *, gfp_t);
340 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
341 struct io_context *);
343 static inline int rq_in_driver(struct cfq_data *cfqd)
345 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
348 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
351 return cic->cfqq[is_sync];
354 static inline void cic_set_cfqq(struct cfq_io_context *cic,
355 struct cfq_queue *cfqq, bool is_sync)
357 cic->cfqq[is_sync] = cfqq;
361 * We regard a request as SYNC, if it's either a read or has the SYNC bit
362 * set (in which case it could also be direct WRITE).
364 static inline bool cfq_bio_sync(struct bio *bio)
366 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
370 * scheduler run of queue, if there are requests pending and no one in the
371 * driver that will restart queueing
373 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
375 if (cfqd->busy_queues) {
376 cfq_log(cfqd, "schedule dispatch");
377 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
381 static int cfq_queue_empty(struct request_queue *q)
383 struct cfq_data *cfqd = q->elevator->elevator_data;
385 return !cfqd->busy_queues;
389 * Scale schedule slice based on io priority. Use the sync time slice only
390 * if a queue is marked sync and has sync io queued. A sync queue with async
391 * io only, should not get full sync slice length.
393 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
396 const int base_slice = cfqd->cfq_slice[sync];
398 WARN_ON(prio >= IOPRIO_BE_NR);
400 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
404 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
406 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
410 * get averaged number of queues of RT/BE priority.
411 * average is updated, with a formula that gives more weight to higher numbers,
412 * to quickly follows sudden increases and decrease slowly
415 static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt)
417 unsigned min_q, max_q;
418 unsigned mult = cfq_hist_divisor - 1;
419 unsigned round = cfq_hist_divisor / 2;
420 unsigned busy = cfq_busy_queues_wl(rt, cfqd);
422 min_q = min(cfqd->busy_queues_avg[rt], busy);
423 max_q = max(cfqd->busy_queues_avg[rt], busy);
424 cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
426 return cfqd->busy_queues_avg[rt];
430 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
432 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
433 if (cfqd->cfq_latency) {
434 /* interested queues (we consider only the ones with the same
436 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
437 unsigned sync_slice = cfqd->cfq_slice[1];
438 unsigned expect_latency = sync_slice * iq;
439 if (expect_latency > cfq_target_latency) {
440 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
441 /* scale low_slice according to IO priority
442 * and sync vs async */
444 min(slice, base_low_slice * slice / sync_slice);
445 /* the adapted slice value is scaled to fit all iqs
446 * into the target latency */
447 slice = max(slice * cfq_target_latency / expect_latency,
451 cfqq->slice_end = jiffies + slice;
452 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
456 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
457 * isn't valid until the first request from the dispatch is activated
458 * and the slice time set.
460 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
462 if (cfq_cfqq_slice_new(cfqq))
464 if (time_before(jiffies, cfqq->slice_end))
471 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
472 * We choose the request that is closest to the head right now. Distance
473 * behind the head is penalized and only allowed to a certain extent.
475 static struct request *
476 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
478 sector_t s1, s2, d1 = 0, d2 = 0;
479 unsigned long back_max;
480 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
481 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
482 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
484 if (rq1 == NULL || rq1 == rq2)
489 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
491 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
493 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
495 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
498 s1 = blk_rq_pos(rq1);
499 s2 = blk_rq_pos(rq2);
502 * by definition, 1KiB is 2 sectors
504 back_max = cfqd->cfq_back_max * 2;
507 * Strict one way elevator _except_ in the case where we allow
508 * short backward seeks which are biased as twice the cost of a
509 * similar forward seek.
513 else if (s1 + back_max >= last)
514 d1 = (last - s1) * cfqd->cfq_back_penalty;
516 wrap |= CFQ_RQ1_WRAP;
520 else if (s2 + back_max >= last)
521 d2 = (last - s2) * cfqd->cfq_back_penalty;
523 wrap |= CFQ_RQ2_WRAP;
525 /* Found required data */
528 * By doing switch() on the bit mask "wrap" we avoid having to
529 * check two variables for all permutations: --> faster!
532 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
548 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
551 * Since both rqs are wrapped,
552 * start with the one that's further behind head
553 * (--> only *one* back seek required),
554 * since back seek takes more time than forward.
564 * The below is leftmost cache rbtree addon
566 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
569 root->left = rb_first(&root->rb);
572 return rb_entry(root->left, struct cfq_queue, rb_node);
577 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
583 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
587 rb_erase_init(n, &root->rb);
592 * would be nice to take fifo expire time into account as well
594 static struct request *
595 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
596 struct request *last)
598 struct rb_node *rbnext = rb_next(&last->rb_node);
599 struct rb_node *rbprev = rb_prev(&last->rb_node);
600 struct request *next = NULL, *prev = NULL;
602 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
605 prev = rb_entry_rq(rbprev);
608 next = rb_entry_rq(rbnext);
610 rbnext = rb_first(&cfqq->sort_list);
611 if (rbnext && rbnext != &last->rb_node)
612 next = rb_entry_rq(rbnext);
615 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
618 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
619 struct cfq_queue *cfqq)
622 * just an approximation, should be ok.
624 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
625 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
629 * The cfqd->service_trees holds all pending cfq_queue's that have
630 * requests waiting to be processed. It is sorted in the order that
631 * we will service the queues.
633 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
636 struct rb_node **p, *parent;
637 struct cfq_queue *__cfqq;
638 unsigned long rb_key;
639 struct cfq_rb_root *service_tree;
642 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
643 cfqq_type(cfqq), cfqd);
644 if (cfq_class_idle(cfqq)) {
645 rb_key = CFQ_IDLE_DELAY;
646 parent = rb_last(&service_tree->rb);
647 if (parent && parent != &cfqq->rb_node) {
648 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
649 rb_key += __cfqq->rb_key;
652 } else if (!add_front) {
654 * Get our rb key offset. Subtract any residual slice
655 * value carried from last service. A negative resid
656 * count indicates slice overrun, and this should position
657 * the next service time further away in the tree.
659 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
660 rb_key -= cfqq->slice_resid;
661 cfqq->slice_resid = 0;
664 __cfqq = cfq_rb_first(service_tree);
665 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
668 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
670 * same position, nothing more to do
672 if (rb_key == cfqq->rb_key &&
673 cfqq->service_tree == service_tree)
676 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
677 cfqq->service_tree = NULL;
682 cfqq->service_tree = service_tree;
683 p = &service_tree->rb.rb_node;
688 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
691 * sort by key, that represents service time.
693 if (time_before(rb_key, __cfqq->rb_key))
704 service_tree->left = &cfqq->rb_node;
706 cfqq->rb_key = rb_key;
707 rb_link_node(&cfqq->rb_node, parent, p);
708 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
709 service_tree->count++;
712 static struct cfq_queue *
713 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
714 sector_t sector, struct rb_node **ret_parent,
715 struct rb_node ***rb_link)
717 struct rb_node **p, *parent;
718 struct cfq_queue *cfqq = NULL;
726 cfqq = rb_entry(parent, struct cfq_queue, p_node);
729 * Sort strictly based on sector. Smallest to the left,
730 * largest to the right.
732 if (sector > blk_rq_pos(cfqq->next_rq))
734 else if (sector < blk_rq_pos(cfqq->next_rq))
742 *ret_parent = parent;
748 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
750 struct rb_node **p, *parent;
751 struct cfq_queue *__cfqq;
754 rb_erase(&cfqq->p_node, cfqq->p_root);
758 if (cfq_class_idle(cfqq))
763 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
764 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
765 blk_rq_pos(cfqq->next_rq), &parent, &p);
767 rb_link_node(&cfqq->p_node, parent, p);
768 rb_insert_color(&cfqq->p_node, cfqq->p_root);
774 * Update cfqq's position in the service tree.
776 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
779 * Resorting requires the cfqq to be on the RR list already.
781 if (cfq_cfqq_on_rr(cfqq)) {
782 cfq_service_tree_add(cfqd, cfqq, 0);
783 cfq_prio_tree_add(cfqd, cfqq);
788 * add to busy list of queues for service, trying to be fair in ordering
789 * the pending list according to last request service
791 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
793 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
794 BUG_ON(cfq_cfqq_on_rr(cfqq));
795 cfq_mark_cfqq_on_rr(cfqq);
798 cfq_resort_rr_list(cfqd, cfqq);
802 * Called when the cfqq no longer has requests pending, remove it from
805 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
807 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
808 BUG_ON(!cfq_cfqq_on_rr(cfqq));
809 cfq_clear_cfqq_on_rr(cfqq);
811 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
812 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
813 cfqq->service_tree = NULL;
816 rb_erase(&cfqq->p_node, cfqq->p_root);
820 BUG_ON(!cfqd->busy_queues);
825 * rb tree support functions
827 static void cfq_del_rq_rb(struct request *rq)
829 struct cfq_queue *cfqq = RQ_CFQQ(rq);
830 struct cfq_data *cfqd = cfqq->cfqd;
831 const int sync = rq_is_sync(rq);
833 BUG_ON(!cfqq->queued[sync]);
834 cfqq->queued[sync]--;
836 elv_rb_del(&cfqq->sort_list, rq);
838 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
839 cfq_del_cfqq_rr(cfqd, cfqq);
842 static void cfq_add_rq_rb(struct request *rq)
844 struct cfq_queue *cfqq = RQ_CFQQ(rq);
845 struct cfq_data *cfqd = cfqq->cfqd;
846 struct request *__alias, *prev;
848 cfqq->queued[rq_is_sync(rq)]++;
851 * looks a little odd, but the first insert might return an alias.
852 * if that happens, put the alias on the dispatch list
854 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
855 cfq_dispatch_insert(cfqd->queue, __alias);
857 if (!cfq_cfqq_on_rr(cfqq))
858 cfq_add_cfqq_rr(cfqd, cfqq);
861 * check if this request is a better next-serve candidate
863 prev = cfqq->next_rq;
864 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
867 * adjust priority tree position, if ->next_rq changes
869 if (prev != cfqq->next_rq)
870 cfq_prio_tree_add(cfqd, cfqq);
872 BUG_ON(!cfqq->next_rq);
875 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
877 elv_rb_del(&cfqq->sort_list, rq);
878 cfqq->queued[rq_is_sync(rq)]--;
882 static struct request *
883 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
885 struct task_struct *tsk = current;
886 struct cfq_io_context *cic;
887 struct cfq_queue *cfqq;
889 cic = cfq_cic_lookup(cfqd, tsk->io_context);
893 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
895 sector_t sector = bio->bi_sector + bio_sectors(bio);
897 return elv_rb_find(&cfqq->sort_list, sector);
903 static void cfq_activate_request(struct request_queue *q, struct request *rq)
905 struct cfq_data *cfqd = q->elevator->elevator_data;
907 cfqd->rq_in_driver[rq_is_sync(rq)]++;
908 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
911 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
914 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
916 struct cfq_data *cfqd = q->elevator->elevator_data;
917 const int sync = rq_is_sync(rq);
919 WARN_ON(!cfqd->rq_in_driver[sync]);
920 cfqd->rq_in_driver[sync]--;
921 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
925 static void cfq_remove_request(struct request *rq)
927 struct cfq_queue *cfqq = RQ_CFQQ(rq);
929 if (cfqq->next_rq == rq)
930 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
932 list_del_init(&rq->queuelist);
935 cfqq->cfqd->rq_queued--;
936 if (rq_is_meta(rq)) {
937 WARN_ON(!cfqq->meta_pending);
938 cfqq->meta_pending--;
942 static int cfq_merge(struct request_queue *q, struct request **req,
945 struct cfq_data *cfqd = q->elevator->elevator_data;
946 struct request *__rq;
948 __rq = cfq_find_rq_fmerge(cfqd, bio);
949 if (__rq && elv_rq_merge_ok(__rq, bio)) {
951 return ELEVATOR_FRONT_MERGE;
954 return ELEVATOR_NO_MERGE;
957 static void cfq_merged_request(struct request_queue *q, struct request *req,
960 if (type == ELEVATOR_FRONT_MERGE) {
961 struct cfq_queue *cfqq = RQ_CFQQ(req);
963 cfq_reposition_rq_rb(cfqq, req);
968 cfq_merged_requests(struct request_queue *q, struct request *rq,
969 struct request *next)
971 struct cfq_queue *cfqq = RQ_CFQQ(rq);
973 * reposition in fifo if next is older than rq
975 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
976 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
977 list_move(&rq->queuelist, &next->queuelist);
978 rq_set_fifo_time(rq, rq_fifo_time(next));
981 if (cfqq->next_rq == next)
983 cfq_remove_request(next);
986 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
989 struct cfq_data *cfqd = q->elevator->elevator_data;
990 struct cfq_io_context *cic;
991 struct cfq_queue *cfqq;
994 * Disallow merge of a sync bio into an async request.
996 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1000 * Lookup the cfqq that this bio will be queued with. Allow
1001 * merge only if rq is queued there.
1003 cic = cfq_cic_lookup(cfqd, current->io_context);
1007 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1008 return cfqq == RQ_CFQQ(rq);
1011 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1012 struct cfq_queue *cfqq)
1015 cfq_log_cfqq(cfqd, cfqq, "set_active");
1016 cfqq->slice_end = 0;
1017 cfqq->slice_dispatch = 0;
1019 cfq_clear_cfqq_wait_request(cfqq);
1020 cfq_clear_cfqq_must_dispatch(cfqq);
1021 cfq_clear_cfqq_must_alloc_slice(cfqq);
1022 cfq_clear_cfqq_fifo_expire(cfqq);
1023 cfq_mark_cfqq_slice_new(cfqq);
1025 del_timer(&cfqd->idle_slice_timer);
1028 cfqd->active_queue = cfqq;
1032 * current cfqq expired its slice (or was too idle), select new one
1035 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1038 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1040 if (cfq_cfqq_wait_request(cfqq))
1041 del_timer(&cfqd->idle_slice_timer);
1043 cfq_clear_cfqq_wait_request(cfqq);
1046 * store what was left of this slice, if the queue idled/timed out
1048 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1049 cfqq->slice_resid = cfqq->slice_end - jiffies;
1050 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1053 cfq_resort_rr_list(cfqd, cfqq);
1055 if (cfqq == cfqd->active_queue)
1056 cfqd->active_queue = NULL;
1058 if (cfqd->active_cic) {
1059 put_io_context(cfqd->active_cic->ioc);
1060 cfqd->active_cic = NULL;
1064 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1066 struct cfq_queue *cfqq = cfqd->active_queue;
1069 __cfq_slice_expired(cfqd, cfqq, timed_out);
1073 * Get next queue for service. Unless we have a queue preemption,
1074 * we'll simply select the first cfqq in the service tree.
1076 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1078 struct cfq_rb_root *service_tree =
1079 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1080 cfqd->serving_type, cfqd);
1082 if (RB_EMPTY_ROOT(&service_tree->rb))
1084 return cfq_rb_first(service_tree);
1088 * Get and set a new active queue for service.
1090 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1091 struct cfq_queue *cfqq)
1094 cfqq = cfq_get_next_queue(cfqd);
1096 __cfq_set_active_queue(cfqd, cfqq);
1100 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1103 if (blk_rq_pos(rq) >= cfqd->last_position)
1104 return blk_rq_pos(rq) - cfqd->last_position;
1106 return cfqd->last_position - blk_rq_pos(rq);
1109 #define CFQQ_SEEK_THR 8 * 1024
1110 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1112 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1115 sector_t sdist = cfqq->seek_mean;
1117 if (!sample_valid(cfqq->seek_samples))
1118 sdist = CFQQ_SEEK_THR;
1120 return cfq_dist_from_last(cfqd, rq) <= sdist;
1123 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1124 struct cfq_queue *cur_cfqq)
1126 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1127 struct rb_node *parent, *node;
1128 struct cfq_queue *__cfqq;
1129 sector_t sector = cfqd->last_position;
1131 if (RB_EMPTY_ROOT(root))
1135 * First, if we find a request starting at the end of the last
1136 * request, choose it.
1138 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1143 * If the exact sector wasn't found, the parent of the NULL leaf
1144 * will contain the closest sector.
1146 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1147 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1150 if (blk_rq_pos(__cfqq->next_rq) < sector)
1151 node = rb_next(&__cfqq->p_node);
1153 node = rb_prev(&__cfqq->p_node);
1157 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1158 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1166 * cur_cfqq - passed in so that we don't decide that the current queue is
1167 * closely cooperating with itself.
1169 * So, basically we're assuming that that cur_cfqq has dispatched at least
1170 * one request, and that cfqd->last_position reflects a position on the disk
1171 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1174 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1175 struct cfq_queue *cur_cfqq)
1177 struct cfq_queue *cfqq;
1179 if (!cfq_cfqq_sync(cur_cfqq))
1181 if (CFQQ_SEEKY(cur_cfqq))
1185 * We should notice if some of the queues are cooperating, eg
1186 * working closely on the same area of the disk. In that case,
1187 * we can group them together and don't waste time idling.
1189 cfqq = cfqq_close(cfqd, cur_cfqq);
1194 * It only makes sense to merge sync queues.
1196 if (!cfq_cfqq_sync(cfqq))
1198 if (CFQQ_SEEKY(cfqq))
1202 * Do not merge queues of different priority classes
1204 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1211 * Determine whether we should enforce idle window for this queue.
1214 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1216 enum wl_prio_t prio = cfqq_prio(cfqq);
1217 struct cfq_rb_root *service_tree = cfqq->service_tree;
1219 /* We never do for idle class queues. */
1220 if (prio == IDLE_WORKLOAD)
1223 /* We do for queues that were marked with idle window flag. */
1224 if (cfq_cfqq_idle_window(cfqq))
1228 * Otherwise, we do only if they are the last ones
1229 * in their service tree.
1232 service_tree = service_tree_for(cfqq->cfqg, prio,
1233 cfqq_type(cfqq), cfqd);
1235 if (service_tree->count == 0)
1238 return (service_tree->count == 1 && cfq_rb_first(service_tree) == cfqq);
1241 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1243 struct cfq_queue *cfqq = cfqd->active_queue;
1244 struct cfq_io_context *cic;
1248 * SSD device without seek penalty, disable idling. But only do so
1249 * for devices that support queuing, otherwise we still have a problem
1250 * with sync vs async workloads.
1252 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1255 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1256 WARN_ON(cfq_cfqq_slice_new(cfqq));
1259 * idle is disabled, either manually or by past process history
1261 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1265 * still active requests from this queue, don't idle
1267 if (cfqq->dispatched)
1271 * task has exited, don't wait
1273 cic = cfqd->active_cic;
1274 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1278 * If our average think time is larger than the remaining time
1279 * slice, then don't idle. This avoids overrunning the allotted
1282 if (sample_valid(cic->ttime_samples) &&
1283 (cfqq->slice_end - jiffies < cic->ttime_mean))
1286 cfq_mark_cfqq_wait_request(cfqq);
1288 sl = cfqd->cfq_slice_idle;
1290 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1291 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1295 * Move request from internal lists to the request queue dispatch list.
1297 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1299 struct cfq_data *cfqd = q->elevator->elevator_data;
1300 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1302 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1304 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1305 cfq_remove_request(rq);
1307 elv_dispatch_sort(q, rq);
1309 if (cfq_cfqq_sync(cfqq))
1310 cfqd->sync_flight++;
1314 * return expired entry, or NULL to just start from scratch in rbtree
1316 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1318 struct request *rq = NULL;
1320 if (cfq_cfqq_fifo_expire(cfqq))
1323 cfq_mark_cfqq_fifo_expire(cfqq);
1325 if (list_empty(&cfqq->fifo))
1328 rq = rq_entry_fifo(cfqq->fifo.next);
1329 if (time_before(jiffies, rq_fifo_time(rq)))
1332 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1337 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1339 const int base_rq = cfqd->cfq_slice_async_rq;
1341 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1343 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1347 * Must be called with the queue_lock held.
1349 static int cfqq_process_refs(struct cfq_queue *cfqq)
1351 int process_refs, io_refs;
1353 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1354 process_refs = atomic_read(&cfqq->ref) - io_refs;
1355 BUG_ON(process_refs < 0);
1356 return process_refs;
1359 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1361 int process_refs, new_process_refs;
1362 struct cfq_queue *__cfqq;
1364 /* Avoid a circular list and skip interim queue merges */
1365 while ((__cfqq = new_cfqq->new_cfqq)) {
1371 process_refs = cfqq_process_refs(cfqq);
1373 * If the process for the cfqq has gone away, there is no
1374 * sense in merging the queues.
1376 if (process_refs == 0)
1380 * Merge in the direction of the lesser amount of work.
1382 new_process_refs = cfqq_process_refs(new_cfqq);
1383 if (new_process_refs >= process_refs) {
1384 cfqq->new_cfqq = new_cfqq;
1385 atomic_add(process_refs, &new_cfqq->ref);
1387 new_cfqq->new_cfqq = cfqq;
1388 atomic_add(new_process_refs, &cfqq->ref);
1392 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
1393 struct cfq_group *cfqg, enum wl_prio_t prio,
1396 struct cfq_queue *queue;
1398 bool key_valid = false;
1399 unsigned long lowest_key = 0;
1400 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1404 * When priorities switched, we prefer starting
1405 * from SYNC_NOIDLE (first choice), or just SYNC
1408 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1410 cur_best = SYNC_WORKLOAD;
1411 if (service_tree_for(cfqg, prio, cur_best, cfqd)->count)
1414 return ASYNC_WORKLOAD;
1417 for (i = 0; i < 3; ++i) {
1418 /* otherwise, select the one with lowest rb_key */
1419 queue = cfq_rb_first(service_tree_for(cfqg, prio, i, cfqd));
1421 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1422 lowest_key = queue->rb_key;
1431 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
1433 enum wl_prio_t previous_prio = cfqd->serving_prio;
1437 struct cfq_rb_root *st;
1439 /* Choose next priority. RT > BE > IDLE */
1440 if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1441 cfqd->serving_prio = RT_WORKLOAD;
1442 else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1443 cfqd->serving_prio = BE_WORKLOAD;
1445 cfqd->serving_prio = IDLE_WORKLOAD;
1446 cfqd->workload_expires = jiffies + 1;
1451 * For RT and BE, we have to choose also the type
1452 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1455 prio_changed = (cfqd->serving_prio != previous_prio);
1456 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1461 * If priority didn't change, check workload expiration,
1462 * and that we still have other queues ready
1464 if (!prio_changed && count &&
1465 !time_after(jiffies, cfqd->workload_expires))
1468 /* otherwise select new workload type */
1469 cfqd->serving_type =
1470 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio, prio_changed);
1471 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type,
1476 * the workload slice is computed as a fraction of target latency
1477 * proportional to the number of queues in that workload, over
1478 * all the queues in the same priority class
1480 slice = cfq_target_latency * count /
1481 max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio],
1482 cfq_busy_queues_wl(cfqd->serving_prio, cfqd));
1484 if (cfqd->serving_type == ASYNC_WORKLOAD)
1485 /* async workload slice is scaled down according to
1486 * the sync/async slice ratio. */
1487 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1489 /* sync workload slice is at least 2 * cfq_slice_idle */
1490 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1492 slice = max_t(unsigned, slice, CFQ_MIN_TT);
1493 cfqd->workload_expires = jiffies + slice;
1494 cfqd->noidle_tree_requires_idle = false;
1497 static void cfq_choose_cfqg(struct cfq_data *cfqd)
1499 cfqd->serving_group = &cfqd->root_group;
1500 choose_service_tree(cfqd, &cfqd->root_group);
1504 * Select a queue for service. If we have a current active queue,
1505 * check whether to continue servicing it, or retrieve and set a new one.
1507 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1509 struct cfq_queue *cfqq, *new_cfqq = NULL;
1511 cfqq = cfqd->active_queue;
1516 * The active queue has run out of time, expire it and select new.
1518 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1522 * The active queue has requests and isn't expired, allow it to
1525 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1529 * If another queue has a request waiting within our mean seek
1530 * distance, let it run. The expire code will check for close
1531 * cooperators and put the close queue at the front of the service
1532 * tree. If possible, merge the expiring queue with the new cfqq.
1534 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1536 if (!cfqq->new_cfqq)
1537 cfq_setup_merge(cfqq, new_cfqq);
1542 * No requests pending. If the active queue still has requests in
1543 * flight or is idling for a new request, allow either of these
1544 * conditions to happen (or time out) before selecting a new queue.
1546 if (timer_pending(&cfqd->idle_slice_timer) ||
1547 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1553 cfq_slice_expired(cfqd, 0);
1556 * Current queue expired. Check if we have to switch to a new
1560 cfq_choose_cfqg(cfqd);
1562 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1567 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1571 while (cfqq->next_rq) {
1572 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1576 BUG_ON(!list_empty(&cfqq->fifo));
1581 * Drain our current requests. Used for barriers and when switching
1582 * io schedulers on-the-fly.
1584 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1586 struct cfq_queue *cfqq;
1589 struct cfq_group *cfqg = &cfqd->root_group;
1591 for (i = 0; i < 2; ++i)
1592 for (j = 0; j < 3; ++j)
1593 while ((cfqq = cfq_rb_first(&cfqg->service_trees[i][j]))
1595 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1597 while ((cfqq = cfq_rb_first(&cfqg->service_tree_idle)) != NULL)
1598 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1600 cfq_slice_expired(cfqd, 0);
1602 BUG_ON(cfqd->busy_queues);
1604 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1608 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1610 unsigned int max_dispatch;
1613 * Drain async requests before we start sync IO
1615 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1619 * If this is an async queue and we have sync IO in flight, let it wait
1621 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1624 max_dispatch = cfqd->cfq_quantum;
1625 if (cfq_class_idle(cfqq))
1629 * Does this cfqq already have too much IO in flight?
1631 if (cfqq->dispatched >= max_dispatch) {
1633 * idle queue must always only have a single IO in flight
1635 if (cfq_class_idle(cfqq))
1639 * We have other queues, don't allow more IO from this one
1641 if (cfqd->busy_queues > 1)
1645 * Sole queue user, no limit
1651 * Async queues must wait a bit before being allowed dispatch.
1652 * We also ramp up the dispatch depth gradually for async IO,
1653 * based on the last sync IO we serviced
1655 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1656 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1659 depth = last_sync / cfqd->cfq_slice[1];
1660 if (!depth && !cfqq->dispatched)
1662 if (depth < max_dispatch)
1663 max_dispatch = depth;
1667 * If we're below the current max, allow a dispatch
1669 return cfqq->dispatched < max_dispatch;
1673 * Dispatch a request from cfqq, moving them to the request queue
1676 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1680 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1682 if (!cfq_may_dispatch(cfqd, cfqq))
1686 * follow expired path, else get first next available
1688 rq = cfq_check_fifo(cfqq);
1693 * insert request into driver dispatch list
1695 cfq_dispatch_insert(cfqd->queue, rq);
1697 if (!cfqd->active_cic) {
1698 struct cfq_io_context *cic = RQ_CIC(rq);
1700 atomic_long_inc(&cic->ioc->refcount);
1701 cfqd->active_cic = cic;
1708 * Find the cfqq that we need to service and move a request from that to the
1711 static int cfq_dispatch_requests(struct request_queue *q, int force)
1713 struct cfq_data *cfqd = q->elevator->elevator_data;
1714 struct cfq_queue *cfqq;
1716 if (!cfqd->busy_queues)
1719 if (unlikely(force))
1720 return cfq_forced_dispatch(cfqd);
1722 cfqq = cfq_select_queue(cfqd);
1727 * Dispatch a request from this cfqq, if it is allowed
1729 if (!cfq_dispatch_request(cfqd, cfqq))
1732 cfqq->slice_dispatch++;
1733 cfq_clear_cfqq_must_dispatch(cfqq);
1736 * expire an async queue immediately if it has used up its slice. idle
1737 * queue always expire after 1 dispatch round.
1739 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1740 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1741 cfq_class_idle(cfqq))) {
1742 cfqq->slice_end = jiffies + 1;
1743 cfq_slice_expired(cfqd, 0);
1746 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1751 * task holds one reference to the queue, dropped when task exits. each rq
1752 * in-flight on this queue also holds a reference, dropped when rq is freed.
1754 * queue lock must be held here.
1756 static void cfq_put_queue(struct cfq_queue *cfqq)
1758 struct cfq_data *cfqd = cfqq->cfqd;
1760 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1762 if (!atomic_dec_and_test(&cfqq->ref))
1765 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1766 BUG_ON(rb_first(&cfqq->sort_list));
1767 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1768 BUG_ON(cfq_cfqq_on_rr(cfqq));
1770 if (unlikely(cfqd->active_queue == cfqq)) {
1771 __cfq_slice_expired(cfqd, cfqq, 0);
1772 cfq_schedule_dispatch(cfqd);
1775 kmem_cache_free(cfq_pool, cfqq);
1779 * Must always be called with the rcu_read_lock() held
1782 __call_for_each_cic(struct io_context *ioc,
1783 void (*func)(struct io_context *, struct cfq_io_context *))
1785 struct cfq_io_context *cic;
1786 struct hlist_node *n;
1788 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1793 * Call func for each cic attached to this ioc.
1796 call_for_each_cic(struct io_context *ioc,
1797 void (*func)(struct io_context *, struct cfq_io_context *))
1800 __call_for_each_cic(ioc, func);
1804 static void cfq_cic_free_rcu(struct rcu_head *head)
1806 struct cfq_io_context *cic;
1808 cic = container_of(head, struct cfq_io_context, rcu_head);
1810 kmem_cache_free(cfq_ioc_pool, cic);
1811 elv_ioc_count_dec(cfq_ioc_count);
1815 * CFQ scheduler is exiting, grab exit lock and check
1816 * the pending io context count. If it hits zero,
1817 * complete ioc_gone and set it back to NULL
1819 spin_lock(&ioc_gone_lock);
1820 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1824 spin_unlock(&ioc_gone_lock);
1828 static void cfq_cic_free(struct cfq_io_context *cic)
1830 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1833 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1835 unsigned long flags;
1837 BUG_ON(!cic->dead_key);
1839 spin_lock_irqsave(&ioc->lock, flags);
1840 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1841 hlist_del_rcu(&cic->cic_list);
1842 spin_unlock_irqrestore(&ioc->lock, flags);
1848 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1849 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1850 * and ->trim() which is called with the task lock held
1852 static void cfq_free_io_context(struct io_context *ioc)
1855 * ioc->refcount is zero here, or we are called from elv_unregister(),
1856 * so no more cic's are allowed to be linked into this ioc. So it
1857 * should be ok to iterate over the known list, we will see all cic's
1858 * since no new ones are added.
1860 __call_for_each_cic(ioc, cic_free_func);
1863 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1865 struct cfq_queue *__cfqq, *next;
1867 if (unlikely(cfqq == cfqd->active_queue)) {
1868 __cfq_slice_expired(cfqd, cfqq, 0);
1869 cfq_schedule_dispatch(cfqd);
1873 * If this queue was scheduled to merge with another queue, be
1874 * sure to drop the reference taken on that queue (and others in
1875 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1877 __cfqq = cfqq->new_cfqq;
1879 if (__cfqq == cfqq) {
1880 WARN(1, "cfqq->new_cfqq loop detected\n");
1883 next = __cfqq->new_cfqq;
1884 cfq_put_queue(__cfqq);
1888 cfq_put_queue(cfqq);
1891 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1892 struct cfq_io_context *cic)
1894 struct io_context *ioc = cic->ioc;
1896 list_del_init(&cic->queue_list);
1899 * Make sure key == NULL is seen for dead queues
1902 cic->dead_key = (unsigned long) cic->key;
1905 if (ioc->ioc_data == cic)
1906 rcu_assign_pointer(ioc->ioc_data, NULL);
1908 if (cic->cfqq[BLK_RW_ASYNC]) {
1909 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1910 cic->cfqq[BLK_RW_ASYNC] = NULL;
1913 if (cic->cfqq[BLK_RW_SYNC]) {
1914 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1915 cic->cfqq[BLK_RW_SYNC] = NULL;
1919 static void cfq_exit_single_io_context(struct io_context *ioc,
1920 struct cfq_io_context *cic)
1922 struct cfq_data *cfqd = cic->key;
1925 struct request_queue *q = cfqd->queue;
1926 unsigned long flags;
1928 spin_lock_irqsave(q->queue_lock, flags);
1931 * Ensure we get a fresh copy of the ->key to prevent
1932 * race between exiting task and queue
1934 smp_read_barrier_depends();
1936 __cfq_exit_single_io_context(cfqd, cic);
1938 spin_unlock_irqrestore(q->queue_lock, flags);
1943 * The process that ioc belongs to has exited, we need to clean up
1944 * and put the internal structures we have that belongs to that process.
1946 static void cfq_exit_io_context(struct io_context *ioc)
1948 call_for_each_cic(ioc, cfq_exit_single_io_context);
1951 static struct cfq_io_context *
1952 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1954 struct cfq_io_context *cic;
1956 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1959 cic->last_end_request = jiffies;
1960 INIT_LIST_HEAD(&cic->queue_list);
1961 INIT_HLIST_NODE(&cic->cic_list);
1962 cic->dtor = cfq_free_io_context;
1963 cic->exit = cfq_exit_io_context;
1964 elv_ioc_count_inc(cfq_ioc_count);
1970 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1972 struct task_struct *tsk = current;
1975 if (!cfq_cfqq_prio_changed(cfqq))
1978 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1979 switch (ioprio_class) {
1981 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1982 case IOPRIO_CLASS_NONE:
1984 * no prio set, inherit CPU scheduling settings
1986 cfqq->ioprio = task_nice_ioprio(tsk);
1987 cfqq->ioprio_class = task_nice_ioclass(tsk);
1989 case IOPRIO_CLASS_RT:
1990 cfqq->ioprio = task_ioprio(ioc);
1991 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1993 case IOPRIO_CLASS_BE:
1994 cfqq->ioprio = task_ioprio(ioc);
1995 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1997 case IOPRIO_CLASS_IDLE:
1998 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2000 cfq_clear_cfqq_idle_window(cfqq);
2005 * keep track of original prio settings in case we have to temporarily
2006 * elevate the priority of this queue
2008 cfqq->org_ioprio = cfqq->ioprio;
2009 cfqq->org_ioprio_class = cfqq->ioprio_class;
2010 cfq_clear_cfqq_prio_changed(cfqq);
2013 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2015 struct cfq_data *cfqd = cic->key;
2016 struct cfq_queue *cfqq;
2017 unsigned long flags;
2019 if (unlikely(!cfqd))
2022 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2024 cfqq = cic->cfqq[BLK_RW_ASYNC];
2026 struct cfq_queue *new_cfqq;
2027 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2030 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2031 cfq_put_queue(cfqq);
2035 cfqq = cic->cfqq[BLK_RW_SYNC];
2037 cfq_mark_cfqq_prio_changed(cfqq);
2039 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2042 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2044 call_for_each_cic(ioc, changed_ioprio);
2045 ioc->ioprio_changed = 0;
2048 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2049 pid_t pid, bool is_sync)
2051 RB_CLEAR_NODE(&cfqq->rb_node);
2052 RB_CLEAR_NODE(&cfqq->p_node);
2053 INIT_LIST_HEAD(&cfqq->fifo);
2055 atomic_set(&cfqq->ref, 0);
2058 cfq_mark_cfqq_prio_changed(cfqq);
2061 if (!cfq_class_idle(cfqq))
2062 cfq_mark_cfqq_idle_window(cfqq);
2063 cfq_mark_cfqq_sync(cfqq);
2068 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
2073 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
2075 return &cfqd->root_group;
2078 static struct cfq_queue *
2079 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2080 struct io_context *ioc, gfp_t gfp_mask)
2082 struct cfq_queue *cfqq, *new_cfqq = NULL;
2083 struct cfq_io_context *cic;
2084 struct cfq_group *cfqg;
2087 cfqg = cfq_get_cfqg(cfqd, 1);
2088 cic = cfq_cic_lookup(cfqd, ioc);
2089 /* cic always exists here */
2090 cfqq = cic_to_cfqq(cic, is_sync);
2093 * Always try a new alloc if we fell back to the OOM cfqq
2094 * originally, since it should just be a temporary situation.
2096 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2101 } else if (gfp_mask & __GFP_WAIT) {
2102 spin_unlock_irq(cfqd->queue->queue_lock);
2103 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2104 gfp_mask | __GFP_ZERO,
2106 spin_lock_irq(cfqd->queue->queue_lock);
2110 cfqq = kmem_cache_alloc_node(cfq_pool,
2111 gfp_mask | __GFP_ZERO,
2116 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2117 cfq_init_prio_data(cfqq, ioc);
2118 cfq_link_cfqq_cfqg(cfqq, cfqg);
2119 cfq_log_cfqq(cfqd, cfqq, "alloced");
2121 cfqq = &cfqd->oom_cfqq;
2125 kmem_cache_free(cfq_pool, new_cfqq);
2130 static struct cfq_queue **
2131 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2133 switch (ioprio_class) {
2134 case IOPRIO_CLASS_RT:
2135 return &cfqd->async_cfqq[0][ioprio];
2136 case IOPRIO_CLASS_BE:
2137 return &cfqd->async_cfqq[1][ioprio];
2138 case IOPRIO_CLASS_IDLE:
2139 return &cfqd->async_idle_cfqq;
2145 static struct cfq_queue *
2146 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2149 const int ioprio = task_ioprio(ioc);
2150 const int ioprio_class = task_ioprio_class(ioc);
2151 struct cfq_queue **async_cfqq = NULL;
2152 struct cfq_queue *cfqq = NULL;
2155 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2160 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2163 * pin the queue now that it's allocated, scheduler exit will prune it
2165 if (!is_sync && !(*async_cfqq)) {
2166 atomic_inc(&cfqq->ref);
2170 atomic_inc(&cfqq->ref);
2175 * We drop cfq io contexts lazily, so we may find a dead one.
2178 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2179 struct cfq_io_context *cic)
2181 unsigned long flags;
2183 WARN_ON(!list_empty(&cic->queue_list));
2185 spin_lock_irqsave(&ioc->lock, flags);
2187 BUG_ON(ioc->ioc_data == cic);
2189 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2190 hlist_del_rcu(&cic->cic_list);
2191 spin_unlock_irqrestore(&ioc->lock, flags);
2196 static struct cfq_io_context *
2197 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2199 struct cfq_io_context *cic;
2200 unsigned long flags;
2209 * we maintain a last-hit cache, to avoid browsing over the tree
2211 cic = rcu_dereference(ioc->ioc_data);
2212 if (cic && cic->key == cfqd) {
2218 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2222 /* ->key must be copied to avoid race with cfq_exit_queue() */
2225 cfq_drop_dead_cic(cfqd, ioc, cic);
2230 spin_lock_irqsave(&ioc->lock, flags);
2231 rcu_assign_pointer(ioc->ioc_data, cic);
2232 spin_unlock_irqrestore(&ioc->lock, flags);
2240 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2241 * the process specific cfq io context when entered from the block layer.
2242 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2244 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2245 struct cfq_io_context *cic, gfp_t gfp_mask)
2247 unsigned long flags;
2250 ret = radix_tree_preload(gfp_mask);
2255 spin_lock_irqsave(&ioc->lock, flags);
2256 ret = radix_tree_insert(&ioc->radix_root,
2257 (unsigned long) cfqd, cic);
2259 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2260 spin_unlock_irqrestore(&ioc->lock, flags);
2262 radix_tree_preload_end();
2265 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2266 list_add(&cic->queue_list, &cfqd->cic_list);
2267 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2272 printk(KERN_ERR "cfq: cic link failed!\n");
2278 * Setup general io context and cfq io context. There can be several cfq
2279 * io contexts per general io context, if this process is doing io to more
2280 * than one device managed by cfq.
2282 static struct cfq_io_context *
2283 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2285 struct io_context *ioc = NULL;
2286 struct cfq_io_context *cic;
2288 might_sleep_if(gfp_mask & __GFP_WAIT);
2290 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2294 cic = cfq_cic_lookup(cfqd, ioc);
2298 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2302 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2306 smp_read_barrier_depends();
2307 if (unlikely(ioc->ioprio_changed))
2308 cfq_ioc_set_ioprio(ioc);
2314 put_io_context(ioc);
2319 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2321 unsigned long elapsed = jiffies - cic->last_end_request;
2322 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2324 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2325 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2326 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2330 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2336 if (!cfqq->last_request_pos)
2338 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2339 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2341 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2344 * Don't allow the seek distance to get too large from the
2345 * odd fragment, pagein, etc
2347 if (cfqq->seek_samples <= 60) /* second&third seek */
2348 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2350 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2352 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2353 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2354 total = cfqq->seek_total + (cfqq->seek_samples/2);
2355 do_div(total, cfqq->seek_samples);
2356 cfqq->seek_mean = (sector_t)total;
2359 * If this cfqq is shared between multiple processes, check to
2360 * make sure that those processes are still issuing I/Os within
2361 * the mean seek distance. If not, it may be time to break the
2362 * queues apart again.
2364 if (cfq_cfqq_coop(cfqq)) {
2365 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2366 cfqq->seeky_start = jiffies;
2367 else if (!CFQQ_SEEKY(cfqq))
2368 cfqq->seeky_start = 0;
2373 * Disable idle window if the process thinks too long or seeks so much that
2377 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2378 struct cfq_io_context *cic)
2380 int old_idle, enable_idle;
2383 * Don't idle for async or idle io prio class
2385 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2388 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2390 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
2391 cfq_mark_cfqq_deep(cfqq);
2393 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2394 (!cfq_cfqq_deep(cfqq) && sample_valid(cfqq->seek_samples)
2395 && CFQQ_SEEKY(cfqq)))
2397 else if (sample_valid(cic->ttime_samples)) {
2398 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2404 if (old_idle != enable_idle) {
2405 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2407 cfq_mark_cfqq_idle_window(cfqq);
2409 cfq_clear_cfqq_idle_window(cfqq);
2414 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2415 * no or if we aren't sure, a 1 will cause a preempt.
2418 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2421 struct cfq_queue *cfqq;
2423 cfqq = cfqd->active_queue;
2427 if (cfq_slice_used(cfqq))
2430 if (cfq_class_idle(new_cfqq))
2433 if (cfq_class_idle(cfqq))
2436 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
2437 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
2438 new_cfqq->service_tree->count == 1)
2442 * if the new request is sync, but the currently running queue is
2443 * not, let the sync request have priority.
2445 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2449 * So both queues are sync. Let the new request get disk time if
2450 * it's a metadata request and the current queue is doing regular IO.
2452 if (rq_is_meta(rq) && !cfqq->meta_pending)
2456 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2458 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2461 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2465 * if this request is as-good as one we would expect from the
2466 * current cfqq, let it preempt
2468 if (cfq_rq_close(cfqd, cfqq, rq))
2475 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2476 * let it have half of its nominal slice.
2478 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2480 cfq_log_cfqq(cfqd, cfqq, "preempt");
2481 cfq_slice_expired(cfqd, 1);
2484 * Put the new queue at the front of the of the current list,
2485 * so we know that it will be selected next.
2487 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2489 cfq_service_tree_add(cfqd, cfqq, 1);
2491 cfqq->slice_end = 0;
2492 cfq_mark_cfqq_slice_new(cfqq);
2496 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2497 * something we should do about it
2500 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2503 struct cfq_io_context *cic = RQ_CIC(rq);
2507 cfqq->meta_pending++;
2509 cfq_update_io_thinktime(cfqd, cic);
2510 cfq_update_io_seektime(cfqd, cfqq, rq);
2511 cfq_update_idle_window(cfqd, cfqq, cic);
2513 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2515 if (cfqq == cfqd->active_queue) {
2517 * Remember that we saw a request from this process, but
2518 * don't start queuing just yet. Otherwise we risk seeing lots
2519 * of tiny requests, because we disrupt the normal plugging
2520 * and merging. If the request is already larger than a single
2521 * page, let it rip immediately. For that case we assume that
2522 * merging is already done. Ditto for a busy system that
2523 * has other work pending, don't risk delaying until the
2524 * idle timer unplug to continue working.
2526 if (cfq_cfqq_wait_request(cfqq)) {
2527 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2528 cfqd->busy_queues > 1) {
2529 del_timer(&cfqd->idle_slice_timer);
2530 __blk_run_queue(cfqd->queue);
2532 cfq_mark_cfqq_must_dispatch(cfqq);
2534 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2536 * not the active queue - expire current slice if it is
2537 * idle and has expired it's mean thinktime or this new queue
2538 * has some old slice time left and is of higher priority or
2539 * this new queue is RT and the current one is BE
2541 cfq_preempt_queue(cfqd, cfqq);
2542 __blk_run_queue(cfqd->queue);
2546 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2548 struct cfq_data *cfqd = q->elevator->elevator_data;
2549 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2551 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2552 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2554 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2555 list_add_tail(&rq->queuelist, &cfqq->fifo);
2558 cfq_rq_enqueued(cfqd, cfqq, rq);
2562 * Update hw_tag based on peak queue depth over 50 samples under
2565 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2567 struct cfq_queue *cfqq = cfqd->active_queue;
2569 if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
2570 cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
2572 if (cfqd->hw_tag == 1)
2575 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2576 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2580 * If active queue hasn't enough requests and can idle, cfq might not
2581 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2584 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2585 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2586 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2589 if (cfqd->hw_tag_samples++ < 50)
2592 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
2598 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2600 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2601 struct cfq_data *cfqd = cfqq->cfqd;
2602 const int sync = rq_is_sync(rq);
2606 cfq_log_cfqq(cfqd, cfqq, "complete");
2608 cfq_update_hw_tag(cfqd);
2610 WARN_ON(!cfqd->rq_in_driver[sync]);
2611 WARN_ON(!cfqq->dispatched);
2612 cfqd->rq_in_driver[sync]--;
2615 if (cfq_cfqq_sync(cfqq))
2616 cfqd->sync_flight--;
2619 RQ_CIC(rq)->last_end_request = now;
2620 cfqd->last_end_sync_rq = now;
2624 * If this is the active queue, check if it needs to be expired,
2625 * or if we want to idle in case it has no pending requests.
2627 if (cfqd->active_queue == cfqq) {
2628 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2630 if (cfq_cfqq_slice_new(cfqq)) {
2631 cfq_set_prio_slice(cfqd, cfqq);
2632 cfq_clear_cfqq_slice_new(cfqq);
2635 * Idling is not enabled on:
2637 * - idle-priority queues
2639 * - queues with still some requests queued
2640 * - when there is a close cooperator
2642 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2643 cfq_slice_expired(cfqd, 1);
2644 else if (sync && cfqq_empty &&
2645 !cfq_close_cooperator(cfqd, cfqq)) {
2646 cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
2648 * Idling is enabled for SYNC_WORKLOAD.
2649 * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
2650 * only if we processed at least one !rq_noidle request
2652 if (cfqd->serving_type == SYNC_WORKLOAD
2653 || cfqd->noidle_tree_requires_idle)
2654 cfq_arm_slice_timer(cfqd);
2658 if (!rq_in_driver(cfqd))
2659 cfq_schedule_dispatch(cfqd);
2663 * we temporarily boost lower priority queues if they are holding fs exclusive
2664 * resources. they are boosted to normal prio (CLASS_BE/4)
2666 static void cfq_prio_boost(struct cfq_queue *cfqq)
2668 if (has_fs_excl()) {
2670 * boost idle prio on transactions that would lock out other
2671 * users of the filesystem
2673 if (cfq_class_idle(cfqq))
2674 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2675 if (cfqq->ioprio > IOPRIO_NORM)
2676 cfqq->ioprio = IOPRIO_NORM;
2679 * unboost the queue (if needed)
2681 cfqq->ioprio_class = cfqq->org_ioprio_class;
2682 cfqq->ioprio = cfqq->org_ioprio;
2686 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2688 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2689 cfq_mark_cfqq_must_alloc_slice(cfqq);
2690 return ELV_MQUEUE_MUST;
2693 return ELV_MQUEUE_MAY;
2696 static int cfq_may_queue(struct request_queue *q, int rw)
2698 struct cfq_data *cfqd = q->elevator->elevator_data;
2699 struct task_struct *tsk = current;
2700 struct cfq_io_context *cic;
2701 struct cfq_queue *cfqq;
2704 * don't force setup of a queue from here, as a call to may_queue
2705 * does not necessarily imply that a request actually will be queued.
2706 * so just lookup a possibly existing queue, or return 'may queue'
2709 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2711 return ELV_MQUEUE_MAY;
2713 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2715 cfq_init_prio_data(cfqq, cic->ioc);
2716 cfq_prio_boost(cfqq);
2718 return __cfq_may_queue(cfqq);
2721 return ELV_MQUEUE_MAY;
2725 * queue lock held here
2727 static void cfq_put_request(struct request *rq)
2729 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2732 const int rw = rq_data_dir(rq);
2734 BUG_ON(!cfqq->allocated[rw]);
2735 cfqq->allocated[rw]--;
2737 put_io_context(RQ_CIC(rq)->ioc);
2739 rq->elevator_private = NULL;
2740 rq->elevator_private2 = NULL;
2742 cfq_put_queue(cfqq);
2746 static struct cfq_queue *
2747 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2748 struct cfq_queue *cfqq)
2750 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2751 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2752 cfq_mark_cfqq_coop(cfqq->new_cfqq);
2753 cfq_put_queue(cfqq);
2754 return cic_to_cfqq(cic, 1);
2757 static int should_split_cfqq(struct cfq_queue *cfqq)
2759 if (cfqq->seeky_start &&
2760 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2766 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2767 * was the last process referring to said cfqq.
2769 static struct cfq_queue *
2770 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2772 if (cfqq_process_refs(cfqq) == 1) {
2773 cfqq->seeky_start = 0;
2774 cfqq->pid = current->pid;
2775 cfq_clear_cfqq_coop(cfqq);
2779 cic_set_cfqq(cic, NULL, 1);
2780 cfq_put_queue(cfqq);
2784 * Allocate cfq data structures associated with this request.
2787 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2789 struct cfq_data *cfqd = q->elevator->elevator_data;
2790 struct cfq_io_context *cic;
2791 const int rw = rq_data_dir(rq);
2792 const bool is_sync = rq_is_sync(rq);
2793 struct cfq_queue *cfqq;
2794 unsigned long flags;
2796 might_sleep_if(gfp_mask & __GFP_WAIT);
2798 cic = cfq_get_io_context(cfqd, gfp_mask);
2800 spin_lock_irqsave(q->queue_lock, flags);
2806 cfqq = cic_to_cfqq(cic, is_sync);
2807 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2808 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2809 cic_set_cfqq(cic, cfqq, is_sync);
2812 * If the queue was seeky for too long, break it apart.
2814 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2815 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2816 cfqq = split_cfqq(cic, cfqq);
2822 * Check to see if this queue is scheduled to merge with
2823 * another, closely cooperating queue. The merging of
2824 * queues happens here as it must be done in process context.
2825 * The reference on new_cfqq was taken in merge_cfqqs.
2828 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2831 cfqq->allocated[rw]++;
2832 atomic_inc(&cfqq->ref);
2834 spin_unlock_irqrestore(q->queue_lock, flags);
2836 rq->elevator_private = cic;
2837 rq->elevator_private2 = cfqq;
2842 put_io_context(cic->ioc);
2844 cfq_schedule_dispatch(cfqd);
2845 spin_unlock_irqrestore(q->queue_lock, flags);
2846 cfq_log(cfqd, "set_request fail");
2850 static void cfq_kick_queue(struct work_struct *work)
2852 struct cfq_data *cfqd =
2853 container_of(work, struct cfq_data, unplug_work);
2854 struct request_queue *q = cfqd->queue;
2856 spin_lock_irq(q->queue_lock);
2857 __blk_run_queue(cfqd->queue);
2858 spin_unlock_irq(q->queue_lock);
2862 * Timer running if the active_queue is currently idling inside its time slice
2864 static void cfq_idle_slice_timer(unsigned long data)
2866 struct cfq_data *cfqd = (struct cfq_data *) data;
2867 struct cfq_queue *cfqq;
2868 unsigned long flags;
2871 cfq_log(cfqd, "idle timer fired");
2873 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2875 cfqq = cfqd->active_queue;
2880 * We saw a request before the queue expired, let it through
2882 if (cfq_cfqq_must_dispatch(cfqq))
2888 if (cfq_slice_used(cfqq))
2892 * only expire and reinvoke request handler, if there are
2893 * other queues with pending requests
2895 if (!cfqd->busy_queues)
2899 * not expired and it has a request pending, let it dispatch
2901 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2905 * Queue depth flag is reset only when the idle didn't succeed
2907 cfq_clear_cfqq_deep(cfqq);
2910 cfq_slice_expired(cfqd, timed_out);
2912 cfq_schedule_dispatch(cfqd);
2914 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2917 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2919 del_timer_sync(&cfqd->idle_slice_timer);
2920 cancel_work_sync(&cfqd->unplug_work);
2923 static void cfq_put_async_queues(struct cfq_data *cfqd)
2927 for (i = 0; i < IOPRIO_BE_NR; i++) {
2928 if (cfqd->async_cfqq[0][i])
2929 cfq_put_queue(cfqd->async_cfqq[0][i]);
2930 if (cfqd->async_cfqq[1][i])
2931 cfq_put_queue(cfqd->async_cfqq[1][i]);
2934 if (cfqd->async_idle_cfqq)
2935 cfq_put_queue(cfqd->async_idle_cfqq);
2938 static void cfq_exit_queue(struct elevator_queue *e)
2940 struct cfq_data *cfqd = e->elevator_data;
2941 struct request_queue *q = cfqd->queue;
2943 cfq_shutdown_timer_wq(cfqd);
2945 spin_lock_irq(q->queue_lock);
2947 if (cfqd->active_queue)
2948 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2950 while (!list_empty(&cfqd->cic_list)) {
2951 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2952 struct cfq_io_context,
2955 __cfq_exit_single_io_context(cfqd, cic);
2958 cfq_put_async_queues(cfqd);
2960 spin_unlock_irq(q->queue_lock);
2962 cfq_shutdown_timer_wq(cfqd);
2967 static void *cfq_init_queue(struct request_queue *q)
2969 struct cfq_data *cfqd;
2971 struct cfq_group *cfqg;
2973 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2977 /* Init root group */
2978 cfqg = &cfqd->root_group;
2980 for (i = 0; i < 2; ++i)
2981 for (j = 0; j < 3; ++j)
2982 cfqg->service_trees[i][j] = CFQ_RB_ROOT;
2983 cfqg->service_tree_idle = CFQ_RB_ROOT;
2986 * Not strictly needed (since RB_ROOT just clears the node and we
2987 * zeroed cfqd on alloc), but better be safe in case someone decides
2988 * to add magic to the rb code
2990 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2991 cfqd->prio_trees[i] = RB_ROOT;
2994 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2995 * Grab a permanent reference to it, so that the normal code flow
2996 * will not attempt to free it.
2998 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2999 atomic_inc(&cfqd->oom_cfqq.ref);
3000 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3002 INIT_LIST_HEAD(&cfqd->cic_list);
3006 init_timer(&cfqd->idle_slice_timer);
3007 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3008 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3010 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3012 cfqd->cfq_quantum = cfq_quantum;
3013 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3014 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3015 cfqd->cfq_back_max = cfq_back_max;
3016 cfqd->cfq_back_penalty = cfq_back_penalty;
3017 cfqd->cfq_slice[0] = cfq_slice_async;
3018 cfqd->cfq_slice[1] = cfq_slice_sync;
3019 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3020 cfqd->cfq_slice_idle = cfq_slice_idle;
3021 cfqd->cfq_latency = 1;
3023 cfqd->last_end_sync_rq = jiffies;
3027 static void cfq_slab_kill(void)
3030 * Caller already ensured that pending RCU callbacks are completed,
3031 * so we should have no busy allocations at this point.
3034 kmem_cache_destroy(cfq_pool);
3036 kmem_cache_destroy(cfq_ioc_pool);
3039 static int __init cfq_slab_setup(void)
3041 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3045 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3056 * sysfs parts below -->
3059 cfq_var_show(unsigned int var, char *page)
3061 return sprintf(page, "%d\n", var);
3065 cfq_var_store(unsigned int *var, const char *page, size_t count)
3067 char *p = (char *) page;
3069 *var = simple_strtoul(p, &p, 10);
3073 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3074 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3076 struct cfq_data *cfqd = e->elevator_data; \
3077 unsigned int __data = __VAR; \
3079 __data = jiffies_to_msecs(__data); \
3080 return cfq_var_show(__data, (page)); \
3082 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3083 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3084 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3085 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3086 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3087 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3088 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3089 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3090 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3091 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3092 #undef SHOW_FUNCTION
3094 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3095 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3097 struct cfq_data *cfqd = e->elevator_data; \
3098 unsigned int __data; \
3099 int ret = cfq_var_store(&__data, (page), count); \
3100 if (__data < (MIN)) \
3102 else if (__data > (MAX)) \
3105 *(__PTR) = msecs_to_jiffies(__data); \
3107 *(__PTR) = __data; \
3110 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3111 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3113 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3115 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3116 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3118 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3119 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3120 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3121 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3123 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3124 #undef STORE_FUNCTION
3126 #define CFQ_ATTR(name) \
3127 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3129 static struct elv_fs_entry cfq_attrs[] = {
3131 CFQ_ATTR(fifo_expire_sync),
3132 CFQ_ATTR(fifo_expire_async),
3133 CFQ_ATTR(back_seek_max),
3134 CFQ_ATTR(back_seek_penalty),
3135 CFQ_ATTR(slice_sync),
3136 CFQ_ATTR(slice_async),
3137 CFQ_ATTR(slice_async_rq),
3138 CFQ_ATTR(slice_idle),
3139 CFQ_ATTR(low_latency),
3143 static struct elevator_type iosched_cfq = {
3145 .elevator_merge_fn = cfq_merge,
3146 .elevator_merged_fn = cfq_merged_request,
3147 .elevator_merge_req_fn = cfq_merged_requests,
3148 .elevator_allow_merge_fn = cfq_allow_merge,
3149 .elevator_dispatch_fn = cfq_dispatch_requests,
3150 .elevator_add_req_fn = cfq_insert_request,
3151 .elevator_activate_req_fn = cfq_activate_request,
3152 .elevator_deactivate_req_fn = cfq_deactivate_request,
3153 .elevator_queue_empty_fn = cfq_queue_empty,
3154 .elevator_completed_req_fn = cfq_completed_request,
3155 .elevator_former_req_fn = elv_rb_former_request,
3156 .elevator_latter_req_fn = elv_rb_latter_request,
3157 .elevator_set_req_fn = cfq_set_request,
3158 .elevator_put_req_fn = cfq_put_request,
3159 .elevator_may_queue_fn = cfq_may_queue,
3160 .elevator_init_fn = cfq_init_queue,
3161 .elevator_exit_fn = cfq_exit_queue,
3162 .trim = cfq_free_io_context,
3164 .elevator_attrs = cfq_attrs,
3165 .elevator_name = "cfq",
3166 .elevator_owner = THIS_MODULE,
3169 static int __init cfq_init(void)
3172 * could be 0 on HZ < 1000 setups
3174 if (!cfq_slice_async)
3175 cfq_slice_async = 1;
3176 if (!cfq_slice_idle)
3179 if (cfq_slab_setup())
3182 elv_register(&iosched_cfq);
3187 static void __exit cfq_exit(void)
3189 DECLARE_COMPLETION_ONSTACK(all_gone);
3190 elv_unregister(&iosched_cfq);
3191 ioc_gone = &all_gone;
3192 /* ioc_gone's update must be visible before reading ioc_count */
3196 * this also protects us from entering cfq_slab_kill() with
3197 * pending RCU callbacks
3199 if (elv_ioc_count_read(cfq_ioc_count))
3200 wait_for_completion(&all_gone);
3204 module_init(cfq_init);
3205 module_exit(cfq_exit);
3207 MODULE_AUTHOR("Jens Axboe");
3208 MODULE_LICENSE("GPL");
3209 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");