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/rbtree.h>
13 #include <linux/ioprio.h>
14 #include <linux/blktrace_api.h>
19 /* max queue in one round of service */
20 static const int cfq_quantum = 4;
21 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
22 /* maximum backwards seek, in KiB */
23 static const int cfq_back_max = 16 * 1024;
24 /* penalty of a backwards seek */
25 static const int cfq_back_penalty = 2;
26 static const int cfq_slice_sync = HZ / 10;
27 static int cfq_slice_async = HZ / 25;
28 static const int cfq_slice_async_rq = 2;
29 static int cfq_slice_idle = HZ / 125;
30 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
31 static const int cfq_hist_divisor = 4;
34 * offset from end of service tree
36 #define CFQ_IDLE_DELAY (HZ / 5)
39 * below this threshold, we consider thinktime immediate
41 #define CFQ_MIN_TT (2)
44 * Allow merged cfqqs to perform this amount of seeky I/O before
45 * deciding to break the queues up again.
47 #define CFQQ_COOP_TOUT (HZ)
49 #define CFQ_SLICE_SCALE (5)
50 #define CFQ_HW_QUEUE_MIN (5)
53 ((struct cfq_io_context *) (rq)->elevator_private)
54 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private2)
56 static struct kmem_cache *cfq_pool;
57 static struct kmem_cache *cfq_ioc_pool;
59 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
60 static struct completion *ioc_gone;
61 static DEFINE_SPINLOCK(ioc_gone_lock);
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
70 * Most of our rbtree usage is for sorting with min extraction, so
71 * if we cache the leftmost node we don't have to walk down the tree
72 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
73 * move this into the elevator for the rq sorting as well.
80 #define CFQ_RB_ROOT (struct cfq_rb_root) { RB_ROOT, NULL, 0, }
83 * Per process-grouping structure
88 /* various state flags, see below */
91 struct cfq_data *cfqd;
92 /* service_tree member */
93 struct rb_node rb_node;
94 /* service_tree key */
96 /* prio tree member */
97 struct rb_node p_node;
98 /* prio tree root we belong to, if any */
99 struct rb_root *p_root;
100 /* sorted list of pending requests */
101 struct rb_root sort_list;
102 /* if fifo isn't expired, next request to serve */
103 struct request *next_rq;
104 /* requests queued in sort_list */
106 /* currently allocated requests */
108 /* fifo list of requests in sort_list */
109 struct list_head fifo;
111 unsigned long slice_end;
113 unsigned int slice_dispatch;
115 /* pending metadata requests */
117 /* number of requests that are on the dispatch list or inside driver */
120 /* io prio of this group */
121 unsigned short ioprio, org_ioprio;
122 unsigned short ioprio_class, org_ioprio_class;
124 unsigned int seek_samples;
127 sector_t last_request_pos;
128 unsigned long seeky_start;
132 struct cfq_rb_root *service_tree;
133 struct cfq_queue *new_cfqq;
137 * First index in the service_trees.
138 * IDLE is handled separately, so it has negative index
147 * Second index in the service_trees.
151 SYNC_NOIDLE_WORKLOAD = 1,
157 * Per block device queue structure
160 struct request_queue *queue;
163 * rr lists of queues with requests, onle rr for each priority class.
164 * Counts are embedded in the cfq_rb_root
166 struct cfq_rb_root service_trees[2][3];
167 struct cfq_rb_root service_tree_idle;
169 * The priority currently being served
171 enum wl_prio_t serving_prio;
172 enum wl_type_t serving_type;
173 unsigned long workload_expires;
176 * Each priority tree is sorted by next_request position. These
177 * trees are used when determining if two or more queues are
178 * interleaving requests (see cfq_close_cooperator).
180 struct rb_root prio_trees[CFQ_PRIO_LISTS];
182 unsigned int busy_queues;
183 unsigned int busy_queues_avg[2];
189 * queue-depth detection
194 int rq_in_driver_peak;
197 * idle window management
199 struct timer_list idle_slice_timer;
200 struct work_struct unplug_work;
202 struct cfq_queue *active_queue;
203 struct cfq_io_context *active_cic;
206 * async queue for each priority case
208 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
209 struct cfq_queue *async_idle_cfqq;
211 sector_t last_position;
214 * tunables, see top of file
216 unsigned int cfq_quantum;
217 unsigned int cfq_fifo_expire[2];
218 unsigned int cfq_back_penalty;
219 unsigned int cfq_back_max;
220 unsigned int cfq_slice[2];
221 unsigned int cfq_slice_async_rq;
222 unsigned int cfq_slice_idle;
223 unsigned int cfq_latency;
225 struct list_head cic_list;
228 * Fallback dummy cfqq for extreme OOM conditions
230 struct cfq_queue oom_cfqq;
232 unsigned long last_end_sync_rq;
235 static struct cfq_rb_root *service_tree_for(enum wl_prio_t prio,
237 struct cfq_data *cfqd)
239 if (prio == IDLE_WORKLOAD)
240 return &cfqd->service_tree_idle;
242 return &cfqd->service_trees[prio][type];
245 enum cfqq_state_flags {
246 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
247 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
248 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
249 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
250 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
251 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
252 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
253 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
254 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
255 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
256 CFQ_CFQQ_FLAG_coop_preempt, /* coop preempt */
259 #define CFQ_CFQQ_FNS(name) \
260 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
262 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
264 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
266 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
268 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
270 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
274 CFQ_CFQQ_FNS(wait_request);
275 CFQ_CFQQ_FNS(must_dispatch);
276 CFQ_CFQQ_FNS(must_alloc_slice);
277 CFQ_CFQQ_FNS(fifo_expire);
278 CFQ_CFQQ_FNS(idle_window);
279 CFQ_CFQQ_FNS(prio_changed);
280 CFQ_CFQQ_FNS(slice_new);
283 CFQ_CFQQ_FNS(coop_preempt);
286 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
287 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
288 #define cfq_log(cfqd, fmt, args...) \
289 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
291 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
293 if (cfq_class_idle(cfqq))
294 return IDLE_WORKLOAD;
295 if (cfq_class_rt(cfqq))
301 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
303 if (!cfq_cfqq_sync(cfqq))
304 return ASYNC_WORKLOAD;
305 if (!cfq_cfqq_idle_window(cfqq))
306 return SYNC_NOIDLE_WORKLOAD;
307 return SYNC_WORKLOAD;
310 static inline int cfq_busy_queues_wl(enum wl_prio_t wl, struct cfq_data *cfqd)
312 if (wl == IDLE_WORKLOAD)
313 return cfqd->service_tree_idle.count;
315 return cfqd->service_trees[wl][ASYNC_WORKLOAD].count
316 + cfqd->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
317 + cfqd->service_trees[wl][SYNC_WORKLOAD].count;
320 static void cfq_dispatch_insert(struct request_queue *, struct request *);
321 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
322 struct io_context *, gfp_t);
323 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
324 struct io_context *);
326 static inline int rq_in_driver(struct cfq_data *cfqd)
328 return cfqd->rq_in_driver[0] + cfqd->rq_in_driver[1];
331 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
334 return cic->cfqq[is_sync];
337 static inline void cic_set_cfqq(struct cfq_io_context *cic,
338 struct cfq_queue *cfqq, bool is_sync)
340 cic->cfqq[is_sync] = cfqq;
344 * We regard a request as SYNC, if it's either a read or has the SYNC bit
345 * set (in which case it could also be direct WRITE).
347 static inline bool cfq_bio_sync(struct bio *bio)
349 return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
353 * scheduler run of queue, if there are requests pending and no one in the
354 * driver that will restart queueing
356 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
358 if (cfqd->busy_queues) {
359 cfq_log(cfqd, "schedule dispatch");
360 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
364 static int cfq_queue_empty(struct request_queue *q)
366 struct cfq_data *cfqd = q->elevator->elevator_data;
368 return !cfqd->busy_queues;
372 * Scale schedule slice based on io priority. Use the sync time slice only
373 * if a queue is marked sync and has sync io queued. A sync queue with async
374 * io only, should not get full sync slice length.
376 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
379 const int base_slice = cfqd->cfq_slice[sync];
381 WARN_ON(prio >= IOPRIO_BE_NR);
383 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
387 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
389 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
393 * get averaged number of queues of RT/BE priority.
394 * average is updated, with a formula that gives more weight to higher numbers,
395 * to quickly follows sudden increases and decrease slowly
398 static inline unsigned cfq_get_avg_queues(struct cfq_data *cfqd, bool rt)
400 unsigned min_q, max_q;
401 unsigned mult = cfq_hist_divisor - 1;
402 unsigned round = cfq_hist_divisor / 2;
403 unsigned busy = cfq_busy_queues_wl(rt, cfqd);
405 min_q = min(cfqd->busy_queues_avg[rt], busy);
406 max_q = max(cfqd->busy_queues_avg[rt], busy);
407 cfqd->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
409 return cfqd->busy_queues_avg[rt];
413 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
415 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
416 if (cfqd->cfq_latency) {
417 /* interested queues (we consider only the ones with the same
419 unsigned iq = cfq_get_avg_queues(cfqd, cfq_class_rt(cfqq));
420 unsigned sync_slice = cfqd->cfq_slice[1];
421 unsigned expect_latency = sync_slice * iq;
422 if (expect_latency > cfq_target_latency) {
423 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
424 /* scale low_slice according to IO priority
425 * and sync vs async */
427 min(slice, base_low_slice * slice / sync_slice);
428 /* the adapted slice value is scaled to fit all iqs
429 * into the target latency */
430 slice = max(slice * cfq_target_latency / expect_latency,
434 cfqq->slice_end = jiffies + slice;
435 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
439 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
440 * isn't valid until the first request from the dispatch is activated
441 * and the slice time set.
443 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
445 if (cfq_cfqq_slice_new(cfqq))
447 if (time_before(jiffies, cfqq->slice_end))
454 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
455 * We choose the request that is closest to the head right now. Distance
456 * behind the head is penalized and only allowed to a certain extent.
458 static struct request *
459 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2)
461 sector_t last, s1, s2, d1 = 0, d2 = 0;
462 unsigned long back_max;
463 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
464 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
465 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
467 if (rq1 == NULL || rq1 == rq2)
472 if (rq_is_sync(rq1) && !rq_is_sync(rq2))
474 else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
476 if (rq_is_meta(rq1) && !rq_is_meta(rq2))
478 else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
481 s1 = blk_rq_pos(rq1);
482 s2 = blk_rq_pos(rq2);
484 last = cfqd->last_position;
487 * by definition, 1KiB is 2 sectors
489 back_max = cfqd->cfq_back_max * 2;
492 * Strict one way elevator _except_ in the case where we allow
493 * short backward seeks which are biased as twice the cost of a
494 * similar forward seek.
498 else if (s1 + back_max >= last)
499 d1 = (last - s1) * cfqd->cfq_back_penalty;
501 wrap |= CFQ_RQ1_WRAP;
505 else if (s2 + back_max >= last)
506 d2 = (last - s2) * cfqd->cfq_back_penalty;
508 wrap |= CFQ_RQ2_WRAP;
510 /* Found required data */
513 * By doing switch() on the bit mask "wrap" we avoid having to
514 * check two variables for all permutations: --> faster!
517 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
533 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
536 * Since both rqs are wrapped,
537 * start with the one that's further behind head
538 * (--> only *one* back seek required),
539 * since back seek takes more time than forward.
549 * The below is leftmost cache rbtree addon
551 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
554 root->left = rb_first(&root->rb);
557 return rb_entry(root->left, struct cfq_queue, rb_node);
562 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
568 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
572 rb_erase_init(n, &root->rb);
577 * would be nice to take fifo expire time into account as well
579 static struct request *
580 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
581 struct request *last)
583 struct rb_node *rbnext = rb_next(&last->rb_node);
584 struct rb_node *rbprev = rb_prev(&last->rb_node);
585 struct request *next = NULL, *prev = NULL;
587 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
590 prev = rb_entry_rq(rbprev);
593 next = rb_entry_rq(rbnext);
595 rbnext = rb_first(&cfqq->sort_list);
596 if (rbnext && rbnext != &last->rb_node)
597 next = rb_entry_rq(rbnext);
600 return cfq_choose_req(cfqd, next, prev);
603 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
604 struct cfq_queue *cfqq)
607 * just an approximation, should be ok.
609 return (cfqd->busy_queues - 1) * (cfq_prio_slice(cfqd, 1, 0) -
610 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
614 * The cfqd->service_trees holds all pending cfq_queue's that have
615 * requests waiting to be processed. It is sorted in the order that
616 * we will service the queues.
618 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
621 struct rb_node **p, *parent;
622 struct cfq_queue *__cfqq;
623 unsigned long rb_key;
624 struct cfq_rb_root *service_tree;
627 service_tree = service_tree_for(cfqq_prio(cfqq), cfqq_type(cfqq), cfqd);
628 if (cfq_class_idle(cfqq)) {
629 rb_key = CFQ_IDLE_DELAY;
630 parent = rb_last(&service_tree->rb);
631 if (parent && parent != &cfqq->rb_node) {
632 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
633 rb_key += __cfqq->rb_key;
636 } else if (!add_front) {
638 * Get our rb key offset. Subtract any residual slice
639 * value carried from last service. A negative resid
640 * count indicates slice overrun, and this should position
641 * the next service time further away in the tree.
643 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
644 rb_key -= cfqq->slice_resid;
645 cfqq->slice_resid = 0;
648 __cfqq = cfq_rb_first(service_tree);
649 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
652 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
654 * same position, nothing more to do
656 if (rb_key == cfqq->rb_key &&
657 cfqq->service_tree == service_tree)
660 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
661 cfqq->service_tree = NULL;
666 cfqq->service_tree = service_tree;
667 p = &service_tree->rb.rb_node;
672 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
675 * sort by key, that represents service time.
677 if (time_before(rb_key, __cfqq->rb_key))
688 service_tree->left = &cfqq->rb_node;
690 cfqq->rb_key = rb_key;
691 rb_link_node(&cfqq->rb_node, parent, p);
692 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
693 service_tree->count++;
696 static struct cfq_queue *
697 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
698 sector_t sector, struct rb_node **ret_parent,
699 struct rb_node ***rb_link)
701 struct rb_node **p, *parent;
702 struct cfq_queue *cfqq = NULL;
710 cfqq = rb_entry(parent, struct cfq_queue, p_node);
713 * Sort strictly based on sector. Smallest to the left,
714 * largest to the right.
716 if (sector > blk_rq_pos(cfqq->next_rq))
718 else if (sector < blk_rq_pos(cfqq->next_rq))
726 *ret_parent = parent;
732 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
734 struct rb_node **p, *parent;
735 struct cfq_queue *__cfqq;
738 rb_erase(&cfqq->p_node, cfqq->p_root);
742 if (cfq_class_idle(cfqq))
747 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
748 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
749 blk_rq_pos(cfqq->next_rq), &parent, &p);
751 rb_link_node(&cfqq->p_node, parent, p);
752 rb_insert_color(&cfqq->p_node, cfqq->p_root);
758 * Update cfqq's position in the service tree.
760 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
763 * Resorting requires the cfqq to be on the RR list already.
765 if (cfq_cfqq_on_rr(cfqq)) {
766 cfq_service_tree_add(cfqd, cfqq, 0);
767 cfq_prio_tree_add(cfqd, cfqq);
772 * add to busy list of queues for service, trying to be fair in ordering
773 * the pending list according to last request service
775 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
777 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
778 BUG_ON(cfq_cfqq_on_rr(cfqq));
779 cfq_mark_cfqq_on_rr(cfqq);
782 cfq_resort_rr_list(cfqd, cfqq);
786 * Called when the cfqq no longer has requests pending, remove it from
789 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
791 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
792 BUG_ON(!cfq_cfqq_on_rr(cfqq));
793 cfq_clear_cfqq_on_rr(cfqq);
795 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
796 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
797 cfqq->service_tree = NULL;
800 rb_erase(&cfqq->p_node, cfqq->p_root);
804 BUG_ON(!cfqd->busy_queues);
809 * rb tree support functions
811 static void cfq_del_rq_rb(struct request *rq)
813 struct cfq_queue *cfqq = RQ_CFQQ(rq);
814 struct cfq_data *cfqd = cfqq->cfqd;
815 const int sync = rq_is_sync(rq);
817 BUG_ON(!cfqq->queued[sync]);
818 cfqq->queued[sync]--;
820 elv_rb_del(&cfqq->sort_list, rq);
822 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
823 cfq_del_cfqq_rr(cfqd, cfqq);
826 static void cfq_add_rq_rb(struct request *rq)
828 struct cfq_queue *cfqq = RQ_CFQQ(rq);
829 struct cfq_data *cfqd = cfqq->cfqd;
830 struct request *__alias, *prev;
832 cfqq->queued[rq_is_sync(rq)]++;
835 * looks a little odd, but the first insert might return an alias.
836 * if that happens, put the alias on the dispatch list
838 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
839 cfq_dispatch_insert(cfqd->queue, __alias);
841 if (!cfq_cfqq_on_rr(cfqq))
842 cfq_add_cfqq_rr(cfqd, cfqq);
845 * check if this request is a better next-serve candidate
847 prev = cfqq->next_rq;
848 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq);
851 * adjust priority tree position, if ->next_rq changes
853 if (prev != cfqq->next_rq)
854 cfq_prio_tree_add(cfqd, cfqq);
856 BUG_ON(!cfqq->next_rq);
859 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
861 elv_rb_del(&cfqq->sort_list, rq);
862 cfqq->queued[rq_is_sync(rq)]--;
866 static struct request *
867 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
869 struct task_struct *tsk = current;
870 struct cfq_io_context *cic;
871 struct cfq_queue *cfqq;
873 cic = cfq_cic_lookup(cfqd, tsk->io_context);
877 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
879 sector_t sector = bio->bi_sector + bio_sectors(bio);
881 return elv_rb_find(&cfqq->sort_list, sector);
887 static void cfq_activate_request(struct request_queue *q, struct request *rq)
889 struct cfq_data *cfqd = q->elevator->elevator_data;
891 cfqd->rq_in_driver[rq_is_sync(rq)]++;
892 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
895 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
898 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
900 struct cfq_data *cfqd = q->elevator->elevator_data;
901 const int sync = rq_is_sync(rq);
903 WARN_ON(!cfqd->rq_in_driver[sync]);
904 cfqd->rq_in_driver[sync]--;
905 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
909 static void cfq_remove_request(struct request *rq)
911 struct cfq_queue *cfqq = RQ_CFQQ(rq);
913 if (cfqq->next_rq == rq)
914 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
916 list_del_init(&rq->queuelist);
919 cfqq->cfqd->rq_queued--;
920 if (rq_is_meta(rq)) {
921 WARN_ON(!cfqq->meta_pending);
922 cfqq->meta_pending--;
926 static int cfq_merge(struct request_queue *q, struct request **req,
929 struct cfq_data *cfqd = q->elevator->elevator_data;
930 struct request *__rq;
932 __rq = cfq_find_rq_fmerge(cfqd, bio);
933 if (__rq && elv_rq_merge_ok(__rq, bio)) {
935 return ELEVATOR_FRONT_MERGE;
938 return ELEVATOR_NO_MERGE;
941 static void cfq_merged_request(struct request_queue *q, struct request *req,
944 if (type == ELEVATOR_FRONT_MERGE) {
945 struct cfq_queue *cfqq = RQ_CFQQ(req);
947 cfq_reposition_rq_rb(cfqq, req);
952 cfq_merged_requests(struct request_queue *q, struct request *rq,
953 struct request *next)
956 * reposition in fifo if next is older than rq
958 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
959 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
960 list_move(&rq->queuelist, &next->queuelist);
961 rq_set_fifo_time(rq, rq_fifo_time(next));
964 cfq_remove_request(next);
967 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
970 struct cfq_data *cfqd = q->elevator->elevator_data;
971 struct cfq_io_context *cic;
972 struct cfq_queue *cfqq;
975 * Disallow merge of a sync bio into an async request.
977 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
981 * Lookup the cfqq that this bio will be queued with. Allow
982 * merge only if rq is queued there.
984 cic = cfq_cic_lookup(cfqd, current->io_context);
988 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
989 return cfqq == RQ_CFQQ(rq);
992 static void __cfq_set_active_queue(struct cfq_data *cfqd,
993 struct cfq_queue *cfqq)
996 cfq_log_cfqq(cfqd, cfqq, "set_active");
998 cfqq->slice_dispatch = 0;
1000 cfq_clear_cfqq_wait_request(cfqq);
1001 cfq_clear_cfqq_must_dispatch(cfqq);
1002 cfq_clear_cfqq_must_alloc_slice(cfqq);
1003 cfq_clear_cfqq_fifo_expire(cfqq);
1004 cfq_mark_cfqq_slice_new(cfqq);
1006 del_timer(&cfqd->idle_slice_timer);
1009 cfqd->active_queue = cfqq;
1013 * current cfqq expired its slice (or was too idle), select new one
1016 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1019 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1021 if (cfq_cfqq_wait_request(cfqq))
1022 del_timer(&cfqd->idle_slice_timer);
1024 cfq_clear_cfqq_wait_request(cfqq);
1027 * store what was left of this slice, if the queue idled/timed out
1029 if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1030 cfqq->slice_resid = cfqq->slice_end - jiffies;
1031 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1034 cfq_resort_rr_list(cfqd, cfqq);
1036 if (cfqq == cfqd->active_queue)
1037 cfqd->active_queue = NULL;
1039 if (cfqd->active_cic) {
1040 put_io_context(cfqd->active_cic->ioc);
1041 cfqd->active_cic = NULL;
1045 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1047 struct cfq_queue *cfqq = cfqd->active_queue;
1050 __cfq_slice_expired(cfqd, cfqq, timed_out);
1054 * Get next queue for service. Unless we have a queue preemption,
1055 * we'll simply select the first cfqq in the service tree.
1057 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1059 struct cfq_rb_root *service_tree =
1060 service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd);
1062 if (RB_EMPTY_ROOT(&service_tree->rb))
1064 return cfq_rb_first(service_tree);
1068 * Get and set a new active queue for service.
1070 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1071 struct cfq_queue *cfqq)
1074 cfqq = cfq_get_next_queue(cfqd);
1076 if (cfqq && !cfq_cfqq_coop_preempt(cfqq))
1077 cfq_clear_cfqq_coop(cfqq);
1081 cfq_clear_cfqq_coop_preempt(cfqq);
1083 __cfq_set_active_queue(cfqd, cfqq);
1087 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1090 if (blk_rq_pos(rq) >= cfqd->last_position)
1091 return blk_rq_pos(rq) - cfqd->last_position;
1093 return cfqd->last_position - blk_rq_pos(rq);
1096 #define CFQQ_SEEK_THR 8 * 1024
1097 #define CFQQ_SEEKY(cfqq) ((cfqq)->seek_mean > CFQQ_SEEK_THR)
1099 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1102 sector_t sdist = cfqq->seek_mean;
1104 if (!sample_valid(cfqq->seek_samples))
1105 sdist = CFQQ_SEEK_THR;
1107 return cfq_dist_from_last(cfqd, rq) <= sdist;
1110 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1111 struct cfq_queue *cur_cfqq)
1113 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1114 struct rb_node *parent, *node;
1115 struct cfq_queue *__cfqq;
1116 sector_t sector = cfqd->last_position;
1118 if (RB_EMPTY_ROOT(root))
1122 * First, if we find a request starting at the end of the last
1123 * request, choose it.
1125 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1130 * If the exact sector wasn't found, the parent of the NULL leaf
1131 * will contain the closest sector.
1133 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1134 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1137 if (blk_rq_pos(__cfqq->next_rq) < sector)
1138 node = rb_next(&__cfqq->p_node);
1140 node = rb_prev(&__cfqq->p_node);
1144 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1145 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1153 * cur_cfqq - passed in so that we don't decide that the current queue is
1154 * closely cooperating with itself.
1156 * So, basically we're assuming that that cur_cfqq has dispatched at least
1157 * one request, and that cfqd->last_position reflects a position on the disk
1158 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1161 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1162 struct cfq_queue *cur_cfqq)
1164 struct cfq_queue *cfqq;
1166 if (!cfq_cfqq_sync(cur_cfqq))
1168 if (CFQQ_SEEKY(cur_cfqq))
1172 * We should notice if some of the queues are cooperating, eg
1173 * working closely on the same area of the disk. In that case,
1174 * we can group them together and don't waste time idling.
1176 cfqq = cfqq_close(cfqd, cur_cfqq);
1181 * It only makes sense to merge sync queues.
1183 if (!cfq_cfqq_sync(cfqq))
1185 if (CFQQ_SEEKY(cfqq))
1189 * Do not merge queues of different priority classes
1191 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1198 * Determine whether we should enforce idle window for this queue.
1201 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1203 enum wl_prio_t prio = cfqq_prio(cfqq);
1204 struct cfq_rb_root *service_tree = cfqq->service_tree;
1206 /* We never do for idle class queues. */
1207 if (prio == IDLE_WORKLOAD)
1210 /* We do for queues that were marked with idle window flag. */
1211 if (cfq_cfqq_idle_window(cfqq))
1215 * Otherwise, we do only if they are the last ones
1216 * in their service tree.
1219 service_tree = service_tree_for(prio, cfqq_type(cfqq), cfqd);
1221 if (service_tree->count == 0)
1224 return (service_tree->count == 1 && cfq_rb_first(service_tree) == cfqq);
1227 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1229 struct cfq_queue *cfqq = cfqd->active_queue;
1230 struct cfq_io_context *cic;
1234 * SSD device without seek penalty, disable idling. But only do so
1235 * for devices that support queuing, otherwise we still have a problem
1236 * with sync vs async workloads.
1238 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1241 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1242 WARN_ON(cfq_cfqq_slice_new(cfqq));
1245 * idle is disabled, either manually or by past process history
1247 if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1251 * still requests with the driver, don't idle
1253 if (rq_in_driver(cfqd))
1257 * task has exited, don't wait
1259 cic = cfqd->active_cic;
1260 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1264 * If our average think time is larger than the remaining time
1265 * slice, then don't idle. This avoids overrunning the allotted
1268 if (sample_valid(cic->ttime_samples) &&
1269 (cfqq->slice_end - jiffies < cic->ttime_mean))
1272 cfq_mark_cfqq_wait_request(cfqq);
1274 sl = cfqd->cfq_slice_idle;
1275 /* are we servicing noidle tree, and there are more queues?
1276 * non-rotational or NCQ: no idle
1277 * non-NCQ rotational : very small idle, to allow
1278 * fair distribution of slice time for a process doing back-to-back
1281 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
1282 service_tree_for(cfqd->serving_prio, SYNC_NOIDLE_WORKLOAD, cfqd)
1284 if (blk_queue_nonrot(cfqd->queue) || cfqd->hw_tag)
1286 sl = min(sl, msecs_to_jiffies(CFQ_MIN_TT));
1289 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1290 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1294 * Move request from internal lists to the request queue dispatch list.
1296 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1298 struct cfq_data *cfqd = q->elevator->elevator_data;
1299 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1301 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1303 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1304 cfq_remove_request(rq);
1306 elv_dispatch_sort(q, rq);
1308 if (cfq_cfqq_sync(cfqq))
1309 cfqd->sync_flight++;
1313 * return expired entry, or NULL to just start from scratch in rbtree
1315 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1317 struct request *rq = NULL;
1319 if (cfq_cfqq_fifo_expire(cfqq))
1322 cfq_mark_cfqq_fifo_expire(cfqq);
1324 if (list_empty(&cfqq->fifo))
1327 rq = rq_entry_fifo(cfqq->fifo.next);
1328 if (time_before(jiffies, rq_fifo_time(rq)))
1331 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1336 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1338 const int base_rq = cfqd->cfq_slice_async_rq;
1340 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1342 return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1346 * Must be called with the queue_lock held.
1348 static int cfqq_process_refs(struct cfq_queue *cfqq)
1350 int process_refs, io_refs;
1352 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1353 process_refs = atomic_read(&cfqq->ref) - io_refs;
1354 BUG_ON(process_refs < 0);
1355 return process_refs;
1358 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1360 int process_refs, new_process_refs;
1361 struct cfq_queue *__cfqq;
1363 /* Avoid a circular list and skip interim queue merges */
1364 while ((__cfqq = new_cfqq->new_cfqq)) {
1370 process_refs = cfqq_process_refs(cfqq);
1372 * If the process for the cfqq has gone away, there is no
1373 * sense in merging the queues.
1375 if (process_refs == 0)
1379 * Merge in the direction of the lesser amount of work.
1381 new_process_refs = cfqq_process_refs(new_cfqq);
1382 if (new_process_refs >= process_refs) {
1383 cfqq->new_cfqq = new_cfqq;
1384 atomic_add(process_refs, &new_cfqq->ref);
1386 new_cfqq->new_cfqq = cfqq;
1387 atomic_add(new_process_refs, &cfqq->ref);
1391 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd, enum wl_prio_t prio,
1394 struct cfq_queue *queue;
1396 bool key_valid = false;
1397 unsigned long lowest_key = 0;
1398 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
1402 * When priorities switched, we prefer starting
1403 * from SYNC_NOIDLE (first choice), or just SYNC
1406 if (service_tree_for(prio, cur_best, cfqd)->count)
1408 cur_best = SYNC_WORKLOAD;
1409 if (service_tree_for(prio, cur_best, cfqd)->count)
1412 return ASYNC_WORKLOAD;
1415 for (i = 0; i < 3; ++i) {
1416 /* otherwise, select the one with lowest rb_key */
1417 queue = cfq_rb_first(service_tree_for(prio, i, cfqd));
1419 (!key_valid || time_before(queue->rb_key, lowest_key))) {
1420 lowest_key = queue->rb_key;
1429 static void choose_service_tree(struct cfq_data *cfqd)
1431 enum wl_prio_t previous_prio = cfqd->serving_prio;
1436 /* Choose next priority. RT > BE > IDLE */
1437 if (cfq_busy_queues_wl(RT_WORKLOAD, cfqd))
1438 cfqd->serving_prio = RT_WORKLOAD;
1439 else if (cfq_busy_queues_wl(BE_WORKLOAD, cfqd))
1440 cfqd->serving_prio = BE_WORKLOAD;
1442 cfqd->serving_prio = IDLE_WORKLOAD;
1443 cfqd->workload_expires = jiffies + 1;
1448 * For RT and BE, we have to choose also the type
1449 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
1452 prio_changed = (cfqd->serving_prio != previous_prio);
1453 count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
1457 * If priority didn't change, check workload expiration,
1458 * and that we still have other queues ready
1460 if (!prio_changed && count &&
1461 !time_after(jiffies, cfqd->workload_expires))
1464 /* otherwise select new workload type */
1465 cfqd->serving_type =
1466 cfq_choose_wl(cfqd, cfqd->serving_prio, prio_changed);
1467 count = service_tree_for(cfqd->serving_prio, cfqd->serving_type, cfqd)
1471 * the workload slice is computed as a fraction of target latency
1472 * proportional to the number of queues in that workload, over
1473 * all the queues in the same priority class
1475 slice = cfq_target_latency * count /
1476 max_t(unsigned, cfqd->busy_queues_avg[cfqd->serving_prio],
1477 cfq_busy_queues_wl(cfqd->serving_prio, cfqd));
1479 if (cfqd->serving_type == ASYNC_WORKLOAD)
1480 /* async workload slice is scaled down according to
1481 * the sync/async slice ratio. */
1482 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
1484 /* sync workload slice is at least 2 * cfq_slice_idle */
1485 slice = max(slice, 2 * cfqd->cfq_slice_idle);
1487 slice = max_t(unsigned, slice, CFQ_MIN_TT);
1488 cfqd->workload_expires = jiffies + slice;
1492 * Select a queue for service. If we have a current active queue,
1493 * check whether to continue servicing it, or retrieve and set a new one.
1495 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
1497 struct cfq_queue *cfqq, *new_cfqq = NULL;
1499 cfqq = cfqd->active_queue;
1504 * The active queue has run out of time, expire it and select new.
1506 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq))
1510 * The active queue has requests and isn't expired, allow it to
1513 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
1517 * If another queue has a request waiting within our mean seek
1518 * distance, let it run. The expire code will check for close
1519 * cooperators and put the close queue at the front of the service
1520 * tree. If possible, merge the expiring queue with the new cfqq.
1522 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
1524 if (!cfqq->new_cfqq)
1525 cfq_setup_merge(cfqq, new_cfqq);
1530 * No requests pending. If the active queue still has requests in
1531 * flight or is idling for a new request, allow either of these
1532 * conditions to happen (or time out) before selecting a new queue.
1534 if (timer_pending(&cfqd->idle_slice_timer) ||
1535 (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
1541 cfq_slice_expired(cfqd, 0);
1544 * Current queue expired. Check if we have to switch to a new
1548 choose_service_tree(cfqd);
1550 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
1555 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
1559 while (cfqq->next_rq) {
1560 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
1564 BUG_ON(!list_empty(&cfqq->fifo));
1569 * Drain our current requests. Used for barriers and when switching
1570 * io schedulers on-the-fly.
1572 static int cfq_forced_dispatch(struct cfq_data *cfqd)
1574 struct cfq_queue *cfqq;
1577 for (i = 0; i < 2; ++i)
1578 for (j = 0; j < 3; ++j)
1579 while ((cfqq = cfq_rb_first(&cfqd->service_trees[i][j]))
1581 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1583 while ((cfqq = cfq_rb_first(&cfqd->service_tree_idle)) != NULL)
1584 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
1586 cfq_slice_expired(cfqd, 0);
1588 BUG_ON(cfqd->busy_queues);
1590 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
1594 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1596 unsigned int max_dispatch;
1599 * Drain async requests before we start sync IO
1601 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_driver[BLK_RW_ASYNC])
1605 * If this is an async queue and we have sync IO in flight, let it wait
1607 if (cfqd->sync_flight && !cfq_cfqq_sync(cfqq))
1610 max_dispatch = cfqd->cfq_quantum;
1611 if (cfq_class_idle(cfqq))
1615 * Does this cfqq already have too much IO in flight?
1617 if (cfqq->dispatched >= max_dispatch) {
1619 * idle queue must always only have a single IO in flight
1621 if (cfq_class_idle(cfqq))
1625 * We have other queues, don't allow more IO from this one
1627 if (cfqd->busy_queues > 1)
1631 * Sole queue user, allow bigger slice
1637 * Async queues must wait a bit before being allowed dispatch.
1638 * We also ramp up the dispatch depth gradually for async IO,
1639 * based on the last sync IO we serviced
1641 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
1642 unsigned long last_sync = jiffies - cfqd->last_end_sync_rq;
1645 depth = last_sync / cfqd->cfq_slice[1];
1646 if (!depth && !cfqq->dispatched)
1648 if (depth < max_dispatch)
1649 max_dispatch = depth;
1653 * If we're below the current max, allow a dispatch
1655 return cfqq->dispatched < max_dispatch;
1659 * Dispatch a request from cfqq, moving them to the request queue
1662 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1666 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
1668 if (!cfq_may_dispatch(cfqd, cfqq))
1672 * follow expired path, else get first next available
1674 rq = cfq_check_fifo(cfqq);
1679 * insert request into driver dispatch list
1681 cfq_dispatch_insert(cfqd->queue, rq);
1683 if (!cfqd->active_cic) {
1684 struct cfq_io_context *cic = RQ_CIC(rq);
1686 atomic_long_inc(&cic->ioc->refcount);
1687 cfqd->active_cic = cic;
1694 * Find the cfqq that we need to service and move a request from that to the
1697 static int cfq_dispatch_requests(struct request_queue *q, int force)
1699 struct cfq_data *cfqd = q->elevator->elevator_data;
1700 struct cfq_queue *cfqq;
1702 if (!cfqd->busy_queues)
1705 if (unlikely(force))
1706 return cfq_forced_dispatch(cfqd);
1708 cfqq = cfq_select_queue(cfqd);
1713 * Dispatch a request from this cfqq, if it is allowed
1715 if (!cfq_dispatch_request(cfqd, cfqq))
1718 cfqq->slice_dispatch++;
1719 cfq_clear_cfqq_must_dispatch(cfqq);
1722 * expire an async queue immediately if it has used up its slice. idle
1723 * queue always expire after 1 dispatch round.
1725 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
1726 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
1727 cfq_class_idle(cfqq))) {
1728 cfqq->slice_end = jiffies + 1;
1729 cfq_slice_expired(cfqd, 0);
1732 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
1737 * task holds one reference to the queue, dropped when task exits. each rq
1738 * in-flight on this queue also holds a reference, dropped when rq is freed.
1740 * queue lock must be held here.
1742 static void cfq_put_queue(struct cfq_queue *cfqq)
1744 struct cfq_data *cfqd = cfqq->cfqd;
1746 BUG_ON(atomic_read(&cfqq->ref) <= 0);
1748 if (!atomic_dec_and_test(&cfqq->ref))
1751 cfq_log_cfqq(cfqd, cfqq, "put_queue");
1752 BUG_ON(rb_first(&cfqq->sort_list));
1753 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
1754 BUG_ON(cfq_cfqq_on_rr(cfqq));
1756 if (unlikely(cfqd->active_queue == cfqq)) {
1757 __cfq_slice_expired(cfqd, cfqq, 0);
1758 cfq_schedule_dispatch(cfqd);
1761 kmem_cache_free(cfq_pool, cfqq);
1765 * Must always be called with the rcu_read_lock() held
1768 __call_for_each_cic(struct io_context *ioc,
1769 void (*func)(struct io_context *, struct cfq_io_context *))
1771 struct cfq_io_context *cic;
1772 struct hlist_node *n;
1774 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
1779 * Call func for each cic attached to this ioc.
1782 call_for_each_cic(struct io_context *ioc,
1783 void (*func)(struct io_context *, struct cfq_io_context *))
1786 __call_for_each_cic(ioc, func);
1790 static void cfq_cic_free_rcu(struct rcu_head *head)
1792 struct cfq_io_context *cic;
1794 cic = container_of(head, struct cfq_io_context, rcu_head);
1796 kmem_cache_free(cfq_ioc_pool, cic);
1797 elv_ioc_count_dec(cfq_ioc_count);
1801 * CFQ scheduler is exiting, grab exit lock and check
1802 * the pending io context count. If it hits zero,
1803 * complete ioc_gone and set it back to NULL
1805 spin_lock(&ioc_gone_lock);
1806 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
1810 spin_unlock(&ioc_gone_lock);
1814 static void cfq_cic_free(struct cfq_io_context *cic)
1816 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
1819 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
1821 unsigned long flags;
1823 BUG_ON(!cic->dead_key);
1825 spin_lock_irqsave(&ioc->lock, flags);
1826 radix_tree_delete(&ioc->radix_root, cic->dead_key);
1827 hlist_del_rcu(&cic->cic_list);
1828 spin_unlock_irqrestore(&ioc->lock, flags);
1834 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
1835 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
1836 * and ->trim() which is called with the task lock held
1838 static void cfq_free_io_context(struct io_context *ioc)
1841 * ioc->refcount is zero here, or we are called from elv_unregister(),
1842 * so no more cic's are allowed to be linked into this ioc. So it
1843 * should be ok to iterate over the known list, we will see all cic's
1844 * since no new ones are added.
1846 __call_for_each_cic(ioc, cic_free_func);
1849 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1851 struct cfq_queue *__cfqq, *next;
1853 if (unlikely(cfqq == cfqd->active_queue)) {
1854 __cfq_slice_expired(cfqd, cfqq, 0);
1855 cfq_schedule_dispatch(cfqd);
1859 * If this queue was scheduled to merge with another queue, be
1860 * sure to drop the reference taken on that queue (and others in
1861 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
1863 __cfqq = cfqq->new_cfqq;
1865 if (__cfqq == cfqq) {
1866 WARN(1, "cfqq->new_cfqq loop detected\n");
1869 next = __cfqq->new_cfqq;
1870 cfq_put_queue(__cfqq);
1874 cfq_put_queue(cfqq);
1877 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
1878 struct cfq_io_context *cic)
1880 struct io_context *ioc = cic->ioc;
1882 list_del_init(&cic->queue_list);
1885 * Make sure key == NULL is seen for dead queues
1888 cic->dead_key = (unsigned long) cic->key;
1891 if (ioc->ioc_data == cic)
1892 rcu_assign_pointer(ioc->ioc_data, NULL);
1894 if (cic->cfqq[BLK_RW_ASYNC]) {
1895 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
1896 cic->cfqq[BLK_RW_ASYNC] = NULL;
1899 if (cic->cfqq[BLK_RW_SYNC]) {
1900 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
1901 cic->cfqq[BLK_RW_SYNC] = NULL;
1905 static void cfq_exit_single_io_context(struct io_context *ioc,
1906 struct cfq_io_context *cic)
1908 struct cfq_data *cfqd = cic->key;
1911 struct request_queue *q = cfqd->queue;
1912 unsigned long flags;
1914 spin_lock_irqsave(q->queue_lock, flags);
1917 * Ensure we get a fresh copy of the ->key to prevent
1918 * race between exiting task and queue
1920 smp_read_barrier_depends();
1922 __cfq_exit_single_io_context(cfqd, cic);
1924 spin_unlock_irqrestore(q->queue_lock, flags);
1929 * The process that ioc belongs to has exited, we need to clean up
1930 * and put the internal structures we have that belongs to that process.
1932 static void cfq_exit_io_context(struct io_context *ioc)
1934 call_for_each_cic(ioc, cfq_exit_single_io_context);
1937 static struct cfq_io_context *
1938 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
1940 struct cfq_io_context *cic;
1942 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
1945 cic->last_end_request = jiffies;
1946 INIT_LIST_HEAD(&cic->queue_list);
1947 INIT_HLIST_NODE(&cic->cic_list);
1948 cic->dtor = cfq_free_io_context;
1949 cic->exit = cfq_exit_io_context;
1950 elv_ioc_count_inc(cfq_ioc_count);
1956 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
1958 struct task_struct *tsk = current;
1961 if (!cfq_cfqq_prio_changed(cfqq))
1964 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
1965 switch (ioprio_class) {
1967 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
1968 case IOPRIO_CLASS_NONE:
1970 * no prio set, inherit CPU scheduling settings
1972 cfqq->ioprio = task_nice_ioprio(tsk);
1973 cfqq->ioprio_class = task_nice_ioclass(tsk);
1975 case IOPRIO_CLASS_RT:
1976 cfqq->ioprio = task_ioprio(ioc);
1977 cfqq->ioprio_class = IOPRIO_CLASS_RT;
1979 case IOPRIO_CLASS_BE:
1980 cfqq->ioprio = task_ioprio(ioc);
1981 cfqq->ioprio_class = IOPRIO_CLASS_BE;
1983 case IOPRIO_CLASS_IDLE:
1984 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
1986 cfq_clear_cfqq_idle_window(cfqq);
1991 * keep track of original prio settings in case we have to temporarily
1992 * elevate the priority of this queue
1994 cfqq->org_ioprio = cfqq->ioprio;
1995 cfqq->org_ioprio_class = cfqq->ioprio_class;
1996 cfq_clear_cfqq_prio_changed(cfqq);
1999 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2001 struct cfq_data *cfqd = cic->key;
2002 struct cfq_queue *cfqq;
2003 unsigned long flags;
2005 if (unlikely(!cfqd))
2008 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2010 cfqq = cic->cfqq[BLK_RW_ASYNC];
2012 struct cfq_queue *new_cfqq;
2013 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2016 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2017 cfq_put_queue(cfqq);
2021 cfqq = cic->cfqq[BLK_RW_SYNC];
2023 cfq_mark_cfqq_prio_changed(cfqq);
2025 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2028 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2030 call_for_each_cic(ioc, changed_ioprio);
2031 ioc->ioprio_changed = 0;
2034 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2035 pid_t pid, bool is_sync)
2037 RB_CLEAR_NODE(&cfqq->rb_node);
2038 RB_CLEAR_NODE(&cfqq->p_node);
2039 INIT_LIST_HEAD(&cfqq->fifo);
2041 atomic_set(&cfqq->ref, 0);
2044 cfq_mark_cfqq_prio_changed(cfqq);
2047 if (!cfq_class_idle(cfqq))
2048 cfq_mark_cfqq_idle_window(cfqq);
2049 cfq_mark_cfqq_sync(cfqq);
2054 static struct cfq_queue *
2055 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2056 struct io_context *ioc, gfp_t gfp_mask)
2058 struct cfq_queue *cfqq, *new_cfqq = NULL;
2059 struct cfq_io_context *cic;
2062 cic = cfq_cic_lookup(cfqd, ioc);
2063 /* cic always exists here */
2064 cfqq = cic_to_cfqq(cic, is_sync);
2067 * Always try a new alloc if we fell back to the OOM cfqq
2068 * originally, since it should just be a temporary situation.
2070 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2075 } else if (gfp_mask & __GFP_WAIT) {
2076 spin_unlock_irq(cfqd->queue->queue_lock);
2077 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2078 gfp_mask | __GFP_ZERO,
2080 spin_lock_irq(cfqd->queue->queue_lock);
2084 cfqq = kmem_cache_alloc_node(cfq_pool,
2085 gfp_mask | __GFP_ZERO,
2090 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2091 cfq_init_prio_data(cfqq, ioc);
2092 cfq_log_cfqq(cfqd, cfqq, "alloced");
2094 cfqq = &cfqd->oom_cfqq;
2098 kmem_cache_free(cfq_pool, new_cfqq);
2103 static struct cfq_queue **
2104 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2106 switch (ioprio_class) {
2107 case IOPRIO_CLASS_RT:
2108 return &cfqd->async_cfqq[0][ioprio];
2109 case IOPRIO_CLASS_BE:
2110 return &cfqd->async_cfqq[1][ioprio];
2111 case IOPRIO_CLASS_IDLE:
2112 return &cfqd->async_idle_cfqq;
2118 static struct cfq_queue *
2119 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2122 const int ioprio = task_ioprio(ioc);
2123 const int ioprio_class = task_ioprio_class(ioc);
2124 struct cfq_queue **async_cfqq = NULL;
2125 struct cfq_queue *cfqq = NULL;
2128 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2133 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2136 * pin the queue now that it's allocated, scheduler exit will prune it
2138 if (!is_sync && !(*async_cfqq)) {
2139 atomic_inc(&cfqq->ref);
2143 atomic_inc(&cfqq->ref);
2148 * We drop cfq io contexts lazily, so we may find a dead one.
2151 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2152 struct cfq_io_context *cic)
2154 unsigned long flags;
2156 WARN_ON(!list_empty(&cic->queue_list));
2158 spin_lock_irqsave(&ioc->lock, flags);
2160 BUG_ON(ioc->ioc_data == cic);
2162 radix_tree_delete(&ioc->radix_root, (unsigned long) cfqd);
2163 hlist_del_rcu(&cic->cic_list);
2164 spin_unlock_irqrestore(&ioc->lock, flags);
2169 static struct cfq_io_context *
2170 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2172 struct cfq_io_context *cic;
2173 unsigned long flags;
2182 * we maintain a last-hit cache, to avoid browsing over the tree
2184 cic = rcu_dereference(ioc->ioc_data);
2185 if (cic && cic->key == cfqd) {
2191 cic = radix_tree_lookup(&ioc->radix_root, (unsigned long) cfqd);
2195 /* ->key must be copied to avoid race with cfq_exit_queue() */
2198 cfq_drop_dead_cic(cfqd, ioc, cic);
2203 spin_lock_irqsave(&ioc->lock, flags);
2204 rcu_assign_pointer(ioc->ioc_data, cic);
2205 spin_unlock_irqrestore(&ioc->lock, flags);
2213 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2214 * the process specific cfq io context when entered from the block layer.
2215 * Also adds the cic to a per-cfqd list, used when this queue is removed.
2217 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2218 struct cfq_io_context *cic, gfp_t gfp_mask)
2220 unsigned long flags;
2223 ret = radix_tree_preload(gfp_mask);
2228 spin_lock_irqsave(&ioc->lock, flags);
2229 ret = radix_tree_insert(&ioc->radix_root,
2230 (unsigned long) cfqd, cic);
2232 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2233 spin_unlock_irqrestore(&ioc->lock, flags);
2235 radix_tree_preload_end();
2238 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2239 list_add(&cic->queue_list, &cfqd->cic_list);
2240 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2245 printk(KERN_ERR "cfq: cic link failed!\n");
2251 * Setup general io context and cfq io context. There can be several cfq
2252 * io contexts per general io context, if this process is doing io to more
2253 * than one device managed by cfq.
2255 static struct cfq_io_context *
2256 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2258 struct io_context *ioc = NULL;
2259 struct cfq_io_context *cic;
2261 might_sleep_if(gfp_mask & __GFP_WAIT);
2263 ioc = get_io_context(gfp_mask, cfqd->queue->node);
2267 cic = cfq_cic_lookup(cfqd, ioc);
2271 cic = cfq_alloc_io_context(cfqd, gfp_mask);
2275 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
2279 smp_read_barrier_depends();
2280 if (unlikely(ioc->ioprio_changed))
2281 cfq_ioc_set_ioprio(ioc);
2287 put_io_context(ioc);
2292 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
2294 unsigned long elapsed = jiffies - cic->last_end_request;
2295 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
2297 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
2298 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
2299 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
2303 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2309 if (!cfqq->last_request_pos)
2311 else if (cfqq->last_request_pos < blk_rq_pos(rq))
2312 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2314 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2317 * Don't allow the seek distance to get too large from the
2318 * odd fragment, pagein, etc
2320 if (cfqq->seek_samples <= 60) /* second&third seek */
2321 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*1024);
2323 sdist = min(sdist, (cfqq->seek_mean * 4) + 2*1024*64);
2325 cfqq->seek_samples = (7*cfqq->seek_samples + 256) / 8;
2326 cfqq->seek_total = (7*cfqq->seek_total + (u64)256*sdist) / 8;
2327 total = cfqq->seek_total + (cfqq->seek_samples/2);
2328 do_div(total, cfqq->seek_samples);
2329 cfqq->seek_mean = (sector_t)total;
2332 * If this cfqq is shared between multiple processes, check to
2333 * make sure that those processes are still issuing I/Os within
2334 * the mean seek distance. If not, it may be time to break the
2335 * queues apart again.
2337 if (cfq_cfqq_coop(cfqq)) {
2338 if (CFQQ_SEEKY(cfqq) && !cfqq->seeky_start)
2339 cfqq->seeky_start = jiffies;
2340 else if (!CFQQ_SEEKY(cfqq))
2341 cfqq->seeky_start = 0;
2346 * Disable idle window if the process thinks too long or seeks so much that
2350 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2351 struct cfq_io_context *cic)
2353 int old_idle, enable_idle;
2356 * Don't idle for async or idle io prio class
2358 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2361 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
2363 if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
2364 (sample_valid(cfqq->seek_samples) && CFQQ_SEEKY(cfqq)))
2366 else if (sample_valid(cic->ttime_samples)) {
2367 if (cic->ttime_mean > cfqd->cfq_slice_idle)
2373 if (old_idle != enable_idle) {
2374 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
2376 cfq_mark_cfqq_idle_window(cfqq);
2378 cfq_clear_cfqq_idle_window(cfqq);
2383 * Check if new_cfqq should preempt the currently active queue. Return 0 for
2384 * no or if we aren't sure, a 1 will cause a preempt.
2387 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
2390 struct cfq_queue *cfqq;
2392 cfqq = cfqd->active_queue;
2396 if (cfq_slice_used(cfqq))
2399 if (cfq_class_idle(new_cfqq))
2402 if (cfq_class_idle(cfqq))
2405 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD
2406 && new_cfqq->service_tree == cfqq->service_tree)
2410 * if the new request is sync, but the currently running queue is
2411 * not, let the sync request have priority.
2413 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
2417 * So both queues are sync. Let the new request get disk time if
2418 * it's a metadata request and the current queue is doing regular IO.
2420 if (rq_is_meta(rq) && !cfqq->meta_pending)
2424 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
2426 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
2429 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
2433 * if this request is as-good as one we would expect from the
2434 * current cfqq, let it preempt
2436 if (cfq_rq_close(cfqd, cfqq, rq) && (!cfq_cfqq_coop(new_cfqq) ||
2437 cfqd->busy_queues == 1)) {
2439 * Mark new queue coop_preempt, so its coop flag will not be
2440 * cleared when new queue gets scheduled at the very first time
2442 cfq_mark_cfqq_coop_preempt(new_cfqq);
2443 cfq_mark_cfqq_coop(new_cfqq);
2451 * cfqq preempts the active queue. if we allowed preempt with no slice left,
2452 * let it have half of its nominal slice.
2454 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2456 cfq_log_cfqq(cfqd, cfqq, "preempt");
2457 cfq_slice_expired(cfqd, 1);
2460 * Put the new queue at the front of the of the current list,
2461 * so we know that it will be selected next.
2463 BUG_ON(!cfq_cfqq_on_rr(cfqq));
2465 cfq_service_tree_add(cfqd, cfqq, 1);
2467 cfqq->slice_end = 0;
2468 cfq_mark_cfqq_slice_new(cfqq);
2472 * Called when a new fs request (rq) is added (to cfqq). Check if there's
2473 * something we should do about it
2476 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2479 struct cfq_io_context *cic = RQ_CIC(rq);
2483 cfqq->meta_pending++;
2485 cfq_update_io_thinktime(cfqd, cic);
2486 cfq_update_io_seektime(cfqd, cfqq, rq);
2487 cfq_update_idle_window(cfqd, cfqq, cic);
2489 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
2491 if (cfqq == cfqd->active_queue) {
2493 * Remember that we saw a request from this process, but
2494 * don't start queuing just yet. Otherwise we risk seeing lots
2495 * of tiny requests, because we disrupt the normal plugging
2496 * and merging. If the request is already larger than a single
2497 * page, let it rip immediately. For that case we assume that
2498 * merging is already done. Ditto for a busy system that
2499 * has other work pending, don't risk delaying until the
2500 * idle timer unplug to continue working.
2502 if (cfq_cfqq_wait_request(cfqq)) {
2503 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
2504 cfqd->busy_queues > 1) {
2505 del_timer(&cfqd->idle_slice_timer);
2506 __blk_run_queue(cfqd->queue);
2508 cfq_mark_cfqq_must_dispatch(cfqq);
2510 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
2512 * not the active queue - expire current slice if it is
2513 * idle and has expired it's mean thinktime or this new queue
2514 * has some old slice time left and is of higher priority or
2515 * this new queue is RT and the current one is BE
2517 cfq_preempt_queue(cfqd, cfqq);
2518 __blk_run_queue(cfqd->queue);
2522 static void cfq_insert_request(struct request_queue *q, struct request *rq)
2524 struct cfq_data *cfqd = q->elevator->elevator_data;
2525 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2527 cfq_log_cfqq(cfqd, cfqq, "insert_request");
2528 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
2530 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
2531 list_add_tail(&rq->queuelist, &cfqq->fifo);
2534 cfq_rq_enqueued(cfqd, cfqq, rq);
2538 * Update hw_tag based on peak queue depth over 50 samples under
2541 static void cfq_update_hw_tag(struct cfq_data *cfqd)
2543 struct cfq_queue *cfqq = cfqd->active_queue;
2545 if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
2546 cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
2548 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
2549 rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
2553 * If active queue hasn't enough requests and can idle, cfq might not
2554 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
2557 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
2558 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
2559 CFQ_HW_QUEUE_MIN && rq_in_driver(cfqd) < CFQ_HW_QUEUE_MIN)
2562 if (cfqd->hw_tag_samples++ < 50)
2565 if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
2570 cfqd->hw_tag_samples = 0;
2571 cfqd->rq_in_driver_peak = 0;
2574 static void cfq_completed_request(struct request_queue *q, struct request *rq)
2576 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2577 struct cfq_data *cfqd = cfqq->cfqd;
2578 const int sync = rq_is_sync(rq);
2582 cfq_log_cfqq(cfqd, cfqq, "complete");
2584 cfq_update_hw_tag(cfqd);
2586 WARN_ON(!cfqd->rq_in_driver[sync]);
2587 WARN_ON(!cfqq->dispatched);
2588 cfqd->rq_in_driver[sync]--;
2591 if (cfq_cfqq_sync(cfqq))
2592 cfqd->sync_flight--;
2595 RQ_CIC(rq)->last_end_request = now;
2596 cfqd->last_end_sync_rq = now;
2600 * If this is the active queue, check if it needs to be expired,
2601 * or if we want to idle in case it has no pending requests.
2603 if (cfqd->active_queue == cfqq) {
2604 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
2606 if (cfq_cfqq_slice_new(cfqq)) {
2607 cfq_set_prio_slice(cfqd, cfqq);
2608 cfq_clear_cfqq_slice_new(cfqq);
2611 * If there are no requests waiting in this queue, and
2612 * there are other queues ready to issue requests, AND
2613 * those other queues are issuing requests within our
2614 * mean seek distance, give them a chance to run instead
2617 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
2618 cfq_slice_expired(cfqd, 1);
2619 else if (cfqq_empty && !cfq_close_cooperator(cfqd, cfqq) &&
2620 sync && !rq_noidle(rq))
2621 cfq_arm_slice_timer(cfqd);
2624 if (!rq_in_driver(cfqd))
2625 cfq_schedule_dispatch(cfqd);
2629 * we temporarily boost lower priority queues if they are holding fs exclusive
2630 * resources. they are boosted to normal prio (CLASS_BE/4)
2632 static void cfq_prio_boost(struct cfq_queue *cfqq)
2634 if (has_fs_excl()) {
2636 * boost idle prio on transactions that would lock out other
2637 * users of the filesystem
2639 if (cfq_class_idle(cfqq))
2640 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2641 if (cfqq->ioprio > IOPRIO_NORM)
2642 cfqq->ioprio = IOPRIO_NORM;
2645 * unboost the queue (if needed)
2647 cfqq->ioprio_class = cfqq->org_ioprio_class;
2648 cfqq->ioprio = cfqq->org_ioprio;
2652 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
2654 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
2655 cfq_mark_cfqq_must_alloc_slice(cfqq);
2656 return ELV_MQUEUE_MUST;
2659 return ELV_MQUEUE_MAY;
2662 static int cfq_may_queue(struct request_queue *q, int rw)
2664 struct cfq_data *cfqd = q->elevator->elevator_data;
2665 struct task_struct *tsk = current;
2666 struct cfq_io_context *cic;
2667 struct cfq_queue *cfqq;
2670 * don't force setup of a queue from here, as a call to may_queue
2671 * does not necessarily imply that a request actually will be queued.
2672 * so just lookup a possibly existing queue, or return 'may queue'
2675 cic = cfq_cic_lookup(cfqd, tsk->io_context);
2677 return ELV_MQUEUE_MAY;
2679 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
2681 cfq_init_prio_data(cfqq, cic->ioc);
2682 cfq_prio_boost(cfqq);
2684 return __cfq_may_queue(cfqq);
2687 return ELV_MQUEUE_MAY;
2691 * queue lock held here
2693 static void cfq_put_request(struct request *rq)
2695 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2698 const int rw = rq_data_dir(rq);
2700 BUG_ON(!cfqq->allocated[rw]);
2701 cfqq->allocated[rw]--;
2703 put_io_context(RQ_CIC(rq)->ioc);
2705 rq->elevator_private = NULL;
2706 rq->elevator_private2 = NULL;
2708 cfq_put_queue(cfqq);
2712 static struct cfq_queue *
2713 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
2714 struct cfq_queue *cfqq)
2716 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
2717 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
2718 cfq_mark_cfqq_coop(cfqq->new_cfqq);
2719 cfq_put_queue(cfqq);
2720 return cic_to_cfqq(cic, 1);
2723 static int should_split_cfqq(struct cfq_queue *cfqq)
2725 if (cfqq->seeky_start &&
2726 time_after(jiffies, cfqq->seeky_start + CFQQ_COOP_TOUT))
2732 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
2733 * was the last process referring to said cfqq.
2735 static struct cfq_queue *
2736 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
2738 if (cfqq_process_refs(cfqq) == 1) {
2739 cfqq->seeky_start = 0;
2740 cfqq->pid = current->pid;
2741 cfq_clear_cfqq_coop(cfqq);
2745 cic_set_cfqq(cic, NULL, 1);
2746 cfq_put_queue(cfqq);
2750 * Allocate cfq data structures associated with this request.
2753 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
2755 struct cfq_data *cfqd = q->elevator->elevator_data;
2756 struct cfq_io_context *cic;
2757 const int rw = rq_data_dir(rq);
2758 const bool is_sync = rq_is_sync(rq);
2759 struct cfq_queue *cfqq;
2760 unsigned long flags;
2762 might_sleep_if(gfp_mask & __GFP_WAIT);
2764 cic = cfq_get_io_context(cfqd, gfp_mask);
2766 spin_lock_irqsave(q->queue_lock, flags);
2772 cfqq = cic_to_cfqq(cic, is_sync);
2773 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2774 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
2775 cic_set_cfqq(cic, cfqq, is_sync);
2778 * If the queue was seeky for too long, break it apart.
2780 if (cfq_cfqq_coop(cfqq) && should_split_cfqq(cfqq)) {
2781 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
2782 cfqq = split_cfqq(cic, cfqq);
2788 * Check to see if this queue is scheduled to merge with
2789 * another, closely cooperating queue. The merging of
2790 * queues happens here as it must be done in process context.
2791 * The reference on new_cfqq was taken in merge_cfqqs.
2794 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
2797 cfqq->allocated[rw]++;
2798 atomic_inc(&cfqq->ref);
2800 spin_unlock_irqrestore(q->queue_lock, flags);
2802 rq->elevator_private = cic;
2803 rq->elevator_private2 = cfqq;
2808 put_io_context(cic->ioc);
2810 cfq_schedule_dispatch(cfqd);
2811 spin_unlock_irqrestore(q->queue_lock, flags);
2812 cfq_log(cfqd, "set_request fail");
2816 static void cfq_kick_queue(struct work_struct *work)
2818 struct cfq_data *cfqd =
2819 container_of(work, struct cfq_data, unplug_work);
2820 struct request_queue *q = cfqd->queue;
2822 spin_lock_irq(q->queue_lock);
2823 __blk_run_queue(cfqd->queue);
2824 spin_unlock_irq(q->queue_lock);
2828 * Timer running if the active_queue is currently idling inside its time slice
2830 static void cfq_idle_slice_timer(unsigned long data)
2832 struct cfq_data *cfqd = (struct cfq_data *) data;
2833 struct cfq_queue *cfqq;
2834 unsigned long flags;
2837 cfq_log(cfqd, "idle timer fired");
2839 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2841 cfqq = cfqd->active_queue;
2846 * We saw a request before the queue expired, let it through
2848 if (cfq_cfqq_must_dispatch(cfqq))
2854 if (cfq_slice_used(cfqq))
2858 * only expire and reinvoke request handler, if there are
2859 * other queues with pending requests
2861 if (!cfqd->busy_queues)
2865 * not expired and it has a request pending, let it dispatch
2867 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2871 cfq_slice_expired(cfqd, timed_out);
2873 cfq_schedule_dispatch(cfqd);
2875 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2878 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
2880 del_timer_sync(&cfqd->idle_slice_timer);
2881 cancel_work_sync(&cfqd->unplug_work);
2884 static void cfq_put_async_queues(struct cfq_data *cfqd)
2888 for (i = 0; i < IOPRIO_BE_NR; i++) {
2889 if (cfqd->async_cfqq[0][i])
2890 cfq_put_queue(cfqd->async_cfqq[0][i]);
2891 if (cfqd->async_cfqq[1][i])
2892 cfq_put_queue(cfqd->async_cfqq[1][i]);
2895 if (cfqd->async_idle_cfqq)
2896 cfq_put_queue(cfqd->async_idle_cfqq);
2899 static void cfq_exit_queue(struct elevator_queue *e)
2901 struct cfq_data *cfqd = e->elevator_data;
2902 struct request_queue *q = cfqd->queue;
2904 cfq_shutdown_timer_wq(cfqd);
2906 spin_lock_irq(q->queue_lock);
2908 if (cfqd->active_queue)
2909 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
2911 while (!list_empty(&cfqd->cic_list)) {
2912 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
2913 struct cfq_io_context,
2916 __cfq_exit_single_io_context(cfqd, cic);
2919 cfq_put_async_queues(cfqd);
2921 spin_unlock_irq(q->queue_lock);
2923 cfq_shutdown_timer_wq(cfqd);
2928 static void *cfq_init_queue(struct request_queue *q)
2930 struct cfq_data *cfqd;
2933 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
2937 for (i = 0; i < 2; ++i)
2938 for (j = 0; j < 3; ++j)
2939 cfqd->service_trees[i][j] = CFQ_RB_ROOT;
2940 cfqd->service_tree_idle = CFQ_RB_ROOT;
2943 * Not strictly needed (since RB_ROOT just clears the node and we
2944 * zeroed cfqd on alloc), but better be safe in case someone decides
2945 * to add magic to the rb code
2947 for (i = 0; i < CFQ_PRIO_LISTS; i++)
2948 cfqd->prio_trees[i] = RB_ROOT;
2951 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
2952 * Grab a permanent reference to it, so that the normal code flow
2953 * will not attempt to free it.
2955 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
2956 atomic_inc(&cfqd->oom_cfqq.ref);
2958 INIT_LIST_HEAD(&cfqd->cic_list);
2962 init_timer(&cfqd->idle_slice_timer);
2963 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
2964 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
2966 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
2968 cfqd->cfq_quantum = cfq_quantum;
2969 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
2970 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
2971 cfqd->cfq_back_max = cfq_back_max;
2972 cfqd->cfq_back_penalty = cfq_back_penalty;
2973 cfqd->cfq_slice[0] = cfq_slice_async;
2974 cfqd->cfq_slice[1] = cfq_slice_sync;
2975 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
2976 cfqd->cfq_slice_idle = cfq_slice_idle;
2977 cfqd->cfq_latency = 1;
2979 cfqd->last_end_sync_rq = jiffies;
2983 static void cfq_slab_kill(void)
2986 * Caller already ensured that pending RCU callbacks are completed,
2987 * so we should have no busy allocations at this point.
2990 kmem_cache_destroy(cfq_pool);
2992 kmem_cache_destroy(cfq_ioc_pool);
2995 static int __init cfq_slab_setup(void)
2997 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3001 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3012 * sysfs parts below -->
3015 cfq_var_show(unsigned int var, char *page)
3017 return sprintf(page, "%d\n", var);
3021 cfq_var_store(unsigned int *var, const char *page, size_t count)
3023 char *p = (char *) page;
3025 *var = simple_strtoul(p, &p, 10);
3029 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3030 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3032 struct cfq_data *cfqd = e->elevator_data; \
3033 unsigned int __data = __VAR; \
3035 __data = jiffies_to_msecs(__data); \
3036 return cfq_var_show(__data, (page)); \
3038 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3039 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3040 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3041 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3042 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3043 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3044 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3045 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3046 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3047 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3048 #undef SHOW_FUNCTION
3050 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3051 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3053 struct cfq_data *cfqd = e->elevator_data; \
3054 unsigned int __data; \
3055 int ret = cfq_var_store(&__data, (page), count); \
3056 if (__data < (MIN)) \
3058 else if (__data > (MAX)) \
3061 *(__PTR) = msecs_to_jiffies(__data); \
3063 *(__PTR) = __data; \
3066 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3067 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3069 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3071 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3072 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3074 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3075 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3076 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3077 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3079 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3080 #undef STORE_FUNCTION
3082 #define CFQ_ATTR(name) \
3083 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3085 static struct elv_fs_entry cfq_attrs[] = {
3087 CFQ_ATTR(fifo_expire_sync),
3088 CFQ_ATTR(fifo_expire_async),
3089 CFQ_ATTR(back_seek_max),
3090 CFQ_ATTR(back_seek_penalty),
3091 CFQ_ATTR(slice_sync),
3092 CFQ_ATTR(slice_async),
3093 CFQ_ATTR(slice_async_rq),
3094 CFQ_ATTR(slice_idle),
3095 CFQ_ATTR(low_latency),
3099 static struct elevator_type iosched_cfq = {
3101 .elevator_merge_fn = cfq_merge,
3102 .elevator_merged_fn = cfq_merged_request,
3103 .elevator_merge_req_fn = cfq_merged_requests,
3104 .elevator_allow_merge_fn = cfq_allow_merge,
3105 .elevator_dispatch_fn = cfq_dispatch_requests,
3106 .elevator_add_req_fn = cfq_insert_request,
3107 .elevator_activate_req_fn = cfq_activate_request,
3108 .elevator_deactivate_req_fn = cfq_deactivate_request,
3109 .elevator_queue_empty_fn = cfq_queue_empty,
3110 .elevator_completed_req_fn = cfq_completed_request,
3111 .elevator_former_req_fn = elv_rb_former_request,
3112 .elevator_latter_req_fn = elv_rb_latter_request,
3113 .elevator_set_req_fn = cfq_set_request,
3114 .elevator_put_req_fn = cfq_put_request,
3115 .elevator_may_queue_fn = cfq_may_queue,
3116 .elevator_init_fn = cfq_init_queue,
3117 .elevator_exit_fn = cfq_exit_queue,
3118 .trim = cfq_free_io_context,
3120 .elevator_attrs = cfq_attrs,
3121 .elevator_name = "cfq",
3122 .elevator_owner = THIS_MODULE,
3125 static int __init cfq_init(void)
3128 * could be 0 on HZ < 1000 setups
3130 if (!cfq_slice_async)
3131 cfq_slice_async = 1;
3132 if (!cfq_slice_idle)
3135 if (cfq_slab_setup())
3138 elv_register(&iosched_cfq);
3143 static void __exit cfq_exit(void)
3145 DECLARE_COMPLETION_ONSTACK(all_gone);
3146 elv_unregister(&iosched_cfq);
3147 ioc_gone = &all_gone;
3148 /* ioc_gone's update must be visible before reading ioc_count */
3152 * this also protects us from entering cfq_slab_kill() with
3153 * pending RCU callbacks
3155 if (elv_ioc_count_read(cfq_ioc_count))
3156 wait_for_completion(&all_gone);
3160 module_init(cfq_init);
3161 module_exit(cfq_exit);
3163 MODULE_AUTHOR("Jens Axboe");
3164 MODULE_LICENSE("GPL");
3165 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");