gspca: Fix falling back to lower isoc alt settings
[zen-stable.git] / block / cfq-iosched.c
blob3548705b04e482a4405097217011256105c4bdca
1 /*
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>
8 */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
20 * tunables
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
38 * offset from end of service tree
40 #define CFQ_IDLE_DELAY (HZ / 5)
43 * below this threshold, we consider thinktime immediate
45 #define CFQ_MIN_TT (2)
47 #define CFQ_SLICE_SCALE (5)
48 #define CFQ_HW_QUEUE_MIN (5)
49 #define CFQ_SERVICE_SHIFT 12
51 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
56 #define RQ_CIC(rq) \
57 ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elevator_private[2])
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
71 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
75 #define sample_valid(samples) ((samples) > 80)
76 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
84 struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
87 unsigned count;
88 unsigned total_weight;
89 u64 min_vdisktime;
90 struct cfq_ttime ttime;
92 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
93 .ttime = {.last_end_request = jiffies,},}
96 * Per process-grouping structure
98 struct cfq_queue {
99 /* reference count */
100 int ref;
101 /* various state flags, see below */
102 unsigned int flags;
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
118 int queued[2];
119 /* currently allocated requests */
120 int allocated[2];
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
126 unsigned int allocated_slice;
127 unsigned int slice_dispatch;
128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
130 unsigned long slice_end;
131 long slice_resid;
133 /* pending priority requests */
134 int prio_pending;
135 /* number of requests that are on the dispatch list or inside driver */
136 int dispatched;
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
140 unsigned short ioprio_class;
142 pid_t pid;
144 u32 seek_history;
145 sector_t last_request_pos;
147 struct cfq_rb_root *service_tree;
148 struct cfq_queue *new_cfqq;
149 struct cfq_group *cfqg;
150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
155 * First index in the service_trees.
156 * IDLE is handled separately, so it has negative index
158 enum wl_prio_t {
159 BE_WORKLOAD = 0,
160 RT_WORKLOAD = 1,
161 IDLE_WORKLOAD = 2,
162 CFQ_PRIO_NR,
166 * Second index in the service_trees.
168 enum wl_type_t {
169 ASYNC_WORKLOAD = 0,
170 SYNC_NOIDLE_WORKLOAD = 1,
171 SYNC_WORKLOAD = 2
174 /* This is per cgroup per device grouping structure */
175 struct cfq_group {
176 /* group service_tree member */
177 struct rb_node rb_node;
179 /* group service_tree key */
180 u64 vdisktime;
181 unsigned int weight;
182 unsigned int new_weight;
183 bool needs_update;
185 /* number of cfqq currently on this group */
186 int nr_cfqq;
189 * Per group busy queues average. Useful for workload slice calc. We
190 * create the array for each prio class but at run time it is used
191 * only for RT and BE class and slot for IDLE class remains unused.
192 * This is primarily done to avoid confusion and a gcc warning.
194 unsigned int busy_queues_avg[CFQ_PRIO_NR];
196 * rr lists of queues with requests. We maintain service trees for
197 * RT and BE classes. These trees are subdivided in subclasses
198 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199 * class there is no subclassification and all the cfq queues go on
200 * a single tree service_tree_idle.
201 * Counts are embedded in the cfq_rb_root
203 struct cfq_rb_root service_trees[2][3];
204 struct cfq_rb_root service_tree_idle;
206 unsigned long saved_workload_slice;
207 enum wl_type_t saved_workload;
208 enum wl_prio_t saved_serving_prio;
209 struct blkio_group blkg;
210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
211 struct hlist_node cfqd_node;
212 int ref;
213 #endif
214 /* number of requests that are on the dispatch list or inside driver */
215 int dispatched;
216 struct cfq_ttime ttime;
220 * Per block device queue structure
222 struct cfq_data {
223 struct request_queue *queue;
224 /* Root service tree for cfq_groups */
225 struct cfq_rb_root grp_service_tree;
226 struct cfq_group root_group;
229 * The priority currently being served
231 enum wl_prio_t serving_prio;
232 enum wl_type_t serving_type;
233 unsigned long workload_expires;
234 struct cfq_group *serving_group;
237 * Each priority tree is sorted by next_request position. These
238 * trees are used when determining if two or more queues are
239 * interleaving requests (see cfq_close_cooperator).
241 struct rb_root prio_trees[CFQ_PRIO_LISTS];
243 unsigned int busy_queues;
244 unsigned int busy_sync_queues;
246 int rq_in_driver;
247 int rq_in_flight[2];
250 * queue-depth detection
252 int rq_queued;
253 int hw_tag;
255 * hw_tag can be
256 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
257 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
258 * 0 => no NCQ
260 int hw_tag_est_depth;
261 unsigned int hw_tag_samples;
264 * idle window management
266 struct timer_list idle_slice_timer;
267 struct work_struct unplug_work;
269 struct cfq_queue *active_queue;
270 struct cfq_io_context *active_cic;
273 * async queue for each priority case
275 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
276 struct cfq_queue *async_idle_cfqq;
278 sector_t last_position;
281 * tunables, see top of file
283 unsigned int cfq_quantum;
284 unsigned int cfq_fifo_expire[2];
285 unsigned int cfq_back_penalty;
286 unsigned int cfq_back_max;
287 unsigned int cfq_slice[2];
288 unsigned int cfq_slice_async_rq;
289 unsigned int cfq_slice_idle;
290 unsigned int cfq_group_idle;
291 unsigned int cfq_latency;
293 unsigned int cic_index;
294 struct list_head cic_list;
297 * Fallback dummy cfqq for extreme OOM conditions
299 struct cfq_queue oom_cfqq;
301 unsigned long last_delayed_sync;
303 /* List of cfq groups being managed on this device*/
304 struct hlist_head cfqg_list;
306 /* Number of groups which are on blkcg->blkg_list */
307 unsigned int nr_blkcg_linked_grps;
310 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
312 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
313 enum wl_prio_t prio,
314 enum wl_type_t type)
316 if (!cfqg)
317 return NULL;
319 if (prio == IDLE_WORKLOAD)
320 return &cfqg->service_tree_idle;
322 return &cfqg->service_trees[prio][type];
325 enum cfqq_state_flags {
326 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
327 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
328 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
329 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
330 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
331 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
332 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
333 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
334 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
335 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
336 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
337 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
338 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
341 #define CFQ_CFQQ_FNS(name) \
342 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
344 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
346 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
348 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
350 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
352 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
355 CFQ_CFQQ_FNS(on_rr);
356 CFQ_CFQQ_FNS(wait_request);
357 CFQ_CFQQ_FNS(must_dispatch);
358 CFQ_CFQQ_FNS(must_alloc_slice);
359 CFQ_CFQQ_FNS(fifo_expire);
360 CFQ_CFQQ_FNS(idle_window);
361 CFQ_CFQQ_FNS(prio_changed);
362 CFQ_CFQQ_FNS(slice_new);
363 CFQ_CFQQ_FNS(sync);
364 CFQ_CFQQ_FNS(coop);
365 CFQ_CFQQ_FNS(split_coop);
366 CFQ_CFQQ_FNS(deep);
367 CFQ_CFQQ_FNS(wait_busy);
368 #undef CFQ_CFQQ_FNS
370 #ifdef CONFIG_CFQ_GROUP_IOSCHED
371 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
372 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
373 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
374 blkg_path(&(cfqq)->cfqg->blkg), ##args)
376 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
377 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
378 blkg_path(&(cfqg)->blkg), ##args) \
380 #else
381 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
382 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
383 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
384 #endif
385 #define cfq_log(cfqd, fmt, args...) \
386 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
388 /* Traverses through cfq group service trees */
389 #define for_each_cfqg_st(cfqg, i, j, st) \
390 for (i = 0; i <= IDLE_WORKLOAD; i++) \
391 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
392 : &cfqg->service_tree_idle; \
393 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
394 (i == IDLE_WORKLOAD && j == 0); \
395 j++, st = i < IDLE_WORKLOAD ? \
396 &cfqg->service_trees[i][j]: NULL) \
398 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
399 struct cfq_ttime *ttime, bool group_idle)
401 unsigned long slice;
402 if (!sample_valid(ttime->ttime_samples))
403 return false;
404 if (group_idle)
405 slice = cfqd->cfq_group_idle;
406 else
407 slice = cfqd->cfq_slice_idle;
408 return ttime->ttime_mean > slice;
411 static inline bool iops_mode(struct cfq_data *cfqd)
414 * If we are not idling on queues and it is a NCQ drive, parallel
415 * execution of requests is on and measuring time is not possible
416 * in most of the cases until and unless we drive shallower queue
417 * depths and that becomes a performance bottleneck. In such cases
418 * switch to start providing fairness in terms of number of IOs.
420 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
421 return true;
422 else
423 return false;
426 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
428 if (cfq_class_idle(cfqq))
429 return IDLE_WORKLOAD;
430 if (cfq_class_rt(cfqq))
431 return RT_WORKLOAD;
432 return BE_WORKLOAD;
436 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
438 if (!cfq_cfqq_sync(cfqq))
439 return ASYNC_WORKLOAD;
440 if (!cfq_cfqq_idle_window(cfqq))
441 return SYNC_NOIDLE_WORKLOAD;
442 return SYNC_WORKLOAD;
445 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
446 struct cfq_data *cfqd,
447 struct cfq_group *cfqg)
449 if (wl == IDLE_WORKLOAD)
450 return cfqg->service_tree_idle.count;
452 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
453 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
454 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
457 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
458 struct cfq_group *cfqg)
460 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
461 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
464 static void cfq_dispatch_insert(struct request_queue *, struct request *);
465 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
466 struct io_context *, gfp_t);
467 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
468 struct io_context *);
470 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
471 bool is_sync)
473 return cic->cfqq[is_sync];
476 static inline void cic_set_cfqq(struct cfq_io_context *cic,
477 struct cfq_queue *cfqq, bool is_sync)
479 cic->cfqq[is_sync] = cfqq;
482 #define CIC_DEAD_KEY 1ul
483 #define CIC_DEAD_INDEX_SHIFT 1
485 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
487 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
490 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
492 struct cfq_data *cfqd = cic->key;
494 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
495 return NULL;
497 return cfqd;
501 * We regard a request as SYNC, if it's either a read or has the SYNC bit
502 * set (in which case it could also be direct WRITE).
504 static inline bool cfq_bio_sync(struct bio *bio)
506 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
510 * scheduler run of queue, if there are requests pending and no one in the
511 * driver that will restart queueing
513 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
515 if (cfqd->busy_queues) {
516 cfq_log(cfqd, "schedule dispatch");
517 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
522 * Scale schedule slice based on io priority. Use the sync time slice only
523 * if a queue is marked sync and has sync io queued. A sync queue with async
524 * io only, should not get full sync slice length.
526 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
527 unsigned short prio)
529 const int base_slice = cfqd->cfq_slice[sync];
531 WARN_ON(prio >= IOPRIO_BE_NR);
533 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
536 static inline int
537 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
539 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
542 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
544 u64 d = delta << CFQ_SERVICE_SHIFT;
546 d = d * BLKIO_WEIGHT_DEFAULT;
547 do_div(d, cfqg->weight);
548 return d;
551 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
553 s64 delta = (s64)(vdisktime - min_vdisktime);
554 if (delta > 0)
555 min_vdisktime = vdisktime;
557 return min_vdisktime;
560 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
562 s64 delta = (s64)(vdisktime - min_vdisktime);
563 if (delta < 0)
564 min_vdisktime = vdisktime;
566 return min_vdisktime;
569 static void update_min_vdisktime(struct cfq_rb_root *st)
571 struct cfq_group *cfqg;
573 if (st->left) {
574 cfqg = rb_entry_cfqg(st->left);
575 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
576 cfqg->vdisktime);
581 * get averaged number of queues of RT/BE priority.
582 * average is updated, with a formula that gives more weight to higher numbers,
583 * to quickly follows sudden increases and decrease slowly
586 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
587 struct cfq_group *cfqg, bool rt)
589 unsigned min_q, max_q;
590 unsigned mult = cfq_hist_divisor - 1;
591 unsigned round = cfq_hist_divisor / 2;
592 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
594 min_q = min(cfqg->busy_queues_avg[rt], busy);
595 max_q = max(cfqg->busy_queues_avg[rt], busy);
596 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
597 cfq_hist_divisor;
598 return cfqg->busy_queues_avg[rt];
601 static inline unsigned
602 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
604 struct cfq_rb_root *st = &cfqd->grp_service_tree;
606 return cfq_target_latency * cfqg->weight / st->total_weight;
609 static inline unsigned
610 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
612 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
613 if (cfqd->cfq_latency) {
615 * interested queues (we consider only the ones with the same
616 * priority class in the cfq group)
618 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
619 cfq_class_rt(cfqq));
620 unsigned sync_slice = cfqd->cfq_slice[1];
621 unsigned expect_latency = sync_slice * iq;
622 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
624 if (expect_latency > group_slice) {
625 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
626 /* scale low_slice according to IO priority
627 * and sync vs async */
628 unsigned low_slice =
629 min(slice, base_low_slice * slice / sync_slice);
630 /* the adapted slice value is scaled to fit all iqs
631 * into the target latency */
632 slice = max(slice * group_slice / expect_latency,
633 low_slice);
636 return slice;
639 static inline void
640 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
642 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
644 cfqq->slice_start = jiffies;
645 cfqq->slice_end = jiffies + slice;
646 cfqq->allocated_slice = slice;
647 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
651 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
652 * isn't valid until the first request from the dispatch is activated
653 * and the slice time set.
655 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
657 if (cfq_cfqq_slice_new(cfqq))
658 return false;
659 if (time_before(jiffies, cfqq->slice_end))
660 return false;
662 return true;
666 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
667 * We choose the request that is closest to the head right now. Distance
668 * behind the head is penalized and only allowed to a certain extent.
670 static struct request *
671 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
673 sector_t s1, s2, d1 = 0, d2 = 0;
674 unsigned long back_max;
675 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
676 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
677 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
679 if (rq1 == NULL || rq1 == rq2)
680 return rq2;
681 if (rq2 == NULL)
682 return rq1;
684 if (rq_is_sync(rq1) != rq_is_sync(rq2))
685 return rq_is_sync(rq1) ? rq1 : rq2;
687 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
688 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
690 s1 = blk_rq_pos(rq1);
691 s2 = blk_rq_pos(rq2);
694 * by definition, 1KiB is 2 sectors
696 back_max = cfqd->cfq_back_max * 2;
699 * Strict one way elevator _except_ in the case where we allow
700 * short backward seeks which are biased as twice the cost of a
701 * similar forward seek.
703 if (s1 >= last)
704 d1 = s1 - last;
705 else if (s1 + back_max >= last)
706 d1 = (last - s1) * cfqd->cfq_back_penalty;
707 else
708 wrap |= CFQ_RQ1_WRAP;
710 if (s2 >= last)
711 d2 = s2 - last;
712 else if (s2 + back_max >= last)
713 d2 = (last - s2) * cfqd->cfq_back_penalty;
714 else
715 wrap |= CFQ_RQ2_WRAP;
717 /* Found required data */
720 * By doing switch() on the bit mask "wrap" we avoid having to
721 * check two variables for all permutations: --> faster!
723 switch (wrap) {
724 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
725 if (d1 < d2)
726 return rq1;
727 else if (d2 < d1)
728 return rq2;
729 else {
730 if (s1 >= s2)
731 return rq1;
732 else
733 return rq2;
736 case CFQ_RQ2_WRAP:
737 return rq1;
738 case CFQ_RQ1_WRAP:
739 return rq2;
740 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
741 default:
743 * Since both rqs are wrapped,
744 * start with the one that's further behind head
745 * (--> only *one* back seek required),
746 * since back seek takes more time than forward.
748 if (s1 <= s2)
749 return rq1;
750 else
751 return rq2;
756 * The below is leftmost cache rbtree addon
758 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
760 /* Service tree is empty */
761 if (!root->count)
762 return NULL;
764 if (!root->left)
765 root->left = rb_first(&root->rb);
767 if (root->left)
768 return rb_entry(root->left, struct cfq_queue, rb_node);
770 return NULL;
773 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
775 if (!root->left)
776 root->left = rb_first(&root->rb);
778 if (root->left)
779 return rb_entry_cfqg(root->left);
781 return NULL;
784 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
786 rb_erase(n, root);
787 RB_CLEAR_NODE(n);
790 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
792 if (root->left == n)
793 root->left = NULL;
794 rb_erase_init(n, &root->rb);
795 --root->count;
799 * would be nice to take fifo expire time into account as well
801 static struct request *
802 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
803 struct request *last)
805 struct rb_node *rbnext = rb_next(&last->rb_node);
806 struct rb_node *rbprev = rb_prev(&last->rb_node);
807 struct request *next = NULL, *prev = NULL;
809 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
811 if (rbprev)
812 prev = rb_entry_rq(rbprev);
814 if (rbnext)
815 next = rb_entry_rq(rbnext);
816 else {
817 rbnext = rb_first(&cfqq->sort_list);
818 if (rbnext && rbnext != &last->rb_node)
819 next = rb_entry_rq(rbnext);
822 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
825 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
826 struct cfq_queue *cfqq)
829 * just an approximation, should be ok.
831 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
832 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
835 static inline s64
836 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
838 return cfqg->vdisktime - st->min_vdisktime;
841 static void
842 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
844 struct rb_node **node = &st->rb.rb_node;
845 struct rb_node *parent = NULL;
846 struct cfq_group *__cfqg;
847 s64 key = cfqg_key(st, cfqg);
848 int left = 1;
850 while (*node != NULL) {
851 parent = *node;
852 __cfqg = rb_entry_cfqg(parent);
854 if (key < cfqg_key(st, __cfqg))
855 node = &parent->rb_left;
856 else {
857 node = &parent->rb_right;
858 left = 0;
862 if (left)
863 st->left = &cfqg->rb_node;
865 rb_link_node(&cfqg->rb_node, parent, node);
866 rb_insert_color(&cfqg->rb_node, &st->rb);
869 static void
870 cfq_update_group_weight(struct cfq_group *cfqg)
872 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
873 if (cfqg->needs_update) {
874 cfqg->weight = cfqg->new_weight;
875 cfqg->needs_update = false;
879 static void
880 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
882 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
884 cfq_update_group_weight(cfqg);
885 __cfq_group_service_tree_add(st, cfqg);
886 st->total_weight += cfqg->weight;
889 static void
890 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
892 struct cfq_rb_root *st = &cfqd->grp_service_tree;
893 struct cfq_group *__cfqg;
894 struct rb_node *n;
896 cfqg->nr_cfqq++;
897 if (!RB_EMPTY_NODE(&cfqg->rb_node))
898 return;
901 * Currently put the group at the end. Later implement something
902 * so that groups get lesser vtime based on their weights, so that
903 * if group does not loose all if it was not continuously backlogged.
905 n = rb_last(&st->rb);
906 if (n) {
907 __cfqg = rb_entry_cfqg(n);
908 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
909 } else
910 cfqg->vdisktime = st->min_vdisktime;
911 cfq_group_service_tree_add(st, cfqg);
914 static void
915 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
917 st->total_weight -= cfqg->weight;
918 if (!RB_EMPTY_NODE(&cfqg->rb_node))
919 cfq_rb_erase(&cfqg->rb_node, st);
922 static void
923 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
925 struct cfq_rb_root *st = &cfqd->grp_service_tree;
927 BUG_ON(cfqg->nr_cfqq < 1);
928 cfqg->nr_cfqq--;
930 /* If there are other cfq queues under this group, don't delete it */
931 if (cfqg->nr_cfqq)
932 return;
934 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
935 cfq_group_service_tree_del(st, cfqg);
936 cfqg->saved_workload_slice = 0;
937 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
940 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
941 unsigned int *unaccounted_time)
943 unsigned int slice_used;
946 * Queue got expired before even a single request completed or
947 * got expired immediately after first request completion.
949 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
951 * Also charge the seek time incurred to the group, otherwise
952 * if there are mutiple queues in the group, each can dispatch
953 * a single request on seeky media and cause lots of seek time
954 * and group will never know it.
956 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
958 } else {
959 slice_used = jiffies - cfqq->slice_start;
960 if (slice_used > cfqq->allocated_slice) {
961 *unaccounted_time = slice_used - cfqq->allocated_slice;
962 slice_used = cfqq->allocated_slice;
964 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
965 *unaccounted_time += cfqq->slice_start -
966 cfqq->dispatch_start;
969 return slice_used;
972 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
973 struct cfq_queue *cfqq)
975 struct cfq_rb_root *st = &cfqd->grp_service_tree;
976 unsigned int used_sl, charge, unaccounted_sl = 0;
977 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
978 - cfqg->service_tree_idle.count;
980 BUG_ON(nr_sync < 0);
981 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
983 if (iops_mode(cfqd))
984 charge = cfqq->slice_dispatch;
985 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
986 charge = cfqq->allocated_slice;
988 /* Can't update vdisktime while group is on service tree */
989 cfq_group_service_tree_del(st, cfqg);
990 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
991 /* If a new weight was requested, update now, off tree */
992 cfq_group_service_tree_add(st, cfqg);
994 /* This group is being expired. Save the context */
995 if (time_after(cfqd->workload_expires, jiffies)) {
996 cfqg->saved_workload_slice = cfqd->workload_expires
997 - jiffies;
998 cfqg->saved_workload = cfqd->serving_type;
999 cfqg->saved_serving_prio = cfqd->serving_prio;
1000 } else
1001 cfqg->saved_workload_slice = 0;
1003 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1004 st->min_vdisktime);
1005 cfq_log_cfqq(cfqq->cfqd, cfqq,
1006 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1007 used_sl, cfqq->slice_dispatch, charge,
1008 iops_mode(cfqd), cfqq->nr_sectors);
1009 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1010 unaccounted_sl);
1011 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1014 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1015 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1017 if (blkg)
1018 return container_of(blkg, struct cfq_group, blkg);
1019 return NULL;
1022 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1023 unsigned int weight)
1025 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1026 cfqg->new_weight = weight;
1027 cfqg->needs_update = true;
1030 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1031 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1033 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1034 unsigned int major, minor;
1037 * Add group onto cgroup list. It might happen that bdi->dev is
1038 * not initialized yet. Initialize this new group without major
1039 * and minor info and this info will be filled in once a new thread
1040 * comes for IO.
1042 if (bdi->dev) {
1043 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1044 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1045 (void *)cfqd, MKDEV(major, minor));
1046 } else
1047 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1048 (void *)cfqd, 0);
1050 cfqd->nr_blkcg_linked_grps++;
1051 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1053 /* Add group on cfqd list */
1054 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1058 * Should be called from sleepable context. No request queue lock as per
1059 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1060 * from sleepable context.
1062 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1064 struct cfq_group *cfqg = NULL;
1065 int i, j, ret;
1066 struct cfq_rb_root *st;
1068 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1069 if (!cfqg)
1070 return NULL;
1072 for_each_cfqg_st(cfqg, i, j, st)
1073 *st = CFQ_RB_ROOT;
1074 RB_CLEAR_NODE(&cfqg->rb_node);
1076 cfqg->ttime.last_end_request = jiffies;
1079 * Take the initial reference that will be released on destroy
1080 * This can be thought of a joint reference by cgroup and
1081 * elevator which will be dropped by either elevator exit
1082 * or cgroup deletion path depending on who is exiting first.
1084 cfqg->ref = 1;
1086 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1087 if (ret) {
1088 kfree(cfqg);
1089 return NULL;
1092 return cfqg;
1095 static struct cfq_group *
1096 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1098 struct cfq_group *cfqg = NULL;
1099 void *key = cfqd;
1100 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1101 unsigned int major, minor;
1104 * This is the common case when there are no blkio cgroups.
1105 * Avoid lookup in this case
1107 if (blkcg == &blkio_root_cgroup)
1108 cfqg = &cfqd->root_group;
1109 else
1110 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1112 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1113 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1114 cfqg->blkg.dev = MKDEV(major, minor);
1117 return cfqg;
1121 * Search for the cfq group current task belongs to. request_queue lock must
1122 * be held.
1124 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1126 struct blkio_cgroup *blkcg;
1127 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1128 struct request_queue *q = cfqd->queue;
1130 rcu_read_lock();
1131 blkcg = task_blkio_cgroup(current);
1132 cfqg = cfq_find_cfqg(cfqd, blkcg);
1133 if (cfqg) {
1134 rcu_read_unlock();
1135 return cfqg;
1139 * Need to allocate a group. Allocation of group also needs allocation
1140 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1141 * we need to drop rcu lock and queue_lock before we call alloc.
1143 * Not taking any queue reference here and assuming that queue is
1144 * around by the time we return. CFQ queue allocation code does
1145 * the same. It might be racy though.
1148 rcu_read_unlock();
1149 spin_unlock_irq(q->queue_lock);
1151 cfqg = cfq_alloc_cfqg(cfqd);
1153 spin_lock_irq(q->queue_lock);
1155 rcu_read_lock();
1156 blkcg = task_blkio_cgroup(current);
1159 * If some other thread already allocated the group while we were
1160 * not holding queue lock, free up the group
1162 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1164 if (__cfqg) {
1165 kfree(cfqg);
1166 rcu_read_unlock();
1167 return __cfqg;
1170 if (!cfqg)
1171 cfqg = &cfqd->root_group;
1173 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1174 rcu_read_unlock();
1175 return cfqg;
1178 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1180 cfqg->ref++;
1181 return cfqg;
1184 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1186 /* Currently, all async queues are mapped to root group */
1187 if (!cfq_cfqq_sync(cfqq))
1188 cfqg = &cfqq->cfqd->root_group;
1190 cfqq->cfqg = cfqg;
1191 /* cfqq reference on cfqg */
1192 cfqq->cfqg->ref++;
1195 static void cfq_put_cfqg(struct cfq_group *cfqg)
1197 struct cfq_rb_root *st;
1198 int i, j;
1200 BUG_ON(cfqg->ref <= 0);
1201 cfqg->ref--;
1202 if (cfqg->ref)
1203 return;
1204 for_each_cfqg_st(cfqg, i, j, st)
1205 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1206 free_percpu(cfqg->blkg.stats_cpu);
1207 kfree(cfqg);
1210 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1212 /* Something wrong if we are trying to remove same group twice */
1213 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1215 hlist_del_init(&cfqg->cfqd_node);
1217 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1218 cfqd->nr_blkcg_linked_grps--;
1221 * Put the reference taken at the time of creation so that when all
1222 * queues are gone, group can be destroyed.
1224 cfq_put_cfqg(cfqg);
1227 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1229 struct hlist_node *pos, *n;
1230 struct cfq_group *cfqg;
1232 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1234 * If cgroup removal path got to blk_group first and removed
1235 * it from cgroup list, then it will take care of destroying
1236 * cfqg also.
1238 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1239 cfq_destroy_cfqg(cfqd, cfqg);
1244 * Blk cgroup controller notification saying that blkio_group object is being
1245 * delinked as associated cgroup object is going away. That also means that
1246 * no new IO will come in this group. So get rid of this group as soon as
1247 * any pending IO in the group is finished.
1249 * This function is called under rcu_read_lock(). key is the rcu protected
1250 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1251 * read lock.
1253 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1254 * it should not be NULL as even if elevator was exiting, cgroup deltion
1255 * path got to it first.
1257 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1259 unsigned long flags;
1260 struct cfq_data *cfqd = key;
1262 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1263 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1264 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1267 #else /* GROUP_IOSCHED */
1268 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1270 return &cfqd->root_group;
1273 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1275 return cfqg;
1278 static inline void
1279 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1280 cfqq->cfqg = cfqg;
1283 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1284 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1286 #endif /* GROUP_IOSCHED */
1289 * The cfqd->service_trees holds all pending cfq_queue's that have
1290 * requests waiting to be processed. It is sorted in the order that
1291 * we will service the queues.
1293 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1294 bool add_front)
1296 struct rb_node **p, *parent;
1297 struct cfq_queue *__cfqq;
1298 unsigned long rb_key;
1299 struct cfq_rb_root *service_tree;
1300 int left;
1301 int new_cfqq = 1;
1303 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1304 cfqq_type(cfqq));
1305 if (cfq_class_idle(cfqq)) {
1306 rb_key = CFQ_IDLE_DELAY;
1307 parent = rb_last(&service_tree->rb);
1308 if (parent && parent != &cfqq->rb_node) {
1309 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1310 rb_key += __cfqq->rb_key;
1311 } else
1312 rb_key += jiffies;
1313 } else if (!add_front) {
1315 * Get our rb key offset. Subtract any residual slice
1316 * value carried from last service. A negative resid
1317 * count indicates slice overrun, and this should position
1318 * the next service time further away in the tree.
1320 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1321 rb_key -= cfqq->slice_resid;
1322 cfqq->slice_resid = 0;
1323 } else {
1324 rb_key = -HZ;
1325 __cfqq = cfq_rb_first(service_tree);
1326 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1329 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1330 new_cfqq = 0;
1332 * same position, nothing more to do
1334 if (rb_key == cfqq->rb_key &&
1335 cfqq->service_tree == service_tree)
1336 return;
1338 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1339 cfqq->service_tree = NULL;
1342 left = 1;
1343 parent = NULL;
1344 cfqq->service_tree = service_tree;
1345 p = &service_tree->rb.rb_node;
1346 while (*p) {
1347 struct rb_node **n;
1349 parent = *p;
1350 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1353 * sort by key, that represents service time.
1355 if (time_before(rb_key, __cfqq->rb_key))
1356 n = &(*p)->rb_left;
1357 else {
1358 n = &(*p)->rb_right;
1359 left = 0;
1362 p = n;
1365 if (left)
1366 service_tree->left = &cfqq->rb_node;
1368 cfqq->rb_key = rb_key;
1369 rb_link_node(&cfqq->rb_node, parent, p);
1370 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1371 service_tree->count++;
1372 if (add_front || !new_cfqq)
1373 return;
1374 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1377 static struct cfq_queue *
1378 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1379 sector_t sector, struct rb_node **ret_parent,
1380 struct rb_node ***rb_link)
1382 struct rb_node **p, *parent;
1383 struct cfq_queue *cfqq = NULL;
1385 parent = NULL;
1386 p = &root->rb_node;
1387 while (*p) {
1388 struct rb_node **n;
1390 parent = *p;
1391 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1394 * Sort strictly based on sector. Smallest to the left,
1395 * largest to the right.
1397 if (sector > blk_rq_pos(cfqq->next_rq))
1398 n = &(*p)->rb_right;
1399 else if (sector < blk_rq_pos(cfqq->next_rq))
1400 n = &(*p)->rb_left;
1401 else
1402 break;
1403 p = n;
1404 cfqq = NULL;
1407 *ret_parent = parent;
1408 if (rb_link)
1409 *rb_link = p;
1410 return cfqq;
1413 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1415 struct rb_node **p, *parent;
1416 struct cfq_queue *__cfqq;
1418 if (cfqq->p_root) {
1419 rb_erase(&cfqq->p_node, cfqq->p_root);
1420 cfqq->p_root = NULL;
1423 if (cfq_class_idle(cfqq))
1424 return;
1425 if (!cfqq->next_rq)
1426 return;
1428 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1429 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1430 blk_rq_pos(cfqq->next_rq), &parent, &p);
1431 if (!__cfqq) {
1432 rb_link_node(&cfqq->p_node, parent, p);
1433 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1434 } else
1435 cfqq->p_root = NULL;
1439 * Update cfqq's position in the service tree.
1441 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1444 * Resorting requires the cfqq to be on the RR list already.
1446 if (cfq_cfqq_on_rr(cfqq)) {
1447 cfq_service_tree_add(cfqd, cfqq, 0);
1448 cfq_prio_tree_add(cfqd, cfqq);
1453 * add to busy list of queues for service, trying to be fair in ordering
1454 * the pending list according to last request service
1456 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1458 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1459 BUG_ON(cfq_cfqq_on_rr(cfqq));
1460 cfq_mark_cfqq_on_rr(cfqq);
1461 cfqd->busy_queues++;
1462 if (cfq_cfqq_sync(cfqq))
1463 cfqd->busy_sync_queues++;
1465 cfq_resort_rr_list(cfqd, cfqq);
1469 * Called when the cfqq no longer has requests pending, remove it from
1470 * the service tree.
1472 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1474 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1475 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1476 cfq_clear_cfqq_on_rr(cfqq);
1478 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1479 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1480 cfqq->service_tree = NULL;
1482 if (cfqq->p_root) {
1483 rb_erase(&cfqq->p_node, cfqq->p_root);
1484 cfqq->p_root = NULL;
1487 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1488 BUG_ON(!cfqd->busy_queues);
1489 cfqd->busy_queues--;
1490 if (cfq_cfqq_sync(cfqq))
1491 cfqd->busy_sync_queues--;
1495 * rb tree support functions
1497 static void cfq_del_rq_rb(struct request *rq)
1499 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1500 const int sync = rq_is_sync(rq);
1502 BUG_ON(!cfqq->queued[sync]);
1503 cfqq->queued[sync]--;
1505 elv_rb_del(&cfqq->sort_list, rq);
1507 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1509 * Queue will be deleted from service tree when we actually
1510 * expire it later. Right now just remove it from prio tree
1511 * as it is empty.
1513 if (cfqq->p_root) {
1514 rb_erase(&cfqq->p_node, cfqq->p_root);
1515 cfqq->p_root = NULL;
1520 static void cfq_add_rq_rb(struct request *rq)
1522 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1523 struct cfq_data *cfqd = cfqq->cfqd;
1524 struct request *prev;
1526 cfqq->queued[rq_is_sync(rq)]++;
1528 elv_rb_add(&cfqq->sort_list, rq);
1530 if (!cfq_cfqq_on_rr(cfqq))
1531 cfq_add_cfqq_rr(cfqd, cfqq);
1534 * check if this request is a better next-serve candidate
1536 prev = cfqq->next_rq;
1537 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1540 * adjust priority tree position, if ->next_rq changes
1542 if (prev != cfqq->next_rq)
1543 cfq_prio_tree_add(cfqd, cfqq);
1545 BUG_ON(!cfqq->next_rq);
1548 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1550 elv_rb_del(&cfqq->sort_list, rq);
1551 cfqq->queued[rq_is_sync(rq)]--;
1552 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1553 rq_data_dir(rq), rq_is_sync(rq));
1554 cfq_add_rq_rb(rq);
1555 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1556 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1557 rq_is_sync(rq));
1560 static struct request *
1561 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1563 struct task_struct *tsk = current;
1564 struct cfq_io_context *cic;
1565 struct cfq_queue *cfqq;
1567 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1568 if (!cic)
1569 return NULL;
1571 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1572 if (cfqq) {
1573 sector_t sector = bio->bi_sector + bio_sectors(bio);
1575 return elv_rb_find(&cfqq->sort_list, sector);
1578 return NULL;
1581 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1583 struct cfq_data *cfqd = q->elevator->elevator_data;
1585 cfqd->rq_in_driver++;
1586 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1587 cfqd->rq_in_driver);
1589 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1592 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1594 struct cfq_data *cfqd = q->elevator->elevator_data;
1596 WARN_ON(!cfqd->rq_in_driver);
1597 cfqd->rq_in_driver--;
1598 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1599 cfqd->rq_in_driver);
1602 static void cfq_remove_request(struct request *rq)
1604 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1606 if (cfqq->next_rq == rq)
1607 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1609 list_del_init(&rq->queuelist);
1610 cfq_del_rq_rb(rq);
1612 cfqq->cfqd->rq_queued--;
1613 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1614 rq_data_dir(rq), rq_is_sync(rq));
1615 if (rq->cmd_flags & REQ_PRIO) {
1616 WARN_ON(!cfqq->prio_pending);
1617 cfqq->prio_pending--;
1621 static int cfq_merge(struct request_queue *q, struct request **req,
1622 struct bio *bio)
1624 struct cfq_data *cfqd = q->elevator->elevator_data;
1625 struct request *__rq;
1627 __rq = cfq_find_rq_fmerge(cfqd, bio);
1628 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1629 *req = __rq;
1630 return ELEVATOR_FRONT_MERGE;
1633 return ELEVATOR_NO_MERGE;
1636 static void cfq_merged_request(struct request_queue *q, struct request *req,
1637 int type)
1639 if (type == ELEVATOR_FRONT_MERGE) {
1640 struct cfq_queue *cfqq = RQ_CFQQ(req);
1642 cfq_reposition_rq_rb(cfqq, req);
1646 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1647 struct bio *bio)
1649 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1650 bio_data_dir(bio), cfq_bio_sync(bio));
1653 static void
1654 cfq_merged_requests(struct request_queue *q, struct request *rq,
1655 struct request *next)
1657 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1658 struct cfq_data *cfqd = q->elevator->elevator_data;
1661 * reposition in fifo if next is older than rq
1663 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1664 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1665 list_move(&rq->queuelist, &next->queuelist);
1666 rq_set_fifo_time(rq, rq_fifo_time(next));
1669 if (cfqq->next_rq == next)
1670 cfqq->next_rq = rq;
1671 cfq_remove_request(next);
1672 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1673 rq_data_dir(next), rq_is_sync(next));
1675 cfqq = RQ_CFQQ(next);
1677 * all requests of this queue are merged to other queues, delete it
1678 * from the service tree. If it's the active_queue,
1679 * cfq_dispatch_requests() will choose to expire it or do idle
1681 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1682 cfqq != cfqd->active_queue)
1683 cfq_del_cfqq_rr(cfqd, cfqq);
1686 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1687 struct bio *bio)
1689 struct cfq_data *cfqd = q->elevator->elevator_data;
1690 struct cfq_io_context *cic;
1691 struct cfq_queue *cfqq;
1694 * Disallow merge of a sync bio into an async request.
1696 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1697 return false;
1700 * Lookup the cfqq that this bio will be queued with. Allow
1701 * merge only if rq is queued there.
1703 cic = cfq_cic_lookup(cfqd, current->io_context);
1704 if (!cic)
1705 return false;
1707 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1708 return cfqq == RQ_CFQQ(rq);
1711 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1713 del_timer(&cfqd->idle_slice_timer);
1714 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1717 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1718 struct cfq_queue *cfqq)
1720 if (cfqq) {
1721 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1722 cfqd->serving_prio, cfqd->serving_type);
1723 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1724 cfqq->slice_start = 0;
1725 cfqq->dispatch_start = jiffies;
1726 cfqq->allocated_slice = 0;
1727 cfqq->slice_end = 0;
1728 cfqq->slice_dispatch = 0;
1729 cfqq->nr_sectors = 0;
1731 cfq_clear_cfqq_wait_request(cfqq);
1732 cfq_clear_cfqq_must_dispatch(cfqq);
1733 cfq_clear_cfqq_must_alloc_slice(cfqq);
1734 cfq_clear_cfqq_fifo_expire(cfqq);
1735 cfq_mark_cfqq_slice_new(cfqq);
1737 cfq_del_timer(cfqd, cfqq);
1740 cfqd->active_queue = cfqq;
1744 * current cfqq expired its slice (or was too idle), select new one
1746 static void
1747 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1748 bool timed_out)
1750 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1752 if (cfq_cfqq_wait_request(cfqq))
1753 cfq_del_timer(cfqd, cfqq);
1755 cfq_clear_cfqq_wait_request(cfqq);
1756 cfq_clear_cfqq_wait_busy(cfqq);
1759 * If this cfqq is shared between multiple processes, check to
1760 * make sure that those processes are still issuing I/Os within
1761 * the mean seek distance. If not, it may be time to break the
1762 * queues apart again.
1764 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1765 cfq_mark_cfqq_split_coop(cfqq);
1768 * store what was left of this slice, if the queue idled/timed out
1770 if (timed_out) {
1771 if (cfq_cfqq_slice_new(cfqq))
1772 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1773 else
1774 cfqq->slice_resid = cfqq->slice_end - jiffies;
1775 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1778 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1780 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1781 cfq_del_cfqq_rr(cfqd, cfqq);
1783 cfq_resort_rr_list(cfqd, cfqq);
1785 if (cfqq == cfqd->active_queue)
1786 cfqd->active_queue = NULL;
1788 if (cfqd->active_cic) {
1789 put_io_context(cfqd->active_cic->ioc);
1790 cfqd->active_cic = NULL;
1794 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1796 struct cfq_queue *cfqq = cfqd->active_queue;
1798 if (cfqq)
1799 __cfq_slice_expired(cfqd, cfqq, timed_out);
1803 * Get next queue for service. Unless we have a queue preemption,
1804 * we'll simply select the first cfqq in the service tree.
1806 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1808 struct cfq_rb_root *service_tree =
1809 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1810 cfqd->serving_type);
1812 if (!cfqd->rq_queued)
1813 return NULL;
1815 /* There is nothing to dispatch */
1816 if (!service_tree)
1817 return NULL;
1818 if (RB_EMPTY_ROOT(&service_tree->rb))
1819 return NULL;
1820 return cfq_rb_first(service_tree);
1823 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1825 struct cfq_group *cfqg;
1826 struct cfq_queue *cfqq;
1827 int i, j;
1828 struct cfq_rb_root *st;
1830 if (!cfqd->rq_queued)
1831 return NULL;
1833 cfqg = cfq_get_next_cfqg(cfqd);
1834 if (!cfqg)
1835 return NULL;
1837 for_each_cfqg_st(cfqg, i, j, st)
1838 if ((cfqq = cfq_rb_first(st)) != NULL)
1839 return cfqq;
1840 return NULL;
1844 * Get and set a new active queue for service.
1846 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1847 struct cfq_queue *cfqq)
1849 if (!cfqq)
1850 cfqq = cfq_get_next_queue(cfqd);
1852 __cfq_set_active_queue(cfqd, cfqq);
1853 return cfqq;
1856 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1857 struct request *rq)
1859 if (blk_rq_pos(rq) >= cfqd->last_position)
1860 return blk_rq_pos(rq) - cfqd->last_position;
1861 else
1862 return cfqd->last_position - blk_rq_pos(rq);
1865 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1866 struct request *rq)
1868 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1871 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1872 struct cfq_queue *cur_cfqq)
1874 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1875 struct rb_node *parent, *node;
1876 struct cfq_queue *__cfqq;
1877 sector_t sector = cfqd->last_position;
1879 if (RB_EMPTY_ROOT(root))
1880 return NULL;
1883 * First, if we find a request starting at the end of the last
1884 * request, choose it.
1886 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1887 if (__cfqq)
1888 return __cfqq;
1891 * If the exact sector wasn't found, the parent of the NULL leaf
1892 * will contain the closest sector.
1894 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1895 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1896 return __cfqq;
1898 if (blk_rq_pos(__cfqq->next_rq) < sector)
1899 node = rb_next(&__cfqq->p_node);
1900 else
1901 node = rb_prev(&__cfqq->p_node);
1902 if (!node)
1903 return NULL;
1905 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1906 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1907 return __cfqq;
1909 return NULL;
1913 * cfqd - obvious
1914 * cur_cfqq - passed in so that we don't decide that the current queue is
1915 * closely cooperating with itself.
1917 * So, basically we're assuming that that cur_cfqq has dispatched at least
1918 * one request, and that cfqd->last_position reflects a position on the disk
1919 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1920 * assumption.
1922 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1923 struct cfq_queue *cur_cfqq)
1925 struct cfq_queue *cfqq;
1927 if (cfq_class_idle(cur_cfqq))
1928 return NULL;
1929 if (!cfq_cfqq_sync(cur_cfqq))
1930 return NULL;
1931 if (CFQQ_SEEKY(cur_cfqq))
1932 return NULL;
1935 * Don't search priority tree if it's the only queue in the group.
1937 if (cur_cfqq->cfqg->nr_cfqq == 1)
1938 return NULL;
1941 * We should notice if some of the queues are cooperating, eg
1942 * working closely on the same area of the disk. In that case,
1943 * we can group them together and don't waste time idling.
1945 cfqq = cfqq_close(cfqd, cur_cfqq);
1946 if (!cfqq)
1947 return NULL;
1949 /* If new queue belongs to different cfq_group, don't choose it */
1950 if (cur_cfqq->cfqg != cfqq->cfqg)
1951 return NULL;
1954 * It only makes sense to merge sync queues.
1956 if (!cfq_cfqq_sync(cfqq))
1957 return NULL;
1958 if (CFQQ_SEEKY(cfqq))
1959 return NULL;
1962 * Do not merge queues of different priority classes
1964 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1965 return NULL;
1967 return cfqq;
1971 * Determine whether we should enforce idle window for this queue.
1974 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1976 enum wl_prio_t prio = cfqq_prio(cfqq);
1977 struct cfq_rb_root *service_tree = cfqq->service_tree;
1979 BUG_ON(!service_tree);
1980 BUG_ON(!service_tree->count);
1982 if (!cfqd->cfq_slice_idle)
1983 return false;
1985 /* We never do for idle class queues. */
1986 if (prio == IDLE_WORKLOAD)
1987 return false;
1989 /* We do for queues that were marked with idle window flag. */
1990 if (cfq_cfqq_idle_window(cfqq) &&
1991 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1992 return true;
1995 * Otherwise, we do only if they are the last ones
1996 * in their service tree.
1998 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1999 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2000 return true;
2001 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2002 service_tree->count);
2003 return false;
2006 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2008 struct cfq_queue *cfqq = cfqd->active_queue;
2009 struct cfq_io_context *cic;
2010 unsigned long sl, group_idle = 0;
2013 * SSD device without seek penalty, disable idling. But only do so
2014 * for devices that support queuing, otherwise we still have a problem
2015 * with sync vs async workloads.
2017 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2018 return;
2020 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2021 WARN_ON(cfq_cfqq_slice_new(cfqq));
2024 * idle is disabled, either manually or by past process history
2026 if (!cfq_should_idle(cfqd, cfqq)) {
2027 /* no queue idling. Check for group idling */
2028 if (cfqd->cfq_group_idle)
2029 group_idle = cfqd->cfq_group_idle;
2030 else
2031 return;
2035 * still active requests from this queue, don't idle
2037 if (cfqq->dispatched)
2038 return;
2041 * task has exited, don't wait
2043 cic = cfqd->active_cic;
2044 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2045 return;
2048 * If our average think time is larger than the remaining time
2049 * slice, then don't idle. This avoids overrunning the allotted
2050 * time slice.
2052 if (sample_valid(cic->ttime.ttime_samples) &&
2053 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2054 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2055 cic->ttime.ttime_mean);
2056 return;
2059 /* There are other queues in the group, don't do group idle */
2060 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2061 return;
2063 cfq_mark_cfqq_wait_request(cfqq);
2065 if (group_idle)
2066 sl = cfqd->cfq_group_idle;
2067 else
2068 sl = cfqd->cfq_slice_idle;
2070 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2071 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2072 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2073 group_idle ? 1 : 0);
2077 * Move request from internal lists to the request queue dispatch list.
2079 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2081 struct cfq_data *cfqd = q->elevator->elevator_data;
2082 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2084 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2086 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2087 cfq_remove_request(rq);
2088 cfqq->dispatched++;
2089 (RQ_CFQG(rq))->dispatched++;
2090 elv_dispatch_sort(q, rq);
2092 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2093 cfqq->nr_sectors += blk_rq_sectors(rq);
2094 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2095 rq_data_dir(rq), rq_is_sync(rq));
2099 * return expired entry, or NULL to just start from scratch in rbtree
2101 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2103 struct request *rq = NULL;
2105 if (cfq_cfqq_fifo_expire(cfqq))
2106 return NULL;
2108 cfq_mark_cfqq_fifo_expire(cfqq);
2110 if (list_empty(&cfqq->fifo))
2111 return NULL;
2113 rq = rq_entry_fifo(cfqq->fifo.next);
2114 if (time_before(jiffies, rq_fifo_time(rq)))
2115 rq = NULL;
2117 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2118 return rq;
2121 static inline int
2122 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2124 const int base_rq = cfqd->cfq_slice_async_rq;
2126 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2128 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2132 * Must be called with the queue_lock held.
2134 static int cfqq_process_refs(struct cfq_queue *cfqq)
2136 int process_refs, io_refs;
2138 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2139 process_refs = cfqq->ref - io_refs;
2140 BUG_ON(process_refs < 0);
2141 return process_refs;
2144 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2146 int process_refs, new_process_refs;
2147 struct cfq_queue *__cfqq;
2150 * If there are no process references on the new_cfqq, then it is
2151 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2152 * chain may have dropped their last reference (not just their
2153 * last process reference).
2155 if (!cfqq_process_refs(new_cfqq))
2156 return;
2158 /* Avoid a circular list and skip interim queue merges */
2159 while ((__cfqq = new_cfqq->new_cfqq)) {
2160 if (__cfqq == cfqq)
2161 return;
2162 new_cfqq = __cfqq;
2165 process_refs = cfqq_process_refs(cfqq);
2166 new_process_refs = cfqq_process_refs(new_cfqq);
2168 * If the process for the cfqq has gone away, there is no
2169 * sense in merging the queues.
2171 if (process_refs == 0 || new_process_refs == 0)
2172 return;
2175 * Merge in the direction of the lesser amount of work.
2177 if (new_process_refs >= process_refs) {
2178 cfqq->new_cfqq = new_cfqq;
2179 new_cfqq->ref += process_refs;
2180 } else {
2181 new_cfqq->new_cfqq = cfqq;
2182 cfqq->ref += new_process_refs;
2186 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2187 struct cfq_group *cfqg, enum wl_prio_t prio)
2189 struct cfq_queue *queue;
2190 int i;
2191 bool key_valid = false;
2192 unsigned long lowest_key = 0;
2193 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2195 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2196 /* select the one with lowest rb_key */
2197 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2198 if (queue &&
2199 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2200 lowest_key = queue->rb_key;
2201 cur_best = i;
2202 key_valid = true;
2206 return cur_best;
2209 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2211 unsigned slice;
2212 unsigned count;
2213 struct cfq_rb_root *st;
2214 unsigned group_slice;
2215 enum wl_prio_t original_prio = cfqd->serving_prio;
2217 /* Choose next priority. RT > BE > IDLE */
2218 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2219 cfqd->serving_prio = RT_WORKLOAD;
2220 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2221 cfqd->serving_prio = BE_WORKLOAD;
2222 else {
2223 cfqd->serving_prio = IDLE_WORKLOAD;
2224 cfqd->workload_expires = jiffies + 1;
2225 return;
2228 if (original_prio != cfqd->serving_prio)
2229 goto new_workload;
2232 * For RT and BE, we have to choose also the type
2233 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2234 * expiration time
2236 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2237 count = st->count;
2240 * check workload expiration, and that we still have other queues ready
2242 if (count && !time_after(jiffies, cfqd->workload_expires))
2243 return;
2245 new_workload:
2246 /* otherwise select new workload type */
2247 cfqd->serving_type =
2248 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2249 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2250 count = st->count;
2253 * the workload slice is computed as a fraction of target latency
2254 * proportional to the number of queues in that workload, over
2255 * all the queues in the same priority class
2257 group_slice = cfq_group_slice(cfqd, cfqg);
2259 slice = group_slice * count /
2260 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2261 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2263 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2264 unsigned int tmp;
2267 * Async queues are currently system wide. Just taking
2268 * proportion of queues with-in same group will lead to higher
2269 * async ratio system wide as generally root group is going
2270 * to have higher weight. A more accurate thing would be to
2271 * calculate system wide asnc/sync ratio.
2273 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2274 tmp = tmp/cfqd->busy_queues;
2275 slice = min_t(unsigned, slice, tmp);
2277 /* async workload slice is scaled down according to
2278 * the sync/async slice ratio. */
2279 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2280 } else
2281 /* sync workload slice is at least 2 * cfq_slice_idle */
2282 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2284 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2285 cfq_log(cfqd, "workload slice:%d", slice);
2286 cfqd->workload_expires = jiffies + slice;
2289 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2291 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2292 struct cfq_group *cfqg;
2294 if (RB_EMPTY_ROOT(&st->rb))
2295 return NULL;
2296 cfqg = cfq_rb_first_group(st);
2297 update_min_vdisktime(st);
2298 return cfqg;
2301 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2303 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2305 cfqd->serving_group = cfqg;
2307 /* Restore the workload type data */
2308 if (cfqg->saved_workload_slice) {
2309 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2310 cfqd->serving_type = cfqg->saved_workload;
2311 cfqd->serving_prio = cfqg->saved_serving_prio;
2312 } else
2313 cfqd->workload_expires = jiffies - 1;
2315 choose_service_tree(cfqd, cfqg);
2319 * Select a queue for service. If we have a current active queue,
2320 * check whether to continue servicing it, or retrieve and set a new one.
2322 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2324 struct cfq_queue *cfqq, *new_cfqq = NULL;
2326 cfqq = cfqd->active_queue;
2327 if (!cfqq)
2328 goto new_queue;
2330 if (!cfqd->rq_queued)
2331 return NULL;
2334 * We were waiting for group to get backlogged. Expire the queue
2336 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2337 goto expire;
2340 * The active queue has run out of time, expire it and select new.
2342 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2344 * If slice had not expired at the completion of last request
2345 * we might not have turned on wait_busy flag. Don't expire
2346 * the queue yet. Allow the group to get backlogged.
2348 * The very fact that we have used the slice, that means we
2349 * have been idling all along on this queue and it should be
2350 * ok to wait for this request to complete.
2352 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2353 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2354 cfqq = NULL;
2355 goto keep_queue;
2356 } else
2357 goto check_group_idle;
2361 * The active queue has requests and isn't expired, allow it to
2362 * dispatch.
2364 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2365 goto keep_queue;
2368 * If another queue has a request waiting within our mean seek
2369 * distance, let it run. The expire code will check for close
2370 * cooperators and put the close queue at the front of the service
2371 * tree. If possible, merge the expiring queue with the new cfqq.
2373 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2374 if (new_cfqq) {
2375 if (!cfqq->new_cfqq)
2376 cfq_setup_merge(cfqq, new_cfqq);
2377 goto expire;
2381 * No requests pending. If the active queue still has requests in
2382 * flight or is idling for a new request, allow either of these
2383 * conditions to happen (or time out) before selecting a new queue.
2385 if (timer_pending(&cfqd->idle_slice_timer)) {
2386 cfqq = NULL;
2387 goto keep_queue;
2391 * This is a deep seek queue, but the device is much faster than
2392 * the queue can deliver, don't idle
2394 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2395 (cfq_cfqq_slice_new(cfqq) ||
2396 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2397 cfq_clear_cfqq_deep(cfqq);
2398 cfq_clear_cfqq_idle_window(cfqq);
2401 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2402 cfqq = NULL;
2403 goto keep_queue;
2407 * If group idle is enabled and there are requests dispatched from
2408 * this group, wait for requests to complete.
2410 check_group_idle:
2411 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2412 cfqq->cfqg->dispatched &&
2413 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2414 cfqq = NULL;
2415 goto keep_queue;
2418 expire:
2419 cfq_slice_expired(cfqd, 0);
2420 new_queue:
2422 * Current queue expired. Check if we have to switch to a new
2423 * service tree
2425 if (!new_cfqq)
2426 cfq_choose_cfqg(cfqd);
2428 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2429 keep_queue:
2430 return cfqq;
2433 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2435 int dispatched = 0;
2437 while (cfqq->next_rq) {
2438 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2439 dispatched++;
2442 BUG_ON(!list_empty(&cfqq->fifo));
2444 /* By default cfqq is not expired if it is empty. Do it explicitly */
2445 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2446 return dispatched;
2450 * Drain our current requests. Used for barriers and when switching
2451 * io schedulers on-the-fly.
2453 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2455 struct cfq_queue *cfqq;
2456 int dispatched = 0;
2458 /* Expire the timeslice of the current active queue first */
2459 cfq_slice_expired(cfqd, 0);
2460 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2461 __cfq_set_active_queue(cfqd, cfqq);
2462 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2465 BUG_ON(cfqd->busy_queues);
2467 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2468 return dispatched;
2471 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2472 struct cfq_queue *cfqq)
2474 /* the queue hasn't finished any request, can't estimate */
2475 if (cfq_cfqq_slice_new(cfqq))
2476 return true;
2477 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2478 cfqq->slice_end))
2479 return true;
2481 return false;
2484 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2486 unsigned int max_dispatch;
2489 * Drain async requests before we start sync IO
2491 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2492 return false;
2495 * If this is an async queue and we have sync IO in flight, let it wait
2497 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2498 return false;
2500 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2501 if (cfq_class_idle(cfqq))
2502 max_dispatch = 1;
2505 * Does this cfqq already have too much IO in flight?
2507 if (cfqq->dispatched >= max_dispatch) {
2508 bool promote_sync = false;
2510 * idle queue must always only have a single IO in flight
2512 if (cfq_class_idle(cfqq))
2513 return false;
2516 * If there is only one sync queue
2517 * we can ignore async queue here and give the sync
2518 * queue no dispatch limit. The reason is a sync queue can
2519 * preempt async queue, limiting the sync queue doesn't make
2520 * sense. This is useful for aiostress test.
2522 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2523 promote_sync = true;
2526 * We have other queues, don't allow more IO from this one
2528 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2529 !promote_sync)
2530 return false;
2533 * Sole queue user, no limit
2535 if (cfqd->busy_queues == 1 || promote_sync)
2536 max_dispatch = -1;
2537 else
2539 * Normally we start throttling cfqq when cfq_quantum/2
2540 * requests have been dispatched. But we can drive
2541 * deeper queue depths at the beginning of slice
2542 * subjected to upper limit of cfq_quantum.
2543 * */
2544 max_dispatch = cfqd->cfq_quantum;
2548 * Async queues must wait a bit before being allowed dispatch.
2549 * We also ramp up the dispatch depth gradually for async IO,
2550 * based on the last sync IO we serviced
2552 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2553 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2554 unsigned int depth;
2556 depth = last_sync / cfqd->cfq_slice[1];
2557 if (!depth && !cfqq->dispatched)
2558 depth = 1;
2559 if (depth < max_dispatch)
2560 max_dispatch = depth;
2564 * If we're below the current max, allow a dispatch
2566 return cfqq->dispatched < max_dispatch;
2570 * Dispatch a request from cfqq, moving them to the request queue
2571 * dispatch list.
2573 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2575 struct request *rq;
2577 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2579 if (!cfq_may_dispatch(cfqd, cfqq))
2580 return false;
2583 * follow expired path, else get first next available
2585 rq = cfq_check_fifo(cfqq);
2586 if (!rq)
2587 rq = cfqq->next_rq;
2590 * insert request into driver dispatch list
2592 cfq_dispatch_insert(cfqd->queue, rq);
2594 if (!cfqd->active_cic) {
2595 struct cfq_io_context *cic = RQ_CIC(rq);
2597 atomic_long_inc(&cic->ioc->refcount);
2598 cfqd->active_cic = cic;
2601 return true;
2605 * Find the cfqq that we need to service and move a request from that to the
2606 * dispatch list
2608 static int cfq_dispatch_requests(struct request_queue *q, int force)
2610 struct cfq_data *cfqd = q->elevator->elevator_data;
2611 struct cfq_queue *cfqq;
2613 if (!cfqd->busy_queues)
2614 return 0;
2616 if (unlikely(force))
2617 return cfq_forced_dispatch(cfqd);
2619 cfqq = cfq_select_queue(cfqd);
2620 if (!cfqq)
2621 return 0;
2624 * Dispatch a request from this cfqq, if it is allowed
2626 if (!cfq_dispatch_request(cfqd, cfqq))
2627 return 0;
2629 cfqq->slice_dispatch++;
2630 cfq_clear_cfqq_must_dispatch(cfqq);
2633 * expire an async queue immediately if it has used up its slice. idle
2634 * queue always expire after 1 dispatch round.
2636 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2637 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2638 cfq_class_idle(cfqq))) {
2639 cfqq->slice_end = jiffies + 1;
2640 cfq_slice_expired(cfqd, 0);
2643 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2644 return 1;
2648 * task holds one reference to the queue, dropped when task exits. each rq
2649 * in-flight on this queue also holds a reference, dropped when rq is freed.
2651 * Each cfq queue took a reference on the parent group. Drop it now.
2652 * queue lock must be held here.
2654 static void cfq_put_queue(struct cfq_queue *cfqq)
2656 struct cfq_data *cfqd = cfqq->cfqd;
2657 struct cfq_group *cfqg;
2659 BUG_ON(cfqq->ref <= 0);
2661 cfqq->ref--;
2662 if (cfqq->ref)
2663 return;
2665 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2666 BUG_ON(rb_first(&cfqq->sort_list));
2667 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2668 cfqg = cfqq->cfqg;
2670 if (unlikely(cfqd->active_queue == cfqq)) {
2671 __cfq_slice_expired(cfqd, cfqq, 0);
2672 cfq_schedule_dispatch(cfqd);
2675 BUG_ON(cfq_cfqq_on_rr(cfqq));
2676 kmem_cache_free(cfq_pool, cfqq);
2677 cfq_put_cfqg(cfqg);
2681 * Call func for each cic attached to this ioc.
2683 static void
2684 call_for_each_cic(struct io_context *ioc,
2685 void (*func)(struct io_context *, struct cfq_io_context *))
2687 struct cfq_io_context *cic;
2688 struct hlist_node *n;
2690 rcu_read_lock();
2692 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2693 func(ioc, cic);
2695 rcu_read_unlock();
2698 static void cfq_cic_free_rcu(struct rcu_head *head)
2700 struct cfq_io_context *cic;
2702 cic = container_of(head, struct cfq_io_context, rcu_head);
2704 kmem_cache_free(cfq_ioc_pool, cic);
2705 elv_ioc_count_dec(cfq_ioc_count);
2707 if (ioc_gone) {
2709 * CFQ scheduler is exiting, grab exit lock and check
2710 * the pending io context count. If it hits zero,
2711 * complete ioc_gone and set it back to NULL
2713 spin_lock(&ioc_gone_lock);
2714 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2715 complete(ioc_gone);
2716 ioc_gone = NULL;
2718 spin_unlock(&ioc_gone_lock);
2722 static void cfq_cic_free(struct cfq_io_context *cic)
2724 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2727 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2729 unsigned long flags;
2730 unsigned long dead_key = (unsigned long) cic->key;
2732 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2734 spin_lock_irqsave(&ioc->lock, flags);
2735 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2736 hlist_del_rcu(&cic->cic_list);
2737 spin_unlock_irqrestore(&ioc->lock, flags);
2739 cfq_cic_free(cic);
2743 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2744 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2745 * and ->trim() which is called with the task lock held
2747 static void cfq_free_io_context(struct io_context *ioc)
2750 * ioc->refcount is zero here, or we are called from elv_unregister(),
2751 * so no more cic's are allowed to be linked into this ioc. So it
2752 * should be ok to iterate over the known list, we will see all cic's
2753 * since no new ones are added.
2755 call_for_each_cic(ioc, cic_free_func);
2758 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2760 struct cfq_queue *__cfqq, *next;
2763 * If this queue was scheduled to merge with another queue, be
2764 * sure to drop the reference taken on that queue (and others in
2765 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2767 __cfqq = cfqq->new_cfqq;
2768 while (__cfqq) {
2769 if (__cfqq == cfqq) {
2770 WARN(1, "cfqq->new_cfqq loop detected\n");
2771 break;
2773 next = __cfqq->new_cfqq;
2774 cfq_put_queue(__cfqq);
2775 __cfqq = next;
2779 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2781 if (unlikely(cfqq == cfqd->active_queue)) {
2782 __cfq_slice_expired(cfqd, cfqq, 0);
2783 cfq_schedule_dispatch(cfqd);
2786 cfq_put_cooperator(cfqq);
2788 cfq_put_queue(cfqq);
2791 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2792 struct cfq_io_context *cic)
2794 struct io_context *ioc = cic->ioc;
2796 list_del_init(&cic->queue_list);
2799 * Make sure dead mark is seen for dead queues
2801 smp_wmb();
2802 cic->key = cfqd_dead_key(cfqd);
2804 rcu_read_lock();
2805 if (rcu_dereference(ioc->ioc_data) == cic) {
2806 rcu_read_unlock();
2807 spin_lock(&ioc->lock);
2808 rcu_assign_pointer(ioc->ioc_data, NULL);
2809 spin_unlock(&ioc->lock);
2810 } else
2811 rcu_read_unlock();
2813 if (cic->cfqq[BLK_RW_ASYNC]) {
2814 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2815 cic->cfqq[BLK_RW_ASYNC] = NULL;
2818 if (cic->cfqq[BLK_RW_SYNC]) {
2819 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2820 cic->cfqq[BLK_RW_SYNC] = NULL;
2824 static void cfq_exit_single_io_context(struct io_context *ioc,
2825 struct cfq_io_context *cic)
2827 struct cfq_data *cfqd = cic_to_cfqd(cic);
2829 if (cfqd) {
2830 struct request_queue *q = cfqd->queue;
2831 unsigned long flags;
2833 spin_lock_irqsave(q->queue_lock, flags);
2836 * Ensure we get a fresh copy of the ->key to prevent
2837 * race between exiting task and queue
2839 smp_read_barrier_depends();
2840 if (cic->key == cfqd)
2841 __cfq_exit_single_io_context(cfqd, cic);
2843 spin_unlock_irqrestore(q->queue_lock, flags);
2848 * The process that ioc belongs to has exited, we need to clean up
2849 * and put the internal structures we have that belongs to that process.
2851 static void cfq_exit_io_context(struct io_context *ioc)
2853 call_for_each_cic(ioc, cfq_exit_single_io_context);
2856 static struct cfq_io_context *
2857 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2859 struct cfq_io_context *cic;
2861 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2862 cfqd->queue->node);
2863 if (cic) {
2864 cic->ttime.last_end_request = jiffies;
2865 INIT_LIST_HEAD(&cic->queue_list);
2866 INIT_HLIST_NODE(&cic->cic_list);
2867 cic->dtor = cfq_free_io_context;
2868 cic->exit = cfq_exit_io_context;
2869 elv_ioc_count_inc(cfq_ioc_count);
2872 return cic;
2875 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2877 struct task_struct *tsk = current;
2878 int ioprio_class;
2880 if (!cfq_cfqq_prio_changed(cfqq))
2881 return;
2883 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2884 switch (ioprio_class) {
2885 default:
2886 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2887 case IOPRIO_CLASS_NONE:
2889 * no prio set, inherit CPU scheduling settings
2891 cfqq->ioprio = task_nice_ioprio(tsk);
2892 cfqq->ioprio_class = task_nice_ioclass(tsk);
2893 break;
2894 case IOPRIO_CLASS_RT:
2895 cfqq->ioprio = task_ioprio(ioc);
2896 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2897 break;
2898 case IOPRIO_CLASS_BE:
2899 cfqq->ioprio = task_ioprio(ioc);
2900 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2901 break;
2902 case IOPRIO_CLASS_IDLE:
2903 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2904 cfqq->ioprio = 7;
2905 cfq_clear_cfqq_idle_window(cfqq);
2906 break;
2910 * keep track of original prio settings in case we have to temporarily
2911 * elevate the priority of this queue
2913 cfqq->org_ioprio = cfqq->ioprio;
2914 cfq_clear_cfqq_prio_changed(cfqq);
2917 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2919 struct cfq_data *cfqd = cic_to_cfqd(cic);
2920 struct cfq_queue *cfqq;
2921 unsigned long flags;
2923 if (unlikely(!cfqd))
2924 return;
2926 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2928 cfqq = cic->cfqq[BLK_RW_ASYNC];
2929 if (cfqq) {
2930 struct cfq_queue *new_cfqq;
2931 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2932 GFP_ATOMIC);
2933 if (new_cfqq) {
2934 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2935 cfq_put_queue(cfqq);
2939 cfqq = cic->cfqq[BLK_RW_SYNC];
2940 if (cfqq)
2941 cfq_mark_cfqq_prio_changed(cfqq);
2943 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2946 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2948 call_for_each_cic(ioc, changed_ioprio);
2949 ioc->ioprio_changed = 0;
2952 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2953 pid_t pid, bool is_sync)
2955 RB_CLEAR_NODE(&cfqq->rb_node);
2956 RB_CLEAR_NODE(&cfqq->p_node);
2957 INIT_LIST_HEAD(&cfqq->fifo);
2959 cfqq->ref = 0;
2960 cfqq->cfqd = cfqd;
2962 cfq_mark_cfqq_prio_changed(cfqq);
2964 if (is_sync) {
2965 if (!cfq_class_idle(cfqq))
2966 cfq_mark_cfqq_idle_window(cfqq);
2967 cfq_mark_cfqq_sync(cfqq);
2969 cfqq->pid = pid;
2972 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2973 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2975 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2976 struct cfq_data *cfqd = cic_to_cfqd(cic);
2977 unsigned long flags;
2978 struct request_queue *q;
2980 if (unlikely(!cfqd))
2981 return;
2983 q = cfqd->queue;
2985 spin_lock_irqsave(q->queue_lock, flags);
2987 if (sync_cfqq) {
2989 * Drop reference to sync queue. A new sync queue will be
2990 * assigned in new group upon arrival of a fresh request.
2992 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2993 cic_set_cfqq(cic, NULL, 1);
2994 cfq_put_queue(sync_cfqq);
2997 spin_unlock_irqrestore(q->queue_lock, flags);
3000 static void cfq_ioc_set_cgroup(struct io_context *ioc)
3002 call_for_each_cic(ioc, changed_cgroup);
3003 ioc->cgroup_changed = 0;
3005 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
3007 static struct cfq_queue *
3008 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
3009 struct io_context *ioc, gfp_t gfp_mask)
3011 struct cfq_queue *cfqq, *new_cfqq = NULL;
3012 struct cfq_io_context *cic;
3013 struct cfq_group *cfqg;
3015 retry:
3016 cfqg = cfq_get_cfqg(cfqd);
3017 cic = cfq_cic_lookup(cfqd, ioc);
3018 /* cic always exists here */
3019 cfqq = cic_to_cfqq(cic, is_sync);
3022 * Always try a new alloc if we fell back to the OOM cfqq
3023 * originally, since it should just be a temporary situation.
3025 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3026 cfqq = NULL;
3027 if (new_cfqq) {
3028 cfqq = new_cfqq;
3029 new_cfqq = NULL;
3030 } else if (gfp_mask & __GFP_WAIT) {
3031 spin_unlock_irq(cfqd->queue->queue_lock);
3032 new_cfqq = kmem_cache_alloc_node(cfq_pool,
3033 gfp_mask | __GFP_ZERO,
3034 cfqd->queue->node);
3035 spin_lock_irq(cfqd->queue->queue_lock);
3036 if (new_cfqq)
3037 goto retry;
3038 } else {
3039 cfqq = kmem_cache_alloc_node(cfq_pool,
3040 gfp_mask | __GFP_ZERO,
3041 cfqd->queue->node);
3044 if (cfqq) {
3045 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3046 cfq_init_prio_data(cfqq, ioc);
3047 cfq_link_cfqq_cfqg(cfqq, cfqg);
3048 cfq_log_cfqq(cfqd, cfqq, "alloced");
3049 } else
3050 cfqq = &cfqd->oom_cfqq;
3053 if (new_cfqq)
3054 kmem_cache_free(cfq_pool, new_cfqq);
3056 return cfqq;
3059 static struct cfq_queue **
3060 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3062 switch (ioprio_class) {
3063 case IOPRIO_CLASS_RT:
3064 return &cfqd->async_cfqq[0][ioprio];
3065 case IOPRIO_CLASS_BE:
3066 return &cfqd->async_cfqq[1][ioprio];
3067 case IOPRIO_CLASS_IDLE:
3068 return &cfqd->async_idle_cfqq;
3069 default:
3070 BUG();
3074 static struct cfq_queue *
3075 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3076 gfp_t gfp_mask)
3078 const int ioprio = task_ioprio(ioc);
3079 const int ioprio_class = task_ioprio_class(ioc);
3080 struct cfq_queue **async_cfqq = NULL;
3081 struct cfq_queue *cfqq = NULL;
3083 if (!is_sync) {
3084 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3085 cfqq = *async_cfqq;
3088 if (!cfqq)
3089 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3092 * pin the queue now that it's allocated, scheduler exit will prune it
3094 if (!is_sync && !(*async_cfqq)) {
3095 cfqq->ref++;
3096 *async_cfqq = cfqq;
3099 cfqq->ref++;
3100 return cfqq;
3104 * We drop cfq io contexts lazily, so we may find a dead one.
3106 static void
3107 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3108 struct cfq_io_context *cic)
3110 unsigned long flags;
3112 WARN_ON(!list_empty(&cic->queue_list));
3113 BUG_ON(cic->key != cfqd_dead_key(cfqd));
3115 spin_lock_irqsave(&ioc->lock, flags);
3117 BUG_ON(rcu_dereference_check(ioc->ioc_data,
3118 lockdep_is_held(&ioc->lock)) == cic);
3120 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3121 hlist_del_rcu(&cic->cic_list);
3122 spin_unlock_irqrestore(&ioc->lock, flags);
3124 cfq_cic_free(cic);
3127 static struct cfq_io_context *
3128 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3130 struct cfq_io_context *cic;
3131 unsigned long flags;
3133 if (unlikely(!ioc))
3134 return NULL;
3136 rcu_read_lock();
3139 * we maintain a last-hit cache, to avoid browsing over the tree
3141 cic = rcu_dereference(ioc->ioc_data);
3142 if (cic && cic->key == cfqd) {
3143 rcu_read_unlock();
3144 return cic;
3147 do {
3148 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3149 rcu_read_unlock();
3150 if (!cic)
3151 break;
3152 if (unlikely(cic->key != cfqd)) {
3153 cfq_drop_dead_cic(cfqd, ioc, cic);
3154 rcu_read_lock();
3155 continue;
3158 spin_lock_irqsave(&ioc->lock, flags);
3159 rcu_assign_pointer(ioc->ioc_data, cic);
3160 spin_unlock_irqrestore(&ioc->lock, flags);
3161 break;
3162 } while (1);
3164 return cic;
3168 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3169 * the process specific cfq io context when entered from the block layer.
3170 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3172 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3173 struct cfq_io_context *cic, gfp_t gfp_mask)
3175 unsigned long flags;
3176 int ret;
3178 ret = radix_tree_preload(gfp_mask);
3179 if (!ret) {
3180 cic->ioc = ioc;
3181 cic->key = cfqd;
3183 spin_lock_irqsave(&ioc->lock, flags);
3184 ret = radix_tree_insert(&ioc->radix_root,
3185 cfqd->cic_index, cic);
3186 if (!ret)
3187 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3188 spin_unlock_irqrestore(&ioc->lock, flags);
3190 radix_tree_preload_end();
3192 if (!ret) {
3193 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3194 list_add(&cic->queue_list, &cfqd->cic_list);
3195 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3199 if (ret && ret != -EEXIST)
3200 printk(KERN_ERR "cfq: cic link failed!\n");
3202 return ret;
3206 * Setup general io context and cfq io context. There can be several cfq
3207 * io contexts per general io context, if this process is doing io to more
3208 * than one device managed by cfq.
3210 static struct cfq_io_context *
3211 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3213 struct io_context *ioc = NULL;
3214 struct cfq_io_context *cic;
3215 int ret;
3217 might_sleep_if(gfp_mask & __GFP_WAIT);
3219 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3220 if (!ioc)
3221 return NULL;
3223 retry:
3224 cic = cfq_cic_lookup(cfqd, ioc);
3225 if (cic)
3226 goto out;
3228 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3229 if (cic == NULL)
3230 goto err;
3232 ret = cfq_cic_link(cfqd, ioc, cic, gfp_mask);
3233 if (ret == -EEXIST) {
3234 /* someone has linked cic to ioc already */
3235 cfq_cic_free(cic);
3236 goto retry;
3237 } else if (ret)
3238 goto err_free;
3240 out:
3241 smp_read_barrier_depends();
3242 if (unlikely(ioc->ioprio_changed))
3243 cfq_ioc_set_ioprio(ioc);
3245 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3246 if (unlikely(ioc->cgroup_changed))
3247 cfq_ioc_set_cgroup(ioc);
3248 #endif
3249 return cic;
3250 err_free:
3251 cfq_cic_free(cic);
3252 err:
3253 put_io_context(ioc);
3254 return NULL;
3257 static void
3258 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3260 unsigned long elapsed = jiffies - ttime->last_end_request;
3261 elapsed = min(elapsed, 2UL * slice_idle);
3263 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3264 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3265 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3268 static void
3269 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3270 struct cfq_io_context *cic)
3272 if (cfq_cfqq_sync(cfqq)) {
3273 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3274 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3275 cfqd->cfq_slice_idle);
3277 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3278 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3279 #endif
3282 static void
3283 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3284 struct request *rq)
3286 sector_t sdist = 0;
3287 sector_t n_sec = blk_rq_sectors(rq);
3288 if (cfqq->last_request_pos) {
3289 if (cfqq->last_request_pos < blk_rq_pos(rq))
3290 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3291 else
3292 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3295 cfqq->seek_history <<= 1;
3296 if (blk_queue_nonrot(cfqd->queue))
3297 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3298 else
3299 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3303 * Disable idle window if the process thinks too long or seeks so much that
3304 * it doesn't matter
3306 static void
3307 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3308 struct cfq_io_context *cic)
3310 int old_idle, enable_idle;
3313 * Don't idle for async or idle io prio class
3315 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3316 return;
3318 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3320 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3321 cfq_mark_cfqq_deep(cfqq);
3323 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3324 enable_idle = 0;
3325 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3326 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3327 enable_idle = 0;
3328 else if (sample_valid(cic->ttime.ttime_samples)) {
3329 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3330 enable_idle = 0;
3331 else
3332 enable_idle = 1;
3335 if (old_idle != enable_idle) {
3336 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3337 if (enable_idle)
3338 cfq_mark_cfqq_idle_window(cfqq);
3339 else
3340 cfq_clear_cfqq_idle_window(cfqq);
3345 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3346 * no or if we aren't sure, a 1 will cause a preempt.
3348 static bool
3349 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3350 struct request *rq)
3352 struct cfq_queue *cfqq;
3354 cfqq = cfqd->active_queue;
3355 if (!cfqq)
3356 return false;
3358 if (cfq_class_idle(new_cfqq))
3359 return false;
3361 if (cfq_class_idle(cfqq))
3362 return true;
3365 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3367 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3368 return false;
3371 * if the new request is sync, but the currently running queue is
3372 * not, let the sync request have priority.
3374 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3375 return true;
3377 if (new_cfqq->cfqg != cfqq->cfqg)
3378 return false;
3380 if (cfq_slice_used(cfqq))
3381 return true;
3383 /* Allow preemption only if we are idling on sync-noidle tree */
3384 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3385 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3386 new_cfqq->service_tree->count == 2 &&
3387 RB_EMPTY_ROOT(&cfqq->sort_list))
3388 return true;
3391 * So both queues are sync. Let the new request get disk time if
3392 * it's a metadata request and the current queue is doing regular IO.
3394 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3395 return true;
3398 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3400 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3401 return true;
3403 /* An idle queue should not be idle now for some reason */
3404 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3405 return true;
3407 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3408 return false;
3411 * if this request is as-good as one we would expect from the
3412 * current cfqq, let it preempt
3414 if (cfq_rq_close(cfqd, cfqq, rq))
3415 return true;
3417 return false;
3421 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3422 * let it have half of its nominal slice.
3424 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3426 struct cfq_queue *old_cfqq = cfqd->active_queue;
3428 cfq_log_cfqq(cfqd, cfqq, "preempt");
3429 cfq_slice_expired(cfqd, 1);
3432 * workload type is changed, don't save slice, otherwise preempt
3433 * doesn't happen
3435 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3436 cfqq->cfqg->saved_workload_slice = 0;
3439 * Put the new queue at the front of the of the current list,
3440 * so we know that it will be selected next.
3442 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3444 cfq_service_tree_add(cfqd, cfqq, 1);
3446 cfqq->slice_end = 0;
3447 cfq_mark_cfqq_slice_new(cfqq);
3451 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3452 * something we should do about it
3454 static void
3455 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3456 struct request *rq)
3458 struct cfq_io_context *cic = RQ_CIC(rq);
3460 cfqd->rq_queued++;
3461 if (rq->cmd_flags & REQ_PRIO)
3462 cfqq->prio_pending++;
3464 cfq_update_io_thinktime(cfqd, cfqq, cic);
3465 cfq_update_io_seektime(cfqd, cfqq, rq);
3466 cfq_update_idle_window(cfqd, cfqq, cic);
3468 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3470 if (cfqq == cfqd->active_queue) {
3472 * Remember that we saw a request from this process, but
3473 * don't start queuing just yet. Otherwise we risk seeing lots
3474 * of tiny requests, because we disrupt the normal plugging
3475 * and merging. If the request is already larger than a single
3476 * page, let it rip immediately. For that case we assume that
3477 * merging is already done. Ditto for a busy system that
3478 * has other work pending, don't risk delaying until the
3479 * idle timer unplug to continue working.
3481 if (cfq_cfqq_wait_request(cfqq)) {
3482 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3483 cfqd->busy_queues > 1) {
3484 cfq_del_timer(cfqd, cfqq);
3485 cfq_clear_cfqq_wait_request(cfqq);
3486 __blk_run_queue(cfqd->queue);
3487 } else {
3488 cfq_blkiocg_update_idle_time_stats(
3489 &cfqq->cfqg->blkg);
3490 cfq_mark_cfqq_must_dispatch(cfqq);
3493 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3495 * not the active queue - expire current slice if it is
3496 * idle and has expired it's mean thinktime or this new queue
3497 * has some old slice time left and is of higher priority or
3498 * this new queue is RT and the current one is BE
3500 cfq_preempt_queue(cfqd, cfqq);
3501 __blk_run_queue(cfqd->queue);
3505 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3507 struct cfq_data *cfqd = q->elevator->elevator_data;
3508 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3510 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3511 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3513 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3514 list_add_tail(&rq->queuelist, &cfqq->fifo);
3515 cfq_add_rq_rb(rq);
3516 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3517 &cfqd->serving_group->blkg, rq_data_dir(rq),
3518 rq_is_sync(rq));
3519 cfq_rq_enqueued(cfqd, cfqq, rq);
3523 * Update hw_tag based on peak queue depth over 50 samples under
3524 * sufficient load.
3526 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3528 struct cfq_queue *cfqq = cfqd->active_queue;
3530 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3531 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3533 if (cfqd->hw_tag == 1)
3534 return;
3536 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3537 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3538 return;
3541 * If active queue hasn't enough requests and can idle, cfq might not
3542 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3543 * case
3545 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3546 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3547 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3548 return;
3550 if (cfqd->hw_tag_samples++ < 50)
3551 return;
3553 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3554 cfqd->hw_tag = 1;
3555 else
3556 cfqd->hw_tag = 0;
3559 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3561 struct cfq_io_context *cic = cfqd->active_cic;
3563 /* If the queue already has requests, don't wait */
3564 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3565 return false;
3567 /* If there are other queues in the group, don't wait */
3568 if (cfqq->cfqg->nr_cfqq > 1)
3569 return false;
3571 /* the only queue in the group, but think time is big */
3572 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3573 return false;
3575 if (cfq_slice_used(cfqq))
3576 return true;
3578 /* if slice left is less than think time, wait busy */
3579 if (cic && sample_valid(cic->ttime.ttime_samples)
3580 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3581 return true;
3584 * If think times is less than a jiffy than ttime_mean=0 and above
3585 * will not be true. It might happen that slice has not expired yet
3586 * but will expire soon (4-5 ns) during select_queue(). To cover the
3587 * case where think time is less than a jiffy, mark the queue wait
3588 * busy if only 1 jiffy is left in the slice.
3590 if (cfqq->slice_end - jiffies == 1)
3591 return true;
3593 return false;
3596 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3598 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3599 struct cfq_data *cfqd = cfqq->cfqd;
3600 const int sync = rq_is_sync(rq);
3601 unsigned long now;
3603 now = jiffies;
3604 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3605 !!(rq->cmd_flags & REQ_NOIDLE));
3607 cfq_update_hw_tag(cfqd);
3609 WARN_ON(!cfqd->rq_in_driver);
3610 WARN_ON(!cfqq->dispatched);
3611 cfqd->rq_in_driver--;
3612 cfqq->dispatched--;
3613 (RQ_CFQG(rq))->dispatched--;
3614 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3615 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3616 rq_data_dir(rq), rq_is_sync(rq));
3618 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3620 if (sync) {
3621 struct cfq_rb_root *service_tree;
3623 RQ_CIC(rq)->ttime.last_end_request = now;
3625 if (cfq_cfqq_on_rr(cfqq))
3626 service_tree = cfqq->service_tree;
3627 else
3628 service_tree = service_tree_for(cfqq->cfqg,
3629 cfqq_prio(cfqq), cfqq_type(cfqq));
3630 service_tree->ttime.last_end_request = now;
3631 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3632 cfqd->last_delayed_sync = now;
3635 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3636 cfqq->cfqg->ttime.last_end_request = now;
3637 #endif
3640 * If this is the active queue, check if it needs to be expired,
3641 * or if we want to idle in case it has no pending requests.
3643 if (cfqd->active_queue == cfqq) {
3644 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3646 if (cfq_cfqq_slice_new(cfqq)) {
3647 cfq_set_prio_slice(cfqd, cfqq);
3648 cfq_clear_cfqq_slice_new(cfqq);
3652 * Should we wait for next request to come in before we expire
3653 * the queue.
3655 if (cfq_should_wait_busy(cfqd, cfqq)) {
3656 unsigned long extend_sl = cfqd->cfq_slice_idle;
3657 if (!cfqd->cfq_slice_idle)
3658 extend_sl = cfqd->cfq_group_idle;
3659 cfqq->slice_end = jiffies + extend_sl;
3660 cfq_mark_cfqq_wait_busy(cfqq);
3661 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3665 * Idling is not enabled on:
3666 * - expired queues
3667 * - idle-priority queues
3668 * - async queues
3669 * - queues with still some requests queued
3670 * - when there is a close cooperator
3672 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3673 cfq_slice_expired(cfqd, 1);
3674 else if (sync && cfqq_empty &&
3675 !cfq_close_cooperator(cfqd, cfqq)) {
3676 cfq_arm_slice_timer(cfqd);
3680 if (!cfqd->rq_in_driver)
3681 cfq_schedule_dispatch(cfqd);
3684 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3686 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3687 cfq_mark_cfqq_must_alloc_slice(cfqq);
3688 return ELV_MQUEUE_MUST;
3691 return ELV_MQUEUE_MAY;
3694 static int cfq_may_queue(struct request_queue *q, int rw)
3696 struct cfq_data *cfqd = q->elevator->elevator_data;
3697 struct task_struct *tsk = current;
3698 struct cfq_io_context *cic;
3699 struct cfq_queue *cfqq;
3702 * don't force setup of a queue from here, as a call to may_queue
3703 * does not necessarily imply that a request actually will be queued.
3704 * so just lookup a possibly existing queue, or return 'may queue'
3705 * if that fails
3707 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3708 if (!cic)
3709 return ELV_MQUEUE_MAY;
3711 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3712 if (cfqq) {
3713 cfq_init_prio_data(cfqq, cic->ioc);
3715 return __cfq_may_queue(cfqq);
3718 return ELV_MQUEUE_MAY;
3722 * queue lock held here
3724 static void cfq_put_request(struct request *rq)
3726 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3728 if (cfqq) {
3729 const int rw = rq_data_dir(rq);
3731 BUG_ON(!cfqq->allocated[rw]);
3732 cfqq->allocated[rw]--;
3734 put_io_context(RQ_CIC(rq)->ioc);
3736 rq->elevator_private[0] = NULL;
3737 rq->elevator_private[1] = NULL;
3739 /* Put down rq reference on cfqg */
3740 cfq_put_cfqg(RQ_CFQG(rq));
3741 rq->elevator_private[2] = NULL;
3743 cfq_put_queue(cfqq);
3747 static struct cfq_queue *
3748 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3749 struct cfq_queue *cfqq)
3751 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3752 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3753 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3754 cfq_put_queue(cfqq);
3755 return cic_to_cfqq(cic, 1);
3759 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3760 * was the last process referring to said cfqq.
3762 static struct cfq_queue *
3763 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3765 if (cfqq_process_refs(cfqq) == 1) {
3766 cfqq->pid = current->pid;
3767 cfq_clear_cfqq_coop(cfqq);
3768 cfq_clear_cfqq_split_coop(cfqq);
3769 return cfqq;
3772 cic_set_cfqq(cic, NULL, 1);
3774 cfq_put_cooperator(cfqq);
3776 cfq_put_queue(cfqq);
3777 return NULL;
3780 * Allocate cfq data structures associated with this request.
3782 static int
3783 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3785 struct cfq_data *cfqd = q->elevator->elevator_data;
3786 struct cfq_io_context *cic;
3787 const int rw = rq_data_dir(rq);
3788 const bool is_sync = rq_is_sync(rq);
3789 struct cfq_queue *cfqq;
3790 unsigned long flags;
3792 might_sleep_if(gfp_mask & __GFP_WAIT);
3794 cic = cfq_get_io_context(cfqd, gfp_mask);
3796 spin_lock_irqsave(q->queue_lock, flags);
3798 if (!cic)
3799 goto queue_fail;
3801 new_queue:
3802 cfqq = cic_to_cfqq(cic, is_sync);
3803 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3804 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3805 cic_set_cfqq(cic, cfqq, is_sync);
3806 } else {
3808 * If the queue was seeky for too long, break it apart.
3810 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3811 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3812 cfqq = split_cfqq(cic, cfqq);
3813 if (!cfqq)
3814 goto new_queue;
3818 * Check to see if this queue is scheduled to merge with
3819 * another, closely cooperating queue. The merging of
3820 * queues happens here as it must be done in process context.
3821 * The reference on new_cfqq was taken in merge_cfqqs.
3823 if (cfqq->new_cfqq)
3824 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3827 cfqq->allocated[rw]++;
3829 cfqq->ref++;
3830 rq->elevator_private[0] = cic;
3831 rq->elevator_private[1] = cfqq;
3832 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3833 spin_unlock_irqrestore(q->queue_lock, flags);
3834 return 0;
3836 queue_fail:
3837 cfq_schedule_dispatch(cfqd);
3838 spin_unlock_irqrestore(q->queue_lock, flags);
3839 cfq_log(cfqd, "set_request fail");
3840 return 1;
3843 static void cfq_kick_queue(struct work_struct *work)
3845 struct cfq_data *cfqd =
3846 container_of(work, struct cfq_data, unplug_work);
3847 struct request_queue *q = cfqd->queue;
3849 spin_lock_irq(q->queue_lock);
3850 __blk_run_queue(cfqd->queue);
3851 spin_unlock_irq(q->queue_lock);
3855 * Timer running if the active_queue is currently idling inside its time slice
3857 static void cfq_idle_slice_timer(unsigned long data)
3859 struct cfq_data *cfqd = (struct cfq_data *) data;
3860 struct cfq_queue *cfqq;
3861 unsigned long flags;
3862 int timed_out = 1;
3864 cfq_log(cfqd, "idle timer fired");
3866 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3868 cfqq = cfqd->active_queue;
3869 if (cfqq) {
3870 timed_out = 0;
3873 * We saw a request before the queue expired, let it through
3875 if (cfq_cfqq_must_dispatch(cfqq))
3876 goto out_kick;
3879 * expired
3881 if (cfq_slice_used(cfqq))
3882 goto expire;
3885 * only expire and reinvoke request handler, if there are
3886 * other queues with pending requests
3888 if (!cfqd->busy_queues)
3889 goto out_cont;
3892 * not expired and it has a request pending, let it dispatch
3894 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3895 goto out_kick;
3898 * Queue depth flag is reset only when the idle didn't succeed
3900 cfq_clear_cfqq_deep(cfqq);
3902 expire:
3903 cfq_slice_expired(cfqd, timed_out);
3904 out_kick:
3905 cfq_schedule_dispatch(cfqd);
3906 out_cont:
3907 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3910 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3912 del_timer_sync(&cfqd->idle_slice_timer);
3913 cancel_work_sync(&cfqd->unplug_work);
3916 static void cfq_put_async_queues(struct cfq_data *cfqd)
3918 int i;
3920 for (i = 0; i < IOPRIO_BE_NR; i++) {
3921 if (cfqd->async_cfqq[0][i])
3922 cfq_put_queue(cfqd->async_cfqq[0][i]);
3923 if (cfqd->async_cfqq[1][i])
3924 cfq_put_queue(cfqd->async_cfqq[1][i]);
3927 if (cfqd->async_idle_cfqq)
3928 cfq_put_queue(cfqd->async_idle_cfqq);
3931 static void cfq_exit_queue(struct elevator_queue *e)
3933 struct cfq_data *cfqd = e->elevator_data;
3934 struct request_queue *q = cfqd->queue;
3935 bool wait = false;
3937 cfq_shutdown_timer_wq(cfqd);
3939 spin_lock_irq(q->queue_lock);
3941 if (cfqd->active_queue)
3942 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3944 while (!list_empty(&cfqd->cic_list)) {
3945 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3946 struct cfq_io_context,
3947 queue_list);
3949 __cfq_exit_single_io_context(cfqd, cic);
3952 cfq_put_async_queues(cfqd);
3953 cfq_release_cfq_groups(cfqd);
3956 * If there are groups which we could not unlink from blkcg list,
3957 * wait for a rcu period for them to be freed.
3959 if (cfqd->nr_blkcg_linked_grps)
3960 wait = true;
3962 spin_unlock_irq(q->queue_lock);
3964 cfq_shutdown_timer_wq(cfqd);
3966 spin_lock(&cic_index_lock);
3967 ida_remove(&cic_index_ida, cfqd->cic_index);
3968 spin_unlock(&cic_index_lock);
3971 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3972 * Do this wait only if there are other unlinked groups out
3973 * there. This can happen if cgroup deletion path claimed the
3974 * responsibility of cleaning up a group before queue cleanup code
3975 * get to the group.
3977 * Do not call synchronize_rcu() unconditionally as there are drivers
3978 * which create/delete request queue hundreds of times during scan/boot
3979 * and synchronize_rcu() can take significant time and slow down boot.
3981 if (wait)
3982 synchronize_rcu();
3984 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3985 /* Free up per cpu stats for root group */
3986 free_percpu(cfqd->root_group.blkg.stats_cpu);
3987 #endif
3988 kfree(cfqd);
3991 static int cfq_alloc_cic_index(void)
3993 int index, error;
3995 do {
3996 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3997 return -ENOMEM;
3999 spin_lock(&cic_index_lock);
4000 error = ida_get_new(&cic_index_ida, &index);
4001 spin_unlock(&cic_index_lock);
4002 if (error && error != -EAGAIN)
4003 return error;
4004 } while (error);
4006 return index;
4009 static void *cfq_init_queue(struct request_queue *q)
4011 struct cfq_data *cfqd;
4012 int i, j;
4013 struct cfq_group *cfqg;
4014 struct cfq_rb_root *st;
4016 i = cfq_alloc_cic_index();
4017 if (i < 0)
4018 return NULL;
4020 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4021 if (!cfqd) {
4022 spin_lock(&cic_index_lock);
4023 ida_remove(&cic_index_ida, i);
4024 spin_unlock(&cic_index_lock);
4025 return NULL;
4029 * Don't need take queue_lock in the routine, since we are
4030 * initializing the ioscheduler, and nobody is using cfqd
4032 cfqd->cic_index = i;
4034 /* Init root service tree */
4035 cfqd->grp_service_tree = CFQ_RB_ROOT;
4037 /* Init root group */
4038 cfqg = &cfqd->root_group;
4039 for_each_cfqg_st(cfqg, i, j, st)
4040 *st = CFQ_RB_ROOT;
4041 RB_CLEAR_NODE(&cfqg->rb_node);
4043 /* Give preference to root group over other groups */
4044 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
4046 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4048 * Set root group reference to 2. One reference will be dropped when
4049 * all groups on cfqd->cfqg_list are being deleted during queue exit.
4050 * Other reference will remain there as we don't want to delete this
4051 * group as it is statically allocated and gets destroyed when
4052 * throtl_data goes away.
4054 cfqg->ref = 2;
4056 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4057 kfree(cfqg);
4059 spin_lock(&cic_index_lock);
4060 ida_remove(&cic_index_ida, cfqd->cic_index);
4061 spin_unlock(&cic_index_lock);
4063 kfree(cfqd);
4064 return NULL;
4067 rcu_read_lock();
4069 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4070 (void *)cfqd, 0);
4071 rcu_read_unlock();
4072 cfqd->nr_blkcg_linked_grps++;
4074 /* Add group on cfqd->cfqg_list */
4075 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4076 #endif
4078 * Not strictly needed (since RB_ROOT just clears the node and we
4079 * zeroed cfqd on alloc), but better be safe in case someone decides
4080 * to add magic to the rb code
4082 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4083 cfqd->prio_trees[i] = RB_ROOT;
4086 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4087 * Grab a permanent reference to it, so that the normal code flow
4088 * will not attempt to free it.
4090 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4091 cfqd->oom_cfqq.ref++;
4092 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4094 INIT_LIST_HEAD(&cfqd->cic_list);
4096 cfqd->queue = q;
4098 init_timer(&cfqd->idle_slice_timer);
4099 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4100 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4102 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4104 cfqd->cfq_quantum = cfq_quantum;
4105 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4106 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4107 cfqd->cfq_back_max = cfq_back_max;
4108 cfqd->cfq_back_penalty = cfq_back_penalty;
4109 cfqd->cfq_slice[0] = cfq_slice_async;
4110 cfqd->cfq_slice[1] = cfq_slice_sync;
4111 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4112 cfqd->cfq_slice_idle = cfq_slice_idle;
4113 cfqd->cfq_group_idle = cfq_group_idle;
4114 cfqd->cfq_latency = 1;
4115 cfqd->hw_tag = -1;
4117 * we optimistically start assuming sync ops weren't delayed in last
4118 * second, in order to have larger depth for async operations.
4120 cfqd->last_delayed_sync = jiffies - HZ;
4121 return cfqd;
4124 static void cfq_slab_kill(void)
4127 * Caller already ensured that pending RCU callbacks are completed,
4128 * so we should have no busy allocations at this point.
4130 if (cfq_pool)
4131 kmem_cache_destroy(cfq_pool);
4132 if (cfq_ioc_pool)
4133 kmem_cache_destroy(cfq_ioc_pool);
4136 static int __init cfq_slab_setup(void)
4138 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4139 if (!cfq_pool)
4140 goto fail;
4142 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4143 if (!cfq_ioc_pool)
4144 goto fail;
4146 return 0;
4147 fail:
4148 cfq_slab_kill();
4149 return -ENOMEM;
4153 * sysfs parts below -->
4155 static ssize_t
4156 cfq_var_show(unsigned int var, char *page)
4158 return sprintf(page, "%d\n", var);
4161 static ssize_t
4162 cfq_var_store(unsigned int *var, const char *page, size_t count)
4164 char *p = (char *) page;
4166 *var = simple_strtoul(p, &p, 10);
4167 return count;
4170 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4171 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4173 struct cfq_data *cfqd = e->elevator_data; \
4174 unsigned int __data = __VAR; \
4175 if (__CONV) \
4176 __data = jiffies_to_msecs(__data); \
4177 return cfq_var_show(__data, (page)); \
4179 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4180 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4181 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4182 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4183 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4184 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4185 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4186 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4187 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4188 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4189 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4190 #undef SHOW_FUNCTION
4192 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4193 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4195 struct cfq_data *cfqd = e->elevator_data; \
4196 unsigned int __data; \
4197 int ret = cfq_var_store(&__data, (page), count); \
4198 if (__data < (MIN)) \
4199 __data = (MIN); \
4200 else if (__data > (MAX)) \
4201 __data = (MAX); \
4202 if (__CONV) \
4203 *(__PTR) = msecs_to_jiffies(__data); \
4204 else \
4205 *(__PTR) = __data; \
4206 return ret; \
4208 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4209 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4210 UINT_MAX, 1);
4211 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4212 UINT_MAX, 1);
4213 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4214 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4215 UINT_MAX, 0);
4216 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4217 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4218 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4219 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4220 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4221 UINT_MAX, 0);
4222 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4223 #undef STORE_FUNCTION
4225 #define CFQ_ATTR(name) \
4226 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4228 static struct elv_fs_entry cfq_attrs[] = {
4229 CFQ_ATTR(quantum),
4230 CFQ_ATTR(fifo_expire_sync),
4231 CFQ_ATTR(fifo_expire_async),
4232 CFQ_ATTR(back_seek_max),
4233 CFQ_ATTR(back_seek_penalty),
4234 CFQ_ATTR(slice_sync),
4235 CFQ_ATTR(slice_async),
4236 CFQ_ATTR(slice_async_rq),
4237 CFQ_ATTR(slice_idle),
4238 CFQ_ATTR(group_idle),
4239 CFQ_ATTR(low_latency),
4240 __ATTR_NULL
4243 static struct elevator_type iosched_cfq = {
4244 .ops = {
4245 .elevator_merge_fn = cfq_merge,
4246 .elevator_merged_fn = cfq_merged_request,
4247 .elevator_merge_req_fn = cfq_merged_requests,
4248 .elevator_allow_merge_fn = cfq_allow_merge,
4249 .elevator_bio_merged_fn = cfq_bio_merged,
4250 .elevator_dispatch_fn = cfq_dispatch_requests,
4251 .elevator_add_req_fn = cfq_insert_request,
4252 .elevator_activate_req_fn = cfq_activate_request,
4253 .elevator_deactivate_req_fn = cfq_deactivate_request,
4254 .elevator_completed_req_fn = cfq_completed_request,
4255 .elevator_former_req_fn = elv_rb_former_request,
4256 .elevator_latter_req_fn = elv_rb_latter_request,
4257 .elevator_set_req_fn = cfq_set_request,
4258 .elevator_put_req_fn = cfq_put_request,
4259 .elevator_may_queue_fn = cfq_may_queue,
4260 .elevator_init_fn = cfq_init_queue,
4261 .elevator_exit_fn = cfq_exit_queue,
4262 .trim = cfq_free_io_context,
4264 .elevator_attrs = cfq_attrs,
4265 .elevator_name = "cfq",
4266 .elevator_owner = THIS_MODULE,
4269 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4270 static struct blkio_policy_type blkio_policy_cfq = {
4271 .ops = {
4272 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4273 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4275 .plid = BLKIO_POLICY_PROP,
4277 #else
4278 static struct blkio_policy_type blkio_policy_cfq;
4279 #endif
4281 static int __init cfq_init(void)
4284 * could be 0 on HZ < 1000 setups
4286 if (!cfq_slice_async)
4287 cfq_slice_async = 1;
4288 if (!cfq_slice_idle)
4289 cfq_slice_idle = 1;
4291 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4292 if (!cfq_group_idle)
4293 cfq_group_idle = 1;
4294 #else
4295 cfq_group_idle = 0;
4296 #endif
4297 if (cfq_slab_setup())
4298 return -ENOMEM;
4300 elv_register(&iosched_cfq);
4301 blkio_policy_register(&blkio_policy_cfq);
4303 return 0;
4306 static void __exit cfq_exit(void)
4308 DECLARE_COMPLETION_ONSTACK(all_gone);
4309 blkio_policy_unregister(&blkio_policy_cfq);
4310 elv_unregister(&iosched_cfq);
4311 ioc_gone = &all_gone;
4312 /* ioc_gone's update must be visible before reading ioc_count */
4313 smp_wmb();
4316 * this also protects us from entering cfq_slab_kill() with
4317 * pending RCU callbacks
4319 if (elv_ioc_count_read(cfq_ioc_count))
4320 wait_for_completion(&all_gone);
4321 ida_destroy(&cic_index_ida);
4322 cfq_slab_kill();
4325 module_init(cfq_init);
4326 module_exit(cfq_exit);
4328 MODULE_AUTHOR("Jens Axboe");
4329 MODULE_LICENSE("GPL");
4330 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");