ALSA: hda - Remove ALC268 Dell, Toshiba and Zapto model quirks
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / block / cfq-iosched.c
blob1f96ad6254f1eea18478b80f3a58de4173c9f8ef
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 /* number of requests that are on the dispatch list or inside driver */
134 int dispatched;
136 /* io prio of this group */
137 unsigned short ioprio, org_ioprio;
138 unsigned short ioprio_class;
140 pid_t pid;
142 u32 seek_history;
143 sector_t last_request_pos;
145 struct cfq_rb_root *service_tree;
146 struct cfq_queue *new_cfqq;
147 struct cfq_group *cfqg;
148 /* Number of sectors dispatched from queue in single dispatch round */
149 unsigned long nr_sectors;
153 * First index in the service_trees.
154 * IDLE is handled separately, so it has negative index
156 enum wl_prio_t {
157 BE_WORKLOAD = 0,
158 RT_WORKLOAD = 1,
159 IDLE_WORKLOAD = 2,
160 CFQ_PRIO_NR,
164 * Second index in the service_trees.
166 enum wl_type_t {
167 ASYNC_WORKLOAD = 0,
168 SYNC_NOIDLE_WORKLOAD = 1,
169 SYNC_WORKLOAD = 2
172 /* This is per cgroup per device grouping structure */
173 struct cfq_group {
174 /* group service_tree member */
175 struct rb_node rb_node;
177 /* group service_tree key */
178 u64 vdisktime;
179 unsigned int weight;
180 unsigned int new_weight;
181 bool needs_update;
183 /* number of cfqq currently on this group */
184 int nr_cfqq;
187 * Per group busy queues average. Useful for workload slice calc. We
188 * create the array for each prio class but at run time it is used
189 * only for RT and BE class and slot for IDLE class remains unused.
190 * This is primarily done to avoid confusion and a gcc warning.
192 unsigned int busy_queues_avg[CFQ_PRIO_NR];
194 * rr lists of queues with requests. We maintain service trees for
195 * RT and BE classes. These trees are subdivided in subclasses
196 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
197 * class there is no subclassification and all the cfq queues go on
198 * a single tree service_tree_idle.
199 * Counts are embedded in the cfq_rb_root
201 struct cfq_rb_root service_trees[2][3];
202 struct cfq_rb_root service_tree_idle;
204 unsigned long saved_workload_slice;
205 enum wl_type_t saved_workload;
206 enum wl_prio_t saved_serving_prio;
207 struct blkio_group blkg;
208 #ifdef CONFIG_CFQ_GROUP_IOSCHED
209 struct hlist_node cfqd_node;
210 int ref;
211 #endif
212 /* number of requests that are on the dispatch list or inside driver */
213 int dispatched;
214 struct cfq_ttime ttime;
218 * Per block device queue structure
220 struct cfq_data {
221 struct request_queue *queue;
222 /* Root service tree for cfq_groups */
223 struct cfq_rb_root grp_service_tree;
224 struct cfq_group root_group;
227 * The priority currently being served
229 enum wl_prio_t serving_prio;
230 enum wl_type_t serving_type;
231 unsigned long workload_expires;
232 struct cfq_group *serving_group;
235 * Each priority tree is sorted by next_request position. These
236 * trees are used when determining if two or more queues are
237 * interleaving requests (see cfq_close_cooperator).
239 struct rb_root prio_trees[CFQ_PRIO_LISTS];
241 unsigned int busy_queues;
242 unsigned int busy_sync_queues;
244 int rq_in_driver;
245 int rq_in_flight[2];
248 * queue-depth detection
250 int rq_queued;
251 int hw_tag;
253 * hw_tag can be
254 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256 * 0 => no NCQ
258 int hw_tag_est_depth;
259 unsigned int hw_tag_samples;
262 * idle window management
264 struct timer_list idle_slice_timer;
265 struct work_struct unplug_work;
267 struct cfq_queue *active_queue;
268 struct cfq_io_context *active_cic;
271 * async queue for each priority case
273 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274 struct cfq_queue *async_idle_cfqq;
276 sector_t last_position;
279 * tunables, see top of file
281 unsigned int cfq_quantum;
282 unsigned int cfq_fifo_expire[2];
283 unsigned int cfq_back_penalty;
284 unsigned int cfq_back_max;
285 unsigned int cfq_slice[2];
286 unsigned int cfq_slice_async_rq;
287 unsigned int cfq_slice_idle;
288 unsigned int cfq_group_idle;
289 unsigned int cfq_latency;
291 unsigned int cic_index;
292 struct list_head cic_list;
295 * Fallback dummy cfqq for extreme OOM conditions
297 struct cfq_queue oom_cfqq;
299 unsigned long last_delayed_sync;
301 /* List of cfq groups being managed on this device*/
302 struct hlist_head cfqg_list;
304 /* Number of groups which are on blkcg->blkg_list */
305 unsigned int nr_blkcg_linked_grps;
308 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
310 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
311 enum wl_prio_t prio,
312 enum wl_type_t type)
314 if (!cfqg)
315 return NULL;
317 if (prio == IDLE_WORKLOAD)
318 return &cfqg->service_tree_idle;
320 return &cfqg->service_trees[prio][type];
323 enum cfqq_state_flags {
324 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
325 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
326 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
327 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
328 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
329 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
330 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
331 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
332 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
333 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
334 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
335 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
336 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
339 #define CFQ_CFQQ_FNS(name) \
340 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
342 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
344 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
346 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
348 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
350 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
353 CFQ_CFQQ_FNS(on_rr);
354 CFQ_CFQQ_FNS(wait_request);
355 CFQ_CFQQ_FNS(must_dispatch);
356 CFQ_CFQQ_FNS(must_alloc_slice);
357 CFQ_CFQQ_FNS(fifo_expire);
358 CFQ_CFQQ_FNS(idle_window);
359 CFQ_CFQQ_FNS(prio_changed);
360 CFQ_CFQQ_FNS(slice_new);
361 CFQ_CFQQ_FNS(sync);
362 CFQ_CFQQ_FNS(coop);
363 CFQ_CFQQ_FNS(split_coop);
364 CFQ_CFQQ_FNS(deep);
365 CFQ_CFQQ_FNS(wait_busy);
366 #undef CFQ_CFQQ_FNS
368 #ifdef CONFIG_CFQ_GROUP_IOSCHED
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
370 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
371 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
372 blkg_path(&(cfqq)->cfqg->blkg), ##args)
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
376 blkg_path(&(cfqg)->blkg), ##args) \
378 #else
379 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
382 #endif
383 #define cfq_log(cfqd, fmt, args...) \
384 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
386 /* Traverses through cfq group service trees */
387 #define for_each_cfqg_st(cfqg, i, j, st) \
388 for (i = 0; i <= IDLE_WORKLOAD; i++) \
389 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
390 : &cfqg->service_tree_idle; \
391 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
392 (i == IDLE_WORKLOAD && j == 0); \
393 j++, st = i < IDLE_WORKLOAD ? \
394 &cfqg->service_trees[i][j]: NULL) \
396 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
397 struct cfq_ttime *ttime, bool group_idle)
399 unsigned long slice;
400 if (!sample_valid(ttime->ttime_samples))
401 return false;
402 if (group_idle)
403 slice = cfqd->cfq_group_idle;
404 else
405 slice = cfqd->cfq_slice_idle;
406 return ttime->ttime_mean > slice;
409 static inline bool iops_mode(struct cfq_data *cfqd)
412 * If we are not idling on queues and it is a NCQ drive, parallel
413 * execution of requests is on and measuring time is not possible
414 * in most of the cases until and unless we drive shallower queue
415 * depths and that becomes a performance bottleneck. In such cases
416 * switch to start providing fairness in terms of number of IOs.
418 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
419 return true;
420 else
421 return false;
424 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
426 if (cfq_class_idle(cfqq))
427 return IDLE_WORKLOAD;
428 if (cfq_class_rt(cfqq))
429 return RT_WORKLOAD;
430 return BE_WORKLOAD;
434 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
436 if (!cfq_cfqq_sync(cfqq))
437 return ASYNC_WORKLOAD;
438 if (!cfq_cfqq_idle_window(cfqq))
439 return SYNC_NOIDLE_WORKLOAD;
440 return SYNC_WORKLOAD;
443 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
444 struct cfq_data *cfqd,
445 struct cfq_group *cfqg)
447 if (wl == IDLE_WORKLOAD)
448 return cfqg->service_tree_idle.count;
450 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
451 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
452 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
455 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
456 struct cfq_group *cfqg)
458 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
459 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
462 static void cfq_dispatch_insert(struct request_queue *, struct request *);
463 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
464 struct io_context *, gfp_t);
465 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
466 struct io_context *);
468 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
469 bool is_sync)
471 return cic->cfqq[is_sync];
474 static inline void cic_set_cfqq(struct cfq_io_context *cic,
475 struct cfq_queue *cfqq, bool is_sync)
477 cic->cfqq[is_sync] = cfqq;
480 #define CIC_DEAD_KEY 1ul
481 #define CIC_DEAD_INDEX_SHIFT 1
483 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
485 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
488 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
490 struct cfq_data *cfqd = cic->key;
492 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
493 return NULL;
495 return cfqd;
499 * We regard a request as SYNC, if it's either a read or has the SYNC bit
500 * set (in which case it could also be direct WRITE).
502 static inline bool cfq_bio_sync(struct bio *bio)
504 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
508 * scheduler run of queue, if there are requests pending and no one in the
509 * driver that will restart queueing
511 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
513 if (cfqd->busy_queues) {
514 cfq_log(cfqd, "schedule dispatch");
515 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
520 * Scale schedule slice based on io priority. Use the sync time slice only
521 * if a queue is marked sync and has sync io queued. A sync queue with async
522 * io only, should not get full sync slice length.
524 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
525 unsigned short prio)
527 const int base_slice = cfqd->cfq_slice[sync];
529 WARN_ON(prio >= IOPRIO_BE_NR);
531 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
534 static inline int
535 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
537 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
540 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
542 u64 d = delta << CFQ_SERVICE_SHIFT;
544 d = d * BLKIO_WEIGHT_DEFAULT;
545 do_div(d, cfqg->weight);
546 return d;
549 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
551 s64 delta = (s64)(vdisktime - min_vdisktime);
552 if (delta > 0)
553 min_vdisktime = vdisktime;
555 return min_vdisktime;
558 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
560 s64 delta = (s64)(vdisktime - min_vdisktime);
561 if (delta < 0)
562 min_vdisktime = vdisktime;
564 return min_vdisktime;
567 static void update_min_vdisktime(struct cfq_rb_root *st)
569 struct cfq_group *cfqg;
571 if (st->left) {
572 cfqg = rb_entry_cfqg(st->left);
573 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
574 cfqg->vdisktime);
579 * get averaged number of queues of RT/BE priority.
580 * average is updated, with a formula that gives more weight to higher numbers,
581 * to quickly follows sudden increases and decrease slowly
584 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
585 struct cfq_group *cfqg, bool rt)
587 unsigned min_q, max_q;
588 unsigned mult = cfq_hist_divisor - 1;
589 unsigned round = cfq_hist_divisor / 2;
590 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
592 min_q = min(cfqg->busy_queues_avg[rt], busy);
593 max_q = max(cfqg->busy_queues_avg[rt], busy);
594 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
595 cfq_hist_divisor;
596 return cfqg->busy_queues_avg[rt];
599 static inline unsigned
600 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
602 struct cfq_rb_root *st = &cfqd->grp_service_tree;
604 return cfq_target_latency * cfqg->weight / st->total_weight;
607 static inline unsigned
608 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
610 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
611 if (cfqd->cfq_latency) {
613 * interested queues (we consider only the ones with the same
614 * priority class in the cfq group)
616 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
617 cfq_class_rt(cfqq));
618 unsigned sync_slice = cfqd->cfq_slice[1];
619 unsigned expect_latency = sync_slice * iq;
620 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
622 if (expect_latency > group_slice) {
623 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
624 /* scale low_slice according to IO priority
625 * and sync vs async */
626 unsigned low_slice =
627 min(slice, base_low_slice * slice / sync_slice);
628 /* the adapted slice value is scaled to fit all iqs
629 * into the target latency */
630 slice = max(slice * group_slice / expect_latency,
631 low_slice);
634 return slice;
637 static inline void
638 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
640 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
642 cfqq->slice_start = jiffies;
643 cfqq->slice_end = jiffies + slice;
644 cfqq->allocated_slice = slice;
645 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
649 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
650 * isn't valid until the first request from the dispatch is activated
651 * and the slice time set.
653 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
655 if (cfq_cfqq_slice_new(cfqq))
656 return false;
657 if (time_before(jiffies, cfqq->slice_end))
658 return false;
660 return true;
664 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
665 * We choose the request that is closest to the head right now. Distance
666 * behind the head is penalized and only allowed to a certain extent.
668 static struct request *
669 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
671 sector_t s1, s2, d1 = 0, d2 = 0;
672 unsigned long back_max;
673 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
674 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
675 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
677 if (rq1 == NULL || rq1 == rq2)
678 return rq2;
679 if (rq2 == NULL)
680 return rq1;
682 if (rq_is_sync(rq1) != rq_is_sync(rq2))
683 return rq_is_sync(rq1) ? rq1 : rq2;
685 s1 = blk_rq_pos(rq1);
686 s2 = blk_rq_pos(rq2);
689 * by definition, 1KiB is 2 sectors
691 back_max = cfqd->cfq_back_max * 2;
694 * Strict one way elevator _except_ in the case where we allow
695 * short backward seeks which are biased as twice the cost of a
696 * similar forward seek.
698 if (s1 >= last)
699 d1 = s1 - last;
700 else if (s1 + back_max >= last)
701 d1 = (last - s1) * cfqd->cfq_back_penalty;
702 else
703 wrap |= CFQ_RQ1_WRAP;
705 if (s2 >= last)
706 d2 = s2 - last;
707 else if (s2 + back_max >= last)
708 d2 = (last - s2) * cfqd->cfq_back_penalty;
709 else
710 wrap |= CFQ_RQ2_WRAP;
712 /* Found required data */
715 * By doing switch() on the bit mask "wrap" we avoid having to
716 * check two variables for all permutations: --> faster!
718 switch (wrap) {
719 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
720 if (d1 < d2)
721 return rq1;
722 else if (d2 < d1)
723 return rq2;
724 else {
725 if (s1 >= s2)
726 return rq1;
727 else
728 return rq2;
731 case CFQ_RQ2_WRAP:
732 return rq1;
733 case CFQ_RQ1_WRAP:
734 return rq2;
735 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
736 default:
738 * Since both rqs are wrapped,
739 * start with the one that's further behind head
740 * (--> only *one* back seek required),
741 * since back seek takes more time than forward.
743 if (s1 <= s2)
744 return rq1;
745 else
746 return rq2;
751 * The below is leftmost cache rbtree addon
753 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
755 /* Service tree is empty */
756 if (!root->count)
757 return NULL;
759 if (!root->left)
760 root->left = rb_first(&root->rb);
762 if (root->left)
763 return rb_entry(root->left, struct cfq_queue, rb_node);
765 return NULL;
768 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
770 if (!root->left)
771 root->left = rb_first(&root->rb);
773 if (root->left)
774 return rb_entry_cfqg(root->left);
776 return NULL;
779 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
781 rb_erase(n, root);
782 RB_CLEAR_NODE(n);
785 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
787 if (root->left == n)
788 root->left = NULL;
789 rb_erase_init(n, &root->rb);
790 --root->count;
794 * would be nice to take fifo expire time into account as well
796 static struct request *
797 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
798 struct request *last)
800 struct rb_node *rbnext = rb_next(&last->rb_node);
801 struct rb_node *rbprev = rb_prev(&last->rb_node);
802 struct request *next = NULL, *prev = NULL;
804 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
806 if (rbprev)
807 prev = rb_entry_rq(rbprev);
809 if (rbnext)
810 next = rb_entry_rq(rbnext);
811 else {
812 rbnext = rb_first(&cfqq->sort_list);
813 if (rbnext && rbnext != &last->rb_node)
814 next = rb_entry_rq(rbnext);
817 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
820 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
821 struct cfq_queue *cfqq)
824 * just an approximation, should be ok.
826 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
827 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
830 static inline s64
831 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
833 return cfqg->vdisktime - st->min_vdisktime;
836 static void
837 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
839 struct rb_node **node = &st->rb.rb_node;
840 struct rb_node *parent = NULL;
841 struct cfq_group *__cfqg;
842 s64 key = cfqg_key(st, cfqg);
843 int left = 1;
845 while (*node != NULL) {
846 parent = *node;
847 __cfqg = rb_entry_cfqg(parent);
849 if (key < cfqg_key(st, __cfqg))
850 node = &parent->rb_left;
851 else {
852 node = &parent->rb_right;
853 left = 0;
857 if (left)
858 st->left = &cfqg->rb_node;
860 rb_link_node(&cfqg->rb_node, parent, node);
861 rb_insert_color(&cfqg->rb_node, &st->rb);
864 static void
865 cfq_update_group_weight(struct cfq_group *cfqg)
867 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
868 if (cfqg->needs_update) {
869 cfqg->weight = cfqg->new_weight;
870 cfqg->needs_update = false;
874 static void
875 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
877 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
879 cfq_update_group_weight(cfqg);
880 __cfq_group_service_tree_add(st, cfqg);
881 st->total_weight += cfqg->weight;
884 static void
885 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
887 struct cfq_rb_root *st = &cfqd->grp_service_tree;
888 struct cfq_group *__cfqg;
889 struct rb_node *n;
891 cfqg->nr_cfqq++;
892 if (!RB_EMPTY_NODE(&cfqg->rb_node))
893 return;
896 * Currently put the group at the end. Later implement something
897 * so that groups get lesser vtime based on their weights, so that
898 * if group does not loose all if it was not continuously backlogged.
900 n = rb_last(&st->rb);
901 if (n) {
902 __cfqg = rb_entry_cfqg(n);
903 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
904 } else
905 cfqg->vdisktime = st->min_vdisktime;
906 cfq_group_service_tree_add(st, cfqg);
909 static void
910 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
912 st->total_weight -= cfqg->weight;
913 if (!RB_EMPTY_NODE(&cfqg->rb_node))
914 cfq_rb_erase(&cfqg->rb_node, st);
917 static void
918 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
920 struct cfq_rb_root *st = &cfqd->grp_service_tree;
922 BUG_ON(cfqg->nr_cfqq < 1);
923 cfqg->nr_cfqq--;
925 /* If there are other cfq queues under this group, don't delete it */
926 if (cfqg->nr_cfqq)
927 return;
929 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
930 cfq_group_service_tree_del(st, cfqg);
931 cfqg->saved_workload_slice = 0;
932 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
935 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
936 unsigned int *unaccounted_time)
938 unsigned int slice_used;
941 * Queue got expired before even a single request completed or
942 * got expired immediately after first request completion.
944 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
946 * Also charge the seek time incurred to the group, otherwise
947 * if there are mutiple queues in the group, each can dispatch
948 * a single request on seeky media and cause lots of seek time
949 * and group will never know it.
951 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
953 } else {
954 slice_used = jiffies - cfqq->slice_start;
955 if (slice_used > cfqq->allocated_slice) {
956 *unaccounted_time = slice_used - cfqq->allocated_slice;
957 slice_used = cfqq->allocated_slice;
959 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
960 *unaccounted_time += cfqq->slice_start -
961 cfqq->dispatch_start;
964 return slice_used;
967 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
968 struct cfq_queue *cfqq)
970 struct cfq_rb_root *st = &cfqd->grp_service_tree;
971 unsigned int used_sl, charge, unaccounted_sl = 0;
972 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
973 - cfqg->service_tree_idle.count;
975 BUG_ON(nr_sync < 0);
976 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
978 if (iops_mode(cfqd))
979 charge = cfqq->slice_dispatch;
980 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
981 charge = cfqq->allocated_slice;
983 /* Can't update vdisktime while group is on service tree */
984 cfq_group_service_tree_del(st, cfqg);
985 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
986 /* If a new weight was requested, update now, off tree */
987 cfq_group_service_tree_add(st, cfqg);
989 /* This group is being expired. Save the context */
990 if (time_after(cfqd->workload_expires, jiffies)) {
991 cfqg->saved_workload_slice = cfqd->workload_expires
992 - jiffies;
993 cfqg->saved_workload = cfqd->serving_type;
994 cfqg->saved_serving_prio = cfqd->serving_prio;
995 } else
996 cfqg->saved_workload_slice = 0;
998 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
999 st->min_vdisktime);
1000 cfq_log_cfqq(cfqq->cfqd, cfqq,
1001 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1002 used_sl, cfqq->slice_dispatch, charge,
1003 iops_mode(cfqd), cfqq->nr_sectors);
1004 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1005 unaccounted_sl);
1006 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1009 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1010 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1012 if (blkg)
1013 return container_of(blkg, struct cfq_group, blkg);
1014 return NULL;
1017 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1018 unsigned int weight)
1020 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1021 cfqg->new_weight = weight;
1022 cfqg->needs_update = true;
1025 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1026 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1028 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1029 unsigned int major, minor;
1032 * Add group onto cgroup list. It might happen that bdi->dev is
1033 * not initialized yet. Initialize this new group without major
1034 * and minor info and this info will be filled in once a new thread
1035 * comes for IO.
1037 if (bdi->dev) {
1038 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1039 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1040 (void *)cfqd, MKDEV(major, minor));
1041 } else
1042 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1043 (void *)cfqd, 0);
1045 cfqd->nr_blkcg_linked_grps++;
1046 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1048 /* Add group on cfqd list */
1049 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1053 * Should be called from sleepable context. No request queue lock as per
1054 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1055 * from sleepable context.
1057 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1059 struct cfq_group *cfqg = NULL;
1060 int i, j, ret;
1061 struct cfq_rb_root *st;
1063 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1064 if (!cfqg)
1065 return NULL;
1067 for_each_cfqg_st(cfqg, i, j, st)
1068 *st = CFQ_RB_ROOT;
1069 RB_CLEAR_NODE(&cfqg->rb_node);
1071 cfqg->ttime.last_end_request = jiffies;
1074 * Take the initial reference that will be released on destroy
1075 * This can be thought of a joint reference by cgroup and
1076 * elevator which will be dropped by either elevator exit
1077 * or cgroup deletion path depending on who is exiting first.
1079 cfqg->ref = 1;
1081 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1082 if (ret) {
1083 kfree(cfqg);
1084 return NULL;
1087 return cfqg;
1090 static struct cfq_group *
1091 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1093 struct cfq_group *cfqg = NULL;
1094 void *key = cfqd;
1095 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1096 unsigned int major, minor;
1099 * This is the common case when there are no blkio cgroups.
1100 * Avoid lookup in this case
1102 if (blkcg == &blkio_root_cgroup)
1103 cfqg = &cfqd->root_group;
1104 else
1105 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1107 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1108 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1109 cfqg->blkg.dev = MKDEV(major, minor);
1112 return cfqg;
1116 * Search for the cfq group current task belongs to. request_queue lock must
1117 * be held.
1119 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1121 struct blkio_cgroup *blkcg;
1122 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1123 struct request_queue *q = cfqd->queue;
1125 rcu_read_lock();
1126 blkcg = task_blkio_cgroup(current);
1127 cfqg = cfq_find_cfqg(cfqd, blkcg);
1128 if (cfqg) {
1129 rcu_read_unlock();
1130 return cfqg;
1134 * Need to allocate a group. Allocation of group also needs allocation
1135 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1136 * we need to drop rcu lock and queue_lock before we call alloc.
1138 * Not taking any queue reference here and assuming that queue is
1139 * around by the time we return. CFQ queue allocation code does
1140 * the same. It might be racy though.
1143 rcu_read_unlock();
1144 spin_unlock_irq(q->queue_lock);
1146 cfqg = cfq_alloc_cfqg(cfqd);
1148 spin_lock_irq(q->queue_lock);
1150 rcu_read_lock();
1151 blkcg = task_blkio_cgroup(current);
1154 * If some other thread already allocated the group while we were
1155 * not holding queue lock, free up the group
1157 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1159 if (__cfqg) {
1160 kfree(cfqg);
1161 rcu_read_unlock();
1162 return __cfqg;
1165 if (!cfqg)
1166 cfqg = &cfqd->root_group;
1168 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1169 rcu_read_unlock();
1170 return cfqg;
1173 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1175 cfqg->ref++;
1176 return cfqg;
1179 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1181 /* Currently, all async queues are mapped to root group */
1182 if (!cfq_cfqq_sync(cfqq))
1183 cfqg = &cfqq->cfqd->root_group;
1185 cfqq->cfqg = cfqg;
1186 /* cfqq reference on cfqg */
1187 cfqq->cfqg->ref++;
1190 static void cfq_put_cfqg(struct cfq_group *cfqg)
1192 struct cfq_rb_root *st;
1193 int i, j;
1195 BUG_ON(cfqg->ref <= 0);
1196 cfqg->ref--;
1197 if (cfqg->ref)
1198 return;
1199 for_each_cfqg_st(cfqg, i, j, st)
1200 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1201 free_percpu(cfqg->blkg.stats_cpu);
1202 kfree(cfqg);
1205 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1207 /* Something wrong if we are trying to remove same group twice */
1208 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1210 hlist_del_init(&cfqg->cfqd_node);
1213 * Put the reference taken at the time of creation so that when all
1214 * queues are gone, group can be destroyed.
1216 cfq_put_cfqg(cfqg);
1219 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1221 struct hlist_node *pos, *n;
1222 struct cfq_group *cfqg;
1224 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1226 * If cgroup removal path got to blk_group first and removed
1227 * it from cgroup list, then it will take care of destroying
1228 * cfqg also.
1230 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1231 cfq_destroy_cfqg(cfqd, cfqg);
1236 * Blk cgroup controller notification saying that blkio_group object is being
1237 * delinked as associated cgroup object is going away. That also means that
1238 * no new IO will come in this group. So get rid of this group as soon as
1239 * any pending IO in the group is finished.
1241 * This function is called under rcu_read_lock(). key is the rcu protected
1242 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1243 * read lock.
1245 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1246 * it should not be NULL as even if elevator was exiting, cgroup deltion
1247 * path got to it first.
1249 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1251 unsigned long flags;
1252 struct cfq_data *cfqd = key;
1254 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1255 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1256 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1259 #else /* GROUP_IOSCHED */
1260 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1262 return &cfqd->root_group;
1265 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1267 return cfqg;
1270 static inline void
1271 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1272 cfqq->cfqg = cfqg;
1275 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1276 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1278 #endif /* GROUP_IOSCHED */
1281 * The cfqd->service_trees holds all pending cfq_queue's that have
1282 * requests waiting to be processed. It is sorted in the order that
1283 * we will service the queues.
1285 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1286 bool add_front)
1288 struct rb_node **p, *parent;
1289 struct cfq_queue *__cfqq;
1290 unsigned long rb_key;
1291 struct cfq_rb_root *service_tree;
1292 int left;
1293 int new_cfqq = 1;
1295 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1296 cfqq_type(cfqq));
1297 if (cfq_class_idle(cfqq)) {
1298 rb_key = CFQ_IDLE_DELAY;
1299 parent = rb_last(&service_tree->rb);
1300 if (parent && parent != &cfqq->rb_node) {
1301 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1302 rb_key += __cfqq->rb_key;
1303 } else
1304 rb_key += jiffies;
1305 } else if (!add_front) {
1307 * Get our rb key offset. Subtract any residual slice
1308 * value carried from last service. A negative resid
1309 * count indicates slice overrun, and this should position
1310 * the next service time further away in the tree.
1312 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1313 rb_key -= cfqq->slice_resid;
1314 cfqq->slice_resid = 0;
1315 } else {
1316 rb_key = -HZ;
1317 __cfqq = cfq_rb_first(service_tree);
1318 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1321 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1322 new_cfqq = 0;
1324 * same position, nothing more to do
1326 if (rb_key == cfqq->rb_key &&
1327 cfqq->service_tree == service_tree)
1328 return;
1330 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1331 cfqq->service_tree = NULL;
1334 left = 1;
1335 parent = NULL;
1336 cfqq->service_tree = service_tree;
1337 p = &service_tree->rb.rb_node;
1338 while (*p) {
1339 struct rb_node **n;
1341 parent = *p;
1342 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1345 * sort by key, that represents service time.
1347 if (time_before(rb_key, __cfqq->rb_key))
1348 n = &(*p)->rb_left;
1349 else {
1350 n = &(*p)->rb_right;
1351 left = 0;
1354 p = n;
1357 if (left)
1358 service_tree->left = &cfqq->rb_node;
1360 cfqq->rb_key = rb_key;
1361 rb_link_node(&cfqq->rb_node, parent, p);
1362 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1363 service_tree->count++;
1364 if (add_front || !new_cfqq)
1365 return;
1366 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1369 static struct cfq_queue *
1370 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1371 sector_t sector, struct rb_node **ret_parent,
1372 struct rb_node ***rb_link)
1374 struct rb_node **p, *parent;
1375 struct cfq_queue *cfqq = NULL;
1377 parent = NULL;
1378 p = &root->rb_node;
1379 while (*p) {
1380 struct rb_node **n;
1382 parent = *p;
1383 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1386 * Sort strictly based on sector. Smallest to the left,
1387 * largest to the right.
1389 if (sector > blk_rq_pos(cfqq->next_rq))
1390 n = &(*p)->rb_right;
1391 else if (sector < blk_rq_pos(cfqq->next_rq))
1392 n = &(*p)->rb_left;
1393 else
1394 break;
1395 p = n;
1396 cfqq = NULL;
1399 *ret_parent = parent;
1400 if (rb_link)
1401 *rb_link = p;
1402 return cfqq;
1405 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1407 struct rb_node **p, *parent;
1408 struct cfq_queue *__cfqq;
1410 if (cfqq->p_root) {
1411 rb_erase(&cfqq->p_node, cfqq->p_root);
1412 cfqq->p_root = NULL;
1415 if (cfq_class_idle(cfqq))
1416 return;
1417 if (!cfqq->next_rq)
1418 return;
1420 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1421 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1422 blk_rq_pos(cfqq->next_rq), &parent, &p);
1423 if (!__cfqq) {
1424 rb_link_node(&cfqq->p_node, parent, p);
1425 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1426 } else
1427 cfqq->p_root = NULL;
1431 * Update cfqq's position in the service tree.
1433 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1436 * Resorting requires the cfqq to be on the RR list already.
1438 if (cfq_cfqq_on_rr(cfqq)) {
1439 cfq_service_tree_add(cfqd, cfqq, 0);
1440 cfq_prio_tree_add(cfqd, cfqq);
1445 * add to busy list of queues for service, trying to be fair in ordering
1446 * the pending list according to last request service
1448 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1450 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1451 BUG_ON(cfq_cfqq_on_rr(cfqq));
1452 cfq_mark_cfqq_on_rr(cfqq);
1453 cfqd->busy_queues++;
1454 if (cfq_cfqq_sync(cfqq))
1455 cfqd->busy_sync_queues++;
1457 cfq_resort_rr_list(cfqd, cfqq);
1461 * Called when the cfqq no longer has requests pending, remove it from
1462 * the service tree.
1464 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1466 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1467 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1468 cfq_clear_cfqq_on_rr(cfqq);
1470 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1471 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1472 cfqq->service_tree = NULL;
1474 if (cfqq->p_root) {
1475 rb_erase(&cfqq->p_node, cfqq->p_root);
1476 cfqq->p_root = NULL;
1479 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1480 BUG_ON(!cfqd->busy_queues);
1481 cfqd->busy_queues--;
1482 if (cfq_cfqq_sync(cfqq))
1483 cfqd->busy_sync_queues--;
1487 * rb tree support functions
1489 static void cfq_del_rq_rb(struct request *rq)
1491 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1492 const int sync = rq_is_sync(rq);
1494 BUG_ON(!cfqq->queued[sync]);
1495 cfqq->queued[sync]--;
1497 elv_rb_del(&cfqq->sort_list, rq);
1499 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1501 * Queue will be deleted from service tree when we actually
1502 * expire it later. Right now just remove it from prio tree
1503 * as it is empty.
1505 if (cfqq->p_root) {
1506 rb_erase(&cfqq->p_node, cfqq->p_root);
1507 cfqq->p_root = NULL;
1512 static void cfq_add_rq_rb(struct request *rq)
1514 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1515 struct cfq_data *cfqd = cfqq->cfqd;
1516 struct request *prev;
1518 cfqq->queued[rq_is_sync(rq)]++;
1520 elv_rb_add(&cfqq->sort_list, rq);
1522 if (!cfq_cfqq_on_rr(cfqq))
1523 cfq_add_cfqq_rr(cfqd, cfqq);
1526 * check if this request is a better next-serve candidate
1528 prev = cfqq->next_rq;
1529 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1532 * adjust priority tree position, if ->next_rq changes
1534 if (prev != cfqq->next_rq)
1535 cfq_prio_tree_add(cfqd, cfqq);
1537 BUG_ON(!cfqq->next_rq);
1540 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1542 elv_rb_del(&cfqq->sort_list, rq);
1543 cfqq->queued[rq_is_sync(rq)]--;
1544 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1545 rq_data_dir(rq), rq_is_sync(rq));
1546 cfq_add_rq_rb(rq);
1547 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1548 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1549 rq_is_sync(rq));
1552 static struct request *
1553 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1555 struct task_struct *tsk = current;
1556 struct cfq_io_context *cic;
1557 struct cfq_queue *cfqq;
1559 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1560 if (!cic)
1561 return NULL;
1563 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1564 if (cfqq) {
1565 sector_t sector = bio->bi_sector + bio_sectors(bio);
1567 return elv_rb_find(&cfqq->sort_list, sector);
1570 return NULL;
1573 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1575 struct cfq_data *cfqd = q->elevator->elevator_data;
1577 cfqd->rq_in_driver++;
1578 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1579 cfqd->rq_in_driver);
1581 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1584 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1586 struct cfq_data *cfqd = q->elevator->elevator_data;
1588 WARN_ON(!cfqd->rq_in_driver);
1589 cfqd->rq_in_driver--;
1590 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1591 cfqd->rq_in_driver);
1594 static void cfq_remove_request(struct request *rq)
1596 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1598 if (cfqq->next_rq == rq)
1599 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1601 list_del_init(&rq->queuelist);
1602 cfq_del_rq_rb(rq);
1604 cfqq->cfqd->rq_queued--;
1605 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1606 rq_data_dir(rq), rq_is_sync(rq));
1609 static int cfq_merge(struct request_queue *q, struct request **req,
1610 struct bio *bio)
1612 struct cfq_data *cfqd = q->elevator->elevator_data;
1613 struct request *__rq;
1615 __rq = cfq_find_rq_fmerge(cfqd, bio);
1616 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1617 *req = __rq;
1618 return ELEVATOR_FRONT_MERGE;
1621 return ELEVATOR_NO_MERGE;
1624 static void cfq_merged_request(struct request_queue *q, struct request *req,
1625 int type)
1627 if (type == ELEVATOR_FRONT_MERGE) {
1628 struct cfq_queue *cfqq = RQ_CFQQ(req);
1630 cfq_reposition_rq_rb(cfqq, req);
1634 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1635 struct bio *bio)
1637 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1638 bio_data_dir(bio), cfq_bio_sync(bio));
1641 static void
1642 cfq_merged_requests(struct request_queue *q, struct request *rq,
1643 struct request *next)
1645 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1647 * reposition in fifo if next is older than rq
1649 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1650 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1651 list_move(&rq->queuelist, &next->queuelist);
1652 rq_set_fifo_time(rq, rq_fifo_time(next));
1655 if (cfqq->next_rq == next)
1656 cfqq->next_rq = rq;
1657 cfq_remove_request(next);
1658 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1659 rq_data_dir(next), rq_is_sync(next));
1662 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1663 struct bio *bio)
1665 struct cfq_data *cfqd = q->elevator->elevator_data;
1666 struct cfq_io_context *cic;
1667 struct cfq_queue *cfqq;
1670 * Disallow merge of a sync bio into an async request.
1672 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1673 return false;
1676 * Lookup the cfqq that this bio will be queued with. Allow
1677 * merge only if rq is queued there.
1679 cic = cfq_cic_lookup(cfqd, current->io_context);
1680 if (!cic)
1681 return false;
1683 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1684 return cfqq == RQ_CFQQ(rq);
1687 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1689 del_timer(&cfqd->idle_slice_timer);
1690 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1693 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1694 struct cfq_queue *cfqq)
1696 if (cfqq) {
1697 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1698 cfqd->serving_prio, cfqd->serving_type);
1699 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1700 cfqq->slice_start = 0;
1701 cfqq->dispatch_start = jiffies;
1702 cfqq->allocated_slice = 0;
1703 cfqq->slice_end = 0;
1704 cfqq->slice_dispatch = 0;
1705 cfqq->nr_sectors = 0;
1707 cfq_clear_cfqq_wait_request(cfqq);
1708 cfq_clear_cfqq_must_dispatch(cfqq);
1709 cfq_clear_cfqq_must_alloc_slice(cfqq);
1710 cfq_clear_cfqq_fifo_expire(cfqq);
1711 cfq_mark_cfqq_slice_new(cfqq);
1713 cfq_del_timer(cfqd, cfqq);
1716 cfqd->active_queue = cfqq;
1720 * current cfqq expired its slice (or was too idle), select new one
1722 static void
1723 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1724 bool timed_out)
1726 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1728 if (cfq_cfqq_wait_request(cfqq))
1729 cfq_del_timer(cfqd, cfqq);
1731 cfq_clear_cfqq_wait_request(cfqq);
1732 cfq_clear_cfqq_wait_busy(cfqq);
1735 * If this cfqq is shared between multiple processes, check to
1736 * make sure that those processes are still issuing I/Os within
1737 * the mean seek distance. If not, it may be time to break the
1738 * queues apart again.
1740 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1741 cfq_mark_cfqq_split_coop(cfqq);
1744 * store what was left of this slice, if the queue idled/timed out
1746 if (timed_out) {
1747 if (cfq_cfqq_slice_new(cfqq))
1748 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1749 else
1750 cfqq->slice_resid = cfqq->slice_end - jiffies;
1751 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1754 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1756 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1757 cfq_del_cfqq_rr(cfqd, cfqq);
1759 cfq_resort_rr_list(cfqd, cfqq);
1761 if (cfqq == cfqd->active_queue)
1762 cfqd->active_queue = NULL;
1764 if (cfqd->active_cic) {
1765 put_io_context(cfqd->active_cic->ioc);
1766 cfqd->active_cic = NULL;
1770 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1772 struct cfq_queue *cfqq = cfqd->active_queue;
1774 if (cfqq)
1775 __cfq_slice_expired(cfqd, cfqq, timed_out);
1779 * Get next queue for service. Unless we have a queue preemption,
1780 * we'll simply select the first cfqq in the service tree.
1782 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1784 struct cfq_rb_root *service_tree =
1785 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1786 cfqd->serving_type);
1788 if (!cfqd->rq_queued)
1789 return NULL;
1791 /* There is nothing to dispatch */
1792 if (!service_tree)
1793 return NULL;
1794 if (RB_EMPTY_ROOT(&service_tree->rb))
1795 return NULL;
1796 return cfq_rb_first(service_tree);
1799 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1801 struct cfq_group *cfqg;
1802 struct cfq_queue *cfqq;
1803 int i, j;
1804 struct cfq_rb_root *st;
1806 if (!cfqd->rq_queued)
1807 return NULL;
1809 cfqg = cfq_get_next_cfqg(cfqd);
1810 if (!cfqg)
1811 return NULL;
1813 for_each_cfqg_st(cfqg, i, j, st)
1814 if ((cfqq = cfq_rb_first(st)) != NULL)
1815 return cfqq;
1816 return NULL;
1820 * Get and set a new active queue for service.
1822 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1823 struct cfq_queue *cfqq)
1825 if (!cfqq)
1826 cfqq = cfq_get_next_queue(cfqd);
1828 __cfq_set_active_queue(cfqd, cfqq);
1829 return cfqq;
1832 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1833 struct request *rq)
1835 if (blk_rq_pos(rq) >= cfqd->last_position)
1836 return blk_rq_pos(rq) - cfqd->last_position;
1837 else
1838 return cfqd->last_position - blk_rq_pos(rq);
1841 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1842 struct request *rq)
1844 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1847 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1848 struct cfq_queue *cur_cfqq)
1850 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1851 struct rb_node *parent, *node;
1852 struct cfq_queue *__cfqq;
1853 sector_t sector = cfqd->last_position;
1855 if (RB_EMPTY_ROOT(root))
1856 return NULL;
1859 * First, if we find a request starting at the end of the last
1860 * request, choose it.
1862 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1863 if (__cfqq)
1864 return __cfqq;
1867 * If the exact sector wasn't found, the parent of the NULL leaf
1868 * will contain the closest sector.
1870 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1871 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1872 return __cfqq;
1874 if (blk_rq_pos(__cfqq->next_rq) < sector)
1875 node = rb_next(&__cfqq->p_node);
1876 else
1877 node = rb_prev(&__cfqq->p_node);
1878 if (!node)
1879 return NULL;
1881 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1882 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1883 return __cfqq;
1885 return NULL;
1889 * cfqd - obvious
1890 * cur_cfqq - passed in so that we don't decide that the current queue is
1891 * closely cooperating with itself.
1893 * So, basically we're assuming that that cur_cfqq has dispatched at least
1894 * one request, and that cfqd->last_position reflects a position on the disk
1895 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1896 * assumption.
1898 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1899 struct cfq_queue *cur_cfqq)
1901 struct cfq_queue *cfqq;
1903 if (cfq_class_idle(cur_cfqq))
1904 return NULL;
1905 if (!cfq_cfqq_sync(cur_cfqq))
1906 return NULL;
1907 if (CFQQ_SEEKY(cur_cfqq))
1908 return NULL;
1911 * Don't search priority tree if it's the only queue in the group.
1913 if (cur_cfqq->cfqg->nr_cfqq == 1)
1914 return NULL;
1917 * We should notice if some of the queues are cooperating, eg
1918 * working closely on the same area of the disk. In that case,
1919 * we can group them together and don't waste time idling.
1921 cfqq = cfqq_close(cfqd, cur_cfqq);
1922 if (!cfqq)
1923 return NULL;
1925 /* If new queue belongs to different cfq_group, don't choose it */
1926 if (cur_cfqq->cfqg != cfqq->cfqg)
1927 return NULL;
1930 * It only makes sense to merge sync queues.
1932 if (!cfq_cfqq_sync(cfqq))
1933 return NULL;
1934 if (CFQQ_SEEKY(cfqq))
1935 return NULL;
1938 * Do not merge queues of different priority classes
1940 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1941 return NULL;
1943 return cfqq;
1947 * Determine whether we should enforce idle window for this queue.
1950 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1952 enum wl_prio_t prio = cfqq_prio(cfqq);
1953 struct cfq_rb_root *service_tree = cfqq->service_tree;
1955 BUG_ON(!service_tree);
1956 BUG_ON(!service_tree->count);
1958 if (!cfqd->cfq_slice_idle)
1959 return false;
1961 /* We never do for idle class queues. */
1962 if (prio == IDLE_WORKLOAD)
1963 return false;
1965 /* We do for queues that were marked with idle window flag. */
1966 if (cfq_cfqq_idle_window(cfqq) &&
1967 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1968 return true;
1971 * Otherwise, we do only if they are the last ones
1972 * in their service tree.
1974 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1975 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1976 return true;
1977 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1978 service_tree->count);
1979 return false;
1982 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1984 struct cfq_queue *cfqq = cfqd->active_queue;
1985 struct cfq_io_context *cic;
1986 unsigned long sl, group_idle = 0;
1989 * SSD device without seek penalty, disable idling. But only do so
1990 * for devices that support queuing, otherwise we still have a problem
1991 * with sync vs async workloads.
1993 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1994 return;
1996 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1997 WARN_ON(cfq_cfqq_slice_new(cfqq));
2000 * idle is disabled, either manually or by past process history
2002 if (!cfq_should_idle(cfqd, cfqq)) {
2003 /* no queue idling. Check for group idling */
2004 if (cfqd->cfq_group_idle)
2005 group_idle = cfqd->cfq_group_idle;
2006 else
2007 return;
2011 * still active requests from this queue, don't idle
2013 if (cfqq->dispatched)
2014 return;
2017 * task has exited, don't wait
2019 cic = cfqd->active_cic;
2020 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2021 return;
2024 * If our average think time is larger than the remaining time
2025 * slice, then don't idle. This avoids overrunning the allotted
2026 * time slice.
2028 if (sample_valid(cic->ttime.ttime_samples) &&
2029 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2030 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2031 cic->ttime.ttime_mean);
2032 return;
2035 /* There are other queues in the group, don't do group idle */
2036 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2037 return;
2039 cfq_mark_cfqq_wait_request(cfqq);
2041 if (group_idle)
2042 sl = cfqd->cfq_group_idle;
2043 else
2044 sl = cfqd->cfq_slice_idle;
2046 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2047 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2048 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2049 group_idle ? 1 : 0);
2053 * Move request from internal lists to the request queue dispatch list.
2055 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2057 struct cfq_data *cfqd = q->elevator->elevator_data;
2058 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2060 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2062 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2063 cfq_remove_request(rq);
2064 cfqq->dispatched++;
2065 (RQ_CFQG(rq))->dispatched++;
2066 elv_dispatch_sort(q, rq);
2068 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2069 cfqq->nr_sectors += blk_rq_sectors(rq);
2070 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2071 rq_data_dir(rq), rq_is_sync(rq));
2075 * return expired entry, or NULL to just start from scratch in rbtree
2077 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2079 struct request *rq = NULL;
2081 if (cfq_cfqq_fifo_expire(cfqq))
2082 return NULL;
2084 cfq_mark_cfqq_fifo_expire(cfqq);
2086 if (list_empty(&cfqq->fifo))
2087 return NULL;
2089 rq = rq_entry_fifo(cfqq->fifo.next);
2090 if (time_before(jiffies, rq_fifo_time(rq)))
2091 rq = NULL;
2093 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2094 return rq;
2097 static inline int
2098 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2100 const int base_rq = cfqd->cfq_slice_async_rq;
2102 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2104 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2108 * Must be called with the queue_lock held.
2110 static int cfqq_process_refs(struct cfq_queue *cfqq)
2112 int process_refs, io_refs;
2114 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2115 process_refs = cfqq->ref - io_refs;
2116 BUG_ON(process_refs < 0);
2117 return process_refs;
2120 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2122 int process_refs, new_process_refs;
2123 struct cfq_queue *__cfqq;
2126 * If there are no process references on the new_cfqq, then it is
2127 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2128 * chain may have dropped their last reference (not just their
2129 * last process reference).
2131 if (!cfqq_process_refs(new_cfqq))
2132 return;
2134 /* Avoid a circular list and skip interim queue merges */
2135 while ((__cfqq = new_cfqq->new_cfqq)) {
2136 if (__cfqq == cfqq)
2137 return;
2138 new_cfqq = __cfqq;
2141 process_refs = cfqq_process_refs(cfqq);
2142 new_process_refs = cfqq_process_refs(new_cfqq);
2144 * If the process for the cfqq has gone away, there is no
2145 * sense in merging the queues.
2147 if (process_refs == 0 || new_process_refs == 0)
2148 return;
2151 * Merge in the direction of the lesser amount of work.
2153 if (new_process_refs >= process_refs) {
2154 cfqq->new_cfqq = new_cfqq;
2155 new_cfqq->ref += process_refs;
2156 } else {
2157 new_cfqq->new_cfqq = cfqq;
2158 cfqq->ref += new_process_refs;
2162 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2163 struct cfq_group *cfqg, enum wl_prio_t prio)
2165 struct cfq_queue *queue;
2166 int i;
2167 bool key_valid = false;
2168 unsigned long lowest_key = 0;
2169 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2171 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2172 /* select the one with lowest rb_key */
2173 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2174 if (queue &&
2175 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2176 lowest_key = queue->rb_key;
2177 cur_best = i;
2178 key_valid = true;
2182 return cur_best;
2185 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2187 unsigned slice;
2188 unsigned count;
2189 struct cfq_rb_root *st;
2190 unsigned group_slice;
2191 enum wl_prio_t original_prio = cfqd->serving_prio;
2193 /* Choose next priority. RT > BE > IDLE */
2194 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2195 cfqd->serving_prio = RT_WORKLOAD;
2196 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2197 cfqd->serving_prio = BE_WORKLOAD;
2198 else {
2199 cfqd->serving_prio = IDLE_WORKLOAD;
2200 cfqd->workload_expires = jiffies + 1;
2201 return;
2204 if (original_prio != cfqd->serving_prio)
2205 goto new_workload;
2208 * For RT and BE, we have to choose also the type
2209 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2210 * expiration time
2212 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2213 count = st->count;
2216 * check workload expiration, and that we still have other queues ready
2218 if (count && !time_after(jiffies, cfqd->workload_expires))
2219 return;
2221 new_workload:
2222 /* otherwise select new workload type */
2223 cfqd->serving_type =
2224 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2225 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2226 count = st->count;
2229 * the workload slice is computed as a fraction of target latency
2230 * proportional to the number of queues in that workload, over
2231 * all the queues in the same priority class
2233 group_slice = cfq_group_slice(cfqd, cfqg);
2235 slice = group_slice * count /
2236 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2237 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2239 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2240 unsigned int tmp;
2243 * Async queues are currently system wide. Just taking
2244 * proportion of queues with-in same group will lead to higher
2245 * async ratio system wide as generally root group is going
2246 * to have higher weight. A more accurate thing would be to
2247 * calculate system wide asnc/sync ratio.
2249 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2250 tmp = tmp/cfqd->busy_queues;
2251 slice = min_t(unsigned, slice, tmp);
2253 /* async workload slice is scaled down according to
2254 * the sync/async slice ratio. */
2255 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2256 } else
2257 /* sync workload slice is at least 2 * cfq_slice_idle */
2258 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2260 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2261 cfq_log(cfqd, "workload slice:%d", slice);
2262 cfqd->workload_expires = jiffies + slice;
2265 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2267 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2268 struct cfq_group *cfqg;
2270 if (RB_EMPTY_ROOT(&st->rb))
2271 return NULL;
2272 cfqg = cfq_rb_first_group(st);
2273 update_min_vdisktime(st);
2274 return cfqg;
2277 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2279 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2281 cfqd->serving_group = cfqg;
2283 /* Restore the workload type data */
2284 if (cfqg->saved_workload_slice) {
2285 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2286 cfqd->serving_type = cfqg->saved_workload;
2287 cfqd->serving_prio = cfqg->saved_serving_prio;
2288 } else
2289 cfqd->workload_expires = jiffies - 1;
2291 choose_service_tree(cfqd, cfqg);
2295 * Select a queue for service. If we have a current active queue,
2296 * check whether to continue servicing it, or retrieve and set a new one.
2298 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2300 struct cfq_queue *cfqq, *new_cfqq = NULL;
2302 cfqq = cfqd->active_queue;
2303 if (!cfqq)
2304 goto new_queue;
2306 if (!cfqd->rq_queued)
2307 return NULL;
2310 * We were waiting for group to get backlogged. Expire the queue
2312 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2313 goto expire;
2316 * The active queue has run out of time, expire it and select new.
2318 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2320 * If slice had not expired at the completion of last request
2321 * we might not have turned on wait_busy flag. Don't expire
2322 * the queue yet. Allow the group to get backlogged.
2324 * The very fact that we have used the slice, that means we
2325 * have been idling all along on this queue and it should be
2326 * ok to wait for this request to complete.
2328 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2329 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2330 cfqq = NULL;
2331 goto keep_queue;
2332 } else
2333 goto check_group_idle;
2337 * The active queue has requests and isn't expired, allow it to
2338 * dispatch.
2340 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2341 goto keep_queue;
2344 * If another queue has a request waiting within our mean seek
2345 * distance, let it run. The expire code will check for close
2346 * cooperators and put the close queue at the front of the service
2347 * tree. If possible, merge the expiring queue with the new cfqq.
2349 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2350 if (new_cfqq) {
2351 if (!cfqq->new_cfqq)
2352 cfq_setup_merge(cfqq, new_cfqq);
2353 goto expire;
2357 * No requests pending. If the active queue still has requests in
2358 * flight or is idling for a new request, allow either of these
2359 * conditions to happen (or time out) before selecting a new queue.
2361 if (timer_pending(&cfqd->idle_slice_timer)) {
2362 cfqq = NULL;
2363 goto keep_queue;
2367 * This is a deep seek queue, but the device is much faster than
2368 * the queue can deliver, don't idle
2370 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2371 (cfq_cfqq_slice_new(cfqq) ||
2372 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2373 cfq_clear_cfqq_deep(cfqq);
2374 cfq_clear_cfqq_idle_window(cfqq);
2377 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2378 cfqq = NULL;
2379 goto keep_queue;
2383 * If group idle is enabled and there are requests dispatched from
2384 * this group, wait for requests to complete.
2386 check_group_idle:
2387 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2388 cfqq->cfqg->dispatched &&
2389 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2390 cfqq = NULL;
2391 goto keep_queue;
2394 expire:
2395 cfq_slice_expired(cfqd, 0);
2396 new_queue:
2398 * Current queue expired. Check if we have to switch to a new
2399 * service tree
2401 if (!new_cfqq)
2402 cfq_choose_cfqg(cfqd);
2404 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2405 keep_queue:
2406 return cfqq;
2409 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2411 int dispatched = 0;
2413 while (cfqq->next_rq) {
2414 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2415 dispatched++;
2418 BUG_ON(!list_empty(&cfqq->fifo));
2420 /* By default cfqq is not expired if it is empty. Do it explicitly */
2421 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2422 return dispatched;
2426 * Drain our current requests. Used for barriers and when switching
2427 * io schedulers on-the-fly.
2429 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2431 struct cfq_queue *cfqq;
2432 int dispatched = 0;
2434 /* Expire the timeslice of the current active queue first */
2435 cfq_slice_expired(cfqd, 0);
2436 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2437 __cfq_set_active_queue(cfqd, cfqq);
2438 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2441 BUG_ON(cfqd->busy_queues);
2443 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2444 return dispatched;
2447 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2448 struct cfq_queue *cfqq)
2450 /* the queue hasn't finished any request, can't estimate */
2451 if (cfq_cfqq_slice_new(cfqq))
2452 return true;
2453 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2454 cfqq->slice_end))
2455 return true;
2457 return false;
2460 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2462 unsigned int max_dispatch;
2465 * Drain async requests before we start sync IO
2467 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2468 return false;
2471 * If this is an async queue and we have sync IO in flight, let it wait
2473 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2474 return false;
2476 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2477 if (cfq_class_idle(cfqq))
2478 max_dispatch = 1;
2481 * Does this cfqq already have too much IO in flight?
2483 if (cfqq->dispatched >= max_dispatch) {
2484 bool promote_sync = false;
2486 * idle queue must always only have a single IO in flight
2488 if (cfq_class_idle(cfqq))
2489 return false;
2492 * If there is only one sync queue
2493 * we can ignore async queue here and give the sync
2494 * queue no dispatch limit. The reason is a sync queue can
2495 * preempt async queue, limiting the sync queue doesn't make
2496 * sense. This is useful for aiostress test.
2498 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2499 promote_sync = true;
2502 * We have other queues, don't allow more IO from this one
2504 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2505 !promote_sync)
2506 return false;
2509 * Sole queue user, no limit
2511 if (cfqd->busy_queues == 1 || promote_sync)
2512 max_dispatch = -1;
2513 else
2515 * Normally we start throttling cfqq when cfq_quantum/2
2516 * requests have been dispatched. But we can drive
2517 * deeper queue depths at the beginning of slice
2518 * subjected to upper limit of cfq_quantum.
2519 * */
2520 max_dispatch = cfqd->cfq_quantum;
2524 * Async queues must wait a bit before being allowed dispatch.
2525 * We also ramp up the dispatch depth gradually for async IO,
2526 * based on the last sync IO we serviced
2528 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2529 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2530 unsigned int depth;
2532 depth = last_sync / cfqd->cfq_slice[1];
2533 if (!depth && !cfqq->dispatched)
2534 depth = 1;
2535 if (depth < max_dispatch)
2536 max_dispatch = depth;
2540 * If we're below the current max, allow a dispatch
2542 return cfqq->dispatched < max_dispatch;
2546 * Dispatch a request from cfqq, moving them to the request queue
2547 * dispatch list.
2549 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2551 struct request *rq;
2553 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2555 if (!cfq_may_dispatch(cfqd, cfqq))
2556 return false;
2559 * follow expired path, else get first next available
2561 rq = cfq_check_fifo(cfqq);
2562 if (!rq)
2563 rq = cfqq->next_rq;
2566 * insert request into driver dispatch list
2568 cfq_dispatch_insert(cfqd->queue, rq);
2570 if (!cfqd->active_cic) {
2571 struct cfq_io_context *cic = RQ_CIC(rq);
2573 atomic_long_inc(&cic->ioc->refcount);
2574 cfqd->active_cic = cic;
2577 return true;
2581 * Find the cfqq that we need to service and move a request from that to the
2582 * dispatch list
2584 static int cfq_dispatch_requests(struct request_queue *q, int force)
2586 struct cfq_data *cfqd = q->elevator->elevator_data;
2587 struct cfq_queue *cfqq;
2589 if (!cfqd->busy_queues)
2590 return 0;
2592 if (unlikely(force))
2593 return cfq_forced_dispatch(cfqd);
2595 cfqq = cfq_select_queue(cfqd);
2596 if (!cfqq)
2597 return 0;
2600 * Dispatch a request from this cfqq, if it is allowed
2602 if (!cfq_dispatch_request(cfqd, cfqq))
2603 return 0;
2605 cfqq->slice_dispatch++;
2606 cfq_clear_cfqq_must_dispatch(cfqq);
2609 * expire an async queue immediately if it has used up its slice. idle
2610 * queue always expire after 1 dispatch round.
2612 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2613 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2614 cfq_class_idle(cfqq))) {
2615 cfqq->slice_end = jiffies + 1;
2616 cfq_slice_expired(cfqd, 0);
2619 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2620 return 1;
2624 * task holds one reference to the queue, dropped when task exits. each rq
2625 * in-flight on this queue also holds a reference, dropped when rq is freed.
2627 * Each cfq queue took a reference on the parent group. Drop it now.
2628 * queue lock must be held here.
2630 static void cfq_put_queue(struct cfq_queue *cfqq)
2632 struct cfq_data *cfqd = cfqq->cfqd;
2633 struct cfq_group *cfqg;
2635 BUG_ON(cfqq->ref <= 0);
2637 cfqq->ref--;
2638 if (cfqq->ref)
2639 return;
2641 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2642 BUG_ON(rb_first(&cfqq->sort_list));
2643 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2644 cfqg = cfqq->cfqg;
2646 if (unlikely(cfqd->active_queue == cfqq)) {
2647 __cfq_slice_expired(cfqd, cfqq, 0);
2648 cfq_schedule_dispatch(cfqd);
2651 BUG_ON(cfq_cfqq_on_rr(cfqq));
2652 kmem_cache_free(cfq_pool, cfqq);
2653 cfq_put_cfqg(cfqg);
2657 * Call func for each cic attached to this ioc.
2659 static void
2660 call_for_each_cic(struct io_context *ioc,
2661 void (*func)(struct io_context *, struct cfq_io_context *))
2663 struct cfq_io_context *cic;
2664 struct hlist_node *n;
2666 rcu_read_lock();
2668 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2669 func(ioc, cic);
2671 rcu_read_unlock();
2674 static void cfq_cic_free_rcu(struct rcu_head *head)
2676 struct cfq_io_context *cic;
2678 cic = container_of(head, struct cfq_io_context, rcu_head);
2680 kmem_cache_free(cfq_ioc_pool, cic);
2681 elv_ioc_count_dec(cfq_ioc_count);
2683 if (ioc_gone) {
2685 * CFQ scheduler is exiting, grab exit lock and check
2686 * the pending io context count. If it hits zero,
2687 * complete ioc_gone and set it back to NULL
2689 spin_lock(&ioc_gone_lock);
2690 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2691 complete(ioc_gone);
2692 ioc_gone = NULL;
2694 spin_unlock(&ioc_gone_lock);
2698 static void cfq_cic_free(struct cfq_io_context *cic)
2700 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2703 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2705 unsigned long flags;
2706 unsigned long dead_key = (unsigned long) cic->key;
2708 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2710 spin_lock_irqsave(&ioc->lock, flags);
2711 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2712 hlist_del_rcu(&cic->cic_list);
2713 spin_unlock_irqrestore(&ioc->lock, flags);
2715 cfq_cic_free(cic);
2719 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2720 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2721 * and ->trim() which is called with the task lock held
2723 static void cfq_free_io_context(struct io_context *ioc)
2726 * ioc->refcount is zero here, or we are called from elv_unregister(),
2727 * so no more cic's are allowed to be linked into this ioc. So it
2728 * should be ok to iterate over the known list, we will see all cic's
2729 * since no new ones are added.
2731 call_for_each_cic(ioc, cic_free_func);
2734 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2736 struct cfq_queue *__cfqq, *next;
2739 * If this queue was scheduled to merge with another queue, be
2740 * sure to drop the reference taken on that queue (and others in
2741 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2743 __cfqq = cfqq->new_cfqq;
2744 while (__cfqq) {
2745 if (__cfqq == cfqq) {
2746 WARN(1, "cfqq->new_cfqq loop detected\n");
2747 break;
2749 next = __cfqq->new_cfqq;
2750 cfq_put_queue(__cfqq);
2751 __cfqq = next;
2755 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2757 if (unlikely(cfqq == cfqd->active_queue)) {
2758 __cfq_slice_expired(cfqd, cfqq, 0);
2759 cfq_schedule_dispatch(cfqd);
2762 cfq_put_cooperator(cfqq);
2764 cfq_put_queue(cfqq);
2767 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2768 struct cfq_io_context *cic)
2770 struct io_context *ioc = cic->ioc;
2772 list_del_init(&cic->queue_list);
2775 * Make sure dead mark is seen for dead queues
2777 smp_wmb();
2778 cic->key = cfqd_dead_key(cfqd);
2780 rcu_read_lock();
2781 if (rcu_dereference(ioc->ioc_data) == cic) {
2782 rcu_read_unlock();
2783 spin_lock(&ioc->lock);
2784 rcu_assign_pointer(ioc->ioc_data, NULL);
2785 spin_unlock(&ioc->lock);
2786 } else
2787 rcu_read_unlock();
2789 if (cic->cfqq[BLK_RW_ASYNC]) {
2790 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2791 cic->cfqq[BLK_RW_ASYNC] = NULL;
2794 if (cic->cfqq[BLK_RW_SYNC]) {
2795 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2796 cic->cfqq[BLK_RW_SYNC] = NULL;
2800 static void cfq_exit_single_io_context(struct io_context *ioc,
2801 struct cfq_io_context *cic)
2803 struct cfq_data *cfqd = cic_to_cfqd(cic);
2805 if (cfqd) {
2806 struct request_queue *q = cfqd->queue;
2807 unsigned long flags;
2809 spin_lock_irqsave(q->queue_lock, flags);
2812 * Ensure we get a fresh copy of the ->key to prevent
2813 * race between exiting task and queue
2815 smp_read_barrier_depends();
2816 if (cic->key == cfqd)
2817 __cfq_exit_single_io_context(cfqd, cic);
2819 spin_unlock_irqrestore(q->queue_lock, flags);
2824 * The process that ioc belongs to has exited, we need to clean up
2825 * and put the internal structures we have that belongs to that process.
2827 static void cfq_exit_io_context(struct io_context *ioc)
2829 call_for_each_cic(ioc, cfq_exit_single_io_context);
2832 static struct cfq_io_context *
2833 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2835 struct cfq_io_context *cic;
2837 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2838 cfqd->queue->node);
2839 if (cic) {
2840 cic->ttime.last_end_request = jiffies;
2841 INIT_LIST_HEAD(&cic->queue_list);
2842 INIT_HLIST_NODE(&cic->cic_list);
2843 cic->dtor = cfq_free_io_context;
2844 cic->exit = cfq_exit_io_context;
2845 elv_ioc_count_inc(cfq_ioc_count);
2848 return cic;
2851 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2853 struct task_struct *tsk = current;
2854 int ioprio_class;
2856 if (!cfq_cfqq_prio_changed(cfqq))
2857 return;
2859 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2860 switch (ioprio_class) {
2861 default:
2862 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2863 case IOPRIO_CLASS_NONE:
2865 * no prio set, inherit CPU scheduling settings
2867 cfqq->ioprio = task_nice_ioprio(tsk);
2868 cfqq->ioprio_class = task_nice_ioclass(tsk);
2869 break;
2870 case IOPRIO_CLASS_RT:
2871 cfqq->ioprio = task_ioprio(ioc);
2872 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2873 break;
2874 case IOPRIO_CLASS_BE:
2875 cfqq->ioprio = task_ioprio(ioc);
2876 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2877 break;
2878 case IOPRIO_CLASS_IDLE:
2879 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2880 cfqq->ioprio = 7;
2881 cfq_clear_cfqq_idle_window(cfqq);
2882 break;
2886 * keep track of original prio settings in case we have to temporarily
2887 * elevate the priority of this queue
2889 cfqq->org_ioprio = cfqq->ioprio;
2890 cfq_clear_cfqq_prio_changed(cfqq);
2893 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2895 struct cfq_data *cfqd = cic_to_cfqd(cic);
2896 struct cfq_queue *cfqq;
2897 unsigned long flags;
2899 if (unlikely(!cfqd))
2900 return;
2902 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2904 cfqq = cic->cfqq[BLK_RW_ASYNC];
2905 if (cfqq) {
2906 struct cfq_queue *new_cfqq;
2907 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2908 GFP_ATOMIC);
2909 if (new_cfqq) {
2910 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2911 cfq_put_queue(cfqq);
2915 cfqq = cic->cfqq[BLK_RW_SYNC];
2916 if (cfqq)
2917 cfq_mark_cfqq_prio_changed(cfqq);
2919 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2922 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2924 call_for_each_cic(ioc, changed_ioprio);
2925 ioc->ioprio_changed = 0;
2928 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2929 pid_t pid, bool is_sync)
2931 RB_CLEAR_NODE(&cfqq->rb_node);
2932 RB_CLEAR_NODE(&cfqq->p_node);
2933 INIT_LIST_HEAD(&cfqq->fifo);
2935 cfqq->ref = 0;
2936 cfqq->cfqd = cfqd;
2938 cfq_mark_cfqq_prio_changed(cfqq);
2940 if (is_sync) {
2941 if (!cfq_class_idle(cfqq))
2942 cfq_mark_cfqq_idle_window(cfqq);
2943 cfq_mark_cfqq_sync(cfqq);
2945 cfqq->pid = pid;
2948 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2949 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2951 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2952 struct cfq_data *cfqd = cic_to_cfqd(cic);
2953 unsigned long flags;
2954 struct request_queue *q;
2956 if (unlikely(!cfqd))
2957 return;
2959 q = cfqd->queue;
2961 spin_lock_irqsave(q->queue_lock, flags);
2963 if (sync_cfqq) {
2965 * Drop reference to sync queue. A new sync queue will be
2966 * assigned in new group upon arrival of a fresh request.
2968 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2969 cic_set_cfqq(cic, NULL, 1);
2970 cfq_put_queue(sync_cfqq);
2973 spin_unlock_irqrestore(q->queue_lock, flags);
2976 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2978 call_for_each_cic(ioc, changed_cgroup);
2979 ioc->cgroup_changed = 0;
2981 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2983 static struct cfq_queue *
2984 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2985 struct io_context *ioc, gfp_t gfp_mask)
2987 struct cfq_queue *cfqq, *new_cfqq = NULL;
2988 struct cfq_io_context *cic;
2989 struct cfq_group *cfqg;
2991 retry:
2992 cfqg = cfq_get_cfqg(cfqd);
2993 cic = cfq_cic_lookup(cfqd, ioc);
2994 /* cic always exists here */
2995 cfqq = cic_to_cfqq(cic, is_sync);
2998 * Always try a new alloc if we fell back to the OOM cfqq
2999 * originally, since it should just be a temporary situation.
3001 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3002 cfqq = NULL;
3003 if (new_cfqq) {
3004 cfqq = new_cfqq;
3005 new_cfqq = NULL;
3006 } else if (gfp_mask & __GFP_WAIT) {
3007 spin_unlock_irq(cfqd->queue->queue_lock);
3008 new_cfqq = kmem_cache_alloc_node(cfq_pool,
3009 gfp_mask | __GFP_ZERO,
3010 cfqd->queue->node);
3011 spin_lock_irq(cfqd->queue->queue_lock);
3012 if (new_cfqq)
3013 goto retry;
3014 } else {
3015 cfqq = kmem_cache_alloc_node(cfq_pool,
3016 gfp_mask | __GFP_ZERO,
3017 cfqd->queue->node);
3020 if (cfqq) {
3021 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3022 cfq_init_prio_data(cfqq, ioc);
3023 cfq_link_cfqq_cfqg(cfqq, cfqg);
3024 cfq_log_cfqq(cfqd, cfqq, "alloced");
3025 } else
3026 cfqq = &cfqd->oom_cfqq;
3029 if (new_cfqq)
3030 kmem_cache_free(cfq_pool, new_cfqq);
3032 return cfqq;
3035 static struct cfq_queue **
3036 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3038 switch (ioprio_class) {
3039 case IOPRIO_CLASS_RT:
3040 return &cfqd->async_cfqq[0][ioprio];
3041 case IOPRIO_CLASS_BE:
3042 return &cfqd->async_cfqq[1][ioprio];
3043 case IOPRIO_CLASS_IDLE:
3044 return &cfqd->async_idle_cfqq;
3045 default:
3046 BUG();
3050 static struct cfq_queue *
3051 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3052 gfp_t gfp_mask)
3054 const int ioprio = task_ioprio(ioc);
3055 const int ioprio_class = task_ioprio_class(ioc);
3056 struct cfq_queue **async_cfqq = NULL;
3057 struct cfq_queue *cfqq = NULL;
3059 if (!is_sync) {
3060 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3061 cfqq = *async_cfqq;
3064 if (!cfqq)
3065 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3068 * pin the queue now that it's allocated, scheduler exit will prune it
3070 if (!is_sync && !(*async_cfqq)) {
3071 cfqq->ref++;
3072 *async_cfqq = cfqq;
3075 cfqq->ref++;
3076 return cfqq;
3080 * We drop cfq io contexts lazily, so we may find a dead one.
3082 static void
3083 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3084 struct cfq_io_context *cic)
3086 unsigned long flags;
3088 WARN_ON(!list_empty(&cic->queue_list));
3089 BUG_ON(cic->key != cfqd_dead_key(cfqd));
3091 spin_lock_irqsave(&ioc->lock, flags);
3093 BUG_ON(rcu_dereference_check(ioc->ioc_data,
3094 lockdep_is_held(&ioc->lock)) == cic);
3096 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3097 hlist_del_rcu(&cic->cic_list);
3098 spin_unlock_irqrestore(&ioc->lock, flags);
3100 cfq_cic_free(cic);
3103 static struct cfq_io_context *
3104 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3106 struct cfq_io_context *cic;
3107 unsigned long flags;
3109 if (unlikely(!ioc))
3110 return NULL;
3112 rcu_read_lock();
3115 * we maintain a last-hit cache, to avoid browsing over the tree
3117 cic = rcu_dereference(ioc->ioc_data);
3118 if (cic && cic->key == cfqd) {
3119 rcu_read_unlock();
3120 return cic;
3123 do {
3124 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3125 rcu_read_unlock();
3126 if (!cic)
3127 break;
3128 if (unlikely(cic->key != cfqd)) {
3129 cfq_drop_dead_cic(cfqd, ioc, cic);
3130 rcu_read_lock();
3131 continue;
3134 spin_lock_irqsave(&ioc->lock, flags);
3135 rcu_assign_pointer(ioc->ioc_data, cic);
3136 spin_unlock_irqrestore(&ioc->lock, flags);
3137 break;
3138 } while (1);
3140 return cic;
3144 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3145 * the process specific cfq io context when entered from the block layer.
3146 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3148 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3149 struct cfq_io_context *cic, gfp_t gfp_mask)
3151 unsigned long flags;
3152 int ret;
3154 ret = radix_tree_preload(gfp_mask);
3155 if (!ret) {
3156 cic->ioc = ioc;
3157 cic->key = cfqd;
3159 spin_lock_irqsave(&ioc->lock, flags);
3160 ret = radix_tree_insert(&ioc->radix_root,
3161 cfqd->cic_index, cic);
3162 if (!ret)
3163 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3164 spin_unlock_irqrestore(&ioc->lock, flags);
3166 radix_tree_preload_end();
3168 if (!ret) {
3169 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3170 list_add(&cic->queue_list, &cfqd->cic_list);
3171 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3175 if (ret)
3176 printk(KERN_ERR "cfq: cic link failed!\n");
3178 return ret;
3182 * Setup general io context and cfq io context. There can be several cfq
3183 * io contexts per general io context, if this process is doing io to more
3184 * than one device managed by cfq.
3186 static struct cfq_io_context *
3187 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3189 struct io_context *ioc = NULL;
3190 struct cfq_io_context *cic;
3192 might_sleep_if(gfp_mask & __GFP_WAIT);
3194 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3195 if (!ioc)
3196 return NULL;
3198 cic = cfq_cic_lookup(cfqd, ioc);
3199 if (cic)
3200 goto out;
3202 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3203 if (cic == NULL)
3204 goto err;
3206 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3207 goto err_free;
3209 out:
3210 smp_read_barrier_depends();
3211 if (unlikely(ioc->ioprio_changed))
3212 cfq_ioc_set_ioprio(ioc);
3214 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3215 if (unlikely(ioc->cgroup_changed))
3216 cfq_ioc_set_cgroup(ioc);
3217 #endif
3218 return cic;
3219 err_free:
3220 cfq_cic_free(cic);
3221 err:
3222 put_io_context(ioc);
3223 return NULL;
3226 static void
3227 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3229 unsigned long elapsed = jiffies - ttime->last_end_request;
3230 elapsed = min(elapsed, 2UL * slice_idle);
3232 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3233 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3234 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3237 static void
3238 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3239 struct cfq_io_context *cic)
3241 if (cfq_cfqq_sync(cfqq)) {
3242 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3243 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3244 cfqd->cfq_slice_idle);
3246 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3247 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3248 #endif
3251 static void
3252 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3253 struct request *rq)
3255 sector_t sdist = 0;
3256 sector_t n_sec = blk_rq_sectors(rq);
3257 if (cfqq->last_request_pos) {
3258 if (cfqq->last_request_pos < blk_rq_pos(rq))
3259 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3260 else
3261 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3264 cfqq->seek_history <<= 1;
3265 if (blk_queue_nonrot(cfqd->queue))
3266 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3267 else
3268 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3272 * Disable idle window if the process thinks too long or seeks so much that
3273 * it doesn't matter
3275 static void
3276 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3277 struct cfq_io_context *cic)
3279 int old_idle, enable_idle;
3282 * Don't idle for async or idle io prio class
3284 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3285 return;
3287 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3289 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3290 cfq_mark_cfqq_deep(cfqq);
3292 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3293 enable_idle = 0;
3294 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3295 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3296 enable_idle = 0;
3297 else if (sample_valid(cic->ttime.ttime_samples)) {
3298 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3299 enable_idle = 0;
3300 else
3301 enable_idle = 1;
3304 if (old_idle != enable_idle) {
3305 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3306 if (enable_idle)
3307 cfq_mark_cfqq_idle_window(cfqq);
3308 else
3309 cfq_clear_cfqq_idle_window(cfqq);
3314 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3315 * no or if we aren't sure, a 1 will cause a preempt.
3317 static bool
3318 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3319 struct request *rq)
3321 struct cfq_queue *cfqq;
3323 cfqq = cfqd->active_queue;
3324 if (!cfqq)
3325 return false;
3327 if (cfq_class_idle(new_cfqq))
3328 return false;
3330 if (cfq_class_idle(cfqq))
3331 return true;
3334 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3336 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3337 return false;
3340 * if the new request is sync, but the currently running queue is
3341 * not, let the sync request have priority.
3343 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3344 return true;
3346 if (new_cfqq->cfqg != cfqq->cfqg)
3347 return false;
3349 if (cfq_slice_used(cfqq))
3350 return true;
3352 /* Allow preemption only if we are idling on sync-noidle tree */
3353 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3354 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3355 new_cfqq->service_tree->count == 2 &&
3356 RB_EMPTY_ROOT(&cfqq->sort_list))
3357 return true;
3360 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3362 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3363 return true;
3365 /* An idle queue should not be idle now for some reason */
3366 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3367 return true;
3369 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3370 return false;
3373 * if this request is as-good as one we would expect from the
3374 * current cfqq, let it preempt
3376 if (cfq_rq_close(cfqd, cfqq, rq))
3377 return true;
3379 return false;
3383 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3384 * let it have half of its nominal slice.
3386 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3388 struct cfq_queue *old_cfqq = cfqd->active_queue;
3390 cfq_log_cfqq(cfqd, cfqq, "preempt");
3391 cfq_slice_expired(cfqd, 1);
3394 * workload type is changed, don't save slice, otherwise preempt
3395 * doesn't happen
3397 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3398 cfqq->cfqg->saved_workload_slice = 0;
3401 * Put the new queue at the front of the of the current list,
3402 * so we know that it will be selected next.
3404 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3406 cfq_service_tree_add(cfqd, cfqq, 1);
3408 cfqq->slice_end = 0;
3409 cfq_mark_cfqq_slice_new(cfqq);
3413 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3414 * something we should do about it
3416 static void
3417 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3418 struct request *rq)
3420 struct cfq_io_context *cic = RQ_CIC(rq);
3422 cfqd->rq_queued++;
3424 cfq_update_io_thinktime(cfqd, cfqq, cic);
3425 cfq_update_io_seektime(cfqd, cfqq, rq);
3426 cfq_update_idle_window(cfqd, cfqq, cic);
3428 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3430 if (cfqq == cfqd->active_queue) {
3432 * Remember that we saw a request from this process, but
3433 * don't start queuing just yet. Otherwise we risk seeing lots
3434 * of tiny requests, because we disrupt the normal plugging
3435 * and merging. If the request is already larger than a single
3436 * page, let it rip immediately. For that case we assume that
3437 * merging is already done. Ditto for a busy system that
3438 * has other work pending, don't risk delaying until the
3439 * idle timer unplug to continue working.
3441 if (cfq_cfqq_wait_request(cfqq)) {
3442 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3443 cfqd->busy_queues > 1) {
3444 cfq_del_timer(cfqd, cfqq);
3445 cfq_clear_cfqq_wait_request(cfqq);
3446 __blk_run_queue(cfqd->queue);
3447 } else {
3448 cfq_blkiocg_update_idle_time_stats(
3449 &cfqq->cfqg->blkg);
3450 cfq_mark_cfqq_must_dispatch(cfqq);
3453 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3455 * not the active queue - expire current slice if it is
3456 * idle and has expired it's mean thinktime or this new queue
3457 * has some old slice time left and is of higher priority or
3458 * this new queue is RT and the current one is BE
3460 cfq_preempt_queue(cfqd, cfqq);
3461 __blk_run_queue(cfqd->queue);
3465 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3467 struct cfq_data *cfqd = q->elevator->elevator_data;
3468 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3470 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3471 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3473 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3474 list_add_tail(&rq->queuelist, &cfqq->fifo);
3475 cfq_add_rq_rb(rq);
3476 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3477 &cfqd->serving_group->blkg, rq_data_dir(rq),
3478 rq_is_sync(rq));
3479 cfq_rq_enqueued(cfqd, cfqq, rq);
3483 * Update hw_tag based on peak queue depth over 50 samples under
3484 * sufficient load.
3486 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3488 struct cfq_queue *cfqq = cfqd->active_queue;
3490 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3491 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3493 if (cfqd->hw_tag == 1)
3494 return;
3496 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3497 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3498 return;
3501 * If active queue hasn't enough requests and can idle, cfq might not
3502 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3503 * case
3505 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3506 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3507 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3508 return;
3510 if (cfqd->hw_tag_samples++ < 50)
3511 return;
3513 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3514 cfqd->hw_tag = 1;
3515 else
3516 cfqd->hw_tag = 0;
3519 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3521 struct cfq_io_context *cic = cfqd->active_cic;
3523 /* If the queue already has requests, don't wait */
3524 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3525 return false;
3527 /* If there are other queues in the group, don't wait */
3528 if (cfqq->cfqg->nr_cfqq > 1)
3529 return false;
3531 /* the only queue in the group, but think time is big */
3532 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3533 return false;
3535 if (cfq_slice_used(cfqq))
3536 return true;
3538 /* if slice left is less than think time, wait busy */
3539 if (cic && sample_valid(cic->ttime.ttime_samples)
3540 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3541 return true;
3544 * If think times is less than a jiffy than ttime_mean=0 and above
3545 * will not be true. It might happen that slice has not expired yet
3546 * but will expire soon (4-5 ns) during select_queue(). To cover the
3547 * case where think time is less than a jiffy, mark the queue wait
3548 * busy if only 1 jiffy is left in the slice.
3550 if (cfqq->slice_end - jiffies == 1)
3551 return true;
3553 return false;
3556 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3558 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3559 struct cfq_data *cfqd = cfqq->cfqd;
3560 const int sync = rq_is_sync(rq);
3561 unsigned long now;
3563 now = jiffies;
3564 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3565 !!(rq->cmd_flags & REQ_NOIDLE));
3567 cfq_update_hw_tag(cfqd);
3569 WARN_ON(!cfqd->rq_in_driver);
3570 WARN_ON(!cfqq->dispatched);
3571 cfqd->rq_in_driver--;
3572 cfqq->dispatched--;
3573 (RQ_CFQG(rq))->dispatched--;
3574 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3575 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3576 rq_data_dir(rq), rq_is_sync(rq));
3578 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3580 if (sync) {
3581 struct cfq_rb_root *service_tree;
3583 RQ_CIC(rq)->ttime.last_end_request = now;
3585 if (cfq_cfqq_on_rr(cfqq))
3586 service_tree = cfqq->service_tree;
3587 else
3588 service_tree = service_tree_for(cfqq->cfqg,
3589 cfqq_prio(cfqq), cfqq_type(cfqq));
3590 service_tree->ttime.last_end_request = now;
3591 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3592 cfqd->last_delayed_sync = now;
3595 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3596 cfqq->cfqg->ttime.last_end_request = now;
3597 #endif
3600 * If this is the active queue, check if it needs to be expired,
3601 * or if we want to idle in case it has no pending requests.
3603 if (cfqd->active_queue == cfqq) {
3604 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3606 if (cfq_cfqq_slice_new(cfqq)) {
3607 cfq_set_prio_slice(cfqd, cfqq);
3608 cfq_clear_cfqq_slice_new(cfqq);
3612 * Should we wait for next request to come in before we expire
3613 * the queue.
3615 if (cfq_should_wait_busy(cfqd, cfqq)) {
3616 unsigned long extend_sl = cfqd->cfq_slice_idle;
3617 if (!cfqd->cfq_slice_idle)
3618 extend_sl = cfqd->cfq_group_idle;
3619 cfqq->slice_end = jiffies + extend_sl;
3620 cfq_mark_cfqq_wait_busy(cfqq);
3621 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3625 * Idling is not enabled on:
3626 * - expired queues
3627 * - idle-priority queues
3628 * - async queues
3629 * - queues with still some requests queued
3630 * - when there is a close cooperator
3632 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3633 cfq_slice_expired(cfqd, 1);
3634 else if (sync && cfqq_empty &&
3635 !cfq_close_cooperator(cfqd, cfqq)) {
3636 cfq_arm_slice_timer(cfqd);
3640 if (!cfqd->rq_in_driver)
3641 cfq_schedule_dispatch(cfqd);
3644 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3646 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3647 cfq_mark_cfqq_must_alloc_slice(cfqq);
3648 return ELV_MQUEUE_MUST;
3651 return ELV_MQUEUE_MAY;
3654 static int cfq_may_queue(struct request_queue *q, int rw)
3656 struct cfq_data *cfqd = q->elevator->elevator_data;
3657 struct task_struct *tsk = current;
3658 struct cfq_io_context *cic;
3659 struct cfq_queue *cfqq;
3662 * don't force setup of a queue from here, as a call to may_queue
3663 * does not necessarily imply that a request actually will be queued.
3664 * so just lookup a possibly existing queue, or return 'may queue'
3665 * if that fails
3667 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3668 if (!cic)
3669 return ELV_MQUEUE_MAY;
3671 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3672 if (cfqq) {
3673 cfq_init_prio_data(cfqq, cic->ioc);
3675 return __cfq_may_queue(cfqq);
3678 return ELV_MQUEUE_MAY;
3682 * queue lock held here
3684 static void cfq_put_request(struct request *rq)
3686 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3688 if (cfqq) {
3689 const int rw = rq_data_dir(rq);
3691 BUG_ON(!cfqq->allocated[rw]);
3692 cfqq->allocated[rw]--;
3694 put_io_context(RQ_CIC(rq)->ioc);
3696 rq->elevator_private[0] = NULL;
3697 rq->elevator_private[1] = NULL;
3699 /* Put down rq reference on cfqg */
3700 cfq_put_cfqg(RQ_CFQG(rq));
3701 rq->elevator_private[2] = NULL;
3703 cfq_put_queue(cfqq);
3707 static struct cfq_queue *
3708 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3709 struct cfq_queue *cfqq)
3711 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3712 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3713 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3714 cfq_put_queue(cfqq);
3715 return cic_to_cfqq(cic, 1);
3719 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3720 * was the last process referring to said cfqq.
3722 static struct cfq_queue *
3723 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3725 if (cfqq_process_refs(cfqq) == 1) {
3726 cfqq->pid = current->pid;
3727 cfq_clear_cfqq_coop(cfqq);
3728 cfq_clear_cfqq_split_coop(cfqq);
3729 return cfqq;
3732 cic_set_cfqq(cic, NULL, 1);
3734 cfq_put_cooperator(cfqq);
3736 cfq_put_queue(cfqq);
3737 return NULL;
3740 * Allocate cfq data structures associated with this request.
3742 static int
3743 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3745 struct cfq_data *cfqd = q->elevator->elevator_data;
3746 struct cfq_io_context *cic;
3747 const int rw = rq_data_dir(rq);
3748 const bool is_sync = rq_is_sync(rq);
3749 struct cfq_queue *cfqq;
3750 unsigned long flags;
3752 might_sleep_if(gfp_mask & __GFP_WAIT);
3754 cic = cfq_get_io_context(cfqd, gfp_mask);
3756 spin_lock_irqsave(q->queue_lock, flags);
3758 if (!cic)
3759 goto queue_fail;
3761 new_queue:
3762 cfqq = cic_to_cfqq(cic, is_sync);
3763 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3764 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3765 cic_set_cfqq(cic, cfqq, is_sync);
3766 } else {
3768 * If the queue was seeky for too long, break it apart.
3770 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3771 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3772 cfqq = split_cfqq(cic, cfqq);
3773 if (!cfqq)
3774 goto new_queue;
3778 * Check to see if this queue is scheduled to merge with
3779 * another, closely cooperating queue. The merging of
3780 * queues happens here as it must be done in process context.
3781 * The reference on new_cfqq was taken in merge_cfqqs.
3783 if (cfqq->new_cfqq)
3784 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3787 cfqq->allocated[rw]++;
3789 cfqq->ref++;
3790 rq->elevator_private[0] = cic;
3791 rq->elevator_private[1] = cfqq;
3792 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3793 spin_unlock_irqrestore(q->queue_lock, flags);
3794 return 0;
3796 queue_fail:
3797 cfq_schedule_dispatch(cfqd);
3798 spin_unlock_irqrestore(q->queue_lock, flags);
3799 cfq_log(cfqd, "set_request fail");
3800 return 1;
3803 static void cfq_kick_queue(struct work_struct *work)
3805 struct cfq_data *cfqd =
3806 container_of(work, struct cfq_data, unplug_work);
3807 struct request_queue *q = cfqd->queue;
3809 spin_lock_irq(q->queue_lock);
3810 __blk_run_queue(cfqd->queue);
3811 spin_unlock_irq(q->queue_lock);
3815 * Timer running if the active_queue is currently idling inside its time slice
3817 static void cfq_idle_slice_timer(unsigned long data)
3819 struct cfq_data *cfqd = (struct cfq_data *) data;
3820 struct cfq_queue *cfqq;
3821 unsigned long flags;
3822 int timed_out = 1;
3824 cfq_log(cfqd, "idle timer fired");
3826 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3828 cfqq = cfqd->active_queue;
3829 if (cfqq) {
3830 timed_out = 0;
3833 * We saw a request before the queue expired, let it through
3835 if (cfq_cfqq_must_dispatch(cfqq))
3836 goto out_kick;
3839 * expired
3841 if (cfq_slice_used(cfqq))
3842 goto expire;
3845 * only expire and reinvoke request handler, if there are
3846 * other queues with pending requests
3848 if (!cfqd->busy_queues)
3849 goto out_cont;
3852 * not expired and it has a request pending, let it dispatch
3854 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3855 goto out_kick;
3858 * Queue depth flag is reset only when the idle didn't succeed
3860 cfq_clear_cfqq_deep(cfqq);
3862 expire:
3863 cfq_slice_expired(cfqd, timed_out);
3864 out_kick:
3865 cfq_schedule_dispatch(cfqd);
3866 out_cont:
3867 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3870 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3872 del_timer_sync(&cfqd->idle_slice_timer);
3873 cancel_work_sync(&cfqd->unplug_work);
3876 static void cfq_put_async_queues(struct cfq_data *cfqd)
3878 int i;
3880 for (i = 0; i < IOPRIO_BE_NR; i++) {
3881 if (cfqd->async_cfqq[0][i])
3882 cfq_put_queue(cfqd->async_cfqq[0][i]);
3883 if (cfqd->async_cfqq[1][i])
3884 cfq_put_queue(cfqd->async_cfqq[1][i]);
3887 if (cfqd->async_idle_cfqq)
3888 cfq_put_queue(cfqd->async_idle_cfqq);
3891 static void cfq_exit_queue(struct elevator_queue *e)
3893 struct cfq_data *cfqd = e->elevator_data;
3894 struct request_queue *q = cfqd->queue;
3895 bool wait = false;
3897 cfq_shutdown_timer_wq(cfqd);
3899 spin_lock_irq(q->queue_lock);
3901 if (cfqd->active_queue)
3902 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3904 while (!list_empty(&cfqd->cic_list)) {
3905 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3906 struct cfq_io_context,
3907 queue_list);
3909 __cfq_exit_single_io_context(cfqd, cic);
3912 cfq_put_async_queues(cfqd);
3913 cfq_release_cfq_groups(cfqd);
3916 * If there are groups which we could not unlink from blkcg list,
3917 * wait for a rcu period for them to be freed.
3919 if (cfqd->nr_blkcg_linked_grps)
3920 wait = true;
3922 spin_unlock_irq(q->queue_lock);
3924 cfq_shutdown_timer_wq(cfqd);
3926 spin_lock(&cic_index_lock);
3927 ida_remove(&cic_index_ida, cfqd->cic_index);
3928 spin_unlock(&cic_index_lock);
3931 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3932 * Do this wait only if there are other unlinked groups out
3933 * there. This can happen if cgroup deletion path claimed the
3934 * responsibility of cleaning up a group before queue cleanup code
3935 * get to the group.
3937 * Do not call synchronize_rcu() unconditionally as there are drivers
3938 * which create/delete request queue hundreds of times during scan/boot
3939 * and synchronize_rcu() can take significant time and slow down boot.
3941 if (wait)
3942 synchronize_rcu();
3944 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3945 /* Free up per cpu stats for root group */
3946 free_percpu(cfqd->root_group.blkg.stats_cpu);
3947 #endif
3948 kfree(cfqd);
3951 static int cfq_alloc_cic_index(void)
3953 int index, error;
3955 do {
3956 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3957 return -ENOMEM;
3959 spin_lock(&cic_index_lock);
3960 error = ida_get_new(&cic_index_ida, &index);
3961 spin_unlock(&cic_index_lock);
3962 if (error && error != -EAGAIN)
3963 return error;
3964 } while (error);
3966 return index;
3969 static void *cfq_init_queue(struct request_queue *q)
3971 struct cfq_data *cfqd;
3972 int i, j;
3973 struct cfq_group *cfqg;
3974 struct cfq_rb_root *st;
3976 i = cfq_alloc_cic_index();
3977 if (i < 0)
3978 return NULL;
3980 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3981 if (!cfqd) {
3982 spin_lock(&cic_index_lock);
3983 ida_remove(&cic_index_ida, i);
3984 spin_unlock(&cic_index_lock);
3985 return NULL;
3989 * Don't need take queue_lock in the routine, since we are
3990 * initializing the ioscheduler, and nobody is using cfqd
3992 cfqd->cic_index = i;
3994 /* Init root service tree */
3995 cfqd->grp_service_tree = CFQ_RB_ROOT;
3997 /* Init root group */
3998 cfqg = &cfqd->root_group;
3999 for_each_cfqg_st(cfqg, i, j, st)
4000 *st = CFQ_RB_ROOT;
4001 RB_CLEAR_NODE(&cfqg->rb_node);
4003 /* Give preference to root group over other groups */
4004 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
4006 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4008 * Set root group reference to 2. One reference will be dropped when
4009 * all groups on cfqd->cfqg_list are being deleted during queue exit.
4010 * Other reference will remain there as we don't want to delete this
4011 * group as it is statically allocated and gets destroyed when
4012 * throtl_data goes away.
4014 cfqg->ref = 2;
4016 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4017 kfree(cfqg);
4018 kfree(cfqd);
4019 return NULL;
4022 rcu_read_lock();
4024 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4025 (void *)cfqd, 0);
4026 rcu_read_unlock();
4027 cfqd->nr_blkcg_linked_grps++;
4029 /* Add group on cfqd->cfqg_list */
4030 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4031 #endif
4033 * Not strictly needed (since RB_ROOT just clears the node and we
4034 * zeroed cfqd on alloc), but better be safe in case someone decides
4035 * to add magic to the rb code
4037 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4038 cfqd->prio_trees[i] = RB_ROOT;
4041 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4042 * Grab a permanent reference to it, so that the normal code flow
4043 * will not attempt to free it.
4045 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4046 cfqd->oom_cfqq.ref++;
4047 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4049 INIT_LIST_HEAD(&cfqd->cic_list);
4051 cfqd->queue = q;
4053 init_timer(&cfqd->idle_slice_timer);
4054 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4055 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4057 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4059 cfqd->cfq_quantum = cfq_quantum;
4060 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4061 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4062 cfqd->cfq_back_max = cfq_back_max;
4063 cfqd->cfq_back_penalty = cfq_back_penalty;
4064 cfqd->cfq_slice[0] = cfq_slice_async;
4065 cfqd->cfq_slice[1] = cfq_slice_sync;
4066 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4067 cfqd->cfq_slice_idle = cfq_slice_idle;
4068 cfqd->cfq_group_idle = cfq_group_idle;
4069 cfqd->cfq_latency = 1;
4070 cfqd->hw_tag = -1;
4072 * we optimistically start assuming sync ops weren't delayed in last
4073 * second, in order to have larger depth for async operations.
4075 cfqd->last_delayed_sync = jiffies - HZ;
4076 return cfqd;
4079 static void cfq_slab_kill(void)
4082 * Caller already ensured that pending RCU callbacks are completed,
4083 * so we should have no busy allocations at this point.
4085 if (cfq_pool)
4086 kmem_cache_destroy(cfq_pool);
4087 if (cfq_ioc_pool)
4088 kmem_cache_destroy(cfq_ioc_pool);
4091 static int __init cfq_slab_setup(void)
4093 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4094 if (!cfq_pool)
4095 goto fail;
4097 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4098 if (!cfq_ioc_pool)
4099 goto fail;
4101 return 0;
4102 fail:
4103 cfq_slab_kill();
4104 return -ENOMEM;
4108 * sysfs parts below -->
4110 static ssize_t
4111 cfq_var_show(unsigned int var, char *page)
4113 return sprintf(page, "%d\n", var);
4116 static ssize_t
4117 cfq_var_store(unsigned int *var, const char *page, size_t count)
4119 char *p = (char *) page;
4121 *var = simple_strtoul(p, &p, 10);
4122 return count;
4125 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4126 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4128 struct cfq_data *cfqd = e->elevator_data; \
4129 unsigned int __data = __VAR; \
4130 if (__CONV) \
4131 __data = jiffies_to_msecs(__data); \
4132 return cfq_var_show(__data, (page)); \
4134 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4135 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4136 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4137 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4138 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4139 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4140 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4141 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4142 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4143 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4144 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4145 #undef SHOW_FUNCTION
4147 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4148 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4150 struct cfq_data *cfqd = e->elevator_data; \
4151 unsigned int __data; \
4152 int ret = cfq_var_store(&__data, (page), count); \
4153 if (__data < (MIN)) \
4154 __data = (MIN); \
4155 else if (__data > (MAX)) \
4156 __data = (MAX); \
4157 if (__CONV) \
4158 *(__PTR) = msecs_to_jiffies(__data); \
4159 else \
4160 *(__PTR) = __data; \
4161 return ret; \
4163 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4164 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4165 UINT_MAX, 1);
4166 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4167 UINT_MAX, 1);
4168 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4169 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4170 UINT_MAX, 0);
4171 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4172 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4173 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4174 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4175 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4176 UINT_MAX, 0);
4177 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4178 #undef STORE_FUNCTION
4180 #define CFQ_ATTR(name) \
4181 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4183 static struct elv_fs_entry cfq_attrs[] = {
4184 CFQ_ATTR(quantum),
4185 CFQ_ATTR(fifo_expire_sync),
4186 CFQ_ATTR(fifo_expire_async),
4187 CFQ_ATTR(back_seek_max),
4188 CFQ_ATTR(back_seek_penalty),
4189 CFQ_ATTR(slice_sync),
4190 CFQ_ATTR(slice_async),
4191 CFQ_ATTR(slice_async_rq),
4192 CFQ_ATTR(slice_idle),
4193 CFQ_ATTR(group_idle),
4194 CFQ_ATTR(low_latency),
4195 __ATTR_NULL
4198 static struct elevator_type iosched_cfq = {
4199 .ops = {
4200 .elevator_merge_fn = cfq_merge,
4201 .elevator_merged_fn = cfq_merged_request,
4202 .elevator_merge_req_fn = cfq_merged_requests,
4203 .elevator_allow_merge_fn = cfq_allow_merge,
4204 .elevator_bio_merged_fn = cfq_bio_merged,
4205 .elevator_dispatch_fn = cfq_dispatch_requests,
4206 .elevator_add_req_fn = cfq_insert_request,
4207 .elevator_activate_req_fn = cfq_activate_request,
4208 .elevator_deactivate_req_fn = cfq_deactivate_request,
4209 .elevator_completed_req_fn = cfq_completed_request,
4210 .elevator_former_req_fn = elv_rb_former_request,
4211 .elevator_latter_req_fn = elv_rb_latter_request,
4212 .elevator_set_req_fn = cfq_set_request,
4213 .elevator_put_req_fn = cfq_put_request,
4214 .elevator_may_queue_fn = cfq_may_queue,
4215 .elevator_init_fn = cfq_init_queue,
4216 .elevator_exit_fn = cfq_exit_queue,
4217 .trim = cfq_free_io_context,
4219 .elevator_attrs = cfq_attrs,
4220 .elevator_name = "cfq",
4221 .elevator_owner = THIS_MODULE,
4224 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4225 static struct blkio_policy_type blkio_policy_cfq = {
4226 .ops = {
4227 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4228 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4230 .plid = BLKIO_POLICY_PROP,
4232 #else
4233 static struct blkio_policy_type blkio_policy_cfq;
4234 #endif
4236 static int __init cfq_init(void)
4239 * could be 0 on HZ < 1000 setups
4241 if (!cfq_slice_async)
4242 cfq_slice_async = 1;
4243 if (!cfq_slice_idle)
4244 cfq_slice_idle = 1;
4246 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4247 if (!cfq_group_idle)
4248 cfq_group_idle = 1;
4249 #else
4250 cfq_group_idle = 0;
4251 #endif
4252 if (cfq_slab_setup())
4253 return -ENOMEM;
4255 elv_register(&iosched_cfq);
4256 blkio_policy_register(&blkio_policy_cfq);
4258 return 0;
4261 static void __exit cfq_exit(void)
4263 DECLARE_COMPLETION_ONSTACK(all_gone);
4264 blkio_policy_unregister(&blkio_policy_cfq);
4265 elv_unregister(&iosched_cfq);
4266 ioc_gone = &all_gone;
4267 /* ioc_gone's update must be visible before reading ioc_count */
4268 smp_wmb();
4271 * this also protects us from entering cfq_slab_kill() with
4272 * pending RCU callbacks
4274 if (elv_ioc_count_read(cfq_ioc_count))
4275 wait_for_completion(&all_gone);
4276 ida_destroy(&cic_index_ida);
4277 cfq_slab_kill();
4280 module_init(cfq_init);
4281 module_exit(cfq_exit);
4283 MODULE_AUTHOR("Jens Axboe");
4284 MODULE_LICENSE("GPL");
4285 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");