isci: add some type safety to the state machine interface
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / block / cfq-iosched.c
blob3c7b537bf9081f9f43b90c7cdcd9a95073f1c02b
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;
91 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92 .count = 0, .min_vdisktime = 0, }
95 * Per process-grouping structure
97 struct cfq_queue {
98 /* reference count */
99 int ref;
100 /* various state flags, see below */
101 unsigned int flags;
102 /* parent cfq_data */
103 struct cfq_data *cfqd;
104 /* service_tree member */
105 struct rb_node rb_node;
106 /* service_tree key */
107 unsigned long rb_key;
108 /* prio tree member */
109 struct rb_node p_node;
110 /* prio tree root we belong to, if any */
111 struct rb_root *p_root;
112 /* sorted list of pending requests */
113 struct rb_root sort_list;
114 /* if fifo isn't expired, next request to serve */
115 struct request *next_rq;
116 /* requests queued in sort_list */
117 int queued[2];
118 /* currently allocated requests */
119 int allocated[2];
120 /* fifo list of requests in sort_list */
121 struct list_head fifo;
123 /* time when queue got scheduled in to dispatch first request. */
124 unsigned long dispatch_start;
125 unsigned int allocated_slice;
126 unsigned int slice_dispatch;
127 /* time when first request from queue completed and slice started. */
128 unsigned long slice_start;
129 unsigned long slice_end;
130 long slice_resid;
132 /* pending metadata requests */
133 int meta_pending;
134 /* number of requests that are on the dispatch list or inside driver */
135 int dispatched;
137 /* io prio of this group */
138 unsigned short ioprio, org_ioprio;
139 unsigned short ioprio_class, org_ioprio_class;
141 pid_t pid;
143 u32 seek_history;
144 sector_t last_request_pos;
146 struct cfq_rb_root *service_tree;
147 struct cfq_queue *new_cfqq;
148 struct cfq_group *cfqg;
149 /* Number of sectors dispatched from queue in single dispatch round */
150 unsigned long nr_sectors;
154 * First index in the service_trees.
155 * IDLE is handled separately, so it has negative index
157 enum wl_prio_t {
158 BE_WORKLOAD = 0,
159 RT_WORKLOAD = 1,
160 IDLE_WORKLOAD = 2,
161 CFQ_PRIO_NR,
165 * Second index in the service_trees.
167 enum wl_type_t {
168 ASYNC_WORKLOAD = 0,
169 SYNC_NOIDLE_WORKLOAD = 1,
170 SYNC_WORKLOAD = 2
173 /* This is per cgroup per device grouping structure */
174 struct cfq_group {
175 /* group service_tree member */
176 struct rb_node rb_node;
178 /* group service_tree key */
179 u64 vdisktime;
180 unsigned int weight;
181 unsigned int new_weight;
182 bool needs_update;
184 /* number of cfqq currently on this group */
185 int nr_cfqq;
188 * Per group busy queues average. Useful for workload slice calc. We
189 * create the array for each prio class but at run time it is used
190 * only for RT and BE class and slot for IDLE class remains unused.
191 * This is primarily done to avoid confusion and a gcc warning.
193 unsigned int busy_queues_avg[CFQ_PRIO_NR];
195 * rr lists of queues with requests. We maintain service trees for
196 * RT and BE classes. These trees are subdivided in subclasses
197 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198 * class there is no subclassification and all the cfq queues go on
199 * a single tree service_tree_idle.
200 * Counts are embedded in the cfq_rb_root
202 struct cfq_rb_root service_trees[2][3];
203 struct cfq_rb_root service_tree_idle;
205 unsigned long saved_workload_slice;
206 enum wl_type_t saved_workload;
207 enum wl_prio_t saved_serving_prio;
208 struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210 struct hlist_node cfqd_node;
211 int ref;
212 #endif
213 /* number of requests that are on the dispatch list or inside driver */
214 int dispatched;
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) \
397 static inline bool iops_mode(struct cfq_data *cfqd)
400 * If we are not idling on queues and it is a NCQ drive, parallel
401 * execution of requests is on and measuring time is not possible
402 * in most of the cases until and unless we drive shallower queue
403 * depths and that becomes a performance bottleneck. In such cases
404 * switch to start providing fairness in terms of number of IOs.
406 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
407 return true;
408 else
409 return false;
412 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
414 if (cfq_class_idle(cfqq))
415 return IDLE_WORKLOAD;
416 if (cfq_class_rt(cfqq))
417 return RT_WORKLOAD;
418 return BE_WORKLOAD;
422 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
424 if (!cfq_cfqq_sync(cfqq))
425 return ASYNC_WORKLOAD;
426 if (!cfq_cfqq_idle_window(cfqq))
427 return SYNC_NOIDLE_WORKLOAD;
428 return SYNC_WORKLOAD;
431 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
432 struct cfq_data *cfqd,
433 struct cfq_group *cfqg)
435 if (wl == IDLE_WORKLOAD)
436 return cfqg->service_tree_idle.count;
438 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
439 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
440 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
443 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
444 struct cfq_group *cfqg)
446 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
447 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
450 static void cfq_dispatch_insert(struct request_queue *, struct request *);
451 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
452 struct io_context *, gfp_t);
453 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
454 struct io_context *);
456 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
457 bool is_sync)
459 return cic->cfqq[is_sync];
462 static inline void cic_set_cfqq(struct cfq_io_context *cic,
463 struct cfq_queue *cfqq, bool is_sync)
465 cic->cfqq[is_sync] = cfqq;
468 #define CIC_DEAD_KEY 1ul
469 #define CIC_DEAD_INDEX_SHIFT 1
471 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
473 return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
476 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
478 struct cfq_data *cfqd = cic->key;
480 if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
481 return NULL;
483 return cfqd;
487 * We regard a request as SYNC, if it's either a read or has the SYNC bit
488 * set (in which case it could also be direct WRITE).
490 static inline bool cfq_bio_sync(struct bio *bio)
492 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
496 * scheduler run of queue, if there are requests pending and no one in the
497 * driver that will restart queueing
499 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
501 if (cfqd->busy_queues) {
502 cfq_log(cfqd, "schedule dispatch");
503 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
508 * Scale schedule slice based on io priority. Use the sync time slice only
509 * if a queue is marked sync and has sync io queued. A sync queue with async
510 * io only, should not get full sync slice length.
512 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
513 unsigned short prio)
515 const int base_slice = cfqd->cfq_slice[sync];
517 WARN_ON(prio >= IOPRIO_BE_NR);
519 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
522 static inline int
523 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
525 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
528 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
530 u64 d = delta << CFQ_SERVICE_SHIFT;
532 d = d * BLKIO_WEIGHT_DEFAULT;
533 do_div(d, cfqg->weight);
534 return d;
537 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
539 s64 delta = (s64)(vdisktime - min_vdisktime);
540 if (delta > 0)
541 min_vdisktime = vdisktime;
543 return min_vdisktime;
546 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
548 s64 delta = (s64)(vdisktime - min_vdisktime);
549 if (delta < 0)
550 min_vdisktime = vdisktime;
552 return min_vdisktime;
555 static void update_min_vdisktime(struct cfq_rb_root *st)
557 struct cfq_group *cfqg;
559 if (st->left) {
560 cfqg = rb_entry_cfqg(st->left);
561 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
562 cfqg->vdisktime);
567 * get averaged number of queues of RT/BE priority.
568 * average is updated, with a formula that gives more weight to higher numbers,
569 * to quickly follows sudden increases and decrease slowly
572 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
573 struct cfq_group *cfqg, bool rt)
575 unsigned min_q, max_q;
576 unsigned mult = cfq_hist_divisor - 1;
577 unsigned round = cfq_hist_divisor / 2;
578 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
580 min_q = min(cfqg->busy_queues_avg[rt], busy);
581 max_q = max(cfqg->busy_queues_avg[rt], busy);
582 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
583 cfq_hist_divisor;
584 return cfqg->busy_queues_avg[rt];
587 static inline unsigned
588 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
590 struct cfq_rb_root *st = &cfqd->grp_service_tree;
592 return cfq_target_latency * cfqg->weight / st->total_weight;
595 static inline unsigned
596 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
598 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
599 if (cfqd->cfq_latency) {
601 * interested queues (we consider only the ones with the same
602 * priority class in the cfq group)
604 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
605 cfq_class_rt(cfqq));
606 unsigned sync_slice = cfqd->cfq_slice[1];
607 unsigned expect_latency = sync_slice * iq;
608 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
610 if (expect_latency > group_slice) {
611 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
612 /* scale low_slice according to IO priority
613 * and sync vs async */
614 unsigned low_slice =
615 min(slice, base_low_slice * slice / sync_slice);
616 /* the adapted slice value is scaled to fit all iqs
617 * into the target latency */
618 slice = max(slice * group_slice / expect_latency,
619 low_slice);
622 return slice;
625 static inline void
626 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
628 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
630 cfqq->slice_start = jiffies;
631 cfqq->slice_end = jiffies + slice;
632 cfqq->allocated_slice = slice;
633 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
637 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
638 * isn't valid until the first request from the dispatch is activated
639 * and the slice time set.
641 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
643 if (cfq_cfqq_slice_new(cfqq))
644 return false;
645 if (time_before(jiffies, cfqq->slice_end))
646 return false;
648 return true;
652 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
653 * We choose the request that is closest to the head right now. Distance
654 * behind the head is penalized and only allowed to a certain extent.
656 static struct request *
657 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
659 sector_t s1, s2, d1 = 0, d2 = 0;
660 unsigned long back_max;
661 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
662 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
663 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
665 if (rq1 == NULL || rq1 == rq2)
666 return rq2;
667 if (rq2 == NULL)
668 return rq1;
670 if (rq_is_sync(rq1) != rq_is_sync(rq2))
671 return rq_is_sync(rq1) ? rq1 : rq2;
673 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META)
674 return rq1->cmd_flags & REQ_META ? rq1 : rq2;
676 s1 = blk_rq_pos(rq1);
677 s2 = blk_rq_pos(rq2);
680 * by definition, 1KiB is 2 sectors
682 back_max = cfqd->cfq_back_max * 2;
685 * Strict one way elevator _except_ in the case where we allow
686 * short backward seeks which are biased as twice the cost of a
687 * similar forward seek.
689 if (s1 >= last)
690 d1 = s1 - last;
691 else if (s1 + back_max >= last)
692 d1 = (last - s1) * cfqd->cfq_back_penalty;
693 else
694 wrap |= CFQ_RQ1_WRAP;
696 if (s2 >= last)
697 d2 = s2 - last;
698 else if (s2 + back_max >= last)
699 d2 = (last - s2) * cfqd->cfq_back_penalty;
700 else
701 wrap |= CFQ_RQ2_WRAP;
703 /* Found required data */
706 * By doing switch() on the bit mask "wrap" we avoid having to
707 * check two variables for all permutations: --> faster!
709 switch (wrap) {
710 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
711 if (d1 < d2)
712 return rq1;
713 else if (d2 < d1)
714 return rq2;
715 else {
716 if (s1 >= s2)
717 return rq1;
718 else
719 return rq2;
722 case CFQ_RQ2_WRAP:
723 return rq1;
724 case CFQ_RQ1_WRAP:
725 return rq2;
726 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
727 default:
729 * Since both rqs are wrapped,
730 * start with the one that's further behind head
731 * (--> only *one* back seek required),
732 * since back seek takes more time than forward.
734 if (s1 <= s2)
735 return rq1;
736 else
737 return rq2;
742 * The below is leftmost cache rbtree addon
744 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
746 /* Service tree is empty */
747 if (!root->count)
748 return NULL;
750 if (!root->left)
751 root->left = rb_first(&root->rb);
753 if (root->left)
754 return rb_entry(root->left, struct cfq_queue, rb_node);
756 return NULL;
759 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
761 if (!root->left)
762 root->left = rb_first(&root->rb);
764 if (root->left)
765 return rb_entry_cfqg(root->left);
767 return NULL;
770 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
772 rb_erase(n, root);
773 RB_CLEAR_NODE(n);
776 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
778 if (root->left == n)
779 root->left = NULL;
780 rb_erase_init(n, &root->rb);
781 --root->count;
785 * would be nice to take fifo expire time into account as well
787 static struct request *
788 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789 struct request *last)
791 struct rb_node *rbnext = rb_next(&last->rb_node);
792 struct rb_node *rbprev = rb_prev(&last->rb_node);
793 struct request *next = NULL, *prev = NULL;
795 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
797 if (rbprev)
798 prev = rb_entry_rq(rbprev);
800 if (rbnext)
801 next = rb_entry_rq(rbnext);
802 else {
803 rbnext = rb_first(&cfqq->sort_list);
804 if (rbnext && rbnext != &last->rb_node)
805 next = rb_entry_rq(rbnext);
808 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
811 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812 struct cfq_queue *cfqq)
815 * just an approximation, should be ok.
817 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
818 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
821 static inline s64
822 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
824 return cfqg->vdisktime - st->min_vdisktime;
827 static void
828 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
830 struct rb_node **node = &st->rb.rb_node;
831 struct rb_node *parent = NULL;
832 struct cfq_group *__cfqg;
833 s64 key = cfqg_key(st, cfqg);
834 int left = 1;
836 while (*node != NULL) {
837 parent = *node;
838 __cfqg = rb_entry_cfqg(parent);
840 if (key < cfqg_key(st, __cfqg))
841 node = &parent->rb_left;
842 else {
843 node = &parent->rb_right;
844 left = 0;
848 if (left)
849 st->left = &cfqg->rb_node;
851 rb_link_node(&cfqg->rb_node, parent, node);
852 rb_insert_color(&cfqg->rb_node, &st->rb);
855 static void
856 cfq_update_group_weight(struct cfq_group *cfqg)
858 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
859 if (cfqg->needs_update) {
860 cfqg->weight = cfqg->new_weight;
861 cfqg->needs_update = false;
865 static void
866 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
868 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
870 cfq_update_group_weight(cfqg);
871 __cfq_group_service_tree_add(st, cfqg);
872 st->total_weight += cfqg->weight;
875 static void
876 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
878 struct cfq_rb_root *st = &cfqd->grp_service_tree;
879 struct cfq_group *__cfqg;
880 struct rb_node *n;
882 cfqg->nr_cfqq++;
883 if (!RB_EMPTY_NODE(&cfqg->rb_node))
884 return;
887 * Currently put the group at the end. Later implement something
888 * so that groups get lesser vtime based on their weights, so that
889 * if group does not loose all if it was not continuously backlogged.
891 n = rb_last(&st->rb);
892 if (n) {
893 __cfqg = rb_entry_cfqg(n);
894 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
895 } else
896 cfqg->vdisktime = st->min_vdisktime;
897 cfq_group_service_tree_add(st, cfqg);
900 static void
901 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
903 st->total_weight -= cfqg->weight;
904 if (!RB_EMPTY_NODE(&cfqg->rb_node))
905 cfq_rb_erase(&cfqg->rb_node, st);
908 static void
909 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
911 struct cfq_rb_root *st = &cfqd->grp_service_tree;
913 BUG_ON(cfqg->nr_cfqq < 1);
914 cfqg->nr_cfqq--;
916 /* If there are other cfq queues under this group, don't delete it */
917 if (cfqg->nr_cfqq)
918 return;
920 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
921 cfq_group_service_tree_del(st, cfqg);
922 cfqg->saved_workload_slice = 0;
923 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
926 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
927 unsigned int *unaccounted_time)
929 unsigned int slice_used;
932 * Queue got expired before even a single request completed or
933 * got expired immediately after first request completion.
935 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
937 * Also charge the seek time incurred to the group, otherwise
938 * if there are mutiple queues in the group, each can dispatch
939 * a single request on seeky media and cause lots of seek time
940 * and group will never know it.
942 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
944 } else {
945 slice_used = jiffies - cfqq->slice_start;
946 if (slice_used > cfqq->allocated_slice) {
947 *unaccounted_time = slice_used - cfqq->allocated_slice;
948 slice_used = cfqq->allocated_slice;
950 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
951 *unaccounted_time += cfqq->slice_start -
952 cfqq->dispatch_start;
955 return slice_used;
958 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
959 struct cfq_queue *cfqq)
961 struct cfq_rb_root *st = &cfqd->grp_service_tree;
962 unsigned int used_sl, charge, unaccounted_sl = 0;
963 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
964 - cfqg->service_tree_idle.count;
966 BUG_ON(nr_sync < 0);
967 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
969 if (iops_mode(cfqd))
970 charge = cfqq->slice_dispatch;
971 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
972 charge = cfqq->allocated_slice;
974 /* Can't update vdisktime while group is on service tree */
975 cfq_group_service_tree_del(st, cfqg);
976 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
977 /* If a new weight was requested, update now, off tree */
978 cfq_group_service_tree_add(st, cfqg);
980 /* This group is being expired. Save the context */
981 if (time_after(cfqd->workload_expires, jiffies)) {
982 cfqg->saved_workload_slice = cfqd->workload_expires
983 - jiffies;
984 cfqg->saved_workload = cfqd->serving_type;
985 cfqg->saved_serving_prio = cfqd->serving_prio;
986 } else
987 cfqg->saved_workload_slice = 0;
989 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
990 st->min_vdisktime);
991 cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
992 " sect=%u", used_sl, cfqq->slice_dispatch, charge,
993 iops_mode(cfqd), cfqq->nr_sectors);
994 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
995 unaccounted_sl);
996 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
999 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1000 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1002 if (blkg)
1003 return container_of(blkg, struct cfq_group, blkg);
1004 return NULL;
1007 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1008 unsigned int weight)
1010 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1011 cfqg->new_weight = weight;
1012 cfqg->needs_update = true;
1015 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1016 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1018 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1019 unsigned int major, minor;
1022 * Add group onto cgroup list. It might happen that bdi->dev is
1023 * not initialized yet. Initialize this new group without major
1024 * and minor info and this info will be filled in once a new thread
1025 * comes for IO.
1027 if (bdi->dev) {
1028 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1029 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1030 (void *)cfqd, MKDEV(major, minor));
1031 } else
1032 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1033 (void *)cfqd, 0);
1035 cfqd->nr_blkcg_linked_grps++;
1036 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1038 /* Add group on cfqd list */
1039 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1043 * Should be called from sleepable context. No request queue lock as per
1044 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1045 * from sleepable context.
1047 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1049 struct cfq_group *cfqg = NULL;
1050 int i, j, ret;
1051 struct cfq_rb_root *st;
1053 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1054 if (!cfqg)
1055 return NULL;
1057 for_each_cfqg_st(cfqg, i, j, st)
1058 *st = CFQ_RB_ROOT;
1059 RB_CLEAR_NODE(&cfqg->rb_node);
1062 * Take the initial reference that will be released on destroy
1063 * This can be thought of a joint reference by cgroup and
1064 * elevator which will be dropped by either elevator exit
1065 * or cgroup deletion path depending on who is exiting first.
1067 cfqg->ref = 1;
1069 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1070 if (ret) {
1071 kfree(cfqg);
1072 return NULL;
1075 return cfqg;
1078 static struct cfq_group *
1079 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1081 struct cfq_group *cfqg = NULL;
1082 void *key = cfqd;
1083 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1084 unsigned int major, minor;
1087 * This is the common case when there are no blkio cgroups.
1088 * Avoid lookup in this case
1090 if (blkcg == &blkio_root_cgroup)
1091 cfqg = &cfqd->root_group;
1092 else
1093 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1095 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1096 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1097 cfqg->blkg.dev = MKDEV(major, minor);
1100 return cfqg;
1104 * Search for the cfq group current task belongs to. request_queue lock must
1105 * be held.
1107 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1109 struct blkio_cgroup *blkcg;
1110 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1111 struct request_queue *q = cfqd->queue;
1113 rcu_read_lock();
1114 blkcg = task_blkio_cgroup(current);
1115 cfqg = cfq_find_cfqg(cfqd, blkcg);
1116 if (cfqg) {
1117 rcu_read_unlock();
1118 return cfqg;
1122 * Need to allocate a group. Allocation of group also needs allocation
1123 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1124 * we need to drop rcu lock and queue_lock before we call alloc.
1126 * Not taking any queue reference here and assuming that queue is
1127 * around by the time we return. CFQ queue allocation code does
1128 * the same. It might be racy though.
1131 rcu_read_unlock();
1132 spin_unlock_irq(q->queue_lock);
1134 cfqg = cfq_alloc_cfqg(cfqd);
1136 spin_lock_irq(q->queue_lock);
1138 rcu_read_lock();
1139 blkcg = task_blkio_cgroup(current);
1142 * If some other thread already allocated the group while we were
1143 * not holding queue lock, free up the group
1145 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1147 if (__cfqg) {
1148 kfree(cfqg);
1149 rcu_read_unlock();
1150 return __cfqg;
1153 if (!cfqg)
1154 cfqg = &cfqd->root_group;
1156 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1157 rcu_read_unlock();
1158 return cfqg;
1161 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1163 cfqg->ref++;
1164 return cfqg;
1167 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1169 /* Currently, all async queues are mapped to root group */
1170 if (!cfq_cfqq_sync(cfqq))
1171 cfqg = &cfqq->cfqd->root_group;
1173 cfqq->cfqg = cfqg;
1174 /* cfqq reference on cfqg */
1175 cfqq->cfqg->ref++;
1178 static void cfq_put_cfqg(struct cfq_group *cfqg)
1180 struct cfq_rb_root *st;
1181 int i, j;
1183 BUG_ON(cfqg->ref <= 0);
1184 cfqg->ref--;
1185 if (cfqg->ref)
1186 return;
1187 for_each_cfqg_st(cfqg, i, j, st)
1188 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1189 free_percpu(cfqg->blkg.stats_cpu);
1190 kfree(cfqg);
1193 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1195 /* Something wrong if we are trying to remove same group twice */
1196 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1198 hlist_del_init(&cfqg->cfqd_node);
1201 * Put the reference taken at the time of creation so that when all
1202 * queues are gone, group can be destroyed.
1204 cfq_put_cfqg(cfqg);
1207 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1209 struct hlist_node *pos, *n;
1210 struct cfq_group *cfqg;
1212 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1214 * If cgroup removal path got to blk_group first and removed
1215 * it from cgroup list, then it will take care of destroying
1216 * cfqg also.
1218 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1219 cfq_destroy_cfqg(cfqd, cfqg);
1224 * Blk cgroup controller notification saying that blkio_group object is being
1225 * delinked as associated cgroup object is going away. That also means that
1226 * no new IO will come in this group. So get rid of this group as soon as
1227 * any pending IO in the group is finished.
1229 * This function is called under rcu_read_lock(). key is the rcu protected
1230 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1231 * read lock.
1233 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1234 * it should not be NULL as even if elevator was exiting, cgroup deltion
1235 * path got to it first.
1237 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1239 unsigned long flags;
1240 struct cfq_data *cfqd = key;
1242 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1243 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1244 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1247 #else /* GROUP_IOSCHED */
1248 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1250 return &cfqd->root_group;
1253 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1255 return cfqg;
1258 static inline void
1259 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1260 cfqq->cfqg = cfqg;
1263 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1264 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1266 #endif /* GROUP_IOSCHED */
1269 * The cfqd->service_trees holds all pending cfq_queue's that have
1270 * requests waiting to be processed. It is sorted in the order that
1271 * we will service the queues.
1273 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1274 bool add_front)
1276 struct rb_node **p, *parent;
1277 struct cfq_queue *__cfqq;
1278 unsigned long rb_key;
1279 struct cfq_rb_root *service_tree;
1280 int left;
1281 int new_cfqq = 1;
1283 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1284 cfqq_type(cfqq));
1285 if (cfq_class_idle(cfqq)) {
1286 rb_key = CFQ_IDLE_DELAY;
1287 parent = rb_last(&service_tree->rb);
1288 if (parent && parent != &cfqq->rb_node) {
1289 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1290 rb_key += __cfqq->rb_key;
1291 } else
1292 rb_key += jiffies;
1293 } else if (!add_front) {
1295 * Get our rb key offset. Subtract any residual slice
1296 * value carried from last service. A negative resid
1297 * count indicates slice overrun, and this should position
1298 * the next service time further away in the tree.
1300 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1301 rb_key -= cfqq->slice_resid;
1302 cfqq->slice_resid = 0;
1303 } else {
1304 rb_key = -HZ;
1305 __cfqq = cfq_rb_first(service_tree);
1306 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1309 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1310 new_cfqq = 0;
1312 * same position, nothing more to do
1314 if (rb_key == cfqq->rb_key &&
1315 cfqq->service_tree == service_tree)
1316 return;
1318 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1319 cfqq->service_tree = NULL;
1322 left = 1;
1323 parent = NULL;
1324 cfqq->service_tree = service_tree;
1325 p = &service_tree->rb.rb_node;
1326 while (*p) {
1327 struct rb_node **n;
1329 parent = *p;
1330 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1333 * sort by key, that represents service time.
1335 if (time_before(rb_key, __cfqq->rb_key))
1336 n = &(*p)->rb_left;
1337 else {
1338 n = &(*p)->rb_right;
1339 left = 0;
1342 p = n;
1345 if (left)
1346 service_tree->left = &cfqq->rb_node;
1348 cfqq->rb_key = rb_key;
1349 rb_link_node(&cfqq->rb_node, parent, p);
1350 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1351 service_tree->count++;
1352 if (add_front || !new_cfqq)
1353 return;
1354 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1357 static struct cfq_queue *
1358 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1359 sector_t sector, struct rb_node **ret_parent,
1360 struct rb_node ***rb_link)
1362 struct rb_node **p, *parent;
1363 struct cfq_queue *cfqq = NULL;
1365 parent = NULL;
1366 p = &root->rb_node;
1367 while (*p) {
1368 struct rb_node **n;
1370 parent = *p;
1371 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1374 * Sort strictly based on sector. Smallest to the left,
1375 * largest to the right.
1377 if (sector > blk_rq_pos(cfqq->next_rq))
1378 n = &(*p)->rb_right;
1379 else if (sector < blk_rq_pos(cfqq->next_rq))
1380 n = &(*p)->rb_left;
1381 else
1382 break;
1383 p = n;
1384 cfqq = NULL;
1387 *ret_parent = parent;
1388 if (rb_link)
1389 *rb_link = p;
1390 return cfqq;
1393 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1395 struct rb_node **p, *parent;
1396 struct cfq_queue *__cfqq;
1398 if (cfqq->p_root) {
1399 rb_erase(&cfqq->p_node, cfqq->p_root);
1400 cfqq->p_root = NULL;
1403 if (cfq_class_idle(cfqq))
1404 return;
1405 if (!cfqq->next_rq)
1406 return;
1408 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1409 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1410 blk_rq_pos(cfqq->next_rq), &parent, &p);
1411 if (!__cfqq) {
1412 rb_link_node(&cfqq->p_node, parent, p);
1413 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1414 } else
1415 cfqq->p_root = NULL;
1419 * Update cfqq's position in the service tree.
1421 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1424 * Resorting requires the cfqq to be on the RR list already.
1426 if (cfq_cfqq_on_rr(cfqq)) {
1427 cfq_service_tree_add(cfqd, cfqq, 0);
1428 cfq_prio_tree_add(cfqd, cfqq);
1433 * add to busy list of queues for service, trying to be fair in ordering
1434 * the pending list according to last request service
1436 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1438 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1439 BUG_ON(cfq_cfqq_on_rr(cfqq));
1440 cfq_mark_cfqq_on_rr(cfqq);
1441 cfqd->busy_queues++;
1442 if (cfq_cfqq_sync(cfqq))
1443 cfqd->busy_sync_queues++;
1445 cfq_resort_rr_list(cfqd, cfqq);
1449 * Called when the cfqq no longer has requests pending, remove it from
1450 * the service tree.
1452 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1454 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1455 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1456 cfq_clear_cfqq_on_rr(cfqq);
1458 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1459 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1460 cfqq->service_tree = NULL;
1462 if (cfqq->p_root) {
1463 rb_erase(&cfqq->p_node, cfqq->p_root);
1464 cfqq->p_root = NULL;
1467 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1468 BUG_ON(!cfqd->busy_queues);
1469 cfqd->busy_queues--;
1470 if (cfq_cfqq_sync(cfqq))
1471 cfqd->busy_sync_queues--;
1475 * rb tree support functions
1477 static void cfq_del_rq_rb(struct request *rq)
1479 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1480 const int sync = rq_is_sync(rq);
1482 BUG_ON(!cfqq->queued[sync]);
1483 cfqq->queued[sync]--;
1485 elv_rb_del(&cfqq->sort_list, rq);
1487 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1489 * Queue will be deleted from service tree when we actually
1490 * expire it later. Right now just remove it from prio tree
1491 * as it is empty.
1493 if (cfqq->p_root) {
1494 rb_erase(&cfqq->p_node, cfqq->p_root);
1495 cfqq->p_root = NULL;
1500 static void cfq_add_rq_rb(struct request *rq)
1502 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1503 struct cfq_data *cfqd = cfqq->cfqd;
1504 struct request *__alias, *prev;
1506 cfqq->queued[rq_is_sync(rq)]++;
1509 * looks a little odd, but the first insert might return an alias.
1510 * if that happens, put the alias on the dispatch list
1512 while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1513 cfq_dispatch_insert(cfqd->queue, __alias);
1515 if (!cfq_cfqq_on_rr(cfqq))
1516 cfq_add_cfqq_rr(cfqd, cfqq);
1519 * check if this request is a better next-serve candidate
1521 prev = cfqq->next_rq;
1522 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1525 * adjust priority tree position, if ->next_rq changes
1527 if (prev != cfqq->next_rq)
1528 cfq_prio_tree_add(cfqd, cfqq);
1530 BUG_ON(!cfqq->next_rq);
1533 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1535 elv_rb_del(&cfqq->sort_list, rq);
1536 cfqq->queued[rq_is_sync(rq)]--;
1537 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1538 rq_data_dir(rq), rq_is_sync(rq));
1539 cfq_add_rq_rb(rq);
1540 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1541 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1542 rq_is_sync(rq));
1545 static struct request *
1546 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1548 struct task_struct *tsk = current;
1549 struct cfq_io_context *cic;
1550 struct cfq_queue *cfqq;
1552 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1553 if (!cic)
1554 return NULL;
1556 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1557 if (cfqq) {
1558 sector_t sector = bio->bi_sector + bio_sectors(bio);
1560 return elv_rb_find(&cfqq->sort_list, sector);
1563 return NULL;
1566 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1568 struct cfq_data *cfqd = q->elevator->elevator_data;
1570 cfqd->rq_in_driver++;
1571 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1572 cfqd->rq_in_driver);
1574 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1577 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1579 struct cfq_data *cfqd = q->elevator->elevator_data;
1581 WARN_ON(!cfqd->rq_in_driver);
1582 cfqd->rq_in_driver--;
1583 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1584 cfqd->rq_in_driver);
1587 static void cfq_remove_request(struct request *rq)
1589 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1591 if (cfqq->next_rq == rq)
1592 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1594 list_del_init(&rq->queuelist);
1595 cfq_del_rq_rb(rq);
1597 cfqq->cfqd->rq_queued--;
1598 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1599 rq_data_dir(rq), rq_is_sync(rq));
1600 if (rq->cmd_flags & REQ_META) {
1601 WARN_ON(!cfqq->meta_pending);
1602 cfqq->meta_pending--;
1606 static int cfq_merge(struct request_queue *q, struct request **req,
1607 struct bio *bio)
1609 struct cfq_data *cfqd = q->elevator->elevator_data;
1610 struct request *__rq;
1612 __rq = cfq_find_rq_fmerge(cfqd, bio);
1613 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1614 *req = __rq;
1615 return ELEVATOR_FRONT_MERGE;
1618 return ELEVATOR_NO_MERGE;
1621 static void cfq_merged_request(struct request_queue *q, struct request *req,
1622 int type)
1624 if (type == ELEVATOR_FRONT_MERGE) {
1625 struct cfq_queue *cfqq = RQ_CFQQ(req);
1627 cfq_reposition_rq_rb(cfqq, req);
1631 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1632 struct bio *bio)
1634 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1635 bio_data_dir(bio), cfq_bio_sync(bio));
1638 static void
1639 cfq_merged_requests(struct request_queue *q, struct request *rq,
1640 struct request *next)
1642 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1644 * reposition in fifo if next is older than rq
1646 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1647 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1648 list_move(&rq->queuelist, &next->queuelist);
1649 rq_set_fifo_time(rq, rq_fifo_time(next));
1652 if (cfqq->next_rq == next)
1653 cfqq->next_rq = rq;
1654 cfq_remove_request(next);
1655 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1656 rq_data_dir(next), rq_is_sync(next));
1659 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1660 struct bio *bio)
1662 struct cfq_data *cfqd = q->elevator->elevator_data;
1663 struct cfq_io_context *cic;
1664 struct cfq_queue *cfqq;
1667 * Disallow merge of a sync bio into an async request.
1669 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1670 return false;
1673 * Lookup the cfqq that this bio will be queued with. Allow
1674 * merge only if rq is queued there.
1676 cic = cfq_cic_lookup(cfqd, current->io_context);
1677 if (!cic)
1678 return false;
1680 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1681 return cfqq == RQ_CFQQ(rq);
1684 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1686 del_timer(&cfqd->idle_slice_timer);
1687 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1690 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1691 struct cfq_queue *cfqq)
1693 if (cfqq) {
1694 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1695 cfqd->serving_prio, cfqd->serving_type);
1696 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1697 cfqq->slice_start = 0;
1698 cfqq->dispatch_start = jiffies;
1699 cfqq->allocated_slice = 0;
1700 cfqq->slice_end = 0;
1701 cfqq->slice_dispatch = 0;
1702 cfqq->nr_sectors = 0;
1704 cfq_clear_cfqq_wait_request(cfqq);
1705 cfq_clear_cfqq_must_dispatch(cfqq);
1706 cfq_clear_cfqq_must_alloc_slice(cfqq);
1707 cfq_clear_cfqq_fifo_expire(cfqq);
1708 cfq_mark_cfqq_slice_new(cfqq);
1710 cfq_del_timer(cfqd, cfqq);
1713 cfqd->active_queue = cfqq;
1717 * current cfqq expired its slice (or was too idle), select new one
1719 static void
1720 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1721 bool timed_out)
1723 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1725 if (cfq_cfqq_wait_request(cfqq))
1726 cfq_del_timer(cfqd, cfqq);
1728 cfq_clear_cfqq_wait_request(cfqq);
1729 cfq_clear_cfqq_wait_busy(cfqq);
1732 * If this cfqq is shared between multiple processes, check to
1733 * make sure that those processes are still issuing I/Os within
1734 * the mean seek distance. If not, it may be time to break the
1735 * queues apart again.
1737 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1738 cfq_mark_cfqq_split_coop(cfqq);
1741 * store what was left of this slice, if the queue idled/timed out
1743 if (timed_out) {
1744 if (cfq_cfqq_slice_new(cfqq))
1745 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1746 else
1747 cfqq->slice_resid = cfqq->slice_end - jiffies;
1748 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1751 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1753 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1754 cfq_del_cfqq_rr(cfqd, cfqq);
1756 cfq_resort_rr_list(cfqd, cfqq);
1758 if (cfqq == cfqd->active_queue)
1759 cfqd->active_queue = NULL;
1761 if (cfqd->active_cic) {
1762 put_io_context(cfqd->active_cic->ioc);
1763 cfqd->active_cic = NULL;
1767 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1769 struct cfq_queue *cfqq = cfqd->active_queue;
1771 if (cfqq)
1772 __cfq_slice_expired(cfqd, cfqq, timed_out);
1776 * Get next queue for service. Unless we have a queue preemption,
1777 * we'll simply select the first cfqq in the service tree.
1779 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1781 struct cfq_rb_root *service_tree =
1782 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1783 cfqd->serving_type);
1785 if (!cfqd->rq_queued)
1786 return NULL;
1788 /* There is nothing to dispatch */
1789 if (!service_tree)
1790 return NULL;
1791 if (RB_EMPTY_ROOT(&service_tree->rb))
1792 return NULL;
1793 return cfq_rb_first(service_tree);
1796 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1798 struct cfq_group *cfqg;
1799 struct cfq_queue *cfqq;
1800 int i, j;
1801 struct cfq_rb_root *st;
1803 if (!cfqd->rq_queued)
1804 return NULL;
1806 cfqg = cfq_get_next_cfqg(cfqd);
1807 if (!cfqg)
1808 return NULL;
1810 for_each_cfqg_st(cfqg, i, j, st)
1811 if ((cfqq = cfq_rb_first(st)) != NULL)
1812 return cfqq;
1813 return NULL;
1817 * Get and set a new active queue for service.
1819 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1820 struct cfq_queue *cfqq)
1822 if (!cfqq)
1823 cfqq = cfq_get_next_queue(cfqd);
1825 __cfq_set_active_queue(cfqd, cfqq);
1826 return cfqq;
1829 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1830 struct request *rq)
1832 if (blk_rq_pos(rq) >= cfqd->last_position)
1833 return blk_rq_pos(rq) - cfqd->last_position;
1834 else
1835 return cfqd->last_position - blk_rq_pos(rq);
1838 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1839 struct request *rq)
1841 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1844 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1845 struct cfq_queue *cur_cfqq)
1847 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1848 struct rb_node *parent, *node;
1849 struct cfq_queue *__cfqq;
1850 sector_t sector = cfqd->last_position;
1852 if (RB_EMPTY_ROOT(root))
1853 return NULL;
1856 * First, if we find a request starting at the end of the last
1857 * request, choose it.
1859 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1860 if (__cfqq)
1861 return __cfqq;
1864 * If the exact sector wasn't found, the parent of the NULL leaf
1865 * will contain the closest sector.
1867 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1868 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1869 return __cfqq;
1871 if (blk_rq_pos(__cfqq->next_rq) < sector)
1872 node = rb_next(&__cfqq->p_node);
1873 else
1874 node = rb_prev(&__cfqq->p_node);
1875 if (!node)
1876 return NULL;
1878 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1879 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1880 return __cfqq;
1882 return NULL;
1886 * cfqd - obvious
1887 * cur_cfqq - passed in so that we don't decide that the current queue is
1888 * closely cooperating with itself.
1890 * So, basically we're assuming that that cur_cfqq has dispatched at least
1891 * one request, and that cfqd->last_position reflects a position on the disk
1892 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1893 * assumption.
1895 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1896 struct cfq_queue *cur_cfqq)
1898 struct cfq_queue *cfqq;
1900 if (cfq_class_idle(cur_cfqq))
1901 return NULL;
1902 if (!cfq_cfqq_sync(cur_cfqq))
1903 return NULL;
1904 if (CFQQ_SEEKY(cur_cfqq))
1905 return NULL;
1908 * Don't search priority tree if it's the only queue in the group.
1910 if (cur_cfqq->cfqg->nr_cfqq == 1)
1911 return NULL;
1914 * We should notice if some of the queues are cooperating, eg
1915 * working closely on the same area of the disk. In that case,
1916 * we can group them together and don't waste time idling.
1918 cfqq = cfqq_close(cfqd, cur_cfqq);
1919 if (!cfqq)
1920 return NULL;
1922 /* If new queue belongs to different cfq_group, don't choose it */
1923 if (cur_cfqq->cfqg != cfqq->cfqg)
1924 return NULL;
1927 * It only makes sense to merge sync queues.
1929 if (!cfq_cfqq_sync(cfqq))
1930 return NULL;
1931 if (CFQQ_SEEKY(cfqq))
1932 return NULL;
1935 * Do not merge queues of different priority classes
1937 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1938 return NULL;
1940 return cfqq;
1944 * Determine whether we should enforce idle window for this queue.
1947 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1949 enum wl_prio_t prio = cfqq_prio(cfqq);
1950 struct cfq_rb_root *service_tree = cfqq->service_tree;
1952 BUG_ON(!service_tree);
1953 BUG_ON(!service_tree->count);
1955 if (!cfqd->cfq_slice_idle)
1956 return false;
1958 /* We never do for idle class queues. */
1959 if (prio == IDLE_WORKLOAD)
1960 return false;
1962 /* We do for queues that were marked with idle window flag. */
1963 if (cfq_cfqq_idle_window(cfqq) &&
1964 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1965 return true;
1968 * Otherwise, we do only if they are the last ones
1969 * in their service tree.
1971 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1972 return true;
1973 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1974 service_tree->count);
1975 return false;
1978 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1980 struct cfq_queue *cfqq = cfqd->active_queue;
1981 struct cfq_io_context *cic;
1982 unsigned long sl, group_idle = 0;
1985 * SSD device without seek penalty, disable idling. But only do so
1986 * for devices that support queuing, otherwise we still have a problem
1987 * with sync vs async workloads.
1989 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1990 return;
1992 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1993 WARN_ON(cfq_cfqq_slice_new(cfqq));
1996 * idle is disabled, either manually or by past process history
1998 if (!cfq_should_idle(cfqd, cfqq)) {
1999 /* no queue idling. Check for group idling */
2000 if (cfqd->cfq_group_idle)
2001 group_idle = cfqd->cfq_group_idle;
2002 else
2003 return;
2007 * still active requests from this queue, don't idle
2009 if (cfqq->dispatched)
2010 return;
2013 * task has exited, don't wait
2015 cic = cfqd->active_cic;
2016 if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2017 return;
2020 * If our average think time is larger than the remaining time
2021 * slice, then don't idle. This avoids overrunning the allotted
2022 * time slice.
2024 if (sample_valid(cic->ttime_samples) &&
2025 (cfqq->slice_end - jiffies < cic->ttime_mean)) {
2026 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
2027 cic->ttime_mean);
2028 return;
2031 /* There are other queues in the group, don't do group idle */
2032 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2033 return;
2035 cfq_mark_cfqq_wait_request(cfqq);
2037 if (group_idle)
2038 sl = cfqd->cfq_group_idle;
2039 else
2040 sl = cfqd->cfq_slice_idle;
2042 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2043 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2044 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2045 group_idle ? 1 : 0);
2049 * Move request from internal lists to the request queue dispatch list.
2051 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2053 struct cfq_data *cfqd = q->elevator->elevator_data;
2054 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2056 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2058 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2059 cfq_remove_request(rq);
2060 cfqq->dispatched++;
2061 (RQ_CFQG(rq))->dispatched++;
2062 elv_dispatch_sort(q, rq);
2064 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2065 cfqq->nr_sectors += blk_rq_sectors(rq);
2066 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2067 rq_data_dir(rq), rq_is_sync(rq));
2071 * return expired entry, or NULL to just start from scratch in rbtree
2073 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2075 struct request *rq = NULL;
2077 if (cfq_cfqq_fifo_expire(cfqq))
2078 return NULL;
2080 cfq_mark_cfqq_fifo_expire(cfqq);
2082 if (list_empty(&cfqq->fifo))
2083 return NULL;
2085 rq = rq_entry_fifo(cfqq->fifo.next);
2086 if (time_before(jiffies, rq_fifo_time(rq)))
2087 rq = NULL;
2089 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2090 return rq;
2093 static inline int
2094 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2096 const int base_rq = cfqd->cfq_slice_async_rq;
2098 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2100 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2104 * Must be called with the queue_lock held.
2106 static int cfqq_process_refs(struct cfq_queue *cfqq)
2108 int process_refs, io_refs;
2110 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2111 process_refs = cfqq->ref - io_refs;
2112 BUG_ON(process_refs < 0);
2113 return process_refs;
2116 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2118 int process_refs, new_process_refs;
2119 struct cfq_queue *__cfqq;
2122 * If there are no process references on the new_cfqq, then it is
2123 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2124 * chain may have dropped their last reference (not just their
2125 * last process reference).
2127 if (!cfqq_process_refs(new_cfqq))
2128 return;
2130 /* Avoid a circular list and skip interim queue merges */
2131 while ((__cfqq = new_cfqq->new_cfqq)) {
2132 if (__cfqq == cfqq)
2133 return;
2134 new_cfqq = __cfqq;
2137 process_refs = cfqq_process_refs(cfqq);
2138 new_process_refs = cfqq_process_refs(new_cfqq);
2140 * If the process for the cfqq has gone away, there is no
2141 * sense in merging the queues.
2143 if (process_refs == 0 || new_process_refs == 0)
2144 return;
2147 * Merge in the direction of the lesser amount of work.
2149 if (new_process_refs >= process_refs) {
2150 cfqq->new_cfqq = new_cfqq;
2151 new_cfqq->ref += process_refs;
2152 } else {
2153 new_cfqq->new_cfqq = cfqq;
2154 cfqq->ref += new_process_refs;
2158 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2159 struct cfq_group *cfqg, enum wl_prio_t prio)
2161 struct cfq_queue *queue;
2162 int i;
2163 bool key_valid = false;
2164 unsigned long lowest_key = 0;
2165 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2167 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2168 /* select the one with lowest rb_key */
2169 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2170 if (queue &&
2171 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2172 lowest_key = queue->rb_key;
2173 cur_best = i;
2174 key_valid = true;
2178 return cur_best;
2181 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2183 unsigned slice;
2184 unsigned count;
2185 struct cfq_rb_root *st;
2186 unsigned group_slice;
2187 enum wl_prio_t original_prio = cfqd->serving_prio;
2189 /* Choose next priority. RT > BE > IDLE */
2190 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2191 cfqd->serving_prio = RT_WORKLOAD;
2192 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2193 cfqd->serving_prio = BE_WORKLOAD;
2194 else {
2195 cfqd->serving_prio = IDLE_WORKLOAD;
2196 cfqd->workload_expires = jiffies + 1;
2197 return;
2200 if (original_prio != cfqd->serving_prio)
2201 goto new_workload;
2204 * For RT and BE, we have to choose also the type
2205 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2206 * expiration time
2208 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2209 count = st->count;
2212 * check workload expiration, and that we still have other queues ready
2214 if (count && !time_after(jiffies, cfqd->workload_expires))
2215 return;
2217 new_workload:
2218 /* otherwise select new workload type */
2219 cfqd->serving_type =
2220 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2221 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2222 count = st->count;
2225 * the workload slice is computed as a fraction of target latency
2226 * proportional to the number of queues in that workload, over
2227 * all the queues in the same priority class
2229 group_slice = cfq_group_slice(cfqd, cfqg);
2231 slice = group_slice * count /
2232 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2233 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2235 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2236 unsigned int tmp;
2239 * Async queues are currently system wide. Just taking
2240 * proportion of queues with-in same group will lead to higher
2241 * async ratio system wide as generally root group is going
2242 * to have higher weight. A more accurate thing would be to
2243 * calculate system wide asnc/sync ratio.
2245 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2246 tmp = tmp/cfqd->busy_queues;
2247 slice = min_t(unsigned, slice, tmp);
2249 /* async workload slice is scaled down according to
2250 * the sync/async slice ratio. */
2251 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2252 } else
2253 /* sync workload slice is at least 2 * cfq_slice_idle */
2254 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2256 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2257 cfq_log(cfqd, "workload slice:%d", slice);
2258 cfqd->workload_expires = jiffies + slice;
2261 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2263 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2264 struct cfq_group *cfqg;
2266 if (RB_EMPTY_ROOT(&st->rb))
2267 return NULL;
2268 cfqg = cfq_rb_first_group(st);
2269 update_min_vdisktime(st);
2270 return cfqg;
2273 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2275 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2277 cfqd->serving_group = cfqg;
2279 /* Restore the workload type data */
2280 if (cfqg->saved_workload_slice) {
2281 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2282 cfqd->serving_type = cfqg->saved_workload;
2283 cfqd->serving_prio = cfqg->saved_serving_prio;
2284 } else
2285 cfqd->workload_expires = jiffies - 1;
2287 choose_service_tree(cfqd, cfqg);
2291 * Select a queue for service. If we have a current active queue,
2292 * check whether to continue servicing it, or retrieve and set a new one.
2294 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2296 struct cfq_queue *cfqq, *new_cfqq = NULL;
2298 cfqq = cfqd->active_queue;
2299 if (!cfqq)
2300 goto new_queue;
2302 if (!cfqd->rq_queued)
2303 return NULL;
2306 * We were waiting for group to get backlogged. Expire the queue
2308 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2309 goto expire;
2312 * The active queue has run out of time, expire it and select new.
2314 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2316 * If slice had not expired at the completion of last request
2317 * we might not have turned on wait_busy flag. Don't expire
2318 * the queue yet. Allow the group to get backlogged.
2320 * The very fact that we have used the slice, that means we
2321 * have been idling all along on this queue and it should be
2322 * ok to wait for this request to complete.
2324 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2325 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2326 cfqq = NULL;
2327 goto keep_queue;
2328 } else
2329 goto check_group_idle;
2333 * The active queue has requests and isn't expired, allow it to
2334 * dispatch.
2336 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2337 goto keep_queue;
2340 * If another queue has a request waiting within our mean seek
2341 * distance, let it run. The expire code will check for close
2342 * cooperators and put the close queue at the front of the service
2343 * tree. If possible, merge the expiring queue with the new cfqq.
2345 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2346 if (new_cfqq) {
2347 if (!cfqq->new_cfqq)
2348 cfq_setup_merge(cfqq, new_cfqq);
2349 goto expire;
2353 * No requests pending. If the active queue still has requests in
2354 * flight or is idling for a new request, allow either of these
2355 * conditions to happen (or time out) before selecting a new queue.
2357 if (timer_pending(&cfqd->idle_slice_timer)) {
2358 cfqq = NULL;
2359 goto keep_queue;
2363 * This is a deep seek queue, but the device is much faster than
2364 * the queue can deliver, don't idle
2366 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2367 (cfq_cfqq_slice_new(cfqq) ||
2368 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2369 cfq_clear_cfqq_deep(cfqq);
2370 cfq_clear_cfqq_idle_window(cfqq);
2373 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2374 cfqq = NULL;
2375 goto keep_queue;
2379 * If group idle is enabled and there are requests dispatched from
2380 * this group, wait for requests to complete.
2382 check_group_idle:
2383 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2384 && cfqq->cfqg->dispatched) {
2385 cfqq = NULL;
2386 goto keep_queue;
2389 expire:
2390 cfq_slice_expired(cfqd, 0);
2391 new_queue:
2393 * Current queue expired. Check if we have to switch to a new
2394 * service tree
2396 if (!new_cfqq)
2397 cfq_choose_cfqg(cfqd);
2399 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2400 keep_queue:
2401 return cfqq;
2404 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2406 int dispatched = 0;
2408 while (cfqq->next_rq) {
2409 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2410 dispatched++;
2413 BUG_ON(!list_empty(&cfqq->fifo));
2415 /* By default cfqq is not expired if it is empty. Do it explicitly */
2416 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2417 return dispatched;
2421 * Drain our current requests. Used for barriers and when switching
2422 * io schedulers on-the-fly.
2424 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2426 struct cfq_queue *cfqq;
2427 int dispatched = 0;
2429 /* Expire the timeslice of the current active queue first */
2430 cfq_slice_expired(cfqd, 0);
2431 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2432 __cfq_set_active_queue(cfqd, cfqq);
2433 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2436 BUG_ON(cfqd->busy_queues);
2438 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2439 return dispatched;
2442 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2443 struct cfq_queue *cfqq)
2445 /* the queue hasn't finished any request, can't estimate */
2446 if (cfq_cfqq_slice_new(cfqq))
2447 return true;
2448 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2449 cfqq->slice_end))
2450 return true;
2452 return false;
2455 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2457 unsigned int max_dispatch;
2460 * Drain async requests before we start sync IO
2462 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2463 return false;
2466 * If this is an async queue and we have sync IO in flight, let it wait
2468 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2469 return false;
2471 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2472 if (cfq_class_idle(cfqq))
2473 max_dispatch = 1;
2476 * Does this cfqq already have too much IO in flight?
2478 if (cfqq->dispatched >= max_dispatch) {
2479 bool promote_sync = false;
2481 * idle queue must always only have a single IO in flight
2483 if (cfq_class_idle(cfqq))
2484 return false;
2487 * If there is only one sync queue
2488 * we can ignore async queue here and give the sync
2489 * queue no dispatch limit. The reason is a sync queue can
2490 * preempt async queue, limiting the sync queue doesn't make
2491 * sense. This is useful for aiostress test.
2493 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2494 promote_sync = true;
2497 * We have other queues, don't allow more IO from this one
2499 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2500 !promote_sync)
2501 return false;
2504 * Sole queue user, no limit
2506 if (cfqd->busy_queues == 1 || promote_sync)
2507 max_dispatch = -1;
2508 else
2510 * Normally we start throttling cfqq when cfq_quantum/2
2511 * requests have been dispatched. But we can drive
2512 * deeper queue depths at the beginning of slice
2513 * subjected to upper limit of cfq_quantum.
2514 * */
2515 max_dispatch = cfqd->cfq_quantum;
2519 * Async queues must wait a bit before being allowed dispatch.
2520 * We also ramp up the dispatch depth gradually for async IO,
2521 * based on the last sync IO we serviced
2523 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2524 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2525 unsigned int depth;
2527 depth = last_sync / cfqd->cfq_slice[1];
2528 if (!depth && !cfqq->dispatched)
2529 depth = 1;
2530 if (depth < max_dispatch)
2531 max_dispatch = depth;
2535 * If we're below the current max, allow a dispatch
2537 return cfqq->dispatched < max_dispatch;
2541 * Dispatch a request from cfqq, moving them to the request queue
2542 * dispatch list.
2544 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2546 struct request *rq;
2548 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2550 if (!cfq_may_dispatch(cfqd, cfqq))
2551 return false;
2554 * follow expired path, else get first next available
2556 rq = cfq_check_fifo(cfqq);
2557 if (!rq)
2558 rq = cfqq->next_rq;
2561 * insert request into driver dispatch list
2563 cfq_dispatch_insert(cfqd->queue, rq);
2565 if (!cfqd->active_cic) {
2566 struct cfq_io_context *cic = RQ_CIC(rq);
2568 atomic_long_inc(&cic->ioc->refcount);
2569 cfqd->active_cic = cic;
2572 return true;
2576 * Find the cfqq that we need to service and move a request from that to the
2577 * dispatch list
2579 static int cfq_dispatch_requests(struct request_queue *q, int force)
2581 struct cfq_data *cfqd = q->elevator->elevator_data;
2582 struct cfq_queue *cfqq;
2584 if (!cfqd->busy_queues)
2585 return 0;
2587 if (unlikely(force))
2588 return cfq_forced_dispatch(cfqd);
2590 cfqq = cfq_select_queue(cfqd);
2591 if (!cfqq)
2592 return 0;
2595 * Dispatch a request from this cfqq, if it is allowed
2597 if (!cfq_dispatch_request(cfqd, cfqq))
2598 return 0;
2600 cfqq->slice_dispatch++;
2601 cfq_clear_cfqq_must_dispatch(cfqq);
2604 * expire an async queue immediately if it has used up its slice. idle
2605 * queue always expire after 1 dispatch round.
2607 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2608 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2609 cfq_class_idle(cfqq))) {
2610 cfqq->slice_end = jiffies + 1;
2611 cfq_slice_expired(cfqd, 0);
2614 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2615 return 1;
2619 * task holds one reference to the queue, dropped when task exits. each rq
2620 * in-flight on this queue also holds a reference, dropped when rq is freed.
2622 * Each cfq queue took a reference on the parent group. Drop it now.
2623 * queue lock must be held here.
2625 static void cfq_put_queue(struct cfq_queue *cfqq)
2627 struct cfq_data *cfqd = cfqq->cfqd;
2628 struct cfq_group *cfqg;
2630 BUG_ON(cfqq->ref <= 0);
2632 cfqq->ref--;
2633 if (cfqq->ref)
2634 return;
2636 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2637 BUG_ON(rb_first(&cfqq->sort_list));
2638 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2639 cfqg = cfqq->cfqg;
2641 if (unlikely(cfqd->active_queue == cfqq)) {
2642 __cfq_slice_expired(cfqd, cfqq, 0);
2643 cfq_schedule_dispatch(cfqd);
2646 BUG_ON(cfq_cfqq_on_rr(cfqq));
2647 kmem_cache_free(cfq_pool, cfqq);
2648 cfq_put_cfqg(cfqg);
2652 * Call func for each cic attached to this ioc.
2654 static void
2655 call_for_each_cic(struct io_context *ioc,
2656 void (*func)(struct io_context *, struct cfq_io_context *))
2658 struct cfq_io_context *cic;
2659 struct hlist_node *n;
2661 rcu_read_lock();
2663 hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2664 func(ioc, cic);
2666 rcu_read_unlock();
2669 static void cfq_cic_free_rcu(struct rcu_head *head)
2671 struct cfq_io_context *cic;
2673 cic = container_of(head, struct cfq_io_context, rcu_head);
2675 kmem_cache_free(cfq_ioc_pool, cic);
2676 elv_ioc_count_dec(cfq_ioc_count);
2678 if (ioc_gone) {
2680 * CFQ scheduler is exiting, grab exit lock and check
2681 * the pending io context count. If it hits zero,
2682 * complete ioc_gone and set it back to NULL
2684 spin_lock(&ioc_gone_lock);
2685 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2686 complete(ioc_gone);
2687 ioc_gone = NULL;
2689 spin_unlock(&ioc_gone_lock);
2693 static void cfq_cic_free(struct cfq_io_context *cic)
2695 call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2698 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2700 unsigned long flags;
2701 unsigned long dead_key = (unsigned long) cic->key;
2703 BUG_ON(!(dead_key & CIC_DEAD_KEY));
2705 spin_lock_irqsave(&ioc->lock, flags);
2706 radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2707 hlist_del_rcu(&cic->cic_list);
2708 spin_unlock_irqrestore(&ioc->lock, flags);
2710 cfq_cic_free(cic);
2714 * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2715 * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2716 * and ->trim() which is called with the task lock held
2718 static void cfq_free_io_context(struct io_context *ioc)
2721 * ioc->refcount is zero here, or we are called from elv_unregister(),
2722 * so no more cic's are allowed to be linked into this ioc. So it
2723 * should be ok to iterate over the known list, we will see all cic's
2724 * since no new ones are added.
2726 call_for_each_cic(ioc, cic_free_func);
2729 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2731 struct cfq_queue *__cfqq, *next;
2734 * If this queue was scheduled to merge with another queue, be
2735 * sure to drop the reference taken on that queue (and others in
2736 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2738 __cfqq = cfqq->new_cfqq;
2739 while (__cfqq) {
2740 if (__cfqq == cfqq) {
2741 WARN(1, "cfqq->new_cfqq loop detected\n");
2742 break;
2744 next = __cfqq->new_cfqq;
2745 cfq_put_queue(__cfqq);
2746 __cfqq = next;
2750 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2752 if (unlikely(cfqq == cfqd->active_queue)) {
2753 __cfq_slice_expired(cfqd, cfqq, 0);
2754 cfq_schedule_dispatch(cfqd);
2757 cfq_put_cooperator(cfqq);
2759 cfq_put_queue(cfqq);
2762 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2763 struct cfq_io_context *cic)
2765 struct io_context *ioc = cic->ioc;
2767 list_del_init(&cic->queue_list);
2770 * Make sure dead mark is seen for dead queues
2772 smp_wmb();
2773 cic->key = cfqd_dead_key(cfqd);
2775 if (ioc->ioc_data == cic)
2776 rcu_assign_pointer(ioc->ioc_data, NULL);
2778 if (cic->cfqq[BLK_RW_ASYNC]) {
2779 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2780 cic->cfqq[BLK_RW_ASYNC] = NULL;
2783 if (cic->cfqq[BLK_RW_SYNC]) {
2784 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2785 cic->cfqq[BLK_RW_SYNC] = NULL;
2789 static void cfq_exit_single_io_context(struct io_context *ioc,
2790 struct cfq_io_context *cic)
2792 struct cfq_data *cfqd = cic_to_cfqd(cic);
2794 if (cfqd) {
2795 struct request_queue *q = cfqd->queue;
2796 unsigned long flags;
2798 spin_lock_irqsave(q->queue_lock, flags);
2801 * Ensure we get a fresh copy of the ->key to prevent
2802 * race between exiting task and queue
2804 smp_read_barrier_depends();
2805 if (cic->key == cfqd)
2806 __cfq_exit_single_io_context(cfqd, cic);
2808 spin_unlock_irqrestore(q->queue_lock, flags);
2813 * The process that ioc belongs to has exited, we need to clean up
2814 * and put the internal structures we have that belongs to that process.
2816 static void cfq_exit_io_context(struct io_context *ioc)
2818 call_for_each_cic(ioc, cfq_exit_single_io_context);
2821 static struct cfq_io_context *
2822 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2824 struct cfq_io_context *cic;
2826 cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2827 cfqd->queue->node);
2828 if (cic) {
2829 cic->last_end_request = jiffies;
2830 INIT_LIST_HEAD(&cic->queue_list);
2831 INIT_HLIST_NODE(&cic->cic_list);
2832 cic->dtor = cfq_free_io_context;
2833 cic->exit = cfq_exit_io_context;
2834 elv_ioc_count_inc(cfq_ioc_count);
2837 return cic;
2840 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2842 struct task_struct *tsk = current;
2843 int ioprio_class;
2845 if (!cfq_cfqq_prio_changed(cfqq))
2846 return;
2848 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2849 switch (ioprio_class) {
2850 default:
2851 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2852 case IOPRIO_CLASS_NONE:
2854 * no prio set, inherit CPU scheduling settings
2856 cfqq->ioprio = task_nice_ioprio(tsk);
2857 cfqq->ioprio_class = task_nice_ioclass(tsk);
2858 break;
2859 case IOPRIO_CLASS_RT:
2860 cfqq->ioprio = task_ioprio(ioc);
2861 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2862 break;
2863 case IOPRIO_CLASS_BE:
2864 cfqq->ioprio = task_ioprio(ioc);
2865 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2866 break;
2867 case IOPRIO_CLASS_IDLE:
2868 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2869 cfqq->ioprio = 7;
2870 cfq_clear_cfqq_idle_window(cfqq);
2871 break;
2875 * keep track of original prio settings in case we have to temporarily
2876 * elevate the priority of this queue
2878 cfqq->org_ioprio = cfqq->ioprio;
2879 cfqq->org_ioprio_class = cfqq->ioprio_class;
2880 cfq_clear_cfqq_prio_changed(cfqq);
2883 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2885 struct cfq_data *cfqd = cic_to_cfqd(cic);
2886 struct cfq_queue *cfqq;
2887 unsigned long flags;
2889 if (unlikely(!cfqd))
2890 return;
2892 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2894 cfqq = cic->cfqq[BLK_RW_ASYNC];
2895 if (cfqq) {
2896 struct cfq_queue *new_cfqq;
2897 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2898 GFP_ATOMIC);
2899 if (new_cfqq) {
2900 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2901 cfq_put_queue(cfqq);
2905 cfqq = cic->cfqq[BLK_RW_SYNC];
2906 if (cfqq)
2907 cfq_mark_cfqq_prio_changed(cfqq);
2909 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2912 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2914 call_for_each_cic(ioc, changed_ioprio);
2915 ioc->ioprio_changed = 0;
2918 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2919 pid_t pid, bool is_sync)
2921 RB_CLEAR_NODE(&cfqq->rb_node);
2922 RB_CLEAR_NODE(&cfqq->p_node);
2923 INIT_LIST_HEAD(&cfqq->fifo);
2925 cfqq->ref = 0;
2926 cfqq->cfqd = cfqd;
2928 cfq_mark_cfqq_prio_changed(cfqq);
2930 if (is_sync) {
2931 if (!cfq_class_idle(cfqq))
2932 cfq_mark_cfqq_idle_window(cfqq);
2933 cfq_mark_cfqq_sync(cfqq);
2935 cfqq->pid = pid;
2938 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2939 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2941 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2942 struct cfq_data *cfqd = cic_to_cfqd(cic);
2943 unsigned long flags;
2944 struct request_queue *q;
2946 if (unlikely(!cfqd))
2947 return;
2949 q = cfqd->queue;
2951 spin_lock_irqsave(q->queue_lock, flags);
2953 if (sync_cfqq) {
2955 * Drop reference to sync queue. A new sync queue will be
2956 * assigned in new group upon arrival of a fresh request.
2958 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2959 cic_set_cfqq(cic, NULL, 1);
2960 cfq_put_queue(sync_cfqq);
2963 spin_unlock_irqrestore(q->queue_lock, flags);
2966 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2968 call_for_each_cic(ioc, changed_cgroup);
2969 ioc->cgroup_changed = 0;
2971 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2973 static struct cfq_queue *
2974 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2975 struct io_context *ioc, gfp_t gfp_mask)
2977 struct cfq_queue *cfqq, *new_cfqq = NULL;
2978 struct cfq_io_context *cic;
2979 struct cfq_group *cfqg;
2981 retry:
2982 cfqg = cfq_get_cfqg(cfqd);
2983 cic = cfq_cic_lookup(cfqd, ioc);
2984 /* cic always exists here */
2985 cfqq = cic_to_cfqq(cic, is_sync);
2988 * Always try a new alloc if we fell back to the OOM cfqq
2989 * originally, since it should just be a temporary situation.
2991 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2992 cfqq = NULL;
2993 if (new_cfqq) {
2994 cfqq = new_cfqq;
2995 new_cfqq = NULL;
2996 } else if (gfp_mask & __GFP_WAIT) {
2997 spin_unlock_irq(cfqd->queue->queue_lock);
2998 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2999 gfp_mask | __GFP_ZERO,
3000 cfqd->queue->node);
3001 spin_lock_irq(cfqd->queue->queue_lock);
3002 if (new_cfqq)
3003 goto retry;
3004 } else {
3005 cfqq = kmem_cache_alloc_node(cfq_pool,
3006 gfp_mask | __GFP_ZERO,
3007 cfqd->queue->node);
3010 if (cfqq) {
3011 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3012 cfq_init_prio_data(cfqq, ioc);
3013 cfq_link_cfqq_cfqg(cfqq, cfqg);
3014 cfq_log_cfqq(cfqd, cfqq, "alloced");
3015 } else
3016 cfqq = &cfqd->oom_cfqq;
3019 if (new_cfqq)
3020 kmem_cache_free(cfq_pool, new_cfqq);
3022 return cfqq;
3025 static struct cfq_queue **
3026 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3028 switch (ioprio_class) {
3029 case IOPRIO_CLASS_RT:
3030 return &cfqd->async_cfqq[0][ioprio];
3031 case IOPRIO_CLASS_BE:
3032 return &cfqd->async_cfqq[1][ioprio];
3033 case IOPRIO_CLASS_IDLE:
3034 return &cfqd->async_idle_cfqq;
3035 default:
3036 BUG();
3040 static struct cfq_queue *
3041 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3042 gfp_t gfp_mask)
3044 const int ioprio = task_ioprio(ioc);
3045 const int ioprio_class = task_ioprio_class(ioc);
3046 struct cfq_queue **async_cfqq = NULL;
3047 struct cfq_queue *cfqq = NULL;
3049 if (!is_sync) {
3050 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3051 cfqq = *async_cfqq;
3054 if (!cfqq)
3055 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3058 * pin the queue now that it's allocated, scheduler exit will prune it
3060 if (!is_sync && !(*async_cfqq)) {
3061 cfqq->ref++;
3062 *async_cfqq = cfqq;
3065 cfqq->ref++;
3066 return cfqq;
3070 * We drop cfq io contexts lazily, so we may find a dead one.
3072 static void
3073 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3074 struct cfq_io_context *cic)
3076 unsigned long flags;
3078 WARN_ON(!list_empty(&cic->queue_list));
3079 BUG_ON(cic->key != cfqd_dead_key(cfqd));
3081 spin_lock_irqsave(&ioc->lock, flags);
3083 BUG_ON(ioc->ioc_data == cic);
3085 radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3086 hlist_del_rcu(&cic->cic_list);
3087 spin_unlock_irqrestore(&ioc->lock, flags);
3089 cfq_cic_free(cic);
3092 static struct cfq_io_context *
3093 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3095 struct cfq_io_context *cic;
3096 unsigned long flags;
3098 if (unlikely(!ioc))
3099 return NULL;
3101 rcu_read_lock();
3104 * we maintain a last-hit cache, to avoid browsing over the tree
3106 cic = rcu_dereference(ioc->ioc_data);
3107 if (cic && cic->key == cfqd) {
3108 rcu_read_unlock();
3109 return cic;
3112 do {
3113 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3114 rcu_read_unlock();
3115 if (!cic)
3116 break;
3117 if (unlikely(cic->key != cfqd)) {
3118 cfq_drop_dead_cic(cfqd, ioc, cic);
3119 rcu_read_lock();
3120 continue;
3123 spin_lock_irqsave(&ioc->lock, flags);
3124 rcu_assign_pointer(ioc->ioc_data, cic);
3125 spin_unlock_irqrestore(&ioc->lock, flags);
3126 break;
3127 } while (1);
3129 return cic;
3133 * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3134 * the process specific cfq io context when entered from the block layer.
3135 * Also adds the cic to a per-cfqd list, used when this queue is removed.
3137 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3138 struct cfq_io_context *cic, gfp_t gfp_mask)
3140 unsigned long flags;
3141 int ret;
3143 ret = radix_tree_preload(gfp_mask);
3144 if (!ret) {
3145 cic->ioc = ioc;
3146 cic->key = cfqd;
3148 spin_lock_irqsave(&ioc->lock, flags);
3149 ret = radix_tree_insert(&ioc->radix_root,
3150 cfqd->cic_index, cic);
3151 if (!ret)
3152 hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3153 spin_unlock_irqrestore(&ioc->lock, flags);
3155 radix_tree_preload_end();
3157 if (!ret) {
3158 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3159 list_add(&cic->queue_list, &cfqd->cic_list);
3160 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3164 if (ret)
3165 printk(KERN_ERR "cfq: cic link failed!\n");
3167 return ret;
3171 * Setup general io context and cfq io context. There can be several cfq
3172 * io contexts per general io context, if this process is doing io to more
3173 * than one device managed by cfq.
3175 static struct cfq_io_context *
3176 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3178 struct io_context *ioc = NULL;
3179 struct cfq_io_context *cic;
3181 might_sleep_if(gfp_mask & __GFP_WAIT);
3183 ioc = get_io_context(gfp_mask, cfqd->queue->node);
3184 if (!ioc)
3185 return NULL;
3187 cic = cfq_cic_lookup(cfqd, ioc);
3188 if (cic)
3189 goto out;
3191 cic = cfq_alloc_io_context(cfqd, gfp_mask);
3192 if (cic == NULL)
3193 goto err;
3195 if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3196 goto err_free;
3198 out:
3199 smp_read_barrier_depends();
3200 if (unlikely(ioc->ioprio_changed))
3201 cfq_ioc_set_ioprio(ioc);
3203 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3204 if (unlikely(ioc->cgroup_changed))
3205 cfq_ioc_set_cgroup(ioc);
3206 #endif
3207 return cic;
3208 err_free:
3209 cfq_cic_free(cic);
3210 err:
3211 put_io_context(ioc);
3212 return NULL;
3215 static void
3216 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3218 unsigned long elapsed = jiffies - cic->last_end_request;
3219 unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3221 cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3222 cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3223 cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3226 static void
3227 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3228 struct request *rq)
3230 sector_t sdist = 0;
3231 sector_t n_sec = blk_rq_sectors(rq);
3232 if (cfqq->last_request_pos) {
3233 if (cfqq->last_request_pos < blk_rq_pos(rq))
3234 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3235 else
3236 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3239 cfqq->seek_history <<= 1;
3240 if (blk_queue_nonrot(cfqd->queue))
3241 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3242 else
3243 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3247 * Disable idle window if the process thinks too long or seeks so much that
3248 * it doesn't matter
3250 static void
3251 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3252 struct cfq_io_context *cic)
3254 int old_idle, enable_idle;
3257 * Don't idle for async or idle io prio class
3259 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3260 return;
3262 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3264 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3265 cfq_mark_cfqq_deep(cfqq);
3267 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3268 enable_idle = 0;
3269 else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3270 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3271 enable_idle = 0;
3272 else if (sample_valid(cic->ttime_samples)) {
3273 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3274 enable_idle = 0;
3275 else
3276 enable_idle = 1;
3279 if (old_idle != enable_idle) {
3280 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3281 if (enable_idle)
3282 cfq_mark_cfqq_idle_window(cfqq);
3283 else
3284 cfq_clear_cfqq_idle_window(cfqq);
3289 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3290 * no or if we aren't sure, a 1 will cause a preempt.
3292 static bool
3293 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3294 struct request *rq)
3296 struct cfq_queue *cfqq;
3298 cfqq = cfqd->active_queue;
3299 if (!cfqq)
3300 return false;
3302 if (cfq_class_idle(new_cfqq))
3303 return false;
3305 if (cfq_class_idle(cfqq))
3306 return true;
3309 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3311 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3312 return false;
3315 * if the new request is sync, but the currently running queue is
3316 * not, let the sync request have priority.
3318 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3319 return true;
3321 if (new_cfqq->cfqg != cfqq->cfqg)
3322 return false;
3324 if (cfq_slice_used(cfqq))
3325 return true;
3327 /* Allow preemption only if we are idling on sync-noidle tree */
3328 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3329 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3330 new_cfqq->service_tree->count == 2 &&
3331 RB_EMPTY_ROOT(&cfqq->sort_list))
3332 return true;
3335 * So both queues are sync. Let the new request get disk time if
3336 * it's a metadata request and the current queue is doing regular IO.
3338 if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3339 return true;
3342 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3344 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3345 return true;
3347 /* An idle queue should not be idle now for some reason */
3348 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3349 return true;
3351 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3352 return false;
3355 * if this request is as-good as one we would expect from the
3356 * current cfqq, let it preempt
3358 if (cfq_rq_close(cfqd, cfqq, rq))
3359 return true;
3361 return false;
3365 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3366 * let it have half of its nominal slice.
3368 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3370 struct cfq_queue *old_cfqq = cfqd->active_queue;
3372 cfq_log_cfqq(cfqd, cfqq, "preempt");
3373 cfq_slice_expired(cfqd, 1);
3376 * workload type is changed, don't save slice, otherwise preempt
3377 * doesn't happen
3379 if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3380 cfqq->cfqg->saved_workload_slice = 0;
3383 * Put the new queue at the front of the of the current list,
3384 * so we know that it will be selected next.
3386 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3388 cfq_service_tree_add(cfqd, cfqq, 1);
3390 cfqq->slice_end = 0;
3391 cfq_mark_cfqq_slice_new(cfqq);
3395 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3396 * something we should do about it
3398 static void
3399 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3400 struct request *rq)
3402 struct cfq_io_context *cic = RQ_CIC(rq);
3404 cfqd->rq_queued++;
3405 if (rq->cmd_flags & REQ_META)
3406 cfqq->meta_pending++;
3408 cfq_update_io_thinktime(cfqd, cic);
3409 cfq_update_io_seektime(cfqd, cfqq, rq);
3410 cfq_update_idle_window(cfqd, cfqq, cic);
3412 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3414 if (cfqq == cfqd->active_queue) {
3416 * Remember that we saw a request from this process, but
3417 * don't start queuing just yet. Otherwise we risk seeing lots
3418 * of tiny requests, because we disrupt the normal plugging
3419 * and merging. If the request is already larger than a single
3420 * page, let it rip immediately. For that case we assume that
3421 * merging is already done. Ditto for a busy system that
3422 * has other work pending, don't risk delaying until the
3423 * idle timer unplug to continue working.
3425 if (cfq_cfqq_wait_request(cfqq)) {
3426 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3427 cfqd->busy_queues > 1) {
3428 cfq_del_timer(cfqd, cfqq);
3429 cfq_clear_cfqq_wait_request(cfqq);
3430 __blk_run_queue(cfqd->queue);
3431 } else {
3432 cfq_blkiocg_update_idle_time_stats(
3433 &cfqq->cfqg->blkg);
3434 cfq_mark_cfqq_must_dispatch(cfqq);
3437 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3439 * not the active queue - expire current slice if it is
3440 * idle and has expired it's mean thinktime or this new queue
3441 * has some old slice time left and is of higher priority or
3442 * this new queue is RT and the current one is BE
3444 cfq_preempt_queue(cfqd, cfqq);
3445 __blk_run_queue(cfqd->queue);
3449 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3451 struct cfq_data *cfqd = q->elevator->elevator_data;
3452 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3454 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3455 cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3457 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3458 list_add_tail(&rq->queuelist, &cfqq->fifo);
3459 cfq_add_rq_rb(rq);
3460 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3461 &cfqd->serving_group->blkg, rq_data_dir(rq),
3462 rq_is_sync(rq));
3463 cfq_rq_enqueued(cfqd, cfqq, rq);
3467 * Update hw_tag based on peak queue depth over 50 samples under
3468 * sufficient load.
3470 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3472 struct cfq_queue *cfqq = cfqd->active_queue;
3474 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3475 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3477 if (cfqd->hw_tag == 1)
3478 return;
3480 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3481 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3482 return;
3485 * If active queue hasn't enough requests and can idle, cfq might not
3486 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3487 * case
3489 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3490 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3491 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3492 return;
3494 if (cfqd->hw_tag_samples++ < 50)
3495 return;
3497 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3498 cfqd->hw_tag = 1;
3499 else
3500 cfqd->hw_tag = 0;
3503 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3505 struct cfq_io_context *cic = cfqd->active_cic;
3507 /* If the queue already has requests, don't wait */
3508 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3509 return false;
3511 /* If there are other queues in the group, don't wait */
3512 if (cfqq->cfqg->nr_cfqq > 1)
3513 return false;
3515 if (cfq_slice_used(cfqq))
3516 return true;
3518 /* if slice left is less than think time, wait busy */
3519 if (cic && sample_valid(cic->ttime_samples)
3520 && (cfqq->slice_end - jiffies < cic->ttime_mean))
3521 return true;
3524 * If think times is less than a jiffy than ttime_mean=0 and above
3525 * will not be true. It might happen that slice has not expired yet
3526 * but will expire soon (4-5 ns) during select_queue(). To cover the
3527 * case where think time is less than a jiffy, mark the queue wait
3528 * busy if only 1 jiffy is left in the slice.
3530 if (cfqq->slice_end - jiffies == 1)
3531 return true;
3533 return false;
3536 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3538 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3539 struct cfq_data *cfqd = cfqq->cfqd;
3540 const int sync = rq_is_sync(rq);
3541 unsigned long now;
3543 now = jiffies;
3544 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3545 !!(rq->cmd_flags & REQ_NOIDLE));
3547 cfq_update_hw_tag(cfqd);
3549 WARN_ON(!cfqd->rq_in_driver);
3550 WARN_ON(!cfqq->dispatched);
3551 cfqd->rq_in_driver--;
3552 cfqq->dispatched--;
3553 (RQ_CFQG(rq))->dispatched--;
3554 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3555 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3556 rq_data_dir(rq), rq_is_sync(rq));
3558 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3560 if (sync) {
3561 RQ_CIC(rq)->last_end_request = now;
3562 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3563 cfqd->last_delayed_sync = now;
3567 * If this is the active queue, check if it needs to be expired,
3568 * or if we want to idle in case it has no pending requests.
3570 if (cfqd->active_queue == cfqq) {
3571 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3573 if (cfq_cfqq_slice_new(cfqq)) {
3574 cfq_set_prio_slice(cfqd, cfqq);
3575 cfq_clear_cfqq_slice_new(cfqq);
3579 * Should we wait for next request to come in before we expire
3580 * the queue.
3582 if (cfq_should_wait_busy(cfqd, cfqq)) {
3583 unsigned long extend_sl = cfqd->cfq_slice_idle;
3584 if (!cfqd->cfq_slice_idle)
3585 extend_sl = cfqd->cfq_group_idle;
3586 cfqq->slice_end = jiffies + extend_sl;
3587 cfq_mark_cfqq_wait_busy(cfqq);
3588 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3592 * Idling is not enabled on:
3593 * - expired queues
3594 * - idle-priority queues
3595 * - async queues
3596 * - queues with still some requests queued
3597 * - when there is a close cooperator
3599 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3600 cfq_slice_expired(cfqd, 1);
3601 else if (sync && cfqq_empty &&
3602 !cfq_close_cooperator(cfqd, cfqq)) {
3603 cfq_arm_slice_timer(cfqd);
3607 if (!cfqd->rq_in_driver)
3608 cfq_schedule_dispatch(cfqd);
3612 * we temporarily boost lower priority queues if they are holding fs exclusive
3613 * resources. they are boosted to normal prio (CLASS_BE/4)
3615 static void cfq_prio_boost(struct cfq_queue *cfqq)
3617 if (has_fs_excl()) {
3619 * boost idle prio on transactions that would lock out other
3620 * users of the filesystem
3622 if (cfq_class_idle(cfqq))
3623 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3624 if (cfqq->ioprio > IOPRIO_NORM)
3625 cfqq->ioprio = IOPRIO_NORM;
3626 } else {
3628 * unboost the queue (if needed)
3630 cfqq->ioprio_class = cfqq->org_ioprio_class;
3631 cfqq->ioprio = cfqq->org_ioprio;
3635 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3637 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3638 cfq_mark_cfqq_must_alloc_slice(cfqq);
3639 return ELV_MQUEUE_MUST;
3642 return ELV_MQUEUE_MAY;
3645 static int cfq_may_queue(struct request_queue *q, int rw)
3647 struct cfq_data *cfqd = q->elevator->elevator_data;
3648 struct task_struct *tsk = current;
3649 struct cfq_io_context *cic;
3650 struct cfq_queue *cfqq;
3653 * don't force setup of a queue from here, as a call to may_queue
3654 * does not necessarily imply that a request actually will be queued.
3655 * so just lookup a possibly existing queue, or return 'may queue'
3656 * if that fails
3658 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3659 if (!cic)
3660 return ELV_MQUEUE_MAY;
3662 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3663 if (cfqq) {
3664 cfq_init_prio_data(cfqq, cic->ioc);
3665 cfq_prio_boost(cfqq);
3667 return __cfq_may_queue(cfqq);
3670 return ELV_MQUEUE_MAY;
3674 * queue lock held here
3676 static void cfq_put_request(struct request *rq)
3678 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3680 if (cfqq) {
3681 const int rw = rq_data_dir(rq);
3683 BUG_ON(!cfqq->allocated[rw]);
3684 cfqq->allocated[rw]--;
3686 put_io_context(RQ_CIC(rq)->ioc);
3688 rq->elevator_private[0] = NULL;
3689 rq->elevator_private[1] = NULL;
3691 /* Put down rq reference on cfqg */
3692 cfq_put_cfqg(RQ_CFQG(rq));
3693 rq->elevator_private[2] = NULL;
3695 cfq_put_queue(cfqq);
3699 static struct cfq_queue *
3700 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3701 struct cfq_queue *cfqq)
3703 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3704 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3705 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3706 cfq_put_queue(cfqq);
3707 return cic_to_cfqq(cic, 1);
3711 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3712 * was the last process referring to said cfqq.
3714 static struct cfq_queue *
3715 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3717 if (cfqq_process_refs(cfqq) == 1) {
3718 cfqq->pid = current->pid;
3719 cfq_clear_cfqq_coop(cfqq);
3720 cfq_clear_cfqq_split_coop(cfqq);
3721 return cfqq;
3724 cic_set_cfqq(cic, NULL, 1);
3726 cfq_put_cooperator(cfqq);
3728 cfq_put_queue(cfqq);
3729 return NULL;
3732 * Allocate cfq data structures associated with this request.
3734 static int
3735 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3737 struct cfq_data *cfqd = q->elevator->elevator_data;
3738 struct cfq_io_context *cic;
3739 const int rw = rq_data_dir(rq);
3740 const bool is_sync = rq_is_sync(rq);
3741 struct cfq_queue *cfqq;
3742 unsigned long flags;
3744 might_sleep_if(gfp_mask & __GFP_WAIT);
3746 cic = cfq_get_io_context(cfqd, gfp_mask);
3748 spin_lock_irqsave(q->queue_lock, flags);
3750 if (!cic)
3751 goto queue_fail;
3753 new_queue:
3754 cfqq = cic_to_cfqq(cic, is_sync);
3755 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3756 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3757 cic_set_cfqq(cic, cfqq, is_sync);
3758 } else {
3760 * If the queue was seeky for too long, break it apart.
3762 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3763 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3764 cfqq = split_cfqq(cic, cfqq);
3765 if (!cfqq)
3766 goto new_queue;
3770 * Check to see if this queue is scheduled to merge with
3771 * another, closely cooperating queue. The merging of
3772 * queues happens here as it must be done in process context.
3773 * The reference on new_cfqq was taken in merge_cfqqs.
3775 if (cfqq->new_cfqq)
3776 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3779 cfqq->allocated[rw]++;
3781 cfqq->ref++;
3782 rq->elevator_private[0] = cic;
3783 rq->elevator_private[1] = cfqq;
3784 rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3785 spin_unlock_irqrestore(q->queue_lock, flags);
3786 return 0;
3788 queue_fail:
3789 cfq_schedule_dispatch(cfqd);
3790 spin_unlock_irqrestore(q->queue_lock, flags);
3791 cfq_log(cfqd, "set_request fail");
3792 return 1;
3795 static void cfq_kick_queue(struct work_struct *work)
3797 struct cfq_data *cfqd =
3798 container_of(work, struct cfq_data, unplug_work);
3799 struct request_queue *q = cfqd->queue;
3801 spin_lock_irq(q->queue_lock);
3802 __blk_run_queue(cfqd->queue);
3803 spin_unlock_irq(q->queue_lock);
3807 * Timer running if the active_queue is currently idling inside its time slice
3809 static void cfq_idle_slice_timer(unsigned long data)
3811 struct cfq_data *cfqd = (struct cfq_data *) data;
3812 struct cfq_queue *cfqq;
3813 unsigned long flags;
3814 int timed_out = 1;
3816 cfq_log(cfqd, "idle timer fired");
3818 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3820 cfqq = cfqd->active_queue;
3821 if (cfqq) {
3822 timed_out = 0;
3825 * We saw a request before the queue expired, let it through
3827 if (cfq_cfqq_must_dispatch(cfqq))
3828 goto out_kick;
3831 * expired
3833 if (cfq_slice_used(cfqq))
3834 goto expire;
3837 * only expire and reinvoke request handler, if there are
3838 * other queues with pending requests
3840 if (!cfqd->busy_queues)
3841 goto out_cont;
3844 * not expired and it has a request pending, let it dispatch
3846 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3847 goto out_kick;
3850 * Queue depth flag is reset only when the idle didn't succeed
3852 cfq_clear_cfqq_deep(cfqq);
3854 expire:
3855 cfq_slice_expired(cfqd, timed_out);
3856 out_kick:
3857 cfq_schedule_dispatch(cfqd);
3858 out_cont:
3859 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3862 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3864 del_timer_sync(&cfqd->idle_slice_timer);
3865 cancel_work_sync(&cfqd->unplug_work);
3868 static void cfq_put_async_queues(struct cfq_data *cfqd)
3870 int i;
3872 for (i = 0; i < IOPRIO_BE_NR; i++) {
3873 if (cfqd->async_cfqq[0][i])
3874 cfq_put_queue(cfqd->async_cfqq[0][i]);
3875 if (cfqd->async_cfqq[1][i])
3876 cfq_put_queue(cfqd->async_cfqq[1][i]);
3879 if (cfqd->async_idle_cfqq)
3880 cfq_put_queue(cfqd->async_idle_cfqq);
3883 static void cfq_exit_queue(struct elevator_queue *e)
3885 struct cfq_data *cfqd = e->elevator_data;
3886 struct request_queue *q = cfqd->queue;
3887 bool wait = false;
3889 cfq_shutdown_timer_wq(cfqd);
3891 spin_lock_irq(q->queue_lock);
3893 if (cfqd->active_queue)
3894 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3896 while (!list_empty(&cfqd->cic_list)) {
3897 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3898 struct cfq_io_context,
3899 queue_list);
3901 __cfq_exit_single_io_context(cfqd, cic);
3904 cfq_put_async_queues(cfqd);
3905 cfq_release_cfq_groups(cfqd);
3908 * If there are groups which we could not unlink from blkcg list,
3909 * wait for a rcu period for them to be freed.
3911 if (cfqd->nr_blkcg_linked_grps)
3912 wait = true;
3914 spin_unlock_irq(q->queue_lock);
3916 cfq_shutdown_timer_wq(cfqd);
3918 spin_lock(&cic_index_lock);
3919 ida_remove(&cic_index_ida, cfqd->cic_index);
3920 spin_unlock(&cic_index_lock);
3923 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3924 * Do this wait only if there are other unlinked groups out
3925 * there. This can happen if cgroup deletion path claimed the
3926 * responsibility of cleaning up a group before queue cleanup code
3927 * get to the group.
3929 * Do not call synchronize_rcu() unconditionally as there are drivers
3930 * which create/delete request queue hundreds of times during scan/boot
3931 * and synchronize_rcu() can take significant time and slow down boot.
3933 if (wait)
3934 synchronize_rcu();
3936 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3937 /* Free up per cpu stats for root group */
3938 free_percpu(cfqd->root_group.blkg.stats_cpu);
3939 #endif
3940 kfree(cfqd);
3943 static int cfq_alloc_cic_index(void)
3945 int index, error;
3947 do {
3948 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3949 return -ENOMEM;
3951 spin_lock(&cic_index_lock);
3952 error = ida_get_new(&cic_index_ida, &index);
3953 spin_unlock(&cic_index_lock);
3954 if (error && error != -EAGAIN)
3955 return error;
3956 } while (error);
3958 return index;
3961 static void *cfq_init_queue(struct request_queue *q)
3963 struct cfq_data *cfqd;
3964 int i, j;
3965 struct cfq_group *cfqg;
3966 struct cfq_rb_root *st;
3968 i = cfq_alloc_cic_index();
3969 if (i < 0)
3970 return NULL;
3972 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3973 if (!cfqd) {
3974 spin_lock(&cic_index_lock);
3975 ida_remove(&cic_index_ida, i);
3976 spin_unlock(&cic_index_lock);
3977 return NULL;
3981 * Don't need take queue_lock in the routine, since we are
3982 * initializing the ioscheduler, and nobody is using cfqd
3984 cfqd->cic_index = i;
3986 /* Init root service tree */
3987 cfqd->grp_service_tree = CFQ_RB_ROOT;
3989 /* Init root group */
3990 cfqg = &cfqd->root_group;
3991 for_each_cfqg_st(cfqg, i, j, st)
3992 *st = CFQ_RB_ROOT;
3993 RB_CLEAR_NODE(&cfqg->rb_node);
3995 /* Give preference to root group over other groups */
3996 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3998 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4000 * Set root group reference to 2. One reference will be dropped when
4001 * all groups on cfqd->cfqg_list are being deleted during queue exit.
4002 * Other reference will remain there as we don't want to delete this
4003 * group as it is statically allocated and gets destroyed when
4004 * throtl_data goes away.
4006 cfqg->ref = 2;
4008 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4009 kfree(cfqg);
4010 kfree(cfqd);
4011 return NULL;
4014 rcu_read_lock();
4016 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4017 (void *)cfqd, 0);
4018 rcu_read_unlock();
4019 cfqd->nr_blkcg_linked_grps++;
4021 /* Add group on cfqd->cfqg_list */
4022 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4023 #endif
4025 * Not strictly needed (since RB_ROOT just clears the node and we
4026 * zeroed cfqd on alloc), but better be safe in case someone decides
4027 * to add magic to the rb code
4029 for (i = 0; i < CFQ_PRIO_LISTS; i++)
4030 cfqd->prio_trees[i] = RB_ROOT;
4033 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4034 * Grab a permanent reference to it, so that the normal code flow
4035 * will not attempt to free it.
4037 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4038 cfqd->oom_cfqq.ref++;
4039 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4041 INIT_LIST_HEAD(&cfqd->cic_list);
4043 cfqd->queue = q;
4045 init_timer(&cfqd->idle_slice_timer);
4046 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4047 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4049 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4051 cfqd->cfq_quantum = cfq_quantum;
4052 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4053 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4054 cfqd->cfq_back_max = cfq_back_max;
4055 cfqd->cfq_back_penalty = cfq_back_penalty;
4056 cfqd->cfq_slice[0] = cfq_slice_async;
4057 cfqd->cfq_slice[1] = cfq_slice_sync;
4058 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4059 cfqd->cfq_slice_idle = cfq_slice_idle;
4060 cfqd->cfq_group_idle = cfq_group_idle;
4061 cfqd->cfq_latency = 1;
4062 cfqd->hw_tag = -1;
4064 * we optimistically start assuming sync ops weren't delayed in last
4065 * second, in order to have larger depth for async operations.
4067 cfqd->last_delayed_sync = jiffies - HZ;
4068 return cfqd;
4071 static void cfq_slab_kill(void)
4074 * Caller already ensured that pending RCU callbacks are completed,
4075 * so we should have no busy allocations at this point.
4077 if (cfq_pool)
4078 kmem_cache_destroy(cfq_pool);
4079 if (cfq_ioc_pool)
4080 kmem_cache_destroy(cfq_ioc_pool);
4083 static int __init cfq_slab_setup(void)
4085 cfq_pool = KMEM_CACHE(cfq_queue, 0);
4086 if (!cfq_pool)
4087 goto fail;
4089 cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4090 if (!cfq_ioc_pool)
4091 goto fail;
4093 return 0;
4094 fail:
4095 cfq_slab_kill();
4096 return -ENOMEM;
4100 * sysfs parts below -->
4102 static ssize_t
4103 cfq_var_show(unsigned int var, char *page)
4105 return sprintf(page, "%d\n", var);
4108 static ssize_t
4109 cfq_var_store(unsigned int *var, const char *page, size_t count)
4111 char *p = (char *) page;
4113 *var = simple_strtoul(p, &p, 10);
4114 return count;
4117 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
4118 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
4120 struct cfq_data *cfqd = e->elevator_data; \
4121 unsigned int __data = __VAR; \
4122 if (__CONV) \
4123 __data = jiffies_to_msecs(__data); \
4124 return cfq_var_show(__data, (page)); \
4126 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4127 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4128 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4129 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4130 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4131 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4132 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4133 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4134 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4135 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4136 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4137 #undef SHOW_FUNCTION
4139 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
4140 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4142 struct cfq_data *cfqd = e->elevator_data; \
4143 unsigned int __data; \
4144 int ret = cfq_var_store(&__data, (page), count); \
4145 if (__data < (MIN)) \
4146 __data = (MIN); \
4147 else if (__data > (MAX)) \
4148 __data = (MAX); \
4149 if (__CONV) \
4150 *(__PTR) = msecs_to_jiffies(__data); \
4151 else \
4152 *(__PTR) = __data; \
4153 return ret; \
4155 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4156 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4157 UINT_MAX, 1);
4158 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4159 UINT_MAX, 1);
4160 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4161 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4162 UINT_MAX, 0);
4163 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4164 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4165 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4166 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4167 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4168 UINT_MAX, 0);
4169 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4170 #undef STORE_FUNCTION
4172 #define CFQ_ATTR(name) \
4173 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4175 static struct elv_fs_entry cfq_attrs[] = {
4176 CFQ_ATTR(quantum),
4177 CFQ_ATTR(fifo_expire_sync),
4178 CFQ_ATTR(fifo_expire_async),
4179 CFQ_ATTR(back_seek_max),
4180 CFQ_ATTR(back_seek_penalty),
4181 CFQ_ATTR(slice_sync),
4182 CFQ_ATTR(slice_async),
4183 CFQ_ATTR(slice_async_rq),
4184 CFQ_ATTR(slice_idle),
4185 CFQ_ATTR(group_idle),
4186 CFQ_ATTR(low_latency),
4187 __ATTR_NULL
4190 static struct elevator_type iosched_cfq = {
4191 .ops = {
4192 .elevator_merge_fn = cfq_merge,
4193 .elevator_merged_fn = cfq_merged_request,
4194 .elevator_merge_req_fn = cfq_merged_requests,
4195 .elevator_allow_merge_fn = cfq_allow_merge,
4196 .elevator_bio_merged_fn = cfq_bio_merged,
4197 .elevator_dispatch_fn = cfq_dispatch_requests,
4198 .elevator_add_req_fn = cfq_insert_request,
4199 .elevator_activate_req_fn = cfq_activate_request,
4200 .elevator_deactivate_req_fn = cfq_deactivate_request,
4201 .elevator_completed_req_fn = cfq_completed_request,
4202 .elevator_former_req_fn = elv_rb_former_request,
4203 .elevator_latter_req_fn = elv_rb_latter_request,
4204 .elevator_set_req_fn = cfq_set_request,
4205 .elevator_put_req_fn = cfq_put_request,
4206 .elevator_may_queue_fn = cfq_may_queue,
4207 .elevator_init_fn = cfq_init_queue,
4208 .elevator_exit_fn = cfq_exit_queue,
4209 .trim = cfq_free_io_context,
4211 .elevator_attrs = cfq_attrs,
4212 .elevator_name = "cfq",
4213 .elevator_owner = THIS_MODULE,
4216 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4217 static struct blkio_policy_type blkio_policy_cfq = {
4218 .ops = {
4219 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
4220 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4222 .plid = BLKIO_POLICY_PROP,
4224 #else
4225 static struct blkio_policy_type blkio_policy_cfq;
4226 #endif
4228 static int __init cfq_init(void)
4231 * could be 0 on HZ < 1000 setups
4233 if (!cfq_slice_async)
4234 cfq_slice_async = 1;
4235 if (!cfq_slice_idle)
4236 cfq_slice_idle = 1;
4238 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4239 if (!cfq_group_idle)
4240 cfq_group_idle = 1;
4241 #else
4242 cfq_group_idle = 0;
4243 #endif
4244 if (cfq_slab_setup())
4245 return -ENOMEM;
4247 elv_register(&iosched_cfq);
4248 blkio_policy_register(&blkio_policy_cfq);
4250 return 0;
4253 static void __exit cfq_exit(void)
4255 DECLARE_COMPLETION_ONSTACK(all_gone);
4256 blkio_policy_unregister(&blkio_policy_cfq);
4257 elv_unregister(&iosched_cfq);
4258 ioc_gone = &all_gone;
4259 /* ioc_gone's update must be visible before reading ioc_count */
4260 smp_wmb();
4263 * this also protects us from entering cfq_slab_kill() with
4264 * pending RCU callbacks
4266 if (elv_ioc_count_read(cfq_ioc_count))
4267 wait_for_completion(&all_gone);
4268 ida_destroy(&cic_index_ida);
4269 cfq_slab_kill();
4272 module_init(cfq_init);
4273 module_exit(cfq_exit);
4275 MODULE_AUTHOR("Jens Axboe");
4276 MODULE_LICENSE("GPL");
4277 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");