xircom_cb: fix device probe error path.
[linux-2.6.git] / block / cfq-iosched.c
blob457295253566e97179daca0502c57c320962ed24
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 "blk.h"
18 #include "cfq.h"
21 * tunables
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
39 * offset from end of service tree
41 #define CFQ_IDLE_DELAY (HZ / 5)
44 * below this threshold, we consider thinktime immediate
46 #define CFQ_MIN_TT (2)
48 #define CFQ_SLICE_SCALE (5)
49 #define CFQ_HW_QUEUE_MIN (5)
50 #define CFQ_SERVICE_SHIFT 12
52 #define CFQQ_SEEK_THR (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq) (hweight32(cfqq->seek_history) > 32/8)
57 #define RQ_CIC(rq) icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq) (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq) (struct cfq_group *) ((rq)->elv.priv[1])
61 static struct kmem_cache *cfq_pool;
63 #define CFQ_PRIO_LISTS IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq) ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67 #define sample_valid(samples) ((samples) > 80)
68 #define rb_entry_cfqg(node) rb_entry((node), struct cfq_group, rb_node)
70 struct cfq_ttime {
71 unsigned long last_end_request;
73 unsigned long ttime_total;
74 unsigned long ttime_samples;
75 unsigned long ttime_mean;
79 * Most of our rbtree usage is for sorting with min extraction, so
80 * if we cache the leftmost node we don't have to walk down the tree
81 * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82 * move this into the elevator for the rq sorting as well.
84 struct cfq_rb_root {
85 struct rb_root rb;
86 struct rb_node *left;
87 unsigned count;
88 unsigned total_weight;
89 u64 min_vdisktime;
90 struct cfq_ttime ttime;
92 #define CFQ_RB_ROOT (struct cfq_rb_root) { .rb = RB_ROOT, \
93 .ttime = {.last_end_request = jiffies,},}
96 * Per process-grouping structure
98 struct cfq_queue {
99 /* reference count */
100 int ref;
101 /* various state flags, see below */
102 unsigned int flags;
103 /* parent cfq_data */
104 struct cfq_data *cfqd;
105 /* service_tree member */
106 struct rb_node rb_node;
107 /* service_tree key */
108 unsigned long rb_key;
109 /* prio tree member */
110 struct rb_node p_node;
111 /* prio tree root we belong to, if any */
112 struct rb_root *p_root;
113 /* sorted list of pending requests */
114 struct rb_root sort_list;
115 /* if fifo isn't expired, next request to serve */
116 struct request *next_rq;
117 /* requests queued in sort_list */
118 int queued[2];
119 /* currently allocated requests */
120 int allocated[2];
121 /* fifo list of requests in sort_list */
122 struct list_head fifo;
124 /* time when queue got scheduled in to dispatch first request. */
125 unsigned long dispatch_start;
126 unsigned int allocated_slice;
127 unsigned int slice_dispatch;
128 /* time when first request from queue completed and slice started. */
129 unsigned long slice_start;
130 unsigned long slice_end;
131 long slice_resid;
133 /* pending priority requests */
134 int prio_pending;
135 /* number of requests that are on the dispatch list or inside driver */
136 int dispatched;
138 /* io prio of this group */
139 unsigned short ioprio, org_ioprio;
140 unsigned short ioprio_class;
142 pid_t pid;
144 u32 seek_history;
145 sector_t last_request_pos;
147 struct cfq_rb_root *service_tree;
148 struct cfq_queue *new_cfqq;
149 struct cfq_group *cfqg;
150 /* Number of sectors dispatched from queue in single dispatch round */
151 unsigned long nr_sectors;
155 * First index in the service_trees.
156 * IDLE is handled separately, so it has negative index
158 enum wl_prio_t {
159 BE_WORKLOAD = 0,
160 RT_WORKLOAD = 1,
161 IDLE_WORKLOAD = 2,
162 CFQ_PRIO_NR,
166 * Second index in the service_trees.
168 enum wl_type_t {
169 ASYNC_WORKLOAD = 0,
170 SYNC_NOIDLE_WORKLOAD = 1,
171 SYNC_WORKLOAD = 2
174 /* This is per cgroup per device grouping structure */
175 struct cfq_group {
176 /* group service_tree member */
177 struct rb_node rb_node;
179 /* group service_tree key */
180 u64 vdisktime;
181 unsigned int weight;
182 unsigned int new_weight;
183 bool needs_update;
185 /* number of cfqq currently on this group */
186 int nr_cfqq;
189 * Per group busy queues average. Useful for workload slice calc. We
190 * create the array for each prio class but at run time it is used
191 * only for RT and BE class and slot for IDLE class remains unused.
192 * This is primarily done to avoid confusion and a gcc warning.
194 unsigned int busy_queues_avg[CFQ_PRIO_NR];
196 * rr lists of queues with requests. We maintain service trees for
197 * RT and BE classes. These trees are subdivided in subclasses
198 * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199 * class there is no subclassification and all the cfq queues go on
200 * a single tree service_tree_idle.
201 * Counts are embedded in the cfq_rb_root
203 struct cfq_rb_root service_trees[2][3];
204 struct cfq_rb_root service_tree_idle;
206 unsigned long saved_workload_slice;
207 enum wl_type_t saved_workload;
208 enum wl_prio_t saved_serving_prio;
209 struct blkio_group blkg;
210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
211 struct hlist_node cfqd_node;
212 int ref;
213 #endif
214 /* number of requests that are on the dispatch list or inside driver */
215 int dispatched;
216 struct cfq_ttime ttime;
219 struct cfq_io_cq {
220 struct io_cq icq; /* must be the first member */
221 struct cfq_queue *cfqq[2];
222 struct cfq_ttime ttime;
226 * Per block device queue structure
228 struct cfq_data {
229 struct request_queue *queue;
230 /* Root service tree for cfq_groups */
231 struct cfq_rb_root grp_service_tree;
232 struct cfq_group root_group;
235 * The priority currently being served
237 enum wl_prio_t serving_prio;
238 enum wl_type_t serving_type;
239 unsigned long workload_expires;
240 struct cfq_group *serving_group;
243 * Each priority tree is sorted by next_request position. These
244 * trees are used when determining if two or more queues are
245 * interleaving requests (see cfq_close_cooperator).
247 struct rb_root prio_trees[CFQ_PRIO_LISTS];
249 unsigned int busy_queues;
250 unsigned int busy_sync_queues;
252 int rq_in_driver;
253 int rq_in_flight[2];
256 * queue-depth detection
258 int rq_queued;
259 int hw_tag;
261 * hw_tag can be
262 * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
263 * 1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
264 * 0 => no NCQ
266 int hw_tag_est_depth;
267 unsigned int hw_tag_samples;
270 * idle window management
272 struct timer_list idle_slice_timer;
273 struct work_struct unplug_work;
275 struct cfq_queue *active_queue;
276 struct cfq_io_cq *active_cic;
279 * async queue for each priority case
281 struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
282 struct cfq_queue *async_idle_cfqq;
284 sector_t last_position;
287 * tunables, see top of file
289 unsigned int cfq_quantum;
290 unsigned int cfq_fifo_expire[2];
291 unsigned int cfq_back_penalty;
292 unsigned int cfq_back_max;
293 unsigned int cfq_slice[2];
294 unsigned int cfq_slice_async_rq;
295 unsigned int cfq_slice_idle;
296 unsigned int cfq_group_idle;
297 unsigned int cfq_latency;
300 * Fallback dummy cfqq for extreme OOM conditions
302 struct cfq_queue oom_cfqq;
304 unsigned long last_delayed_sync;
306 /* List of cfq groups being managed on this device*/
307 struct hlist_head cfqg_list;
309 /* Number of groups which are on blkcg->blkg_list */
310 unsigned int nr_blkcg_linked_grps;
313 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
315 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
316 enum wl_prio_t prio,
317 enum wl_type_t type)
319 if (!cfqg)
320 return NULL;
322 if (prio == IDLE_WORKLOAD)
323 return &cfqg->service_tree_idle;
325 return &cfqg->service_trees[prio][type];
328 enum cfqq_state_flags {
329 CFQ_CFQQ_FLAG_on_rr = 0, /* on round-robin busy list */
330 CFQ_CFQQ_FLAG_wait_request, /* waiting for a request */
331 CFQ_CFQQ_FLAG_must_dispatch, /* must be allowed a dispatch */
332 CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
333 CFQ_CFQQ_FLAG_fifo_expire, /* FIFO checked in this slice */
334 CFQ_CFQQ_FLAG_idle_window, /* slice idling enabled */
335 CFQ_CFQQ_FLAG_prio_changed, /* task priority has changed */
336 CFQ_CFQQ_FLAG_slice_new, /* no requests dispatched in slice */
337 CFQ_CFQQ_FLAG_sync, /* synchronous queue */
338 CFQ_CFQQ_FLAG_coop, /* cfqq is shared */
339 CFQ_CFQQ_FLAG_split_coop, /* shared cfqq will be splitted */
340 CFQ_CFQQ_FLAG_deep, /* sync cfqq experienced large depth */
341 CFQ_CFQQ_FLAG_wait_busy, /* Waiting for next request */
344 #define CFQ_CFQQ_FNS(name) \
345 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq) \
347 (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name); \
349 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq) \
351 (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name); \
353 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq) \
355 return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0; \
358 CFQ_CFQQ_FNS(on_rr);
359 CFQ_CFQQ_FNS(wait_request);
360 CFQ_CFQQ_FNS(must_dispatch);
361 CFQ_CFQQ_FNS(must_alloc_slice);
362 CFQ_CFQQ_FNS(fifo_expire);
363 CFQ_CFQQ_FNS(idle_window);
364 CFQ_CFQQ_FNS(prio_changed);
365 CFQ_CFQQ_FNS(slice_new);
366 CFQ_CFQQ_FNS(sync);
367 CFQ_CFQQ_FNS(coop);
368 CFQ_CFQQ_FNS(split_coop);
369 CFQ_CFQQ_FNS(deep);
370 CFQ_CFQQ_FNS(wait_busy);
371 #undef CFQ_CFQQ_FNS
373 #ifdef CONFIG_CFQ_GROUP_IOSCHED
374 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
375 blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
376 cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
377 blkg_path(&(cfqq)->cfqg->blkg), ##args)
379 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) \
380 blk_add_trace_msg((cfqd)->queue, "%s " fmt, \
381 blkg_path(&(cfqg)->blkg), ##args) \
383 #else
384 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...) \
385 blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
386 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...) do {} while (0)
387 #endif
388 #define cfq_log(cfqd, fmt, args...) \
389 blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
391 /* Traverses through cfq group service trees */
392 #define for_each_cfqg_st(cfqg, i, j, st) \
393 for (i = 0; i <= IDLE_WORKLOAD; i++) \
394 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
395 : &cfqg->service_tree_idle; \
396 (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
397 (i == IDLE_WORKLOAD && j == 0); \
398 j++, st = i < IDLE_WORKLOAD ? \
399 &cfqg->service_trees[i][j]: NULL) \
401 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
402 struct cfq_ttime *ttime, bool group_idle)
404 unsigned long slice;
405 if (!sample_valid(ttime->ttime_samples))
406 return false;
407 if (group_idle)
408 slice = cfqd->cfq_group_idle;
409 else
410 slice = cfqd->cfq_slice_idle;
411 return ttime->ttime_mean > slice;
414 static inline bool iops_mode(struct cfq_data *cfqd)
417 * If we are not idling on queues and it is a NCQ drive, parallel
418 * execution of requests is on and measuring time is not possible
419 * in most of the cases until and unless we drive shallower queue
420 * depths and that becomes a performance bottleneck. In such cases
421 * switch to start providing fairness in terms of number of IOs.
423 if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
424 return true;
425 else
426 return false;
429 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
431 if (cfq_class_idle(cfqq))
432 return IDLE_WORKLOAD;
433 if (cfq_class_rt(cfqq))
434 return RT_WORKLOAD;
435 return BE_WORKLOAD;
439 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
441 if (!cfq_cfqq_sync(cfqq))
442 return ASYNC_WORKLOAD;
443 if (!cfq_cfqq_idle_window(cfqq))
444 return SYNC_NOIDLE_WORKLOAD;
445 return SYNC_WORKLOAD;
448 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
449 struct cfq_data *cfqd,
450 struct cfq_group *cfqg)
452 if (wl == IDLE_WORKLOAD)
453 return cfqg->service_tree_idle.count;
455 return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
456 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
457 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
460 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
461 struct cfq_group *cfqg)
463 return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
464 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
467 static void cfq_dispatch_insert(struct request_queue *, struct request *);
468 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
469 struct io_context *, gfp_t);
471 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
473 /* cic->icq is the first member, %NULL will convert to %NULL */
474 return container_of(icq, struct cfq_io_cq, icq);
477 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
478 struct io_context *ioc)
480 if (ioc)
481 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
482 return NULL;
485 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
487 return cic->cfqq[is_sync];
490 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
491 bool is_sync)
493 cic->cfqq[is_sync] = cfqq;
496 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
498 return cic->icq.q->elevator->elevator_data;
502 * We regard a request as SYNC, if it's either a read or has the SYNC bit
503 * set (in which case it could also be direct WRITE).
505 static inline bool cfq_bio_sync(struct bio *bio)
507 return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
511 * scheduler run of queue, if there are requests pending and no one in the
512 * driver that will restart queueing
514 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
516 if (cfqd->busy_queues) {
517 cfq_log(cfqd, "schedule dispatch");
518 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
523 * Scale schedule slice based on io priority. Use the sync time slice only
524 * if a queue is marked sync and has sync io queued. A sync queue with async
525 * io only, should not get full sync slice length.
527 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
528 unsigned short prio)
530 const int base_slice = cfqd->cfq_slice[sync];
532 WARN_ON(prio >= IOPRIO_BE_NR);
534 return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
537 static inline int
538 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
540 return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
543 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
545 u64 d = delta << CFQ_SERVICE_SHIFT;
547 d = d * BLKIO_WEIGHT_DEFAULT;
548 do_div(d, cfqg->weight);
549 return d;
552 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
554 s64 delta = (s64)(vdisktime - min_vdisktime);
555 if (delta > 0)
556 min_vdisktime = vdisktime;
558 return min_vdisktime;
561 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
563 s64 delta = (s64)(vdisktime - min_vdisktime);
564 if (delta < 0)
565 min_vdisktime = vdisktime;
567 return min_vdisktime;
570 static void update_min_vdisktime(struct cfq_rb_root *st)
572 struct cfq_group *cfqg;
574 if (st->left) {
575 cfqg = rb_entry_cfqg(st->left);
576 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
577 cfqg->vdisktime);
582 * get averaged number of queues of RT/BE priority.
583 * average is updated, with a formula that gives more weight to higher numbers,
584 * to quickly follows sudden increases and decrease slowly
587 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
588 struct cfq_group *cfqg, bool rt)
590 unsigned min_q, max_q;
591 unsigned mult = cfq_hist_divisor - 1;
592 unsigned round = cfq_hist_divisor / 2;
593 unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
595 min_q = min(cfqg->busy_queues_avg[rt], busy);
596 max_q = max(cfqg->busy_queues_avg[rt], busy);
597 cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
598 cfq_hist_divisor;
599 return cfqg->busy_queues_avg[rt];
602 static inline unsigned
603 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
605 struct cfq_rb_root *st = &cfqd->grp_service_tree;
607 return cfq_target_latency * cfqg->weight / st->total_weight;
610 static inline unsigned
611 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
613 unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
614 if (cfqd->cfq_latency) {
616 * interested queues (we consider only the ones with the same
617 * priority class in the cfq group)
619 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
620 cfq_class_rt(cfqq));
621 unsigned sync_slice = cfqd->cfq_slice[1];
622 unsigned expect_latency = sync_slice * iq;
623 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
625 if (expect_latency > group_slice) {
626 unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
627 /* scale low_slice according to IO priority
628 * and sync vs async */
629 unsigned low_slice =
630 min(slice, base_low_slice * slice / sync_slice);
631 /* the adapted slice value is scaled to fit all iqs
632 * into the target latency */
633 slice = max(slice * group_slice / expect_latency,
634 low_slice);
637 return slice;
640 static inline void
641 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
643 unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
645 cfqq->slice_start = jiffies;
646 cfqq->slice_end = jiffies + slice;
647 cfqq->allocated_slice = slice;
648 cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
652 * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
653 * isn't valid until the first request from the dispatch is activated
654 * and the slice time set.
656 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
658 if (cfq_cfqq_slice_new(cfqq))
659 return false;
660 if (time_before(jiffies, cfqq->slice_end))
661 return false;
663 return true;
667 * Lifted from AS - choose which of rq1 and rq2 that is best served now.
668 * We choose the request that is closest to the head right now. Distance
669 * behind the head is penalized and only allowed to a certain extent.
671 static struct request *
672 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
674 sector_t s1, s2, d1 = 0, d2 = 0;
675 unsigned long back_max;
676 #define CFQ_RQ1_WRAP 0x01 /* request 1 wraps */
677 #define CFQ_RQ2_WRAP 0x02 /* request 2 wraps */
678 unsigned wrap = 0; /* bit mask: requests behind the disk head? */
680 if (rq1 == NULL || rq1 == rq2)
681 return rq2;
682 if (rq2 == NULL)
683 return rq1;
685 if (rq_is_sync(rq1) != rq_is_sync(rq2))
686 return rq_is_sync(rq1) ? rq1 : rq2;
688 if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
689 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
691 s1 = blk_rq_pos(rq1);
692 s2 = blk_rq_pos(rq2);
695 * by definition, 1KiB is 2 sectors
697 back_max = cfqd->cfq_back_max * 2;
700 * Strict one way elevator _except_ in the case where we allow
701 * short backward seeks which are biased as twice the cost of a
702 * similar forward seek.
704 if (s1 >= last)
705 d1 = s1 - last;
706 else if (s1 + back_max >= last)
707 d1 = (last - s1) * cfqd->cfq_back_penalty;
708 else
709 wrap |= CFQ_RQ1_WRAP;
711 if (s2 >= last)
712 d2 = s2 - last;
713 else if (s2 + back_max >= last)
714 d2 = (last - s2) * cfqd->cfq_back_penalty;
715 else
716 wrap |= CFQ_RQ2_WRAP;
718 /* Found required data */
721 * By doing switch() on the bit mask "wrap" we avoid having to
722 * check two variables for all permutations: --> faster!
724 switch (wrap) {
725 case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
726 if (d1 < d2)
727 return rq1;
728 else if (d2 < d1)
729 return rq2;
730 else {
731 if (s1 >= s2)
732 return rq1;
733 else
734 return rq2;
737 case CFQ_RQ2_WRAP:
738 return rq1;
739 case CFQ_RQ1_WRAP:
740 return rq2;
741 case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
742 default:
744 * Since both rqs are wrapped,
745 * start with the one that's further behind head
746 * (--> only *one* back seek required),
747 * since back seek takes more time than forward.
749 if (s1 <= s2)
750 return rq1;
751 else
752 return rq2;
757 * The below is leftmost cache rbtree addon
759 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
761 /* Service tree is empty */
762 if (!root->count)
763 return NULL;
765 if (!root->left)
766 root->left = rb_first(&root->rb);
768 if (root->left)
769 return rb_entry(root->left, struct cfq_queue, rb_node);
771 return NULL;
774 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
776 if (!root->left)
777 root->left = rb_first(&root->rb);
779 if (root->left)
780 return rb_entry_cfqg(root->left);
782 return NULL;
785 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
787 rb_erase(n, root);
788 RB_CLEAR_NODE(n);
791 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
793 if (root->left == n)
794 root->left = NULL;
795 rb_erase_init(n, &root->rb);
796 --root->count;
800 * would be nice to take fifo expire time into account as well
802 static struct request *
803 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
804 struct request *last)
806 struct rb_node *rbnext = rb_next(&last->rb_node);
807 struct rb_node *rbprev = rb_prev(&last->rb_node);
808 struct request *next = NULL, *prev = NULL;
810 BUG_ON(RB_EMPTY_NODE(&last->rb_node));
812 if (rbprev)
813 prev = rb_entry_rq(rbprev);
815 if (rbnext)
816 next = rb_entry_rq(rbnext);
817 else {
818 rbnext = rb_first(&cfqq->sort_list);
819 if (rbnext && rbnext != &last->rb_node)
820 next = rb_entry_rq(rbnext);
823 return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
826 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
827 struct cfq_queue *cfqq)
830 * just an approximation, should be ok.
832 return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
833 cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
836 static inline s64
837 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
839 return cfqg->vdisktime - st->min_vdisktime;
842 static void
843 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
845 struct rb_node **node = &st->rb.rb_node;
846 struct rb_node *parent = NULL;
847 struct cfq_group *__cfqg;
848 s64 key = cfqg_key(st, cfqg);
849 int left = 1;
851 while (*node != NULL) {
852 parent = *node;
853 __cfqg = rb_entry_cfqg(parent);
855 if (key < cfqg_key(st, __cfqg))
856 node = &parent->rb_left;
857 else {
858 node = &parent->rb_right;
859 left = 0;
863 if (left)
864 st->left = &cfqg->rb_node;
866 rb_link_node(&cfqg->rb_node, parent, node);
867 rb_insert_color(&cfqg->rb_node, &st->rb);
870 static void
871 cfq_update_group_weight(struct cfq_group *cfqg)
873 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
874 if (cfqg->needs_update) {
875 cfqg->weight = cfqg->new_weight;
876 cfqg->needs_update = false;
880 static void
881 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
883 BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
885 cfq_update_group_weight(cfqg);
886 __cfq_group_service_tree_add(st, cfqg);
887 st->total_weight += cfqg->weight;
890 static void
891 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
893 struct cfq_rb_root *st = &cfqd->grp_service_tree;
894 struct cfq_group *__cfqg;
895 struct rb_node *n;
897 cfqg->nr_cfqq++;
898 if (!RB_EMPTY_NODE(&cfqg->rb_node))
899 return;
902 * Currently put the group at the end. Later implement something
903 * so that groups get lesser vtime based on their weights, so that
904 * if group does not loose all if it was not continuously backlogged.
906 n = rb_last(&st->rb);
907 if (n) {
908 __cfqg = rb_entry_cfqg(n);
909 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
910 } else
911 cfqg->vdisktime = st->min_vdisktime;
912 cfq_group_service_tree_add(st, cfqg);
915 static void
916 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
918 st->total_weight -= cfqg->weight;
919 if (!RB_EMPTY_NODE(&cfqg->rb_node))
920 cfq_rb_erase(&cfqg->rb_node, st);
923 static void
924 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
926 struct cfq_rb_root *st = &cfqd->grp_service_tree;
928 BUG_ON(cfqg->nr_cfqq < 1);
929 cfqg->nr_cfqq--;
931 /* If there are other cfq queues under this group, don't delete it */
932 if (cfqg->nr_cfqq)
933 return;
935 cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
936 cfq_group_service_tree_del(st, cfqg);
937 cfqg->saved_workload_slice = 0;
938 cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
941 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
942 unsigned int *unaccounted_time)
944 unsigned int slice_used;
947 * Queue got expired before even a single request completed or
948 * got expired immediately after first request completion.
950 if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
952 * Also charge the seek time incurred to the group, otherwise
953 * if there are mutiple queues in the group, each can dispatch
954 * a single request on seeky media and cause lots of seek time
955 * and group will never know it.
957 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
959 } else {
960 slice_used = jiffies - cfqq->slice_start;
961 if (slice_used > cfqq->allocated_slice) {
962 *unaccounted_time = slice_used - cfqq->allocated_slice;
963 slice_used = cfqq->allocated_slice;
965 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
966 *unaccounted_time += cfqq->slice_start -
967 cfqq->dispatch_start;
970 return slice_used;
973 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
974 struct cfq_queue *cfqq)
976 struct cfq_rb_root *st = &cfqd->grp_service_tree;
977 unsigned int used_sl, charge, unaccounted_sl = 0;
978 int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
979 - cfqg->service_tree_idle.count;
981 BUG_ON(nr_sync < 0);
982 used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
984 if (iops_mode(cfqd))
985 charge = cfqq->slice_dispatch;
986 else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
987 charge = cfqq->allocated_slice;
989 /* Can't update vdisktime while group is on service tree */
990 cfq_group_service_tree_del(st, cfqg);
991 cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
992 /* If a new weight was requested, update now, off tree */
993 cfq_group_service_tree_add(st, cfqg);
995 /* This group is being expired. Save the context */
996 if (time_after(cfqd->workload_expires, jiffies)) {
997 cfqg->saved_workload_slice = cfqd->workload_expires
998 - jiffies;
999 cfqg->saved_workload = cfqd->serving_type;
1000 cfqg->saved_serving_prio = cfqd->serving_prio;
1001 } else
1002 cfqg->saved_workload_slice = 0;
1004 cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1005 st->min_vdisktime);
1006 cfq_log_cfqq(cfqq->cfqd, cfqq,
1007 "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1008 used_sl, cfqq->slice_dispatch, charge,
1009 iops_mode(cfqd), cfqq->nr_sectors);
1010 cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1011 unaccounted_sl);
1012 cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1015 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1016 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1018 if (blkg)
1019 return container_of(blkg, struct cfq_group, blkg);
1020 return NULL;
1023 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1024 unsigned int weight)
1026 struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1027 cfqg->new_weight = weight;
1028 cfqg->needs_update = true;
1031 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1032 struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1034 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1035 unsigned int major, minor;
1038 * Add group onto cgroup list. It might happen that bdi->dev is
1039 * not initialized yet. Initialize this new group without major
1040 * and minor info and this info will be filled in once a new thread
1041 * comes for IO.
1043 if (bdi->dev) {
1044 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1045 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1046 (void *)cfqd, MKDEV(major, minor));
1047 } else
1048 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1049 (void *)cfqd, 0);
1051 cfqd->nr_blkcg_linked_grps++;
1052 cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1054 /* Add group on cfqd list */
1055 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1059 * Should be called from sleepable context. No request queue lock as per
1060 * cpu stats are allocated dynamically and alloc_percpu needs to be called
1061 * from sleepable context.
1063 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1065 struct cfq_group *cfqg = NULL;
1066 int i, j, ret;
1067 struct cfq_rb_root *st;
1069 cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1070 if (!cfqg)
1071 return NULL;
1073 for_each_cfqg_st(cfqg, i, j, st)
1074 *st = CFQ_RB_ROOT;
1075 RB_CLEAR_NODE(&cfqg->rb_node);
1077 cfqg->ttime.last_end_request = jiffies;
1080 * Take the initial reference that will be released on destroy
1081 * This can be thought of a joint reference by cgroup and
1082 * elevator which will be dropped by either elevator exit
1083 * or cgroup deletion path depending on who is exiting first.
1085 cfqg->ref = 1;
1087 ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1088 if (ret) {
1089 kfree(cfqg);
1090 return NULL;
1093 return cfqg;
1096 static struct cfq_group *
1097 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1099 struct cfq_group *cfqg = NULL;
1100 void *key = cfqd;
1101 struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1102 unsigned int major, minor;
1105 * This is the common case when there are no blkio cgroups.
1106 * Avoid lookup in this case
1108 if (blkcg == &blkio_root_cgroup)
1109 cfqg = &cfqd->root_group;
1110 else
1111 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1113 if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1114 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1115 cfqg->blkg.dev = MKDEV(major, minor);
1118 return cfqg;
1122 * Search for the cfq group current task belongs to. request_queue lock must
1123 * be held.
1125 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1127 struct blkio_cgroup *blkcg;
1128 struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1129 struct request_queue *q = cfqd->queue;
1131 rcu_read_lock();
1132 blkcg = task_blkio_cgroup(current);
1133 cfqg = cfq_find_cfqg(cfqd, blkcg);
1134 if (cfqg) {
1135 rcu_read_unlock();
1136 return cfqg;
1140 * Need to allocate a group. Allocation of group also needs allocation
1141 * of per cpu stats which in-turn takes a mutex() and can block. Hence
1142 * we need to drop rcu lock and queue_lock before we call alloc.
1144 * Not taking any queue reference here and assuming that queue is
1145 * around by the time we return. CFQ queue allocation code does
1146 * the same. It might be racy though.
1149 rcu_read_unlock();
1150 spin_unlock_irq(q->queue_lock);
1152 cfqg = cfq_alloc_cfqg(cfqd);
1154 spin_lock_irq(q->queue_lock);
1156 rcu_read_lock();
1157 blkcg = task_blkio_cgroup(current);
1160 * If some other thread already allocated the group while we were
1161 * not holding queue lock, free up the group
1163 __cfqg = cfq_find_cfqg(cfqd, blkcg);
1165 if (__cfqg) {
1166 kfree(cfqg);
1167 rcu_read_unlock();
1168 return __cfqg;
1171 if (!cfqg)
1172 cfqg = &cfqd->root_group;
1174 cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1175 rcu_read_unlock();
1176 return cfqg;
1179 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1181 cfqg->ref++;
1182 return cfqg;
1185 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1187 /* Currently, all async queues are mapped to root group */
1188 if (!cfq_cfqq_sync(cfqq))
1189 cfqg = &cfqq->cfqd->root_group;
1191 cfqq->cfqg = cfqg;
1192 /* cfqq reference on cfqg */
1193 cfqq->cfqg->ref++;
1196 static void cfq_put_cfqg(struct cfq_group *cfqg)
1198 struct cfq_rb_root *st;
1199 int i, j;
1201 BUG_ON(cfqg->ref <= 0);
1202 cfqg->ref--;
1203 if (cfqg->ref)
1204 return;
1205 for_each_cfqg_st(cfqg, i, j, st)
1206 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1207 free_percpu(cfqg->blkg.stats_cpu);
1208 kfree(cfqg);
1211 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1213 /* Something wrong if we are trying to remove same group twice */
1214 BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1216 hlist_del_init(&cfqg->cfqd_node);
1218 BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1219 cfqd->nr_blkcg_linked_grps--;
1222 * Put the reference taken at the time of creation so that when all
1223 * queues are gone, group can be destroyed.
1225 cfq_put_cfqg(cfqg);
1228 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1230 struct hlist_node *pos, *n;
1231 struct cfq_group *cfqg;
1233 hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1235 * If cgroup removal path got to blk_group first and removed
1236 * it from cgroup list, then it will take care of destroying
1237 * cfqg also.
1239 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1240 cfq_destroy_cfqg(cfqd, cfqg);
1245 * Blk cgroup controller notification saying that blkio_group object is being
1246 * delinked as associated cgroup object is going away. That also means that
1247 * no new IO will come in this group. So get rid of this group as soon as
1248 * any pending IO in the group is finished.
1250 * This function is called under rcu_read_lock(). key is the rcu protected
1251 * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1252 * read lock.
1254 * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1255 * it should not be NULL as even if elevator was exiting, cgroup deltion
1256 * path got to it first.
1258 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1260 unsigned long flags;
1261 struct cfq_data *cfqd = key;
1263 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1264 cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1265 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1268 #else /* GROUP_IOSCHED */
1269 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1271 return &cfqd->root_group;
1274 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1276 return cfqg;
1279 static inline void
1280 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1281 cfqq->cfqg = cfqg;
1284 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1285 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1287 #endif /* GROUP_IOSCHED */
1290 * The cfqd->service_trees holds all pending cfq_queue's that have
1291 * requests waiting to be processed. It is sorted in the order that
1292 * we will service the queues.
1294 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1295 bool add_front)
1297 struct rb_node **p, *parent;
1298 struct cfq_queue *__cfqq;
1299 unsigned long rb_key;
1300 struct cfq_rb_root *service_tree;
1301 int left;
1302 int new_cfqq = 1;
1304 service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1305 cfqq_type(cfqq));
1306 if (cfq_class_idle(cfqq)) {
1307 rb_key = CFQ_IDLE_DELAY;
1308 parent = rb_last(&service_tree->rb);
1309 if (parent && parent != &cfqq->rb_node) {
1310 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1311 rb_key += __cfqq->rb_key;
1312 } else
1313 rb_key += jiffies;
1314 } else if (!add_front) {
1316 * Get our rb key offset. Subtract any residual slice
1317 * value carried from last service. A negative resid
1318 * count indicates slice overrun, and this should position
1319 * the next service time further away in the tree.
1321 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1322 rb_key -= cfqq->slice_resid;
1323 cfqq->slice_resid = 0;
1324 } else {
1325 rb_key = -HZ;
1326 __cfqq = cfq_rb_first(service_tree);
1327 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1330 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1331 new_cfqq = 0;
1333 * same position, nothing more to do
1335 if (rb_key == cfqq->rb_key &&
1336 cfqq->service_tree == service_tree)
1337 return;
1339 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1340 cfqq->service_tree = NULL;
1343 left = 1;
1344 parent = NULL;
1345 cfqq->service_tree = service_tree;
1346 p = &service_tree->rb.rb_node;
1347 while (*p) {
1348 struct rb_node **n;
1350 parent = *p;
1351 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1354 * sort by key, that represents service time.
1356 if (time_before(rb_key, __cfqq->rb_key))
1357 n = &(*p)->rb_left;
1358 else {
1359 n = &(*p)->rb_right;
1360 left = 0;
1363 p = n;
1366 if (left)
1367 service_tree->left = &cfqq->rb_node;
1369 cfqq->rb_key = rb_key;
1370 rb_link_node(&cfqq->rb_node, parent, p);
1371 rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1372 service_tree->count++;
1373 if (add_front || !new_cfqq)
1374 return;
1375 cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1378 static struct cfq_queue *
1379 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1380 sector_t sector, struct rb_node **ret_parent,
1381 struct rb_node ***rb_link)
1383 struct rb_node **p, *parent;
1384 struct cfq_queue *cfqq = NULL;
1386 parent = NULL;
1387 p = &root->rb_node;
1388 while (*p) {
1389 struct rb_node **n;
1391 parent = *p;
1392 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1395 * Sort strictly based on sector. Smallest to the left,
1396 * largest to the right.
1398 if (sector > blk_rq_pos(cfqq->next_rq))
1399 n = &(*p)->rb_right;
1400 else if (sector < blk_rq_pos(cfqq->next_rq))
1401 n = &(*p)->rb_left;
1402 else
1403 break;
1404 p = n;
1405 cfqq = NULL;
1408 *ret_parent = parent;
1409 if (rb_link)
1410 *rb_link = p;
1411 return cfqq;
1414 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1416 struct rb_node **p, *parent;
1417 struct cfq_queue *__cfqq;
1419 if (cfqq->p_root) {
1420 rb_erase(&cfqq->p_node, cfqq->p_root);
1421 cfqq->p_root = NULL;
1424 if (cfq_class_idle(cfqq))
1425 return;
1426 if (!cfqq->next_rq)
1427 return;
1429 cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1430 __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1431 blk_rq_pos(cfqq->next_rq), &parent, &p);
1432 if (!__cfqq) {
1433 rb_link_node(&cfqq->p_node, parent, p);
1434 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1435 } else
1436 cfqq->p_root = NULL;
1440 * Update cfqq's position in the service tree.
1442 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1445 * Resorting requires the cfqq to be on the RR list already.
1447 if (cfq_cfqq_on_rr(cfqq)) {
1448 cfq_service_tree_add(cfqd, cfqq, 0);
1449 cfq_prio_tree_add(cfqd, cfqq);
1454 * add to busy list of queues for service, trying to be fair in ordering
1455 * the pending list according to last request service
1457 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1459 cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1460 BUG_ON(cfq_cfqq_on_rr(cfqq));
1461 cfq_mark_cfqq_on_rr(cfqq);
1462 cfqd->busy_queues++;
1463 if (cfq_cfqq_sync(cfqq))
1464 cfqd->busy_sync_queues++;
1466 cfq_resort_rr_list(cfqd, cfqq);
1470 * Called when the cfqq no longer has requests pending, remove it from
1471 * the service tree.
1473 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1475 cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1476 BUG_ON(!cfq_cfqq_on_rr(cfqq));
1477 cfq_clear_cfqq_on_rr(cfqq);
1479 if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1480 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1481 cfqq->service_tree = NULL;
1483 if (cfqq->p_root) {
1484 rb_erase(&cfqq->p_node, cfqq->p_root);
1485 cfqq->p_root = NULL;
1488 cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1489 BUG_ON(!cfqd->busy_queues);
1490 cfqd->busy_queues--;
1491 if (cfq_cfqq_sync(cfqq))
1492 cfqd->busy_sync_queues--;
1496 * rb tree support functions
1498 static void cfq_del_rq_rb(struct request *rq)
1500 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1501 const int sync = rq_is_sync(rq);
1503 BUG_ON(!cfqq->queued[sync]);
1504 cfqq->queued[sync]--;
1506 elv_rb_del(&cfqq->sort_list, rq);
1508 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1510 * Queue will be deleted from service tree when we actually
1511 * expire it later. Right now just remove it from prio tree
1512 * as it is empty.
1514 if (cfqq->p_root) {
1515 rb_erase(&cfqq->p_node, cfqq->p_root);
1516 cfqq->p_root = NULL;
1521 static void cfq_add_rq_rb(struct request *rq)
1523 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1524 struct cfq_data *cfqd = cfqq->cfqd;
1525 struct request *prev;
1527 cfqq->queued[rq_is_sync(rq)]++;
1529 elv_rb_add(&cfqq->sort_list, rq);
1531 if (!cfq_cfqq_on_rr(cfqq))
1532 cfq_add_cfqq_rr(cfqd, cfqq);
1535 * check if this request is a better next-serve candidate
1537 prev = cfqq->next_rq;
1538 cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1541 * adjust priority tree position, if ->next_rq changes
1543 if (prev != cfqq->next_rq)
1544 cfq_prio_tree_add(cfqd, cfqq);
1546 BUG_ON(!cfqq->next_rq);
1549 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1551 elv_rb_del(&cfqq->sort_list, rq);
1552 cfqq->queued[rq_is_sync(rq)]--;
1553 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1554 rq_data_dir(rq), rq_is_sync(rq));
1555 cfq_add_rq_rb(rq);
1556 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1557 &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1558 rq_is_sync(rq));
1561 static struct request *
1562 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1564 struct task_struct *tsk = current;
1565 struct cfq_io_cq *cic;
1566 struct cfq_queue *cfqq;
1568 cic = cfq_cic_lookup(cfqd, tsk->io_context);
1569 if (!cic)
1570 return NULL;
1572 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1573 if (cfqq) {
1574 sector_t sector = bio->bi_sector + bio_sectors(bio);
1576 return elv_rb_find(&cfqq->sort_list, sector);
1579 return NULL;
1582 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1584 struct cfq_data *cfqd = q->elevator->elevator_data;
1586 cfqd->rq_in_driver++;
1587 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1588 cfqd->rq_in_driver);
1590 cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1593 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1595 struct cfq_data *cfqd = q->elevator->elevator_data;
1597 WARN_ON(!cfqd->rq_in_driver);
1598 cfqd->rq_in_driver--;
1599 cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1600 cfqd->rq_in_driver);
1603 static void cfq_remove_request(struct request *rq)
1605 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1607 if (cfqq->next_rq == rq)
1608 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1610 list_del_init(&rq->queuelist);
1611 cfq_del_rq_rb(rq);
1613 cfqq->cfqd->rq_queued--;
1614 cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1615 rq_data_dir(rq), rq_is_sync(rq));
1616 if (rq->cmd_flags & REQ_PRIO) {
1617 WARN_ON(!cfqq->prio_pending);
1618 cfqq->prio_pending--;
1622 static int cfq_merge(struct request_queue *q, struct request **req,
1623 struct bio *bio)
1625 struct cfq_data *cfqd = q->elevator->elevator_data;
1626 struct request *__rq;
1628 __rq = cfq_find_rq_fmerge(cfqd, bio);
1629 if (__rq && elv_rq_merge_ok(__rq, bio)) {
1630 *req = __rq;
1631 return ELEVATOR_FRONT_MERGE;
1634 return ELEVATOR_NO_MERGE;
1637 static void cfq_merged_request(struct request_queue *q, struct request *req,
1638 int type)
1640 if (type == ELEVATOR_FRONT_MERGE) {
1641 struct cfq_queue *cfqq = RQ_CFQQ(req);
1643 cfq_reposition_rq_rb(cfqq, req);
1647 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1648 struct bio *bio)
1650 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1651 bio_data_dir(bio), cfq_bio_sync(bio));
1654 static void
1655 cfq_merged_requests(struct request_queue *q, struct request *rq,
1656 struct request *next)
1658 struct cfq_queue *cfqq = RQ_CFQQ(rq);
1659 struct cfq_data *cfqd = q->elevator->elevator_data;
1662 * reposition in fifo if next is older than rq
1664 if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1665 time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1666 list_move(&rq->queuelist, &next->queuelist);
1667 rq_set_fifo_time(rq, rq_fifo_time(next));
1670 if (cfqq->next_rq == next)
1671 cfqq->next_rq = rq;
1672 cfq_remove_request(next);
1673 cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1674 rq_data_dir(next), rq_is_sync(next));
1676 cfqq = RQ_CFQQ(next);
1678 * all requests of this queue are merged to other queues, delete it
1679 * from the service tree. If it's the active_queue,
1680 * cfq_dispatch_requests() will choose to expire it or do idle
1682 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1683 cfqq != cfqd->active_queue)
1684 cfq_del_cfqq_rr(cfqd, cfqq);
1687 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1688 struct bio *bio)
1690 struct cfq_data *cfqd = q->elevator->elevator_data;
1691 struct cfq_io_cq *cic;
1692 struct cfq_queue *cfqq;
1695 * Disallow merge of a sync bio into an async request.
1697 if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1698 return false;
1701 * Lookup the cfqq that this bio will be queued with and allow
1702 * merge only if rq is queued there.
1704 cic = cfq_cic_lookup(cfqd, current->io_context);
1705 if (!cic)
1706 return false;
1708 cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1709 return cfqq == RQ_CFQQ(rq);
1712 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1714 del_timer(&cfqd->idle_slice_timer);
1715 cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1718 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1719 struct cfq_queue *cfqq)
1721 if (cfqq) {
1722 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1723 cfqd->serving_prio, cfqd->serving_type);
1724 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1725 cfqq->slice_start = 0;
1726 cfqq->dispatch_start = jiffies;
1727 cfqq->allocated_slice = 0;
1728 cfqq->slice_end = 0;
1729 cfqq->slice_dispatch = 0;
1730 cfqq->nr_sectors = 0;
1732 cfq_clear_cfqq_wait_request(cfqq);
1733 cfq_clear_cfqq_must_dispatch(cfqq);
1734 cfq_clear_cfqq_must_alloc_slice(cfqq);
1735 cfq_clear_cfqq_fifo_expire(cfqq);
1736 cfq_mark_cfqq_slice_new(cfqq);
1738 cfq_del_timer(cfqd, cfqq);
1741 cfqd->active_queue = cfqq;
1745 * current cfqq expired its slice (or was too idle), select new one
1747 static void
1748 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1749 bool timed_out)
1751 cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1753 if (cfq_cfqq_wait_request(cfqq))
1754 cfq_del_timer(cfqd, cfqq);
1756 cfq_clear_cfqq_wait_request(cfqq);
1757 cfq_clear_cfqq_wait_busy(cfqq);
1760 * If this cfqq is shared between multiple processes, check to
1761 * make sure that those processes are still issuing I/Os within
1762 * the mean seek distance. If not, it may be time to break the
1763 * queues apart again.
1765 if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1766 cfq_mark_cfqq_split_coop(cfqq);
1769 * store what was left of this slice, if the queue idled/timed out
1771 if (timed_out) {
1772 if (cfq_cfqq_slice_new(cfqq))
1773 cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1774 else
1775 cfqq->slice_resid = cfqq->slice_end - jiffies;
1776 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1779 cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1781 if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1782 cfq_del_cfqq_rr(cfqd, cfqq);
1784 cfq_resort_rr_list(cfqd, cfqq);
1786 if (cfqq == cfqd->active_queue)
1787 cfqd->active_queue = NULL;
1789 if (cfqd->active_cic) {
1790 put_io_context(cfqd->active_cic->icq.ioc);
1791 cfqd->active_cic = NULL;
1795 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1797 struct cfq_queue *cfqq = cfqd->active_queue;
1799 if (cfqq)
1800 __cfq_slice_expired(cfqd, cfqq, timed_out);
1804 * Get next queue for service. Unless we have a queue preemption,
1805 * we'll simply select the first cfqq in the service tree.
1807 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1809 struct cfq_rb_root *service_tree =
1810 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1811 cfqd->serving_type);
1813 if (!cfqd->rq_queued)
1814 return NULL;
1816 /* There is nothing to dispatch */
1817 if (!service_tree)
1818 return NULL;
1819 if (RB_EMPTY_ROOT(&service_tree->rb))
1820 return NULL;
1821 return cfq_rb_first(service_tree);
1824 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1826 struct cfq_group *cfqg;
1827 struct cfq_queue *cfqq;
1828 int i, j;
1829 struct cfq_rb_root *st;
1831 if (!cfqd->rq_queued)
1832 return NULL;
1834 cfqg = cfq_get_next_cfqg(cfqd);
1835 if (!cfqg)
1836 return NULL;
1838 for_each_cfqg_st(cfqg, i, j, st)
1839 if ((cfqq = cfq_rb_first(st)) != NULL)
1840 return cfqq;
1841 return NULL;
1845 * Get and set a new active queue for service.
1847 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1848 struct cfq_queue *cfqq)
1850 if (!cfqq)
1851 cfqq = cfq_get_next_queue(cfqd);
1853 __cfq_set_active_queue(cfqd, cfqq);
1854 return cfqq;
1857 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1858 struct request *rq)
1860 if (blk_rq_pos(rq) >= cfqd->last_position)
1861 return blk_rq_pos(rq) - cfqd->last_position;
1862 else
1863 return cfqd->last_position - blk_rq_pos(rq);
1866 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1867 struct request *rq)
1869 return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1872 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1873 struct cfq_queue *cur_cfqq)
1875 struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1876 struct rb_node *parent, *node;
1877 struct cfq_queue *__cfqq;
1878 sector_t sector = cfqd->last_position;
1880 if (RB_EMPTY_ROOT(root))
1881 return NULL;
1884 * First, if we find a request starting at the end of the last
1885 * request, choose it.
1887 __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1888 if (__cfqq)
1889 return __cfqq;
1892 * If the exact sector wasn't found, the parent of the NULL leaf
1893 * will contain the closest sector.
1895 __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1896 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1897 return __cfqq;
1899 if (blk_rq_pos(__cfqq->next_rq) < sector)
1900 node = rb_next(&__cfqq->p_node);
1901 else
1902 node = rb_prev(&__cfqq->p_node);
1903 if (!node)
1904 return NULL;
1906 __cfqq = rb_entry(node, struct cfq_queue, p_node);
1907 if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1908 return __cfqq;
1910 return NULL;
1914 * cfqd - obvious
1915 * cur_cfqq - passed in so that we don't decide that the current queue is
1916 * closely cooperating with itself.
1918 * So, basically we're assuming that that cur_cfqq has dispatched at least
1919 * one request, and that cfqd->last_position reflects a position on the disk
1920 * associated with the I/O issued by cur_cfqq. I'm not sure this is a valid
1921 * assumption.
1923 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1924 struct cfq_queue *cur_cfqq)
1926 struct cfq_queue *cfqq;
1928 if (cfq_class_idle(cur_cfqq))
1929 return NULL;
1930 if (!cfq_cfqq_sync(cur_cfqq))
1931 return NULL;
1932 if (CFQQ_SEEKY(cur_cfqq))
1933 return NULL;
1936 * Don't search priority tree if it's the only queue in the group.
1938 if (cur_cfqq->cfqg->nr_cfqq == 1)
1939 return NULL;
1942 * We should notice if some of the queues are cooperating, eg
1943 * working closely on the same area of the disk. In that case,
1944 * we can group them together and don't waste time idling.
1946 cfqq = cfqq_close(cfqd, cur_cfqq);
1947 if (!cfqq)
1948 return NULL;
1950 /* If new queue belongs to different cfq_group, don't choose it */
1951 if (cur_cfqq->cfqg != cfqq->cfqg)
1952 return NULL;
1955 * It only makes sense to merge sync queues.
1957 if (!cfq_cfqq_sync(cfqq))
1958 return NULL;
1959 if (CFQQ_SEEKY(cfqq))
1960 return NULL;
1963 * Do not merge queues of different priority classes
1965 if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1966 return NULL;
1968 return cfqq;
1972 * Determine whether we should enforce idle window for this queue.
1975 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1977 enum wl_prio_t prio = cfqq_prio(cfqq);
1978 struct cfq_rb_root *service_tree = cfqq->service_tree;
1980 BUG_ON(!service_tree);
1981 BUG_ON(!service_tree->count);
1983 if (!cfqd->cfq_slice_idle)
1984 return false;
1986 /* We never do for idle class queues. */
1987 if (prio == IDLE_WORKLOAD)
1988 return false;
1990 /* We do for queues that were marked with idle window flag. */
1991 if (cfq_cfqq_idle_window(cfqq) &&
1992 !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1993 return true;
1996 * Otherwise, we do only if they are the last ones
1997 * in their service tree.
1999 if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2000 !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2001 return true;
2002 cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2003 service_tree->count);
2004 return false;
2007 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2009 struct cfq_queue *cfqq = cfqd->active_queue;
2010 struct cfq_io_cq *cic;
2011 unsigned long sl, group_idle = 0;
2014 * SSD device without seek penalty, disable idling. But only do so
2015 * for devices that support queuing, otherwise we still have a problem
2016 * with sync vs async workloads.
2018 if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2019 return;
2021 WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2022 WARN_ON(cfq_cfqq_slice_new(cfqq));
2025 * idle is disabled, either manually or by past process history
2027 if (!cfq_should_idle(cfqd, cfqq)) {
2028 /* no queue idling. Check for group idling */
2029 if (cfqd->cfq_group_idle)
2030 group_idle = cfqd->cfq_group_idle;
2031 else
2032 return;
2036 * still active requests from this queue, don't idle
2038 if (cfqq->dispatched)
2039 return;
2042 * task has exited, don't wait
2044 cic = cfqd->active_cic;
2045 if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
2046 return;
2049 * If our average think time is larger than the remaining time
2050 * slice, then don't idle. This avoids overrunning the allotted
2051 * time slice.
2053 if (sample_valid(cic->ttime.ttime_samples) &&
2054 (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2055 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2056 cic->ttime.ttime_mean);
2057 return;
2060 /* There are other queues in the group, don't do group idle */
2061 if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2062 return;
2064 cfq_mark_cfqq_wait_request(cfqq);
2066 if (group_idle)
2067 sl = cfqd->cfq_group_idle;
2068 else
2069 sl = cfqd->cfq_slice_idle;
2071 mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2072 cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2073 cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2074 group_idle ? 1 : 0);
2078 * Move request from internal lists to the request queue dispatch list.
2080 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2082 struct cfq_data *cfqd = q->elevator->elevator_data;
2083 struct cfq_queue *cfqq = RQ_CFQQ(rq);
2085 cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2087 cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2088 cfq_remove_request(rq);
2089 cfqq->dispatched++;
2090 (RQ_CFQG(rq))->dispatched++;
2091 elv_dispatch_sort(q, rq);
2093 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2094 cfqq->nr_sectors += blk_rq_sectors(rq);
2095 cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2096 rq_data_dir(rq), rq_is_sync(rq));
2100 * return expired entry, or NULL to just start from scratch in rbtree
2102 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2104 struct request *rq = NULL;
2106 if (cfq_cfqq_fifo_expire(cfqq))
2107 return NULL;
2109 cfq_mark_cfqq_fifo_expire(cfqq);
2111 if (list_empty(&cfqq->fifo))
2112 return NULL;
2114 rq = rq_entry_fifo(cfqq->fifo.next);
2115 if (time_before(jiffies, rq_fifo_time(rq)))
2116 rq = NULL;
2118 cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2119 return rq;
2122 static inline int
2123 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2125 const int base_rq = cfqd->cfq_slice_async_rq;
2127 WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2129 return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2133 * Must be called with the queue_lock held.
2135 static int cfqq_process_refs(struct cfq_queue *cfqq)
2137 int process_refs, io_refs;
2139 io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2140 process_refs = cfqq->ref - io_refs;
2141 BUG_ON(process_refs < 0);
2142 return process_refs;
2145 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2147 int process_refs, new_process_refs;
2148 struct cfq_queue *__cfqq;
2151 * If there are no process references on the new_cfqq, then it is
2152 * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2153 * chain may have dropped their last reference (not just their
2154 * last process reference).
2156 if (!cfqq_process_refs(new_cfqq))
2157 return;
2159 /* Avoid a circular list and skip interim queue merges */
2160 while ((__cfqq = new_cfqq->new_cfqq)) {
2161 if (__cfqq == cfqq)
2162 return;
2163 new_cfqq = __cfqq;
2166 process_refs = cfqq_process_refs(cfqq);
2167 new_process_refs = cfqq_process_refs(new_cfqq);
2169 * If the process for the cfqq has gone away, there is no
2170 * sense in merging the queues.
2172 if (process_refs == 0 || new_process_refs == 0)
2173 return;
2176 * Merge in the direction of the lesser amount of work.
2178 if (new_process_refs >= process_refs) {
2179 cfqq->new_cfqq = new_cfqq;
2180 new_cfqq->ref += process_refs;
2181 } else {
2182 new_cfqq->new_cfqq = cfqq;
2183 cfqq->ref += new_process_refs;
2187 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2188 struct cfq_group *cfqg, enum wl_prio_t prio)
2190 struct cfq_queue *queue;
2191 int i;
2192 bool key_valid = false;
2193 unsigned long lowest_key = 0;
2194 enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2196 for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2197 /* select the one with lowest rb_key */
2198 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2199 if (queue &&
2200 (!key_valid || time_before(queue->rb_key, lowest_key))) {
2201 lowest_key = queue->rb_key;
2202 cur_best = i;
2203 key_valid = true;
2207 return cur_best;
2210 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2212 unsigned slice;
2213 unsigned count;
2214 struct cfq_rb_root *st;
2215 unsigned group_slice;
2216 enum wl_prio_t original_prio = cfqd->serving_prio;
2218 /* Choose next priority. RT > BE > IDLE */
2219 if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2220 cfqd->serving_prio = RT_WORKLOAD;
2221 else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2222 cfqd->serving_prio = BE_WORKLOAD;
2223 else {
2224 cfqd->serving_prio = IDLE_WORKLOAD;
2225 cfqd->workload_expires = jiffies + 1;
2226 return;
2229 if (original_prio != cfqd->serving_prio)
2230 goto new_workload;
2233 * For RT and BE, we have to choose also the type
2234 * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2235 * expiration time
2237 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2238 count = st->count;
2241 * check workload expiration, and that we still have other queues ready
2243 if (count && !time_after(jiffies, cfqd->workload_expires))
2244 return;
2246 new_workload:
2247 /* otherwise select new workload type */
2248 cfqd->serving_type =
2249 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2250 st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2251 count = st->count;
2254 * the workload slice is computed as a fraction of target latency
2255 * proportional to the number of queues in that workload, over
2256 * all the queues in the same priority class
2258 group_slice = cfq_group_slice(cfqd, cfqg);
2260 slice = group_slice * count /
2261 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2262 cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2264 if (cfqd->serving_type == ASYNC_WORKLOAD) {
2265 unsigned int tmp;
2268 * Async queues are currently system wide. Just taking
2269 * proportion of queues with-in same group will lead to higher
2270 * async ratio system wide as generally root group is going
2271 * to have higher weight. A more accurate thing would be to
2272 * calculate system wide asnc/sync ratio.
2274 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2275 tmp = tmp/cfqd->busy_queues;
2276 slice = min_t(unsigned, slice, tmp);
2278 /* async workload slice is scaled down according to
2279 * the sync/async slice ratio. */
2280 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2281 } else
2282 /* sync workload slice is at least 2 * cfq_slice_idle */
2283 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2285 slice = max_t(unsigned, slice, CFQ_MIN_TT);
2286 cfq_log(cfqd, "workload slice:%d", slice);
2287 cfqd->workload_expires = jiffies + slice;
2290 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2292 struct cfq_rb_root *st = &cfqd->grp_service_tree;
2293 struct cfq_group *cfqg;
2295 if (RB_EMPTY_ROOT(&st->rb))
2296 return NULL;
2297 cfqg = cfq_rb_first_group(st);
2298 update_min_vdisktime(st);
2299 return cfqg;
2302 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2304 struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2306 cfqd->serving_group = cfqg;
2308 /* Restore the workload type data */
2309 if (cfqg->saved_workload_slice) {
2310 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2311 cfqd->serving_type = cfqg->saved_workload;
2312 cfqd->serving_prio = cfqg->saved_serving_prio;
2313 } else
2314 cfqd->workload_expires = jiffies - 1;
2316 choose_service_tree(cfqd, cfqg);
2320 * Select a queue for service. If we have a current active queue,
2321 * check whether to continue servicing it, or retrieve and set a new one.
2323 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2325 struct cfq_queue *cfqq, *new_cfqq = NULL;
2327 cfqq = cfqd->active_queue;
2328 if (!cfqq)
2329 goto new_queue;
2331 if (!cfqd->rq_queued)
2332 return NULL;
2335 * We were waiting for group to get backlogged. Expire the queue
2337 if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2338 goto expire;
2341 * The active queue has run out of time, expire it and select new.
2343 if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2345 * If slice had not expired at the completion of last request
2346 * we might not have turned on wait_busy flag. Don't expire
2347 * the queue yet. Allow the group to get backlogged.
2349 * The very fact that we have used the slice, that means we
2350 * have been idling all along on this queue and it should be
2351 * ok to wait for this request to complete.
2353 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2354 && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2355 cfqq = NULL;
2356 goto keep_queue;
2357 } else
2358 goto check_group_idle;
2362 * The active queue has requests and isn't expired, allow it to
2363 * dispatch.
2365 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2366 goto keep_queue;
2369 * If another queue has a request waiting within our mean seek
2370 * distance, let it run. The expire code will check for close
2371 * cooperators and put the close queue at the front of the service
2372 * tree. If possible, merge the expiring queue with the new cfqq.
2374 new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2375 if (new_cfqq) {
2376 if (!cfqq->new_cfqq)
2377 cfq_setup_merge(cfqq, new_cfqq);
2378 goto expire;
2382 * No requests pending. If the active queue still has requests in
2383 * flight or is idling for a new request, allow either of these
2384 * conditions to happen (or time out) before selecting a new queue.
2386 if (timer_pending(&cfqd->idle_slice_timer)) {
2387 cfqq = NULL;
2388 goto keep_queue;
2392 * This is a deep seek queue, but the device is much faster than
2393 * the queue can deliver, don't idle
2395 if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2396 (cfq_cfqq_slice_new(cfqq) ||
2397 (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2398 cfq_clear_cfqq_deep(cfqq);
2399 cfq_clear_cfqq_idle_window(cfqq);
2402 if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2403 cfqq = NULL;
2404 goto keep_queue;
2408 * If group idle is enabled and there are requests dispatched from
2409 * this group, wait for requests to complete.
2411 check_group_idle:
2412 if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2413 cfqq->cfqg->dispatched &&
2414 !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2415 cfqq = NULL;
2416 goto keep_queue;
2419 expire:
2420 cfq_slice_expired(cfqd, 0);
2421 new_queue:
2423 * Current queue expired. Check if we have to switch to a new
2424 * service tree
2426 if (!new_cfqq)
2427 cfq_choose_cfqg(cfqd);
2429 cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2430 keep_queue:
2431 return cfqq;
2434 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2436 int dispatched = 0;
2438 while (cfqq->next_rq) {
2439 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2440 dispatched++;
2443 BUG_ON(!list_empty(&cfqq->fifo));
2445 /* By default cfqq is not expired if it is empty. Do it explicitly */
2446 __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2447 return dispatched;
2451 * Drain our current requests. Used for barriers and when switching
2452 * io schedulers on-the-fly.
2454 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2456 struct cfq_queue *cfqq;
2457 int dispatched = 0;
2459 /* Expire the timeslice of the current active queue first */
2460 cfq_slice_expired(cfqd, 0);
2461 while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2462 __cfq_set_active_queue(cfqd, cfqq);
2463 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2466 BUG_ON(cfqd->busy_queues);
2468 cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2469 return dispatched;
2472 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2473 struct cfq_queue *cfqq)
2475 /* the queue hasn't finished any request, can't estimate */
2476 if (cfq_cfqq_slice_new(cfqq))
2477 return true;
2478 if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2479 cfqq->slice_end))
2480 return true;
2482 return false;
2485 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2487 unsigned int max_dispatch;
2490 * Drain async requests before we start sync IO
2492 if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2493 return false;
2496 * If this is an async queue and we have sync IO in flight, let it wait
2498 if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2499 return false;
2501 max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2502 if (cfq_class_idle(cfqq))
2503 max_dispatch = 1;
2506 * Does this cfqq already have too much IO in flight?
2508 if (cfqq->dispatched >= max_dispatch) {
2509 bool promote_sync = false;
2511 * idle queue must always only have a single IO in flight
2513 if (cfq_class_idle(cfqq))
2514 return false;
2517 * If there is only one sync queue
2518 * we can ignore async queue here and give the sync
2519 * queue no dispatch limit. The reason is a sync queue can
2520 * preempt async queue, limiting the sync queue doesn't make
2521 * sense. This is useful for aiostress test.
2523 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2524 promote_sync = true;
2527 * We have other queues, don't allow more IO from this one
2529 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2530 !promote_sync)
2531 return false;
2534 * Sole queue user, no limit
2536 if (cfqd->busy_queues == 1 || promote_sync)
2537 max_dispatch = -1;
2538 else
2540 * Normally we start throttling cfqq when cfq_quantum/2
2541 * requests have been dispatched. But we can drive
2542 * deeper queue depths at the beginning of slice
2543 * subjected to upper limit of cfq_quantum.
2544 * */
2545 max_dispatch = cfqd->cfq_quantum;
2549 * Async queues must wait a bit before being allowed dispatch.
2550 * We also ramp up the dispatch depth gradually for async IO,
2551 * based on the last sync IO we serviced
2553 if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2554 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2555 unsigned int depth;
2557 depth = last_sync / cfqd->cfq_slice[1];
2558 if (!depth && !cfqq->dispatched)
2559 depth = 1;
2560 if (depth < max_dispatch)
2561 max_dispatch = depth;
2565 * If we're below the current max, allow a dispatch
2567 return cfqq->dispatched < max_dispatch;
2571 * Dispatch a request from cfqq, moving them to the request queue
2572 * dispatch list.
2574 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2576 struct request *rq;
2578 BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2580 if (!cfq_may_dispatch(cfqd, cfqq))
2581 return false;
2584 * follow expired path, else get first next available
2586 rq = cfq_check_fifo(cfqq);
2587 if (!rq)
2588 rq = cfqq->next_rq;
2591 * insert request into driver dispatch list
2593 cfq_dispatch_insert(cfqd->queue, rq);
2595 if (!cfqd->active_cic) {
2596 struct cfq_io_cq *cic = RQ_CIC(rq);
2598 atomic_long_inc(&cic->icq.ioc->refcount);
2599 cfqd->active_cic = cic;
2602 return true;
2606 * Find the cfqq that we need to service and move a request from that to the
2607 * dispatch list
2609 static int cfq_dispatch_requests(struct request_queue *q, int force)
2611 struct cfq_data *cfqd = q->elevator->elevator_data;
2612 struct cfq_queue *cfqq;
2614 if (!cfqd->busy_queues)
2615 return 0;
2617 if (unlikely(force))
2618 return cfq_forced_dispatch(cfqd);
2620 cfqq = cfq_select_queue(cfqd);
2621 if (!cfqq)
2622 return 0;
2625 * Dispatch a request from this cfqq, if it is allowed
2627 if (!cfq_dispatch_request(cfqd, cfqq))
2628 return 0;
2630 cfqq->slice_dispatch++;
2631 cfq_clear_cfqq_must_dispatch(cfqq);
2634 * expire an async queue immediately if it has used up its slice. idle
2635 * queue always expire after 1 dispatch round.
2637 if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2638 cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2639 cfq_class_idle(cfqq))) {
2640 cfqq->slice_end = jiffies + 1;
2641 cfq_slice_expired(cfqd, 0);
2644 cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2645 return 1;
2649 * task holds one reference to the queue, dropped when task exits. each rq
2650 * in-flight on this queue also holds a reference, dropped when rq is freed.
2652 * Each cfq queue took a reference on the parent group. Drop it now.
2653 * queue lock must be held here.
2655 static void cfq_put_queue(struct cfq_queue *cfqq)
2657 struct cfq_data *cfqd = cfqq->cfqd;
2658 struct cfq_group *cfqg;
2660 BUG_ON(cfqq->ref <= 0);
2662 cfqq->ref--;
2663 if (cfqq->ref)
2664 return;
2666 cfq_log_cfqq(cfqd, cfqq, "put_queue");
2667 BUG_ON(rb_first(&cfqq->sort_list));
2668 BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2669 cfqg = cfqq->cfqg;
2671 if (unlikely(cfqd->active_queue == cfqq)) {
2672 __cfq_slice_expired(cfqd, cfqq, 0);
2673 cfq_schedule_dispatch(cfqd);
2676 BUG_ON(cfq_cfqq_on_rr(cfqq));
2677 kmem_cache_free(cfq_pool, cfqq);
2678 cfq_put_cfqg(cfqg);
2681 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2683 struct cfq_queue *__cfqq, *next;
2686 * If this queue was scheduled to merge with another queue, be
2687 * sure to drop the reference taken on that queue (and others in
2688 * the merge chain). See cfq_setup_merge and cfq_merge_cfqqs.
2690 __cfqq = cfqq->new_cfqq;
2691 while (__cfqq) {
2692 if (__cfqq == cfqq) {
2693 WARN(1, "cfqq->new_cfqq loop detected\n");
2694 break;
2696 next = __cfqq->new_cfqq;
2697 cfq_put_queue(__cfqq);
2698 __cfqq = next;
2702 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2704 if (unlikely(cfqq == cfqd->active_queue)) {
2705 __cfq_slice_expired(cfqd, cfqq, 0);
2706 cfq_schedule_dispatch(cfqd);
2709 cfq_put_cooperator(cfqq);
2711 cfq_put_queue(cfqq);
2714 static void cfq_init_icq(struct io_cq *icq)
2716 struct cfq_io_cq *cic = icq_to_cic(icq);
2718 cic->ttime.last_end_request = jiffies;
2721 static void cfq_exit_icq(struct io_cq *icq)
2723 struct cfq_io_cq *cic = icq_to_cic(icq);
2724 struct cfq_data *cfqd = cic_to_cfqd(cic);
2726 if (cic->cfqq[BLK_RW_ASYNC]) {
2727 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2728 cic->cfqq[BLK_RW_ASYNC] = NULL;
2731 if (cic->cfqq[BLK_RW_SYNC]) {
2732 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2733 cic->cfqq[BLK_RW_SYNC] = NULL;
2737 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2739 struct task_struct *tsk = current;
2740 int ioprio_class;
2742 if (!cfq_cfqq_prio_changed(cfqq))
2743 return;
2745 ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2746 switch (ioprio_class) {
2747 default:
2748 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2749 case IOPRIO_CLASS_NONE:
2751 * no prio set, inherit CPU scheduling settings
2753 cfqq->ioprio = task_nice_ioprio(tsk);
2754 cfqq->ioprio_class = task_nice_ioclass(tsk);
2755 break;
2756 case IOPRIO_CLASS_RT:
2757 cfqq->ioprio = task_ioprio(ioc);
2758 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2759 break;
2760 case IOPRIO_CLASS_BE:
2761 cfqq->ioprio = task_ioprio(ioc);
2762 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2763 break;
2764 case IOPRIO_CLASS_IDLE:
2765 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2766 cfqq->ioprio = 7;
2767 cfq_clear_cfqq_idle_window(cfqq);
2768 break;
2772 * keep track of original prio settings in case we have to temporarily
2773 * elevate the priority of this queue
2775 cfqq->org_ioprio = cfqq->ioprio;
2776 cfq_clear_cfqq_prio_changed(cfqq);
2779 static void changed_ioprio(struct cfq_io_cq *cic)
2781 struct cfq_data *cfqd = cic_to_cfqd(cic);
2782 struct cfq_queue *cfqq;
2784 if (unlikely(!cfqd))
2785 return;
2787 cfqq = cic->cfqq[BLK_RW_ASYNC];
2788 if (cfqq) {
2789 struct cfq_queue *new_cfqq;
2790 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2791 GFP_ATOMIC);
2792 if (new_cfqq) {
2793 cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2794 cfq_put_queue(cfqq);
2798 cfqq = cic->cfqq[BLK_RW_SYNC];
2799 if (cfqq)
2800 cfq_mark_cfqq_prio_changed(cfqq);
2803 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2804 pid_t pid, bool is_sync)
2806 RB_CLEAR_NODE(&cfqq->rb_node);
2807 RB_CLEAR_NODE(&cfqq->p_node);
2808 INIT_LIST_HEAD(&cfqq->fifo);
2810 cfqq->ref = 0;
2811 cfqq->cfqd = cfqd;
2813 cfq_mark_cfqq_prio_changed(cfqq);
2815 if (is_sync) {
2816 if (!cfq_class_idle(cfqq))
2817 cfq_mark_cfqq_idle_window(cfqq);
2818 cfq_mark_cfqq_sync(cfqq);
2820 cfqq->pid = pid;
2823 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2824 static void changed_cgroup(struct cfq_io_cq *cic)
2826 struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2827 struct cfq_data *cfqd = cic_to_cfqd(cic);
2828 struct request_queue *q;
2830 if (unlikely(!cfqd))
2831 return;
2833 q = cfqd->queue;
2835 if (sync_cfqq) {
2837 * Drop reference to sync queue. A new sync queue will be
2838 * assigned in new group upon arrival of a fresh request.
2840 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2841 cic_set_cfqq(cic, NULL, 1);
2842 cfq_put_queue(sync_cfqq);
2845 #endif /* CONFIG_CFQ_GROUP_IOSCHED */
2847 static struct cfq_queue *
2848 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2849 struct io_context *ioc, gfp_t gfp_mask)
2851 struct cfq_queue *cfqq, *new_cfqq = NULL;
2852 struct cfq_io_cq *cic;
2853 struct cfq_group *cfqg;
2855 retry:
2856 cfqg = cfq_get_cfqg(cfqd);
2857 cic = cfq_cic_lookup(cfqd, ioc);
2858 /* cic always exists here */
2859 cfqq = cic_to_cfqq(cic, is_sync);
2862 * Always try a new alloc if we fell back to the OOM cfqq
2863 * originally, since it should just be a temporary situation.
2865 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2866 cfqq = NULL;
2867 if (new_cfqq) {
2868 cfqq = new_cfqq;
2869 new_cfqq = NULL;
2870 } else if (gfp_mask & __GFP_WAIT) {
2871 spin_unlock_irq(cfqd->queue->queue_lock);
2872 new_cfqq = kmem_cache_alloc_node(cfq_pool,
2873 gfp_mask | __GFP_ZERO,
2874 cfqd->queue->node);
2875 spin_lock_irq(cfqd->queue->queue_lock);
2876 if (new_cfqq)
2877 goto retry;
2878 } else {
2879 cfqq = kmem_cache_alloc_node(cfq_pool,
2880 gfp_mask | __GFP_ZERO,
2881 cfqd->queue->node);
2884 if (cfqq) {
2885 cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2886 cfq_init_prio_data(cfqq, ioc);
2887 cfq_link_cfqq_cfqg(cfqq, cfqg);
2888 cfq_log_cfqq(cfqd, cfqq, "alloced");
2889 } else
2890 cfqq = &cfqd->oom_cfqq;
2893 if (new_cfqq)
2894 kmem_cache_free(cfq_pool, new_cfqq);
2896 return cfqq;
2899 static struct cfq_queue **
2900 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2902 switch (ioprio_class) {
2903 case IOPRIO_CLASS_RT:
2904 return &cfqd->async_cfqq[0][ioprio];
2905 case IOPRIO_CLASS_BE:
2906 return &cfqd->async_cfqq[1][ioprio];
2907 case IOPRIO_CLASS_IDLE:
2908 return &cfqd->async_idle_cfqq;
2909 default:
2910 BUG();
2914 static struct cfq_queue *
2915 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2916 gfp_t gfp_mask)
2918 const int ioprio = task_ioprio(ioc);
2919 const int ioprio_class = task_ioprio_class(ioc);
2920 struct cfq_queue **async_cfqq = NULL;
2921 struct cfq_queue *cfqq = NULL;
2923 if (!is_sync) {
2924 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2925 cfqq = *async_cfqq;
2928 if (!cfqq)
2929 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2932 * pin the queue now that it's allocated, scheduler exit will prune it
2934 if (!is_sync && !(*async_cfqq)) {
2935 cfqq->ref++;
2936 *async_cfqq = cfqq;
2939 cfqq->ref++;
2940 return cfqq;
2943 static void
2944 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
2946 unsigned long elapsed = jiffies - ttime->last_end_request;
2947 elapsed = min(elapsed, 2UL * slice_idle);
2949 ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2950 ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2951 ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2954 static void
2955 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2956 struct cfq_io_cq *cic)
2958 if (cfq_cfqq_sync(cfqq)) {
2959 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
2960 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2961 cfqd->cfq_slice_idle);
2963 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2964 __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2965 #endif
2968 static void
2969 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2970 struct request *rq)
2972 sector_t sdist = 0;
2973 sector_t n_sec = blk_rq_sectors(rq);
2974 if (cfqq->last_request_pos) {
2975 if (cfqq->last_request_pos < blk_rq_pos(rq))
2976 sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2977 else
2978 sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2981 cfqq->seek_history <<= 1;
2982 if (blk_queue_nonrot(cfqd->queue))
2983 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2984 else
2985 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2989 * Disable idle window if the process thinks too long or seeks so much that
2990 * it doesn't matter
2992 static void
2993 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2994 struct cfq_io_cq *cic)
2996 int old_idle, enable_idle;
2999 * Don't idle for async or idle io prio class
3001 if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3002 return;
3004 enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3006 if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3007 cfq_mark_cfqq_deep(cfqq);
3009 if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3010 enable_idle = 0;
3011 else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3012 !cfqd->cfq_slice_idle ||
3013 (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3014 enable_idle = 0;
3015 else if (sample_valid(cic->ttime.ttime_samples)) {
3016 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3017 enable_idle = 0;
3018 else
3019 enable_idle = 1;
3022 if (old_idle != enable_idle) {
3023 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3024 if (enable_idle)
3025 cfq_mark_cfqq_idle_window(cfqq);
3026 else
3027 cfq_clear_cfqq_idle_window(cfqq);
3032 * Check if new_cfqq should preempt the currently active queue. Return 0 for
3033 * no or if we aren't sure, a 1 will cause a preempt.
3035 static bool
3036 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3037 struct request *rq)
3039 struct cfq_queue *cfqq;
3041 cfqq = cfqd->active_queue;
3042 if (!cfqq)
3043 return false;
3045 if (cfq_class_idle(new_cfqq))
3046 return false;
3048 if (cfq_class_idle(cfqq))
3049 return true;
3052 * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3054 if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3055 return false;
3058 * if the new request is sync, but the currently running queue is
3059 * not, let the sync request have priority.
3061 if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3062 return true;
3064 if (new_cfqq->cfqg != cfqq->cfqg)
3065 return false;
3067 if (cfq_slice_used(cfqq))
3068 return true;
3070 /* Allow preemption only if we are idling on sync-noidle tree */
3071 if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3072 cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3073 new_cfqq->service_tree->count == 2 &&
3074 RB_EMPTY_ROOT(&cfqq->sort_list))
3075 return true;
3078 * So both queues are sync. Let the new request get disk time if
3079 * it's a metadata request and the current queue is doing regular IO.
3081 if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3082 return true;
3085 * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3087 if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3088 return true;
3090 /* An idle queue should not be idle now for some reason */
3091 if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3092 return true;
3094 if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3095 return false;
3098 * if this request is as-good as one we would expect from the
3099 * current cfqq, let it preempt
3101 if (cfq_rq_close(cfqd, cfqq, rq))
3102 return true;
3104 return false;
3108 * cfqq preempts the active queue. if we allowed preempt with no slice left,
3109 * let it have half of its nominal slice.
3111 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3113 enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3115 cfq_log_cfqq(cfqd, cfqq, "preempt");
3116 cfq_slice_expired(cfqd, 1);
3119 * workload type is changed, don't save slice, otherwise preempt
3120 * doesn't happen
3122 if (old_type != cfqq_type(cfqq))
3123 cfqq->cfqg->saved_workload_slice = 0;
3126 * Put the new queue at the front of the of the current list,
3127 * so we know that it will be selected next.
3129 BUG_ON(!cfq_cfqq_on_rr(cfqq));
3131 cfq_service_tree_add(cfqd, cfqq, 1);
3133 cfqq->slice_end = 0;
3134 cfq_mark_cfqq_slice_new(cfqq);
3138 * Called when a new fs request (rq) is added (to cfqq). Check if there's
3139 * something we should do about it
3141 static void
3142 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3143 struct request *rq)
3145 struct cfq_io_cq *cic = RQ_CIC(rq);
3147 cfqd->rq_queued++;
3148 if (rq->cmd_flags & REQ_PRIO)
3149 cfqq->prio_pending++;
3151 cfq_update_io_thinktime(cfqd, cfqq, cic);
3152 cfq_update_io_seektime(cfqd, cfqq, rq);
3153 cfq_update_idle_window(cfqd, cfqq, cic);
3155 cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3157 if (cfqq == cfqd->active_queue) {
3159 * Remember that we saw a request from this process, but
3160 * don't start queuing just yet. Otherwise we risk seeing lots
3161 * of tiny requests, because we disrupt the normal plugging
3162 * and merging. If the request is already larger than a single
3163 * page, let it rip immediately. For that case we assume that
3164 * merging is already done. Ditto for a busy system that
3165 * has other work pending, don't risk delaying until the
3166 * idle timer unplug to continue working.
3168 if (cfq_cfqq_wait_request(cfqq)) {
3169 if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3170 cfqd->busy_queues > 1) {
3171 cfq_del_timer(cfqd, cfqq);
3172 cfq_clear_cfqq_wait_request(cfqq);
3173 __blk_run_queue(cfqd->queue);
3174 } else {
3175 cfq_blkiocg_update_idle_time_stats(
3176 &cfqq->cfqg->blkg);
3177 cfq_mark_cfqq_must_dispatch(cfqq);
3180 } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3182 * not the active queue - expire current slice if it is
3183 * idle and has expired it's mean thinktime or this new queue
3184 * has some old slice time left and is of higher priority or
3185 * this new queue is RT and the current one is BE
3187 cfq_preempt_queue(cfqd, cfqq);
3188 __blk_run_queue(cfqd->queue);
3192 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3194 struct cfq_data *cfqd = q->elevator->elevator_data;
3195 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3197 cfq_log_cfqq(cfqd, cfqq, "insert_request");
3198 cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3200 rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3201 list_add_tail(&rq->queuelist, &cfqq->fifo);
3202 cfq_add_rq_rb(rq);
3203 cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3204 &cfqd->serving_group->blkg, rq_data_dir(rq),
3205 rq_is_sync(rq));
3206 cfq_rq_enqueued(cfqd, cfqq, rq);
3210 * Update hw_tag based on peak queue depth over 50 samples under
3211 * sufficient load.
3213 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3215 struct cfq_queue *cfqq = cfqd->active_queue;
3217 if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3218 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3220 if (cfqd->hw_tag == 1)
3221 return;
3223 if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3224 cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3225 return;
3228 * If active queue hasn't enough requests and can idle, cfq might not
3229 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3230 * case
3232 if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3233 cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3234 CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3235 return;
3237 if (cfqd->hw_tag_samples++ < 50)
3238 return;
3240 if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3241 cfqd->hw_tag = 1;
3242 else
3243 cfqd->hw_tag = 0;
3246 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3248 struct cfq_io_cq *cic = cfqd->active_cic;
3250 /* If the queue already has requests, don't wait */
3251 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3252 return false;
3254 /* If there are other queues in the group, don't wait */
3255 if (cfqq->cfqg->nr_cfqq > 1)
3256 return false;
3258 /* the only queue in the group, but think time is big */
3259 if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3260 return false;
3262 if (cfq_slice_used(cfqq))
3263 return true;
3265 /* if slice left is less than think time, wait busy */
3266 if (cic && sample_valid(cic->ttime.ttime_samples)
3267 && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3268 return true;
3271 * If think times is less than a jiffy than ttime_mean=0 and above
3272 * will not be true. It might happen that slice has not expired yet
3273 * but will expire soon (4-5 ns) during select_queue(). To cover the
3274 * case where think time is less than a jiffy, mark the queue wait
3275 * busy if only 1 jiffy is left in the slice.
3277 if (cfqq->slice_end - jiffies == 1)
3278 return true;
3280 return false;
3283 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3285 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3286 struct cfq_data *cfqd = cfqq->cfqd;
3287 const int sync = rq_is_sync(rq);
3288 unsigned long now;
3290 now = jiffies;
3291 cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3292 !!(rq->cmd_flags & REQ_NOIDLE));
3294 cfq_update_hw_tag(cfqd);
3296 WARN_ON(!cfqd->rq_in_driver);
3297 WARN_ON(!cfqq->dispatched);
3298 cfqd->rq_in_driver--;
3299 cfqq->dispatched--;
3300 (RQ_CFQG(rq))->dispatched--;
3301 cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3302 rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3303 rq_data_dir(rq), rq_is_sync(rq));
3305 cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3307 if (sync) {
3308 struct cfq_rb_root *service_tree;
3310 RQ_CIC(rq)->ttime.last_end_request = now;
3312 if (cfq_cfqq_on_rr(cfqq))
3313 service_tree = cfqq->service_tree;
3314 else
3315 service_tree = service_tree_for(cfqq->cfqg,
3316 cfqq_prio(cfqq), cfqq_type(cfqq));
3317 service_tree->ttime.last_end_request = now;
3318 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3319 cfqd->last_delayed_sync = now;
3322 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3323 cfqq->cfqg->ttime.last_end_request = now;
3324 #endif
3327 * If this is the active queue, check if it needs to be expired,
3328 * or if we want to idle in case it has no pending requests.
3330 if (cfqd->active_queue == cfqq) {
3331 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3333 if (cfq_cfqq_slice_new(cfqq)) {
3334 cfq_set_prio_slice(cfqd, cfqq);
3335 cfq_clear_cfqq_slice_new(cfqq);
3339 * Should we wait for next request to come in before we expire
3340 * the queue.
3342 if (cfq_should_wait_busy(cfqd, cfqq)) {
3343 unsigned long extend_sl = cfqd->cfq_slice_idle;
3344 if (!cfqd->cfq_slice_idle)
3345 extend_sl = cfqd->cfq_group_idle;
3346 cfqq->slice_end = jiffies + extend_sl;
3347 cfq_mark_cfqq_wait_busy(cfqq);
3348 cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3352 * Idling is not enabled on:
3353 * - expired queues
3354 * - idle-priority queues
3355 * - async queues
3356 * - queues with still some requests queued
3357 * - when there is a close cooperator
3359 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3360 cfq_slice_expired(cfqd, 1);
3361 else if (sync && cfqq_empty &&
3362 !cfq_close_cooperator(cfqd, cfqq)) {
3363 cfq_arm_slice_timer(cfqd);
3367 if (!cfqd->rq_in_driver)
3368 cfq_schedule_dispatch(cfqd);
3371 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3373 if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3374 cfq_mark_cfqq_must_alloc_slice(cfqq);
3375 return ELV_MQUEUE_MUST;
3378 return ELV_MQUEUE_MAY;
3381 static int cfq_may_queue(struct request_queue *q, int rw)
3383 struct cfq_data *cfqd = q->elevator->elevator_data;
3384 struct task_struct *tsk = current;
3385 struct cfq_io_cq *cic;
3386 struct cfq_queue *cfqq;
3389 * don't force setup of a queue from here, as a call to may_queue
3390 * does not necessarily imply that a request actually will be queued.
3391 * so just lookup a possibly existing queue, or return 'may queue'
3392 * if that fails
3394 cic = cfq_cic_lookup(cfqd, tsk->io_context);
3395 if (!cic)
3396 return ELV_MQUEUE_MAY;
3398 cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3399 if (cfqq) {
3400 cfq_init_prio_data(cfqq, cic->icq.ioc);
3402 return __cfq_may_queue(cfqq);
3405 return ELV_MQUEUE_MAY;
3409 * queue lock held here
3411 static void cfq_put_request(struct request *rq)
3413 struct cfq_queue *cfqq = RQ_CFQQ(rq);
3415 if (cfqq) {
3416 const int rw = rq_data_dir(rq);
3418 BUG_ON(!cfqq->allocated[rw]);
3419 cfqq->allocated[rw]--;
3421 /* Put down rq reference on cfqg */
3422 cfq_put_cfqg(RQ_CFQG(rq));
3423 rq->elv.priv[0] = NULL;
3424 rq->elv.priv[1] = NULL;
3426 cfq_put_queue(cfqq);
3430 static struct cfq_queue *
3431 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3432 struct cfq_queue *cfqq)
3434 cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3435 cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3436 cfq_mark_cfqq_coop(cfqq->new_cfqq);
3437 cfq_put_queue(cfqq);
3438 return cic_to_cfqq(cic, 1);
3442 * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3443 * was the last process referring to said cfqq.
3445 static struct cfq_queue *
3446 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3448 if (cfqq_process_refs(cfqq) == 1) {
3449 cfqq->pid = current->pid;
3450 cfq_clear_cfqq_coop(cfqq);
3451 cfq_clear_cfqq_split_coop(cfqq);
3452 return cfqq;
3455 cic_set_cfqq(cic, NULL, 1);
3457 cfq_put_cooperator(cfqq);
3459 cfq_put_queue(cfqq);
3460 return NULL;
3463 * Allocate cfq data structures associated with this request.
3465 static int
3466 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3468 struct cfq_data *cfqd = q->elevator->elevator_data;
3469 struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3470 const int rw = rq_data_dir(rq);
3471 const bool is_sync = rq_is_sync(rq);
3472 struct cfq_queue *cfqq;
3473 unsigned int changed;
3475 might_sleep_if(gfp_mask & __GFP_WAIT);
3477 spin_lock_irq(q->queue_lock);
3479 /* handle changed notifications */
3480 changed = icq_get_changed(&cic->icq);
3481 if (unlikely(changed & ICQ_IOPRIO_CHANGED))
3482 changed_ioprio(cic);
3483 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3484 if (unlikely(changed & ICQ_CGROUP_CHANGED))
3485 changed_cgroup(cic);
3486 #endif
3488 new_queue:
3489 cfqq = cic_to_cfqq(cic, is_sync);
3490 if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3491 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3492 cic_set_cfqq(cic, cfqq, is_sync);
3493 } else {
3495 * If the queue was seeky for too long, break it apart.
3497 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3498 cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3499 cfqq = split_cfqq(cic, cfqq);
3500 if (!cfqq)
3501 goto new_queue;
3505 * Check to see if this queue is scheduled to merge with
3506 * another, closely cooperating queue. The merging of
3507 * queues happens here as it must be done in process context.
3508 * The reference on new_cfqq was taken in merge_cfqqs.
3510 if (cfqq->new_cfqq)
3511 cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3514 cfqq->allocated[rw]++;
3516 cfqq->ref++;
3517 rq->elv.priv[0] = cfqq;
3518 rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
3519 spin_unlock_irq(q->queue_lock);
3520 return 0;
3523 static void cfq_kick_queue(struct work_struct *work)
3525 struct cfq_data *cfqd =
3526 container_of(work, struct cfq_data, unplug_work);
3527 struct request_queue *q = cfqd->queue;
3529 spin_lock_irq(q->queue_lock);
3530 __blk_run_queue(cfqd->queue);
3531 spin_unlock_irq(q->queue_lock);
3535 * Timer running if the active_queue is currently idling inside its time slice
3537 static void cfq_idle_slice_timer(unsigned long data)
3539 struct cfq_data *cfqd = (struct cfq_data *) data;
3540 struct cfq_queue *cfqq;
3541 unsigned long flags;
3542 int timed_out = 1;
3544 cfq_log(cfqd, "idle timer fired");
3546 spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3548 cfqq = cfqd->active_queue;
3549 if (cfqq) {
3550 timed_out = 0;
3553 * We saw a request before the queue expired, let it through
3555 if (cfq_cfqq_must_dispatch(cfqq))
3556 goto out_kick;
3559 * expired
3561 if (cfq_slice_used(cfqq))
3562 goto expire;
3565 * only expire and reinvoke request handler, if there are
3566 * other queues with pending requests
3568 if (!cfqd->busy_queues)
3569 goto out_cont;
3572 * not expired and it has a request pending, let it dispatch
3574 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3575 goto out_kick;
3578 * Queue depth flag is reset only when the idle didn't succeed
3580 cfq_clear_cfqq_deep(cfqq);
3582 expire:
3583 cfq_slice_expired(cfqd, timed_out);
3584 out_kick:
3585 cfq_schedule_dispatch(cfqd);
3586 out_cont:
3587 spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3590 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3592 del_timer_sync(&cfqd->idle_slice_timer);
3593 cancel_work_sync(&cfqd->unplug_work);
3596 static void cfq_put_async_queues(struct cfq_data *cfqd)
3598 int i;
3600 for (i = 0; i < IOPRIO_BE_NR; i++) {
3601 if (cfqd->async_cfqq[0][i])
3602 cfq_put_queue(cfqd->async_cfqq[0][i]);
3603 if (cfqd->async_cfqq[1][i])
3604 cfq_put_queue(cfqd->async_cfqq[1][i]);
3607 if (cfqd->async_idle_cfqq)
3608 cfq_put_queue(cfqd->async_idle_cfqq);
3611 static void cfq_exit_queue(struct elevator_queue *e)
3613 struct cfq_data *cfqd = e->elevator_data;
3614 struct request_queue *q = cfqd->queue;
3615 bool wait = false;
3617 cfq_shutdown_timer_wq(cfqd);
3619 spin_lock_irq(q->queue_lock);
3621 if (cfqd->active_queue)
3622 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3624 cfq_put_async_queues(cfqd);
3625 cfq_release_cfq_groups(cfqd);
3628 * If there are groups which we could not unlink from blkcg list,
3629 * wait for a rcu period for them to be freed.
3631 if (cfqd->nr_blkcg_linked_grps)
3632 wait = true;
3634 spin_unlock_irq(q->queue_lock);
3636 cfq_shutdown_timer_wq(cfqd);
3639 * Wait for cfqg->blkg->key accessors to exit their grace periods.
3640 * Do this wait only if there are other unlinked groups out
3641 * there. This can happen if cgroup deletion path claimed the
3642 * responsibility of cleaning up a group before queue cleanup code
3643 * get to the group.
3645 * Do not call synchronize_rcu() unconditionally as there are drivers
3646 * which create/delete request queue hundreds of times during scan/boot
3647 * and synchronize_rcu() can take significant time and slow down boot.
3649 if (wait)
3650 synchronize_rcu();
3652 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3653 /* Free up per cpu stats for root group */
3654 free_percpu(cfqd->root_group.blkg.stats_cpu);
3655 #endif
3656 kfree(cfqd);
3659 static void *cfq_init_queue(struct request_queue *q)
3661 struct cfq_data *cfqd;
3662 int i, j;
3663 struct cfq_group *cfqg;
3664 struct cfq_rb_root *st;
3666 cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3667 if (!cfqd)
3668 return NULL;
3670 /* Init root service tree */
3671 cfqd->grp_service_tree = CFQ_RB_ROOT;
3673 /* Init root group */
3674 cfqg = &cfqd->root_group;
3675 for_each_cfqg_st(cfqg, i, j, st)
3676 *st = CFQ_RB_ROOT;
3677 RB_CLEAR_NODE(&cfqg->rb_node);
3679 /* Give preference to root group over other groups */
3680 cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3682 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3684 * Set root group reference to 2. One reference will be dropped when
3685 * all groups on cfqd->cfqg_list are being deleted during queue exit.
3686 * Other reference will remain there as we don't want to delete this
3687 * group as it is statically allocated and gets destroyed when
3688 * throtl_data goes away.
3690 cfqg->ref = 2;
3692 if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3693 kfree(cfqg);
3694 kfree(cfqd);
3695 return NULL;
3698 rcu_read_lock();
3700 cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3701 (void *)cfqd, 0);
3702 rcu_read_unlock();
3703 cfqd->nr_blkcg_linked_grps++;
3705 /* Add group on cfqd->cfqg_list */
3706 hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3707 #endif
3709 * Not strictly needed (since RB_ROOT just clears the node and we
3710 * zeroed cfqd on alloc), but better be safe in case someone decides
3711 * to add magic to the rb code
3713 for (i = 0; i < CFQ_PRIO_LISTS; i++)
3714 cfqd->prio_trees[i] = RB_ROOT;
3717 * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3718 * Grab a permanent reference to it, so that the normal code flow
3719 * will not attempt to free it.
3721 cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3722 cfqd->oom_cfqq.ref++;
3723 cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3725 cfqd->queue = q;
3727 init_timer(&cfqd->idle_slice_timer);
3728 cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3729 cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3731 INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3733 cfqd->cfq_quantum = cfq_quantum;
3734 cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3735 cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3736 cfqd->cfq_back_max = cfq_back_max;
3737 cfqd->cfq_back_penalty = cfq_back_penalty;
3738 cfqd->cfq_slice[0] = cfq_slice_async;
3739 cfqd->cfq_slice[1] = cfq_slice_sync;
3740 cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3741 cfqd->cfq_slice_idle = cfq_slice_idle;
3742 cfqd->cfq_group_idle = cfq_group_idle;
3743 cfqd->cfq_latency = 1;
3744 cfqd->hw_tag = -1;
3746 * we optimistically start assuming sync ops weren't delayed in last
3747 * second, in order to have larger depth for async operations.
3749 cfqd->last_delayed_sync = jiffies - HZ;
3750 return cfqd;
3754 * sysfs parts below -->
3756 static ssize_t
3757 cfq_var_show(unsigned int var, char *page)
3759 return sprintf(page, "%d\n", var);
3762 static ssize_t
3763 cfq_var_store(unsigned int *var, const char *page, size_t count)
3765 char *p = (char *) page;
3767 *var = simple_strtoul(p, &p, 10);
3768 return count;
3771 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV) \
3772 static ssize_t __FUNC(struct elevator_queue *e, char *page) \
3774 struct cfq_data *cfqd = e->elevator_data; \
3775 unsigned int __data = __VAR; \
3776 if (__CONV) \
3777 __data = jiffies_to_msecs(__data); \
3778 return cfq_var_show(__data, (page)); \
3780 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3781 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3782 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3783 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3784 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3785 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3786 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3787 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3788 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3789 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3790 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3791 #undef SHOW_FUNCTION
3793 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV) \
3794 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3796 struct cfq_data *cfqd = e->elevator_data; \
3797 unsigned int __data; \
3798 int ret = cfq_var_store(&__data, (page), count); \
3799 if (__data < (MIN)) \
3800 __data = (MIN); \
3801 else if (__data > (MAX)) \
3802 __data = (MAX); \
3803 if (__CONV) \
3804 *(__PTR) = msecs_to_jiffies(__data); \
3805 else \
3806 *(__PTR) = __data; \
3807 return ret; \
3809 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3810 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3811 UINT_MAX, 1);
3812 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3813 UINT_MAX, 1);
3814 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3815 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3816 UINT_MAX, 0);
3817 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3818 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
3819 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3820 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3821 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3822 UINT_MAX, 0);
3823 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3824 #undef STORE_FUNCTION
3826 #define CFQ_ATTR(name) \
3827 __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3829 static struct elv_fs_entry cfq_attrs[] = {
3830 CFQ_ATTR(quantum),
3831 CFQ_ATTR(fifo_expire_sync),
3832 CFQ_ATTR(fifo_expire_async),
3833 CFQ_ATTR(back_seek_max),
3834 CFQ_ATTR(back_seek_penalty),
3835 CFQ_ATTR(slice_sync),
3836 CFQ_ATTR(slice_async),
3837 CFQ_ATTR(slice_async_rq),
3838 CFQ_ATTR(slice_idle),
3839 CFQ_ATTR(group_idle),
3840 CFQ_ATTR(low_latency),
3841 __ATTR_NULL
3844 static struct elevator_type iosched_cfq = {
3845 .ops = {
3846 .elevator_merge_fn = cfq_merge,
3847 .elevator_merged_fn = cfq_merged_request,
3848 .elevator_merge_req_fn = cfq_merged_requests,
3849 .elevator_allow_merge_fn = cfq_allow_merge,
3850 .elevator_bio_merged_fn = cfq_bio_merged,
3851 .elevator_dispatch_fn = cfq_dispatch_requests,
3852 .elevator_add_req_fn = cfq_insert_request,
3853 .elevator_activate_req_fn = cfq_activate_request,
3854 .elevator_deactivate_req_fn = cfq_deactivate_request,
3855 .elevator_completed_req_fn = cfq_completed_request,
3856 .elevator_former_req_fn = elv_rb_former_request,
3857 .elevator_latter_req_fn = elv_rb_latter_request,
3858 .elevator_init_icq_fn = cfq_init_icq,
3859 .elevator_exit_icq_fn = cfq_exit_icq,
3860 .elevator_set_req_fn = cfq_set_request,
3861 .elevator_put_req_fn = cfq_put_request,
3862 .elevator_may_queue_fn = cfq_may_queue,
3863 .elevator_init_fn = cfq_init_queue,
3864 .elevator_exit_fn = cfq_exit_queue,
3866 .icq_size = sizeof(struct cfq_io_cq),
3867 .icq_align = __alignof__(struct cfq_io_cq),
3868 .elevator_attrs = cfq_attrs,
3869 .elevator_name = "cfq",
3870 .elevator_owner = THIS_MODULE,
3873 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3874 static struct blkio_policy_type blkio_policy_cfq = {
3875 .ops = {
3876 .blkio_unlink_group_fn = cfq_unlink_blkio_group,
3877 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3879 .plid = BLKIO_POLICY_PROP,
3881 #else
3882 static struct blkio_policy_type blkio_policy_cfq;
3883 #endif
3885 static int __init cfq_init(void)
3887 int ret;
3890 * could be 0 on HZ < 1000 setups
3892 if (!cfq_slice_async)
3893 cfq_slice_async = 1;
3894 if (!cfq_slice_idle)
3895 cfq_slice_idle = 1;
3897 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3898 if (!cfq_group_idle)
3899 cfq_group_idle = 1;
3900 #else
3901 cfq_group_idle = 0;
3902 #endif
3903 cfq_pool = KMEM_CACHE(cfq_queue, 0);
3904 if (!cfq_pool)
3905 return -ENOMEM;
3907 ret = elv_register(&iosched_cfq);
3908 if (ret) {
3909 kmem_cache_destroy(cfq_pool);
3910 return ret;
3913 blkio_policy_register(&blkio_policy_cfq);
3915 return 0;
3918 static void __exit cfq_exit(void)
3920 blkio_policy_unregister(&blkio_policy_cfq);
3921 elv_unregister(&iosched_cfq);
3922 kmem_cache_destroy(cfq_pool);
3925 module_init(cfq_init);
3926 module_exit(cfq_exit);
3928 MODULE_AUTHOR("Jens Axboe");
3929 MODULE_LICENSE("GPL");
3930 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");