blkcg: add blkg_policy_data->plid
[linux-3.10.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
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 "blk-cgroup.h"
19
20 /*
21  * tunables
22  */
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;
37
38 /*
39  * offset from end of service tree
40  */
41 #define CFQ_IDLE_DELAY          (HZ / 5)
42
43 /*
44  * below this threshold, we consider thinktime immediate
45  */
46 #define CFQ_MIN_TT              (2)
47
48 #define CFQ_SLICE_SCALE         (5)
49 #define CFQ_HW_QUEUE_MIN        (5)
50 #define CFQ_SERVICE_SHIFT       12
51
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)
56
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])
60
61 static struct kmem_cache *cfq_pool;
62
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)
66
67 #define sample_valid(samples)   ((samples) > 80)
68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
69
70 struct cfq_ttime {
71         unsigned long last_end_request;
72
73         unsigned long ttime_total;
74         unsigned long ttime_samples;
75         unsigned long ttime_mean;
76 };
77
78 /*
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.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         u64 min_vdisktime;
89         struct cfq_ttime ttime;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
92                         .ttime = {.last_end_request = jiffies,},}
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         int ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending priority requests */
133         int prio_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         /* Number of sectors dispatched from queue in single dispatch round */
150         unsigned long nr_sectors;
151 };
152
153 /*
154  * First index in the service_trees.
155  * IDLE is handled separately, so it has negative index
156  */
157 enum wl_class_t {
158         BE_WORKLOAD = 0,
159         RT_WORKLOAD = 1,
160         IDLE_WORKLOAD = 2,
161         CFQ_PRIO_NR,
162 };
163
164 /*
165  * Second index in the service_trees.
166  */
167 enum wl_type_t {
168         ASYNC_WORKLOAD = 0,
169         SYNC_NOIDLE_WORKLOAD = 1,
170         SYNC_WORKLOAD = 2
171 };
172
173 struct cfqg_stats {
174 #ifdef CONFIG_CFQ_GROUP_IOSCHED
175         /* total bytes transferred */
176         struct blkg_rwstat              service_bytes;
177         /* total IOs serviced, post merge */
178         struct blkg_rwstat              serviced;
179         /* number of ios merged */
180         struct blkg_rwstat              merged;
181         /* total time spent on device in ns, may not be accurate w/ queueing */
182         struct blkg_rwstat              service_time;
183         /* total time spent waiting in scheduler queue in ns */
184         struct blkg_rwstat              wait_time;
185         /* number of IOs queued up */
186         struct blkg_rwstat              queued;
187         /* total sectors transferred */
188         struct blkg_stat                sectors;
189         /* total disk time and nr sectors dispatched by this group */
190         struct blkg_stat                time;
191 #ifdef CONFIG_DEBUG_BLK_CGROUP
192         /* time not charged to this cgroup */
193         struct blkg_stat                unaccounted_time;
194         /* sum of number of ios queued across all samples */
195         struct blkg_stat                avg_queue_size_sum;
196         /* count of samples taken for average */
197         struct blkg_stat                avg_queue_size_samples;
198         /* how many times this group has been removed from service tree */
199         struct blkg_stat                dequeue;
200         /* total time spent waiting for it to be assigned a timeslice. */
201         struct blkg_stat                group_wait_time;
202         /* time spent idling for this blkcg_gq */
203         struct blkg_stat                idle_time;
204         /* total time with empty current active q with other requests queued */
205         struct blkg_stat                empty_time;
206         /* fields after this shouldn't be cleared on stat reset */
207         uint64_t                        start_group_wait_time;
208         uint64_t                        start_idle_time;
209         uint64_t                        start_empty_time;
210         uint16_t                        flags;
211 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
212 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
213 };
214
215 /* This is per cgroup per device grouping structure */
216 struct cfq_group {
217         /* must be the first member */
218         struct blkg_policy_data pd;
219
220         /* group service_tree member */
221         struct rb_node rb_node;
222
223         /* group service_tree key */
224         u64 vdisktime;
225
226         /*
227          * The number of active cfqgs and sum of their weights under this
228          * cfqg.  This covers this cfqg's leaf_weight and all children's
229          * weights, but does not cover weights of further descendants.
230          *
231          * If a cfqg is on the service tree, it's active.  An active cfqg
232          * also activates its parent and contributes to the children_weight
233          * of the parent.
234          */
235         int nr_active;
236         unsigned int children_weight;
237
238         /*
239          * vfraction is the fraction of vdisktime that the tasks in this
240          * cfqg are entitled to.  This is determined by compounding the
241          * ratios walking up from this cfqg to the root.
242          *
243          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
244          * vfractions on a service tree is approximately 1.  The sum may
245          * deviate a bit due to rounding errors and fluctuations caused by
246          * cfqgs entering and leaving the service tree.
247          */
248         unsigned int vfraction;
249
250         /*
251          * There are two weights - (internal) weight is the weight of this
252          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
253          * this cfqg against the child cfqgs.  For the root cfqg, both
254          * weights are kept in sync for backward compatibility.
255          */
256         unsigned int weight;
257         unsigned int new_weight;
258         unsigned int dev_weight;
259
260         unsigned int leaf_weight;
261         unsigned int new_leaf_weight;
262         unsigned int dev_leaf_weight;
263
264         /* number of cfqq currently on this group */
265         int nr_cfqq;
266
267         /*
268          * Per group busy queues average. Useful for workload slice calc. We
269          * create the array for each prio class but at run time it is used
270          * only for RT and BE class and slot for IDLE class remains unused.
271          * This is primarily done to avoid confusion and a gcc warning.
272          */
273         unsigned int busy_queues_avg[CFQ_PRIO_NR];
274         /*
275          * rr lists of queues with requests. We maintain service trees for
276          * RT and BE classes. These trees are subdivided in subclasses
277          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
278          * class there is no subclassification and all the cfq queues go on
279          * a single tree service_tree_idle.
280          * Counts are embedded in the cfq_rb_root
281          */
282         struct cfq_rb_root service_trees[2][3];
283         struct cfq_rb_root service_tree_idle;
284
285         unsigned long saved_wl_slice;
286         enum wl_type_t saved_wl_type;
287         enum wl_class_t saved_wl_class;
288
289         /* number of requests that are on the dispatch list or inside driver */
290         int dispatched;
291         struct cfq_ttime ttime;
292         struct cfqg_stats stats;
293 };
294
295 struct cfq_io_cq {
296         struct io_cq            icq;            /* must be the first member */
297         struct cfq_queue        *cfqq[2];
298         struct cfq_ttime        ttime;
299         int                     ioprio;         /* the current ioprio */
300 #ifdef CONFIG_CFQ_GROUP_IOSCHED
301         uint64_t                blkcg_id;       /* the current blkcg ID */
302 #endif
303 };
304
305 /*
306  * Per block device queue structure
307  */
308 struct cfq_data {
309         struct request_queue *queue;
310         /* Root service tree for cfq_groups */
311         struct cfq_rb_root grp_service_tree;
312         struct cfq_group *root_group;
313
314         /*
315          * The priority currently being served
316          */
317         enum wl_class_t serving_wl_class;
318         enum wl_type_t serving_wl_type;
319         unsigned long workload_expires;
320         struct cfq_group *serving_group;
321
322         /*
323          * Each priority tree is sorted by next_request position.  These
324          * trees are used when determining if two or more queues are
325          * interleaving requests (see cfq_close_cooperator).
326          */
327         struct rb_root prio_trees[CFQ_PRIO_LISTS];
328
329         unsigned int busy_queues;
330         unsigned int busy_sync_queues;
331
332         int rq_in_driver;
333         int rq_in_flight[2];
334
335         /*
336          * queue-depth detection
337          */
338         int rq_queued;
339         int hw_tag;
340         /*
341          * hw_tag can be
342          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
343          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
344          *  0 => no NCQ
345          */
346         int hw_tag_est_depth;
347         unsigned int hw_tag_samples;
348
349         /*
350          * idle window management
351          */
352         struct timer_list idle_slice_timer;
353         struct work_struct unplug_work;
354
355         struct cfq_queue *active_queue;
356         struct cfq_io_cq *active_cic;
357
358         /*
359          * async queue for each priority case
360          */
361         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
362         struct cfq_queue *async_idle_cfqq;
363
364         sector_t last_position;
365
366         /*
367          * tunables, see top of file
368          */
369         unsigned int cfq_quantum;
370         unsigned int cfq_fifo_expire[2];
371         unsigned int cfq_back_penalty;
372         unsigned int cfq_back_max;
373         unsigned int cfq_slice[2];
374         unsigned int cfq_slice_async_rq;
375         unsigned int cfq_slice_idle;
376         unsigned int cfq_group_idle;
377         unsigned int cfq_latency;
378         unsigned int cfq_target_latency;
379
380         /*
381          * Fallback dummy cfqq for extreme OOM conditions
382          */
383         struct cfq_queue oom_cfqq;
384
385         unsigned long last_delayed_sync;
386 };
387
388 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
389
390 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
391                                             enum wl_class_t class,
392                                             enum wl_type_t type)
393 {
394         if (!cfqg)
395                 return NULL;
396
397         if (class == IDLE_WORKLOAD)
398                 return &cfqg->service_tree_idle;
399
400         return &cfqg->service_trees[class][type];
401 }
402
403 enum cfqq_state_flags {
404         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
405         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
406         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
407         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
408         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
409         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
410         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
411         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
412         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
413         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
414         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
415         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
416         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
417 };
418
419 #define CFQ_CFQQ_FNS(name)                                              \
420 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
421 {                                                                       \
422         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
423 }                                                                       \
424 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
425 {                                                                       \
426         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
427 }                                                                       \
428 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
429 {                                                                       \
430         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
431 }
432
433 CFQ_CFQQ_FNS(on_rr);
434 CFQ_CFQQ_FNS(wait_request);
435 CFQ_CFQQ_FNS(must_dispatch);
436 CFQ_CFQQ_FNS(must_alloc_slice);
437 CFQ_CFQQ_FNS(fifo_expire);
438 CFQ_CFQQ_FNS(idle_window);
439 CFQ_CFQQ_FNS(prio_changed);
440 CFQ_CFQQ_FNS(slice_new);
441 CFQ_CFQQ_FNS(sync);
442 CFQ_CFQQ_FNS(coop);
443 CFQ_CFQQ_FNS(split_coop);
444 CFQ_CFQQ_FNS(deep);
445 CFQ_CFQQ_FNS(wait_busy);
446 #undef CFQ_CFQQ_FNS
447
448 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
449 {
450         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
451 }
452
453 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
454 {
455         return pd_to_blkg(&cfqg->pd);
456 }
457
458 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
459
460 /* cfqg stats flags */
461 enum cfqg_stats_flags {
462         CFQG_stats_waiting = 0,
463         CFQG_stats_idling,
464         CFQG_stats_empty,
465 };
466
467 #define CFQG_FLAG_FNS(name)                                             \
468 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
469 {                                                                       \
470         stats->flags |= (1 << CFQG_stats_##name);                       \
471 }                                                                       \
472 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
473 {                                                                       \
474         stats->flags &= ~(1 << CFQG_stats_##name);                      \
475 }                                                                       \
476 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
477 {                                                                       \
478         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
479 }                                                                       \
480
481 CFQG_FLAG_FNS(waiting)
482 CFQG_FLAG_FNS(idling)
483 CFQG_FLAG_FNS(empty)
484 #undef CFQG_FLAG_FNS
485
486 /* This should be called with the queue_lock held. */
487 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
488 {
489         unsigned long long now;
490
491         if (!cfqg_stats_waiting(stats))
492                 return;
493
494         now = sched_clock();
495         if (time_after64(now, stats->start_group_wait_time))
496                 blkg_stat_add(&stats->group_wait_time,
497                               now - stats->start_group_wait_time);
498         cfqg_stats_clear_waiting(stats);
499 }
500
501 /* This should be called with the queue_lock held. */
502 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
503                                                  struct cfq_group *curr_cfqg)
504 {
505         struct cfqg_stats *stats = &cfqg->stats;
506
507         if (cfqg_stats_waiting(stats))
508                 return;
509         if (cfqg == curr_cfqg)
510                 return;
511         stats->start_group_wait_time = sched_clock();
512         cfqg_stats_mark_waiting(stats);
513 }
514
515 /* This should be called with the queue_lock held. */
516 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
517 {
518         unsigned long long now;
519
520         if (!cfqg_stats_empty(stats))
521                 return;
522
523         now = sched_clock();
524         if (time_after64(now, stats->start_empty_time))
525                 blkg_stat_add(&stats->empty_time,
526                               now - stats->start_empty_time);
527         cfqg_stats_clear_empty(stats);
528 }
529
530 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
531 {
532         blkg_stat_add(&cfqg->stats.dequeue, 1);
533 }
534
535 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
536 {
537         struct cfqg_stats *stats = &cfqg->stats;
538
539         if (blkg_rwstat_sum(&stats->queued))
540                 return;
541
542         /*
543          * group is already marked empty. This can happen if cfqq got new
544          * request in parent group and moved to this group while being added
545          * to service tree. Just ignore the event and move on.
546          */
547         if (cfqg_stats_empty(stats))
548                 return;
549
550         stats->start_empty_time = sched_clock();
551         cfqg_stats_mark_empty(stats);
552 }
553
554 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
555 {
556         struct cfqg_stats *stats = &cfqg->stats;
557
558         if (cfqg_stats_idling(stats)) {
559                 unsigned long long now = sched_clock();
560
561                 if (time_after64(now, stats->start_idle_time))
562                         blkg_stat_add(&stats->idle_time,
563                                       now - stats->start_idle_time);
564                 cfqg_stats_clear_idling(stats);
565         }
566 }
567
568 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
569 {
570         struct cfqg_stats *stats = &cfqg->stats;
571
572         BUG_ON(cfqg_stats_idling(stats));
573
574         stats->start_idle_time = sched_clock();
575         cfqg_stats_mark_idling(stats);
576 }
577
578 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
579 {
580         struct cfqg_stats *stats = &cfqg->stats;
581
582         blkg_stat_add(&stats->avg_queue_size_sum,
583                       blkg_rwstat_sum(&stats->queued));
584         blkg_stat_add(&stats->avg_queue_size_samples, 1);
585         cfqg_stats_update_group_wait_time(stats);
586 }
587
588 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
589
590 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
591 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
592 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
593 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
594 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
595 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
596 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
597
598 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
599
600 #ifdef CONFIG_CFQ_GROUP_IOSCHED
601
602 static struct blkcg_policy blkcg_policy_cfq;
603
604 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
605 {
606         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
607 }
608
609 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
610 {
611         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
612
613         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
614 }
615
616 static inline void cfqg_get(struct cfq_group *cfqg)
617 {
618         return blkg_get(cfqg_to_blkg(cfqg));
619 }
620
621 static inline void cfqg_put(struct cfq_group *cfqg)
622 {
623         return blkg_put(cfqg_to_blkg(cfqg));
624 }
625
626 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
627         char __pbuf[128];                                               \
628                                                                         \
629         blkg_path(cfqg_to_blkg((cfqq)->cfqg), __pbuf, sizeof(__pbuf));  \
630         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c %s " fmt, (cfqq)->pid, \
631                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
632                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
633                           __pbuf, ##args);                              \
634 } while (0)
635
636 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
637         char __pbuf[128];                                               \
638                                                                         \
639         blkg_path(cfqg_to_blkg(cfqg), __pbuf, sizeof(__pbuf));          \
640         blk_add_trace_msg((cfqd)->queue, "%s " fmt, __pbuf, ##args);    \
641 } while (0)
642
643 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
644                                             struct cfq_group *curr_cfqg, int rw)
645 {
646         blkg_rwstat_add(&cfqg->stats.queued, rw, 1);
647         cfqg_stats_end_empty_time(&cfqg->stats);
648         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
649 }
650
651 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
652                         unsigned long time, unsigned long unaccounted_time)
653 {
654         blkg_stat_add(&cfqg->stats.time, time);
655 #ifdef CONFIG_DEBUG_BLK_CGROUP
656         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
657 #endif
658 }
659
660 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw)
661 {
662         blkg_rwstat_add(&cfqg->stats.queued, rw, -1);
663 }
664
665 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw)
666 {
667         blkg_rwstat_add(&cfqg->stats.merged, rw, 1);
668 }
669
670 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
671                                               uint64_t bytes, int rw)
672 {
673         blkg_stat_add(&cfqg->stats.sectors, bytes >> 9);
674         blkg_rwstat_add(&cfqg->stats.serviced, rw, 1);
675         blkg_rwstat_add(&cfqg->stats.service_bytes, rw, bytes);
676 }
677
678 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
679                         uint64_t start_time, uint64_t io_start_time, int rw)
680 {
681         struct cfqg_stats *stats = &cfqg->stats;
682         unsigned long long now = sched_clock();
683
684         if (time_after64(now, io_start_time))
685                 blkg_rwstat_add(&stats->service_time, rw, now - io_start_time);
686         if (time_after64(io_start_time, start_time))
687                 blkg_rwstat_add(&stats->wait_time, rw,
688                                 io_start_time - start_time);
689 }
690
691 static void cfq_pd_reset_stats(struct blkcg_gq *blkg)
692 {
693         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
694         struct cfqg_stats *stats = &cfqg->stats;
695
696         /* queued stats shouldn't be cleared */
697         blkg_rwstat_reset(&stats->service_bytes);
698         blkg_rwstat_reset(&stats->serviced);
699         blkg_rwstat_reset(&stats->merged);
700         blkg_rwstat_reset(&stats->service_time);
701         blkg_rwstat_reset(&stats->wait_time);
702         blkg_stat_reset(&stats->time);
703 #ifdef CONFIG_DEBUG_BLK_CGROUP
704         blkg_stat_reset(&stats->unaccounted_time);
705         blkg_stat_reset(&stats->avg_queue_size_sum);
706         blkg_stat_reset(&stats->avg_queue_size_samples);
707         blkg_stat_reset(&stats->dequeue);
708         blkg_stat_reset(&stats->group_wait_time);
709         blkg_stat_reset(&stats->idle_time);
710         blkg_stat_reset(&stats->empty_time);
711 #endif
712 }
713
714 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
715
716 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
717 static inline void cfqg_get(struct cfq_group *cfqg) { }
718 static inline void cfqg_put(struct cfq_group *cfqg) { }
719
720 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
721         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
722                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
723                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
724                                 ##args)
725 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
726
727 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
728                         struct cfq_group *curr_cfqg, int rw) { }
729 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
730                         unsigned long time, unsigned long unaccounted_time) { }
731 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg, int rw) { }
732 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg, int rw) { }
733 static inline void cfqg_stats_update_dispatch(struct cfq_group *cfqg,
734                                               uint64_t bytes, int rw) { }
735 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
736                         uint64_t start_time, uint64_t io_start_time, int rw) { }
737
738 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
739
740 #define cfq_log(cfqd, fmt, args...)     \
741         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
742
743 /* Traverses through cfq group service trees */
744 #define for_each_cfqg_st(cfqg, i, j, st) \
745         for (i = 0; i <= IDLE_WORKLOAD; i++) \
746                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
747                         : &cfqg->service_tree_idle; \
748                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
749                         (i == IDLE_WORKLOAD && j == 0); \
750                         j++, st = i < IDLE_WORKLOAD ? \
751                         &cfqg->service_trees[i][j]: NULL) \
752
753 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
754         struct cfq_ttime *ttime, bool group_idle)
755 {
756         unsigned long slice;
757         if (!sample_valid(ttime->ttime_samples))
758                 return false;
759         if (group_idle)
760                 slice = cfqd->cfq_group_idle;
761         else
762                 slice = cfqd->cfq_slice_idle;
763         return ttime->ttime_mean > slice;
764 }
765
766 static inline bool iops_mode(struct cfq_data *cfqd)
767 {
768         /*
769          * If we are not idling on queues and it is a NCQ drive, parallel
770          * execution of requests is on and measuring time is not possible
771          * in most of the cases until and unless we drive shallower queue
772          * depths and that becomes a performance bottleneck. In such cases
773          * switch to start providing fairness in terms of number of IOs.
774          */
775         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
776                 return true;
777         else
778                 return false;
779 }
780
781 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
782 {
783         if (cfq_class_idle(cfqq))
784                 return IDLE_WORKLOAD;
785         if (cfq_class_rt(cfqq))
786                 return RT_WORKLOAD;
787         return BE_WORKLOAD;
788 }
789
790
791 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
792 {
793         if (!cfq_cfqq_sync(cfqq))
794                 return ASYNC_WORKLOAD;
795         if (!cfq_cfqq_idle_window(cfqq))
796                 return SYNC_NOIDLE_WORKLOAD;
797         return SYNC_WORKLOAD;
798 }
799
800 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
801                                         struct cfq_data *cfqd,
802                                         struct cfq_group *cfqg)
803 {
804         if (wl_class == IDLE_WORKLOAD)
805                 return cfqg->service_tree_idle.count;
806
807         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
808                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
809                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
810 }
811
812 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
813                                         struct cfq_group *cfqg)
814 {
815         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
816                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
817 }
818
819 static void cfq_dispatch_insert(struct request_queue *, struct request *);
820 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
821                                        struct cfq_io_cq *cic, struct bio *bio,
822                                        gfp_t gfp_mask);
823
824 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
825 {
826         /* cic->icq is the first member, %NULL will convert to %NULL */
827         return container_of(icq, struct cfq_io_cq, icq);
828 }
829
830 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
831                                                struct io_context *ioc)
832 {
833         if (ioc)
834                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
835         return NULL;
836 }
837
838 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
839 {
840         return cic->cfqq[is_sync];
841 }
842
843 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
844                                 bool is_sync)
845 {
846         cic->cfqq[is_sync] = cfqq;
847 }
848
849 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
850 {
851         return cic->icq.q->elevator->elevator_data;
852 }
853
854 /*
855  * We regard a request as SYNC, if it's either a read or has the SYNC bit
856  * set (in which case it could also be direct WRITE).
857  */
858 static inline bool cfq_bio_sync(struct bio *bio)
859 {
860         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
861 }
862
863 /*
864  * scheduler run of queue, if there are requests pending and no one in the
865  * driver that will restart queueing
866  */
867 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
868 {
869         if (cfqd->busy_queues) {
870                 cfq_log(cfqd, "schedule dispatch");
871                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
872         }
873 }
874
875 /*
876  * Scale schedule slice based on io priority. Use the sync time slice only
877  * if a queue is marked sync and has sync io queued. A sync queue with async
878  * io only, should not get full sync slice length.
879  */
880 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
881                                  unsigned short prio)
882 {
883         const int base_slice = cfqd->cfq_slice[sync];
884
885         WARN_ON(prio >= IOPRIO_BE_NR);
886
887         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
888 }
889
890 static inline int
891 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
892 {
893         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
894 }
895
896 /**
897  * cfqg_scale_charge - scale disk time charge according to cfqg weight
898  * @charge: disk time being charged
899  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
900  *
901  * Scale @charge according to @vfraction, which is in range (0, 1].  The
902  * scaling is inversely proportional.
903  *
904  * scaled = charge / vfraction
905  *
906  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
907  */
908 static inline u64 cfqg_scale_charge(unsigned long charge,
909                                     unsigned int vfraction)
910 {
911         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
912
913         /* charge / vfraction */
914         c <<= CFQ_SERVICE_SHIFT;
915         do_div(c, vfraction);
916         return c;
917 }
918
919 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
920 {
921         s64 delta = (s64)(vdisktime - min_vdisktime);
922         if (delta > 0)
923                 min_vdisktime = vdisktime;
924
925         return min_vdisktime;
926 }
927
928 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
929 {
930         s64 delta = (s64)(vdisktime - min_vdisktime);
931         if (delta < 0)
932                 min_vdisktime = vdisktime;
933
934         return min_vdisktime;
935 }
936
937 static void update_min_vdisktime(struct cfq_rb_root *st)
938 {
939         struct cfq_group *cfqg;
940
941         if (st->left) {
942                 cfqg = rb_entry_cfqg(st->left);
943                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
944                                                   cfqg->vdisktime);
945         }
946 }
947
948 /*
949  * get averaged number of queues of RT/BE priority.
950  * average is updated, with a formula that gives more weight to higher numbers,
951  * to quickly follows sudden increases and decrease slowly
952  */
953
954 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
955                                         struct cfq_group *cfqg, bool rt)
956 {
957         unsigned min_q, max_q;
958         unsigned mult  = cfq_hist_divisor - 1;
959         unsigned round = cfq_hist_divisor / 2;
960         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
961
962         min_q = min(cfqg->busy_queues_avg[rt], busy);
963         max_q = max(cfqg->busy_queues_avg[rt], busy);
964         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
965                 cfq_hist_divisor;
966         return cfqg->busy_queues_avg[rt];
967 }
968
969 static inline unsigned
970 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
971 {
972         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
973 }
974
975 static inline unsigned
976 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
977 {
978         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
979         if (cfqd->cfq_latency) {
980                 /*
981                  * interested queues (we consider only the ones with the same
982                  * priority class in the cfq group)
983                  */
984                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
985                                                 cfq_class_rt(cfqq));
986                 unsigned sync_slice = cfqd->cfq_slice[1];
987                 unsigned expect_latency = sync_slice * iq;
988                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
989
990                 if (expect_latency > group_slice) {
991                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
992                         /* scale low_slice according to IO priority
993                          * and sync vs async */
994                         unsigned low_slice =
995                                 min(slice, base_low_slice * slice / sync_slice);
996                         /* the adapted slice value is scaled to fit all iqs
997                          * into the target latency */
998                         slice = max(slice * group_slice / expect_latency,
999                                     low_slice);
1000                 }
1001         }
1002         return slice;
1003 }
1004
1005 static inline void
1006 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1007 {
1008         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1009
1010         cfqq->slice_start = jiffies;
1011         cfqq->slice_end = jiffies + slice;
1012         cfqq->allocated_slice = slice;
1013         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
1014 }
1015
1016 /*
1017  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1018  * isn't valid until the first request from the dispatch is activated
1019  * and the slice time set.
1020  */
1021 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1022 {
1023         if (cfq_cfqq_slice_new(cfqq))
1024                 return false;
1025         if (time_before(jiffies, cfqq->slice_end))
1026                 return false;
1027
1028         return true;
1029 }
1030
1031 /*
1032  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1033  * We choose the request that is closest to the head right now. Distance
1034  * behind the head is penalized and only allowed to a certain extent.
1035  */
1036 static struct request *
1037 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1038 {
1039         sector_t s1, s2, d1 = 0, d2 = 0;
1040         unsigned long back_max;
1041 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1042 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1043         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1044
1045         if (rq1 == NULL || rq1 == rq2)
1046                 return rq2;
1047         if (rq2 == NULL)
1048                 return rq1;
1049
1050         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1051                 return rq_is_sync(rq1) ? rq1 : rq2;
1052
1053         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1054                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1055
1056         s1 = blk_rq_pos(rq1);
1057         s2 = blk_rq_pos(rq2);
1058
1059         /*
1060          * by definition, 1KiB is 2 sectors
1061          */
1062         back_max = cfqd->cfq_back_max * 2;
1063
1064         /*
1065          * Strict one way elevator _except_ in the case where we allow
1066          * short backward seeks which are biased as twice the cost of a
1067          * similar forward seek.
1068          */
1069         if (s1 >= last)
1070                 d1 = s1 - last;
1071         else if (s1 + back_max >= last)
1072                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1073         else
1074                 wrap |= CFQ_RQ1_WRAP;
1075
1076         if (s2 >= last)
1077                 d2 = s2 - last;
1078         else if (s2 + back_max >= last)
1079                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1080         else
1081                 wrap |= CFQ_RQ2_WRAP;
1082
1083         /* Found required data */
1084
1085         /*
1086          * By doing switch() on the bit mask "wrap" we avoid having to
1087          * check two variables for all permutations: --> faster!
1088          */
1089         switch (wrap) {
1090         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1091                 if (d1 < d2)
1092                         return rq1;
1093                 else if (d2 < d1)
1094                         return rq2;
1095                 else {
1096                         if (s1 >= s2)
1097                                 return rq1;
1098                         else
1099                                 return rq2;
1100                 }
1101
1102         case CFQ_RQ2_WRAP:
1103                 return rq1;
1104         case CFQ_RQ1_WRAP:
1105                 return rq2;
1106         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1107         default:
1108                 /*
1109                  * Since both rqs are wrapped,
1110                  * start with the one that's further behind head
1111                  * (--> only *one* back seek required),
1112                  * since back seek takes more time than forward.
1113                  */
1114                 if (s1 <= s2)
1115                         return rq1;
1116                 else
1117                         return rq2;
1118         }
1119 }
1120
1121 /*
1122  * The below is leftmost cache rbtree addon
1123  */
1124 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1125 {
1126         /* Service tree is empty */
1127         if (!root->count)
1128                 return NULL;
1129
1130         if (!root->left)
1131                 root->left = rb_first(&root->rb);
1132
1133         if (root->left)
1134                 return rb_entry(root->left, struct cfq_queue, rb_node);
1135
1136         return NULL;
1137 }
1138
1139 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1140 {
1141         if (!root->left)
1142                 root->left = rb_first(&root->rb);
1143
1144         if (root->left)
1145                 return rb_entry_cfqg(root->left);
1146
1147         return NULL;
1148 }
1149
1150 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
1151 {
1152         rb_erase(n, root);
1153         RB_CLEAR_NODE(n);
1154 }
1155
1156 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1157 {
1158         if (root->left == n)
1159                 root->left = NULL;
1160         rb_erase_init(n, &root->rb);
1161         --root->count;
1162 }
1163
1164 /*
1165  * would be nice to take fifo expire time into account as well
1166  */
1167 static struct request *
1168 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1169                   struct request *last)
1170 {
1171         struct rb_node *rbnext = rb_next(&last->rb_node);
1172         struct rb_node *rbprev = rb_prev(&last->rb_node);
1173         struct request *next = NULL, *prev = NULL;
1174
1175         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1176
1177         if (rbprev)
1178                 prev = rb_entry_rq(rbprev);
1179
1180         if (rbnext)
1181                 next = rb_entry_rq(rbnext);
1182         else {
1183                 rbnext = rb_first(&cfqq->sort_list);
1184                 if (rbnext && rbnext != &last->rb_node)
1185                         next = rb_entry_rq(rbnext);
1186         }
1187
1188         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1189 }
1190
1191 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
1192                                       struct cfq_queue *cfqq)
1193 {
1194         /*
1195          * just an approximation, should be ok.
1196          */
1197         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1198                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1199 }
1200
1201 static inline s64
1202 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1203 {
1204         return cfqg->vdisktime - st->min_vdisktime;
1205 }
1206
1207 static void
1208 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1209 {
1210         struct rb_node **node = &st->rb.rb_node;
1211         struct rb_node *parent = NULL;
1212         struct cfq_group *__cfqg;
1213         s64 key = cfqg_key(st, cfqg);
1214         int left = 1;
1215
1216         while (*node != NULL) {
1217                 parent = *node;
1218                 __cfqg = rb_entry_cfqg(parent);
1219
1220                 if (key < cfqg_key(st, __cfqg))
1221                         node = &parent->rb_left;
1222                 else {
1223                         node = &parent->rb_right;
1224                         left = 0;
1225                 }
1226         }
1227
1228         if (left)
1229                 st->left = &cfqg->rb_node;
1230
1231         rb_link_node(&cfqg->rb_node, parent, node);
1232         rb_insert_color(&cfqg->rb_node, &st->rb);
1233 }
1234
1235 static void
1236 cfq_update_group_weight(struct cfq_group *cfqg)
1237 {
1238         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1239
1240         if (cfqg->new_weight) {
1241                 cfqg->weight = cfqg->new_weight;
1242                 cfqg->new_weight = 0;
1243         }
1244
1245         if (cfqg->new_leaf_weight) {
1246                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1247                 cfqg->new_leaf_weight = 0;
1248         }
1249 }
1250
1251 static void
1252 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1253 {
1254         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1255         struct cfq_group *pos = cfqg;
1256         struct cfq_group *parent;
1257         bool propagate;
1258
1259         /* add to the service tree */
1260         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1261
1262         cfq_update_group_weight(cfqg);
1263         __cfq_group_service_tree_add(st, cfqg);
1264
1265         /*
1266          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1267          * entitled to.  vfraction is calculated by walking the tree
1268          * towards the root calculating the fraction it has at each level.
1269          * The compounded ratio is how much vfraction @cfqg owns.
1270          *
1271          * Start with the proportion tasks in this cfqg has against active
1272          * children cfqgs - its leaf_weight against children_weight.
1273          */
1274         propagate = !pos->nr_active++;
1275         pos->children_weight += pos->leaf_weight;
1276         vfr = vfr * pos->leaf_weight / pos->children_weight;
1277
1278         /*
1279          * Compound ->weight walking up the tree.  Both activation and
1280          * vfraction calculation are done in the same loop.  Propagation
1281          * stops once an already activated node is met.  vfraction
1282          * calculation should always continue to the root.
1283          */
1284         while ((parent = cfqg_parent(pos))) {
1285                 if (propagate) {
1286                         propagate = !parent->nr_active++;
1287                         parent->children_weight += pos->weight;
1288                 }
1289                 vfr = vfr * pos->weight / parent->children_weight;
1290                 pos = parent;
1291         }
1292
1293         cfqg->vfraction = max_t(unsigned, vfr, 1);
1294 }
1295
1296 static void
1297 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1298 {
1299         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1300         struct cfq_group *__cfqg;
1301         struct rb_node *n;
1302
1303         cfqg->nr_cfqq++;
1304         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1305                 return;
1306
1307         /*
1308          * Currently put the group at the end. Later implement something
1309          * so that groups get lesser vtime based on their weights, so that
1310          * if group does not loose all if it was not continuously backlogged.
1311          */
1312         n = rb_last(&st->rb);
1313         if (n) {
1314                 __cfqg = rb_entry_cfqg(n);
1315                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
1316         } else
1317                 cfqg->vdisktime = st->min_vdisktime;
1318         cfq_group_service_tree_add(st, cfqg);
1319 }
1320
1321 static void
1322 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1323 {
1324         struct cfq_group *pos = cfqg;
1325         bool propagate;
1326
1327         /*
1328          * Undo activation from cfq_group_service_tree_add().  Deactivate
1329          * @cfqg and propagate deactivation upwards.
1330          */
1331         propagate = !--pos->nr_active;
1332         pos->children_weight -= pos->leaf_weight;
1333
1334         while (propagate) {
1335                 struct cfq_group *parent = cfqg_parent(pos);
1336
1337                 /* @pos has 0 nr_active at this point */
1338                 WARN_ON_ONCE(pos->children_weight);
1339                 pos->vfraction = 0;
1340
1341                 if (!parent)
1342                         break;
1343
1344                 propagate = !--parent->nr_active;
1345                 parent->children_weight -= pos->weight;
1346                 pos = parent;
1347         }
1348
1349         /* remove from the service tree */
1350         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1351                 cfq_rb_erase(&cfqg->rb_node, st);
1352 }
1353
1354 static void
1355 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1356 {
1357         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1358
1359         BUG_ON(cfqg->nr_cfqq < 1);
1360         cfqg->nr_cfqq--;
1361
1362         /* If there are other cfq queues under this group, don't delete it */
1363         if (cfqg->nr_cfqq)
1364                 return;
1365
1366         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1367         cfq_group_service_tree_del(st, cfqg);
1368         cfqg->saved_wl_slice = 0;
1369         cfqg_stats_update_dequeue(cfqg);
1370 }
1371
1372 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1373                                                 unsigned int *unaccounted_time)
1374 {
1375         unsigned int slice_used;
1376
1377         /*
1378          * Queue got expired before even a single request completed or
1379          * got expired immediately after first request completion.
1380          */
1381         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
1382                 /*
1383                  * Also charge the seek time incurred to the group, otherwise
1384                  * if there are mutiple queues in the group, each can dispatch
1385                  * a single request on seeky media and cause lots of seek time
1386                  * and group will never know it.
1387                  */
1388                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
1389                                         1);
1390         } else {
1391                 slice_used = jiffies - cfqq->slice_start;
1392                 if (slice_used > cfqq->allocated_slice) {
1393                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1394                         slice_used = cfqq->allocated_slice;
1395                 }
1396                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
1397                         *unaccounted_time += cfqq->slice_start -
1398                                         cfqq->dispatch_start;
1399         }
1400
1401         return slice_used;
1402 }
1403
1404 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1405                                 struct cfq_queue *cfqq)
1406 {
1407         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1408         unsigned int used_sl, charge, unaccounted_sl = 0;
1409         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1410                         - cfqg->service_tree_idle.count;
1411         unsigned int vfr;
1412
1413         BUG_ON(nr_sync < 0);
1414         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1415
1416         if (iops_mode(cfqd))
1417                 charge = cfqq->slice_dispatch;
1418         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1419                 charge = cfqq->allocated_slice;
1420
1421         /*
1422          * Can't update vdisktime while on service tree and cfqg->vfraction
1423          * is valid only while on it.  Cache vfr, leave the service tree,
1424          * update vdisktime and go back on.  The re-addition to the tree
1425          * will also update the weights as necessary.
1426          */
1427         vfr = cfqg->vfraction;
1428         cfq_group_service_tree_del(st, cfqg);
1429         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1430         cfq_group_service_tree_add(st, cfqg);
1431
1432         /* This group is being expired. Save the context */
1433         if (time_after(cfqd->workload_expires, jiffies)) {
1434                 cfqg->saved_wl_slice = cfqd->workload_expires
1435                                                 - jiffies;
1436                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1437                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1438         } else
1439                 cfqg->saved_wl_slice = 0;
1440
1441         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1442                                         st->min_vdisktime);
1443         cfq_log_cfqq(cfqq->cfqd, cfqq,
1444                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1445                      used_sl, cfqq->slice_dispatch, charge,
1446                      iops_mode(cfqd), cfqq->nr_sectors);
1447         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1448         cfqg_stats_set_start_empty_time(cfqg);
1449 }
1450
1451 /**
1452  * cfq_init_cfqg_base - initialize base part of a cfq_group
1453  * @cfqg: cfq_group to initialize
1454  *
1455  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1456  * is enabled or not.
1457  */
1458 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1459 {
1460         struct cfq_rb_root *st;
1461         int i, j;
1462
1463         for_each_cfqg_st(cfqg, i, j, st)
1464                 *st = CFQ_RB_ROOT;
1465         RB_CLEAR_NODE(&cfqg->rb_node);
1466
1467         cfqg->ttime.last_end_request = jiffies;
1468 }
1469
1470 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1471 static void cfq_pd_init(struct blkcg_gq *blkg)
1472 {
1473         struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1474
1475         cfq_init_cfqg_base(cfqg);
1476         cfqg->weight = blkg->blkcg->cfq_weight;
1477         cfqg->leaf_weight = blkg->blkcg->cfq_leaf_weight;
1478 }
1479
1480 /*
1481  * Search for the cfq group current task belongs to. request_queue lock must
1482  * be held.
1483  */
1484 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1485                                                 struct blkcg *blkcg)
1486 {
1487         struct request_queue *q = cfqd->queue;
1488         struct cfq_group *cfqg = NULL;
1489
1490         /* avoid lookup for the common case where there's no blkcg */
1491         if (blkcg == &blkcg_root) {
1492                 cfqg = cfqd->root_group;
1493         } else {
1494                 struct blkcg_gq *blkg;
1495
1496                 blkg = blkg_lookup_create(blkcg, q);
1497                 if (!IS_ERR(blkg))
1498                         cfqg = blkg_to_cfqg(blkg);
1499         }
1500
1501         return cfqg;
1502 }
1503
1504 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1505 {
1506         /* Currently, all async queues are mapped to root group */
1507         if (!cfq_cfqq_sync(cfqq))
1508                 cfqg = cfqq->cfqd->root_group;
1509
1510         cfqq->cfqg = cfqg;
1511         /* cfqq reference on cfqg */
1512         cfqg_get(cfqg);
1513 }
1514
1515 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1516                                      struct blkg_policy_data *pd, int off)
1517 {
1518         struct cfq_group *cfqg = pd_to_cfqg(pd);
1519
1520         if (!cfqg->dev_weight)
1521                 return 0;
1522         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1523 }
1524
1525 static int cfqg_print_weight_device(struct cgroup *cgrp, struct cftype *cft,
1526                                     struct seq_file *sf)
1527 {
1528         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1529                           cfqg_prfill_weight_device, &blkcg_policy_cfq, 0,
1530                           false);
1531         return 0;
1532 }
1533
1534 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1535                                           struct blkg_policy_data *pd, int off)
1536 {
1537         struct cfq_group *cfqg = pd_to_cfqg(pd);
1538
1539         if (!cfqg->dev_leaf_weight)
1540                 return 0;
1541         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1542 }
1543
1544 static int cfqg_print_leaf_weight_device(struct cgroup *cgrp,
1545                                          struct cftype *cft,
1546                                          struct seq_file *sf)
1547 {
1548         blkcg_print_blkgs(sf, cgroup_to_blkcg(cgrp),
1549                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq, 0,
1550                           false);
1551         return 0;
1552 }
1553
1554 static int cfq_print_weight(struct cgroup *cgrp, struct cftype *cft,
1555                             struct seq_file *sf)
1556 {
1557         seq_printf(sf, "%u\n", cgroup_to_blkcg(cgrp)->cfq_weight);
1558         return 0;
1559 }
1560
1561 static int cfq_print_leaf_weight(struct cgroup *cgrp, struct cftype *cft,
1562                                  struct seq_file *sf)
1563 {
1564         seq_printf(sf, "%u\n",
1565                    cgroup_to_blkcg(cgrp)->cfq_leaf_weight);
1566         return 0;
1567 }
1568
1569 static int __cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1570                                     const char *buf, bool is_leaf_weight)
1571 {
1572         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1573         struct blkg_conf_ctx ctx;
1574         struct cfq_group *cfqg;
1575         int ret;
1576
1577         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1578         if (ret)
1579                 return ret;
1580
1581         ret = -EINVAL;
1582         cfqg = blkg_to_cfqg(ctx.blkg);
1583         if (!ctx.v || (ctx.v >= CFQ_WEIGHT_MIN && ctx.v <= CFQ_WEIGHT_MAX)) {
1584                 if (!is_leaf_weight) {
1585                         cfqg->dev_weight = ctx.v;
1586                         cfqg->new_weight = ctx.v ?: blkcg->cfq_weight;
1587                 } else {
1588                         cfqg->dev_leaf_weight = ctx.v;
1589                         cfqg->new_leaf_weight = ctx.v ?: blkcg->cfq_leaf_weight;
1590                 }
1591                 ret = 0;
1592         }
1593
1594         blkg_conf_finish(&ctx);
1595         return ret;
1596 }
1597
1598 static int cfqg_set_weight_device(struct cgroup *cgrp, struct cftype *cft,
1599                                   const char *buf)
1600 {
1601         return __cfqg_set_weight_device(cgrp, cft, buf, false);
1602 }
1603
1604 static int cfqg_set_leaf_weight_device(struct cgroup *cgrp, struct cftype *cft,
1605                                        const char *buf)
1606 {
1607         return __cfqg_set_weight_device(cgrp, cft, buf, true);
1608 }
1609
1610 static int __cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val,
1611                             bool is_leaf_weight)
1612 {
1613         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1614         struct blkcg_gq *blkg;
1615         struct hlist_node *n;
1616
1617         if (val < CFQ_WEIGHT_MIN || val > CFQ_WEIGHT_MAX)
1618                 return -EINVAL;
1619
1620         spin_lock_irq(&blkcg->lock);
1621
1622         if (!is_leaf_weight)
1623                 blkcg->cfq_weight = val;
1624         else
1625                 blkcg->cfq_leaf_weight = val;
1626
1627         hlist_for_each_entry(blkg, n, &blkcg->blkg_list, blkcg_node) {
1628                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1629
1630                 if (!cfqg)
1631                         continue;
1632
1633                 if (!is_leaf_weight) {
1634                         if (!cfqg->dev_weight)
1635                                 cfqg->new_weight = blkcg->cfq_weight;
1636                 } else {
1637                         if (!cfqg->dev_leaf_weight)
1638                                 cfqg->new_leaf_weight = blkcg->cfq_leaf_weight;
1639                 }
1640         }
1641
1642         spin_unlock_irq(&blkcg->lock);
1643         return 0;
1644 }
1645
1646 static int cfq_set_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1647 {
1648         return __cfq_set_weight(cgrp, cft, val, false);
1649 }
1650
1651 static int cfq_set_leaf_weight(struct cgroup *cgrp, struct cftype *cft, u64 val)
1652 {
1653         return __cfq_set_weight(cgrp, cft, val, true);
1654 }
1655
1656 static int cfqg_print_stat(struct cgroup *cgrp, struct cftype *cft,
1657                            struct seq_file *sf)
1658 {
1659         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1660
1661         blkcg_print_blkgs(sf, blkcg, blkg_prfill_stat, &blkcg_policy_cfq,
1662                           cft->private, false);
1663         return 0;
1664 }
1665
1666 static int cfqg_print_rwstat(struct cgroup *cgrp, struct cftype *cft,
1667                              struct seq_file *sf)
1668 {
1669         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1670
1671         blkcg_print_blkgs(sf, blkcg, blkg_prfill_rwstat, &blkcg_policy_cfq,
1672                           cft->private, true);
1673         return 0;
1674 }
1675
1676 #ifdef CONFIG_DEBUG_BLK_CGROUP
1677 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1678                                       struct blkg_policy_data *pd, int off)
1679 {
1680         struct cfq_group *cfqg = pd_to_cfqg(pd);
1681         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1682         u64 v = 0;
1683
1684         if (samples) {
1685                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1686                 do_div(v, samples);
1687         }
1688         __blkg_prfill_u64(sf, pd, v);
1689         return 0;
1690 }
1691
1692 /* print avg_queue_size */
1693 static int cfqg_print_avg_queue_size(struct cgroup *cgrp, struct cftype *cft,
1694                                      struct seq_file *sf)
1695 {
1696         struct blkcg *blkcg = cgroup_to_blkcg(cgrp);
1697
1698         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_avg_queue_size,
1699                           &blkcg_policy_cfq, 0, false);
1700         return 0;
1701 }
1702 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1703
1704 static struct cftype cfq_blkcg_files[] = {
1705         /* on root, weight is mapped to leaf_weight */
1706         {
1707                 .name = "weight_device",
1708                 .flags = CFTYPE_ONLY_ON_ROOT,
1709                 .read_seq_string = cfqg_print_leaf_weight_device,
1710                 .write_string = cfqg_set_leaf_weight_device,
1711                 .max_write_len = 256,
1712         },
1713         {
1714                 .name = "weight",
1715                 .flags = CFTYPE_ONLY_ON_ROOT,
1716                 .read_seq_string = cfq_print_leaf_weight,
1717                 .write_u64 = cfq_set_leaf_weight,
1718         },
1719
1720         /* no such mapping necessary for !roots */
1721         {
1722                 .name = "weight_device",
1723                 .flags = CFTYPE_NOT_ON_ROOT,
1724                 .read_seq_string = cfqg_print_weight_device,
1725                 .write_string = cfqg_set_weight_device,
1726                 .max_write_len = 256,
1727         },
1728         {
1729                 .name = "weight",
1730                 .flags = CFTYPE_NOT_ON_ROOT,
1731                 .read_seq_string = cfq_print_weight,
1732                 .write_u64 = cfq_set_weight,
1733         },
1734
1735         {
1736                 .name = "leaf_weight_device",
1737                 .read_seq_string = cfqg_print_leaf_weight_device,
1738                 .write_string = cfqg_set_leaf_weight_device,
1739                 .max_write_len = 256,
1740         },
1741         {
1742                 .name = "leaf_weight",
1743                 .read_seq_string = cfq_print_leaf_weight,
1744                 .write_u64 = cfq_set_leaf_weight,
1745         },
1746
1747         {
1748                 .name = "time",
1749                 .private = offsetof(struct cfq_group, stats.time),
1750                 .read_seq_string = cfqg_print_stat,
1751         },
1752         {
1753                 .name = "sectors",
1754                 .private = offsetof(struct cfq_group, stats.sectors),
1755                 .read_seq_string = cfqg_print_stat,
1756         },
1757         {
1758                 .name = "io_service_bytes",
1759                 .private = offsetof(struct cfq_group, stats.service_bytes),
1760                 .read_seq_string = cfqg_print_rwstat,
1761         },
1762         {
1763                 .name = "io_serviced",
1764                 .private = offsetof(struct cfq_group, stats.serviced),
1765                 .read_seq_string = cfqg_print_rwstat,
1766         },
1767         {
1768                 .name = "io_service_time",
1769                 .private = offsetof(struct cfq_group, stats.service_time),
1770                 .read_seq_string = cfqg_print_rwstat,
1771         },
1772         {
1773                 .name = "io_wait_time",
1774                 .private = offsetof(struct cfq_group, stats.wait_time),
1775                 .read_seq_string = cfqg_print_rwstat,
1776         },
1777         {
1778                 .name = "io_merged",
1779                 .private = offsetof(struct cfq_group, stats.merged),
1780                 .read_seq_string = cfqg_print_rwstat,
1781         },
1782         {
1783                 .name = "io_queued",
1784                 .private = offsetof(struct cfq_group, stats.queued),
1785                 .read_seq_string = cfqg_print_rwstat,
1786         },
1787 #ifdef CONFIG_DEBUG_BLK_CGROUP
1788         {
1789                 .name = "avg_queue_size",
1790                 .read_seq_string = cfqg_print_avg_queue_size,
1791         },
1792         {
1793                 .name = "group_wait_time",
1794                 .private = offsetof(struct cfq_group, stats.group_wait_time),
1795                 .read_seq_string = cfqg_print_stat,
1796         },
1797         {
1798                 .name = "idle_time",
1799                 .private = offsetof(struct cfq_group, stats.idle_time),
1800                 .read_seq_string = cfqg_print_stat,
1801         },
1802         {
1803                 .name = "empty_time",
1804                 .private = offsetof(struct cfq_group, stats.empty_time),
1805                 .read_seq_string = cfqg_print_stat,
1806         },
1807         {
1808                 .name = "dequeue",
1809                 .private = offsetof(struct cfq_group, stats.dequeue),
1810                 .read_seq_string = cfqg_print_stat,
1811         },
1812         {
1813                 .name = "unaccounted_time",
1814                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
1815                 .read_seq_string = cfqg_print_stat,
1816         },
1817 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1818         { }     /* terminate */
1819 };
1820 #else /* GROUP_IOSCHED */
1821 static struct cfq_group *cfq_lookup_create_cfqg(struct cfq_data *cfqd,
1822                                                 struct blkcg *blkcg)
1823 {
1824         return cfqd->root_group;
1825 }
1826
1827 static inline void
1828 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1829         cfqq->cfqg = cfqg;
1830 }
1831
1832 #endif /* GROUP_IOSCHED */
1833
1834 /*
1835  * The cfqd->service_trees holds all pending cfq_queue's that have
1836  * requests waiting to be processed. It is sorted in the order that
1837  * we will service the queues.
1838  */
1839 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1840                                  bool add_front)
1841 {
1842         struct rb_node **p, *parent;
1843         struct cfq_queue *__cfqq;
1844         unsigned long rb_key;
1845         struct cfq_rb_root *st;
1846         int left;
1847         int new_cfqq = 1;
1848
1849         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
1850         if (cfq_class_idle(cfqq)) {
1851                 rb_key = CFQ_IDLE_DELAY;
1852                 parent = rb_last(&st->rb);
1853                 if (parent && parent != &cfqq->rb_node) {
1854                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1855                         rb_key += __cfqq->rb_key;
1856                 } else
1857                         rb_key += jiffies;
1858         } else if (!add_front) {
1859                 /*
1860                  * Get our rb key offset. Subtract any residual slice
1861                  * value carried from last service. A negative resid
1862                  * count indicates slice overrun, and this should position
1863                  * the next service time further away in the tree.
1864                  */
1865                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1866                 rb_key -= cfqq->slice_resid;
1867                 cfqq->slice_resid = 0;
1868         } else {
1869                 rb_key = -HZ;
1870                 __cfqq = cfq_rb_first(st);
1871                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1872         }
1873
1874         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1875                 new_cfqq = 0;
1876                 /*
1877                  * same position, nothing more to do
1878                  */
1879                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
1880                         return;
1881
1882                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1883                 cfqq->service_tree = NULL;
1884         }
1885
1886         left = 1;
1887         parent = NULL;
1888         cfqq->service_tree = st;
1889         p = &st->rb.rb_node;
1890         while (*p) {
1891                 parent = *p;
1892                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1893
1894                 /*
1895                  * sort by key, that represents service time.
1896                  */
1897                 if (time_before(rb_key, __cfqq->rb_key))
1898                         p = &parent->rb_left;
1899                 else {
1900                         p = &parent->rb_right;
1901                         left = 0;
1902                 }
1903         }
1904
1905         if (left)
1906                 st->left = &cfqq->rb_node;
1907
1908         cfqq->rb_key = rb_key;
1909         rb_link_node(&cfqq->rb_node, parent, p);
1910         rb_insert_color(&cfqq->rb_node, &st->rb);
1911         st->count++;
1912         if (add_front || !new_cfqq)
1913                 return;
1914         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1915 }
1916
1917 static struct cfq_queue *
1918 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1919                      sector_t sector, struct rb_node **ret_parent,
1920                      struct rb_node ***rb_link)
1921 {
1922         struct rb_node **p, *parent;
1923         struct cfq_queue *cfqq = NULL;
1924
1925         parent = NULL;
1926         p = &root->rb_node;
1927         while (*p) {
1928                 struct rb_node **n;
1929
1930                 parent = *p;
1931                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1932
1933                 /*
1934                  * Sort strictly based on sector.  Smallest to the left,
1935                  * largest to the right.
1936                  */
1937                 if (sector > blk_rq_pos(cfqq->next_rq))
1938                         n = &(*p)->rb_right;
1939                 else if (sector < blk_rq_pos(cfqq->next_rq))
1940                         n = &(*p)->rb_left;
1941                 else
1942                         break;
1943                 p = n;
1944                 cfqq = NULL;
1945         }
1946
1947         *ret_parent = parent;
1948         if (rb_link)
1949                 *rb_link = p;
1950         return cfqq;
1951 }
1952
1953 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1954 {
1955         struct rb_node **p, *parent;
1956         struct cfq_queue *__cfqq;
1957
1958         if (cfqq->p_root) {
1959                 rb_erase(&cfqq->p_node, cfqq->p_root);
1960                 cfqq->p_root = NULL;
1961         }
1962
1963         if (cfq_class_idle(cfqq))
1964                 return;
1965         if (!cfqq->next_rq)
1966                 return;
1967
1968         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1969         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1970                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1971         if (!__cfqq) {
1972                 rb_link_node(&cfqq->p_node, parent, p);
1973                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1974         } else
1975                 cfqq->p_root = NULL;
1976 }
1977
1978 /*
1979  * Update cfqq's position in the service tree.
1980  */
1981 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1982 {
1983         /*
1984          * Resorting requires the cfqq to be on the RR list already.
1985          */
1986         if (cfq_cfqq_on_rr(cfqq)) {
1987                 cfq_service_tree_add(cfqd, cfqq, 0);
1988                 cfq_prio_tree_add(cfqd, cfqq);
1989         }
1990 }
1991
1992 /*
1993  * add to busy list of queues for service, trying to be fair in ordering
1994  * the pending list according to last request service
1995  */
1996 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1997 {
1998         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1999         BUG_ON(cfq_cfqq_on_rr(cfqq));
2000         cfq_mark_cfqq_on_rr(cfqq);
2001         cfqd->busy_queues++;
2002         if (cfq_cfqq_sync(cfqq))
2003                 cfqd->busy_sync_queues++;
2004
2005         cfq_resort_rr_list(cfqd, cfqq);
2006 }
2007
2008 /*
2009  * Called when the cfqq no longer has requests pending, remove it from
2010  * the service tree.
2011  */
2012 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2013 {
2014         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2015         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2016         cfq_clear_cfqq_on_rr(cfqq);
2017
2018         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2019                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2020                 cfqq->service_tree = NULL;
2021         }
2022         if (cfqq->p_root) {
2023                 rb_erase(&cfqq->p_node, cfqq->p_root);
2024                 cfqq->p_root = NULL;
2025         }
2026
2027         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2028         BUG_ON(!cfqd->busy_queues);
2029         cfqd->busy_queues--;
2030         if (cfq_cfqq_sync(cfqq))
2031                 cfqd->busy_sync_queues--;
2032 }
2033
2034 /*
2035  * rb tree support functions
2036  */
2037 static void cfq_del_rq_rb(struct request *rq)
2038 {
2039         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2040         const int sync = rq_is_sync(rq);
2041
2042         BUG_ON(!cfqq->queued[sync]);
2043         cfqq->queued[sync]--;
2044
2045         elv_rb_del(&cfqq->sort_list, rq);
2046
2047         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2048                 /*
2049                  * Queue will be deleted from service tree when we actually
2050                  * expire it later. Right now just remove it from prio tree
2051                  * as it is empty.
2052                  */
2053                 if (cfqq->p_root) {
2054                         rb_erase(&cfqq->p_node, cfqq->p_root);
2055                         cfqq->p_root = NULL;
2056                 }
2057         }
2058 }
2059
2060 static void cfq_add_rq_rb(struct request *rq)
2061 {
2062         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2063         struct cfq_data *cfqd = cfqq->cfqd;
2064         struct request *prev;
2065
2066         cfqq->queued[rq_is_sync(rq)]++;
2067
2068         elv_rb_add(&cfqq->sort_list, rq);
2069
2070         if (!cfq_cfqq_on_rr(cfqq))
2071                 cfq_add_cfqq_rr(cfqd, cfqq);
2072
2073         /*
2074          * check if this request is a better next-serve candidate
2075          */
2076         prev = cfqq->next_rq;
2077         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2078
2079         /*
2080          * adjust priority tree position, if ->next_rq changes
2081          */
2082         if (prev != cfqq->next_rq)
2083                 cfq_prio_tree_add(cfqd, cfqq);
2084
2085         BUG_ON(!cfqq->next_rq);
2086 }
2087
2088 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2089 {
2090         elv_rb_del(&cfqq->sort_list, rq);
2091         cfqq->queued[rq_is_sync(rq)]--;
2092         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2093         cfq_add_rq_rb(rq);
2094         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2095                                  rq->cmd_flags);
2096 }
2097
2098 static struct request *
2099 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2100 {
2101         struct task_struct *tsk = current;
2102         struct cfq_io_cq *cic;
2103         struct cfq_queue *cfqq;
2104
2105         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2106         if (!cic)
2107                 return NULL;
2108
2109         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2110         if (cfqq) {
2111                 sector_t sector = bio->bi_sector + bio_sectors(bio);
2112
2113                 return elv_rb_find(&cfqq->sort_list, sector);
2114         }
2115
2116         return NULL;
2117 }
2118
2119 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2120 {
2121         struct cfq_data *cfqd = q->elevator->elevator_data;
2122
2123         cfqd->rq_in_driver++;
2124         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2125                                                 cfqd->rq_in_driver);
2126
2127         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2128 }
2129
2130 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2131 {
2132         struct cfq_data *cfqd = q->elevator->elevator_data;
2133
2134         WARN_ON(!cfqd->rq_in_driver);
2135         cfqd->rq_in_driver--;
2136         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2137                                                 cfqd->rq_in_driver);
2138 }
2139
2140 static void cfq_remove_request(struct request *rq)
2141 {
2142         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2143
2144         if (cfqq->next_rq == rq)
2145                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2146
2147         list_del_init(&rq->queuelist);
2148         cfq_del_rq_rb(rq);
2149
2150         cfqq->cfqd->rq_queued--;
2151         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2152         if (rq->cmd_flags & REQ_PRIO) {
2153                 WARN_ON(!cfqq->prio_pending);
2154                 cfqq->prio_pending--;
2155         }
2156 }
2157
2158 static int cfq_merge(struct request_queue *q, struct request **req,
2159                      struct bio *bio)
2160 {
2161         struct cfq_data *cfqd = q->elevator->elevator_data;
2162         struct request *__rq;
2163
2164         __rq = cfq_find_rq_fmerge(cfqd, bio);
2165         if (__rq && elv_rq_merge_ok(__rq, bio)) {
2166                 *req = __rq;
2167                 return ELEVATOR_FRONT_MERGE;
2168         }
2169
2170         return ELEVATOR_NO_MERGE;
2171 }
2172
2173 static void cfq_merged_request(struct request_queue *q, struct request *req,
2174                                int type)
2175 {
2176         if (type == ELEVATOR_FRONT_MERGE) {
2177                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2178
2179                 cfq_reposition_rq_rb(cfqq, req);
2180         }
2181 }
2182
2183 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2184                                 struct bio *bio)
2185 {
2186         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_rw);
2187 }
2188
2189 static void
2190 cfq_merged_requests(struct request_queue *q, struct request *rq,
2191                     struct request *next)
2192 {
2193         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2194         struct cfq_data *cfqd = q->elevator->elevator_data;
2195
2196         /*
2197          * reposition in fifo if next is older than rq
2198          */
2199         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2200             time_before(rq_fifo_time(next), rq_fifo_time(rq)) &&
2201             cfqq == RQ_CFQQ(next)) {
2202                 list_move(&rq->queuelist, &next->queuelist);
2203                 rq_set_fifo_time(rq, rq_fifo_time(next));
2204         }
2205
2206         if (cfqq->next_rq == next)
2207                 cfqq->next_rq = rq;
2208         cfq_remove_request(next);
2209         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2210
2211         cfqq = RQ_CFQQ(next);
2212         /*
2213          * all requests of this queue are merged to other queues, delete it
2214          * from the service tree. If it's the active_queue,
2215          * cfq_dispatch_requests() will choose to expire it or do idle
2216          */
2217         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2218             cfqq != cfqd->active_queue)
2219                 cfq_del_cfqq_rr(cfqd, cfqq);
2220 }
2221
2222 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
2223                            struct bio *bio)
2224 {
2225         struct cfq_data *cfqd = q->elevator->elevator_data;
2226         struct cfq_io_cq *cic;
2227         struct cfq_queue *cfqq;
2228
2229         /*
2230          * Disallow merge of a sync bio into an async request.
2231          */
2232         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
2233                 return false;
2234
2235         /*
2236          * Lookup the cfqq that this bio will be queued with and allow
2237          * merge only if rq is queued there.
2238          */
2239         cic = cfq_cic_lookup(cfqd, current->io_context);
2240         if (!cic)
2241                 return false;
2242
2243         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
2244         return cfqq == RQ_CFQQ(rq);
2245 }
2246
2247 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2248 {
2249         del_timer(&cfqd->idle_slice_timer);
2250         cfqg_stats_update_idle_time(cfqq->cfqg);
2251 }
2252
2253 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2254                                    struct cfq_queue *cfqq)
2255 {
2256         if (cfqq) {
2257                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2258                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2259                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2260                 cfqq->slice_start = 0;
2261                 cfqq->dispatch_start = jiffies;
2262                 cfqq->allocated_slice = 0;
2263                 cfqq->slice_end = 0;
2264                 cfqq->slice_dispatch = 0;
2265                 cfqq->nr_sectors = 0;
2266
2267                 cfq_clear_cfqq_wait_request(cfqq);
2268                 cfq_clear_cfqq_must_dispatch(cfqq);
2269                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2270                 cfq_clear_cfqq_fifo_expire(cfqq);
2271                 cfq_mark_cfqq_slice_new(cfqq);
2272
2273                 cfq_del_timer(cfqd, cfqq);
2274         }
2275
2276         cfqd->active_queue = cfqq;
2277 }
2278
2279 /*
2280  * current cfqq expired its slice (or was too idle), select new one
2281  */
2282 static void
2283 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2284                     bool timed_out)
2285 {
2286         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2287
2288         if (cfq_cfqq_wait_request(cfqq))
2289                 cfq_del_timer(cfqd, cfqq);
2290
2291         cfq_clear_cfqq_wait_request(cfqq);
2292         cfq_clear_cfqq_wait_busy(cfqq);
2293
2294         /*
2295          * If this cfqq is shared between multiple processes, check to
2296          * make sure that those processes are still issuing I/Os within
2297          * the mean seek distance.  If not, it may be time to break the
2298          * queues apart again.
2299          */
2300         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2301                 cfq_mark_cfqq_split_coop(cfqq);
2302
2303         /*
2304          * store what was left of this slice, if the queue idled/timed out
2305          */
2306         if (timed_out) {
2307                 if (cfq_cfqq_slice_new(cfqq))
2308                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2309                 else
2310                         cfqq->slice_resid = cfqq->slice_end - jiffies;
2311                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
2312         }
2313
2314         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2315
2316         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2317                 cfq_del_cfqq_rr(cfqd, cfqq);
2318
2319         cfq_resort_rr_list(cfqd, cfqq);
2320
2321         if (cfqq == cfqd->active_queue)
2322                 cfqd->active_queue = NULL;
2323
2324         if (cfqd->active_cic) {
2325                 put_io_context(cfqd->active_cic->icq.ioc);
2326                 cfqd->active_cic = NULL;
2327         }
2328 }
2329
2330 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2331 {
2332         struct cfq_queue *cfqq = cfqd->active_queue;
2333
2334         if (cfqq)
2335                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2336 }
2337
2338 /*
2339  * Get next queue for service. Unless we have a queue preemption,
2340  * we'll simply select the first cfqq in the service tree.
2341  */
2342 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2343 {
2344         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2345                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2346
2347         if (!cfqd->rq_queued)
2348                 return NULL;
2349
2350         /* There is nothing to dispatch */
2351         if (!st)
2352                 return NULL;
2353         if (RB_EMPTY_ROOT(&st->rb))
2354                 return NULL;
2355         return cfq_rb_first(st);
2356 }
2357
2358 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2359 {
2360         struct cfq_group *cfqg;
2361         struct cfq_queue *cfqq;
2362         int i, j;
2363         struct cfq_rb_root *st;
2364
2365         if (!cfqd->rq_queued)
2366                 return NULL;
2367
2368         cfqg = cfq_get_next_cfqg(cfqd);
2369         if (!cfqg)
2370                 return NULL;
2371
2372         for_each_cfqg_st(cfqg, i, j, st)
2373                 if ((cfqq = cfq_rb_first(st)) != NULL)
2374                         return cfqq;
2375         return NULL;
2376 }
2377
2378 /*
2379  * Get and set a new active queue for service.
2380  */
2381 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2382                                               struct cfq_queue *cfqq)
2383 {
2384         if (!cfqq)
2385                 cfqq = cfq_get_next_queue(cfqd);
2386
2387         __cfq_set_active_queue(cfqd, cfqq);
2388         return cfqq;
2389 }
2390
2391 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2392                                           struct request *rq)
2393 {
2394         if (blk_rq_pos(rq) >= cfqd->last_position)
2395                 return blk_rq_pos(rq) - cfqd->last_position;
2396         else
2397                 return cfqd->last_position - blk_rq_pos(rq);
2398 }
2399
2400 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2401                                struct request *rq)
2402 {
2403         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2404 }
2405
2406 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2407                                     struct cfq_queue *cur_cfqq)
2408 {
2409         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2410         struct rb_node *parent, *node;
2411         struct cfq_queue *__cfqq;
2412         sector_t sector = cfqd->last_position;
2413
2414         if (RB_EMPTY_ROOT(root))
2415                 return NULL;
2416
2417         /*
2418          * First, if we find a request starting at the end of the last
2419          * request, choose it.
2420          */
2421         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2422         if (__cfqq)
2423                 return __cfqq;
2424
2425         /*
2426          * If the exact sector wasn't found, the parent of the NULL leaf
2427          * will contain the closest sector.
2428          */
2429         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2430         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2431                 return __cfqq;
2432
2433         if (blk_rq_pos(__cfqq->next_rq) < sector)
2434                 node = rb_next(&__cfqq->p_node);
2435         else
2436                 node = rb_prev(&__cfqq->p_node);
2437         if (!node)
2438                 return NULL;
2439
2440         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2441         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2442                 return __cfqq;
2443
2444         return NULL;
2445 }
2446
2447 /*
2448  * cfqd - obvious
2449  * cur_cfqq - passed in so that we don't decide that the current queue is
2450  *            closely cooperating with itself.
2451  *
2452  * So, basically we're assuming that that cur_cfqq has dispatched at least
2453  * one request, and that cfqd->last_position reflects a position on the disk
2454  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2455  * assumption.
2456  */
2457 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2458                                               struct cfq_queue *cur_cfqq)
2459 {
2460         struct cfq_queue *cfqq;
2461
2462         if (cfq_class_idle(cur_cfqq))
2463                 return NULL;
2464         if (!cfq_cfqq_sync(cur_cfqq))
2465                 return NULL;
2466         if (CFQQ_SEEKY(cur_cfqq))
2467                 return NULL;
2468
2469         /*
2470          * Don't search priority tree if it's the only queue in the group.
2471          */
2472         if (cur_cfqq->cfqg->nr_cfqq == 1)
2473                 return NULL;
2474
2475         /*
2476          * We should notice if some of the queues are cooperating, eg
2477          * working closely on the same area of the disk. In that case,
2478          * we can group them together and don't waste time idling.
2479          */
2480         cfqq = cfqq_close(cfqd, cur_cfqq);
2481         if (!cfqq)
2482                 return NULL;
2483
2484         /* If new queue belongs to different cfq_group, don't choose it */
2485         if (cur_cfqq->cfqg != cfqq->cfqg)
2486                 return NULL;
2487
2488         /*
2489          * It only makes sense to merge sync queues.
2490          */
2491         if (!cfq_cfqq_sync(cfqq))
2492                 return NULL;
2493         if (CFQQ_SEEKY(cfqq))
2494                 return NULL;
2495
2496         /*
2497          * Do not merge queues of different priority classes
2498          */
2499         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2500                 return NULL;
2501
2502         return cfqq;
2503 }
2504
2505 /*
2506  * Determine whether we should enforce idle window for this queue.
2507  */
2508
2509 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2510 {
2511         enum wl_class_t wl_class = cfqq_class(cfqq);
2512         struct cfq_rb_root *st = cfqq->service_tree;
2513
2514         BUG_ON(!st);
2515         BUG_ON(!st->count);
2516
2517         if (!cfqd->cfq_slice_idle)
2518                 return false;
2519
2520         /* We never do for idle class queues. */
2521         if (wl_class == IDLE_WORKLOAD)
2522                 return false;
2523
2524         /* We do for queues that were marked with idle window flag. */
2525         if (cfq_cfqq_idle_window(cfqq) &&
2526            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2527                 return true;
2528
2529         /*
2530          * Otherwise, we do only if they are the last ones
2531          * in their service tree.
2532          */
2533         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2534            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2535                 return true;
2536         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2537         return false;
2538 }
2539
2540 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2541 {
2542         struct cfq_queue *cfqq = cfqd->active_queue;
2543         struct cfq_io_cq *cic;
2544         unsigned long sl, group_idle = 0;
2545
2546         /*
2547          * SSD device without seek penalty, disable idling. But only do so
2548          * for devices that support queuing, otherwise we still have a problem
2549          * with sync vs async workloads.
2550          */
2551         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2552                 return;
2553
2554         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2555         WARN_ON(cfq_cfqq_slice_new(cfqq));
2556
2557         /*
2558          * idle is disabled, either manually or by past process history
2559          */
2560         if (!cfq_should_idle(cfqd, cfqq)) {
2561                 /* no queue idling. Check for group idling */
2562                 if (cfqd->cfq_group_idle)
2563                         group_idle = cfqd->cfq_group_idle;
2564                 else
2565                         return;
2566         }
2567
2568         /*
2569          * still active requests from this queue, don't idle
2570          */
2571         if (cfqq->dispatched)
2572                 return;
2573
2574         /*
2575          * task has exited, don't wait
2576          */
2577         cic = cfqd->active_cic;
2578         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2579                 return;
2580
2581         /*
2582          * If our average think time is larger than the remaining time
2583          * slice, then don't idle. This avoids overrunning the allotted
2584          * time slice.
2585          */
2586         if (sample_valid(cic->ttime.ttime_samples) &&
2587             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2588                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2589                              cic->ttime.ttime_mean);
2590                 return;
2591         }
2592
2593         /* There are other queues in the group, don't do group idle */
2594         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2595                 return;
2596
2597         cfq_mark_cfqq_wait_request(cfqq);
2598
2599         if (group_idle)
2600                 sl = cfqd->cfq_group_idle;
2601         else
2602                 sl = cfqd->cfq_slice_idle;
2603
2604         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2605         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2606         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2607                         group_idle ? 1 : 0);
2608 }
2609
2610 /*
2611  * Move request from internal lists to the request queue dispatch list.
2612  */
2613 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2614 {
2615         struct cfq_data *cfqd = q->elevator->elevator_data;
2616         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2617
2618         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2619
2620         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2621         cfq_remove_request(rq);
2622         cfqq->dispatched++;
2623         (RQ_CFQG(rq))->dispatched++;
2624         elv_dispatch_sort(q, rq);
2625
2626         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2627         cfqq->nr_sectors += blk_rq_sectors(rq);
2628         cfqg_stats_update_dispatch(cfqq->cfqg, blk_rq_bytes(rq), rq->cmd_flags);
2629 }
2630
2631 /*
2632  * return expired entry, or NULL to just start from scratch in rbtree
2633  */
2634 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2635 {
2636         struct request *rq = NULL;
2637
2638         if (cfq_cfqq_fifo_expire(cfqq))
2639                 return NULL;
2640
2641         cfq_mark_cfqq_fifo_expire(cfqq);
2642
2643         if (list_empty(&cfqq->fifo))
2644                 return NULL;
2645
2646         rq = rq_entry_fifo(cfqq->fifo.next);
2647         if (time_before(jiffies, rq_fifo_time(rq)))
2648                 rq = NULL;
2649
2650         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2651         return rq;
2652 }
2653
2654 static inline int
2655 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2656 {
2657         const int base_rq = cfqd->cfq_slice_async_rq;
2658
2659         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2660
2661         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2662 }
2663
2664 /*
2665  * Must be called with the queue_lock held.
2666  */
2667 static int cfqq_process_refs(struct cfq_queue *cfqq)
2668 {
2669         int process_refs, io_refs;
2670
2671         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2672         process_refs = cfqq->ref - io_refs;
2673         BUG_ON(process_refs < 0);
2674         return process_refs;
2675 }
2676
2677 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2678 {
2679         int process_refs, new_process_refs;
2680         struct cfq_queue *__cfqq;
2681
2682         /*
2683          * If there are no process references on the new_cfqq, then it is
2684          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2685          * chain may have dropped their last reference (not just their
2686          * last process reference).
2687          */
2688         if (!cfqq_process_refs(new_cfqq))
2689                 return;
2690
2691         /* Avoid a circular list and skip interim queue merges */
2692         while ((__cfqq = new_cfqq->new_cfqq)) {
2693                 if (__cfqq == cfqq)
2694                         return;
2695                 new_cfqq = __cfqq;
2696         }
2697
2698         process_refs = cfqq_process_refs(cfqq);
2699         new_process_refs = cfqq_process_refs(new_cfqq);
2700         /*
2701          * If the process for the cfqq has gone away, there is no
2702          * sense in merging the queues.
2703          */
2704         if (process_refs == 0 || new_process_refs == 0)
2705                 return;
2706
2707         /*
2708          * Merge in the direction of the lesser amount of work.
2709          */
2710         if (new_process_refs >= process_refs) {
2711                 cfqq->new_cfqq = new_cfqq;
2712                 new_cfqq->ref += process_refs;
2713         } else {
2714                 new_cfqq->new_cfqq = cfqq;
2715                 cfqq->ref += new_process_refs;
2716         }
2717 }
2718
2719 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
2720                         struct cfq_group *cfqg, enum wl_class_t wl_class)
2721 {
2722         struct cfq_queue *queue;
2723         int i;
2724         bool key_valid = false;
2725         unsigned long lowest_key = 0;
2726         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2727
2728         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2729                 /* select the one with lowest rb_key */
2730                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
2731                 if (queue &&
2732                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2733                         lowest_key = queue->rb_key;
2734                         cur_best = i;
2735                         key_valid = true;
2736                 }
2737         }
2738
2739         return cur_best;
2740 }
2741
2742 static void
2743 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
2744 {
2745         unsigned slice;
2746         unsigned count;
2747         struct cfq_rb_root *st;
2748         unsigned group_slice;
2749         enum wl_class_t original_class = cfqd->serving_wl_class;
2750
2751         /* Choose next priority. RT > BE > IDLE */
2752         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2753                 cfqd->serving_wl_class = RT_WORKLOAD;
2754         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2755                 cfqd->serving_wl_class = BE_WORKLOAD;
2756         else {
2757                 cfqd->serving_wl_class = IDLE_WORKLOAD;
2758                 cfqd->workload_expires = jiffies + 1;
2759                 return;
2760         }
2761
2762         if (original_class != cfqd->serving_wl_class)
2763                 goto new_workload;
2764
2765         /*
2766          * For RT and BE, we have to choose also the type
2767          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2768          * expiration time
2769          */
2770         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2771         count = st->count;
2772
2773         /*
2774          * check workload expiration, and that we still have other queues ready
2775          */
2776         if (count && !time_after(jiffies, cfqd->workload_expires))
2777                 return;
2778
2779 new_workload:
2780         /* otherwise select new workload type */
2781         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
2782                                         cfqd->serving_wl_class);
2783         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
2784         count = st->count;
2785
2786         /*
2787          * the workload slice is computed as a fraction of target latency
2788          * proportional to the number of queues in that workload, over
2789          * all the queues in the same priority class
2790          */
2791         group_slice = cfq_group_slice(cfqd, cfqg);
2792
2793         slice = group_slice * count /
2794                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
2795                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
2796                                         cfqg));
2797
2798         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
2799                 unsigned int tmp;
2800
2801                 /*
2802                  * Async queues are currently system wide. Just taking
2803                  * proportion of queues with-in same group will lead to higher
2804                  * async ratio system wide as generally root group is going
2805                  * to have higher weight. A more accurate thing would be to
2806                  * calculate system wide asnc/sync ratio.
2807                  */
2808                 tmp = cfqd->cfq_target_latency *
2809                         cfqg_busy_async_queues(cfqd, cfqg);
2810                 tmp = tmp/cfqd->busy_queues;
2811                 slice = min_t(unsigned, slice, tmp);
2812
2813                 /* async workload slice is scaled down according to
2814                  * the sync/async slice ratio. */
2815                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2816         } else
2817                 /* sync workload slice is at least 2 * cfq_slice_idle */
2818                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2819
2820         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2821         cfq_log(cfqd, "workload slice:%d", slice);
2822         cfqd->workload_expires = jiffies + slice;
2823 }
2824
2825 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2826 {
2827         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2828         struct cfq_group *cfqg;
2829
2830         if (RB_EMPTY_ROOT(&st->rb))
2831                 return NULL;
2832         cfqg = cfq_rb_first_group(st);
2833         update_min_vdisktime(st);
2834         return cfqg;
2835 }
2836
2837 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2838 {
2839         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2840
2841         cfqd->serving_group = cfqg;
2842
2843         /* Restore the workload type data */
2844         if (cfqg->saved_wl_slice) {
2845                 cfqd->workload_expires = jiffies + cfqg->saved_wl_slice;
2846                 cfqd->serving_wl_type = cfqg->saved_wl_type;
2847                 cfqd->serving_wl_class = cfqg->saved_wl_class;
2848         } else
2849                 cfqd->workload_expires = jiffies - 1;
2850
2851         choose_wl_class_and_type(cfqd, cfqg);
2852 }
2853
2854 /*
2855  * Select a queue for service. If we have a current active queue,
2856  * check whether to continue servicing it, or retrieve and set a new one.
2857  */
2858 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2859 {
2860         struct cfq_queue *cfqq, *new_cfqq = NULL;
2861
2862         cfqq = cfqd->active_queue;
2863         if (!cfqq)
2864                 goto new_queue;
2865
2866         if (!cfqd->rq_queued)
2867                 return NULL;
2868
2869         /*
2870          * We were waiting for group to get backlogged. Expire the queue
2871          */
2872         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2873                 goto expire;
2874
2875         /*
2876          * The active queue has run out of time, expire it and select new.
2877          */
2878         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2879                 /*
2880                  * If slice had not expired at the completion of last request
2881                  * we might not have turned on wait_busy flag. Don't expire
2882                  * the queue yet. Allow the group to get backlogged.
2883                  *
2884                  * The very fact that we have used the slice, that means we
2885                  * have been idling all along on this queue and it should be
2886                  * ok to wait for this request to complete.
2887                  */
2888                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2889                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2890                         cfqq = NULL;
2891                         goto keep_queue;
2892                 } else
2893                         goto check_group_idle;
2894         }
2895
2896         /*
2897          * The active queue has requests and isn't expired, allow it to
2898          * dispatch.
2899          */
2900         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2901                 goto keep_queue;
2902
2903         /*
2904          * If another queue has a request waiting within our mean seek
2905          * distance, let it run.  The expire code will check for close
2906          * cooperators and put the close queue at the front of the service
2907          * tree.  If possible, merge the expiring queue with the new cfqq.
2908          */
2909         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2910         if (new_cfqq) {
2911                 if (!cfqq->new_cfqq)
2912                         cfq_setup_merge(cfqq, new_cfqq);
2913                 goto expire;
2914         }
2915
2916         /*
2917          * No requests pending. If the active queue still has requests in
2918          * flight or is idling for a new request, allow either of these
2919          * conditions to happen (or time out) before selecting a new queue.
2920          */
2921         if (timer_pending(&cfqd->idle_slice_timer)) {
2922                 cfqq = NULL;
2923                 goto keep_queue;
2924         }
2925
2926         /*
2927          * This is a deep seek queue, but the device is much faster than
2928          * the queue can deliver, don't idle
2929          **/
2930         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2931             (cfq_cfqq_slice_new(cfqq) ||
2932             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2933                 cfq_clear_cfqq_deep(cfqq);
2934                 cfq_clear_cfqq_idle_window(cfqq);
2935         }
2936
2937         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2938                 cfqq = NULL;
2939                 goto keep_queue;
2940         }
2941
2942         /*
2943          * If group idle is enabled and there are requests dispatched from
2944          * this group, wait for requests to complete.
2945          */
2946 check_group_idle:
2947         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2948             cfqq->cfqg->dispatched &&
2949             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2950                 cfqq = NULL;
2951                 goto keep_queue;
2952         }
2953
2954 expire:
2955         cfq_slice_expired(cfqd, 0);
2956 new_queue:
2957         /*
2958          * Current queue expired. Check if we have to switch to a new
2959          * service tree
2960          */
2961         if (!new_cfqq)
2962                 cfq_choose_cfqg(cfqd);
2963
2964         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2965 keep_queue:
2966         return cfqq;
2967 }
2968
2969 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2970 {
2971         int dispatched = 0;
2972
2973         while (cfqq->next_rq) {
2974                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2975                 dispatched++;
2976         }
2977
2978         BUG_ON(!list_empty(&cfqq->fifo));
2979
2980         /* By default cfqq is not expired if it is empty. Do it explicitly */
2981         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2982         return dispatched;
2983 }
2984
2985 /*
2986  * Drain our current requests. Used for barriers and when switching
2987  * io schedulers on-the-fly.
2988  */
2989 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2990 {
2991         struct cfq_queue *cfqq;
2992         int dispatched = 0;
2993
2994         /* Expire the timeslice of the current active queue first */
2995         cfq_slice_expired(cfqd, 0);
2996         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2997                 __cfq_set_active_queue(cfqd, cfqq);
2998                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2999         }
3000
3001         BUG_ON(cfqd->busy_queues);
3002
3003         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
3004         return dispatched;
3005 }
3006
3007 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
3008         struct cfq_queue *cfqq)
3009 {
3010         /* the queue hasn't finished any request, can't estimate */
3011         if (cfq_cfqq_slice_new(cfqq))
3012                 return true;
3013         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
3014                 cfqq->slice_end))
3015                 return true;
3016
3017         return false;
3018 }
3019
3020 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3021 {
3022         unsigned int max_dispatch;
3023
3024         /*
3025          * Drain async requests before we start sync IO
3026          */
3027         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
3028                 return false;
3029
3030         /*
3031          * If this is an async queue and we have sync IO in flight, let it wait
3032          */
3033         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
3034                 return false;
3035
3036         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
3037         if (cfq_class_idle(cfqq))
3038                 max_dispatch = 1;
3039
3040         /*
3041          * Does this cfqq already have too much IO in flight?
3042          */
3043         if (cfqq->dispatched >= max_dispatch) {
3044                 bool promote_sync = false;
3045                 /*
3046                  * idle queue must always only have a single IO in flight
3047                  */
3048                 if (cfq_class_idle(cfqq))
3049                         return false;
3050
3051                 /*
3052                  * If there is only one sync queue
3053                  * we can ignore async queue here and give the sync
3054                  * queue no dispatch limit. The reason is a sync queue can
3055                  * preempt async queue, limiting the sync queue doesn't make
3056                  * sense. This is useful for aiostress test.
3057                  */
3058                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
3059                         promote_sync = true;
3060
3061                 /*
3062                  * We have other queues, don't allow more IO from this one
3063                  */
3064                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
3065                                 !promote_sync)
3066                         return false;
3067
3068                 /*
3069                  * Sole queue user, no limit
3070                  */
3071                 if (cfqd->busy_queues == 1 || promote_sync)
3072                         max_dispatch = -1;
3073                 else
3074                         /*
3075                          * Normally we start throttling cfqq when cfq_quantum/2
3076                          * requests have been dispatched. But we can drive
3077                          * deeper queue depths at the beginning of slice
3078                          * subjected to upper limit of cfq_quantum.
3079                          * */
3080                         max_dispatch = cfqd->cfq_quantum;
3081         }
3082
3083         /*
3084          * Async queues must wait a bit before being allowed dispatch.
3085          * We also ramp up the dispatch depth gradually for async IO,
3086          * based on the last sync IO we serviced
3087          */
3088         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
3089                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
3090                 unsigned int depth;
3091
3092                 depth = last_sync / cfqd->cfq_slice[1];
3093                 if (!depth && !cfqq->dispatched)
3094                         depth = 1;
3095                 if (depth < max_dispatch)
3096                         max_dispatch = depth;
3097         }
3098
3099         /*
3100          * If we're below the current max, allow a dispatch
3101          */
3102         return cfqq->dispatched < max_dispatch;
3103 }
3104
3105 /*
3106  * Dispatch a request from cfqq, moving them to the request queue
3107  * dispatch list.
3108  */
3109 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3110 {
3111         struct request *rq;
3112
3113         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
3114
3115         if (!cfq_may_dispatch(cfqd, cfqq))
3116                 return false;
3117
3118         /*
3119          * follow expired path, else get first next available
3120          */
3121         rq = cfq_check_fifo(cfqq);
3122         if (!rq)
3123                 rq = cfqq->next_rq;
3124
3125         /*
3126          * insert request into driver dispatch list
3127          */
3128         cfq_dispatch_insert(cfqd->queue, rq);
3129
3130         if (!cfqd->active_cic) {
3131                 struct cfq_io_cq *cic = RQ_CIC(rq);
3132
3133                 atomic_long_inc(&cic->icq.ioc->refcount);
3134                 cfqd->active_cic = cic;
3135         }
3136
3137         return true;
3138 }
3139
3140 /*
3141  * Find the cfqq that we need to service and move a request from that to the
3142  * dispatch list
3143  */
3144 static int cfq_dispatch_requests(struct request_queue *q, int force)
3145 {
3146         struct cfq_data *cfqd = q->elevator->elevator_data;
3147         struct cfq_queue *cfqq;
3148
3149         if (!cfqd->busy_queues)
3150                 return 0;
3151
3152         if (unlikely(force))
3153                 return cfq_forced_dispatch(cfqd);
3154
3155         cfqq = cfq_select_queue(cfqd);
3156         if (!cfqq)
3157                 return 0;
3158
3159         /*
3160          * Dispatch a request from this cfqq, if it is allowed
3161          */
3162         if (!cfq_dispatch_request(cfqd, cfqq))
3163                 return 0;
3164
3165         cfqq->slice_dispatch++;
3166         cfq_clear_cfqq_must_dispatch(cfqq);
3167
3168         /*
3169          * expire an async queue immediately if it has used up its slice. idle
3170          * queue always expire after 1 dispatch round.
3171          */
3172         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
3173             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
3174             cfq_class_idle(cfqq))) {
3175                 cfqq->slice_end = jiffies + 1;
3176                 cfq_slice_expired(cfqd, 0);
3177         }
3178
3179         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
3180         return 1;
3181 }
3182
3183 /*
3184  * task holds one reference to the queue, dropped when task exits. each rq
3185  * in-flight on this queue also holds a reference, dropped when rq is freed.
3186  *
3187  * Each cfq queue took a reference on the parent group. Drop it now.
3188  * queue lock must be held here.
3189  */
3190 static void cfq_put_queue(struct cfq_queue *cfqq)
3191 {
3192         struct cfq_data *cfqd = cfqq->cfqd;
3193         struct cfq_group *cfqg;
3194
3195         BUG_ON(cfqq->ref <= 0);
3196
3197         cfqq->ref--;
3198         if (cfqq->ref)
3199                 return;
3200
3201         cfq_log_cfqq(cfqd, cfqq, "put_queue");
3202         BUG_ON(rb_first(&cfqq->sort_list));
3203         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
3204         cfqg = cfqq->cfqg;
3205
3206         if (unlikely(cfqd->active_queue == cfqq)) {
3207                 __cfq_slice_expired(cfqd, cfqq, 0);
3208                 cfq_schedule_dispatch(cfqd);
3209         }
3210
3211         BUG_ON(cfq_cfqq_on_rr(cfqq));
3212         kmem_cache_free(cfq_pool, cfqq);
3213         cfqg_put(cfqg);
3214 }
3215
3216 static void cfq_put_cooperator(struct cfq_queue *cfqq)
3217 {
3218         struct cfq_queue *__cfqq, *next;
3219
3220         /*
3221          * If this queue was scheduled to merge with another queue, be
3222          * sure to drop the reference taken on that queue (and others in
3223          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
3224          */
3225         __cfqq = cfqq->new_cfqq;
3226         while (__cfqq) {
3227                 if (__cfqq == cfqq) {
3228                         WARN(1, "cfqq->new_cfqq loop detected\n");
3229                         break;
3230                 }
3231                 next = __cfqq->new_cfqq;
3232                 cfq_put_queue(__cfqq);
3233                 __cfqq = next;
3234         }
3235 }
3236
3237 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3238 {
3239         if (unlikely(cfqq == cfqd->active_queue)) {
3240                 __cfq_slice_expired(cfqd, cfqq, 0);
3241                 cfq_schedule_dispatch(cfqd);
3242         }
3243
3244         cfq_put_cooperator(cfqq);
3245
3246         cfq_put_queue(cfqq);
3247 }
3248
3249 static void cfq_init_icq(struct io_cq *icq)
3250 {
3251         struct cfq_io_cq *cic = icq_to_cic(icq);
3252
3253         cic->ttime.last_end_request = jiffies;
3254 }
3255
3256 static void cfq_exit_icq(struct io_cq *icq)
3257 {
3258         struct cfq_io_cq *cic = icq_to_cic(icq);
3259         struct cfq_data *cfqd = cic_to_cfqd(cic);
3260
3261         if (cic->cfqq[BLK_RW_ASYNC]) {
3262                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
3263                 cic->cfqq[BLK_RW_ASYNC] = NULL;
3264         }
3265
3266         if (cic->cfqq[BLK_RW_SYNC]) {
3267                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
3268                 cic->cfqq[BLK_RW_SYNC] = NULL;
3269         }
3270 }
3271
3272 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct cfq_io_cq *cic)
3273 {
3274         struct task_struct *tsk = current;
3275         int ioprio_class;
3276
3277         if (!cfq_cfqq_prio_changed(cfqq))
3278                 return;
3279
3280         ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3281         switch (ioprio_class) {
3282         default:
3283                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
3284         case IOPRIO_CLASS_NONE:
3285                 /*
3286                  * no prio set, inherit CPU scheduling settings
3287                  */
3288                 cfqq->ioprio = task_nice_ioprio(tsk);
3289                 cfqq->ioprio_class = task_nice_ioclass(tsk);
3290                 break;
3291         case IOPRIO_CLASS_RT:
3292                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3293                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
3294                 break;
3295         case IOPRIO_CLASS_BE:
3296                 cfqq->ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3297                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
3298                 break;
3299         case IOPRIO_CLASS_IDLE:
3300                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
3301                 cfqq->ioprio = 7;
3302                 cfq_clear_cfqq_idle_window(cfqq);
3303                 break;
3304         }
3305
3306         /*
3307          * keep track of original prio settings in case we have to temporarily
3308          * elevate the priority of this queue
3309          */
3310         cfqq->org_ioprio = cfqq->ioprio;
3311         cfq_clear_cfqq_prio_changed(cfqq);
3312 }
3313
3314 static void check_ioprio_changed(struct cfq_io_cq *cic, struct bio *bio)
3315 {
3316         int ioprio = cic->icq.ioc->ioprio;
3317         struct cfq_data *cfqd = cic_to_cfqd(cic);
3318         struct cfq_queue *cfqq;
3319
3320         /*
3321          * Check whether ioprio has changed.  The condition may trigger
3322          * spuriously on a newly created cic but there's no harm.
3323          */
3324         if (unlikely(!cfqd) || likely(cic->ioprio == ioprio))
3325                 return;
3326
3327         cfqq = cic->cfqq[BLK_RW_ASYNC];
3328         if (cfqq) {
3329                 struct cfq_queue *new_cfqq;
3330                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic, bio,
3331                                          GFP_ATOMIC);
3332                 if (new_cfqq) {
3333                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
3334                         cfq_put_queue(cfqq);
3335                 }
3336         }
3337
3338         cfqq = cic->cfqq[BLK_RW_SYNC];
3339         if (cfqq)
3340                 cfq_mark_cfqq_prio_changed(cfqq);
3341
3342         cic->ioprio = ioprio;
3343 }
3344
3345 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3346                           pid_t pid, bool is_sync)
3347 {
3348         RB_CLEAR_NODE(&cfqq->rb_node);
3349         RB_CLEAR_NODE(&cfqq->p_node);
3350         INIT_LIST_HEAD(&cfqq->fifo);
3351
3352         cfqq->ref = 0;
3353         cfqq->cfqd = cfqd;
3354
3355         cfq_mark_cfqq_prio_changed(cfqq);
3356
3357         if (is_sync) {
3358                 if (!cfq_class_idle(cfqq))
3359                         cfq_mark_cfqq_idle_window(cfqq);
3360                 cfq_mark_cfqq_sync(cfqq);
3361         }
3362         cfqq->pid = pid;
3363 }
3364
3365 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3366 static void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio)
3367 {
3368         struct cfq_data *cfqd = cic_to_cfqd(cic);
3369         struct cfq_queue *sync_cfqq;
3370         uint64_t id;
3371
3372         rcu_read_lock();
3373         id = bio_blkcg(bio)->id;
3374         rcu_read_unlock();
3375
3376         /*
3377          * Check whether blkcg has changed.  The condition may trigger
3378          * spuriously on a newly created cic but there's no harm.
3379          */
3380         if (unlikely(!cfqd) || likely(cic->blkcg_id == id))
3381                 return;
3382
3383         sync_cfqq = cic_to_cfqq(cic, 1);
3384         if (sync_cfqq) {
3385                 /*
3386                  * Drop reference to sync queue. A new sync queue will be
3387                  * assigned in new group upon arrival of a fresh request.
3388                  */
3389                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
3390                 cic_set_cfqq(cic, NULL, 1);
3391                 cfq_put_queue(sync_cfqq);
3392         }
3393
3394         cic->blkcg_id = id;
3395 }
3396 #else
3397 static inline void check_blkcg_changed(struct cfq_io_cq *cic, struct bio *bio) { }
3398 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
3399
3400 static struct cfq_queue *
3401 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3402                      struct bio *bio, gfp_t gfp_mask)
3403 {
3404         struct blkcg *blkcg;
3405         struct cfq_queue *cfqq, *new_cfqq = NULL;
3406         struct cfq_group *cfqg;
3407
3408 retry:
3409         rcu_read_lock();
3410
3411         blkcg = bio_blkcg(bio);
3412         cfqg = cfq_lookup_create_cfqg(cfqd, blkcg);
3413         cfqq = cic_to_cfqq(cic, is_sync);
3414
3415         /*
3416          * Always try a new alloc if we fell back to the OOM cfqq
3417          * originally, since it should just be a temporary situation.
3418          */
3419         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3420                 cfqq = NULL;
3421                 if (new_cfqq) {
3422                         cfqq = new_cfqq;
3423                         new_cfqq = NULL;
3424                 } else if (gfp_mask & __GFP_WAIT) {
3425                         rcu_read_unlock();
3426                         spin_unlock_irq(cfqd->queue->queue_lock);
3427                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3428                                         gfp_mask | __GFP_ZERO,
3429                                         cfqd->queue->node);
3430                         spin_lock_irq(cfqd->queue->queue_lock);
3431                         if (new_cfqq)
3432                                 goto retry;
3433                 } else {
3434                         cfqq = kmem_cache_alloc_node(cfq_pool,
3435                                         gfp_mask | __GFP_ZERO,
3436                                         cfqd->queue->node);
3437                 }
3438
3439                 if (cfqq) {
3440                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3441                         cfq_init_prio_data(cfqq, cic);
3442                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3443                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3444                 } else
3445                         cfqq = &cfqd->oom_cfqq;
3446         }
3447
3448         if (new_cfqq)
3449                 kmem_cache_free(cfq_pool, new_cfqq);
3450
3451         rcu_read_unlock();
3452         return cfqq;
3453 }
3454
3455 static struct cfq_queue **
3456 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3457 {
3458         switch (ioprio_class) {
3459         case IOPRIO_CLASS_RT:
3460                 return &cfqd->async_cfqq[0][ioprio];
3461         case IOPRIO_CLASS_NONE:
3462                 ioprio = IOPRIO_NORM;
3463                 /* fall through */
3464         case IOPRIO_CLASS_BE:
3465                 return &cfqd->async_cfqq[1][ioprio];
3466         case IOPRIO_CLASS_IDLE:
3467                 return &cfqd->async_idle_cfqq;
3468         default:
3469                 BUG();
3470         }
3471 }
3472
3473 static struct cfq_queue *
3474 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct cfq_io_cq *cic,
3475               struct bio *bio, gfp_t gfp_mask)
3476 {
3477         const int ioprio_class = IOPRIO_PRIO_CLASS(cic->ioprio);
3478         const int ioprio = IOPRIO_PRIO_DATA(cic->ioprio);
3479         struct cfq_queue **async_cfqq = NULL;
3480         struct cfq_queue *cfqq = NULL;
3481
3482         if (!is_sync) {
3483                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3484                 cfqq = *async_cfqq;
3485         }
3486
3487         if (!cfqq)
3488                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, cic, bio, gfp_mask);
3489
3490         /*
3491          * pin the queue now that it's allocated, scheduler exit will prune it
3492          */
3493         if (!is_sync && !(*async_cfqq)) {
3494                 cfqq->ref++;
3495                 *async_cfqq = cfqq;
3496         }
3497
3498         cfqq->ref++;
3499         return cfqq;
3500 }
3501
3502 static void
3503 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3504 {
3505         unsigned long elapsed = jiffies - ttime->last_end_request;
3506         elapsed = min(elapsed, 2UL * slice_idle);
3507
3508         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3509         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3510         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3511 }
3512
3513 static void
3514 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3515                         struct cfq_io_cq *cic)
3516 {
3517         if (cfq_cfqq_sync(cfqq)) {
3518                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3519                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3520                         cfqd->cfq_slice_idle);
3521         }
3522 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3523         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3524 #endif
3525 }
3526
3527 static void
3528 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3529                        struct request *rq)
3530 {
3531         sector_t sdist = 0;
3532         sector_t n_sec = blk_rq_sectors(rq);
3533         if (cfqq->last_request_pos) {
3534                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3535                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3536                 else
3537                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3538         }
3539
3540         cfqq->seek_history <<= 1;
3541         if (blk_queue_nonrot(cfqd->queue))
3542                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3543         else
3544                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3545 }
3546
3547 /*
3548  * Disable idle window if the process thinks too long or seeks so much that
3549  * it doesn't matter
3550  */
3551 static void
3552 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3553                        struct cfq_io_cq *cic)
3554 {
3555         int old_idle, enable_idle;
3556
3557         /*
3558          * Don't idle for async or idle io prio class
3559          */
3560         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3561                 return;
3562
3563         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3564
3565         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3566                 cfq_mark_cfqq_deep(cfqq);
3567
3568         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3569                 enable_idle = 0;
3570         else if (!atomic_read(&cic->icq.ioc->active_ref) ||
3571                  !cfqd->cfq_slice_idle ||
3572                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3573                 enable_idle = 0;
3574         else if (sample_valid(cic->ttime.ttime_samples)) {
3575                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3576                         enable_idle = 0;
3577                 else
3578                         enable_idle = 1;
3579         }
3580
3581         if (old_idle != enable_idle) {
3582                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3583                 if (enable_idle)
3584                         cfq_mark_cfqq_idle_window(cfqq);
3585                 else
3586                         cfq_clear_cfqq_idle_window(cfqq);
3587         }
3588 }
3589
3590 /*
3591  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3592  * no or if we aren't sure, a 1 will cause a preempt.
3593  */
3594 static bool
3595 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3596                    struct request *rq)
3597 {
3598         struct cfq_queue *cfqq;
3599
3600         cfqq = cfqd->active_queue;
3601         if (!cfqq)
3602                 return false;
3603
3604         if (cfq_class_idle(new_cfqq))
3605                 return false;
3606
3607         if (cfq_class_idle(cfqq))
3608                 return true;
3609
3610         /*
3611          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3612          */
3613         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3614                 return false;
3615
3616         /*
3617          * if the new request is sync, but the currently running queue is
3618          * not, let the sync request have priority.
3619          */
3620         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3621                 return true;
3622
3623         if (new_cfqq->cfqg != cfqq->cfqg)
3624                 return false;
3625
3626         if (cfq_slice_used(cfqq))
3627                 return true;
3628
3629         /* Allow preemption only if we are idling on sync-noidle tree */
3630         if (cfqd->serving_wl_type == SYNC_NOIDLE_WORKLOAD &&
3631             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3632             new_cfqq->service_tree->count == 2 &&
3633             RB_EMPTY_ROOT(&cfqq->sort_list))
3634                 return true;
3635
3636         /*
3637          * So both queues are sync. Let the new request get disk time if
3638          * it's a metadata request and the current queue is doing regular IO.
3639          */
3640         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3641                 return true;
3642
3643         /*
3644          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3645          */
3646         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3647                 return true;
3648
3649         /* An idle queue should not be idle now for some reason */
3650         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3651                 return true;
3652
3653         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3654                 return false;
3655
3656         /*
3657          * if this request is as-good as one we would expect from the
3658          * current cfqq, let it preempt
3659          */
3660         if (cfq_rq_close(cfqd, cfqq, rq))
3661                 return true;
3662
3663         return false;
3664 }
3665
3666 /*
3667  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3668  * let it have half of its nominal slice.
3669  */
3670 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3671 {
3672         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3673
3674         cfq_log_cfqq(cfqd, cfqq, "preempt");
3675         cfq_slice_expired(cfqd, 1);
3676
3677         /*
3678          * workload type is changed, don't save slice, otherwise preempt
3679          * doesn't happen
3680          */
3681         if (old_type != cfqq_type(cfqq))
3682                 cfqq->cfqg->saved_wl_slice = 0;
3683
3684         /*
3685          * Put the new queue at the front of the of the current list,
3686          * so we know that it will be selected next.
3687          */
3688         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3689
3690         cfq_service_tree_add(cfqd, cfqq, 1);
3691
3692         cfqq->slice_end = 0;
3693         cfq_mark_cfqq_slice_new(cfqq);
3694 }
3695
3696 /*
3697  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3698  * something we should do about it
3699  */
3700 static void
3701 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3702                 struct request *rq)
3703 {
3704         struct cfq_io_cq *cic = RQ_CIC(rq);
3705
3706         cfqd->rq_queued++;
3707         if (rq->cmd_flags & REQ_PRIO)
3708                 cfqq->prio_pending++;
3709
3710         cfq_update_io_thinktime(cfqd, cfqq, cic);
3711         cfq_update_io_seektime(cfqd, cfqq, rq);
3712         cfq_update_idle_window(cfqd, cfqq, cic);
3713
3714         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3715
3716         if (cfqq == cfqd->active_queue) {
3717                 /*
3718                  * Remember that we saw a request from this process, but
3719                  * don't start queuing just yet. Otherwise we risk seeing lots
3720                  * of tiny requests, because we disrupt the normal plugging
3721                  * and merging. If the request is already larger than a single
3722                  * page, let it rip immediately. For that case we assume that
3723                  * merging is already done. Ditto for a busy system that
3724                  * has other work pending, don't risk delaying until the
3725                  * idle timer unplug to continue working.
3726                  */
3727                 if (cfq_cfqq_wait_request(cfqq)) {
3728                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3729                             cfqd->busy_queues > 1) {
3730                                 cfq_del_timer(cfqd, cfqq);
3731                                 cfq_clear_cfqq_wait_request(cfqq);
3732                                 __blk_run_queue(cfqd->queue);
3733                         } else {
3734                                 cfqg_stats_update_idle_time(cfqq->cfqg);
3735                                 cfq_mark_cfqq_must_dispatch(cfqq);
3736                         }
3737                 }
3738         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3739                 /*
3740                  * not the active queue - expire current slice if it is
3741                  * idle and has expired it's mean thinktime or this new queue
3742                  * has some old slice time left and is of higher priority or
3743                  * this new queue is RT and the current one is BE
3744                  */
3745                 cfq_preempt_queue(cfqd, cfqq);
3746                 __blk_run_queue(cfqd->queue);
3747         }
3748 }
3749
3750 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3751 {
3752         struct cfq_data *cfqd = q->elevator->elevator_data;
3753         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3754
3755         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3756         cfq_init_prio_data(cfqq, RQ_CIC(rq));
3757
3758         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3759         list_add_tail(&rq->queuelist, &cfqq->fifo);
3760         cfq_add_rq_rb(rq);
3761         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqd->serving_group,
3762                                  rq->cmd_flags);
3763         cfq_rq_enqueued(cfqd, cfqq, rq);
3764 }
3765
3766 /*
3767  * Update hw_tag based on peak queue depth over 50 samples under
3768  * sufficient load.
3769  */
3770 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3771 {
3772         struct cfq_queue *cfqq = cfqd->active_queue;
3773
3774         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3775                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3776
3777         if (cfqd->hw_tag == 1)
3778                 return;
3779
3780         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3781             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3782                 return;
3783
3784         /*
3785          * If active queue hasn't enough requests and can idle, cfq might not
3786          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3787          * case
3788          */
3789         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3790             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3791             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3792                 return;
3793
3794         if (cfqd->hw_tag_samples++ < 50)
3795                 return;
3796
3797         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3798                 cfqd->hw_tag = 1;
3799         else
3800                 cfqd->hw_tag = 0;
3801 }
3802
3803 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3804 {
3805         struct cfq_io_cq *cic = cfqd->active_cic;
3806
3807         /* If the queue already has requests, don't wait */
3808         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3809                 return false;
3810
3811         /* If there are other queues in the group, don't wait */
3812         if (cfqq->cfqg->nr_cfqq > 1)
3813                 return false;
3814
3815         /* the only queue in the group, but think time is big */
3816         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3817                 return false;
3818
3819         if (cfq_slice_used(cfqq))
3820                 return true;
3821
3822         /* if slice left is less than think time, wait busy */
3823         if (cic && sample_valid(cic->ttime.ttime_samples)
3824             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3825                 return true;
3826
3827         /*
3828          * If think times is less than a jiffy than ttime_mean=0 and above
3829          * will not be true. It might happen that slice has not expired yet
3830          * but will expire soon (4-5 ns) during select_queue(). To cover the
3831          * case where think time is less than a jiffy, mark the queue wait
3832          * busy if only 1 jiffy is left in the slice.
3833          */
3834         if (cfqq->slice_end - jiffies == 1)
3835                 return true;
3836
3837         return false;
3838 }
3839
3840 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3841 {
3842         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3843         struct cfq_data *cfqd = cfqq->cfqd;
3844         const int sync = rq_is_sync(rq);
3845         unsigned long now;
3846
3847         now = jiffies;
3848         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3849                      !!(rq->cmd_flags & REQ_NOIDLE));
3850
3851         cfq_update_hw_tag(cfqd);
3852
3853         WARN_ON(!cfqd->rq_in_driver);
3854         WARN_ON(!cfqq->dispatched);
3855         cfqd->rq_in_driver--;
3856         cfqq->dispatched--;
3857         (RQ_CFQG(rq))->dispatched--;
3858         cfqg_stats_update_completion(cfqq->cfqg, rq_start_time_ns(rq),
3859                                      rq_io_start_time_ns(rq), rq->cmd_flags);
3860
3861         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3862
3863         if (sync) {
3864                 struct cfq_rb_root *st;
3865
3866                 RQ_CIC(rq)->ttime.last_end_request = now;
3867
3868                 if (cfq_cfqq_on_rr(cfqq))
3869                         st = cfqq->service_tree;
3870                 else
3871                         st = st_for(cfqq->cfqg, cfqq_class(cfqq),
3872                                         cfqq_type(cfqq));
3873
3874                 st->ttime.last_end_request = now;
3875                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3876                         cfqd->last_delayed_sync = now;
3877         }
3878
3879 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3880         cfqq->cfqg->ttime.last_end_request = now;
3881 #endif
3882
3883         /*
3884          * If this is the active queue, check if it needs to be expired,
3885          * or if we want to idle in case it has no pending requests.
3886          */
3887         if (cfqd->active_queue == cfqq) {
3888                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3889
3890                 if (cfq_cfqq_slice_new(cfqq)) {
3891                         cfq_set_prio_slice(cfqd, cfqq);
3892                         cfq_clear_cfqq_slice_new(cfqq);
3893                 }
3894
3895                 /*
3896                  * Should we wait for next request to come in before we expire
3897                  * the queue.
3898                  */
3899                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3900                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3901                         if (!cfqd->cfq_slice_idle)
3902                                 extend_sl = cfqd->cfq_group_idle;
3903                         cfqq->slice_end = jiffies + extend_sl;
3904                         cfq_mark_cfqq_wait_busy(cfqq);
3905                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3906                 }
3907
3908                 /*
3909                  * Idling is not enabled on:
3910                  * - expired queues
3911                  * - idle-priority queues
3912                  * - async queues
3913                  * - queues with still some requests queued
3914                  * - when there is a close cooperator
3915                  */
3916                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3917                         cfq_slice_expired(cfqd, 1);
3918                 else if (sync && cfqq_empty &&
3919                          !cfq_close_cooperator(cfqd, cfqq)) {
3920                         cfq_arm_slice_timer(cfqd);
3921                 }
3922         }
3923
3924         if (!cfqd->rq_in_driver)
3925                 cfq_schedule_dispatch(cfqd);
3926 }
3927
3928 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3929 {
3930         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3931                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3932                 return ELV_MQUEUE_MUST;
3933         }
3934
3935         return ELV_MQUEUE_MAY;
3936 }
3937
3938 static int cfq_may_queue(struct request_queue *q, int rw)
3939 {
3940         struct cfq_data *cfqd = q->elevator->elevator_data;
3941         struct task_struct *tsk = current;
3942         struct cfq_io_cq *cic;
3943         struct cfq_queue *cfqq;
3944
3945         /*
3946          * don't force setup of a queue from here, as a call to may_queue
3947          * does not necessarily imply that a request actually will be queued.
3948          * so just lookup a possibly existing queue, or return 'may queue'
3949          * if that fails
3950          */
3951         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3952         if (!cic)
3953                 return ELV_MQUEUE_MAY;
3954
3955         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3956         if (cfqq) {
3957                 cfq_init_prio_data(cfqq, cic);
3958
3959                 return __cfq_may_queue(cfqq);
3960         }
3961
3962         return ELV_MQUEUE_MAY;
3963 }
3964
3965 /*
3966  * queue lock held here
3967  */
3968 static void cfq_put_request(struct request *rq)
3969 {
3970         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3971
3972         if (cfqq) {
3973                 const int rw = rq_data_dir(rq);
3974
3975                 BUG_ON(!cfqq->allocated[rw]);
3976                 cfqq->allocated[rw]--;
3977
3978                 /* Put down rq reference on cfqg */
3979                 cfqg_put(RQ_CFQG(rq));
3980                 rq->elv.priv[0] = NULL;
3981                 rq->elv.priv[1] = NULL;
3982
3983                 cfq_put_queue(cfqq);
3984         }
3985 }
3986
3987 static struct cfq_queue *
3988 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3989                 struct cfq_queue *cfqq)
3990 {
3991         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3992         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3993         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3994         cfq_put_queue(cfqq);
3995         return cic_to_cfqq(cic, 1);
3996 }
3997
3998 /*
3999  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
4000  * was the last process referring to said cfqq.
4001  */
4002 static struct cfq_queue *
4003 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
4004 {
4005         if (cfqq_process_refs(cfqq) == 1) {
4006                 cfqq->pid = current->pid;
4007                 cfq_clear_cfqq_coop(cfqq);
4008                 cfq_clear_cfqq_split_coop(cfqq);
4009                 return cfqq;
4010         }
4011
4012         cic_set_cfqq(cic, NULL, 1);
4013
4014         cfq_put_cooperator(cfqq);
4015
4016         cfq_put_queue(cfqq);
4017         return NULL;
4018 }
4019 /*
4020  * Allocate cfq data structures associated with this request.
4021  */
4022 static int
4023 cfq_set_request(struct request_queue *q, struct request *rq, struct bio *bio,
4024                 gfp_t gfp_mask)
4025 {
4026         struct cfq_data *cfqd = q->elevator->elevator_data;
4027         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
4028         const int rw = rq_data_dir(rq);
4029         const bool is_sync = rq_is_sync(rq);
4030         struct cfq_queue *cfqq;
4031
4032         might_sleep_if(gfp_mask & __GFP_WAIT);
4033
4034         spin_lock_irq(q->queue_lock);
4035
4036         check_ioprio_changed(cic, bio);
4037         check_blkcg_changed(cic, bio);
4038 new_queue:
4039         cfqq = cic_to_cfqq(cic, is_sync);
4040         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
4041                 cfqq = cfq_get_queue(cfqd, is_sync, cic, bio, gfp_mask);
4042                 cic_set_cfqq(cic, cfqq, is_sync);
4043         } else {
4044                 /*
4045                  * If the queue was seeky for too long, break it apart.
4046                  */
4047                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
4048                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
4049                         cfqq = split_cfqq(cic, cfqq);
4050                         if (!cfqq)
4051                                 goto new_queue;
4052                 }
4053
4054                 /*
4055                  * Check to see if this queue is scheduled to merge with
4056                  * another, closely cooperating queue.  The merging of
4057                  * queues happens here as it must be done in process context.
4058                  * The reference on new_cfqq was taken in merge_cfqqs.
4059                  */
4060                 if (cfqq->new_cfqq)
4061                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
4062         }
4063
4064         cfqq->allocated[rw]++;
4065
4066         cfqq->ref++;
4067         cfqg_get(cfqq->cfqg);
4068         rq->elv.priv[0] = cfqq;
4069         rq->elv.priv[1] = cfqq->cfqg;
4070         spin_unlock_irq(q->queue_lock);
4071         return 0;
4072 }
4073
4074 static void cfq_kick_queue(struct work_struct *work)
4075 {
4076         struct cfq_data *cfqd =
4077                 container_of(work, struct cfq_data, unplug_work);
4078         struct request_queue *q = cfqd->queue;
4079
4080         spin_lock_irq(q->queue_lock);
4081         __blk_run_queue(cfqd->queue);
4082         spin_unlock_irq(q->queue_lock);
4083 }
4084
4085 /*
4086  * Timer running if the active_queue is currently idling inside its time slice
4087  */
4088 static void cfq_idle_slice_timer(unsigned long data)
4089 {
4090         struct cfq_data *cfqd = (struct cfq_data *) data;
4091         struct cfq_queue *cfqq;
4092         unsigned long flags;
4093         int timed_out = 1;
4094
4095         cfq_log(cfqd, "idle timer fired");
4096
4097         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
4098
4099         cfqq = cfqd->active_queue;
4100         if (cfqq) {
4101                 timed_out = 0;
4102
4103                 /*
4104                  * We saw a request before the queue expired, let it through
4105                  */
4106                 if (cfq_cfqq_must_dispatch(cfqq))
4107                         goto out_kick;
4108
4109                 /*
4110                  * expired
4111                  */
4112                 if (cfq_slice_used(cfqq))
4113                         goto expire;
4114
4115                 /*
4116                  * only expire and reinvoke request handler, if there are
4117                  * other queues with pending requests
4118                  */
4119                 if (!cfqd->busy_queues)
4120                         goto out_cont;
4121
4122                 /*
4123                  * not expired and it has a request pending, let it dispatch
4124                  */
4125                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
4126                         goto out_kick;
4127
4128                 /*
4129                  * Queue depth flag is reset only when the idle didn't succeed
4130                  */
4131                 cfq_clear_cfqq_deep(cfqq);
4132         }
4133 expire:
4134         cfq_slice_expired(cfqd, timed_out);
4135 out_kick:
4136         cfq_schedule_dispatch(cfqd);
4137 out_cont:
4138         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
4139 }
4140
4141 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
4142 {
4143         del_timer_sync(&cfqd->idle_slice_timer);
4144         cancel_work_sync(&cfqd->unplug_work);
4145 }
4146
4147 static void cfq_put_async_queues(struct cfq_data *cfqd)
4148 {
4149         int i;
4150
4151         for (i = 0; i < IOPRIO_BE_NR; i++) {
4152                 if (cfqd->async_cfqq[0][i])
4153                         cfq_put_queue(cfqd->async_cfqq[0][i]);
4154                 if (cfqd->async_cfqq[1][i])
4155                         cfq_put_queue(cfqd->async_cfqq[1][i]);
4156         }
4157
4158         if (cfqd->async_idle_cfqq)
4159                 cfq_put_queue(cfqd->async_idle_cfqq);
4160 }
4161
4162 static void cfq_exit_queue(struct elevator_queue *e)
4163 {
4164         struct cfq_data *cfqd = e->elevator_data;
4165         struct request_queue *q = cfqd->queue;
4166
4167         cfq_shutdown_timer_wq(cfqd);
4168
4169         spin_lock_irq(q->queue_lock);
4170
4171         if (cfqd->active_queue)
4172                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
4173
4174         cfq_put_async_queues(cfqd);
4175
4176         spin_unlock_irq(q->queue_lock);
4177
4178         cfq_shutdown_timer_wq(cfqd);
4179
4180 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4181         blkcg_deactivate_policy(q, &blkcg_policy_cfq);
4182 #else
4183         kfree(cfqd->root_group);
4184 #endif
4185         kfree(cfqd);
4186 }
4187
4188 static int cfq_init_queue(struct request_queue *q)
4189 {
4190         struct cfq_data *cfqd;
4191         struct blkcg_gq *blkg __maybe_unused;
4192         int i, ret;
4193
4194         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4195         if (!cfqd)
4196                 return -ENOMEM;
4197
4198         cfqd->queue = q;
4199         q->elevator->elevator_data = cfqd;
4200
4201         /* Init root service tree */
4202         cfqd->grp_service_tree = CFQ_RB_ROOT;
4203
4204         /* Init root group and prefer root group over other groups by default */
4205 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4206         ret = blkcg_activate_policy(q, &blkcg_policy_cfq);
4207         if (ret)
4208                 goto out_free;
4209
4210         cfqd->root_group = blkg_to_cfqg(q->root_blkg);
4211 #else
4212         ret = -ENOMEM;
4213         cfqd->root_group = kzalloc_node(sizeof(*cfqd->root_group),
4214                                         GFP_KERNEL, cfqd->queue->node);
4215         if (!cfqd->root_group)
4216                 goto out_free;
4217
4218         cfq_init_cfqg_base(cfqd->root_group);
4219 #endif
4220         cfqd->root_group->weight = 2 * CFQ_WEIGHT_DEFAULT;
4221         cfqd->root_group->leaf_weight = 2 * CFQ_WEIGHT_DEFAULT;
4222
4223         /*
4224          * Not strictly needed (since RB_ROOT just clears the node and we
4225          * zeroed cfqd on alloc), but better be safe in case someone decides
4226          * to add magic to the rb code
4227          */
4228         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4229                 cfqd->prio_trees[i] = RB_ROOT;
4230
4231         /*
4232          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4233          * Grab a permanent reference to it, so that the normal code flow
4234          * will not attempt to free it.  oom_cfqq is linked to root_group
4235          * but shouldn't hold a reference as it'll never be unlinked.  Lose
4236          * the reference from linking right away.
4237          */
4238         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4239         cfqd->oom_cfqq.ref++;
4240
4241         spin_lock_irq(q->queue_lock);
4242         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, cfqd->root_group);
4243         cfqg_put(cfqd->root_group);
4244         spin_unlock_irq(q->queue_lock);
4245
4246         init_timer(&cfqd->idle_slice_timer);
4247         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4248         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4249
4250         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4251
4252         cfqd->cfq_quantum = cfq_quantum;
4253         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4254         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4255         cfqd->cfq_back_max = cfq_back_max;
4256         cfqd->cfq_back_penalty = cfq_back_penalty;
4257         cfqd->cfq_slice[0] = cfq_slice_async;
4258         cfqd->cfq_slice[1] = cfq_slice_sync;
4259         cfqd->cfq_target_latency = cfq_target_latency;
4260         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4261         cfqd->cfq_slice_idle = cfq_slice_idle;
4262         cfqd->cfq_group_idle = cfq_group_idle;
4263         cfqd->cfq_latency = 1;
4264         cfqd->hw_tag = -1;
4265         /*
4266          * we optimistically start assuming sync ops weren't delayed in last
4267          * second, in order to have larger depth for async operations.
4268          */
4269         cfqd->last_delayed_sync = jiffies - HZ;
4270         return 0;
4271
4272 out_free:
4273         kfree(cfqd);
4274         return ret;
4275 }
4276
4277 /*
4278  * sysfs parts below -->
4279  */
4280 static ssize_t
4281 cfq_var_show(unsigned int var, char *page)
4282 {
4283         return sprintf(page, "%d\n", var);
4284 }
4285
4286 static ssize_t
4287 cfq_var_store(unsigned int *var, const char *page, size_t count)
4288 {
4289         char *p = (char *) page;
4290
4291         *var = simple_strtoul(p, &p, 10);
4292         return count;
4293 }
4294
4295 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4296 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4297 {                                                                       \
4298         struct cfq_data *cfqd = e->elevator_data;                       \
4299         unsigned int __data = __VAR;                                    \
4300         if (__CONV)                                                     \
4301                 __data = jiffies_to_msecs(__data);                      \
4302         return cfq_var_show(__data, (page));                            \
4303 }
4304 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4305 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4306 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4307 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4308 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4309 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4310 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4311 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4312 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4313 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4314 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4315 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
4316 #undef SHOW_FUNCTION
4317
4318 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4319 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4320 {                                                                       \
4321         struct cfq_data *cfqd = e->elevator_data;                       \
4322         unsigned int __data;                                            \
4323         int ret = cfq_var_store(&__data, (page), count);                \
4324         if (__data < (MIN))                                             \
4325                 __data = (MIN);                                         \
4326         else if (__data > (MAX))                                        \
4327                 __data = (MAX);                                         \
4328         if (__CONV)                                                     \
4329                 *(__PTR) = msecs_to_jiffies(__data);                    \
4330         else                                                            \
4331                 *(__PTR) = __data;                                      \
4332         return ret;                                                     \
4333 }
4334 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4335 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4336                 UINT_MAX, 1);
4337 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4338                 UINT_MAX, 1);
4339 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4340 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4341                 UINT_MAX, 0);
4342 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4343 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4344 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4345 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4346 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4347                 UINT_MAX, 0);
4348 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4349 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
4350 #undef STORE_FUNCTION
4351
4352 #define CFQ_ATTR(name) \
4353         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4354
4355 static struct elv_fs_entry cfq_attrs[] = {
4356         CFQ_ATTR(quantum),
4357         CFQ_ATTR(fifo_expire_sync),
4358         CFQ_ATTR(fifo_expire_async),
4359         CFQ_ATTR(back_seek_max),
4360         CFQ_ATTR(back_seek_penalty),
4361         CFQ_ATTR(slice_sync),
4362         CFQ_ATTR(slice_async),
4363         CFQ_ATTR(slice_async_rq),
4364         CFQ_ATTR(slice_idle),
4365         CFQ_ATTR(group_idle),
4366         CFQ_ATTR(low_latency),
4367         CFQ_ATTR(target_latency),
4368         __ATTR_NULL
4369 };
4370
4371 static struct elevator_type iosched_cfq = {
4372         .ops = {
4373                 .elevator_merge_fn =            cfq_merge,
4374                 .elevator_merged_fn =           cfq_merged_request,
4375                 .elevator_merge_req_fn =        cfq_merged_requests,
4376                 .elevator_allow_merge_fn =      cfq_allow_merge,
4377                 .elevator_bio_merged_fn =       cfq_bio_merged,
4378                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4379                 .elevator_add_req_fn =          cfq_insert_request,
4380                 .elevator_activate_req_fn =     cfq_activate_request,
4381                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4382                 .elevator_completed_req_fn =    cfq_completed_request,
4383                 .elevator_former_req_fn =       elv_rb_former_request,
4384                 .elevator_latter_req_fn =       elv_rb_latter_request,
4385                 .elevator_init_icq_fn =         cfq_init_icq,
4386                 .elevator_exit_icq_fn =         cfq_exit_icq,
4387                 .elevator_set_req_fn =          cfq_set_request,
4388                 .elevator_put_req_fn =          cfq_put_request,
4389                 .elevator_may_queue_fn =        cfq_may_queue,
4390                 .elevator_init_fn =             cfq_init_queue,
4391                 .elevator_exit_fn =             cfq_exit_queue,
4392         },
4393         .icq_size       =       sizeof(struct cfq_io_cq),
4394         .icq_align      =       __alignof__(struct cfq_io_cq),
4395         .elevator_attrs =       cfq_attrs,
4396         .elevator_name  =       "cfq",
4397         .elevator_owner =       THIS_MODULE,
4398 };
4399
4400 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4401 static struct blkcg_policy blkcg_policy_cfq = {
4402         .pd_size                = sizeof(struct cfq_group),
4403         .cftypes                = cfq_blkcg_files,
4404
4405         .pd_init_fn             = cfq_pd_init,
4406         .pd_reset_stats_fn      = cfq_pd_reset_stats,
4407 };
4408 #endif
4409
4410 static int __init cfq_init(void)
4411 {
4412         int ret;
4413
4414         /*
4415          * could be 0 on HZ < 1000 setups
4416          */
4417         if (!cfq_slice_async)
4418                 cfq_slice_async = 1;
4419         if (!cfq_slice_idle)
4420                 cfq_slice_idle = 1;
4421
4422 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4423         if (!cfq_group_idle)
4424                 cfq_group_idle = 1;
4425
4426         ret = blkcg_policy_register(&blkcg_policy_cfq);
4427         if (ret)
4428                 return ret;
4429 #else
4430         cfq_group_idle = 0;
4431 #endif
4432
4433         ret = -ENOMEM;
4434         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4435         if (!cfq_pool)
4436                 goto err_pol_unreg;
4437
4438         ret = elv_register(&iosched_cfq);
4439         if (ret)
4440                 goto err_free_pool;
4441
4442         return 0;
4443
4444 err_free_pool:
4445         kmem_cache_destroy(cfq_pool);
4446 err_pol_unreg:
4447 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4448         blkcg_policy_unregister(&blkcg_policy_cfq);
4449 #endif
4450         return ret;
4451 }
4452
4453 static void __exit cfq_exit(void)
4454 {
4455 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4456         blkcg_policy_unregister(&blkcg_policy_cfq);
4457 #endif
4458         elv_unregister(&iosched_cfq);
4459         kmem_cache_destroy(cfq_pool);
4460 }
4461
4462 module_init(cfq_init);
4463 module_exit(cfq_exit);
4464
4465 MODULE_AUTHOR("Jens Axboe");
4466 MODULE_LICENSE("GPL");
4467 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");