block/cfq: cache rightmost rb_node
[sfrench/cifs-2.6.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/sched/clock.h>
12 #include <linux/blkdev.h>
13 #include <linux/elevator.h>
14 #include <linux/ktime.h>
15 #include <linux/rbtree.h>
16 #include <linux/ioprio.h>
17 #include <linux/blktrace_api.h>
18 #include <linux/blk-cgroup.h>
19 #include "blk.h"
20 #include "blk-wbt.h"
21
22 /*
23  * tunables
24  */
25 /* max queue in one round of service */
26 static const int cfq_quantum = 8;
27 static const u64 cfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
28 /* maximum backwards seek, in KiB */
29 static const int cfq_back_max = 16 * 1024;
30 /* penalty of a backwards seek */
31 static const int cfq_back_penalty = 2;
32 static const u64 cfq_slice_sync = NSEC_PER_SEC / 10;
33 static u64 cfq_slice_async = NSEC_PER_SEC / 25;
34 static const int cfq_slice_async_rq = 2;
35 static u64 cfq_slice_idle = NSEC_PER_SEC / 125;
36 static u64 cfq_group_idle = NSEC_PER_SEC / 125;
37 static const u64 cfq_target_latency = (u64)NSEC_PER_SEC * 3/10; /* 300 ms */
38 static const int cfq_hist_divisor = 4;
39
40 /*
41  * offset from end of queue service tree for idle class
42  */
43 #define CFQ_IDLE_DELAY          (NSEC_PER_SEC / 5)
44 /* offset from end of group service tree under time slice mode */
45 #define CFQ_SLICE_MODE_GROUP_DELAY (NSEC_PER_SEC / 5)
46 /* offset from end of group service under IOPS mode */
47 #define CFQ_IOPS_MODE_GROUP_DELAY (HZ / 5)
48
49 /*
50  * below this threshold, we consider thinktime immediate
51  */
52 #define CFQ_MIN_TT              (2 * NSEC_PER_SEC / HZ)
53
54 #define CFQ_SLICE_SCALE         (5)
55 #define CFQ_HW_QUEUE_MIN        (5)
56 #define CFQ_SERVICE_SHIFT       12
57
58 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
59 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
60 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
61 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
62
63 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
64 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
65 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
66
67 static struct kmem_cache *cfq_pool;
68
69 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
70 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
71 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
72
73 #define sample_valid(samples)   ((samples) > 80)
74 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
75
76 /* blkio-related constants */
77 #define CFQ_WEIGHT_LEGACY_MIN   10
78 #define CFQ_WEIGHT_LEGACY_DFL   500
79 #define CFQ_WEIGHT_LEGACY_MAX   1000
80
81 struct cfq_ttime {
82         u64 last_end_request;
83
84         u64 ttime_total;
85         u64 ttime_mean;
86         unsigned long ttime_samples;
87 };
88
89 /*
90  * Most of our rbtree usage is for sorting with min extraction, so
91  * if we cache the leftmost node we don't have to walk down the tree
92  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
93  * move this into the elevator for the rq sorting as well.
94  */
95 struct cfq_rb_root {
96         struct rb_root_cached rb;
97         struct rb_node *rb_rightmost;
98         unsigned count;
99         u64 min_vdisktime;
100         struct cfq_ttime ttime;
101 };
102 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT_CACHED, \
103                         .rb_rightmost = NULL,                        \
104                         .ttime = {.last_end_request = ktime_get_ns(),},}
105
106 /*
107  * Per process-grouping structure
108  */
109 struct cfq_queue {
110         /* reference count */
111         int ref;
112         /* various state flags, see below */
113         unsigned int flags;
114         /* parent cfq_data */
115         struct cfq_data *cfqd;
116         /* service_tree member */
117         struct rb_node rb_node;
118         /* service_tree key */
119         u64 rb_key;
120         /* prio tree member */
121         struct rb_node p_node;
122         /* prio tree root we belong to, if any */
123         struct rb_root *p_root;
124         /* sorted list of pending requests */
125         struct rb_root sort_list;
126         /* if fifo isn't expired, next request to serve */
127         struct request *next_rq;
128         /* requests queued in sort_list */
129         int queued[2];
130         /* currently allocated requests */
131         int allocated[2];
132         /* fifo list of requests in sort_list */
133         struct list_head fifo;
134
135         /* time when queue got scheduled in to dispatch first request. */
136         u64 dispatch_start;
137         u64 allocated_slice;
138         u64 slice_dispatch;
139         /* time when first request from queue completed and slice started. */
140         u64 slice_start;
141         u64 slice_end;
142         s64 slice_resid;
143
144         /* pending priority requests */
145         int prio_pending;
146         /* number of requests that are on the dispatch list or inside driver */
147         int dispatched;
148
149         /* io prio of this group */
150         unsigned short ioprio, org_ioprio;
151         unsigned short ioprio_class, org_ioprio_class;
152
153         pid_t pid;
154
155         u32 seek_history;
156         sector_t last_request_pos;
157
158         struct cfq_rb_root *service_tree;
159         struct cfq_queue *new_cfqq;
160         struct cfq_group *cfqg;
161         /* Number of sectors dispatched from queue in single dispatch round */
162         unsigned long nr_sectors;
163 };
164
165 /*
166  * First index in the service_trees.
167  * IDLE is handled separately, so it has negative index
168  */
169 enum wl_class_t {
170         BE_WORKLOAD = 0,
171         RT_WORKLOAD = 1,
172         IDLE_WORKLOAD = 2,
173         CFQ_PRIO_NR,
174 };
175
176 /*
177  * Second index in the service_trees.
178  */
179 enum wl_type_t {
180         ASYNC_WORKLOAD = 0,
181         SYNC_NOIDLE_WORKLOAD = 1,
182         SYNC_WORKLOAD = 2
183 };
184
185 struct cfqg_stats {
186 #ifdef CONFIG_CFQ_GROUP_IOSCHED
187         /* number of ios merged */
188         struct blkg_rwstat              merged;
189         /* total time spent on device in ns, may not be accurate w/ queueing */
190         struct blkg_rwstat              service_time;
191         /* total time spent waiting in scheduler queue in ns */
192         struct blkg_rwstat              wait_time;
193         /* number of IOs queued up */
194         struct blkg_rwstat              queued;
195         /* total disk time and nr sectors dispatched by this group */
196         struct blkg_stat                time;
197 #ifdef CONFIG_DEBUG_BLK_CGROUP
198         /* time not charged to this cgroup */
199         struct blkg_stat                unaccounted_time;
200         /* sum of number of ios queued across all samples */
201         struct blkg_stat                avg_queue_size_sum;
202         /* count of samples taken for average */
203         struct blkg_stat                avg_queue_size_samples;
204         /* how many times this group has been removed from service tree */
205         struct blkg_stat                dequeue;
206         /* total time spent waiting for it to be assigned a timeslice. */
207         struct blkg_stat                group_wait_time;
208         /* time spent idling for this blkcg_gq */
209         struct blkg_stat                idle_time;
210         /* total time with empty current active q with other requests queued */
211         struct blkg_stat                empty_time;
212         /* fields after this shouldn't be cleared on stat reset */
213         uint64_t                        start_group_wait_time;
214         uint64_t                        start_idle_time;
215         uint64_t                        start_empty_time;
216         uint16_t                        flags;
217 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
218 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
219 };
220
221 /* Per-cgroup data */
222 struct cfq_group_data {
223         /* must be the first member */
224         struct blkcg_policy_data cpd;
225
226         unsigned int weight;
227         unsigned int leaf_weight;
228 };
229
230 /* This is per cgroup per device grouping structure */
231 struct cfq_group {
232         /* must be the first member */
233         struct blkg_policy_data pd;
234
235         /* group service_tree member */
236         struct rb_node rb_node;
237
238         /* group service_tree key */
239         u64 vdisktime;
240
241         /*
242          * The number of active cfqgs and sum of their weights under this
243          * cfqg.  This covers this cfqg's leaf_weight and all children's
244          * weights, but does not cover weights of further descendants.
245          *
246          * If a cfqg is on the service tree, it's active.  An active cfqg
247          * also activates its parent and contributes to the children_weight
248          * of the parent.
249          */
250         int nr_active;
251         unsigned int children_weight;
252
253         /*
254          * vfraction is the fraction of vdisktime that the tasks in this
255          * cfqg are entitled to.  This is determined by compounding the
256          * ratios walking up from this cfqg to the root.
257          *
258          * It is in fixed point w/ CFQ_SERVICE_SHIFT and the sum of all
259          * vfractions on a service tree is approximately 1.  The sum may
260          * deviate a bit due to rounding errors and fluctuations caused by
261          * cfqgs entering and leaving the service tree.
262          */
263         unsigned int vfraction;
264
265         /*
266          * There are two weights - (internal) weight is the weight of this
267          * cfqg against the sibling cfqgs.  leaf_weight is the wight of
268          * this cfqg against the child cfqgs.  For the root cfqg, both
269          * weights are kept in sync for backward compatibility.
270          */
271         unsigned int weight;
272         unsigned int new_weight;
273         unsigned int dev_weight;
274
275         unsigned int leaf_weight;
276         unsigned int new_leaf_weight;
277         unsigned int dev_leaf_weight;
278
279         /* number of cfqq currently on this group */
280         int nr_cfqq;
281
282         /*
283          * Per group busy queues average. Useful for workload slice calc. We
284          * create the array for each prio class but at run time it is used
285          * only for RT and BE class and slot for IDLE class remains unused.
286          * This is primarily done to avoid confusion and a gcc warning.
287          */
288         unsigned int busy_queues_avg[CFQ_PRIO_NR];
289         /*
290          * rr lists of queues with requests. We maintain service trees for
291          * RT and BE classes. These trees are subdivided in subclasses
292          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
293          * class there is no subclassification and all the cfq queues go on
294          * a single tree service_tree_idle.
295          * Counts are embedded in the cfq_rb_root
296          */
297         struct cfq_rb_root service_trees[2][3];
298         struct cfq_rb_root service_tree_idle;
299
300         u64 saved_wl_slice;
301         enum wl_type_t saved_wl_type;
302         enum wl_class_t saved_wl_class;
303
304         /* number of requests that are on the dispatch list or inside driver */
305         int dispatched;
306         struct cfq_ttime ttime;
307         struct cfqg_stats stats;        /* stats for this cfqg */
308
309         /* async queue for each priority case */
310         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
311         struct cfq_queue *async_idle_cfqq;
312
313 };
314
315 struct cfq_io_cq {
316         struct io_cq            icq;            /* must be the first member */
317         struct cfq_queue        *cfqq[2];
318         struct cfq_ttime        ttime;
319         int                     ioprio;         /* the current ioprio */
320 #ifdef CONFIG_CFQ_GROUP_IOSCHED
321         uint64_t                blkcg_serial_nr; /* the current blkcg serial */
322 #endif
323 };
324
325 /*
326  * Per block device queue structure
327  */
328 struct cfq_data {
329         struct request_queue *queue;
330         /* Root service tree for cfq_groups */
331         struct cfq_rb_root grp_service_tree;
332         struct cfq_group *root_group;
333
334         /*
335          * The priority currently being served
336          */
337         enum wl_class_t serving_wl_class;
338         enum wl_type_t serving_wl_type;
339         u64 workload_expires;
340         struct cfq_group *serving_group;
341
342         /*
343          * Each priority tree is sorted by next_request position.  These
344          * trees are used when determining if two or more queues are
345          * interleaving requests (see cfq_close_cooperator).
346          */
347         struct rb_root prio_trees[CFQ_PRIO_LISTS];
348
349         unsigned int busy_queues;
350         unsigned int busy_sync_queues;
351
352         int rq_in_driver;
353         int rq_in_flight[2];
354
355         /*
356          * queue-depth detection
357          */
358         int rq_queued;
359         int hw_tag;
360         /*
361          * hw_tag can be
362          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
363          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
364          *  0 => no NCQ
365          */
366         int hw_tag_est_depth;
367         unsigned int hw_tag_samples;
368
369         /*
370          * idle window management
371          */
372         struct hrtimer idle_slice_timer;
373         struct work_struct unplug_work;
374
375         struct cfq_queue *active_queue;
376         struct cfq_io_cq *active_cic;
377
378         sector_t last_position;
379
380         /*
381          * tunables, see top of file
382          */
383         unsigned int cfq_quantum;
384         unsigned int cfq_back_penalty;
385         unsigned int cfq_back_max;
386         unsigned int cfq_slice_async_rq;
387         unsigned int cfq_latency;
388         u64 cfq_fifo_expire[2];
389         u64 cfq_slice[2];
390         u64 cfq_slice_idle;
391         u64 cfq_group_idle;
392         u64 cfq_target_latency;
393
394         /*
395          * Fallback dummy cfqq for extreme OOM conditions
396          */
397         struct cfq_queue oom_cfqq;
398
399         u64 last_delayed_sync;
400 };
401
402 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
403 static void cfq_put_queue(struct cfq_queue *cfqq);
404
405 static struct cfq_rb_root *st_for(struct cfq_group *cfqg,
406                                             enum wl_class_t class,
407                                             enum wl_type_t type)
408 {
409         if (!cfqg)
410                 return NULL;
411
412         if (class == IDLE_WORKLOAD)
413                 return &cfqg->service_tree_idle;
414
415         return &cfqg->service_trees[class][type];
416 }
417
418 enum cfqq_state_flags {
419         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
420         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
421         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
422         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
423         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
424         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
425         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
426         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
427         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
428         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
429         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
430         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
431         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
432 };
433
434 #define CFQ_CFQQ_FNS(name)                                              \
435 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
436 {                                                                       \
437         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
438 }                                                                       \
439 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
440 {                                                                       \
441         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
442 }                                                                       \
443 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
444 {                                                                       \
445         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
446 }
447
448 CFQ_CFQQ_FNS(on_rr);
449 CFQ_CFQQ_FNS(wait_request);
450 CFQ_CFQQ_FNS(must_dispatch);
451 CFQ_CFQQ_FNS(must_alloc_slice);
452 CFQ_CFQQ_FNS(fifo_expire);
453 CFQ_CFQQ_FNS(idle_window);
454 CFQ_CFQQ_FNS(prio_changed);
455 CFQ_CFQQ_FNS(slice_new);
456 CFQ_CFQQ_FNS(sync);
457 CFQ_CFQQ_FNS(coop);
458 CFQ_CFQQ_FNS(split_coop);
459 CFQ_CFQQ_FNS(deep);
460 CFQ_CFQQ_FNS(wait_busy);
461 #undef CFQ_CFQQ_FNS
462
463 #if defined(CONFIG_CFQ_GROUP_IOSCHED) && defined(CONFIG_DEBUG_BLK_CGROUP)
464
465 /* cfqg stats flags */
466 enum cfqg_stats_flags {
467         CFQG_stats_waiting = 0,
468         CFQG_stats_idling,
469         CFQG_stats_empty,
470 };
471
472 #define CFQG_FLAG_FNS(name)                                             \
473 static inline void cfqg_stats_mark_##name(struct cfqg_stats *stats)     \
474 {                                                                       \
475         stats->flags |= (1 << CFQG_stats_##name);                       \
476 }                                                                       \
477 static inline void cfqg_stats_clear_##name(struct cfqg_stats *stats)    \
478 {                                                                       \
479         stats->flags &= ~(1 << CFQG_stats_##name);                      \
480 }                                                                       \
481 static inline int cfqg_stats_##name(struct cfqg_stats *stats)           \
482 {                                                                       \
483         return (stats->flags & (1 << CFQG_stats_##name)) != 0;          \
484 }                                                                       \
485
486 CFQG_FLAG_FNS(waiting)
487 CFQG_FLAG_FNS(idling)
488 CFQG_FLAG_FNS(empty)
489 #undef CFQG_FLAG_FNS
490
491 /* This should be called with the queue_lock held. */
492 static void cfqg_stats_update_group_wait_time(struct cfqg_stats *stats)
493 {
494         unsigned long long now;
495
496         if (!cfqg_stats_waiting(stats))
497                 return;
498
499         now = sched_clock();
500         if (time_after64(now, stats->start_group_wait_time))
501                 blkg_stat_add(&stats->group_wait_time,
502                               now - stats->start_group_wait_time);
503         cfqg_stats_clear_waiting(stats);
504 }
505
506 /* This should be called with the queue_lock held. */
507 static void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg,
508                                                  struct cfq_group *curr_cfqg)
509 {
510         struct cfqg_stats *stats = &cfqg->stats;
511
512         if (cfqg_stats_waiting(stats))
513                 return;
514         if (cfqg == curr_cfqg)
515                 return;
516         stats->start_group_wait_time = sched_clock();
517         cfqg_stats_mark_waiting(stats);
518 }
519
520 /* This should be called with the queue_lock held. */
521 static void cfqg_stats_end_empty_time(struct cfqg_stats *stats)
522 {
523         unsigned long long now;
524
525         if (!cfqg_stats_empty(stats))
526                 return;
527
528         now = sched_clock();
529         if (time_after64(now, stats->start_empty_time))
530                 blkg_stat_add(&stats->empty_time,
531                               now - stats->start_empty_time);
532         cfqg_stats_clear_empty(stats);
533 }
534
535 static void cfqg_stats_update_dequeue(struct cfq_group *cfqg)
536 {
537         blkg_stat_add(&cfqg->stats.dequeue, 1);
538 }
539
540 static void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg)
541 {
542         struct cfqg_stats *stats = &cfqg->stats;
543
544         if (blkg_rwstat_total(&stats->queued))
545                 return;
546
547         /*
548          * group is already marked empty. This can happen if cfqq got new
549          * request in parent group and moved to this group while being added
550          * to service tree. Just ignore the event and move on.
551          */
552         if (cfqg_stats_empty(stats))
553                 return;
554
555         stats->start_empty_time = sched_clock();
556         cfqg_stats_mark_empty(stats);
557 }
558
559 static void cfqg_stats_update_idle_time(struct cfq_group *cfqg)
560 {
561         struct cfqg_stats *stats = &cfqg->stats;
562
563         if (cfqg_stats_idling(stats)) {
564                 unsigned long long now = sched_clock();
565
566                 if (time_after64(now, stats->start_idle_time))
567                         blkg_stat_add(&stats->idle_time,
568                                       now - stats->start_idle_time);
569                 cfqg_stats_clear_idling(stats);
570         }
571 }
572
573 static void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg)
574 {
575         struct cfqg_stats *stats = &cfqg->stats;
576
577         BUG_ON(cfqg_stats_idling(stats));
578
579         stats->start_idle_time = sched_clock();
580         cfqg_stats_mark_idling(stats);
581 }
582
583 static void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg)
584 {
585         struct cfqg_stats *stats = &cfqg->stats;
586
587         blkg_stat_add(&stats->avg_queue_size_sum,
588                       blkg_rwstat_total(&stats->queued));
589         blkg_stat_add(&stats->avg_queue_size_samples, 1);
590         cfqg_stats_update_group_wait_time(stats);
591 }
592
593 #else   /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
594
595 static inline void cfqg_stats_set_start_group_wait_time(struct cfq_group *cfqg, struct cfq_group *curr_cfqg) { }
596 static inline void cfqg_stats_end_empty_time(struct cfqg_stats *stats) { }
597 static inline void cfqg_stats_update_dequeue(struct cfq_group *cfqg) { }
598 static inline void cfqg_stats_set_start_empty_time(struct cfq_group *cfqg) { }
599 static inline void cfqg_stats_update_idle_time(struct cfq_group *cfqg) { }
600 static inline void cfqg_stats_set_start_idle_time(struct cfq_group *cfqg) { }
601 static inline void cfqg_stats_update_avg_queue_size(struct cfq_group *cfqg) { }
602
603 #endif  /* CONFIG_CFQ_GROUP_IOSCHED && CONFIG_DEBUG_BLK_CGROUP */
604
605 #ifdef CONFIG_CFQ_GROUP_IOSCHED
606
607 static inline struct cfq_group *pd_to_cfqg(struct blkg_policy_data *pd)
608 {
609         return pd ? container_of(pd, struct cfq_group, pd) : NULL;
610 }
611
612 static struct cfq_group_data
613 *cpd_to_cfqgd(struct blkcg_policy_data *cpd)
614 {
615         return cpd ? container_of(cpd, struct cfq_group_data, cpd) : NULL;
616 }
617
618 static inline struct blkcg_gq *cfqg_to_blkg(struct cfq_group *cfqg)
619 {
620         return pd_to_blkg(&cfqg->pd);
621 }
622
623 static struct blkcg_policy blkcg_policy_cfq;
624
625 static inline struct cfq_group *blkg_to_cfqg(struct blkcg_gq *blkg)
626 {
627         return pd_to_cfqg(blkg_to_pd(blkg, &blkcg_policy_cfq));
628 }
629
630 static struct cfq_group_data *blkcg_to_cfqgd(struct blkcg *blkcg)
631 {
632         return cpd_to_cfqgd(blkcg_to_cpd(blkcg, &blkcg_policy_cfq));
633 }
634
635 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg)
636 {
637         struct blkcg_gq *pblkg = cfqg_to_blkg(cfqg)->parent;
638
639         return pblkg ? blkg_to_cfqg(pblkg) : NULL;
640 }
641
642 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
643                                       struct cfq_group *ancestor)
644 {
645         return cgroup_is_descendant(cfqg_to_blkg(cfqg)->blkcg->css.cgroup,
646                                     cfqg_to_blkg(ancestor)->blkcg->css.cgroup);
647 }
648
649 static inline void cfqg_get(struct cfq_group *cfqg)
650 {
651         return blkg_get(cfqg_to_blkg(cfqg));
652 }
653
654 static inline void cfqg_put(struct cfq_group *cfqg)
655 {
656         return blkg_put(cfqg_to_blkg(cfqg));
657 }
658
659 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  do {                    \
660         blk_add_cgroup_trace_msg((cfqd)->queue,                         \
661                         cfqg_to_blkg((cfqq)->cfqg)->blkcg,              \
662                         "cfq%d%c%c " fmt, (cfqq)->pid,                  \
663                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
664                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
665                           ##args);                                      \
666 } while (0)
667
668 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)  do {                    \
669         blk_add_cgroup_trace_msg((cfqd)->queue,                         \
670                         cfqg_to_blkg(cfqg)->blkcg, fmt, ##args);        \
671 } while (0)
672
673 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
674                                             struct cfq_group *curr_cfqg,
675                                             unsigned int op)
676 {
677         blkg_rwstat_add(&cfqg->stats.queued, op, 1);
678         cfqg_stats_end_empty_time(&cfqg->stats);
679         cfqg_stats_set_start_group_wait_time(cfqg, curr_cfqg);
680 }
681
682 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
683                         uint64_t time, unsigned long unaccounted_time)
684 {
685         blkg_stat_add(&cfqg->stats.time, time);
686 #ifdef CONFIG_DEBUG_BLK_CGROUP
687         blkg_stat_add(&cfqg->stats.unaccounted_time, unaccounted_time);
688 #endif
689 }
690
691 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
692                                                unsigned int op)
693 {
694         blkg_rwstat_add(&cfqg->stats.queued, op, -1);
695 }
696
697 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
698                                                unsigned int op)
699 {
700         blkg_rwstat_add(&cfqg->stats.merged, op, 1);
701 }
702
703 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
704                         uint64_t start_time, uint64_t io_start_time,
705                         unsigned int op)
706 {
707         struct cfqg_stats *stats = &cfqg->stats;
708         unsigned long long now = sched_clock();
709
710         if (time_after64(now, io_start_time))
711                 blkg_rwstat_add(&stats->service_time, op, now - io_start_time);
712         if (time_after64(io_start_time, start_time))
713                 blkg_rwstat_add(&stats->wait_time, op,
714                                 io_start_time - start_time);
715 }
716
717 /* @stats = 0 */
718 static void cfqg_stats_reset(struct cfqg_stats *stats)
719 {
720         /* queued stats shouldn't be cleared */
721         blkg_rwstat_reset(&stats->merged);
722         blkg_rwstat_reset(&stats->service_time);
723         blkg_rwstat_reset(&stats->wait_time);
724         blkg_stat_reset(&stats->time);
725 #ifdef CONFIG_DEBUG_BLK_CGROUP
726         blkg_stat_reset(&stats->unaccounted_time);
727         blkg_stat_reset(&stats->avg_queue_size_sum);
728         blkg_stat_reset(&stats->avg_queue_size_samples);
729         blkg_stat_reset(&stats->dequeue);
730         blkg_stat_reset(&stats->group_wait_time);
731         blkg_stat_reset(&stats->idle_time);
732         blkg_stat_reset(&stats->empty_time);
733 #endif
734 }
735
736 /* @to += @from */
737 static void cfqg_stats_add_aux(struct cfqg_stats *to, struct cfqg_stats *from)
738 {
739         /* queued stats shouldn't be cleared */
740         blkg_rwstat_add_aux(&to->merged, &from->merged);
741         blkg_rwstat_add_aux(&to->service_time, &from->service_time);
742         blkg_rwstat_add_aux(&to->wait_time, &from->wait_time);
743         blkg_stat_add_aux(&from->time, &from->time);
744 #ifdef CONFIG_DEBUG_BLK_CGROUP
745         blkg_stat_add_aux(&to->unaccounted_time, &from->unaccounted_time);
746         blkg_stat_add_aux(&to->avg_queue_size_sum, &from->avg_queue_size_sum);
747         blkg_stat_add_aux(&to->avg_queue_size_samples, &from->avg_queue_size_samples);
748         blkg_stat_add_aux(&to->dequeue, &from->dequeue);
749         blkg_stat_add_aux(&to->group_wait_time, &from->group_wait_time);
750         blkg_stat_add_aux(&to->idle_time, &from->idle_time);
751         blkg_stat_add_aux(&to->empty_time, &from->empty_time);
752 #endif
753 }
754
755 /*
756  * Transfer @cfqg's stats to its parent's aux counts so that the ancestors'
757  * recursive stats can still account for the amount used by this cfqg after
758  * it's gone.
759  */
760 static void cfqg_stats_xfer_dead(struct cfq_group *cfqg)
761 {
762         struct cfq_group *parent = cfqg_parent(cfqg);
763
764         lockdep_assert_held(cfqg_to_blkg(cfqg)->q->queue_lock);
765
766         if (unlikely(!parent))
767                 return;
768
769         cfqg_stats_add_aux(&parent->stats, &cfqg->stats);
770         cfqg_stats_reset(&cfqg->stats);
771 }
772
773 #else   /* CONFIG_CFQ_GROUP_IOSCHED */
774
775 static inline struct cfq_group *cfqg_parent(struct cfq_group *cfqg) { return NULL; }
776 static inline bool cfqg_is_descendant(struct cfq_group *cfqg,
777                                       struct cfq_group *ancestor)
778 {
779         return true;
780 }
781 static inline void cfqg_get(struct cfq_group *cfqg) { }
782 static inline void cfqg_put(struct cfq_group *cfqg) { }
783
784 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
785         blk_add_trace_msg((cfqd)->queue, "cfq%d%c%c " fmt, (cfqq)->pid, \
786                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A',              \
787                         cfqq_type((cfqq)) == SYNC_NOIDLE_WORKLOAD ? 'N' : ' ',\
788                                 ##args)
789 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
790
791 static inline void cfqg_stats_update_io_add(struct cfq_group *cfqg,
792                         struct cfq_group *curr_cfqg, unsigned int op) { }
793 static inline void cfqg_stats_update_timeslice_used(struct cfq_group *cfqg,
794                         uint64_t time, unsigned long unaccounted_time) { }
795 static inline void cfqg_stats_update_io_remove(struct cfq_group *cfqg,
796                         unsigned int op) { }
797 static inline void cfqg_stats_update_io_merged(struct cfq_group *cfqg,
798                         unsigned int op) { }
799 static inline void cfqg_stats_update_completion(struct cfq_group *cfqg,
800                         uint64_t start_time, uint64_t io_start_time,
801                         unsigned int op) { }
802
803 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
804
805 #define cfq_log(cfqd, fmt, args...)     \
806         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
807
808 /* Traverses through cfq group service trees */
809 #define for_each_cfqg_st(cfqg, i, j, st) \
810         for (i = 0; i <= IDLE_WORKLOAD; i++) \
811                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
812                         : &cfqg->service_tree_idle; \
813                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
814                         (i == IDLE_WORKLOAD && j == 0); \
815                         j++, st = i < IDLE_WORKLOAD ? \
816                         &cfqg->service_trees[i][j]: NULL) \
817
818 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
819         struct cfq_ttime *ttime, bool group_idle)
820 {
821         u64 slice;
822         if (!sample_valid(ttime->ttime_samples))
823                 return false;
824         if (group_idle)
825                 slice = cfqd->cfq_group_idle;
826         else
827                 slice = cfqd->cfq_slice_idle;
828         return ttime->ttime_mean > slice;
829 }
830
831 static inline bool iops_mode(struct cfq_data *cfqd)
832 {
833         /*
834          * If we are not idling on queues and it is a NCQ drive, parallel
835          * execution of requests is on and measuring time is not possible
836          * in most of the cases until and unless we drive shallower queue
837          * depths and that becomes a performance bottleneck. In such cases
838          * switch to start providing fairness in terms of number of IOs.
839          */
840         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
841                 return true;
842         else
843                 return false;
844 }
845
846 static inline enum wl_class_t cfqq_class(struct cfq_queue *cfqq)
847 {
848         if (cfq_class_idle(cfqq))
849                 return IDLE_WORKLOAD;
850         if (cfq_class_rt(cfqq))
851                 return RT_WORKLOAD;
852         return BE_WORKLOAD;
853 }
854
855
856 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
857 {
858         if (!cfq_cfqq_sync(cfqq))
859                 return ASYNC_WORKLOAD;
860         if (!cfq_cfqq_idle_window(cfqq))
861                 return SYNC_NOIDLE_WORKLOAD;
862         return SYNC_WORKLOAD;
863 }
864
865 static inline int cfq_group_busy_queues_wl(enum wl_class_t wl_class,
866                                         struct cfq_data *cfqd,
867                                         struct cfq_group *cfqg)
868 {
869         if (wl_class == IDLE_WORKLOAD)
870                 return cfqg->service_tree_idle.count;
871
872         return cfqg->service_trees[wl_class][ASYNC_WORKLOAD].count +
873                 cfqg->service_trees[wl_class][SYNC_NOIDLE_WORKLOAD].count +
874                 cfqg->service_trees[wl_class][SYNC_WORKLOAD].count;
875 }
876
877 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
878                                         struct cfq_group *cfqg)
879 {
880         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count +
881                 cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
882 }
883
884 static void cfq_dispatch_insert(struct request_queue *, struct request *);
885 static struct cfq_queue *cfq_get_queue(struct cfq_data *cfqd, bool is_sync,
886                                        struct cfq_io_cq *cic, struct bio *bio);
887
888 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
889 {
890         /* cic->icq is the first member, %NULL will convert to %NULL */
891         return container_of(icq, struct cfq_io_cq, icq);
892 }
893
894 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
895                                                struct io_context *ioc)
896 {
897         if (ioc)
898                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
899         return NULL;
900 }
901
902 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
903 {
904         return cic->cfqq[is_sync];
905 }
906
907 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
908                                 bool is_sync)
909 {
910         cic->cfqq[is_sync] = cfqq;
911 }
912
913 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
914 {
915         return cic->icq.q->elevator->elevator_data;
916 }
917
918 /*
919  * scheduler run of queue, if there are requests pending and no one in the
920  * driver that will restart queueing
921  */
922 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
923 {
924         if (cfqd->busy_queues) {
925                 cfq_log(cfqd, "schedule dispatch");
926                 kblockd_schedule_work(&cfqd->unplug_work);
927         }
928 }
929
930 /*
931  * Scale schedule slice based on io priority. Use the sync time slice only
932  * if a queue is marked sync and has sync io queued. A sync queue with async
933  * io only, should not get full sync slice length.
934  */
935 static inline u64 cfq_prio_slice(struct cfq_data *cfqd, bool sync,
936                                  unsigned short prio)
937 {
938         u64 base_slice = cfqd->cfq_slice[sync];
939         u64 slice = div_u64(base_slice, CFQ_SLICE_SCALE);
940
941         WARN_ON(prio >= IOPRIO_BE_NR);
942
943         return base_slice + (slice * (4 - prio));
944 }
945
946 static inline u64
947 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
948 {
949         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
950 }
951
952 /**
953  * cfqg_scale_charge - scale disk time charge according to cfqg weight
954  * @charge: disk time being charged
955  * @vfraction: vfraction of the cfqg, fixed point w/ CFQ_SERVICE_SHIFT
956  *
957  * Scale @charge according to @vfraction, which is in range (0, 1].  The
958  * scaling is inversely proportional.
959  *
960  * scaled = charge / vfraction
961  *
962  * The result is also in fixed point w/ CFQ_SERVICE_SHIFT.
963  */
964 static inline u64 cfqg_scale_charge(u64 charge,
965                                     unsigned int vfraction)
966 {
967         u64 c = charge << CFQ_SERVICE_SHIFT;    /* make it fixed point */
968
969         /* charge / vfraction */
970         c <<= CFQ_SERVICE_SHIFT;
971         return div_u64(c, vfraction);
972 }
973
974 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
975 {
976         s64 delta = (s64)(vdisktime - min_vdisktime);
977         if (delta > 0)
978                 min_vdisktime = vdisktime;
979
980         return min_vdisktime;
981 }
982
983 static void update_min_vdisktime(struct cfq_rb_root *st)
984 {
985         if (!RB_EMPTY_ROOT(&st->rb.rb_root)) {
986                 struct cfq_group *cfqg = rb_entry_cfqg(st->rb.rb_leftmost);
987
988                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
989                                                   cfqg->vdisktime);
990         }
991 }
992
993 /*
994  * get averaged number of queues of RT/BE priority.
995  * average is updated, with a formula that gives more weight to higher numbers,
996  * to quickly follows sudden increases and decrease slowly
997  */
998
999 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
1000                                         struct cfq_group *cfqg, bool rt)
1001 {
1002         unsigned min_q, max_q;
1003         unsigned mult  = cfq_hist_divisor - 1;
1004         unsigned round = cfq_hist_divisor / 2;
1005         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
1006
1007         min_q = min(cfqg->busy_queues_avg[rt], busy);
1008         max_q = max(cfqg->busy_queues_avg[rt], busy);
1009         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
1010                 cfq_hist_divisor;
1011         return cfqg->busy_queues_avg[rt];
1012 }
1013
1014 static inline u64
1015 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
1016 {
1017         return cfqd->cfq_target_latency * cfqg->vfraction >> CFQ_SERVICE_SHIFT;
1018 }
1019
1020 static inline u64
1021 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1022 {
1023         u64 slice = cfq_prio_to_slice(cfqd, cfqq);
1024         if (cfqd->cfq_latency) {
1025                 /*
1026                  * interested queues (we consider only the ones with the same
1027                  * priority class in the cfq group)
1028                  */
1029                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
1030                                                 cfq_class_rt(cfqq));
1031                 u64 sync_slice = cfqd->cfq_slice[1];
1032                 u64 expect_latency = sync_slice * iq;
1033                 u64 group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
1034
1035                 if (expect_latency > group_slice) {
1036                         u64 base_low_slice = 2 * cfqd->cfq_slice_idle;
1037                         u64 low_slice;
1038
1039                         /* scale low_slice according to IO priority
1040                          * and sync vs async */
1041                         low_slice = div64_u64(base_low_slice*slice, sync_slice);
1042                         low_slice = min(slice, low_slice);
1043                         /* the adapted slice value is scaled to fit all iqs
1044                          * into the target latency */
1045                         slice = div64_u64(slice*group_slice, expect_latency);
1046                         slice = max(slice, low_slice);
1047                 }
1048         }
1049         return slice;
1050 }
1051
1052 static inline void
1053 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1054 {
1055         u64 slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
1056         u64 now = ktime_get_ns();
1057
1058         cfqq->slice_start = now;
1059         cfqq->slice_end = now + slice;
1060         cfqq->allocated_slice = slice;
1061         cfq_log_cfqq(cfqd, cfqq, "set_slice=%llu", cfqq->slice_end - now);
1062 }
1063
1064 /*
1065  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
1066  * isn't valid until the first request from the dispatch is activated
1067  * and the slice time set.
1068  */
1069 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
1070 {
1071         if (cfq_cfqq_slice_new(cfqq))
1072                 return false;
1073         if (ktime_get_ns() < cfqq->slice_end)
1074                 return false;
1075
1076         return true;
1077 }
1078
1079 /*
1080  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
1081  * We choose the request that is closest to the head right now. Distance
1082  * behind the head is penalized and only allowed to a certain extent.
1083  */
1084 static struct request *
1085 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
1086 {
1087         sector_t s1, s2, d1 = 0, d2 = 0;
1088         unsigned long back_max;
1089 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
1090 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
1091         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
1092
1093         if (rq1 == NULL || rq1 == rq2)
1094                 return rq2;
1095         if (rq2 == NULL)
1096                 return rq1;
1097
1098         if (rq_is_sync(rq1) != rq_is_sync(rq2))
1099                 return rq_is_sync(rq1) ? rq1 : rq2;
1100
1101         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
1102                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
1103
1104         s1 = blk_rq_pos(rq1);
1105         s2 = blk_rq_pos(rq2);
1106
1107         /*
1108          * by definition, 1KiB is 2 sectors
1109          */
1110         back_max = cfqd->cfq_back_max * 2;
1111
1112         /*
1113          * Strict one way elevator _except_ in the case where we allow
1114          * short backward seeks which are biased as twice the cost of a
1115          * similar forward seek.
1116          */
1117         if (s1 >= last)
1118                 d1 = s1 - last;
1119         else if (s1 + back_max >= last)
1120                 d1 = (last - s1) * cfqd->cfq_back_penalty;
1121         else
1122                 wrap |= CFQ_RQ1_WRAP;
1123
1124         if (s2 >= last)
1125                 d2 = s2 - last;
1126         else if (s2 + back_max >= last)
1127                 d2 = (last - s2) * cfqd->cfq_back_penalty;
1128         else
1129                 wrap |= CFQ_RQ2_WRAP;
1130
1131         /* Found required data */
1132
1133         /*
1134          * By doing switch() on the bit mask "wrap" we avoid having to
1135          * check two variables for all permutations: --> faster!
1136          */
1137         switch (wrap) {
1138         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
1139                 if (d1 < d2)
1140                         return rq1;
1141                 else if (d2 < d1)
1142                         return rq2;
1143                 else {
1144                         if (s1 >= s2)
1145                                 return rq1;
1146                         else
1147                                 return rq2;
1148                 }
1149
1150         case CFQ_RQ2_WRAP:
1151                 return rq1;
1152         case CFQ_RQ1_WRAP:
1153                 return rq2;
1154         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
1155         default:
1156                 /*
1157                  * Since both rqs are wrapped,
1158                  * start with the one that's further behind head
1159                  * (--> only *one* back seek required),
1160                  * since back seek takes more time than forward.
1161                  */
1162                 if (s1 <= s2)
1163                         return rq1;
1164                 else
1165                         return rq2;
1166         }
1167 }
1168
1169 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
1170 {
1171         /* Service tree is empty */
1172         if (!root->count)
1173                 return NULL;
1174
1175         return rb_entry(rb_first_cached(&root->rb), struct cfq_queue, rb_node);
1176 }
1177
1178 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
1179 {
1180         return rb_entry_cfqg(rb_first_cached(&root->rb));
1181 }
1182
1183 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
1184 {
1185         if (root->rb_rightmost == n)
1186                 root->rb_rightmost = rb_prev(n);
1187
1188         rb_erase_cached(n, &root->rb);
1189         RB_CLEAR_NODE(n);
1190
1191         --root->count;
1192 }
1193
1194 /*
1195  * would be nice to take fifo expire time into account as well
1196  */
1197 static struct request *
1198 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1199                   struct request *last)
1200 {
1201         struct rb_node *rbnext = rb_next(&last->rb_node);
1202         struct rb_node *rbprev = rb_prev(&last->rb_node);
1203         struct request *next = NULL, *prev = NULL;
1204
1205         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
1206
1207         if (rbprev)
1208                 prev = rb_entry_rq(rbprev);
1209
1210         if (rbnext)
1211                 next = rb_entry_rq(rbnext);
1212         else {
1213                 rbnext = rb_first(&cfqq->sort_list);
1214                 if (rbnext && rbnext != &last->rb_node)
1215                         next = rb_entry_rq(rbnext);
1216         }
1217
1218         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
1219 }
1220
1221 static u64 cfq_slice_offset(struct cfq_data *cfqd,
1222                             struct cfq_queue *cfqq)
1223 {
1224         /*
1225          * just an approximation, should be ok.
1226          */
1227         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
1228                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
1229 }
1230
1231 static inline s64
1232 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
1233 {
1234         return cfqg->vdisktime - st->min_vdisktime;
1235 }
1236
1237 static void
1238 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1239 {
1240         struct rb_node **node = &st->rb.rb_root.rb_node;
1241         struct rb_node *parent = NULL;
1242         struct cfq_group *__cfqg;
1243         s64 key = cfqg_key(st, cfqg);
1244         bool leftmost = true, rightmost = true;
1245
1246         while (*node != NULL) {
1247                 parent = *node;
1248                 __cfqg = rb_entry_cfqg(parent);
1249
1250                 if (key < cfqg_key(st, __cfqg)) {
1251                         node = &parent->rb_left;
1252                         rightmost = false;
1253                 } else {
1254                         node = &parent->rb_right;
1255                         leftmost = false;
1256                 }
1257         }
1258
1259         if (rightmost)
1260                 st->rb_rightmost = &cfqg->rb_node;
1261
1262         rb_link_node(&cfqg->rb_node, parent, node);
1263         rb_insert_color_cached(&cfqg->rb_node, &st->rb, leftmost);
1264 }
1265
1266 /*
1267  * This has to be called only on activation of cfqg
1268  */
1269 static void
1270 cfq_update_group_weight(struct cfq_group *cfqg)
1271 {
1272         if (cfqg->new_weight) {
1273                 cfqg->weight = cfqg->new_weight;
1274                 cfqg->new_weight = 0;
1275         }
1276 }
1277
1278 static void
1279 cfq_update_group_leaf_weight(struct cfq_group *cfqg)
1280 {
1281         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1282
1283         if (cfqg->new_leaf_weight) {
1284                 cfqg->leaf_weight = cfqg->new_leaf_weight;
1285                 cfqg->new_leaf_weight = 0;
1286         }
1287 }
1288
1289 static void
1290 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
1291 {
1292         unsigned int vfr = 1 << CFQ_SERVICE_SHIFT;      /* start with 1 */
1293         struct cfq_group *pos = cfqg;
1294         struct cfq_group *parent;
1295         bool propagate;
1296
1297         /* add to the service tree */
1298         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
1299
1300         /*
1301          * Update leaf_weight.  We cannot update weight at this point
1302          * because cfqg might already have been activated and is
1303          * contributing its current weight to the parent's child_weight.
1304          */
1305         cfq_update_group_leaf_weight(cfqg);
1306         __cfq_group_service_tree_add(st, cfqg);
1307
1308         /*
1309          * Activate @cfqg and calculate the portion of vfraction @cfqg is
1310          * entitled to.  vfraction is calculated by walking the tree
1311          * towards the root calculating the fraction it has at each level.
1312          * The compounded ratio is how much vfraction @cfqg owns.
1313          *
1314          * Start with the proportion tasks in this cfqg has against active
1315          * children cfqgs - its leaf_weight against children_weight.
1316          */
1317         propagate = !pos->nr_active++;
1318         pos->children_weight += pos->leaf_weight;
1319         vfr = vfr * pos->leaf_weight / pos->children_weight;
1320
1321         /*
1322          * Compound ->weight walking up the tree.  Both activation and
1323          * vfraction calculation are done in the same loop.  Propagation
1324          * stops once an already activated node is met.  vfraction
1325          * calculation should always continue to the root.
1326          */
1327         while ((parent = cfqg_parent(pos))) {
1328                 if (propagate) {
1329                         cfq_update_group_weight(pos);
1330                         propagate = !parent->nr_active++;
1331                         parent->children_weight += pos->weight;
1332                 }
1333                 vfr = vfr * pos->weight / parent->children_weight;
1334                 pos = parent;
1335         }
1336
1337         cfqg->vfraction = max_t(unsigned, vfr, 1);
1338 }
1339
1340 static inline u64 cfq_get_cfqg_vdisktime_delay(struct cfq_data *cfqd)
1341 {
1342         if (!iops_mode(cfqd))
1343                 return CFQ_SLICE_MODE_GROUP_DELAY;
1344         else
1345                 return CFQ_IOPS_MODE_GROUP_DELAY;
1346 }
1347
1348 static void
1349 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
1350 {
1351         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1352         struct cfq_group *__cfqg;
1353         struct rb_node *n;
1354
1355         cfqg->nr_cfqq++;
1356         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1357                 return;
1358
1359         /*
1360          * Currently put the group at the end. Later implement something
1361          * so that groups get lesser vtime based on their weights, so that
1362          * if group does not loose all if it was not continuously backlogged.
1363          */
1364         n = st->rb_rightmost;
1365         if (n) {
1366                 __cfqg = rb_entry_cfqg(n);
1367                 cfqg->vdisktime = __cfqg->vdisktime +
1368                         cfq_get_cfqg_vdisktime_delay(cfqd);
1369         } else
1370                 cfqg->vdisktime = st->min_vdisktime;
1371         cfq_group_service_tree_add(st, cfqg);
1372 }
1373
1374 static void
1375 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
1376 {
1377         struct cfq_group *pos = cfqg;
1378         bool propagate;
1379
1380         /*
1381          * Undo activation from cfq_group_service_tree_add().  Deactivate
1382          * @cfqg and propagate deactivation upwards.
1383          */
1384         propagate = !--pos->nr_active;
1385         pos->children_weight -= pos->leaf_weight;
1386
1387         while (propagate) {
1388                 struct cfq_group *parent = cfqg_parent(pos);
1389
1390                 /* @pos has 0 nr_active at this point */
1391                 WARN_ON_ONCE(pos->children_weight);
1392                 pos->vfraction = 0;
1393
1394                 if (!parent)
1395                         break;
1396
1397                 propagate = !--parent->nr_active;
1398                 parent->children_weight -= pos->weight;
1399                 pos = parent;
1400         }
1401
1402         /* remove from the service tree */
1403         if (!RB_EMPTY_NODE(&cfqg->rb_node))
1404                 cfq_rb_erase(&cfqg->rb_node, st);
1405 }
1406
1407 static void
1408 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
1409 {
1410         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1411
1412         BUG_ON(cfqg->nr_cfqq < 1);
1413         cfqg->nr_cfqq--;
1414
1415         /* If there are other cfq queues under this group, don't delete it */
1416         if (cfqg->nr_cfqq)
1417                 return;
1418
1419         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
1420         cfq_group_service_tree_del(st, cfqg);
1421         cfqg->saved_wl_slice = 0;
1422         cfqg_stats_update_dequeue(cfqg);
1423 }
1424
1425 static inline u64 cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
1426                                        u64 *unaccounted_time)
1427 {
1428         u64 slice_used;
1429         u64 now = ktime_get_ns();
1430
1431         /*
1432          * Queue got expired before even a single request completed or
1433          * got expired immediately after first request completion.
1434          */
1435         if (!cfqq->slice_start || cfqq->slice_start == now) {
1436                 /*
1437                  * Also charge the seek time incurred to the group, otherwise
1438                  * if there are mutiple queues in the group, each can dispatch
1439                  * a single request on seeky media and cause lots of seek time
1440                  * and group will never know it.
1441                  */
1442                 slice_used = max_t(u64, (now - cfqq->dispatch_start),
1443                                         jiffies_to_nsecs(1));
1444         } else {
1445                 slice_used = now - cfqq->slice_start;
1446                 if (slice_used > cfqq->allocated_slice) {
1447                         *unaccounted_time = slice_used - cfqq->allocated_slice;
1448                         slice_used = cfqq->allocated_slice;
1449                 }
1450                 if (cfqq->slice_start > cfqq->dispatch_start)
1451                         *unaccounted_time += cfqq->slice_start -
1452                                         cfqq->dispatch_start;
1453         }
1454
1455         return slice_used;
1456 }
1457
1458 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
1459                                 struct cfq_queue *cfqq)
1460 {
1461         struct cfq_rb_root *st = &cfqd->grp_service_tree;
1462         u64 used_sl, charge, unaccounted_sl = 0;
1463         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
1464                         - cfqg->service_tree_idle.count;
1465         unsigned int vfr;
1466         u64 now = ktime_get_ns();
1467
1468         BUG_ON(nr_sync < 0);
1469         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
1470
1471         if (iops_mode(cfqd))
1472                 charge = cfqq->slice_dispatch;
1473         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
1474                 charge = cfqq->allocated_slice;
1475
1476         /*
1477          * Can't update vdisktime while on service tree and cfqg->vfraction
1478          * is valid only while on it.  Cache vfr, leave the service tree,
1479          * update vdisktime and go back on.  The re-addition to the tree
1480          * will also update the weights as necessary.
1481          */
1482         vfr = cfqg->vfraction;
1483         cfq_group_service_tree_del(st, cfqg);
1484         cfqg->vdisktime += cfqg_scale_charge(charge, vfr);
1485         cfq_group_service_tree_add(st, cfqg);
1486
1487         /* This group is being expired. Save the context */
1488         if (cfqd->workload_expires > now) {
1489                 cfqg->saved_wl_slice = cfqd->workload_expires - now;
1490                 cfqg->saved_wl_type = cfqd->serving_wl_type;
1491                 cfqg->saved_wl_class = cfqd->serving_wl_class;
1492         } else
1493                 cfqg->saved_wl_slice = 0;
1494
1495         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1496                                         st->min_vdisktime);
1497         cfq_log_cfqq(cfqq->cfqd, cfqq,
1498                      "sl_used=%llu disp=%llu charge=%llu iops=%u sect=%lu",
1499                      used_sl, cfqq->slice_dispatch, charge,
1500                      iops_mode(cfqd), cfqq->nr_sectors);
1501         cfqg_stats_update_timeslice_used(cfqg, used_sl, unaccounted_sl);
1502         cfqg_stats_set_start_empty_time(cfqg);
1503 }
1504
1505 /**
1506  * cfq_init_cfqg_base - initialize base part of a cfq_group
1507  * @cfqg: cfq_group to initialize
1508  *
1509  * Initialize the base part which is used whether %CONFIG_CFQ_GROUP_IOSCHED
1510  * is enabled or not.
1511  */
1512 static void cfq_init_cfqg_base(struct cfq_group *cfqg)
1513 {
1514         struct cfq_rb_root *st;
1515         int i, j;
1516
1517         for_each_cfqg_st(cfqg, i, j, st)
1518                 *st = CFQ_RB_ROOT;
1519         RB_CLEAR_NODE(&cfqg->rb_node);
1520
1521         cfqg->ttime.last_end_request = ktime_get_ns();
1522 }
1523
1524 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1525 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1526                             bool on_dfl, bool reset_dev, bool is_leaf_weight);
1527
1528 static void cfqg_stats_exit(struct cfqg_stats *stats)
1529 {
1530         blkg_rwstat_exit(&stats->merged);
1531         blkg_rwstat_exit(&stats->service_time);
1532         blkg_rwstat_exit(&stats->wait_time);
1533         blkg_rwstat_exit(&stats->queued);
1534         blkg_stat_exit(&stats->time);
1535 #ifdef CONFIG_DEBUG_BLK_CGROUP
1536         blkg_stat_exit(&stats->unaccounted_time);
1537         blkg_stat_exit(&stats->avg_queue_size_sum);
1538         blkg_stat_exit(&stats->avg_queue_size_samples);
1539         blkg_stat_exit(&stats->dequeue);
1540         blkg_stat_exit(&stats->group_wait_time);
1541         blkg_stat_exit(&stats->idle_time);
1542         blkg_stat_exit(&stats->empty_time);
1543 #endif
1544 }
1545
1546 static int cfqg_stats_init(struct cfqg_stats *stats, gfp_t gfp)
1547 {
1548         if (blkg_rwstat_init(&stats->merged, gfp) ||
1549             blkg_rwstat_init(&stats->service_time, gfp) ||
1550             blkg_rwstat_init(&stats->wait_time, gfp) ||
1551             blkg_rwstat_init(&stats->queued, gfp) ||
1552             blkg_stat_init(&stats->time, gfp))
1553                 goto err;
1554
1555 #ifdef CONFIG_DEBUG_BLK_CGROUP
1556         if (blkg_stat_init(&stats->unaccounted_time, gfp) ||
1557             blkg_stat_init(&stats->avg_queue_size_sum, gfp) ||
1558             blkg_stat_init(&stats->avg_queue_size_samples, gfp) ||
1559             blkg_stat_init(&stats->dequeue, gfp) ||
1560             blkg_stat_init(&stats->group_wait_time, gfp) ||
1561             blkg_stat_init(&stats->idle_time, gfp) ||
1562             blkg_stat_init(&stats->empty_time, gfp))
1563                 goto err;
1564 #endif
1565         return 0;
1566 err:
1567         cfqg_stats_exit(stats);
1568         return -ENOMEM;
1569 }
1570
1571 static struct blkcg_policy_data *cfq_cpd_alloc(gfp_t gfp)
1572 {
1573         struct cfq_group_data *cgd;
1574
1575         cgd = kzalloc(sizeof(*cgd), gfp);
1576         if (!cgd)
1577                 return NULL;
1578         return &cgd->cpd;
1579 }
1580
1581 static void cfq_cpd_init(struct blkcg_policy_data *cpd)
1582 {
1583         struct cfq_group_data *cgd = cpd_to_cfqgd(cpd);
1584         unsigned int weight = cgroup_subsys_on_dfl(io_cgrp_subsys) ?
1585                               CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1586
1587         if (cpd_to_blkcg(cpd) == &blkcg_root)
1588                 weight *= 2;
1589
1590         cgd->weight = weight;
1591         cgd->leaf_weight = weight;
1592 }
1593
1594 static void cfq_cpd_free(struct blkcg_policy_data *cpd)
1595 {
1596         kfree(cpd_to_cfqgd(cpd));
1597 }
1598
1599 static void cfq_cpd_bind(struct blkcg_policy_data *cpd)
1600 {
1601         struct blkcg *blkcg = cpd_to_blkcg(cpd);
1602         bool on_dfl = cgroup_subsys_on_dfl(io_cgrp_subsys);
1603         unsigned int weight = on_dfl ? CGROUP_WEIGHT_DFL : CFQ_WEIGHT_LEGACY_DFL;
1604
1605         if (blkcg == &blkcg_root)
1606                 weight *= 2;
1607
1608         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, false));
1609         WARN_ON_ONCE(__cfq_set_weight(&blkcg->css, weight, on_dfl, true, true));
1610 }
1611
1612 static struct blkg_policy_data *cfq_pd_alloc(gfp_t gfp, int node)
1613 {
1614         struct cfq_group *cfqg;
1615
1616         cfqg = kzalloc_node(sizeof(*cfqg), gfp, node);
1617         if (!cfqg)
1618                 return NULL;
1619
1620         cfq_init_cfqg_base(cfqg);
1621         if (cfqg_stats_init(&cfqg->stats, gfp)) {
1622                 kfree(cfqg);
1623                 return NULL;
1624         }
1625
1626         return &cfqg->pd;
1627 }
1628
1629 static void cfq_pd_init(struct blkg_policy_data *pd)
1630 {
1631         struct cfq_group *cfqg = pd_to_cfqg(pd);
1632         struct cfq_group_data *cgd = blkcg_to_cfqgd(pd->blkg->blkcg);
1633
1634         cfqg->weight = cgd->weight;
1635         cfqg->leaf_weight = cgd->leaf_weight;
1636 }
1637
1638 static void cfq_pd_offline(struct blkg_policy_data *pd)
1639 {
1640         struct cfq_group *cfqg = pd_to_cfqg(pd);
1641         int i;
1642
1643         for (i = 0; i < IOPRIO_BE_NR; i++) {
1644                 if (cfqg->async_cfqq[0][i])
1645                         cfq_put_queue(cfqg->async_cfqq[0][i]);
1646                 if (cfqg->async_cfqq[1][i])
1647                         cfq_put_queue(cfqg->async_cfqq[1][i]);
1648         }
1649
1650         if (cfqg->async_idle_cfqq)
1651                 cfq_put_queue(cfqg->async_idle_cfqq);
1652
1653         /*
1654          * @blkg is going offline and will be ignored by
1655          * blkg_[rw]stat_recursive_sum().  Transfer stats to the parent so
1656          * that they don't get lost.  If IOs complete after this point, the
1657          * stats for them will be lost.  Oh well...
1658          */
1659         cfqg_stats_xfer_dead(cfqg);
1660 }
1661
1662 static void cfq_pd_free(struct blkg_policy_data *pd)
1663 {
1664         struct cfq_group *cfqg = pd_to_cfqg(pd);
1665
1666         cfqg_stats_exit(&cfqg->stats);
1667         return kfree(cfqg);
1668 }
1669
1670 static void cfq_pd_reset_stats(struct blkg_policy_data *pd)
1671 {
1672         struct cfq_group *cfqg = pd_to_cfqg(pd);
1673
1674         cfqg_stats_reset(&cfqg->stats);
1675 }
1676
1677 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
1678                                          struct blkcg *blkcg)
1679 {
1680         struct blkcg_gq *blkg;
1681
1682         blkg = blkg_lookup(blkcg, cfqd->queue);
1683         if (likely(blkg))
1684                 return blkg_to_cfqg(blkg);
1685         return NULL;
1686 }
1687
1688 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1689 {
1690         cfqq->cfqg = cfqg;
1691         /* cfqq reference on cfqg */
1692         cfqg_get(cfqg);
1693 }
1694
1695 static u64 cfqg_prfill_weight_device(struct seq_file *sf,
1696                                      struct blkg_policy_data *pd, int off)
1697 {
1698         struct cfq_group *cfqg = pd_to_cfqg(pd);
1699
1700         if (!cfqg->dev_weight)
1701                 return 0;
1702         return __blkg_prfill_u64(sf, pd, cfqg->dev_weight);
1703 }
1704
1705 static int cfqg_print_weight_device(struct seq_file *sf, void *v)
1706 {
1707         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1708                           cfqg_prfill_weight_device, &blkcg_policy_cfq,
1709                           0, false);
1710         return 0;
1711 }
1712
1713 static u64 cfqg_prfill_leaf_weight_device(struct seq_file *sf,
1714                                           struct blkg_policy_data *pd, int off)
1715 {
1716         struct cfq_group *cfqg = pd_to_cfqg(pd);
1717
1718         if (!cfqg->dev_leaf_weight)
1719                 return 0;
1720         return __blkg_prfill_u64(sf, pd, cfqg->dev_leaf_weight);
1721 }
1722
1723 static int cfqg_print_leaf_weight_device(struct seq_file *sf, void *v)
1724 {
1725         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1726                           cfqg_prfill_leaf_weight_device, &blkcg_policy_cfq,
1727                           0, false);
1728         return 0;
1729 }
1730
1731 static int cfq_print_weight(struct seq_file *sf, void *v)
1732 {
1733         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1734         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1735         unsigned int val = 0;
1736
1737         if (cgd)
1738                 val = cgd->weight;
1739
1740         seq_printf(sf, "%u\n", val);
1741         return 0;
1742 }
1743
1744 static int cfq_print_leaf_weight(struct seq_file *sf, void *v)
1745 {
1746         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
1747         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
1748         unsigned int val = 0;
1749
1750         if (cgd)
1751                 val = cgd->leaf_weight;
1752
1753         seq_printf(sf, "%u\n", val);
1754         return 0;
1755 }
1756
1757 static ssize_t __cfqg_set_weight_device(struct kernfs_open_file *of,
1758                                         char *buf, size_t nbytes, loff_t off,
1759                                         bool on_dfl, bool is_leaf_weight)
1760 {
1761         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1762         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1763         struct blkcg *blkcg = css_to_blkcg(of_css(of));
1764         struct blkg_conf_ctx ctx;
1765         struct cfq_group *cfqg;
1766         struct cfq_group_data *cfqgd;
1767         int ret;
1768         u64 v;
1769
1770         ret = blkg_conf_prep(blkcg, &blkcg_policy_cfq, buf, &ctx);
1771         if (ret)
1772                 return ret;
1773
1774         if (sscanf(ctx.body, "%llu", &v) == 1) {
1775                 /* require "default" on dfl */
1776                 ret = -ERANGE;
1777                 if (!v && on_dfl)
1778                         goto out_finish;
1779         } else if (!strcmp(strim(ctx.body), "default")) {
1780                 v = 0;
1781         } else {
1782                 ret = -EINVAL;
1783                 goto out_finish;
1784         }
1785
1786         cfqg = blkg_to_cfqg(ctx.blkg);
1787         cfqgd = blkcg_to_cfqgd(blkcg);
1788
1789         ret = -ERANGE;
1790         if (!v || (v >= min && v <= max)) {
1791                 if (!is_leaf_weight) {
1792                         cfqg->dev_weight = v;
1793                         cfqg->new_weight = v ?: cfqgd->weight;
1794                 } else {
1795                         cfqg->dev_leaf_weight = v;
1796                         cfqg->new_leaf_weight = v ?: cfqgd->leaf_weight;
1797                 }
1798                 ret = 0;
1799         }
1800 out_finish:
1801         blkg_conf_finish(&ctx);
1802         return ret ?: nbytes;
1803 }
1804
1805 static ssize_t cfqg_set_weight_device(struct kernfs_open_file *of,
1806                                       char *buf, size_t nbytes, loff_t off)
1807 {
1808         return __cfqg_set_weight_device(of, buf, nbytes, off, false, false);
1809 }
1810
1811 static ssize_t cfqg_set_leaf_weight_device(struct kernfs_open_file *of,
1812                                            char *buf, size_t nbytes, loff_t off)
1813 {
1814         return __cfqg_set_weight_device(of, buf, nbytes, off, false, true);
1815 }
1816
1817 static int __cfq_set_weight(struct cgroup_subsys_state *css, u64 val,
1818                             bool on_dfl, bool reset_dev, bool is_leaf_weight)
1819 {
1820         unsigned int min = on_dfl ? CGROUP_WEIGHT_MIN : CFQ_WEIGHT_LEGACY_MIN;
1821         unsigned int max = on_dfl ? CGROUP_WEIGHT_MAX : CFQ_WEIGHT_LEGACY_MAX;
1822         struct blkcg *blkcg = css_to_blkcg(css);
1823         struct blkcg_gq *blkg;
1824         struct cfq_group_data *cfqgd;
1825         int ret = 0;
1826
1827         if (val < min || val > max)
1828                 return -ERANGE;
1829
1830         spin_lock_irq(&blkcg->lock);
1831         cfqgd = blkcg_to_cfqgd(blkcg);
1832         if (!cfqgd) {
1833                 ret = -EINVAL;
1834                 goto out;
1835         }
1836
1837         if (!is_leaf_weight)
1838                 cfqgd->weight = val;
1839         else
1840                 cfqgd->leaf_weight = val;
1841
1842         hlist_for_each_entry(blkg, &blkcg->blkg_list, blkcg_node) {
1843                 struct cfq_group *cfqg = blkg_to_cfqg(blkg);
1844
1845                 if (!cfqg)
1846                         continue;
1847
1848                 if (!is_leaf_weight) {
1849                         if (reset_dev)
1850                                 cfqg->dev_weight = 0;
1851                         if (!cfqg->dev_weight)
1852                                 cfqg->new_weight = cfqgd->weight;
1853                 } else {
1854                         if (reset_dev)
1855                                 cfqg->dev_leaf_weight = 0;
1856                         if (!cfqg->dev_leaf_weight)
1857                                 cfqg->new_leaf_weight = cfqgd->leaf_weight;
1858                 }
1859         }
1860
1861 out:
1862         spin_unlock_irq(&blkcg->lock);
1863         return ret;
1864 }
1865
1866 static int cfq_set_weight(struct cgroup_subsys_state *css, struct cftype *cft,
1867                           u64 val)
1868 {
1869         return __cfq_set_weight(css, val, false, false, false);
1870 }
1871
1872 static int cfq_set_leaf_weight(struct cgroup_subsys_state *css,
1873                                struct cftype *cft, u64 val)
1874 {
1875         return __cfq_set_weight(css, val, false, false, true);
1876 }
1877
1878 static int cfqg_print_stat(struct seq_file *sf, void *v)
1879 {
1880         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_stat,
1881                           &blkcg_policy_cfq, seq_cft(sf)->private, false);
1882         return 0;
1883 }
1884
1885 static int cfqg_print_rwstat(struct seq_file *sf, void *v)
1886 {
1887         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), blkg_prfill_rwstat,
1888                           &blkcg_policy_cfq, seq_cft(sf)->private, true);
1889         return 0;
1890 }
1891
1892 static u64 cfqg_prfill_stat_recursive(struct seq_file *sf,
1893                                       struct blkg_policy_data *pd, int off)
1894 {
1895         u64 sum = blkg_stat_recursive_sum(pd_to_blkg(pd),
1896                                           &blkcg_policy_cfq, off);
1897         return __blkg_prfill_u64(sf, pd, sum);
1898 }
1899
1900 static u64 cfqg_prfill_rwstat_recursive(struct seq_file *sf,
1901                                         struct blkg_policy_data *pd, int off)
1902 {
1903         struct blkg_rwstat sum = blkg_rwstat_recursive_sum(pd_to_blkg(pd),
1904                                                         &blkcg_policy_cfq, off);
1905         return __blkg_prfill_rwstat(sf, pd, &sum);
1906 }
1907
1908 static int cfqg_print_stat_recursive(struct seq_file *sf, void *v)
1909 {
1910         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1911                           cfqg_prfill_stat_recursive, &blkcg_policy_cfq,
1912                           seq_cft(sf)->private, false);
1913         return 0;
1914 }
1915
1916 static int cfqg_print_rwstat_recursive(struct seq_file *sf, void *v)
1917 {
1918         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1919                           cfqg_prfill_rwstat_recursive, &blkcg_policy_cfq,
1920                           seq_cft(sf)->private, true);
1921         return 0;
1922 }
1923
1924 static u64 cfqg_prfill_sectors(struct seq_file *sf, struct blkg_policy_data *pd,
1925                                int off)
1926 {
1927         u64 sum = blkg_rwstat_total(&pd->blkg->stat_bytes);
1928
1929         return __blkg_prfill_u64(sf, pd, sum >> 9);
1930 }
1931
1932 static int cfqg_print_stat_sectors(struct seq_file *sf, void *v)
1933 {
1934         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1935                           cfqg_prfill_sectors, &blkcg_policy_cfq, 0, false);
1936         return 0;
1937 }
1938
1939 static u64 cfqg_prfill_sectors_recursive(struct seq_file *sf,
1940                                          struct blkg_policy_data *pd, int off)
1941 {
1942         struct blkg_rwstat tmp = blkg_rwstat_recursive_sum(pd->blkg, NULL,
1943                                         offsetof(struct blkcg_gq, stat_bytes));
1944         u64 sum = atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_READ]) +
1945                 atomic64_read(&tmp.aux_cnt[BLKG_RWSTAT_WRITE]);
1946
1947         return __blkg_prfill_u64(sf, pd, sum >> 9);
1948 }
1949
1950 static int cfqg_print_stat_sectors_recursive(struct seq_file *sf, void *v)
1951 {
1952         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1953                           cfqg_prfill_sectors_recursive, &blkcg_policy_cfq, 0,
1954                           false);
1955         return 0;
1956 }
1957
1958 #ifdef CONFIG_DEBUG_BLK_CGROUP
1959 static u64 cfqg_prfill_avg_queue_size(struct seq_file *sf,
1960                                       struct blkg_policy_data *pd, int off)
1961 {
1962         struct cfq_group *cfqg = pd_to_cfqg(pd);
1963         u64 samples = blkg_stat_read(&cfqg->stats.avg_queue_size_samples);
1964         u64 v = 0;
1965
1966         if (samples) {
1967                 v = blkg_stat_read(&cfqg->stats.avg_queue_size_sum);
1968                 v = div64_u64(v, samples);
1969         }
1970         __blkg_prfill_u64(sf, pd, v);
1971         return 0;
1972 }
1973
1974 /* print avg_queue_size */
1975 static int cfqg_print_avg_queue_size(struct seq_file *sf, void *v)
1976 {
1977         blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
1978                           cfqg_prfill_avg_queue_size, &blkcg_policy_cfq,
1979                           0, false);
1980         return 0;
1981 }
1982 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
1983
1984 static struct cftype cfq_blkcg_legacy_files[] = {
1985         /* on root, weight is mapped to leaf_weight */
1986         {
1987                 .name = "weight_device",
1988                 .flags = CFTYPE_ONLY_ON_ROOT,
1989                 .seq_show = cfqg_print_leaf_weight_device,
1990                 .write = cfqg_set_leaf_weight_device,
1991         },
1992         {
1993                 .name = "weight",
1994                 .flags = CFTYPE_ONLY_ON_ROOT,
1995                 .seq_show = cfq_print_leaf_weight,
1996                 .write_u64 = cfq_set_leaf_weight,
1997         },
1998
1999         /* no such mapping necessary for !roots */
2000         {
2001                 .name = "weight_device",
2002                 .flags = CFTYPE_NOT_ON_ROOT,
2003                 .seq_show = cfqg_print_weight_device,
2004                 .write = cfqg_set_weight_device,
2005         },
2006         {
2007                 .name = "weight",
2008                 .flags = CFTYPE_NOT_ON_ROOT,
2009                 .seq_show = cfq_print_weight,
2010                 .write_u64 = cfq_set_weight,
2011         },
2012
2013         {
2014                 .name = "leaf_weight_device",
2015                 .seq_show = cfqg_print_leaf_weight_device,
2016                 .write = cfqg_set_leaf_weight_device,
2017         },
2018         {
2019                 .name = "leaf_weight",
2020                 .seq_show = cfq_print_leaf_weight,
2021                 .write_u64 = cfq_set_leaf_weight,
2022         },
2023
2024         /* statistics, covers only the tasks in the cfqg */
2025         {
2026                 .name = "time",
2027                 .private = offsetof(struct cfq_group, stats.time),
2028                 .seq_show = cfqg_print_stat,
2029         },
2030         {
2031                 .name = "sectors",
2032                 .seq_show = cfqg_print_stat_sectors,
2033         },
2034         {
2035                 .name = "io_service_bytes",
2036                 .private = (unsigned long)&blkcg_policy_cfq,
2037                 .seq_show = blkg_print_stat_bytes,
2038         },
2039         {
2040                 .name = "io_serviced",
2041                 .private = (unsigned long)&blkcg_policy_cfq,
2042                 .seq_show = blkg_print_stat_ios,
2043         },
2044         {
2045                 .name = "io_service_time",
2046                 .private = offsetof(struct cfq_group, stats.service_time),
2047                 .seq_show = cfqg_print_rwstat,
2048         },
2049         {
2050                 .name = "io_wait_time",
2051                 .private = offsetof(struct cfq_group, stats.wait_time),
2052                 .seq_show = cfqg_print_rwstat,
2053         },
2054         {
2055                 .name = "io_merged",
2056                 .private = offsetof(struct cfq_group, stats.merged),
2057                 .seq_show = cfqg_print_rwstat,
2058         },
2059         {
2060                 .name = "io_queued",
2061                 .private = offsetof(struct cfq_group, stats.queued),
2062                 .seq_show = cfqg_print_rwstat,
2063         },
2064
2065         /* the same statictics which cover the cfqg and its descendants */
2066         {
2067                 .name = "time_recursive",
2068                 .private = offsetof(struct cfq_group, stats.time),
2069                 .seq_show = cfqg_print_stat_recursive,
2070         },
2071         {
2072                 .name = "sectors_recursive",
2073                 .seq_show = cfqg_print_stat_sectors_recursive,
2074         },
2075         {
2076                 .name = "io_service_bytes_recursive",
2077                 .private = (unsigned long)&blkcg_policy_cfq,
2078                 .seq_show = blkg_print_stat_bytes_recursive,
2079         },
2080         {
2081                 .name = "io_serviced_recursive",
2082                 .private = (unsigned long)&blkcg_policy_cfq,
2083                 .seq_show = blkg_print_stat_ios_recursive,
2084         },
2085         {
2086                 .name = "io_service_time_recursive",
2087                 .private = offsetof(struct cfq_group, stats.service_time),
2088                 .seq_show = cfqg_print_rwstat_recursive,
2089         },
2090         {
2091                 .name = "io_wait_time_recursive",
2092                 .private = offsetof(struct cfq_group, stats.wait_time),
2093                 .seq_show = cfqg_print_rwstat_recursive,
2094         },
2095         {
2096                 .name = "io_merged_recursive",
2097                 .private = offsetof(struct cfq_group, stats.merged),
2098                 .seq_show = cfqg_print_rwstat_recursive,
2099         },
2100         {
2101                 .name = "io_queued_recursive",
2102                 .private = offsetof(struct cfq_group, stats.queued),
2103                 .seq_show = cfqg_print_rwstat_recursive,
2104         },
2105 #ifdef CONFIG_DEBUG_BLK_CGROUP
2106         {
2107                 .name = "avg_queue_size",
2108                 .seq_show = cfqg_print_avg_queue_size,
2109         },
2110         {
2111                 .name = "group_wait_time",
2112                 .private = offsetof(struct cfq_group, stats.group_wait_time),
2113                 .seq_show = cfqg_print_stat,
2114         },
2115         {
2116                 .name = "idle_time",
2117                 .private = offsetof(struct cfq_group, stats.idle_time),
2118                 .seq_show = cfqg_print_stat,
2119         },
2120         {
2121                 .name = "empty_time",
2122                 .private = offsetof(struct cfq_group, stats.empty_time),
2123                 .seq_show = cfqg_print_stat,
2124         },
2125         {
2126                 .name = "dequeue",
2127                 .private = offsetof(struct cfq_group, stats.dequeue),
2128                 .seq_show = cfqg_print_stat,
2129         },
2130         {
2131                 .name = "unaccounted_time",
2132                 .private = offsetof(struct cfq_group, stats.unaccounted_time),
2133                 .seq_show = cfqg_print_stat,
2134         },
2135 #endif  /* CONFIG_DEBUG_BLK_CGROUP */
2136         { }     /* terminate */
2137 };
2138
2139 static int cfq_print_weight_on_dfl(struct seq_file *sf, void *v)
2140 {
2141         struct blkcg *blkcg = css_to_blkcg(seq_css(sf));
2142         struct cfq_group_data *cgd = blkcg_to_cfqgd(blkcg);
2143
2144         seq_printf(sf, "default %u\n", cgd->weight);
2145         blkcg_print_blkgs(sf, blkcg, cfqg_prfill_weight_device,
2146                           &blkcg_policy_cfq, 0, false);
2147         return 0;
2148 }
2149
2150 static ssize_t cfq_set_weight_on_dfl(struct kernfs_open_file *of,
2151                                      char *buf, size_t nbytes, loff_t off)
2152 {
2153         char *endp;
2154         int ret;
2155         u64 v;
2156
2157         buf = strim(buf);
2158
2159         /* "WEIGHT" or "default WEIGHT" sets the default weight */
2160         v = simple_strtoull(buf, &endp, 0);
2161         if (*endp == '\0' || sscanf(buf, "default %llu", &v) == 1) {
2162                 ret = __cfq_set_weight(of_css(of), v, true, false, false);
2163                 return ret ?: nbytes;
2164         }
2165
2166         /* "MAJ:MIN WEIGHT" */
2167         return __cfqg_set_weight_device(of, buf, nbytes, off, true, false);
2168 }
2169
2170 static struct cftype cfq_blkcg_files[] = {
2171         {
2172                 .name = "weight",
2173                 .flags = CFTYPE_NOT_ON_ROOT,
2174                 .seq_show = cfq_print_weight_on_dfl,
2175                 .write = cfq_set_weight_on_dfl,
2176         },
2177         { }     /* terminate */
2178 };
2179
2180 #else /* GROUP_IOSCHED */
2181 static struct cfq_group *cfq_lookup_cfqg(struct cfq_data *cfqd,
2182                                          struct blkcg *blkcg)
2183 {
2184         return cfqd->root_group;
2185 }
2186
2187 static inline void
2188 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
2189         cfqq->cfqg = cfqg;
2190 }
2191
2192 #endif /* GROUP_IOSCHED */
2193
2194 /*
2195  * The cfqd->service_trees holds all pending cfq_queue's that have
2196  * requests waiting to be processed. It is sorted in the order that
2197  * we will service the queues.
2198  */
2199 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2200                                  bool add_front)
2201 {
2202         struct rb_node **p, *parent;
2203         struct cfq_queue *__cfqq;
2204         u64 rb_key;
2205         struct cfq_rb_root *st;
2206         bool leftmost = true;
2207         int new_cfqq = 1;
2208         u64 now = ktime_get_ns();
2209
2210         st = st_for(cfqq->cfqg, cfqq_class(cfqq), cfqq_type(cfqq));
2211         if (cfq_class_idle(cfqq)) {
2212                 rb_key = CFQ_IDLE_DELAY;
2213                 parent = st->rb_rightmost;
2214                 if (parent && parent != &cfqq->rb_node) {
2215                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2216                         rb_key += __cfqq->rb_key;
2217                 } else
2218                         rb_key += now;
2219         } else if (!add_front) {
2220                 /*
2221                  * Get our rb key offset. Subtract any residual slice
2222                  * value carried from last service. A negative resid
2223                  * count indicates slice overrun, and this should position
2224                  * the next service time further away in the tree.
2225                  */
2226                 rb_key = cfq_slice_offset(cfqd, cfqq) + now;
2227                 rb_key -= cfqq->slice_resid;
2228                 cfqq->slice_resid = 0;
2229         } else {
2230                 rb_key = -NSEC_PER_SEC;
2231                 __cfqq = cfq_rb_first(st);
2232                 rb_key += __cfqq ? __cfqq->rb_key : now;
2233         }
2234
2235         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2236                 new_cfqq = 0;
2237                 /*
2238                  * same position, nothing more to do
2239                  */
2240                 if (rb_key == cfqq->rb_key && cfqq->service_tree == st)
2241                         return;
2242
2243                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2244                 cfqq->service_tree = NULL;
2245         }
2246
2247         parent = NULL;
2248         cfqq->service_tree = st;
2249         p = &st->rb.rb_root.rb_node;
2250         while (*p) {
2251                 parent = *p;
2252                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
2253
2254                 /*
2255                  * sort by key, that represents service time.
2256                  */
2257                 if (rb_key < __cfqq->rb_key)
2258                         p = &parent->rb_left;
2259                 else {
2260                         p = &parent->rb_right;
2261                         leftmost = false;
2262                 }
2263         }
2264
2265         cfqq->rb_key = rb_key;
2266         rb_link_node(&cfqq->rb_node, parent, p);
2267         rb_insert_color_cached(&cfqq->rb_node, &st->rb, leftmost);
2268         st->count++;
2269         if (add_front || !new_cfqq)
2270                 return;
2271         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
2272 }
2273
2274 static struct cfq_queue *
2275 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
2276                      sector_t sector, struct rb_node **ret_parent,
2277                      struct rb_node ***rb_link)
2278 {
2279         struct rb_node **p, *parent;
2280         struct cfq_queue *cfqq = NULL;
2281
2282         parent = NULL;
2283         p = &root->rb_node;
2284         while (*p) {
2285                 struct rb_node **n;
2286
2287                 parent = *p;
2288                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
2289
2290                 /*
2291                  * Sort strictly based on sector.  Smallest to the left,
2292                  * largest to the right.
2293                  */
2294                 if (sector > blk_rq_pos(cfqq->next_rq))
2295                         n = &(*p)->rb_right;
2296                 else if (sector < blk_rq_pos(cfqq->next_rq))
2297                         n = &(*p)->rb_left;
2298                 else
2299                         break;
2300                 p = n;
2301                 cfqq = NULL;
2302         }
2303
2304         *ret_parent = parent;
2305         if (rb_link)
2306                 *rb_link = p;
2307         return cfqq;
2308 }
2309
2310 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2311 {
2312         struct rb_node **p, *parent;
2313         struct cfq_queue *__cfqq;
2314
2315         if (cfqq->p_root) {
2316                 rb_erase(&cfqq->p_node, cfqq->p_root);
2317                 cfqq->p_root = NULL;
2318         }
2319
2320         if (cfq_class_idle(cfqq))
2321                 return;
2322         if (!cfqq->next_rq)
2323                 return;
2324
2325         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
2326         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
2327                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
2328         if (!__cfqq) {
2329                 rb_link_node(&cfqq->p_node, parent, p);
2330                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
2331         } else
2332                 cfqq->p_root = NULL;
2333 }
2334
2335 /*
2336  * Update cfqq's position in the service tree.
2337  */
2338 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2339 {
2340         /*
2341          * Resorting requires the cfqq to be on the RR list already.
2342          */
2343         if (cfq_cfqq_on_rr(cfqq)) {
2344                 cfq_service_tree_add(cfqd, cfqq, 0);
2345                 cfq_prio_tree_add(cfqd, cfqq);
2346         }
2347 }
2348
2349 /*
2350  * add to busy list of queues for service, trying to be fair in ordering
2351  * the pending list according to last request service
2352  */
2353 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2354 {
2355         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
2356         BUG_ON(cfq_cfqq_on_rr(cfqq));
2357         cfq_mark_cfqq_on_rr(cfqq);
2358         cfqd->busy_queues++;
2359         if (cfq_cfqq_sync(cfqq))
2360                 cfqd->busy_sync_queues++;
2361
2362         cfq_resort_rr_list(cfqd, cfqq);
2363 }
2364
2365 /*
2366  * Called when the cfqq no longer has requests pending, remove it from
2367  * the service tree.
2368  */
2369 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2370 {
2371         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
2372         BUG_ON(!cfq_cfqq_on_rr(cfqq));
2373         cfq_clear_cfqq_on_rr(cfqq);
2374
2375         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
2376                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
2377                 cfqq->service_tree = NULL;
2378         }
2379         if (cfqq->p_root) {
2380                 rb_erase(&cfqq->p_node, cfqq->p_root);
2381                 cfqq->p_root = NULL;
2382         }
2383
2384         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
2385         BUG_ON(!cfqd->busy_queues);
2386         cfqd->busy_queues--;
2387         if (cfq_cfqq_sync(cfqq))
2388                 cfqd->busy_sync_queues--;
2389 }
2390
2391 /*
2392  * rb tree support functions
2393  */
2394 static void cfq_del_rq_rb(struct request *rq)
2395 {
2396         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2397         const int sync = rq_is_sync(rq);
2398
2399         BUG_ON(!cfqq->queued[sync]);
2400         cfqq->queued[sync]--;
2401
2402         elv_rb_del(&cfqq->sort_list, rq);
2403
2404         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
2405                 /*
2406                  * Queue will be deleted from service tree when we actually
2407                  * expire it later. Right now just remove it from prio tree
2408                  * as it is empty.
2409                  */
2410                 if (cfqq->p_root) {
2411                         rb_erase(&cfqq->p_node, cfqq->p_root);
2412                         cfqq->p_root = NULL;
2413                 }
2414         }
2415 }
2416
2417 static void cfq_add_rq_rb(struct request *rq)
2418 {
2419         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2420         struct cfq_data *cfqd = cfqq->cfqd;
2421         struct request *prev;
2422
2423         cfqq->queued[rq_is_sync(rq)]++;
2424
2425         elv_rb_add(&cfqq->sort_list, rq);
2426
2427         if (!cfq_cfqq_on_rr(cfqq))
2428                 cfq_add_cfqq_rr(cfqd, cfqq);
2429
2430         /*
2431          * check if this request is a better next-serve candidate
2432          */
2433         prev = cfqq->next_rq;
2434         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
2435
2436         /*
2437          * adjust priority tree position, if ->next_rq changes
2438          */
2439         if (prev != cfqq->next_rq)
2440                 cfq_prio_tree_add(cfqd, cfqq);
2441
2442         BUG_ON(!cfqq->next_rq);
2443 }
2444
2445 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
2446 {
2447         elv_rb_del(&cfqq->sort_list, rq);
2448         cfqq->queued[rq_is_sync(rq)]--;
2449         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2450         cfq_add_rq_rb(rq);
2451         cfqg_stats_update_io_add(RQ_CFQG(rq), cfqq->cfqd->serving_group,
2452                                  rq->cmd_flags);
2453 }
2454
2455 static struct request *
2456 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
2457 {
2458         struct task_struct *tsk = current;
2459         struct cfq_io_cq *cic;
2460         struct cfq_queue *cfqq;
2461
2462         cic = cfq_cic_lookup(cfqd, tsk->io_context);
2463         if (!cic)
2464                 return NULL;
2465
2466         cfqq = cic_to_cfqq(cic, op_is_sync(bio->bi_opf));
2467         if (cfqq)
2468                 return elv_rb_find(&cfqq->sort_list, bio_end_sector(bio));
2469
2470         return NULL;
2471 }
2472
2473 static void cfq_activate_request(struct request_queue *q, struct request *rq)
2474 {
2475         struct cfq_data *cfqd = q->elevator->elevator_data;
2476
2477         cfqd->rq_in_driver++;
2478         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
2479                                                 cfqd->rq_in_driver);
2480
2481         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
2482 }
2483
2484 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
2485 {
2486         struct cfq_data *cfqd = q->elevator->elevator_data;
2487
2488         WARN_ON(!cfqd->rq_in_driver);
2489         cfqd->rq_in_driver--;
2490         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
2491                                                 cfqd->rq_in_driver);
2492 }
2493
2494 static void cfq_remove_request(struct request *rq)
2495 {
2496         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2497
2498         if (cfqq->next_rq == rq)
2499                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
2500
2501         list_del_init(&rq->queuelist);
2502         cfq_del_rq_rb(rq);
2503
2504         cfqq->cfqd->rq_queued--;
2505         cfqg_stats_update_io_remove(RQ_CFQG(rq), rq->cmd_flags);
2506         if (rq->cmd_flags & REQ_PRIO) {
2507                 WARN_ON(!cfqq->prio_pending);
2508                 cfqq->prio_pending--;
2509         }
2510 }
2511
2512 static enum elv_merge cfq_merge(struct request_queue *q, struct request **req,
2513                      struct bio *bio)
2514 {
2515         struct cfq_data *cfqd = q->elevator->elevator_data;
2516         struct request *__rq;
2517
2518         __rq = cfq_find_rq_fmerge(cfqd, bio);
2519         if (__rq && elv_bio_merge_ok(__rq, bio)) {
2520                 *req = __rq;
2521                 return ELEVATOR_FRONT_MERGE;
2522         }
2523
2524         return ELEVATOR_NO_MERGE;
2525 }
2526
2527 static void cfq_merged_request(struct request_queue *q, struct request *req,
2528                                enum elv_merge type)
2529 {
2530         if (type == ELEVATOR_FRONT_MERGE) {
2531                 struct cfq_queue *cfqq = RQ_CFQQ(req);
2532
2533                 cfq_reposition_rq_rb(cfqq, req);
2534         }
2535 }
2536
2537 static void cfq_bio_merged(struct request_queue *q, struct request *req,
2538                                 struct bio *bio)
2539 {
2540         cfqg_stats_update_io_merged(RQ_CFQG(req), bio->bi_opf);
2541 }
2542
2543 static void
2544 cfq_merged_requests(struct request_queue *q, struct request *rq,
2545                     struct request *next)
2546 {
2547         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2548         struct cfq_data *cfqd = q->elevator->elevator_data;
2549
2550         /*
2551          * reposition in fifo if next is older than rq
2552          */
2553         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2554             next->fifo_time < rq->fifo_time &&
2555             cfqq == RQ_CFQQ(next)) {
2556                 list_move(&rq->queuelist, &next->queuelist);
2557                 rq->fifo_time = next->fifo_time;
2558         }
2559
2560         if (cfqq->next_rq == next)
2561                 cfqq->next_rq = rq;
2562         cfq_remove_request(next);
2563         cfqg_stats_update_io_merged(RQ_CFQG(rq), next->cmd_flags);
2564
2565         cfqq = RQ_CFQQ(next);
2566         /*
2567          * all requests of this queue are merged to other queues, delete it
2568          * from the service tree. If it's the active_queue,
2569          * cfq_dispatch_requests() will choose to expire it or do idle
2570          */
2571         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
2572             cfqq != cfqd->active_queue)
2573                 cfq_del_cfqq_rr(cfqd, cfqq);
2574 }
2575
2576 static int cfq_allow_bio_merge(struct request_queue *q, struct request *rq,
2577                                struct bio *bio)
2578 {
2579         struct cfq_data *cfqd = q->elevator->elevator_data;
2580         bool is_sync = op_is_sync(bio->bi_opf);
2581         struct cfq_io_cq *cic;
2582         struct cfq_queue *cfqq;
2583
2584         /*
2585          * Disallow merge of a sync bio into an async request.
2586          */
2587         if (is_sync && !rq_is_sync(rq))
2588                 return false;
2589
2590         /*
2591          * Lookup the cfqq that this bio will be queued with and allow
2592          * merge only if rq is queued there.
2593          */
2594         cic = cfq_cic_lookup(cfqd, current->io_context);
2595         if (!cic)
2596                 return false;
2597
2598         cfqq = cic_to_cfqq(cic, is_sync);
2599         return cfqq == RQ_CFQQ(rq);
2600 }
2601
2602 static int cfq_allow_rq_merge(struct request_queue *q, struct request *rq,
2603                               struct request *next)
2604 {
2605         return RQ_CFQQ(rq) == RQ_CFQQ(next);
2606 }
2607
2608 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2609 {
2610         hrtimer_try_to_cancel(&cfqd->idle_slice_timer);
2611         cfqg_stats_update_idle_time(cfqq->cfqg);
2612 }
2613
2614 static void __cfq_set_active_queue(struct cfq_data *cfqd,
2615                                    struct cfq_queue *cfqq)
2616 {
2617         if (cfqq) {
2618                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_class:%d wl_type:%d",
2619                                 cfqd->serving_wl_class, cfqd->serving_wl_type);
2620                 cfqg_stats_update_avg_queue_size(cfqq->cfqg);
2621                 cfqq->slice_start = 0;
2622                 cfqq->dispatch_start = ktime_get_ns();
2623                 cfqq->allocated_slice = 0;
2624                 cfqq->slice_end = 0;
2625                 cfqq->slice_dispatch = 0;
2626                 cfqq->nr_sectors = 0;
2627
2628                 cfq_clear_cfqq_wait_request(cfqq);
2629                 cfq_clear_cfqq_must_dispatch(cfqq);
2630                 cfq_clear_cfqq_must_alloc_slice(cfqq);
2631                 cfq_clear_cfqq_fifo_expire(cfqq);
2632                 cfq_mark_cfqq_slice_new(cfqq);
2633
2634                 cfq_del_timer(cfqd, cfqq);
2635         }
2636
2637         cfqd->active_queue = cfqq;
2638 }
2639
2640 /*
2641  * current cfqq expired its slice (or was too idle), select new one
2642  */
2643 static void
2644 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2645                     bool timed_out)
2646 {
2647         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
2648
2649         if (cfq_cfqq_wait_request(cfqq))
2650                 cfq_del_timer(cfqd, cfqq);
2651
2652         cfq_clear_cfqq_wait_request(cfqq);
2653         cfq_clear_cfqq_wait_busy(cfqq);
2654
2655         /*
2656          * If this cfqq is shared between multiple processes, check to
2657          * make sure that those processes are still issuing I/Os within
2658          * the mean seek distance.  If not, it may be time to break the
2659          * queues apart again.
2660          */
2661         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
2662                 cfq_mark_cfqq_split_coop(cfqq);
2663
2664         /*
2665          * store what was left of this slice, if the queue idled/timed out
2666          */
2667         if (timed_out) {
2668                 if (cfq_cfqq_slice_new(cfqq))
2669                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
2670                 else
2671                         cfqq->slice_resid = cfqq->slice_end - ktime_get_ns();
2672                 cfq_log_cfqq(cfqd, cfqq, "resid=%lld", cfqq->slice_resid);
2673         }
2674
2675         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
2676
2677         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
2678                 cfq_del_cfqq_rr(cfqd, cfqq);
2679
2680         cfq_resort_rr_list(cfqd, cfqq);
2681
2682         if (cfqq == cfqd->active_queue)
2683                 cfqd->active_queue = NULL;
2684
2685         if (cfqd->active_cic) {
2686                 put_io_context(cfqd->active_cic->icq.ioc);
2687                 cfqd->active_cic = NULL;
2688         }
2689 }
2690
2691 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
2692 {
2693         struct cfq_queue *cfqq = cfqd->active_queue;
2694
2695         if (cfqq)
2696                 __cfq_slice_expired(cfqd, cfqq, timed_out);
2697 }
2698
2699 /*
2700  * Get next queue for service. Unless we have a queue preemption,
2701  * we'll simply select the first cfqq in the service tree.
2702  */
2703 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
2704 {
2705         struct cfq_rb_root *st = st_for(cfqd->serving_group,
2706                         cfqd->serving_wl_class, cfqd->serving_wl_type);
2707
2708         if (!cfqd->rq_queued)
2709                 return NULL;
2710
2711         /* There is nothing to dispatch */
2712         if (!st)
2713                 return NULL;
2714         if (RB_EMPTY_ROOT(&st->rb.rb_root))
2715                 return NULL;
2716         return cfq_rb_first(st);
2717 }
2718
2719 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
2720 {
2721         struct cfq_group *cfqg;
2722         struct cfq_queue *cfqq;
2723         int i, j;
2724         struct cfq_rb_root *st;
2725
2726         if (!cfqd->rq_queued)
2727                 return NULL;
2728
2729         cfqg = cfq_get_next_cfqg(cfqd);
2730         if (!cfqg)
2731                 return NULL;
2732
2733         for_each_cfqg_st(cfqg, i, j, st) {
2734                 cfqq = cfq_rb_first(st);
2735                 if (cfqq)
2736                         return cfqq;
2737         }
2738         return NULL;
2739 }
2740
2741 /*
2742  * Get and set a new active queue for service.
2743  */
2744 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
2745                                               struct cfq_queue *cfqq)
2746 {
2747         if (!cfqq)
2748                 cfqq = cfq_get_next_queue(cfqd);
2749
2750         __cfq_set_active_queue(cfqd, cfqq);
2751         return cfqq;
2752 }
2753
2754 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
2755                                           struct request *rq)
2756 {
2757         if (blk_rq_pos(rq) >= cfqd->last_position)
2758                 return blk_rq_pos(rq) - cfqd->last_position;
2759         else
2760                 return cfqd->last_position - blk_rq_pos(rq);
2761 }
2762
2763 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2764                                struct request *rq)
2765 {
2766         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
2767 }
2768
2769 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
2770                                     struct cfq_queue *cur_cfqq)
2771 {
2772         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
2773         struct rb_node *parent, *node;
2774         struct cfq_queue *__cfqq;
2775         sector_t sector = cfqd->last_position;
2776
2777         if (RB_EMPTY_ROOT(root))
2778                 return NULL;
2779
2780         /*
2781          * First, if we find a request starting at the end of the last
2782          * request, choose it.
2783          */
2784         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
2785         if (__cfqq)
2786                 return __cfqq;
2787
2788         /*
2789          * If the exact sector wasn't found, the parent of the NULL leaf
2790          * will contain the closest sector.
2791          */
2792         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
2793         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2794                 return __cfqq;
2795
2796         if (blk_rq_pos(__cfqq->next_rq) < sector)
2797                 node = rb_next(&__cfqq->p_node);
2798         else
2799                 node = rb_prev(&__cfqq->p_node);
2800         if (!node)
2801                 return NULL;
2802
2803         __cfqq = rb_entry(node, struct cfq_queue, p_node);
2804         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
2805                 return __cfqq;
2806
2807         return NULL;
2808 }
2809
2810 /*
2811  * cfqd - obvious
2812  * cur_cfqq - passed in so that we don't decide that the current queue is
2813  *            closely cooperating with itself.
2814  *
2815  * So, basically we're assuming that that cur_cfqq has dispatched at least
2816  * one request, and that cfqd->last_position reflects a position on the disk
2817  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
2818  * assumption.
2819  */
2820 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
2821                                               struct cfq_queue *cur_cfqq)
2822 {
2823         struct cfq_queue *cfqq;
2824
2825         if (cfq_class_idle(cur_cfqq))
2826                 return NULL;
2827         if (!cfq_cfqq_sync(cur_cfqq))
2828                 return NULL;
2829         if (CFQQ_SEEKY(cur_cfqq))
2830                 return NULL;
2831
2832         /*
2833          * Don't search priority tree if it's the only queue in the group.
2834          */
2835         if (cur_cfqq->cfqg->nr_cfqq == 1)
2836                 return NULL;
2837
2838         /*
2839          * We should notice if some of the queues are cooperating, eg
2840          * working closely on the same area of the disk. In that case,
2841          * we can group them together and don't waste time idling.
2842          */
2843         cfqq = cfqq_close(cfqd, cur_cfqq);
2844         if (!cfqq)
2845                 return NULL;
2846
2847         /* If new queue belongs to different cfq_group, don't choose it */
2848         if (cur_cfqq->cfqg != cfqq->cfqg)
2849                 return NULL;
2850
2851         /*
2852          * It only makes sense to merge sync queues.
2853          */
2854         if (!cfq_cfqq_sync(cfqq))
2855                 return NULL;
2856         if (CFQQ_SEEKY(cfqq))
2857                 return NULL;
2858
2859         /*
2860          * Do not merge queues of different priority classes
2861          */
2862         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
2863                 return NULL;
2864
2865         return cfqq;
2866 }
2867
2868 /*
2869  * Determine whether we should enforce idle window for this queue.
2870  */
2871
2872 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2873 {
2874         enum wl_class_t wl_class = cfqq_class(cfqq);
2875         struct cfq_rb_root *st = cfqq->service_tree;
2876
2877         BUG_ON(!st);
2878         BUG_ON(!st->count);
2879
2880         if (!cfqd->cfq_slice_idle)
2881                 return false;
2882
2883         /* We never do for idle class queues. */
2884         if (wl_class == IDLE_WORKLOAD)
2885                 return false;
2886
2887         /* We do for queues that were marked with idle window flag. */
2888         if (cfq_cfqq_idle_window(cfqq) &&
2889            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2890                 return true;
2891
2892         /*
2893          * Otherwise, we do only if they are the last ones
2894          * in their service tree.
2895          */
2896         if (st->count == 1 && cfq_cfqq_sync(cfqq) &&
2897            !cfq_io_thinktime_big(cfqd, &st->ttime, false))
2898                 return true;
2899         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d", st->count);
2900         return false;
2901 }
2902
2903 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2904 {
2905         struct cfq_queue *cfqq = cfqd->active_queue;
2906         struct cfq_rb_root *st = cfqq->service_tree;
2907         struct cfq_io_cq *cic;
2908         u64 sl, group_idle = 0;
2909         u64 now = ktime_get_ns();
2910
2911         /*
2912          * SSD device without seek penalty, disable idling. But only do so
2913          * for devices that support queuing, otherwise we still have a problem
2914          * with sync vs async workloads.
2915          */
2916         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag &&
2917                 !cfqd->cfq_group_idle)
2918                 return;
2919
2920         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2921         WARN_ON(cfq_cfqq_slice_new(cfqq));
2922
2923         /*
2924          * idle is disabled, either manually or by past process history
2925          */
2926         if (!cfq_should_idle(cfqd, cfqq)) {
2927                 /* no queue idling. Check for group idling */
2928                 if (cfqd->cfq_group_idle)
2929                         group_idle = cfqd->cfq_group_idle;
2930                 else
2931                         return;
2932         }
2933
2934         /*
2935          * still active requests from this queue, don't idle
2936          */
2937         if (cfqq->dispatched)
2938                 return;
2939
2940         /*
2941          * task has exited, don't wait
2942          */
2943         cic = cfqd->active_cic;
2944         if (!cic || !atomic_read(&cic->icq.ioc->active_ref))
2945                 return;
2946
2947         /*
2948          * If our average think time is larger than the remaining time
2949          * slice, then don't idle. This avoids overrunning the allotted
2950          * time slice.
2951          */
2952         if (sample_valid(cic->ttime.ttime_samples) &&
2953             (cfqq->slice_end - now < cic->ttime.ttime_mean)) {
2954                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%llu",
2955                              cic->ttime.ttime_mean);
2956                 return;
2957         }
2958
2959         /*
2960          * There are other queues in the group or this is the only group and
2961          * it has too big thinktime, don't do group idle.
2962          */
2963         if (group_idle &&
2964             (cfqq->cfqg->nr_cfqq > 1 ||
2965              cfq_io_thinktime_big(cfqd, &st->ttime, true)))
2966                 return;
2967
2968         cfq_mark_cfqq_wait_request(cfqq);
2969
2970         if (group_idle)
2971                 sl = cfqd->cfq_group_idle;
2972         else
2973                 sl = cfqd->cfq_slice_idle;
2974
2975         hrtimer_start(&cfqd->idle_slice_timer, ns_to_ktime(sl),
2976                       HRTIMER_MODE_REL);
2977         cfqg_stats_set_start_idle_time(cfqq->cfqg);
2978         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %llu group_idle: %d", sl,
2979                         group_idle ? 1 : 0);
2980 }
2981
2982 /*
2983  * Move request from internal lists to the request queue dispatch list.
2984  */
2985 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2986 {
2987         struct cfq_data *cfqd = q->elevator->elevator_data;
2988         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2989
2990         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2991
2992         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2993         cfq_remove_request(rq);
2994         cfqq->dispatched++;
2995         (RQ_CFQG(rq))->dispatched++;
2996         elv_dispatch_sort(q, rq);
2997
2998         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2999         cfqq->nr_sectors += blk_rq_sectors(rq);
3000 }
3001
3002 /*
3003  * return expired entry, or NULL to just start from scratch in rbtree
3004  */
3005 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
3006 {
3007         struct request *rq = NULL;
3008
3009         if (cfq_cfqq_fifo_expire(cfqq))
3010                 return NULL;
3011
3012         cfq_mark_cfqq_fifo_expire(cfqq);
3013
3014         if (list_empty(&cfqq->fifo))
3015                 return NULL;
3016
3017         rq = rq_entry_fifo(cfqq->fifo.next);
3018         if (ktime_get_ns() < rq->fifo_time)
3019                 rq = NULL;
3020
3021         return rq;
3022 }
3023
3024 static inline int
3025 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3026 {
3027         const int base_rq = cfqd->cfq_slice_async_rq;
3028
3029         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
3030
3031         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
3032 }
3033
3034 /*
3035  * Must be called with the queue_lock held.
3036  */
3037 static int cfqq_process_refs(struct cfq_queue *cfqq)
3038 {
3039         int process_refs, io_refs;
3040
3041         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
3042         process_refs = cfqq->ref - io_refs;
3043         BUG_ON(process_refs < 0);
3044         return process_refs;
3045 }
3046
3047 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
3048 {
3049         int process_refs, new_process_refs;
3050         struct cfq_queue *__cfqq;
3051
3052         /*
3053          * If there are no process references on the new_cfqq, then it is
3054          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
3055          * chain may have dropped their last reference (not just their
3056          * last process reference).
3057          */
3058         if (!cfqq_process_refs(new_cfqq))
3059                 return;
3060
3061         /* Avoid a circular list and skip interim queue merges */
3062         while ((__cfqq = new_cfqq->new_cfqq)) {
3063                 if (__cfqq == cfqq)
3064                         return;
3065                 new_cfqq = __cfqq;
3066         }
3067
3068         process_refs = cfqq_process_refs(cfqq);
3069         new_process_refs = cfqq_process_refs(new_cfqq);
3070         /*
3071          * If the process for the cfqq has gone away, there is no
3072          * sense in merging the queues.
3073          */
3074         if (process_refs == 0 || new_process_refs == 0)
3075                 return;
3076
3077         /*
3078          * Merge in the direction of the lesser amount of work.
3079          */
3080         if (new_process_refs >= process_refs) {
3081                 cfqq->new_cfqq = new_cfqq;
3082                 new_cfqq->ref += process_refs;
3083         } else {
3084                 new_cfqq->new_cfqq = cfqq;
3085                 cfqq->ref += new_process_refs;
3086         }
3087 }
3088
3089 static enum wl_type_t cfq_choose_wl_type(struct cfq_data *cfqd,
3090                         struct cfq_group *cfqg, enum wl_class_t wl_class)
3091 {
3092         struct cfq_queue *queue;
3093         int i;
3094         bool key_valid = false;
3095         u64 lowest_key = 0;
3096         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
3097
3098         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
3099                 /* select the one with lowest rb_key */
3100                 queue = cfq_rb_first(st_for(cfqg, wl_class, i));
3101                 if (queue &&
3102                     (!key_valid || queue->rb_key < lowest_key)) {
3103                         lowest_key = queue->rb_key;
3104                         cur_best = i;
3105                         key_valid = true;
3106                 }
3107         }
3108
3109         return cur_best;
3110 }
3111
3112 static void
3113 choose_wl_class_and_type(struct cfq_data *cfqd, struct cfq_group *cfqg)
3114 {
3115         u64 slice;
3116         unsigned count;
3117         struct cfq_rb_root *st;
3118         u64 group_slice;
3119         enum wl_class_t original_class = cfqd->serving_wl_class;
3120         u64 now = ktime_get_ns();
3121
3122         /* Choose next priority. RT > BE > IDLE */
3123         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
3124                 cfqd->serving_wl_class = RT_WORKLOAD;
3125         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
3126                 cfqd->serving_wl_class = BE_WORKLOAD;
3127         else {
3128                 cfqd->serving_wl_class = IDLE_WORKLOAD;
3129                 cfqd->workload_expires = now + jiffies_to_nsecs(1);
3130                 return;
3131         }
3132
3133         if (original_class != cfqd->serving_wl_class)
3134                 goto new_workload;
3135
3136         /*
3137          * For RT and BE, we have to choose also the type
3138          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
3139          * expiration time
3140          */
3141         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3142         count = st->count;
3143
3144         /*
3145          * check workload expiration, and that we still have other queues ready
3146          */
3147         if (count && !(now > cfqd->workload_expires))
3148                 return;
3149
3150 new_workload:
3151         /* otherwise select new workload type */
3152         cfqd->serving_wl_type = cfq_choose_wl_type(cfqd, cfqg,
3153                                         cfqd->serving_wl_class);
3154         st = st_for(cfqg, cfqd->serving_wl_class, cfqd->serving_wl_type);
3155         count = st->count;
3156
3157         /*
3158          * the workload slice is computed as a fraction of target latency
3159          * proportional to the number of queues in that workload, over
3160          * all the queues in the same priority class
3161          */
3162         group_slice = cfq_group_slice(cfqd, cfqg);
3163
3164         slice = div_u64(group_slice * count,
3165                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_wl_class],
3166                       cfq_group_busy_queues_wl(cfqd->serving_wl_class, cfqd,
3167                                         cfqg)));
3168
3169         if (cfqd->serving_wl_type == ASYNC_WORKLOAD) {
3170                 u64 tmp;
3171
3172                 /*
3173                  * Async queues are currently system wide. Just taking
3174                  * proportion of queues with-in same group will lead to higher
3175                  * async ratio system wide as generally root group is going
3176                  * to have higher weight. A more accurate thing would be to
3177                  * calculate system wide asnc/sync ratio.
3178                  */
3179                 tmp = cfqd->cfq_target_latency *
3180                         cfqg_busy_async_queues(cfqd, cfqg);
3181                 tmp = div_u64(tmp, cfqd->busy_queues);
3182                 slice = min_t(u64, slice, tmp);
3183
3184                 /* async workload slice is scaled down according to
3185                  * the sync/async slice ratio. */
3186                 slice = div64_u64(slice*cfqd->cfq_slice[0], cfqd->cfq_slice[1]);
3187         } else
3188                 /* sync workload slice is at least 2 * cfq_slice_idle */
3189                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
3190
3191         slice = max_t(u64, slice, CFQ_MIN_TT);
3192         cfq_log(cfqd, "workload slice:%llu", slice);
3193         cfqd->workload_expires = now + slice;
3194 }
3195
3196 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
3197 {
3198         struct cfq_rb_root *st = &cfqd->grp_service_tree;
3199         struct cfq_group *cfqg;
3200
3201         if (RB_EMPTY_ROOT(&st->rb.rb_root))
3202                 return NULL;
3203         cfqg = cfq_rb_first_group(st);
3204         update_min_vdisktime(st);
3205         return cfqg;
3206 }
3207
3208 static void cfq_choose_cfqg(struct cfq_data *cfqd)
3209 {
3210         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
3211         u64 now = ktime_get_ns();
3212
3213         cfqd->serving_group = cfqg;
3214
3215         /* Restore the workload type data */
3216         if (cfqg->saved_wl_slice) {
3217                 cfqd->workload_expires = now + cfqg->saved_wl_slice;
3218                 cfqd->serving_wl_type = cfqg->saved_wl_type;
3219                 cfqd->serving_wl_class = cfqg->saved_wl_class;
3220         } else
3221                 cfqd->workload_expires = now - 1;
3222
3223         choose_wl_class_and_type(cfqd, cfqg);
3224 }
3225
3226 /*
3227  * Select a queue for service. If we have a current active queue,
3228  * check whether to continue servicing it, or retrieve and set a new one.
3229  */
3230 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
3231 {
3232         struct cfq_queue *cfqq, *new_cfqq = NULL;
3233         u64 now = ktime_get_ns();
3234
3235         cfqq = cfqd->active_queue;
3236         if (!cfqq)
3237                 goto new_queue;
3238
3239         if (!cfqd->rq_queued)
3240                 return NULL;
3241
3242         /*
3243          * We were waiting for group to get backlogged. Expire the queue
3244          */
3245         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
3246                 goto expire;
3247
3248         /*
3249          * The active queue has run out of time, expire it and select new.
3250          */
3251         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
3252                 /*
3253                  * If slice had not expired at the completion of last request
3254                  * we might not have turned on wait_busy flag. Don't expire
3255                  * the queue yet. Allow the group to get backlogged.
3256                  *
3257                  * The very fact that we have used the slice, that means we
3258                  * have been idling all along on this queue and it should be
3259                  * ok to wait for this request to complete.
3260                  */
3261                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
3262                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3263                         cfqq = NULL;
3264                         goto keep_queue;
3265                 } else
3266                         goto check_group_idle;
3267         }
3268
3269         /*
3270          * The active queue has requests and isn't expired, allow it to
3271          * dispatch.
3272          */
3273         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3274                 goto keep_queue;
3275
3276         /*
3277          * If another queue has a request waiting within our mean seek
3278          * distance, let it run.  The expire code will check for close
3279          * cooperators and put the close queue at the front of the service
3280          * tree.  If possible, merge the expiring queue with the new cfqq.
3281          */
3282         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
3283         if (new_cfqq) {
3284                 if (!cfqq->new_cfqq)
3285                         cfq_setup_merge(cfqq, new_cfqq);
3286                 goto expire;
3287         }
3288
3289         /*
3290          * No requests pending. If the active queue still has requests in
3291          * flight or is idling for a new request, allow either of these
3292          * conditions to happen (or time out) before selecting a new queue.
3293          */
3294         if (hrtimer_active(&cfqd->idle_slice_timer)) {
3295                 cfqq = NULL;
3296                 goto keep_queue;
3297         }
3298
3299         /*
3300          * This is a deep seek queue, but the device is much faster than
3301          * the queue can deliver, don't idle
3302          **/
3303         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
3304             (cfq_cfqq_slice_new(cfqq) ||
3305             (cfqq->slice_end - now > now - cfqq->slice_start))) {
3306                 cfq_clear_cfqq_deep(cfqq);
3307                 cfq_clear_cfqq_idle_window(cfqq);
3308         }
3309
3310         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
3311                 cfqq = NULL;
3312                 goto keep_queue;
3313         }
3314
3315         /*
3316          * If group idle is enabled and there are requests dispatched from
3317          * this group, wait for requests to complete.
3318          */
3319 check_group_idle:
3320         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
3321             cfqq->cfqg->dispatched &&
3322             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
3323                 cfqq = NULL;
3324                 goto keep_queue;
3325         }
3326
3327 expire:
3328         cfq_slice_expired(cfqd, 0);
3329 new_queue:
3330         /*
3331          * Current queue expired. Check if we have to switch to a new
3332          * service tree
3333          */
3334         if (!new_cfqq)
3335                 cfq_choose_cfqg(cfqd);
3336
3337         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
3338 keep_queue:
3339         return cfqq;
3340 }
3341
3342 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
3343 {
3344         int dispatched = 0;
3345
3346         while (cfqq->next_rq) {
3347                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
3348                 dispatched++;
3349         }
3350
3351         BUG_ON(!list_empty(&cfqq->fifo));
3352
3353         /* By default cfqq is not expired if it is empty. Do it explicitly */
3354         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
3355         return dispatched;
3356 }
3357
3358 /*
3359  * Drain our current requests. Used for barriers and when switching
3360  * io schedulers on-the-fly.
3361  */
3362 static int cfq_forced_dispatch(struct cfq_data *cfqd)
3363 {