block: Make cfq_target_latency tunable through sysfs.
[linux-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/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk.h"
18 #include "cfq.h"
19
20 /*
21  * tunables
22  */
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
37
38 /*
39  * offset from end of service tree
40  */
41 #define CFQ_IDLE_DELAY          (HZ / 5)
42
43 /*
44  * below this threshold, we consider thinktime immediate
45  */
46 #define CFQ_MIN_TT              (2)
47
48 #define CFQ_SLICE_SCALE         (5)
49 #define CFQ_HW_QUEUE_MIN        (5)
50 #define CFQ_SERVICE_SHIFT       12
51
52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
56
57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
60
61 static struct kmem_cache *cfq_pool;
62
63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66
67 #define sample_valid(samples)   ((samples) > 80)
68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
69
70 struct cfq_ttime {
71         unsigned long last_end_request;
72
73         unsigned long ttime_total;
74         unsigned long ttime_samples;
75         unsigned long ttime_mean;
76 };
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90         struct cfq_ttime ttime;
91 };
92 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
93                         .ttime = {.last_end_request = jiffies,},}
94
95 /*
96  * Per process-grouping structure
97  */
98 struct cfq_queue {
99         /* reference count */
100         int ref;
101         /* various state flags, see below */
102         unsigned int flags;
103         /* parent cfq_data */
104         struct cfq_data *cfqd;
105         /* service_tree member */
106         struct rb_node rb_node;
107         /* service_tree key */
108         unsigned long rb_key;
109         /* prio tree member */
110         struct rb_node p_node;
111         /* prio tree root we belong to, if any */
112         struct rb_root *p_root;
113         /* sorted list of pending requests */
114         struct rb_root sort_list;
115         /* if fifo isn't expired, next request to serve */
116         struct request *next_rq;
117         /* requests queued in sort_list */
118         int queued[2];
119         /* currently allocated requests */
120         int allocated[2];
121         /* fifo list of requests in sort_list */
122         struct list_head fifo;
123
124         /* time when queue got scheduled in to dispatch first request. */
125         unsigned long dispatch_start;
126         unsigned int allocated_slice;
127         unsigned int slice_dispatch;
128         /* time when first request from queue completed and slice started. */
129         unsigned long slice_start;
130         unsigned long slice_end;
131         long slice_resid;
132
133         /* pending priority requests */
134         int prio_pending;
135         /* number of requests that are on the dispatch list or inside driver */
136         int dispatched;
137
138         /* io prio of this group */
139         unsigned short ioprio, org_ioprio;
140         unsigned short ioprio_class;
141
142         pid_t pid;
143
144         u32 seek_history;
145         sector_t last_request_pos;
146
147         struct cfq_rb_root *service_tree;
148         struct cfq_queue *new_cfqq;
149         struct cfq_group *cfqg;
150         /* Number of sectors dispatched from queue in single dispatch round */
151         unsigned long nr_sectors;
152 };
153
154 /*
155  * First index in the service_trees.
156  * IDLE is handled separately, so it has negative index
157  */
158 enum wl_prio_t {
159         BE_WORKLOAD = 0,
160         RT_WORKLOAD = 1,
161         IDLE_WORKLOAD = 2,
162         CFQ_PRIO_NR,
163 };
164
165 /*
166  * Second index in the service_trees.
167  */
168 enum wl_type_t {
169         ASYNC_WORKLOAD = 0,
170         SYNC_NOIDLE_WORKLOAD = 1,
171         SYNC_WORKLOAD = 2
172 };
173
174 /* This is per cgroup per device grouping structure */
175 struct cfq_group {
176         /* group service_tree member */
177         struct rb_node rb_node;
178
179         /* group service_tree key */
180         u64 vdisktime;
181         unsigned int weight;
182         unsigned int new_weight;
183         bool needs_update;
184
185         /* number of cfqq currently on this group */
186         int nr_cfqq;
187
188         /*
189          * Per group busy queues average. Useful for workload slice calc. We
190          * create the array for each prio class but at run time it is used
191          * only for RT and BE class and slot for IDLE class remains unused.
192          * This is primarily done to avoid confusion and a gcc warning.
193          */
194         unsigned int busy_queues_avg[CFQ_PRIO_NR];
195         /*
196          * rr lists of queues with requests. We maintain service trees for
197          * RT and BE classes. These trees are subdivided in subclasses
198          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199          * class there is no subclassification and all the cfq queues go on
200          * a single tree service_tree_idle.
201          * Counts are embedded in the cfq_rb_root
202          */
203         struct cfq_rb_root service_trees[2][3];
204         struct cfq_rb_root service_tree_idle;
205
206         unsigned long saved_workload_slice;
207         enum wl_type_t saved_workload;
208         enum wl_prio_t saved_serving_prio;
209         struct blkio_group blkg;
210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
211         struct hlist_node cfqd_node;
212         int ref;
213 #endif
214         /* number of requests that are on the dispatch list or inside driver */
215         int dispatched;
216         struct cfq_ttime ttime;
217 };
218
219 struct cfq_io_cq {
220         struct io_cq            icq;            /* must be the first member */
221         struct cfq_queue        *cfqq[2];
222         struct cfq_ttime        ttime;
223 };
224
225 /*
226  * Per block device queue structure
227  */
228 struct cfq_data {
229         struct request_queue *queue;
230         /* Root service tree for cfq_groups */
231         struct cfq_rb_root grp_service_tree;
232         struct cfq_group root_group;
233
234         /*
235          * The priority currently being served
236          */
237         enum wl_prio_t serving_prio;
238         enum wl_type_t serving_type;
239         unsigned long workload_expires;
240         struct cfq_group *serving_group;
241
242         /*
243          * Each priority tree is sorted by next_request position.  These
244          * trees are used when determining if two or more queues are
245          * interleaving requests (see cfq_close_cooperator).
246          */
247         struct rb_root prio_trees[CFQ_PRIO_LISTS];
248
249         unsigned int busy_queues;
250         unsigned int busy_sync_queues;
251
252         int rq_in_driver;
253         int rq_in_flight[2];
254
255         /*
256          * queue-depth detection
257          */
258         int rq_queued;
259         int hw_tag;
260         /*
261          * hw_tag can be
262          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
263          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
264          *  0 => no NCQ
265          */
266         int hw_tag_est_depth;
267         unsigned int hw_tag_samples;
268
269         /*
270          * idle window management
271          */
272         struct timer_list idle_slice_timer;
273         struct work_struct unplug_work;
274
275         struct cfq_queue *active_queue;
276         struct cfq_io_cq *active_cic;
277
278         /*
279          * async queue for each priority case
280          */
281         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
282         struct cfq_queue *async_idle_cfqq;
283
284         sector_t last_position;
285
286         /*
287          * tunables, see top of file
288          */
289         unsigned int cfq_quantum;
290         unsigned int cfq_fifo_expire[2];
291         unsigned int cfq_back_penalty;
292         unsigned int cfq_back_max;
293         unsigned int cfq_slice[2];
294         unsigned int cfq_slice_async_rq;
295         unsigned int cfq_slice_idle;
296         unsigned int cfq_group_idle;
297         unsigned int cfq_latency;
298         unsigned int cfq_target_latency;
299
300         /*
301          * Fallback dummy cfqq for extreme OOM conditions
302          */
303         struct cfq_queue oom_cfqq;
304
305         unsigned long last_delayed_sync;
306
307         /* List of cfq groups being managed on this device*/
308         struct hlist_head cfqg_list;
309
310         /* Number of groups which are on blkcg->blkg_list */
311         unsigned int nr_blkcg_linked_grps;
312 };
313
314 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
315
316 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
317                                             enum wl_prio_t prio,
318                                             enum wl_type_t type)
319 {
320         if (!cfqg)
321                 return NULL;
322
323         if (prio == IDLE_WORKLOAD)
324                 return &cfqg->service_tree_idle;
325
326         return &cfqg->service_trees[prio][type];
327 }
328
329 enum cfqq_state_flags {
330         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
331         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
332         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
333         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
334         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
335         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
336         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
337         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
338         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
339         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
340         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
341         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
342         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
343 };
344
345 #define CFQ_CFQQ_FNS(name)                                              \
346 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
347 {                                                                       \
348         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
349 }                                                                       \
350 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
351 {                                                                       \
352         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
353 }                                                                       \
354 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
355 {                                                                       \
356         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
357 }
358
359 CFQ_CFQQ_FNS(on_rr);
360 CFQ_CFQQ_FNS(wait_request);
361 CFQ_CFQQ_FNS(must_dispatch);
362 CFQ_CFQQ_FNS(must_alloc_slice);
363 CFQ_CFQQ_FNS(fifo_expire);
364 CFQ_CFQQ_FNS(idle_window);
365 CFQ_CFQQ_FNS(prio_changed);
366 CFQ_CFQQ_FNS(slice_new);
367 CFQ_CFQQ_FNS(sync);
368 CFQ_CFQQ_FNS(coop);
369 CFQ_CFQQ_FNS(split_coop);
370 CFQ_CFQQ_FNS(deep);
371 CFQ_CFQQ_FNS(wait_busy);
372 #undef CFQ_CFQQ_FNS
373
374 #ifdef CONFIG_CFQ_GROUP_IOSCHED
375 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
376         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
377                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
378                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
379
380 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
381         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
382                                 blkg_path(&(cfqg)->blkg), ##args)       \
383
384 #else
385 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
386         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
387 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
388 #endif
389 #define cfq_log(cfqd, fmt, args...)     \
390         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
391
392 /* Traverses through cfq group service trees */
393 #define for_each_cfqg_st(cfqg, i, j, st) \
394         for (i = 0; i <= IDLE_WORKLOAD; i++) \
395                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
396                         : &cfqg->service_tree_idle; \
397                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
398                         (i == IDLE_WORKLOAD && j == 0); \
399                         j++, st = i < IDLE_WORKLOAD ? \
400                         &cfqg->service_trees[i][j]: NULL) \
401
402 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
403         struct cfq_ttime *ttime, bool group_idle)
404 {
405         unsigned long slice;
406         if (!sample_valid(ttime->ttime_samples))
407                 return false;
408         if (group_idle)
409                 slice = cfqd->cfq_group_idle;
410         else
411                 slice = cfqd->cfq_slice_idle;
412         return ttime->ttime_mean > slice;
413 }
414
415 static inline bool iops_mode(struct cfq_data *cfqd)
416 {
417         /*
418          * If we are not idling on queues and it is a NCQ drive, parallel
419          * execution of requests is on and measuring time is not possible
420          * in most of the cases until and unless we drive shallower queue
421          * depths and that becomes a performance bottleneck. In such cases
422          * switch to start providing fairness in terms of number of IOs.
423          */
424         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
425                 return true;
426         else
427                 return false;
428 }
429
430 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
431 {
432         if (cfq_class_idle(cfqq))
433                 return IDLE_WORKLOAD;
434         if (cfq_class_rt(cfqq))
435                 return RT_WORKLOAD;
436         return BE_WORKLOAD;
437 }
438
439
440 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
441 {
442         if (!cfq_cfqq_sync(cfqq))
443                 return ASYNC_WORKLOAD;
444         if (!cfq_cfqq_idle_window(cfqq))
445                 return SYNC_NOIDLE_WORKLOAD;
446         return SYNC_WORKLOAD;
447 }
448
449 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
450                                         struct cfq_data *cfqd,
451                                         struct cfq_group *cfqg)
452 {
453         if (wl == IDLE_WORKLOAD)
454                 return cfqg->service_tree_idle.count;
455
456         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
457                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
458                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
459 }
460
461 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
462                                         struct cfq_group *cfqg)
463 {
464         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
465                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
466 }
467
468 static void cfq_dispatch_insert(struct request_queue *, struct request *);
469 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
470                                        struct io_context *, gfp_t);
471
472 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
473 {
474         /* cic->icq is the first member, %NULL will convert to %NULL */
475         return container_of(icq, struct cfq_io_cq, icq);
476 }
477
478 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
479                                                struct io_context *ioc)
480 {
481         if (ioc)
482                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
483         return NULL;
484 }
485
486 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
487 {
488         return cic->cfqq[is_sync];
489 }
490
491 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
492                                 bool is_sync)
493 {
494         cic->cfqq[is_sync] = cfqq;
495 }
496
497 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
498 {
499         return cic->icq.q->elevator->elevator_data;
500 }
501
502 /*
503  * We regard a request as SYNC, if it's either a read or has the SYNC bit
504  * set (in which case it could also be direct WRITE).
505  */
506 static inline bool cfq_bio_sync(struct bio *bio)
507 {
508         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
509 }
510
511 /*
512  * scheduler run of queue, if there are requests pending and no one in the
513  * driver that will restart queueing
514  */
515 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
516 {
517         if (cfqd->busy_queues) {
518                 cfq_log(cfqd, "schedule dispatch");
519                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
520         }
521 }
522
523 /*
524  * Scale schedule slice based on io priority. Use the sync time slice only
525  * if a queue is marked sync and has sync io queued. A sync queue with async
526  * io only, should not get full sync slice length.
527  */
528 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
529                                  unsigned short prio)
530 {
531         const int base_slice = cfqd->cfq_slice[sync];
532
533         WARN_ON(prio >= IOPRIO_BE_NR);
534
535         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
536 }
537
538 static inline int
539 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
540 {
541         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
542 }
543
544 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
545 {
546         u64 d = delta << CFQ_SERVICE_SHIFT;
547
548         d = d * BLKIO_WEIGHT_DEFAULT;
549         do_div(d, cfqg->weight);
550         return d;
551 }
552
553 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
554 {
555         s64 delta = (s64)(vdisktime - min_vdisktime);
556         if (delta > 0)
557                 min_vdisktime = vdisktime;
558
559         return min_vdisktime;
560 }
561
562 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
563 {
564         s64 delta = (s64)(vdisktime - min_vdisktime);
565         if (delta < 0)
566                 min_vdisktime = vdisktime;
567
568         return min_vdisktime;
569 }
570
571 static void update_min_vdisktime(struct cfq_rb_root *st)
572 {
573         struct cfq_group *cfqg;
574
575         if (st->left) {
576                 cfqg = rb_entry_cfqg(st->left);
577                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
578                                                   cfqg->vdisktime);
579         }
580 }
581
582 /*
583  * get averaged number of queues of RT/BE priority.
584  * average is updated, with a formula that gives more weight to higher numbers,
585  * to quickly follows sudden increases and decrease slowly
586  */
587
588 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
589                                         struct cfq_group *cfqg, bool rt)
590 {
591         unsigned min_q, max_q;
592         unsigned mult  = cfq_hist_divisor - 1;
593         unsigned round = cfq_hist_divisor / 2;
594         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
595
596         min_q = min(cfqg->busy_queues_avg[rt], busy);
597         max_q = max(cfqg->busy_queues_avg[rt], busy);
598         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
599                 cfq_hist_divisor;
600         return cfqg->busy_queues_avg[rt];
601 }
602
603 static inline unsigned
604 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
605 {
606         struct cfq_rb_root *st = &cfqd->grp_service_tree;
607
608         return cfqd->cfq_target_latency * cfqg->weight / st->total_weight;
609 }
610
611 static inline unsigned
612 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
613 {
614         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
615         if (cfqd->cfq_latency) {
616                 /*
617                  * interested queues (we consider only the ones with the same
618                  * priority class in the cfq group)
619                  */
620                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
621                                                 cfq_class_rt(cfqq));
622                 unsigned sync_slice = cfqd->cfq_slice[1];
623                 unsigned expect_latency = sync_slice * iq;
624                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
625
626                 if (expect_latency > group_slice) {
627                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
628                         /* scale low_slice according to IO priority
629                          * and sync vs async */
630                         unsigned low_slice =
631                                 min(slice, base_low_slice * slice / sync_slice);
632                         /* the adapted slice value is scaled to fit all iqs
633                          * into the target latency */
634                         slice = max(slice * group_slice / expect_latency,
635                                     low_slice);
636                 }
637         }
638         return slice;
639 }
640
641 static inline void
642 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
643 {
644         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
645
646         cfqq->slice_start = jiffies;
647         cfqq->slice_end = jiffies + slice;
648         cfqq->allocated_slice = slice;
649         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
650 }
651
652 /*
653  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
654  * isn't valid until the first request from the dispatch is activated
655  * and the slice time set.
656  */
657 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
658 {
659         if (cfq_cfqq_slice_new(cfqq))
660                 return false;
661         if (time_before(jiffies, cfqq->slice_end))
662                 return false;
663
664         return true;
665 }
666
667 /*
668  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
669  * We choose the request that is closest to the head right now. Distance
670  * behind the head is penalized and only allowed to a certain extent.
671  */
672 static struct request *
673 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
674 {
675         sector_t s1, s2, d1 = 0, d2 = 0;
676         unsigned long back_max;
677 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
678 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
679         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
680
681         if (rq1 == NULL || rq1 == rq2)
682                 return rq2;
683         if (rq2 == NULL)
684                 return rq1;
685
686         if (rq_is_sync(rq1) != rq_is_sync(rq2))
687                 return rq_is_sync(rq1) ? rq1 : rq2;
688
689         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
690                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
691
692         s1 = blk_rq_pos(rq1);
693         s2 = blk_rq_pos(rq2);
694
695         /*
696          * by definition, 1KiB is 2 sectors
697          */
698         back_max = cfqd->cfq_back_max * 2;
699
700         /*
701          * Strict one way elevator _except_ in the case where we allow
702          * short backward seeks which are biased as twice the cost of a
703          * similar forward seek.
704          */
705         if (s1 >= last)
706                 d1 = s1 - last;
707         else if (s1 + back_max >= last)
708                 d1 = (last - s1) * cfqd->cfq_back_penalty;
709         else
710                 wrap |= CFQ_RQ1_WRAP;
711
712         if (s2 >= last)
713                 d2 = s2 - last;
714         else if (s2 + back_max >= last)
715                 d2 = (last - s2) * cfqd->cfq_back_penalty;
716         else
717                 wrap |= CFQ_RQ2_WRAP;
718
719         /* Found required data */
720
721         /*
722          * By doing switch() on the bit mask "wrap" we avoid having to
723          * check two variables for all permutations: --> faster!
724          */
725         switch (wrap) {
726         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
727                 if (d1 < d2)
728                         return rq1;
729                 else if (d2 < d1)
730                         return rq2;
731                 else {
732                         if (s1 >= s2)
733                                 return rq1;
734                         else
735                                 return rq2;
736                 }
737
738         case CFQ_RQ2_WRAP:
739                 return rq1;
740         case CFQ_RQ1_WRAP:
741                 return rq2;
742         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
743         default:
744                 /*
745                  * Since both rqs are wrapped,
746                  * start with the one that's further behind head
747                  * (--> only *one* back seek required),
748                  * since back seek takes more time than forward.
749                  */
750                 if (s1 <= s2)
751                         return rq1;
752                 else
753                         return rq2;
754         }
755 }
756
757 /*
758  * The below is leftmost cache rbtree addon
759  */
760 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
761 {
762         /* Service tree is empty */
763         if (!root->count)
764                 return NULL;
765
766         if (!root->left)
767                 root->left = rb_first(&root->rb);
768
769         if (root->left)
770                 return rb_entry(root->left, struct cfq_queue, rb_node);
771
772         return NULL;
773 }
774
775 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
776 {
777         if (!root->left)
778                 root->left = rb_first(&root->rb);
779
780         if (root->left)
781                 return rb_entry_cfqg(root->left);
782
783         return NULL;
784 }
785
786 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
787 {
788         rb_erase(n, root);
789         RB_CLEAR_NODE(n);
790 }
791
792 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
793 {
794         if (root->left == n)
795                 root->left = NULL;
796         rb_erase_init(n, &root->rb);
797         --root->count;
798 }
799
800 /*
801  * would be nice to take fifo expire time into account as well
802  */
803 static struct request *
804 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
805                   struct request *last)
806 {
807         struct rb_node *rbnext = rb_next(&last->rb_node);
808         struct rb_node *rbprev = rb_prev(&last->rb_node);
809         struct request *next = NULL, *prev = NULL;
810
811         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
812
813         if (rbprev)
814                 prev = rb_entry_rq(rbprev);
815
816         if (rbnext)
817                 next = rb_entry_rq(rbnext);
818         else {
819                 rbnext = rb_first(&cfqq->sort_list);
820                 if (rbnext && rbnext != &last->rb_node)
821                         next = rb_entry_rq(rbnext);
822         }
823
824         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
825 }
826
827 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
828                                       struct cfq_queue *cfqq)
829 {
830         /*
831          * just an approximation, should be ok.
832          */
833         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
834                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
835 }
836
837 static inline s64
838 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
839 {
840         return cfqg->vdisktime - st->min_vdisktime;
841 }
842
843 static void
844 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
845 {
846         struct rb_node **node = &st->rb.rb_node;
847         struct rb_node *parent = NULL;
848         struct cfq_group *__cfqg;
849         s64 key = cfqg_key(st, cfqg);
850         int left = 1;
851
852         while (*node != NULL) {
853                 parent = *node;
854                 __cfqg = rb_entry_cfqg(parent);
855
856                 if (key < cfqg_key(st, __cfqg))
857                         node = &parent->rb_left;
858                 else {
859                         node = &parent->rb_right;
860                         left = 0;
861                 }
862         }
863
864         if (left)
865                 st->left = &cfqg->rb_node;
866
867         rb_link_node(&cfqg->rb_node, parent, node);
868         rb_insert_color(&cfqg->rb_node, &st->rb);
869 }
870
871 static void
872 cfq_update_group_weight(struct cfq_group *cfqg)
873 {
874         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
875         if (cfqg->needs_update) {
876                 cfqg->weight = cfqg->new_weight;
877                 cfqg->needs_update = false;
878         }
879 }
880
881 static void
882 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
883 {
884         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
885
886         cfq_update_group_weight(cfqg);
887         __cfq_group_service_tree_add(st, cfqg);
888         st->total_weight += cfqg->weight;
889 }
890
891 static void
892 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
893 {
894         struct cfq_rb_root *st = &cfqd->grp_service_tree;
895         struct cfq_group *__cfqg;
896         struct rb_node *n;
897
898         cfqg->nr_cfqq++;
899         if (!RB_EMPTY_NODE(&cfqg->rb_node))
900                 return;
901
902         /*
903          * Currently put the group at the end. Later implement something
904          * so that groups get lesser vtime based on their weights, so that
905          * if group does not loose all if it was not continuously backlogged.
906          */
907         n = rb_last(&st->rb);
908         if (n) {
909                 __cfqg = rb_entry_cfqg(n);
910                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
911         } else
912                 cfqg->vdisktime = st->min_vdisktime;
913         cfq_group_service_tree_add(st, cfqg);
914 }
915
916 static void
917 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
918 {
919         st->total_weight -= cfqg->weight;
920         if (!RB_EMPTY_NODE(&cfqg->rb_node))
921                 cfq_rb_erase(&cfqg->rb_node, st);
922 }
923
924 static void
925 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
926 {
927         struct cfq_rb_root *st = &cfqd->grp_service_tree;
928
929         BUG_ON(cfqg->nr_cfqq < 1);
930         cfqg->nr_cfqq--;
931
932         /* If there are other cfq queues under this group, don't delete it */
933         if (cfqg->nr_cfqq)
934                 return;
935
936         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
937         cfq_group_service_tree_del(st, cfqg);
938         cfqg->saved_workload_slice = 0;
939         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
940 }
941
942 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
943                                                 unsigned int *unaccounted_time)
944 {
945         unsigned int slice_used;
946
947         /*
948          * Queue got expired before even a single request completed or
949          * got expired immediately after first request completion.
950          */
951         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
952                 /*
953                  * Also charge the seek time incurred to the group, otherwise
954                  * if there are mutiple queues in the group, each can dispatch
955                  * a single request on seeky media and cause lots of seek time
956                  * and group will never know it.
957                  */
958                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
959                                         1);
960         } else {
961                 slice_used = jiffies - cfqq->slice_start;
962                 if (slice_used > cfqq->allocated_slice) {
963                         *unaccounted_time = slice_used - cfqq->allocated_slice;
964                         slice_used = cfqq->allocated_slice;
965                 }
966                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
967                         *unaccounted_time += cfqq->slice_start -
968                                         cfqq->dispatch_start;
969         }
970
971         return slice_used;
972 }
973
974 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
975                                 struct cfq_queue *cfqq)
976 {
977         struct cfq_rb_root *st = &cfqd->grp_service_tree;
978         unsigned int used_sl, charge, unaccounted_sl = 0;
979         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
980                         - cfqg->service_tree_idle.count;
981
982         BUG_ON(nr_sync < 0);
983         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
984
985         if (iops_mode(cfqd))
986                 charge = cfqq->slice_dispatch;
987         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
988                 charge = cfqq->allocated_slice;
989
990         /* Can't update vdisktime while group is on service tree */
991         cfq_group_service_tree_del(st, cfqg);
992         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
993         /* If a new weight was requested, update now, off tree */
994         cfq_group_service_tree_add(st, cfqg);
995
996         /* This group is being expired. Save the context */
997         if (time_after(cfqd->workload_expires, jiffies)) {
998                 cfqg->saved_workload_slice = cfqd->workload_expires
999                                                 - jiffies;
1000                 cfqg->saved_workload = cfqd->serving_type;
1001                 cfqg->saved_serving_prio = cfqd->serving_prio;
1002         } else
1003                 cfqg->saved_workload_slice = 0;
1004
1005         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1006                                         st->min_vdisktime);
1007         cfq_log_cfqq(cfqq->cfqd, cfqq,
1008                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1009                      used_sl, cfqq->slice_dispatch, charge,
1010                      iops_mode(cfqd), cfqq->nr_sectors);
1011         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1012                                           unaccounted_sl);
1013         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1014 }
1015
1016 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1017 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1018 {
1019         if (blkg)
1020                 return container_of(blkg, struct cfq_group, blkg);
1021         return NULL;
1022 }
1023
1024 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1025                                           unsigned int weight)
1026 {
1027         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1028         cfqg->new_weight = weight;
1029         cfqg->needs_update = true;
1030 }
1031
1032 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1033                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1034 {
1035         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1036         unsigned int major, minor;
1037
1038         /*
1039          * Add group onto cgroup list. It might happen that bdi->dev is
1040          * not initialized yet. Initialize this new group without major
1041          * and minor info and this info will be filled in once a new thread
1042          * comes for IO.
1043          */
1044         if (bdi->dev) {
1045                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1046                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1047                                         (void *)cfqd, MKDEV(major, minor));
1048         } else
1049                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1050                                         (void *)cfqd, 0);
1051
1052         cfqd->nr_blkcg_linked_grps++;
1053         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1054
1055         /* Add group on cfqd list */
1056         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1057 }
1058
1059 /*
1060  * Should be called from sleepable context. No request queue lock as per
1061  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1062  * from sleepable context.
1063  */
1064 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1065 {
1066         struct cfq_group *cfqg = NULL;
1067         int i, j, ret;
1068         struct cfq_rb_root *st;
1069
1070         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1071         if (!cfqg)
1072                 return NULL;
1073
1074         for_each_cfqg_st(cfqg, i, j, st)
1075                 *st = CFQ_RB_ROOT;
1076         RB_CLEAR_NODE(&cfqg->rb_node);
1077
1078         cfqg->ttime.last_end_request = jiffies;
1079
1080         /*
1081          * Take the initial reference that will be released on destroy
1082          * This can be thought of a joint reference by cgroup and
1083          * elevator which will be dropped by either elevator exit
1084          * or cgroup deletion path depending on who is exiting first.
1085          */
1086         cfqg->ref = 1;
1087
1088         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1089         if (ret) {
1090                 kfree(cfqg);
1091                 return NULL;
1092         }
1093
1094         return cfqg;
1095 }
1096
1097 static struct cfq_group *
1098 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1099 {
1100         struct cfq_group *cfqg = NULL;
1101         void *key = cfqd;
1102         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1103         unsigned int major, minor;
1104
1105         /*
1106          * This is the common case when there are no blkio cgroups.
1107          * Avoid lookup in this case
1108          */
1109         if (blkcg == &blkio_root_cgroup)
1110                 cfqg = &cfqd->root_group;
1111         else
1112                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1113
1114         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1115                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1116                 cfqg->blkg.dev = MKDEV(major, minor);
1117         }
1118
1119         return cfqg;
1120 }
1121
1122 /*
1123  * Search for the cfq group current task belongs to. request_queue lock must
1124  * be held.
1125  */
1126 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1127 {
1128         struct blkio_cgroup *blkcg;
1129         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1130         struct request_queue *q = cfqd->queue;
1131
1132         rcu_read_lock();
1133         blkcg = task_blkio_cgroup(current);
1134         cfqg = cfq_find_cfqg(cfqd, blkcg);
1135         if (cfqg) {
1136                 rcu_read_unlock();
1137                 return cfqg;
1138         }
1139
1140         /*
1141          * Need to allocate a group. Allocation of group also needs allocation
1142          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1143          * we need to drop rcu lock and queue_lock before we call alloc.
1144          *
1145          * Not taking any queue reference here and assuming that queue is
1146          * around by the time we return. CFQ queue allocation code does
1147          * the same. It might be racy though.
1148          */
1149
1150         rcu_read_unlock();
1151         spin_unlock_irq(q->queue_lock);
1152
1153         cfqg = cfq_alloc_cfqg(cfqd);
1154
1155         spin_lock_irq(q->queue_lock);
1156
1157         rcu_read_lock();
1158         blkcg = task_blkio_cgroup(current);
1159
1160         /*
1161          * If some other thread already allocated the group while we were
1162          * not holding queue lock, free up the group
1163          */
1164         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1165
1166         if (__cfqg) {
1167                 kfree(cfqg);
1168                 rcu_read_unlock();
1169                 return __cfqg;
1170         }
1171
1172         if (!cfqg)
1173                 cfqg = &cfqd->root_group;
1174
1175         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1176         rcu_read_unlock();
1177         return cfqg;
1178 }
1179
1180 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1181 {
1182         cfqg->ref++;
1183         return cfqg;
1184 }
1185
1186 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1187 {
1188         /* Currently, all async queues are mapped to root group */
1189         if (!cfq_cfqq_sync(cfqq))
1190                 cfqg = &cfqq->cfqd->root_group;
1191
1192         cfqq->cfqg = cfqg;
1193         /* cfqq reference on cfqg */
1194         cfqq->cfqg->ref++;
1195 }
1196
1197 static void cfq_put_cfqg(struct cfq_group *cfqg)
1198 {
1199         struct cfq_rb_root *st;
1200         int i, j;
1201
1202         BUG_ON(cfqg->ref <= 0);
1203         cfqg->ref--;
1204         if (cfqg->ref)
1205                 return;
1206         for_each_cfqg_st(cfqg, i, j, st)
1207                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1208         free_percpu(cfqg->blkg.stats_cpu);
1209         kfree(cfqg);
1210 }
1211
1212 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1213 {
1214         /* Something wrong if we are trying to remove same group twice */
1215         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1216
1217         hlist_del_init(&cfqg->cfqd_node);
1218
1219         BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1220         cfqd->nr_blkcg_linked_grps--;
1221
1222         /*
1223          * Put the reference taken at the time of creation so that when all
1224          * queues are gone, group can be destroyed.
1225          */
1226         cfq_put_cfqg(cfqg);
1227 }
1228
1229 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1230 {
1231         struct hlist_node *pos, *n;
1232         struct cfq_group *cfqg;
1233
1234         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1235                 /*
1236                  * If cgroup removal path got to blk_group first and removed
1237                  * it from cgroup list, then it will take care of destroying
1238                  * cfqg also.
1239                  */
1240                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1241                         cfq_destroy_cfqg(cfqd, cfqg);
1242         }
1243 }
1244
1245 /*
1246  * Blk cgroup controller notification saying that blkio_group object is being
1247  * delinked as associated cgroup object is going away. That also means that
1248  * no new IO will come in this group. So get rid of this group as soon as
1249  * any pending IO in the group is finished.
1250  *
1251  * This function is called under rcu_read_lock(). key is the rcu protected
1252  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1253  * read lock.
1254  *
1255  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1256  * it should not be NULL as even if elevator was exiting, cgroup deltion
1257  * path got to it first.
1258  */
1259 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1260 {
1261         unsigned long  flags;
1262         struct cfq_data *cfqd = key;
1263
1264         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1265         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1266         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1267 }
1268
1269 #else /* GROUP_IOSCHED */
1270 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1271 {
1272         return &cfqd->root_group;
1273 }
1274
1275 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1276 {
1277         return cfqg;
1278 }
1279
1280 static inline void
1281 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1282         cfqq->cfqg = cfqg;
1283 }
1284
1285 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1286 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1287
1288 #endif /* GROUP_IOSCHED */
1289
1290 /*
1291  * The cfqd->service_trees holds all pending cfq_queue's that have
1292  * requests waiting to be processed. It is sorted in the order that
1293  * we will service the queues.
1294  */
1295 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1296                                  bool add_front)
1297 {
1298         struct rb_node **p, *parent;
1299         struct cfq_queue *__cfqq;
1300         unsigned long rb_key;
1301         struct cfq_rb_root *service_tree;
1302         int left;
1303         int new_cfqq = 1;
1304
1305         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1306                                                 cfqq_type(cfqq));
1307         if (cfq_class_idle(cfqq)) {
1308                 rb_key = CFQ_IDLE_DELAY;
1309                 parent = rb_last(&service_tree->rb);
1310                 if (parent && parent != &cfqq->rb_node) {
1311                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1312                         rb_key += __cfqq->rb_key;
1313                 } else
1314                         rb_key += jiffies;
1315         } else if (!add_front) {
1316                 /*
1317                  * Get our rb key offset. Subtract any residual slice
1318                  * value carried from last service. A negative resid
1319                  * count indicates slice overrun, and this should position
1320                  * the next service time further away in the tree.
1321                  */
1322                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1323                 rb_key -= cfqq->slice_resid;
1324                 cfqq->slice_resid = 0;
1325         } else {
1326                 rb_key = -HZ;
1327                 __cfqq = cfq_rb_first(service_tree);
1328                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1329         }
1330
1331         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1332                 new_cfqq = 0;
1333                 /*
1334                  * same position, nothing more to do
1335                  */
1336                 if (rb_key == cfqq->rb_key &&
1337                     cfqq->service_tree == service_tree)
1338                         return;
1339
1340                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1341                 cfqq->service_tree = NULL;
1342         }
1343
1344         left = 1;
1345         parent = NULL;
1346         cfqq->service_tree = service_tree;
1347         p = &service_tree->rb.rb_node;
1348         while (*p) {
1349                 struct rb_node **n;
1350
1351                 parent = *p;
1352                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1353
1354                 /*
1355                  * sort by key, that represents service time.
1356                  */
1357                 if (time_before(rb_key, __cfqq->rb_key))
1358                         n = &(*p)->rb_left;
1359                 else {
1360                         n = &(*p)->rb_right;
1361                         left = 0;
1362                 }
1363
1364                 p = n;
1365         }
1366
1367         if (left)
1368                 service_tree->left = &cfqq->rb_node;
1369
1370         cfqq->rb_key = rb_key;
1371         rb_link_node(&cfqq->rb_node, parent, p);
1372         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1373         service_tree->count++;
1374         if (add_front || !new_cfqq)
1375                 return;
1376         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1377 }
1378
1379 static struct cfq_queue *
1380 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1381                      sector_t sector, struct rb_node **ret_parent,
1382                      struct rb_node ***rb_link)
1383 {
1384         struct rb_node **p, *parent;
1385         struct cfq_queue *cfqq = NULL;
1386
1387         parent = NULL;
1388         p = &root->rb_node;
1389         while (*p) {
1390                 struct rb_node **n;
1391
1392                 parent = *p;
1393                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1394
1395                 /*
1396                  * Sort strictly based on sector.  Smallest to the left,
1397                  * largest to the right.
1398                  */
1399                 if (sector > blk_rq_pos(cfqq->next_rq))
1400                         n = &(*p)->rb_right;
1401                 else if (sector < blk_rq_pos(cfqq->next_rq))
1402                         n = &(*p)->rb_left;
1403                 else
1404                         break;
1405                 p = n;
1406                 cfqq = NULL;
1407         }
1408
1409         *ret_parent = parent;
1410         if (rb_link)
1411                 *rb_link = p;
1412         return cfqq;
1413 }
1414
1415 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1416 {
1417         struct rb_node **p, *parent;
1418         struct cfq_queue *__cfqq;
1419
1420         if (cfqq->p_root) {
1421                 rb_erase(&cfqq->p_node, cfqq->p_root);
1422                 cfqq->p_root = NULL;
1423         }
1424
1425         if (cfq_class_idle(cfqq))
1426                 return;
1427         if (!cfqq->next_rq)
1428                 return;
1429
1430         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1431         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1432                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1433         if (!__cfqq) {
1434                 rb_link_node(&cfqq->p_node, parent, p);
1435                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1436         } else
1437                 cfqq->p_root = NULL;
1438 }
1439
1440 /*
1441  * Update cfqq's position in the service tree.
1442  */
1443 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1444 {
1445         /*
1446          * Resorting requires the cfqq to be on the RR list already.
1447          */
1448         if (cfq_cfqq_on_rr(cfqq)) {
1449                 cfq_service_tree_add(cfqd, cfqq, 0);
1450                 cfq_prio_tree_add(cfqd, cfqq);
1451         }
1452 }
1453
1454 /*
1455  * add to busy list of queues for service, trying to be fair in ordering
1456  * the pending list according to last request service
1457  */
1458 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1459 {
1460         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1461         BUG_ON(cfq_cfqq_on_rr(cfqq));
1462         cfq_mark_cfqq_on_rr(cfqq);
1463         cfqd->busy_queues++;
1464         if (cfq_cfqq_sync(cfqq))
1465                 cfqd->busy_sync_queues++;
1466
1467         cfq_resort_rr_list(cfqd, cfqq);
1468 }
1469
1470 /*
1471  * Called when the cfqq no longer has requests pending, remove it from
1472  * the service tree.
1473  */
1474 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1475 {
1476         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1477         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1478         cfq_clear_cfqq_on_rr(cfqq);
1479
1480         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1481                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1482                 cfqq->service_tree = NULL;
1483         }
1484         if (cfqq->p_root) {
1485                 rb_erase(&cfqq->p_node, cfqq->p_root);
1486                 cfqq->p_root = NULL;
1487         }
1488
1489         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1490         BUG_ON(!cfqd->busy_queues);
1491         cfqd->busy_queues--;
1492         if (cfq_cfqq_sync(cfqq))
1493                 cfqd->busy_sync_queues--;
1494 }
1495
1496 /*
1497  * rb tree support functions
1498  */
1499 static void cfq_del_rq_rb(struct request *rq)
1500 {
1501         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1502         const int sync = rq_is_sync(rq);
1503
1504         BUG_ON(!cfqq->queued[sync]);
1505         cfqq->queued[sync]--;
1506
1507         elv_rb_del(&cfqq->sort_list, rq);
1508
1509         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1510                 /*
1511                  * Queue will be deleted from service tree when we actually
1512                  * expire it later. Right now just remove it from prio tree
1513                  * as it is empty.
1514                  */
1515                 if (cfqq->p_root) {
1516                         rb_erase(&cfqq->p_node, cfqq->p_root);
1517                         cfqq->p_root = NULL;
1518                 }
1519         }
1520 }
1521
1522 static void cfq_add_rq_rb(struct request *rq)
1523 {
1524         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1525         struct cfq_data *cfqd = cfqq->cfqd;
1526         struct request *prev;
1527
1528         cfqq->queued[rq_is_sync(rq)]++;
1529
1530         elv_rb_add(&cfqq->sort_list, rq);
1531
1532         if (!cfq_cfqq_on_rr(cfqq))
1533                 cfq_add_cfqq_rr(cfqd, cfqq);
1534
1535         /*
1536          * check if this request is a better next-serve candidate
1537          */
1538         prev = cfqq->next_rq;
1539         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1540
1541         /*
1542          * adjust priority tree position, if ->next_rq changes
1543          */
1544         if (prev != cfqq->next_rq)
1545                 cfq_prio_tree_add(cfqd, cfqq);
1546
1547         BUG_ON(!cfqq->next_rq);
1548 }
1549
1550 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1551 {
1552         elv_rb_del(&cfqq->sort_list, rq);
1553         cfqq->queued[rq_is_sync(rq)]--;
1554         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1555                                         rq_data_dir(rq), rq_is_sync(rq));
1556         cfq_add_rq_rb(rq);
1557         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1558                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1559                         rq_is_sync(rq));
1560 }
1561
1562 static struct request *
1563 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1564 {
1565         struct task_struct *tsk = current;
1566         struct cfq_io_cq *cic;
1567         struct cfq_queue *cfqq;
1568
1569         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1570         if (!cic)
1571                 return NULL;
1572
1573         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1574         if (cfqq) {
1575                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1576
1577                 return elv_rb_find(&cfqq->sort_list, sector);
1578         }
1579
1580         return NULL;
1581 }
1582
1583 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1584 {
1585         struct cfq_data *cfqd = q->elevator->elevator_data;
1586
1587         cfqd->rq_in_driver++;
1588         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1589                                                 cfqd->rq_in_driver);
1590
1591         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1592 }
1593
1594 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1595 {
1596         struct cfq_data *cfqd = q->elevator->elevator_data;
1597
1598         WARN_ON(!cfqd->rq_in_driver);
1599         cfqd->rq_in_driver--;
1600         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1601                                                 cfqd->rq_in_driver);
1602 }
1603
1604 static void cfq_remove_request(struct request *rq)
1605 {
1606         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1607
1608         if (cfqq->next_rq == rq)
1609                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1610
1611         list_del_init(&rq->queuelist);
1612         cfq_del_rq_rb(rq);
1613
1614         cfqq->cfqd->rq_queued--;
1615         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1616                                         rq_data_dir(rq), rq_is_sync(rq));
1617         if (rq->cmd_flags & REQ_PRIO) {
1618                 WARN_ON(!cfqq->prio_pending);
1619                 cfqq->prio_pending--;
1620         }
1621 }
1622
1623 static int cfq_merge(struct request_queue *q, struct request **req,
1624                      struct bio *bio)
1625 {
1626         struct cfq_data *cfqd = q->elevator->elevator_data;
1627         struct request *__rq;
1628
1629         __rq = cfq_find_rq_fmerge(cfqd, bio);
1630         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1631                 *req = __rq;
1632                 return ELEVATOR_FRONT_MERGE;
1633         }
1634
1635         return ELEVATOR_NO_MERGE;
1636 }
1637
1638 static void cfq_merged_request(struct request_queue *q, struct request *req,
1639                                int type)
1640 {
1641         if (type == ELEVATOR_FRONT_MERGE) {
1642                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1643
1644                 cfq_reposition_rq_rb(cfqq, req);
1645         }
1646 }
1647
1648 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1649                                 struct bio *bio)
1650 {
1651         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1652                                         bio_data_dir(bio), cfq_bio_sync(bio));
1653 }
1654
1655 static void
1656 cfq_merged_requests(struct request_queue *q, struct request *rq,
1657                     struct request *next)
1658 {
1659         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1660         struct cfq_data *cfqd = q->elevator->elevator_data;
1661
1662         /*
1663          * reposition in fifo if next is older than rq
1664          */
1665         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1666             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1667                 list_move(&rq->queuelist, &next->queuelist);
1668                 rq_set_fifo_time(rq, rq_fifo_time(next));
1669         }
1670
1671         if (cfqq->next_rq == next)
1672                 cfqq->next_rq = rq;
1673         cfq_remove_request(next);
1674         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1675                                         rq_data_dir(next), rq_is_sync(next));
1676
1677         cfqq = RQ_CFQQ(next);
1678         /*
1679          * all requests of this queue are merged to other queues, delete it
1680          * from the service tree. If it's the active_queue,
1681          * cfq_dispatch_requests() will choose to expire it or do idle
1682          */
1683         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1684             cfqq != cfqd->active_queue)
1685                 cfq_del_cfqq_rr(cfqd, cfqq);
1686 }
1687
1688 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1689                            struct bio *bio)
1690 {
1691         struct cfq_data *cfqd = q->elevator->elevator_data;
1692         struct cfq_io_cq *cic;
1693         struct cfq_queue *cfqq;
1694
1695         /*
1696          * Disallow merge of a sync bio into an async request.
1697          */
1698         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1699                 return false;
1700
1701         /*
1702          * Lookup the cfqq that this bio will be queued with and allow
1703          * merge only if rq is queued there.
1704          */
1705         cic = cfq_cic_lookup(cfqd, current->io_context);
1706         if (!cic)
1707                 return false;
1708
1709         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1710         return cfqq == RQ_CFQQ(rq);
1711 }
1712
1713 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1714 {
1715         del_timer(&cfqd->idle_slice_timer);
1716         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1717 }
1718
1719 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1720                                    struct cfq_queue *cfqq)
1721 {
1722         if (cfqq) {
1723                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1724                                 cfqd->serving_prio, cfqd->serving_type);
1725                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1726                 cfqq->slice_start = 0;
1727                 cfqq->dispatch_start = jiffies;
1728                 cfqq->allocated_slice = 0;
1729                 cfqq->slice_end = 0;
1730                 cfqq->slice_dispatch = 0;
1731                 cfqq->nr_sectors = 0;
1732
1733                 cfq_clear_cfqq_wait_request(cfqq);
1734                 cfq_clear_cfqq_must_dispatch(cfqq);
1735                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1736                 cfq_clear_cfqq_fifo_expire(cfqq);
1737                 cfq_mark_cfqq_slice_new(cfqq);
1738
1739                 cfq_del_timer(cfqd, cfqq);
1740         }
1741
1742         cfqd->active_queue = cfqq;
1743 }
1744
1745 /*
1746  * current cfqq expired its slice (or was too idle), select new one
1747  */
1748 static void
1749 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1750                     bool timed_out)
1751 {
1752         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1753
1754         if (cfq_cfqq_wait_request(cfqq))
1755                 cfq_del_timer(cfqd, cfqq);
1756
1757         cfq_clear_cfqq_wait_request(cfqq);
1758         cfq_clear_cfqq_wait_busy(cfqq);
1759
1760         /*
1761          * If this cfqq is shared between multiple processes, check to
1762          * make sure that those processes are still issuing I/Os within
1763          * the mean seek distance.  If not, it may be time to break the
1764          * queues apart again.
1765          */
1766         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1767                 cfq_mark_cfqq_split_coop(cfqq);
1768
1769         /*
1770          * store what was left of this slice, if the queue idled/timed out
1771          */
1772         if (timed_out) {
1773                 if (cfq_cfqq_slice_new(cfqq))
1774                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1775                 else
1776                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1777                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1778         }
1779
1780         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1781
1782         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1783                 cfq_del_cfqq_rr(cfqd, cfqq);
1784
1785         cfq_resort_rr_list(cfqd, cfqq);
1786
1787         if (cfqq == cfqd->active_queue)
1788                 cfqd->active_queue = NULL;
1789
1790         if (cfqd->active_cic) {
1791                 put_io_context(cfqd->active_cic->icq.ioc);
1792                 cfqd->active_cic = NULL;
1793         }
1794 }
1795
1796 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1797 {
1798         struct cfq_queue *cfqq = cfqd->active_queue;
1799
1800         if (cfqq)
1801                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1802 }
1803
1804 /*
1805  * Get next queue for service. Unless we have a queue preemption,
1806  * we'll simply select the first cfqq in the service tree.
1807  */
1808 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1809 {
1810         struct cfq_rb_root *service_tree =
1811                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1812                                         cfqd->serving_type);
1813
1814         if (!cfqd->rq_queued)
1815                 return NULL;
1816
1817         /* There is nothing to dispatch */
1818         if (!service_tree)
1819                 return NULL;
1820         if (RB_EMPTY_ROOT(&service_tree->rb))
1821                 return NULL;
1822         return cfq_rb_first(service_tree);
1823 }
1824
1825 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1826 {
1827         struct cfq_group *cfqg;
1828         struct cfq_queue *cfqq;
1829         int i, j;
1830         struct cfq_rb_root *st;
1831
1832         if (!cfqd->rq_queued)
1833                 return NULL;
1834
1835         cfqg = cfq_get_next_cfqg(cfqd);
1836         if (!cfqg)
1837                 return NULL;
1838
1839         for_each_cfqg_st(cfqg, i, j, st)
1840                 if ((cfqq = cfq_rb_first(st)) != NULL)
1841                         return cfqq;
1842         return NULL;
1843 }
1844
1845 /*
1846  * Get and set a new active queue for service.
1847  */
1848 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1849                                               struct cfq_queue *cfqq)
1850 {
1851         if (!cfqq)
1852                 cfqq = cfq_get_next_queue(cfqd);
1853
1854         __cfq_set_active_queue(cfqd, cfqq);
1855         return cfqq;
1856 }
1857
1858 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1859                                           struct request *rq)
1860 {
1861         if (blk_rq_pos(rq) >= cfqd->last_position)
1862                 return blk_rq_pos(rq) - cfqd->last_position;
1863         else
1864                 return cfqd->last_position - blk_rq_pos(rq);
1865 }
1866
1867 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1868                                struct request *rq)
1869 {
1870         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1871 }
1872
1873 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1874                                     struct cfq_queue *cur_cfqq)
1875 {
1876         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1877         struct rb_node *parent, *node;
1878         struct cfq_queue *__cfqq;
1879         sector_t sector = cfqd->last_position;
1880
1881         if (RB_EMPTY_ROOT(root))
1882                 return NULL;
1883
1884         /*
1885          * First, if we find a request starting at the end of the last
1886          * request, choose it.
1887          */
1888         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1889         if (__cfqq)
1890                 return __cfqq;
1891
1892         /*
1893          * If the exact sector wasn't found, the parent of the NULL leaf
1894          * will contain the closest sector.
1895          */
1896         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1897         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1898                 return __cfqq;
1899
1900         if (blk_rq_pos(__cfqq->next_rq) < sector)
1901                 node = rb_next(&__cfqq->p_node);
1902         else
1903                 node = rb_prev(&__cfqq->p_node);
1904         if (!node)
1905                 return NULL;
1906
1907         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1908         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1909                 return __cfqq;
1910
1911         return NULL;
1912 }
1913
1914 /*
1915  * cfqd - obvious
1916  * cur_cfqq - passed in so that we don't decide that the current queue is
1917  *            closely cooperating with itself.
1918  *
1919  * So, basically we're assuming that that cur_cfqq has dispatched at least
1920  * one request, and that cfqd->last_position reflects a position on the disk
1921  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1922  * assumption.
1923  */
1924 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1925                                               struct cfq_queue *cur_cfqq)
1926 {
1927         struct cfq_queue *cfqq;
1928
1929         if (cfq_class_idle(cur_cfqq))
1930                 return NULL;
1931         if (!cfq_cfqq_sync(cur_cfqq))
1932                 return NULL;
1933         if (CFQQ_SEEKY(cur_cfqq))
1934                 return NULL;
1935
1936         /*
1937          * Don't search priority tree if it's the only queue in the group.
1938          */
1939         if (cur_cfqq->cfqg->nr_cfqq == 1)
1940                 return NULL;
1941
1942         /*
1943          * We should notice if some of the queues are cooperating, eg
1944          * working closely on the same area of the disk. In that case,
1945          * we can group them together and don't waste time idling.
1946          */
1947         cfqq = cfqq_close(cfqd, cur_cfqq);
1948         if (!cfqq)
1949                 return NULL;
1950
1951         /* If new queue belongs to different cfq_group, don't choose it */
1952         if (cur_cfqq->cfqg != cfqq->cfqg)
1953                 return NULL;
1954
1955         /*
1956          * It only makes sense to merge sync queues.
1957          */
1958         if (!cfq_cfqq_sync(cfqq))
1959                 return NULL;
1960         if (CFQQ_SEEKY(cfqq))
1961                 return NULL;
1962
1963         /*
1964          * Do not merge queues of different priority classes
1965          */
1966         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1967                 return NULL;
1968
1969         return cfqq;
1970 }
1971
1972 /*
1973  * Determine whether we should enforce idle window for this queue.
1974  */
1975
1976 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1977 {
1978         enum wl_prio_t prio = cfqq_prio(cfqq);
1979         struct cfq_rb_root *service_tree = cfqq->service_tree;
1980
1981         BUG_ON(!service_tree);
1982         BUG_ON(!service_tree->count);
1983
1984         if (!cfqd->cfq_slice_idle)
1985                 return false;
1986
1987         /* We never do for idle class queues. */
1988         if (prio == IDLE_WORKLOAD)
1989                 return false;
1990
1991         /* We do for queues that were marked with idle window flag. */
1992         if (cfq_cfqq_idle_window(cfqq) &&
1993            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1994                 return true;
1995
1996         /*
1997          * Otherwise, we do only if they are the last ones
1998          * in their service tree.
1999          */
2000         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2001            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2002                 return true;
2003         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2004                         service_tree->count);
2005         return false;
2006 }
2007
2008 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2009 {
2010         struct cfq_queue *cfqq = cfqd->active_queue;
2011         struct cfq_io_cq *cic;
2012         unsigned long sl, group_idle = 0;
2013
2014         /*
2015          * SSD device without seek penalty, disable idling. But only do so
2016          * for devices that support queuing, otherwise we still have a problem
2017          * with sync vs async workloads.
2018          */
2019         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2020                 return;
2021
2022         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2023         WARN_ON(cfq_cfqq_slice_new(cfqq));
2024
2025         /*
2026          * idle is disabled, either manually or by past process history
2027          */
2028         if (!cfq_should_idle(cfqd, cfqq)) {
2029                 /* no queue idling. Check for group idling */
2030                 if (cfqd->cfq_group_idle)
2031                         group_idle = cfqd->cfq_group_idle;
2032                 else
2033                         return;
2034         }
2035
2036         /*
2037          * still active requests from this queue, don't idle
2038          */
2039         if (cfqq->dispatched)
2040                 return;
2041
2042         /*
2043          * task has exited, don't wait
2044          */
2045         cic = cfqd->active_cic;
2046         if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
2047                 return;
2048
2049         /*
2050          * If our average think time is larger than the remaining time
2051          * slice, then don't idle. This avoids overrunning the allotted
2052          * time slice.
2053          */
2054         if (sample_valid(cic->ttime.ttime_samples) &&
2055             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2056                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2057                              cic->ttime.ttime_mean);
2058                 return;
2059         }
2060
2061         /* There are other queues in the group, don't do group idle */
2062         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2063                 return;
2064
2065         cfq_mark_cfqq_wait_request(cfqq);
2066
2067         if (group_idle)
2068                 sl = cfqd->cfq_group_idle;
2069         else
2070                 sl = cfqd->cfq_slice_idle;
2071
2072         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2073         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2074         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2075                         group_idle ? 1 : 0);
2076 }
2077
2078 /*
2079  * Move request from internal lists to the request queue dispatch list.
2080  */
2081 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2082 {
2083         struct cfq_data *cfqd = q->elevator->elevator_data;
2084         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2085
2086         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2087
2088         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2089         cfq_remove_request(rq);
2090         cfqq->dispatched++;
2091         (RQ_CFQG(rq))->dispatched++;
2092         elv_dispatch_sort(q, rq);
2093
2094         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2095         cfqq->nr_sectors += blk_rq_sectors(rq);
2096         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2097                                         rq_data_dir(rq), rq_is_sync(rq));
2098 }
2099
2100 /*
2101  * return expired entry, or NULL to just start from scratch in rbtree
2102  */
2103 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2104 {
2105         struct request *rq = NULL;
2106
2107         if (cfq_cfqq_fifo_expire(cfqq))
2108                 return NULL;
2109
2110         cfq_mark_cfqq_fifo_expire(cfqq);
2111
2112         if (list_empty(&cfqq->fifo))
2113                 return NULL;
2114
2115         rq = rq_entry_fifo(cfqq->fifo.next);
2116         if (time_before(jiffies, rq_fifo_time(rq)))
2117                 rq = NULL;
2118
2119         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2120         return rq;
2121 }
2122
2123 static inline int
2124 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2125 {
2126         const int base_rq = cfqd->cfq_slice_async_rq;
2127
2128         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2129
2130         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2131 }
2132
2133 /*
2134  * Must be called with the queue_lock held.
2135  */
2136 static int cfqq_process_refs(struct cfq_queue *cfqq)
2137 {
2138         int process_refs, io_refs;
2139
2140         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2141         process_refs = cfqq->ref - io_refs;
2142         BUG_ON(process_refs < 0);
2143         return process_refs;
2144 }
2145
2146 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2147 {
2148         int process_refs, new_process_refs;
2149         struct cfq_queue *__cfqq;
2150
2151         /*
2152          * If there are no process references on the new_cfqq, then it is
2153          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2154          * chain may have dropped their last reference (not just their
2155          * last process reference).
2156          */
2157         if (!cfqq_process_refs(new_cfqq))
2158                 return;
2159
2160         /* Avoid a circular list and skip interim queue merges */
2161         while ((__cfqq = new_cfqq->new_cfqq)) {
2162                 if (__cfqq == cfqq)
2163                         return;
2164                 new_cfqq = __cfqq;
2165         }
2166
2167         process_refs = cfqq_process_refs(cfqq);
2168         new_process_refs = cfqq_process_refs(new_cfqq);
2169         /*
2170          * If the process for the cfqq has gone away, there is no
2171          * sense in merging the queues.
2172          */
2173         if (process_refs == 0 || new_process_refs == 0)
2174                 return;
2175
2176         /*
2177          * Merge in the direction of the lesser amount of work.
2178          */
2179         if (new_process_refs >= process_refs) {
2180                 cfqq->new_cfqq = new_cfqq;
2181                 new_cfqq->ref += process_refs;
2182         } else {
2183                 new_cfqq->new_cfqq = cfqq;
2184                 cfqq->ref += new_process_refs;
2185         }
2186 }
2187
2188 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2189                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2190 {
2191         struct cfq_queue *queue;
2192         int i;
2193         bool key_valid = false;
2194         unsigned long lowest_key = 0;
2195         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2196
2197         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2198                 /* select the one with lowest rb_key */
2199                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2200                 if (queue &&
2201                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2202                         lowest_key = queue->rb_key;
2203                         cur_best = i;
2204                         key_valid = true;
2205                 }
2206         }
2207
2208         return cur_best;
2209 }
2210
2211 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2212 {
2213         unsigned slice;
2214         unsigned count;
2215         struct cfq_rb_root *st;
2216         unsigned group_slice;
2217         enum wl_prio_t original_prio = cfqd->serving_prio;
2218
2219         /* Choose next priority. RT > BE > IDLE */
2220         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2221                 cfqd->serving_prio = RT_WORKLOAD;
2222         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2223                 cfqd->serving_prio = BE_WORKLOAD;
2224         else {
2225                 cfqd->serving_prio = IDLE_WORKLOAD;
2226                 cfqd->workload_expires = jiffies + 1;
2227                 return;
2228         }
2229
2230         if (original_prio != cfqd->serving_prio)
2231                 goto new_workload;
2232
2233         /*
2234          * For RT and BE, we have to choose also the type
2235          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2236          * expiration time
2237          */
2238         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2239         count = st->count;
2240
2241         /*
2242          * check workload expiration, and that we still have other queues ready
2243          */
2244         if (count && !time_after(jiffies, cfqd->workload_expires))
2245                 return;
2246
2247 new_workload:
2248         /* otherwise select new workload type */
2249         cfqd->serving_type =
2250                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2251         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2252         count = st->count;
2253
2254         /*
2255          * the workload slice is computed as a fraction of target latency
2256          * proportional to the number of queues in that workload, over
2257          * all the queues in the same priority class
2258          */
2259         group_slice = cfq_group_slice(cfqd, cfqg);
2260
2261         slice = group_slice * count /
2262                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2263                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2264
2265         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2266                 unsigned int tmp;
2267
2268                 /*
2269                  * Async queues are currently system wide. Just taking
2270                  * proportion of queues with-in same group will lead to higher
2271                  * async ratio system wide as generally root group is going
2272                  * to have higher weight. A more accurate thing would be to
2273                  * calculate system wide asnc/sync ratio.
2274                  */
2275                 tmp = cfqd->cfq_target_latency *
2276                         cfqg_busy_async_queues(cfqd, cfqg);
2277                 tmp = tmp/cfqd->busy_queues;
2278                 slice = min_t(unsigned, slice, tmp);
2279
2280                 /* async workload slice is scaled down according to
2281                  * the sync/async slice ratio. */
2282                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2283         } else
2284                 /* sync workload slice is at least 2 * cfq_slice_idle */
2285                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2286
2287         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2288         cfq_log(cfqd, "workload slice:%d", slice);
2289         cfqd->workload_expires = jiffies + slice;
2290 }
2291
2292 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2293 {
2294         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2295         struct cfq_group *cfqg;
2296
2297         if (RB_EMPTY_ROOT(&st->rb))
2298                 return NULL;
2299         cfqg = cfq_rb_first_group(st);
2300         update_min_vdisktime(st);
2301         return cfqg;
2302 }
2303
2304 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2305 {
2306         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2307
2308         cfqd->serving_group = cfqg;
2309
2310         /* Restore the workload type data */
2311         if (cfqg->saved_workload_slice) {
2312                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2313                 cfqd->serving_type = cfqg->saved_workload;
2314                 cfqd->serving_prio = cfqg->saved_serving_prio;
2315         } else
2316                 cfqd->workload_expires = jiffies - 1;
2317
2318         choose_service_tree(cfqd, cfqg);
2319 }
2320
2321 /*
2322  * Select a queue for service. If we have a current active queue,
2323  * check whether to continue servicing it, or retrieve and set a new one.
2324  */
2325 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2326 {
2327         struct cfq_queue *cfqq, *new_cfqq = NULL;
2328
2329         cfqq = cfqd->active_queue;
2330         if (!cfqq)
2331                 goto new_queue;
2332
2333         if (!cfqd->rq_queued)
2334                 return NULL;
2335
2336         /*
2337          * We were waiting for group to get backlogged. Expire the queue
2338          */
2339         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2340                 goto expire;
2341
2342         /*
2343          * The active queue has run out of time, expire it and select new.
2344          */
2345         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2346                 /*
2347                  * If slice had not expired at the completion of last request
2348                  * we might not have turned on wait_busy flag. Don't expire
2349                  * the queue yet. Allow the group to get backlogged.
2350                  *
2351                  * The very fact that we have used the slice, that means we
2352                  * have been idling all along on this queue and it should be
2353                  * ok to wait for this request to complete.
2354                  */
2355                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2356                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2357                         cfqq = NULL;
2358                         goto keep_queue;
2359                 } else
2360                         goto check_group_idle;
2361         }
2362
2363         /*
2364          * The active queue has requests and isn't expired, allow it to
2365          * dispatch.
2366          */
2367         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2368                 goto keep_queue;
2369
2370         /*
2371          * If another queue has a request waiting within our mean seek
2372          * distance, let it run.  The expire code will check for close
2373          * cooperators and put the close queue at the front of the service
2374          * tree.  If possible, merge the expiring queue with the new cfqq.
2375          */
2376         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2377         if (new_cfqq) {
2378                 if (!cfqq->new_cfqq)
2379                         cfq_setup_merge(cfqq, new_cfqq);
2380                 goto expire;
2381         }
2382
2383         /*
2384          * No requests pending. If the active queue still has requests in
2385          * flight or is idling for a new request, allow either of these
2386          * conditions to happen (or time out) before selecting a new queue.
2387          */
2388         if (timer_pending(&cfqd->idle_slice_timer)) {
2389                 cfqq = NULL;
2390                 goto keep_queue;
2391         }
2392
2393         /*
2394          * This is a deep seek queue, but the device is much faster than
2395          * the queue can deliver, don't idle
2396          **/
2397         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2398             (cfq_cfqq_slice_new(cfqq) ||
2399             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2400                 cfq_clear_cfqq_deep(cfqq);
2401                 cfq_clear_cfqq_idle_window(cfqq);
2402         }
2403
2404         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2405                 cfqq = NULL;
2406                 goto keep_queue;
2407         }
2408
2409         /*
2410          * If group idle is enabled and there are requests dispatched from
2411          * this group, wait for requests to complete.
2412          */
2413 check_group_idle:
2414         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2415             cfqq->cfqg->dispatched &&
2416             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2417                 cfqq = NULL;
2418                 goto keep_queue;
2419         }
2420
2421 expire:
2422         cfq_slice_expired(cfqd, 0);
2423 new_queue:
2424         /*
2425          * Current queue expired. Check if we have to switch to a new
2426          * service tree
2427          */
2428         if (!new_cfqq)
2429                 cfq_choose_cfqg(cfqd);
2430
2431         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2432 keep_queue:
2433         return cfqq;
2434 }
2435
2436 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2437 {
2438         int dispatched = 0;
2439
2440         while (cfqq->next_rq) {
2441                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2442                 dispatched++;
2443         }
2444
2445         BUG_ON(!list_empty(&cfqq->fifo));
2446
2447         /* By default cfqq is not expired if it is empty. Do it explicitly */
2448         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2449         return dispatched;
2450 }
2451
2452 /*
2453  * Drain our current requests. Used for barriers and when switching
2454  * io schedulers on-the-fly.
2455  */
2456 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2457 {
2458         struct cfq_queue *cfqq;
2459         int dispatched = 0;
2460
2461         /* Expire the timeslice of the current active queue first */
2462         cfq_slice_expired(cfqd, 0);
2463         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2464                 __cfq_set_active_queue(cfqd, cfqq);
2465                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2466         }
2467
2468         BUG_ON(cfqd->busy_queues);
2469
2470         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2471         return dispatched;
2472 }
2473
2474 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2475         struct cfq_queue *cfqq)
2476 {
2477         /* the queue hasn't finished any request, can't estimate */
2478         if (cfq_cfqq_slice_new(cfqq))
2479                 return true;
2480         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2481                 cfqq->slice_end))
2482                 return true;
2483
2484         return false;
2485 }
2486
2487 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2488 {
2489         unsigned int max_dispatch;
2490
2491         /*
2492          * Drain async requests before we start sync IO
2493          */
2494         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2495                 return false;
2496
2497         /*
2498          * If this is an async queue and we have sync IO in flight, let it wait
2499          */
2500         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2501                 return false;
2502
2503         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2504         if (cfq_class_idle(cfqq))
2505                 max_dispatch = 1;
2506
2507         /*
2508          * Does this cfqq already have too much IO in flight?
2509          */
2510         if (cfqq->dispatched >= max_dispatch) {
2511                 bool promote_sync = false;
2512                 /*
2513                  * idle queue must always only have a single IO in flight
2514                  */
2515                 if (cfq_class_idle(cfqq))
2516                         return false;
2517
2518                 /*
2519                  * If there is only one sync queue
2520                  * we can ignore async queue here and give the sync
2521                  * queue no dispatch limit. The reason is a sync queue can
2522                  * preempt async queue, limiting the sync queue doesn't make
2523                  * sense. This is useful for aiostress test.
2524                  */
2525                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2526                         promote_sync = true;
2527
2528                 /*
2529                  * We have other queues, don't allow more IO from this one
2530                  */
2531                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2532                                 !promote_sync)
2533                         return false;
2534
2535                 /*
2536                  * Sole queue user, no limit
2537                  */
2538                 if (cfqd->busy_queues == 1 || promote_sync)
2539                         max_dispatch = -1;
2540                 else
2541                         /*
2542                          * Normally we start throttling cfqq when cfq_quantum/2
2543                          * requests have been dispatched. But we can drive
2544                          * deeper queue depths at the beginning of slice
2545                          * subjected to upper limit of cfq_quantum.
2546                          * */
2547                         max_dispatch = cfqd->cfq_quantum;
2548         }
2549
2550         /*
2551          * Async queues must wait a bit before being allowed dispatch.
2552          * We also ramp up the dispatch depth gradually for async IO,
2553          * based on the last sync IO we serviced
2554          */
2555         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2556                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2557                 unsigned int depth;
2558
2559                 depth = last_sync / cfqd->cfq_slice[1];
2560                 if (!depth && !cfqq->dispatched)
2561                         depth = 1;
2562                 if (depth < max_dispatch)
2563                         max_dispatch = depth;
2564         }
2565
2566         /*
2567          * If we're below the current max, allow a dispatch
2568          */
2569         return cfqq->dispatched < max_dispatch;
2570 }
2571
2572 /*
2573  * Dispatch a request from cfqq, moving them to the request queue
2574  * dispatch list.
2575  */
2576 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2577 {
2578         struct request *rq;
2579
2580         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2581
2582         if (!cfq_may_dispatch(cfqd, cfqq))
2583                 return false;
2584
2585         /*
2586          * follow expired path, else get first next available
2587          */
2588         rq = cfq_check_fifo(cfqq);
2589         if (!rq)
2590                 rq = cfqq->next_rq;
2591
2592         /*
2593          * insert request into driver dispatch list
2594          */
2595         cfq_dispatch_insert(cfqd->queue, rq);
2596
2597         if (!cfqd->active_cic) {
2598                 struct cfq_io_cq *cic = RQ_CIC(rq);
2599
2600                 atomic_long_inc(&cic->icq.ioc->refcount);
2601                 cfqd->active_cic = cic;
2602         }
2603
2604         return true;
2605 }
2606
2607 /*
2608  * Find the cfqq that we need to service and move a request from that to the
2609  * dispatch list
2610  */
2611 static int cfq_dispatch_requests(struct request_queue *q, int force)
2612 {
2613         struct cfq_data *cfqd = q->elevator->elevator_data;
2614         struct cfq_queue *cfqq;
2615
2616         if (!cfqd->busy_queues)
2617                 return 0;
2618
2619         if (unlikely(force))
2620                 return cfq_forced_dispatch(cfqd);
2621
2622         cfqq = cfq_select_queue(cfqd);
2623         if (!cfqq)
2624                 return 0;
2625
2626         /*
2627          * Dispatch a request from this cfqq, if it is allowed
2628          */
2629         if (!cfq_dispatch_request(cfqd, cfqq))
2630                 return 0;
2631
2632         cfqq->slice_dispatch++;
2633         cfq_clear_cfqq_must_dispatch(cfqq);
2634
2635         /*
2636          * expire an async queue immediately if it has used up its slice. idle
2637          * queue always expire after 1 dispatch round.
2638          */
2639         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2640             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2641             cfq_class_idle(cfqq))) {
2642                 cfqq->slice_end = jiffies + 1;
2643                 cfq_slice_expired(cfqd, 0);
2644         }
2645
2646         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2647         return 1;
2648 }
2649
2650 /*
2651  * task holds one reference to the queue, dropped when task exits. each rq
2652  * in-flight on this queue also holds a reference, dropped when rq is freed.
2653  *
2654  * Each cfq queue took a reference on the parent group. Drop it now.
2655  * queue lock must be held here.
2656  */
2657 static void cfq_put_queue(struct cfq_queue *cfqq)
2658 {
2659         struct cfq_data *cfqd = cfqq->cfqd;
2660         struct cfq_group *cfqg;
2661
2662         BUG_ON(cfqq->ref <= 0);
2663
2664         cfqq->ref--;
2665         if (cfqq->ref)
2666                 return;
2667
2668         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2669         BUG_ON(rb_first(&cfqq->sort_list));
2670         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2671         cfqg = cfqq->cfqg;
2672
2673         if (unlikely(cfqd->active_queue == cfqq)) {
2674                 __cfq_slice_expired(cfqd, cfqq, 0);
2675                 cfq_schedule_dispatch(cfqd);
2676         }
2677
2678         BUG_ON(cfq_cfqq_on_rr(cfqq));
2679         kmem_cache_free(cfq_pool, cfqq);
2680         cfq_put_cfqg(cfqg);
2681 }
2682
2683 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2684 {
2685         struct cfq_queue *__cfqq, *next;
2686
2687         /*
2688          * If this queue was scheduled to merge with another queue, be
2689          * sure to drop the reference taken on that queue (and others in
2690          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2691          */
2692         __cfqq = cfqq->new_cfqq;
2693         while (__cfqq) {
2694                 if (__cfqq == cfqq) {
2695                         WARN(1, "cfqq->new_cfqq loop detected\n");
2696                         break;
2697                 }
2698                 next = __cfqq->new_cfqq;
2699                 cfq_put_queue(__cfqq);
2700                 __cfqq = next;
2701         }
2702 }
2703
2704 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2705 {
2706         if (unlikely(cfqq == cfqd->active_queue)) {
2707                 __cfq_slice_expired(cfqd, cfqq, 0);
2708                 cfq_schedule_dispatch(cfqd);
2709         }
2710
2711         cfq_put_cooperator(cfqq);
2712
2713         cfq_put_queue(cfqq);
2714 }
2715
2716 static void cfq_init_icq(struct io_cq *icq)
2717 {
2718         struct cfq_io_cq *cic = icq_to_cic(icq);
2719
2720         cic->ttime.last_end_request = jiffies;
2721 }
2722
2723 static void cfq_exit_icq(struct io_cq *icq)
2724 {
2725         struct cfq_io_cq *cic = icq_to_cic(icq);
2726         struct cfq_data *cfqd = cic_to_cfqd(cic);
2727
2728         if (cic->cfqq[BLK_RW_ASYNC]) {
2729                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2730                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2731         }
2732
2733         if (cic->cfqq[BLK_RW_SYNC]) {
2734                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2735                 cic->cfqq[BLK_RW_SYNC] = NULL;
2736         }
2737 }
2738
2739 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2740 {
2741         struct task_struct *tsk = current;
2742         int ioprio_class;
2743
2744         if (!cfq_cfqq_prio_changed(cfqq))
2745                 return;
2746
2747         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2748         switch (ioprio_class) {
2749         default:
2750                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2751         case IOPRIO_CLASS_NONE:
2752                 /*
2753                  * no prio set, inherit CPU scheduling settings
2754                  */
2755                 cfqq->ioprio = task_nice_ioprio(tsk);
2756                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2757                 break;
2758         case IOPRIO_CLASS_RT:
2759                 cfqq->ioprio = task_ioprio(ioc);
2760                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2761                 break;
2762         case IOPRIO_CLASS_BE:
2763                 cfqq->ioprio = task_ioprio(ioc);
2764                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2765                 break;
2766         case IOPRIO_CLASS_IDLE:
2767                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2768                 cfqq->ioprio = 7;
2769                 cfq_clear_cfqq_idle_window(cfqq);
2770                 break;
2771         }
2772
2773         /*
2774          * keep track of original prio settings in case we have to temporarily
2775          * elevate the priority of this queue
2776          */
2777         cfqq->org_ioprio = cfqq->ioprio;
2778         cfq_clear_cfqq_prio_changed(cfqq);
2779 }
2780
2781 static void changed_ioprio(struct cfq_io_cq *cic)
2782 {
2783         struct cfq_data *cfqd = cic_to_cfqd(cic);
2784         struct cfq_queue *cfqq;
2785
2786         if (unlikely(!cfqd))
2787                 return;
2788
2789         cfqq = cic->cfqq[BLK_RW_ASYNC];
2790         if (cfqq) {
2791                 struct cfq_queue *new_cfqq;
2792                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2793                                                 GFP_ATOMIC);
2794                 if (new_cfqq) {
2795                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2796                         cfq_put_queue(cfqq);
2797                 }
2798         }
2799
2800         cfqq = cic->cfqq[BLK_RW_SYNC];
2801         if (cfqq)
2802                 cfq_mark_cfqq_prio_changed(cfqq);
2803 }
2804
2805 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2806                           pid_t pid, bool is_sync)
2807 {
2808         RB_CLEAR_NODE(&cfqq->rb_node);
2809         RB_CLEAR_NODE(&cfqq->p_node);
2810         INIT_LIST_HEAD(&cfqq->fifo);
2811
2812         cfqq->ref = 0;
2813         cfqq->cfqd = cfqd;
2814
2815         cfq_mark_cfqq_prio_changed(cfqq);
2816
2817         if (is_sync) {
2818                 if (!cfq_class_idle(cfqq))
2819                         cfq_mark_cfqq_idle_window(cfqq);
2820                 cfq_mark_cfqq_sync(cfqq);
2821         }
2822         cfqq->pid = pid;
2823 }
2824
2825 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2826 static void changed_cgroup(struct cfq_io_cq *cic)
2827 {
2828         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2829         struct cfq_data *cfqd = cic_to_cfqd(cic);
2830         struct request_queue *q;
2831
2832         if (unlikely(!cfqd))
2833                 return;
2834
2835         q = cfqd->queue;
2836
2837         if (sync_cfqq) {
2838                 /*
2839                  * Drop reference to sync queue. A new sync queue will be
2840                  * assigned in new group upon arrival of a fresh request.
2841                  */
2842                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2843                 cic_set_cfqq(cic, NULL, 1);
2844                 cfq_put_queue(sync_cfqq);
2845         }
2846 }
2847 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2848
2849 static struct cfq_queue *
2850 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2851                      struct io_context *ioc, gfp_t gfp_mask)
2852 {
2853         struct cfq_queue *cfqq, *new_cfqq = NULL;
2854         struct cfq_io_cq *cic;
2855         struct cfq_group *cfqg;
2856
2857 retry:
2858         cfqg = cfq_get_cfqg(cfqd);
2859         cic = cfq_cic_lookup(cfqd, ioc);
2860         /* cic always exists here */
2861         cfqq = cic_to_cfqq(cic, is_sync);
2862
2863         /*
2864          * Always try a new alloc if we fell back to the OOM cfqq
2865          * originally, since it should just be a temporary situation.
2866          */
2867         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2868                 cfqq = NULL;
2869                 if (new_cfqq) {
2870                         cfqq = new_cfqq;
2871                         new_cfqq = NULL;
2872                 } else if (gfp_mask & __GFP_WAIT) {
2873                         spin_unlock_irq(cfqd->queue->queue_lock);
2874                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2875                                         gfp_mask | __GFP_ZERO,
2876                                         cfqd->queue->node);
2877                         spin_lock_irq(cfqd->queue->queue_lock);
2878                         if (new_cfqq)
2879                                 goto retry;
2880                 } else {
2881                         cfqq = kmem_cache_alloc_node(cfq_pool,
2882                                         gfp_mask | __GFP_ZERO,
2883                                         cfqd->queue->node);
2884                 }
2885
2886                 if (cfqq) {
2887                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2888                         cfq_init_prio_data(cfqq, ioc);
2889                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2890                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2891                 } else
2892                         cfqq = &cfqd->oom_cfqq;
2893         }
2894
2895         if (new_cfqq)
2896                 kmem_cache_free(cfq_pool, new_cfqq);
2897
2898         return cfqq;
2899 }
2900
2901 static struct cfq_queue **
2902 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2903 {
2904         switch (ioprio_class) {
2905         case IOPRIO_CLASS_RT:
2906                 return &cfqd->async_cfqq[0][ioprio];
2907         case IOPRIO_CLASS_BE:
2908                 return &cfqd->async_cfqq[1][ioprio];
2909         case IOPRIO_CLASS_IDLE:
2910                 return &cfqd->async_idle_cfqq;
2911         default:
2912                 BUG();
2913         }
2914 }
2915
2916 static struct cfq_queue *
2917 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2918               gfp_t gfp_mask)
2919 {
2920         const int ioprio = task_ioprio(ioc);
2921         const int ioprio_class = task_ioprio_class(ioc);
2922         struct cfq_queue **async_cfqq = NULL;
2923         struct cfq_queue *cfqq = NULL;
2924
2925         if (!is_sync) {
2926                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2927                 cfqq = *async_cfqq;
2928         }
2929
2930         if (!cfqq)
2931                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2932
2933         /*
2934          * pin the queue now that it's allocated, scheduler exit will prune it
2935          */
2936         if (!is_sync && !(*async_cfqq)) {
2937                 cfqq->ref++;
2938                 *async_cfqq = cfqq;
2939         }
2940
2941         cfqq->ref++;
2942         return cfqq;
2943 }
2944
2945 static void
2946 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
2947 {
2948         unsigned long elapsed = jiffies - ttime->last_end_request;
2949         elapsed = min(elapsed, 2UL * slice_idle);
2950
2951         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2952         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2953         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2954 }
2955
2956 static void
2957 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2958                         struct cfq_io_cq *cic)
2959 {
2960         if (cfq_cfqq_sync(cfqq)) {
2961                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
2962                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2963                         cfqd->cfq_slice_idle);
2964         }
2965 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2966         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2967 #endif
2968 }
2969
2970 static void
2971 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2972                        struct request *rq)
2973 {
2974         sector_t sdist = 0;
2975         sector_t n_sec = blk_rq_sectors(rq);
2976         if (cfqq->last_request_pos) {
2977                 if (cfqq->last_request_pos < blk_rq_pos(rq))
2978                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2979                 else
2980                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2981         }
2982
2983         cfqq->seek_history <<= 1;
2984         if (blk_queue_nonrot(cfqd->queue))
2985                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2986         else
2987                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2988 }
2989
2990 /*
2991  * Disable idle window if the process thinks too long or seeks so much that
2992  * it doesn't matter
2993  */
2994 static void
2995 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2996                        struct cfq_io_cq *cic)
2997 {
2998         int old_idle, enable_idle;
2999
3000         /*
3001          * Don't idle for async or idle io prio class
3002          */
3003         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3004                 return;
3005
3006         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3007
3008         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3009                 cfq_mark_cfqq_deep(cfqq);
3010
3011         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3012                 enable_idle = 0;
3013         else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3014                  !cfqd->cfq_slice_idle ||
3015                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3016                 enable_idle = 0;
3017         else if (sample_valid(cic->ttime.ttime_samples)) {
3018                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3019                         enable_idle = 0;
3020                 else
3021                         enable_idle = 1;
3022         }
3023
3024         if (old_idle != enable_idle) {
3025                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3026                 if (enable_idle)
3027                         cfq_mark_cfqq_idle_window(cfqq);
3028                 else
3029                         cfq_clear_cfqq_idle_window(cfqq);
3030         }
3031 }
3032
3033 /*
3034  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3035  * no or if we aren't sure, a 1 will cause a preempt.
3036  */
3037 static bool
3038 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3039                    struct request *rq)
3040 {
3041         struct cfq_queue *cfqq;
3042
3043         cfqq = cfqd->active_queue;
3044         if (!cfqq)
3045                 return false;
3046
3047         if (cfq_class_idle(new_cfqq))
3048                 return false;
3049
3050         if (cfq_class_idle(cfqq))
3051                 return true;
3052
3053         /*
3054          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3055          */
3056         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3057                 return false;
3058
3059         /*
3060          * if the new request is sync, but the currently running queue is
3061          * not, let the sync request have priority.
3062          */
3063         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3064                 return true;
3065
3066         if (new_cfqq->cfqg != cfqq->cfqg)
3067                 return false;
3068
3069         if (cfq_slice_used(cfqq))
3070                 return true;
3071
3072         /* Allow preemption only if we are idling on sync-noidle tree */
3073         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3074             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3075             new_cfqq->service_tree->count == 2 &&
3076             RB_EMPTY_ROOT(&cfqq->sort_list))
3077                 return true;
3078
3079         /*
3080          * So both queues are sync. Let the new request get disk time if
3081          * it's a metadata request and the current queue is doing regular IO.
3082          */
3083         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3084                 return true;
3085
3086         /*
3087          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3088          */
3089         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3090                 return true;
3091
3092         /* An idle queue should not be idle now for some reason */
3093         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3094                 return true;
3095
3096         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3097                 return false;
3098
3099         /*
3100          * if this request is as-good as one we would expect from the
3101          * current cfqq, let it preempt
3102          */
3103         if (cfq_rq_close(cfqd, cfqq, rq))
3104                 return true;
3105
3106         return false;
3107 }
3108
3109 /*
3110  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3111  * let it have half of its nominal slice.
3112  */
3113 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3114 {
3115         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3116
3117         cfq_log_cfqq(cfqd, cfqq, "preempt");
3118         cfq_slice_expired(cfqd, 1);
3119
3120         /*
3121          * workload type is changed, don't save slice, otherwise preempt
3122          * doesn't happen
3123          */
3124         if (old_type != cfqq_type(cfqq))
3125                 cfqq->cfqg->saved_workload_slice = 0;
3126
3127         /*
3128          * Put the new queue at the front of the of the current list,
3129          * so we know that it will be selected next.
3130          */
3131         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3132
3133         cfq_service_tree_add(cfqd, cfqq, 1);
3134
3135         cfqq->slice_end = 0;
3136         cfq_mark_cfqq_slice_new(cfqq);
3137 }
3138
3139 /*
3140  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3141  * something we should do about it
3142  */
3143 static void
3144 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3145                 struct request *rq)
3146 {
3147         struct cfq_io_cq *cic = RQ_CIC(rq);
3148
3149         cfqd->rq_queued++;
3150         if (rq->cmd_flags & REQ_PRIO)
3151                 cfqq->prio_pending++;
3152
3153         cfq_update_io_thinktime(cfqd, cfqq, cic);
3154         cfq_update_io_seektime(cfqd, cfqq, rq);
3155         cfq_update_idle_window(cfqd, cfqq, cic);
3156
3157         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3158
3159         if (cfqq == cfqd->active_queue) {
3160                 /*
3161                  * Remember that we saw a request from this process, but
3162                  * don't start queuing just yet. Otherwise we risk seeing lots
3163                  * of tiny requests, because we disrupt the normal plugging
3164                  * and merging. If the request is already larger than a single
3165                  * page, let it rip immediately. For that case we assume that
3166                  * merging is already done. Ditto for a busy system that
3167                  * has other work pending, don't risk delaying until the
3168                  * idle timer unplug to continue working.
3169                  */
3170                 if (cfq_cfqq_wait_request(cfqq)) {
3171                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3172                             cfqd->busy_queues > 1) {
3173                                 cfq_del_timer(cfqd, cfqq);
3174                                 cfq_clear_cfqq_wait_request(cfqq);
3175                                 __blk_run_queue(cfqd->queue);
3176                         } else {
3177                                 cfq_blkiocg_update_idle_time_stats(
3178                                                 &cfqq->cfqg->blkg);
3179                                 cfq_mark_cfqq_must_dispatch(cfqq);
3180                         }
3181                 }
3182         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3183                 /*
3184                  * not the active queue - expire current slice if it is
3185                  * idle and has expired it's mean thinktime or this new queue
3186                  * has some old slice time left and is of higher priority or
3187                  * this new queue is RT and the current one is BE
3188                  */
3189                 cfq_preempt_queue(cfqd, cfqq);
3190                 __blk_run_queue(cfqd->queue);
3191         }
3192 }
3193
3194 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3195 {
3196         struct cfq_data *cfqd = q->elevator->elevator_data;
3197         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3198
3199         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3200         cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3201
3202         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3203         list_add_tail(&rq->queuelist, &cfqq->fifo);
3204         cfq_add_rq_rb(rq);
3205         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3206                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3207                         rq_is_sync(rq));
3208         cfq_rq_enqueued(cfqd, cfqq, rq);
3209 }
3210
3211 /*
3212  * Update hw_tag based on peak queue depth over 50 samples under
3213  * sufficient load.
3214  */
3215 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3216 {
3217         struct cfq_queue *cfqq = cfqd->active_queue;
3218
3219         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3220                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3221
3222         if (cfqd->hw_tag == 1)
3223                 return;
3224
3225         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3226             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3227                 return;
3228
3229         /*
3230          * If active queue hasn't enough requests and can idle, cfq might not
3231          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3232          * case
3233          */
3234         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3235             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3236             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3237                 return;
3238
3239         if (cfqd->hw_tag_samples++ < 50)
3240                 return;
3241
3242         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3243                 cfqd->hw_tag = 1;
3244         else
3245                 cfqd->hw_tag = 0;
3246 }
3247
3248 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3249 {
3250         struct cfq_io_cq *cic = cfqd->active_cic;
3251
3252         /* If the queue already has requests, don't wait */
3253         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3254                 return false;
3255
3256         /* If there are other queues in the group, don't wait */
3257         if (cfqq->cfqg->nr_cfqq > 1)
3258                 return false;
3259
3260         /* the only queue in the group, but think time is big */
3261         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3262                 return false;
3263
3264         if (cfq_slice_used(cfqq))
3265                 return true;
3266
3267         /* if slice left is less than think time, wait busy */
3268         if (cic && sample_valid(cic->ttime.ttime_samples)
3269             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3270                 return true;
3271
3272         /*
3273          * If think times is less than a jiffy than ttime_mean=0 and above
3274          * will not be true. It might happen that slice has not expired yet
3275          * but will expire soon (4-5 ns) during select_queue(). To cover the
3276          * case where think time is less than a jiffy, mark the queue wait
3277          * busy if only 1 jiffy is left in the slice.
3278          */
3279         if (cfqq->slice_end - jiffies == 1)
3280                 return true;
3281
3282         return false;
3283 }
3284
3285 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3286 {
3287         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3288         struct cfq_data *cfqd = cfqq->cfqd;
3289         const int sync = rq_is_sync(rq);
3290         unsigned long now;
3291
3292         now = jiffies;
3293         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3294                      !!(rq->cmd_flags & REQ_NOIDLE));
3295
3296         cfq_update_hw_tag(cfqd);
3297
3298         WARN_ON(!cfqd->rq_in_driver);
3299         WARN_ON(!cfqq->dispatched);
3300         cfqd->rq_in_driver--;
3301         cfqq->dispatched--;
3302         (RQ_CFQG(rq))->dispatched--;
3303         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3304                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3305                         rq_data_dir(rq), rq_is_sync(rq));
3306
3307         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3308
3309         if (sync) {
3310                 struct cfq_rb_root *service_tree;
3311
3312                 RQ_CIC(rq)->ttime.last_end_request = now;
3313
3314                 if (cfq_cfqq_on_rr(cfqq))
3315                         service_tree = cfqq->service_tree;
3316                 else
3317                         service_tree = service_tree_for(cfqq->cfqg,
3318                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3319                 service_tree->ttime.last_end_request = now;
3320                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3321                         cfqd->last_delayed_sync = now;
3322         }
3323
3324 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3325         cfqq->cfqg->ttime.last_end_request = now;
3326 #endif
3327
3328         /*
3329          * If this is the active queue, check if it needs to be expired,
3330          * or if we want to idle in case it has no pending requests.
3331          */
3332         if (cfqd->active_queue == cfqq) {
3333                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3334
3335                 if (cfq_cfqq_slice_new(cfqq)) {
3336                         cfq_set_prio_slice(cfqd, cfqq);
3337                         cfq_clear_cfqq_slice_new(cfqq);
3338                 }
3339
3340                 /*
3341                  * Should we wait for next request to come in before we expire
3342                  * the queue.
3343                  */
3344                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3345                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3346                         if (!cfqd->cfq_slice_idle)
3347                                 extend_sl = cfqd->cfq_group_idle;
3348                         cfqq->slice_end = jiffies + extend_sl;
3349                         cfq_mark_cfqq_wait_busy(cfqq);
3350                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3351                 }
3352
3353                 /*
3354                  * Idling is not enabled on:
3355                  * - expired queues
3356                  * - idle-priority queues
3357                  * - async queues
3358                  * - queues with still some requests queued
3359                  * - when there is a close cooperator
3360                  */
3361                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3362                         cfq_slice_expired(cfqd, 1);
3363                 else if (sync && cfqq_empty &&
3364                          !cfq_close_cooperator(cfqd, cfqq)) {
3365                         cfq_arm_slice_timer(cfqd);
3366                 }
3367         }
3368
3369         if (!cfqd->rq_in_driver)
3370                 cfq_schedule_dispatch(cfqd);
3371 }
3372
3373 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3374 {
3375         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3376                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3377                 return ELV_MQUEUE_MUST;
3378         }
3379
3380         return ELV_MQUEUE_MAY;
3381 }
3382
3383 static int cfq_may_queue(struct request_queue *q, int rw)
3384 {
3385         struct cfq_data *cfqd = q->elevator->elevator_data;
3386         struct task_struct *tsk = current;
3387         struct cfq_io_cq *cic;
3388         struct cfq_queue *cfqq;
3389
3390         /*
3391          * don't force setup of a queue from here, as a call to may_queue
3392          * does not necessarily imply that a request actually will be queued.
3393          * so just lookup a possibly existing queue, or return 'may queue'
3394          * if that fails
3395          */
3396         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3397         if (!cic)
3398                 return ELV_MQUEUE_MAY;
3399
3400         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3401         if (cfqq) {
3402                 cfq_init_prio_data(cfqq, cic->icq.ioc);
3403
3404                 return __cfq_may_queue(cfqq);
3405         }
3406
3407         return ELV_MQUEUE_MAY;
3408 }
3409
3410 /*
3411  * queue lock held here
3412  */
3413 static void cfq_put_request(struct request *rq)
3414 {
3415         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3416
3417         if (cfqq) {
3418                 const int rw = rq_data_dir(rq);
3419
3420                 BUG_ON(!cfqq->allocated[rw]);
3421                 cfqq->allocated[rw]--;
3422
3423                 /* Put down rq reference on cfqg */
3424                 cfq_put_cfqg(RQ_CFQG(rq));
3425                 rq->elv.priv[0] = NULL;
3426                 rq->elv.priv[1] = NULL;
3427
3428                 cfq_put_queue(cfqq);
3429         }
3430 }
3431
3432 static struct cfq_queue *
3433 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3434                 struct cfq_queue *cfqq)
3435 {
3436         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3437         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3438         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3439         cfq_put_queue(cfqq);
3440         return cic_to_cfqq(cic, 1);
3441 }
3442
3443 /*
3444  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3445  * was the last process referring to said cfqq.
3446  */
3447 static struct cfq_queue *
3448 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3449 {
3450         if (cfqq_process_refs(cfqq) == 1) {
3451                 cfqq->pid = current->pid;
3452                 cfq_clear_cfqq_coop(cfqq);
3453                 cfq_clear_cfqq_split_coop(cfqq);
3454                 return cfqq;
3455         }
3456
3457         cic_set_cfqq(cic, NULL, 1);
3458
3459         cfq_put_cooperator(cfqq);
3460
3461         cfq_put_queue(cfqq);
3462         return NULL;
3463 }
3464 /*
3465  * Allocate cfq data structures associated with this request.
3466  */
3467 static int
3468 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3469 {
3470         struct cfq_data *cfqd = q->elevator->elevator_data;
3471         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3472         const int rw = rq_data_dir(rq);
3473         const bool is_sync = rq_is_sync(rq);
3474         struct cfq_queue *cfqq;
3475         unsigned int changed;
3476
3477         might_sleep_if(gfp_mask & __GFP_WAIT);
3478
3479         spin_lock_irq(q->queue_lock);
3480
3481         /* handle changed notifications */
3482         changed = icq_get_changed(&cic->icq);
3483         if (unlikely(changed & ICQ_IOPRIO_CHANGED))
3484                 changed_ioprio(cic);
3485 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3486         if (unlikely(changed & ICQ_CGROUP_CHANGED))
3487                 changed_cgroup(cic);
3488 #endif
3489
3490 new_queue:
3491         cfqq = cic_to_cfqq(cic, is_sync);
3492         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3493                 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3494                 cic_set_cfqq(cic, cfqq, is_sync);
3495         } else {
3496                 /*
3497                  * If the queue was seeky for too long, break it apart.
3498                  */
3499                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3500                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3501                         cfqq = split_cfqq(cic, cfqq);
3502                         if (!cfqq)
3503                                 goto new_queue;
3504                 }
3505
3506                 /*
3507                  * Check to see if this queue is scheduled to merge with
3508                  * another, closely cooperating queue.  The merging of
3509                  * queues happens here as it must be done in process context.
3510                  * The reference on new_cfqq was taken in merge_cfqqs.
3511                  */
3512                 if (cfqq->new_cfqq)
3513                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3514         }
3515
3516         cfqq->allocated[rw]++;
3517
3518         cfqq->ref++;
3519         rq->elv.priv[0] = cfqq;
3520         rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
3521         spin_unlock_irq(q->queue_lock);
3522         return 0;
3523 }
3524
3525 static void cfq_kick_queue(struct work_struct *work)
3526 {
3527         struct cfq_data *cfqd =
3528                 container_of(work, struct cfq_data, unplug_work);
3529         struct request_queue *q = cfqd->queue;
3530
3531         spin_lock_irq(q->queue_lock);
3532         __blk_run_queue(cfqd->queue);
3533         spin_unlock_irq(q->queue_lock);
3534 }
3535
3536 /*
3537  * Timer running if the active_queue is currently idling inside its time slice
3538  */
3539 static void cfq_idle_slice_timer(unsigned long data)
3540 {
3541         struct cfq_data *cfqd = (struct cfq_data *) data;
3542         struct cfq_queue *cfqq;
3543         unsigned long flags;
3544         int timed_out = 1;
3545
3546         cfq_log(cfqd, "idle timer fired");
3547
3548         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3549
3550         cfqq = cfqd->active_queue;
3551         if (cfqq) {
3552                 timed_out = 0;
3553
3554                 /*
3555                  * We saw a request before the queue expired, let it through
3556                  */
3557                 if (cfq_cfqq_must_dispatch(cfqq))
3558                         goto out_kick;
3559
3560                 /*
3561                  * expired
3562                  */
3563                 if (cfq_slice_used(cfqq))
3564                         goto expire;
3565
3566                 /*
3567                  * only expire and reinvoke request handler, if there are
3568                  * other queues with pending requests
3569                  */
3570                 if (!cfqd->busy_queues)
3571                         goto out_cont;
3572
3573                 /*
3574                  * not expired and it has a request pending, let it dispatch
3575                  */
3576                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3577                         goto out_kick;
3578
3579                 /*
3580                  * Queue depth flag is reset only when the idle didn't succeed
3581                  */
3582                 cfq_clear_cfqq_deep(cfqq);
3583         }
3584 expire:
3585         cfq_slice_expired(cfqd, timed_out);
3586 out_kick:
3587         cfq_schedule_dispatch(cfqd);
3588 out_cont:
3589         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3590 }
3591
3592 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3593 {
3594         del_timer_sync(&cfqd->idle_slice_timer);
3595         cancel_work_sync(&cfqd->unplug_work);
3596 }
3597
3598 static void cfq_put_async_queues(struct cfq_data *cfqd)
3599 {
3600         int i;
3601
3602         for (i = 0; i < IOPRIO_BE_NR; i++) {
3603                 if (cfqd->async_cfqq[0][i])
3604                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3605                 if (cfqd->async_cfqq[1][i])
3606                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3607         }
3608
3609         if (cfqd->async_idle_cfqq)
3610                 cfq_put_queue(cfqd->async_idle_cfqq);
3611 }
3612
3613 static void cfq_exit_queue(struct elevator_queue *e)
3614 {
3615         struct cfq_data *cfqd = e->elevator_data;
3616         struct request_queue *q = cfqd->queue;
3617         bool wait = false;
3618
3619         cfq_shutdown_timer_wq(cfqd);
3620
3621         spin_lock_irq(q->queue_lock);
3622
3623         if (cfqd->active_queue)
3624                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3625
3626         cfq_put_async_queues(cfqd);
3627         cfq_release_cfq_groups(cfqd);
3628
3629         /*
3630          * If there are groups which we could not unlink from blkcg list,
3631          * wait for a rcu period for them to be freed.
3632          */
3633         if (cfqd->nr_blkcg_linked_grps)
3634                 wait = true;
3635
3636         spin_unlock_irq(q->queue_lock);
3637
3638         cfq_shutdown_timer_wq(cfqd);
3639
3640         /*
3641          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3642          * Do this wait only if there are other unlinked groups out
3643          * there. This can happen if cgroup deletion path claimed the
3644          * responsibility of cleaning up a group before queue cleanup code
3645          * get to the group.
3646          *
3647          * Do not call synchronize_rcu() unconditionally as there are drivers
3648          * which create/delete request queue hundreds of times during scan/boot
3649          * and synchronize_rcu() can take significant time and slow down boot.
3650          */
3651         if (wait)
3652                 synchronize_rcu();
3653
3654 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3655         /* Free up per cpu stats for root group */
3656         free_percpu(cfqd->root_group.blkg.stats_cpu);
3657 #endif
3658         kfree(cfqd);
3659 }
3660
3661 static void *cfq_init_queue(struct request_queue *q)
3662 {
3663         struct cfq_data *cfqd;
3664         int i, j;
3665         struct cfq_group *cfqg;
3666         struct cfq_rb_root *st;
3667
3668         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3669         if (!cfqd)
3670                 return NULL;
3671
3672         /* Init root service tree */
3673         cfqd->grp_service_tree = CFQ_RB_ROOT;
3674
3675         /* Init root group */
3676         cfqg = &cfqd->root_group;
3677         for_each_cfqg_st(cfqg, i, j, st)
3678                 *st = CFQ_RB_ROOT;
3679         RB_CLEAR_NODE(&cfqg->rb_node);
3680
3681         /* Give preference to root group over other groups */
3682         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3683
3684 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3685         /*
3686          * Set root group reference to 2. One reference will be dropped when
3687          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3688          * Other reference will remain there as we don't want to delete this
3689          * group as it is statically allocated and gets destroyed when
3690          * throtl_data goes away.
3691          */
3692         cfqg->ref = 2;
3693
3694         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3695                 kfree(cfqg);
3696                 kfree(cfqd);
3697                 return NULL;
3698         }
3699
3700         rcu_read_lock();
3701
3702         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3703                                         (void *)cfqd, 0);
3704         rcu_read_unlock();
3705         cfqd->nr_blkcg_linked_grps++;
3706
3707         /* Add group on cfqd->cfqg_list */
3708         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3709 #endif
3710         /*
3711          * Not strictly needed (since RB_ROOT just clears the node and we
3712          * zeroed cfqd on alloc), but better be safe in case someone decides
3713          * to add magic to the rb code
3714          */
3715         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3716                 cfqd->prio_trees[i] = RB_ROOT;
3717
3718         /*
3719          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3720          * Grab a permanent reference to it, so that the normal code flow
3721          * will not attempt to free it.
3722          */
3723         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3724         cfqd->oom_cfqq.ref++;
3725         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3726
3727         cfqd->queue = q;
3728
3729         init_timer(&cfqd->idle_slice_timer);
3730         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3731         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3732
3733         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3734
3735         cfqd->cfq_quantum = cfq_quantum;
3736         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3737         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3738         cfqd->cfq_back_max = cfq_back_max;
3739         cfqd->cfq_back_penalty = cfq_back_penalty;
3740         cfqd->cfq_slice[0] = cfq_slice_async;
3741         cfqd->cfq_slice[1] = cfq_slice_sync;
3742         cfqd->cfq_target_latency = cfq_target_latency;
3743         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3744         cfqd->cfq_slice_idle = cfq_slice_idle;
3745         cfqd->cfq_group_idle = cfq_group_idle;
3746         cfqd->cfq_latency = 1;
3747         cfqd->hw_tag = -1;
3748         /*
3749          * we optimistically start assuming sync ops weren't delayed in last
3750          * second, in order to have larger depth for async operations.
3751          */
3752         cfqd->last_delayed_sync = jiffies - HZ;
3753         return cfqd;
3754 }
3755
3756 /*
3757  * sysfs parts below -->
3758  */
3759 static ssize_t
3760 cfq_var_show(unsigned int var, char *page)
3761 {
3762         return sprintf(page, "%d\n", var);
3763 }
3764
3765 static ssize_t
3766 cfq_var_store(unsigned int *var, const char *page, size_t count)
3767 {
3768         char *p = (char *) page;
3769
3770         *var = simple_strtoul(p, &p, 10);
3771         return count;
3772 }
3773
3774 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3775 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3776 {                                                                       \
3777         struct cfq_data *cfqd = e->elevator_data;                       \
3778         unsigned int __data = __VAR;                                    \
3779         if (__CONV)                                                     \
3780                 __data = jiffies_to_msecs(__data);                      \
3781         return cfq_var_show(__data, (page));                            \
3782 }
3783 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3784 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3785 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3786 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3787 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3788 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3789 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3790 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3791 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3792 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3793 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3794 SHOW_FUNCTION(cfq_target_latency_show, cfqd->cfq_target_latency, 1);
3795 #undef SHOW_FUNCTION
3796
3797 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3798 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3799 {                                                                       \
3800         struct cfq_data *cfqd = e->elevator_data;                       \
3801         unsigned int __data;                                            \
3802         int ret = cfq_var_store(&__data, (page), count);                \
3803         if (__data < (MIN))                                             \
3804                 __data = (MIN);                                         \
3805         else if (__data > (MAX))                                        \
3806                 __data = (MAX);                                         \
3807         if (__CONV)                                                     \
3808                 *(__PTR) = msecs_to_jiffies(__data);                    \
3809         else                                                            \
3810                 *(__PTR) = __data;                                      \
3811         return ret;                                                     \
3812 }
3813 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3814 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3815                 UINT_MAX, 1);
3816 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3817                 UINT_MAX, 1);
3818 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3819 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3820                 UINT_MAX, 0);
3821 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3822 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
3823 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3824 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3825 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3826                 UINT_MAX, 0);
3827 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3828 STORE_FUNCTION(cfq_target_latency_store, &cfqd->cfq_target_latency, 1, UINT_MAX, 1);
3829 #undef STORE_FUNCTION
3830
3831 #define CFQ_ATTR(name) \
3832         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3833
3834 static struct elv_fs_entry cfq_attrs[] = {
3835         CFQ_ATTR(quantum),
3836         CFQ_ATTR(fifo_expire_sync),
3837         CFQ_ATTR(fifo_expire_async),
3838         CFQ_ATTR(back_seek_max),
3839         CFQ_ATTR(back_seek_penalty),
3840         CFQ_ATTR(slice_sync),
3841         CFQ_ATTR(slice_async),
3842         CFQ_ATTR(slice_async_rq),
3843         CFQ_ATTR(slice_idle),
3844         CFQ_ATTR(group_idle),
3845         CFQ_ATTR(low_latency),
3846         CFQ_ATTR(target_latency),
3847         __ATTR_NULL
3848 };
3849
3850 static struct elevator_type iosched_cfq = {
3851         .ops = {
3852                 .elevator_merge_fn =            cfq_merge,
3853                 .elevator_merged_fn =           cfq_merged_request,
3854                 .elevator_merge_req_fn =        cfq_merged_requests,
3855                 .elevator_allow_merge_fn =      cfq_allow_merge,
3856                 .elevator_bio_merged_fn =       cfq_bio_merged,
3857                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3858                 .elevator_add_req_fn =          cfq_insert_request,
3859                 .elevator_activate_req_fn =     cfq_activate_request,
3860                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3861                 .elevator_completed_req_fn =    cfq_completed_request,
3862                 .elevator_former_req_fn =       elv_rb_former_request,
3863                 .elevator_latter_req_fn =       elv_rb_latter_request,
3864                 .elevator_init_icq_fn =         cfq_init_icq,
3865                 .elevator_exit_icq_fn =         cfq_exit_icq,
3866                 .elevator_set_req_fn =          cfq_set_request,
3867                 .elevator_put_req_fn =          cfq_put_request,
3868                 .elevator_may_queue_fn =        cfq_may_queue,
3869                 .elevator_init_fn =             cfq_init_queue,
3870                 .elevator_exit_fn =             cfq_exit_queue,
3871         },
3872         .icq_size       =       sizeof(struct cfq_io_cq),
3873         .icq_align      =       __alignof__(struct cfq_io_cq),
3874         .elevator_attrs =       cfq_attrs,
3875         .elevator_name  =       "cfq",
3876         .elevator_owner =       THIS_MODULE,
3877 };
3878
3879 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3880 static struct blkio_policy_type blkio_policy_cfq = {
3881         .ops = {
3882                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3883                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3884         },
3885         .plid = BLKIO_POLICY_PROP,
3886 };
3887 #else
3888 static struct blkio_policy_type blkio_policy_cfq;
3889 #endif
3890
3891 static int __init cfq_init(void)
3892 {
3893         int ret;
3894
3895         /*
3896          * could be 0 on HZ < 1000 setups
3897          */
3898         if (!cfq_slice_async)
3899                 cfq_slice_async = 1;
3900         if (!cfq_slice_idle)
3901                 cfq_slice_idle = 1;
3902
3903 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3904         if (!cfq_group_idle)
3905                 cfq_group_idle = 1;
3906 #else
3907                 cfq_group_idle = 0;
3908 #endif
3909         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3910         if (!cfq_pool)
3911                 return -ENOMEM;
3912
3913         ret = elv_register(&iosched_cfq);
3914         if (ret) {
3915                 kmem_cache_destroy(cfq_pool);
3916                 return ret;
3917         }
3918
3919         blkio_policy_register(&blkio_policy_cfq);
3920
3921         return 0;
3922 }
3923
3924 static void __exit cfq_exit(void)
3925 {
3926         blkio_policy_unregister(&blkio_policy_cfq);
3927         elv_unregister(&iosched_cfq);
3928         kmem_cache_destroy(cfq_pool);
3929 }
3930
3931 module_init(cfq_init);
3932 module_exit(cfq_exit);
3933
3934 MODULE_AUTHOR("Jens Axboe");
3935 MODULE_LICENSE("GPL");
3936 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");