block, cfq: move icq cache management to block core
[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 static struct kmem_cache *cfq_icq_pool;
63
64 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67
68 #define sample_valid(samples)   ((samples) > 80)
69 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
70
71 struct cfq_ttime {
72         unsigned long last_end_request;
73
74         unsigned long ttime_total;
75         unsigned long ttime_samples;
76         unsigned long ttime_mean;
77 };
78
79 /*
80  * Most of our rbtree usage is for sorting with min extraction, so
81  * if we cache the leftmost node we don't have to walk down the tree
82  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
83  * move this into the elevator for the rq sorting as well.
84  */
85 struct cfq_rb_root {
86         struct rb_root rb;
87         struct rb_node *left;
88         unsigned count;
89         unsigned total_weight;
90         u64 min_vdisktime;
91         struct cfq_ttime ttime;
92 };
93 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
94                         .ttime = {.last_end_request = jiffies,},}
95
96 /*
97  * Per process-grouping structure
98  */
99 struct cfq_queue {
100         /* reference count */
101         int ref;
102         /* various state flags, see below */
103         unsigned int flags;
104         /* parent cfq_data */
105         struct cfq_data *cfqd;
106         /* service_tree member */
107         struct rb_node rb_node;
108         /* service_tree key */
109         unsigned long rb_key;
110         /* prio tree member */
111         struct rb_node p_node;
112         /* prio tree root we belong to, if any */
113         struct rb_root *p_root;
114         /* sorted list of pending requests */
115         struct rb_root sort_list;
116         /* if fifo isn't expired, next request to serve */
117         struct request *next_rq;
118         /* requests queued in sort_list */
119         int queued[2];
120         /* currently allocated requests */
121         int allocated[2];
122         /* fifo list of requests in sort_list */
123         struct list_head fifo;
124
125         /* time when queue got scheduled in to dispatch first request. */
126         unsigned long dispatch_start;
127         unsigned int allocated_slice;
128         unsigned int slice_dispatch;
129         /* time when first request from queue completed and slice started. */
130         unsigned long slice_start;
131         unsigned long slice_end;
132         long slice_resid;
133
134         /* pending priority requests */
135         int prio_pending;
136         /* number of requests that are on the dispatch list or inside driver */
137         int dispatched;
138
139         /* io prio of this group */
140         unsigned short ioprio, org_ioprio;
141         unsigned short ioprio_class;
142
143         pid_t pid;
144
145         u32 seek_history;
146         sector_t last_request_pos;
147
148         struct cfq_rb_root *service_tree;
149         struct cfq_queue *new_cfqq;
150         struct cfq_group *cfqg;
151         /* Number of sectors dispatched from queue in single dispatch round */
152         unsigned long nr_sectors;
153 };
154
155 /*
156  * First index in the service_trees.
157  * IDLE is handled separately, so it has negative index
158  */
159 enum wl_prio_t {
160         BE_WORKLOAD = 0,
161         RT_WORKLOAD = 1,
162         IDLE_WORKLOAD = 2,
163         CFQ_PRIO_NR,
164 };
165
166 /*
167  * Second index in the service_trees.
168  */
169 enum wl_type_t {
170         ASYNC_WORKLOAD = 0,
171         SYNC_NOIDLE_WORKLOAD = 1,
172         SYNC_WORKLOAD = 2
173 };
174
175 /* This is per cgroup per device grouping structure */
176 struct cfq_group {
177         /* group service_tree member */
178         struct rb_node rb_node;
179
180         /* group service_tree key */
181         u64 vdisktime;
182         unsigned int weight;
183         unsigned int new_weight;
184         bool needs_update;
185
186         /* number of cfqq currently on this group */
187         int nr_cfqq;
188
189         /*
190          * Per group busy queues average. Useful for workload slice calc. We
191          * create the array for each prio class but at run time it is used
192          * only for RT and BE class and slot for IDLE class remains unused.
193          * This is primarily done to avoid confusion and a gcc warning.
194          */
195         unsigned int busy_queues_avg[CFQ_PRIO_NR];
196         /*
197          * rr lists of queues with requests. We maintain service trees for
198          * RT and BE classes. These trees are subdivided in subclasses
199          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
200          * class there is no subclassification and all the cfq queues go on
201          * a single tree service_tree_idle.
202          * Counts are embedded in the cfq_rb_root
203          */
204         struct cfq_rb_root service_trees[2][3];
205         struct cfq_rb_root service_tree_idle;
206
207         unsigned long saved_workload_slice;
208         enum wl_type_t saved_workload;
209         enum wl_prio_t saved_serving_prio;
210         struct blkio_group blkg;
211 #ifdef CONFIG_CFQ_GROUP_IOSCHED
212         struct hlist_node cfqd_node;
213         int ref;
214 #endif
215         /* number of requests that are on the dispatch list or inside driver */
216         int dispatched;
217         struct cfq_ttime ttime;
218 };
219
220 struct cfq_io_cq {
221         struct io_cq            icq;            /* must be the first member */
222         struct cfq_queue        *cfqq[2];
223         struct cfq_ttime        ttime;
224 };
225
226 /*
227  * Per block device queue structure
228  */
229 struct cfq_data {
230         struct request_queue *queue;
231         /* Root service tree for cfq_groups */
232         struct cfq_rb_root grp_service_tree;
233         struct cfq_group root_group;
234
235         /*
236          * The priority currently being served
237          */
238         enum wl_prio_t serving_prio;
239         enum wl_type_t serving_type;
240         unsigned long workload_expires;
241         struct cfq_group *serving_group;
242
243         /*
244          * Each priority tree is sorted by next_request position.  These
245          * trees are used when determining if two or more queues are
246          * interleaving requests (see cfq_close_cooperator).
247          */
248         struct rb_root prio_trees[CFQ_PRIO_LISTS];
249
250         unsigned int busy_queues;
251         unsigned int busy_sync_queues;
252
253         int rq_in_driver;
254         int rq_in_flight[2];
255
256         /*
257          * queue-depth detection
258          */
259         int rq_queued;
260         int hw_tag;
261         /*
262          * hw_tag can be
263          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
264          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
265          *  0 => no NCQ
266          */
267         int hw_tag_est_depth;
268         unsigned int hw_tag_samples;
269
270         /*
271          * idle window management
272          */
273         struct timer_list idle_slice_timer;
274         struct work_struct unplug_work;
275
276         struct cfq_queue *active_queue;
277         struct cfq_io_cq *active_cic;
278
279         /*
280          * async queue for each priority case
281          */
282         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
283         struct cfq_queue *async_idle_cfqq;
284
285         sector_t last_position;
286
287         /*
288          * tunables, see top of file
289          */
290         unsigned int cfq_quantum;
291         unsigned int cfq_fifo_expire[2];
292         unsigned int cfq_back_penalty;
293         unsigned int cfq_back_max;
294         unsigned int cfq_slice[2];
295         unsigned int cfq_slice_async_rq;
296         unsigned int cfq_slice_idle;
297         unsigned int cfq_group_idle;
298         unsigned int cfq_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 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         /*
1661          * reposition in fifo if next is older than rq
1662          */
1663         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1664             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1665                 list_move(&rq->queuelist, &next->queuelist);
1666                 rq_set_fifo_time(rq, rq_fifo_time(next));
1667         }
1668
1669         if (cfqq->next_rq == next)
1670                 cfqq->next_rq = rq;
1671         cfq_remove_request(next);
1672         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1673                                         rq_data_dir(next), rq_is_sync(next));
1674 }
1675
1676 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1677                            struct bio *bio)
1678 {
1679         struct cfq_data *cfqd = q->elevator->elevator_data;
1680         struct cfq_io_cq *cic;
1681         struct cfq_queue *cfqq;
1682
1683         /*
1684          * Disallow merge of a sync bio into an async request.
1685          */
1686         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1687                 return false;
1688
1689         /*
1690          * Lookup the cfqq that this bio will be queued with and allow
1691          * merge only if rq is queued there.  This function can be called
1692          * from plug merge without queue_lock.  In such cases, ioc of @rq
1693          * and %current are guaranteed to be equal.  Avoid lookup which
1694          * requires queue_lock by using @rq's cic.
1695          */
1696         if (current->io_context == RQ_CIC(rq)->icq.ioc) {
1697                 cic = RQ_CIC(rq);
1698         } else {
1699                 cic = cfq_cic_lookup(cfqd, current->io_context);
1700                 if (!cic)
1701                         return false;
1702         }
1703
1704         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1705         return cfqq == RQ_CFQQ(rq);
1706 }
1707
1708 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1709 {
1710         del_timer(&cfqd->idle_slice_timer);
1711         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1712 }
1713
1714 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1715                                    struct cfq_queue *cfqq)
1716 {
1717         if (cfqq) {
1718                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1719                                 cfqd->serving_prio, cfqd->serving_type);
1720                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1721                 cfqq->slice_start = 0;
1722                 cfqq->dispatch_start = jiffies;
1723                 cfqq->allocated_slice = 0;
1724                 cfqq->slice_end = 0;
1725                 cfqq->slice_dispatch = 0;
1726                 cfqq->nr_sectors = 0;
1727
1728                 cfq_clear_cfqq_wait_request(cfqq);
1729                 cfq_clear_cfqq_must_dispatch(cfqq);
1730                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1731                 cfq_clear_cfqq_fifo_expire(cfqq);
1732                 cfq_mark_cfqq_slice_new(cfqq);
1733
1734                 cfq_del_timer(cfqd, cfqq);
1735         }
1736
1737         cfqd->active_queue = cfqq;
1738 }
1739
1740 /*
1741  * current cfqq expired its slice (or was too idle), select new one
1742  */
1743 static void
1744 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1745                     bool timed_out)
1746 {
1747         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1748
1749         if (cfq_cfqq_wait_request(cfqq))
1750                 cfq_del_timer(cfqd, cfqq);
1751
1752         cfq_clear_cfqq_wait_request(cfqq);
1753         cfq_clear_cfqq_wait_busy(cfqq);
1754
1755         /*
1756          * If this cfqq is shared between multiple processes, check to
1757          * make sure that those processes are still issuing I/Os within
1758          * the mean seek distance.  If not, it may be time to break the
1759          * queues apart again.
1760          */
1761         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1762                 cfq_mark_cfqq_split_coop(cfqq);
1763
1764         /*
1765          * store what was left of this slice, if the queue idled/timed out
1766          */
1767         if (timed_out) {
1768                 if (cfq_cfqq_slice_new(cfqq))
1769                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1770                 else
1771                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1772                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1773         }
1774
1775         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1776
1777         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1778                 cfq_del_cfqq_rr(cfqd, cfqq);
1779
1780         cfq_resort_rr_list(cfqd, cfqq);
1781
1782         if (cfqq == cfqd->active_queue)
1783                 cfqd->active_queue = NULL;
1784
1785         if (cfqd->active_cic) {
1786                 put_io_context(cfqd->active_cic->icq.ioc, cfqd->queue);
1787                 cfqd->active_cic = NULL;
1788         }
1789 }
1790
1791 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1792 {
1793         struct cfq_queue *cfqq = cfqd->active_queue;
1794
1795         if (cfqq)
1796                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1797 }
1798
1799 /*
1800  * Get next queue for service. Unless we have a queue preemption,
1801  * we'll simply select the first cfqq in the service tree.
1802  */
1803 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1804 {
1805         struct cfq_rb_root *service_tree =
1806                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1807                                         cfqd->serving_type);
1808
1809         if (!cfqd->rq_queued)
1810                 return NULL;
1811
1812         /* There is nothing to dispatch */
1813         if (!service_tree)
1814                 return NULL;
1815         if (RB_EMPTY_ROOT(&service_tree->rb))
1816                 return NULL;
1817         return cfq_rb_first(service_tree);
1818 }
1819
1820 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1821 {
1822         struct cfq_group *cfqg;
1823         struct cfq_queue *cfqq;
1824         int i, j;
1825         struct cfq_rb_root *st;
1826
1827         if (!cfqd->rq_queued)
1828                 return NULL;
1829
1830         cfqg = cfq_get_next_cfqg(cfqd);
1831         if (!cfqg)
1832                 return NULL;
1833
1834         for_each_cfqg_st(cfqg, i, j, st)
1835                 if ((cfqq = cfq_rb_first(st)) != NULL)
1836                         return cfqq;
1837         return NULL;
1838 }
1839
1840 /*
1841  * Get and set a new active queue for service.
1842  */
1843 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1844                                               struct cfq_queue *cfqq)
1845 {
1846         if (!cfqq)
1847                 cfqq = cfq_get_next_queue(cfqd);
1848
1849         __cfq_set_active_queue(cfqd, cfqq);
1850         return cfqq;
1851 }
1852
1853 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1854                                           struct request *rq)
1855 {
1856         if (blk_rq_pos(rq) >= cfqd->last_position)
1857                 return blk_rq_pos(rq) - cfqd->last_position;
1858         else
1859                 return cfqd->last_position - blk_rq_pos(rq);
1860 }
1861
1862 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1863                                struct request *rq)
1864 {
1865         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1866 }
1867
1868 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1869                                     struct cfq_queue *cur_cfqq)
1870 {
1871         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1872         struct rb_node *parent, *node;
1873         struct cfq_queue *__cfqq;
1874         sector_t sector = cfqd->last_position;
1875
1876         if (RB_EMPTY_ROOT(root))
1877                 return NULL;
1878
1879         /*
1880          * First, if we find a request starting at the end of the last
1881          * request, choose it.
1882          */
1883         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1884         if (__cfqq)
1885                 return __cfqq;
1886
1887         /*
1888          * If the exact sector wasn't found, the parent of the NULL leaf
1889          * will contain the closest sector.
1890          */
1891         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1892         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1893                 return __cfqq;
1894
1895         if (blk_rq_pos(__cfqq->next_rq) < sector)
1896                 node = rb_next(&__cfqq->p_node);
1897         else
1898                 node = rb_prev(&__cfqq->p_node);
1899         if (!node)
1900                 return NULL;
1901
1902         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1903         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1904                 return __cfqq;
1905
1906         return NULL;
1907 }
1908
1909 /*
1910  * cfqd - obvious
1911  * cur_cfqq - passed in so that we don't decide that the current queue is
1912  *            closely cooperating with itself.
1913  *
1914  * So, basically we're assuming that that cur_cfqq has dispatched at least
1915  * one request, and that cfqd->last_position reflects a position on the disk
1916  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1917  * assumption.
1918  */
1919 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1920                                               struct cfq_queue *cur_cfqq)
1921 {
1922         struct cfq_queue *cfqq;
1923
1924         if (cfq_class_idle(cur_cfqq))
1925                 return NULL;
1926         if (!cfq_cfqq_sync(cur_cfqq))
1927                 return NULL;
1928         if (CFQQ_SEEKY(cur_cfqq))
1929                 return NULL;
1930
1931         /*
1932          * Don't search priority tree if it's the only queue in the group.
1933          */
1934         if (cur_cfqq->cfqg->nr_cfqq == 1)
1935                 return NULL;
1936
1937         /*
1938          * We should notice if some of the queues are cooperating, eg
1939          * working closely on the same area of the disk. In that case,
1940          * we can group them together and don't waste time idling.
1941          */
1942         cfqq = cfqq_close(cfqd, cur_cfqq);
1943         if (!cfqq)
1944                 return NULL;
1945
1946         /* If new queue belongs to different cfq_group, don't choose it */
1947         if (cur_cfqq->cfqg != cfqq->cfqg)
1948                 return NULL;
1949
1950         /*
1951          * It only makes sense to merge sync queues.
1952          */
1953         if (!cfq_cfqq_sync(cfqq))
1954                 return NULL;
1955         if (CFQQ_SEEKY(cfqq))
1956                 return NULL;
1957
1958         /*
1959          * Do not merge queues of different priority classes
1960          */
1961         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1962                 return NULL;
1963
1964         return cfqq;
1965 }
1966
1967 /*
1968  * Determine whether we should enforce idle window for this queue.
1969  */
1970
1971 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1972 {
1973         enum wl_prio_t prio = cfqq_prio(cfqq);
1974         struct cfq_rb_root *service_tree = cfqq->service_tree;
1975
1976         BUG_ON(!service_tree);
1977         BUG_ON(!service_tree->count);
1978
1979         if (!cfqd->cfq_slice_idle)
1980                 return false;
1981
1982         /* We never do for idle class queues. */
1983         if (prio == IDLE_WORKLOAD)
1984                 return false;
1985
1986         /* We do for queues that were marked with idle window flag. */
1987         if (cfq_cfqq_idle_window(cfqq) &&
1988            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1989                 return true;
1990
1991         /*
1992          * Otherwise, we do only if they are the last ones
1993          * in their service tree.
1994          */
1995         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1996            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1997                 return true;
1998         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1999                         service_tree->count);
2000         return false;
2001 }
2002
2003 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2004 {
2005         struct cfq_queue *cfqq = cfqd->active_queue;
2006         struct cfq_io_cq *cic;
2007         unsigned long sl, group_idle = 0;
2008
2009         /*
2010          * SSD device without seek penalty, disable idling. But only do so
2011          * for devices that support queuing, otherwise we still have a problem
2012          * with sync vs async workloads.
2013          */
2014         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2015                 return;
2016
2017         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2018         WARN_ON(cfq_cfqq_slice_new(cfqq));
2019
2020         /*
2021          * idle is disabled, either manually or by past process history
2022          */
2023         if (!cfq_should_idle(cfqd, cfqq)) {
2024                 /* no queue idling. Check for group idling */
2025                 if (cfqd->cfq_group_idle)
2026                         group_idle = cfqd->cfq_group_idle;
2027                 else
2028                         return;
2029         }
2030
2031         /*
2032          * still active requests from this queue, don't idle
2033          */
2034         if (cfqq->dispatched)
2035                 return;
2036
2037         /*
2038          * task has exited, don't wait
2039          */
2040         cic = cfqd->active_cic;
2041         if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
2042                 return;
2043
2044         /*
2045          * If our average think time is larger than the remaining time
2046          * slice, then don't idle. This avoids overrunning the allotted
2047          * time slice.
2048          */
2049         if (sample_valid(cic->ttime.ttime_samples) &&
2050             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2051                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2052                              cic->ttime.ttime_mean);
2053                 return;
2054         }
2055
2056         /* There are other queues in the group, don't do group idle */
2057         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2058                 return;
2059
2060         cfq_mark_cfqq_wait_request(cfqq);
2061
2062         if (group_idle)
2063                 sl = cfqd->cfq_group_idle;
2064         else
2065                 sl = cfqd->cfq_slice_idle;
2066
2067         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2068         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2069         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2070                         group_idle ? 1 : 0);
2071 }
2072
2073 /*
2074  * Move request from internal lists to the request queue dispatch list.
2075  */
2076 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2077 {
2078         struct cfq_data *cfqd = q->elevator->elevator_data;
2079         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2080
2081         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2082
2083         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2084         cfq_remove_request(rq);
2085         cfqq->dispatched++;
2086         (RQ_CFQG(rq))->dispatched++;
2087         elv_dispatch_sort(q, rq);
2088
2089         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2090         cfqq->nr_sectors += blk_rq_sectors(rq);
2091         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2092                                         rq_data_dir(rq), rq_is_sync(rq));
2093 }
2094
2095 /*
2096  * return expired entry, or NULL to just start from scratch in rbtree
2097  */
2098 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2099 {
2100         struct request *rq = NULL;
2101
2102         if (cfq_cfqq_fifo_expire(cfqq))
2103                 return NULL;
2104
2105         cfq_mark_cfqq_fifo_expire(cfqq);
2106
2107         if (list_empty(&cfqq->fifo))
2108                 return NULL;
2109
2110         rq = rq_entry_fifo(cfqq->fifo.next);
2111         if (time_before(jiffies, rq_fifo_time(rq)))
2112                 rq = NULL;
2113
2114         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2115         return rq;
2116 }
2117
2118 static inline int
2119 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2120 {
2121         const int base_rq = cfqd->cfq_slice_async_rq;
2122
2123         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2124
2125         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2126 }
2127
2128 /*
2129  * Must be called with the queue_lock held.
2130  */
2131 static int cfqq_process_refs(struct cfq_queue *cfqq)
2132 {
2133         int process_refs, io_refs;
2134
2135         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2136         process_refs = cfqq->ref - io_refs;
2137         BUG_ON(process_refs < 0);
2138         return process_refs;
2139 }
2140
2141 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2142 {
2143         int process_refs, new_process_refs;
2144         struct cfq_queue *__cfqq;
2145
2146         /*
2147          * If there are no process references on the new_cfqq, then it is
2148          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2149          * chain may have dropped their last reference (not just their
2150          * last process reference).
2151          */
2152         if (!cfqq_process_refs(new_cfqq))
2153                 return;
2154
2155         /* Avoid a circular list and skip interim queue merges */
2156         while ((__cfqq = new_cfqq->new_cfqq)) {
2157                 if (__cfqq == cfqq)
2158                         return;
2159                 new_cfqq = __cfqq;
2160         }
2161
2162         process_refs = cfqq_process_refs(cfqq);
2163         new_process_refs = cfqq_process_refs(new_cfqq);
2164         /*
2165          * If the process for the cfqq has gone away, there is no
2166          * sense in merging the queues.
2167          */
2168         if (process_refs == 0 || new_process_refs == 0)
2169                 return;
2170
2171         /*
2172          * Merge in the direction of the lesser amount of work.
2173          */
2174         if (new_process_refs >= process_refs) {
2175                 cfqq->new_cfqq = new_cfqq;
2176                 new_cfqq->ref += process_refs;
2177         } else {
2178                 new_cfqq->new_cfqq = cfqq;
2179                 cfqq->ref += new_process_refs;
2180         }
2181 }
2182
2183 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2184                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2185 {
2186         struct cfq_queue *queue;
2187         int i;
2188         bool key_valid = false;
2189         unsigned long lowest_key = 0;
2190         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2191
2192         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2193                 /* select the one with lowest rb_key */
2194                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2195                 if (queue &&
2196                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2197                         lowest_key = queue->rb_key;
2198                         cur_best = i;
2199                         key_valid = true;
2200                 }
2201         }
2202
2203         return cur_best;
2204 }
2205
2206 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2207 {
2208         unsigned slice;
2209         unsigned count;
2210         struct cfq_rb_root *st;
2211         unsigned group_slice;
2212         enum wl_prio_t original_prio = cfqd->serving_prio;
2213
2214         /* Choose next priority. RT > BE > IDLE */
2215         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2216                 cfqd->serving_prio = RT_WORKLOAD;
2217         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2218                 cfqd->serving_prio = BE_WORKLOAD;
2219         else {
2220                 cfqd->serving_prio = IDLE_WORKLOAD;
2221                 cfqd->workload_expires = jiffies + 1;
2222                 return;
2223         }
2224
2225         if (original_prio != cfqd->serving_prio)
2226                 goto new_workload;
2227
2228         /*
2229          * For RT and BE, we have to choose also the type
2230          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2231          * expiration time
2232          */
2233         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2234         count = st->count;
2235
2236         /*
2237          * check workload expiration, and that we still have other queues ready
2238          */
2239         if (count && !time_after(jiffies, cfqd->workload_expires))
2240                 return;
2241
2242 new_workload:
2243         /* otherwise select new workload type */
2244         cfqd->serving_type =
2245                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2246         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2247         count = st->count;
2248
2249         /*
2250          * the workload slice is computed as a fraction of target latency
2251          * proportional to the number of queues in that workload, over
2252          * all the queues in the same priority class
2253          */
2254         group_slice = cfq_group_slice(cfqd, cfqg);
2255
2256         slice = group_slice * count /
2257                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2258                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2259
2260         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2261                 unsigned int tmp;
2262
2263                 /*
2264                  * Async queues are currently system wide. Just taking
2265                  * proportion of queues with-in same group will lead to higher
2266                  * async ratio system wide as generally root group is going
2267                  * to have higher weight. A more accurate thing would be to
2268                  * calculate system wide asnc/sync ratio.
2269                  */
2270                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2271                 tmp = tmp/cfqd->busy_queues;
2272                 slice = min_t(unsigned, slice, tmp);
2273
2274                 /* async workload slice is scaled down according to
2275                  * the sync/async slice ratio. */
2276                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2277         } else
2278                 /* sync workload slice is at least 2 * cfq_slice_idle */
2279                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2280
2281         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2282         cfq_log(cfqd, "workload slice:%d", slice);
2283         cfqd->workload_expires = jiffies + slice;
2284 }
2285
2286 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2287 {
2288         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2289         struct cfq_group *cfqg;
2290
2291         if (RB_EMPTY_ROOT(&st->rb))
2292                 return NULL;
2293         cfqg = cfq_rb_first_group(st);
2294         update_min_vdisktime(st);
2295         return cfqg;
2296 }
2297
2298 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2299 {
2300         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2301
2302         cfqd->serving_group = cfqg;
2303
2304         /* Restore the workload type data */
2305         if (cfqg->saved_workload_slice) {
2306                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2307                 cfqd->serving_type = cfqg->saved_workload;
2308                 cfqd->serving_prio = cfqg->saved_serving_prio;
2309         } else
2310                 cfqd->workload_expires = jiffies - 1;
2311
2312         choose_service_tree(cfqd, cfqg);
2313 }
2314
2315 /*
2316  * Select a queue for service. If we have a current active queue,
2317  * check whether to continue servicing it, or retrieve and set a new one.
2318  */
2319 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2320 {
2321         struct cfq_queue *cfqq, *new_cfqq = NULL;
2322
2323         cfqq = cfqd->active_queue;
2324         if (!cfqq)
2325                 goto new_queue;
2326
2327         if (!cfqd->rq_queued)
2328                 return NULL;
2329
2330         /*
2331          * We were waiting for group to get backlogged. Expire the queue
2332          */
2333         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2334                 goto expire;
2335
2336         /*
2337          * The active queue has run out of time, expire it and select new.
2338          */
2339         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2340                 /*
2341                  * If slice had not expired at the completion of last request
2342                  * we might not have turned on wait_busy flag. Don't expire
2343                  * the queue yet. Allow the group to get backlogged.
2344                  *
2345                  * The very fact that we have used the slice, that means we
2346                  * have been idling all along on this queue and it should be
2347                  * ok to wait for this request to complete.
2348                  */
2349                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2350                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2351                         cfqq = NULL;
2352                         goto keep_queue;
2353                 } else
2354                         goto check_group_idle;
2355         }
2356
2357         /*
2358          * The active queue has requests and isn't expired, allow it to
2359          * dispatch.
2360          */
2361         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2362                 goto keep_queue;
2363
2364         /*
2365          * If another queue has a request waiting within our mean seek
2366          * distance, let it run.  The expire code will check for close
2367          * cooperators and put the close queue at the front of the service
2368          * tree.  If possible, merge the expiring queue with the new cfqq.
2369          */
2370         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2371         if (new_cfqq) {
2372                 if (!cfqq->new_cfqq)
2373                         cfq_setup_merge(cfqq, new_cfqq);
2374                 goto expire;
2375         }
2376
2377         /*
2378          * No requests pending. If the active queue still has requests in
2379          * flight or is idling for a new request, allow either of these
2380          * conditions to happen (or time out) before selecting a new queue.
2381          */
2382         if (timer_pending(&cfqd->idle_slice_timer)) {
2383                 cfqq = NULL;
2384                 goto keep_queue;
2385         }
2386
2387         /*
2388          * This is a deep seek queue, but the device is much faster than
2389          * the queue can deliver, don't idle
2390          **/
2391         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2392             (cfq_cfqq_slice_new(cfqq) ||
2393             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2394                 cfq_clear_cfqq_deep(cfqq);
2395                 cfq_clear_cfqq_idle_window(cfqq);
2396         }
2397
2398         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2399                 cfqq = NULL;
2400                 goto keep_queue;
2401         }
2402
2403         /*
2404          * If group idle is enabled and there are requests dispatched from
2405          * this group, wait for requests to complete.
2406          */
2407 check_group_idle:
2408         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2409             cfqq->cfqg->dispatched &&
2410             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2411                 cfqq = NULL;
2412                 goto keep_queue;
2413         }
2414
2415 expire:
2416         cfq_slice_expired(cfqd, 0);
2417 new_queue:
2418         /*
2419          * Current queue expired. Check if we have to switch to a new
2420          * service tree
2421          */
2422         if (!new_cfqq)
2423                 cfq_choose_cfqg(cfqd);
2424
2425         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2426 keep_queue:
2427         return cfqq;
2428 }
2429
2430 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2431 {
2432         int dispatched = 0;
2433
2434         while (cfqq->next_rq) {
2435                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2436                 dispatched++;
2437         }
2438
2439         BUG_ON(!list_empty(&cfqq->fifo));
2440
2441         /* By default cfqq is not expired if it is empty. Do it explicitly */
2442         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2443         return dispatched;
2444 }
2445
2446 /*
2447  * Drain our current requests. Used for barriers and when switching
2448  * io schedulers on-the-fly.
2449  */
2450 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2451 {
2452         struct cfq_queue *cfqq;
2453         int dispatched = 0;
2454
2455         /* Expire the timeslice of the current active queue first */
2456         cfq_slice_expired(cfqd, 0);
2457         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2458                 __cfq_set_active_queue(cfqd, cfqq);
2459                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2460         }
2461
2462         BUG_ON(cfqd->busy_queues);
2463
2464         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2465         return dispatched;
2466 }
2467
2468 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2469         struct cfq_queue *cfqq)
2470 {
2471         /* the queue hasn't finished any request, can't estimate */
2472         if (cfq_cfqq_slice_new(cfqq))
2473                 return true;
2474         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2475                 cfqq->slice_end))
2476                 return true;
2477
2478         return false;
2479 }
2480
2481 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2482 {
2483         unsigned int max_dispatch;
2484
2485         /*
2486          * Drain async requests before we start sync IO
2487          */
2488         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2489                 return false;
2490
2491         /*
2492          * If this is an async queue and we have sync IO in flight, let it wait
2493          */
2494         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2495                 return false;
2496
2497         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2498         if (cfq_class_idle(cfqq))
2499                 max_dispatch = 1;
2500
2501         /*
2502          * Does this cfqq already have too much IO in flight?
2503          */
2504         if (cfqq->dispatched >= max_dispatch) {
2505                 bool promote_sync = false;
2506                 /*
2507                  * idle queue must always only have a single IO in flight
2508                  */
2509                 if (cfq_class_idle(cfqq))
2510                         return false;
2511
2512                 /*
2513                  * If there is only one sync queue
2514                  * we can ignore async queue here and give the sync
2515                  * queue no dispatch limit. The reason is a sync queue can
2516                  * preempt async queue, limiting the sync queue doesn't make
2517                  * sense. This is useful for aiostress test.
2518                  */
2519                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2520                         promote_sync = true;
2521
2522                 /*
2523                  * We have other queues, don't allow more IO from this one
2524                  */
2525                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2526                                 !promote_sync)
2527                         return false;
2528
2529                 /*
2530                  * Sole queue user, no limit
2531                  */
2532                 if (cfqd->busy_queues == 1 || promote_sync)
2533                         max_dispatch = -1;
2534                 else
2535                         /*
2536                          * Normally we start throttling cfqq when cfq_quantum/2
2537                          * requests have been dispatched. But we can drive
2538                          * deeper queue depths at the beginning of slice
2539                          * subjected to upper limit of cfq_quantum.
2540                          * */
2541                         max_dispatch = cfqd->cfq_quantum;
2542         }
2543
2544         /*
2545          * Async queues must wait a bit before being allowed dispatch.
2546          * We also ramp up the dispatch depth gradually for async IO,
2547          * based on the last sync IO we serviced
2548          */
2549         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2550                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2551                 unsigned int depth;
2552
2553                 depth = last_sync / cfqd->cfq_slice[1];
2554                 if (!depth && !cfqq->dispatched)
2555                         depth = 1;
2556                 if (depth < max_dispatch)
2557                         max_dispatch = depth;
2558         }
2559
2560         /*
2561          * If we're below the current max, allow a dispatch
2562          */
2563         return cfqq->dispatched < max_dispatch;
2564 }
2565
2566 /*
2567  * Dispatch a request from cfqq, moving them to the request queue
2568  * dispatch list.
2569  */
2570 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2571 {
2572         struct request *rq;
2573
2574         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2575
2576         if (!cfq_may_dispatch(cfqd, cfqq))
2577                 return false;
2578
2579         /*
2580          * follow expired path, else get first next available
2581          */
2582         rq = cfq_check_fifo(cfqq);
2583         if (!rq)
2584                 rq = cfqq->next_rq;
2585
2586         /*
2587          * insert request into driver dispatch list
2588          */
2589         cfq_dispatch_insert(cfqd->queue, rq);
2590
2591         if (!cfqd->active_cic) {
2592                 struct cfq_io_cq *cic = RQ_CIC(rq);
2593
2594                 atomic_long_inc(&cic->icq.ioc->refcount);
2595                 cfqd->active_cic = cic;
2596         }
2597
2598         return true;
2599 }
2600
2601 /*
2602  * Find the cfqq that we need to service and move a request from that to the
2603  * dispatch list
2604  */
2605 static int cfq_dispatch_requests(struct request_queue *q, int force)
2606 {
2607         struct cfq_data *cfqd = q->elevator->elevator_data;
2608         struct cfq_queue *cfqq;
2609
2610         if (!cfqd->busy_queues)
2611                 return 0;
2612
2613         if (unlikely(force))
2614                 return cfq_forced_dispatch(cfqd);
2615
2616         cfqq = cfq_select_queue(cfqd);
2617         if (!cfqq)
2618                 return 0;
2619
2620         /*
2621          * Dispatch a request from this cfqq, if it is allowed
2622          */
2623         if (!cfq_dispatch_request(cfqd, cfqq))
2624                 return 0;
2625
2626         cfqq->slice_dispatch++;
2627         cfq_clear_cfqq_must_dispatch(cfqq);
2628
2629         /*
2630          * expire an async queue immediately if it has used up its slice. idle
2631          * queue always expire after 1 dispatch round.
2632          */
2633         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2634             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2635             cfq_class_idle(cfqq))) {
2636                 cfqq->slice_end = jiffies + 1;
2637                 cfq_slice_expired(cfqd, 0);
2638         }
2639
2640         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2641         return 1;
2642 }
2643
2644 /*
2645  * task holds one reference to the queue, dropped when task exits. each rq
2646  * in-flight on this queue also holds a reference, dropped when rq is freed.
2647  *
2648  * Each cfq queue took a reference on the parent group. Drop it now.
2649  * queue lock must be held here.
2650  */
2651 static void cfq_put_queue(struct cfq_queue *cfqq)
2652 {
2653         struct cfq_data *cfqd = cfqq->cfqd;
2654         struct cfq_group *cfqg;
2655
2656         BUG_ON(cfqq->ref <= 0);
2657
2658         cfqq->ref--;
2659         if (cfqq->ref)
2660                 return;
2661
2662         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2663         BUG_ON(rb_first(&cfqq->sort_list));
2664         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2665         cfqg = cfqq->cfqg;
2666
2667         if (unlikely(cfqd->active_queue == cfqq)) {
2668                 __cfq_slice_expired(cfqd, cfqq, 0);
2669                 cfq_schedule_dispatch(cfqd);
2670         }
2671
2672         BUG_ON(cfq_cfqq_on_rr(cfqq));
2673         kmem_cache_free(cfq_pool, cfqq);
2674         cfq_put_cfqg(cfqg);
2675 }
2676
2677 static void cfq_icq_free_rcu(struct rcu_head *head)
2678 {
2679         kmem_cache_free(cfq_icq_pool,
2680                         icq_to_cic(container_of(head, struct io_cq, rcu_head)));
2681 }
2682
2683 static void cfq_icq_free(struct io_cq *icq)
2684 {
2685         call_rcu(&icq->rcu_head, cfq_icq_free_rcu);
2686 }
2687
2688 static void cfq_release_icq(struct io_cq *icq)
2689 {
2690         struct io_context *ioc = icq->ioc;
2691
2692         radix_tree_delete(&ioc->icq_tree, icq->q->id);
2693         hlist_del(&icq->ioc_node);
2694         cfq_icq_free(icq);
2695 }
2696
2697 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2698 {
2699         struct cfq_queue *__cfqq, *next;
2700
2701         /*
2702          * If this queue was scheduled to merge with another queue, be
2703          * sure to drop the reference taken on that queue (and others in
2704          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2705          */
2706         __cfqq = cfqq->new_cfqq;
2707         while (__cfqq) {
2708                 if (__cfqq == cfqq) {
2709                         WARN(1, "cfqq->new_cfqq loop detected\n");
2710                         break;
2711                 }
2712                 next = __cfqq->new_cfqq;
2713                 cfq_put_queue(__cfqq);
2714                 __cfqq = next;
2715         }
2716 }
2717
2718 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2719 {
2720         if (unlikely(cfqq == cfqd->active_queue)) {
2721                 __cfq_slice_expired(cfqd, cfqq, 0);
2722                 cfq_schedule_dispatch(cfqd);
2723         }
2724
2725         cfq_put_cooperator(cfqq);
2726
2727         cfq_put_queue(cfqq);
2728 }
2729
2730 static void cfq_exit_icq(struct io_cq *icq)
2731 {
2732         struct cfq_io_cq *cic = icq_to_cic(icq);
2733         struct cfq_data *cfqd = cic_to_cfqd(cic);
2734         struct io_context *ioc = icq->ioc;
2735
2736         list_del_init(&icq->q_node);
2737
2738         /*
2739          * Both setting lookup hint to and clearing it from @icq are done
2740          * under queue_lock.  If it's not pointing to @icq now, it never
2741          * will.  Hint assignment itself can race safely.
2742          */
2743         if (rcu_dereference_raw(ioc->icq_hint) == icq)
2744                 rcu_assign_pointer(ioc->icq_hint, NULL);
2745
2746         if (cic->cfqq[BLK_RW_ASYNC]) {
2747                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2748                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2749         }
2750
2751         if (cic->cfqq[BLK_RW_SYNC]) {
2752                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2753                 cic->cfqq[BLK_RW_SYNC] = NULL;
2754         }
2755 }
2756
2757 static struct cfq_io_cq *cfq_alloc_cic(struct cfq_data *cfqd, gfp_t gfp_mask)
2758 {
2759         struct cfq_io_cq *cic;
2760
2761         cic = kmem_cache_alloc_node(cfq_icq_pool, gfp_mask | __GFP_ZERO,
2762                                                         cfqd->queue->node);
2763         if (cic) {
2764                 cic->ttime.last_end_request = jiffies;
2765                 INIT_LIST_HEAD(&cic->icq.q_node);
2766                 INIT_HLIST_NODE(&cic->icq.ioc_node);
2767                 cic->icq.exit = cfq_exit_icq;
2768                 cic->icq.release = cfq_release_icq;
2769         }
2770
2771         return cic;
2772 }
2773
2774 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2775 {
2776         struct task_struct *tsk = current;
2777         int ioprio_class;
2778
2779         if (!cfq_cfqq_prio_changed(cfqq))
2780                 return;
2781
2782         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2783         switch (ioprio_class) {
2784         default:
2785                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2786         case IOPRIO_CLASS_NONE:
2787                 /*
2788                  * no prio set, inherit CPU scheduling settings
2789                  */
2790                 cfqq->ioprio = task_nice_ioprio(tsk);
2791                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2792                 break;
2793         case IOPRIO_CLASS_RT:
2794                 cfqq->ioprio = task_ioprio(ioc);
2795                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2796                 break;
2797         case IOPRIO_CLASS_BE:
2798                 cfqq->ioprio = task_ioprio(ioc);
2799                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2800                 break;
2801         case IOPRIO_CLASS_IDLE:
2802                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2803                 cfqq->ioprio = 7;
2804                 cfq_clear_cfqq_idle_window(cfqq);
2805                 break;
2806         }
2807
2808         /*
2809          * keep track of original prio settings in case we have to temporarily
2810          * elevate the priority of this queue
2811          */
2812         cfqq->org_ioprio = cfqq->ioprio;
2813         cfq_clear_cfqq_prio_changed(cfqq);
2814 }
2815
2816 static void changed_ioprio(struct cfq_io_cq *cic)
2817 {
2818         struct cfq_data *cfqd = cic_to_cfqd(cic);
2819         struct cfq_queue *cfqq;
2820
2821         if (unlikely(!cfqd))
2822                 return;
2823
2824         cfqq = cic->cfqq[BLK_RW_ASYNC];
2825         if (cfqq) {
2826                 struct cfq_queue *new_cfqq;
2827                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2828                                                 GFP_ATOMIC);
2829                 if (new_cfqq) {
2830                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2831                         cfq_put_queue(cfqq);
2832                 }
2833         }
2834
2835         cfqq = cic->cfqq[BLK_RW_SYNC];
2836         if (cfqq)
2837                 cfq_mark_cfqq_prio_changed(cfqq);
2838 }
2839
2840 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2841                           pid_t pid, bool is_sync)
2842 {
2843         RB_CLEAR_NODE(&cfqq->rb_node);
2844         RB_CLEAR_NODE(&cfqq->p_node);
2845         INIT_LIST_HEAD(&cfqq->fifo);
2846
2847         cfqq->ref = 0;
2848         cfqq->cfqd = cfqd;
2849
2850         cfq_mark_cfqq_prio_changed(cfqq);
2851
2852         if (is_sync) {
2853                 if (!cfq_class_idle(cfqq))
2854                         cfq_mark_cfqq_idle_window(cfqq);
2855                 cfq_mark_cfqq_sync(cfqq);
2856         }
2857         cfqq->pid = pid;
2858 }
2859
2860 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2861 static void changed_cgroup(struct cfq_io_cq *cic)
2862 {
2863         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2864         struct cfq_data *cfqd = cic_to_cfqd(cic);
2865         struct request_queue *q;
2866
2867         if (unlikely(!cfqd))
2868                 return;
2869
2870         q = cfqd->queue;
2871
2872         if (sync_cfqq) {
2873                 /*
2874                  * Drop reference to sync queue. A new sync queue will be
2875                  * assigned in new group upon arrival of a fresh request.
2876                  */
2877                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2878                 cic_set_cfqq(cic, NULL, 1);
2879                 cfq_put_queue(sync_cfqq);
2880         }
2881 }
2882 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2883
2884 static struct cfq_queue *
2885 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2886                      struct io_context *ioc, gfp_t gfp_mask)
2887 {
2888         struct cfq_queue *cfqq, *new_cfqq = NULL;
2889         struct cfq_io_cq *cic;
2890         struct cfq_group *cfqg;
2891
2892 retry:
2893         cfqg = cfq_get_cfqg(cfqd);
2894         cic = cfq_cic_lookup(cfqd, ioc);
2895         /* cic always exists here */
2896         cfqq = cic_to_cfqq(cic, is_sync);
2897
2898         /*
2899          * Always try a new alloc if we fell back to the OOM cfqq
2900          * originally, since it should just be a temporary situation.
2901          */
2902         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2903                 cfqq = NULL;
2904                 if (new_cfqq) {
2905                         cfqq = new_cfqq;
2906                         new_cfqq = NULL;
2907                 } else if (gfp_mask & __GFP_WAIT) {
2908                         spin_unlock_irq(cfqd->queue->queue_lock);
2909                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2910                                         gfp_mask | __GFP_ZERO,
2911                                         cfqd->queue->node);
2912                         spin_lock_irq(cfqd->queue->queue_lock);
2913                         if (new_cfqq)
2914                                 goto retry;
2915                 } else {
2916                         cfqq = kmem_cache_alloc_node(cfq_pool,
2917                                         gfp_mask | __GFP_ZERO,
2918                                         cfqd->queue->node);
2919                 }
2920
2921                 if (cfqq) {
2922                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2923                         cfq_init_prio_data(cfqq, ioc);
2924                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2925                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2926                 } else
2927                         cfqq = &cfqd->oom_cfqq;
2928         }
2929
2930         if (new_cfqq)
2931                 kmem_cache_free(cfq_pool, new_cfqq);
2932
2933         return cfqq;
2934 }
2935
2936 static struct cfq_queue **
2937 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2938 {
2939         switch (ioprio_class) {
2940         case IOPRIO_CLASS_RT:
2941                 return &cfqd->async_cfqq[0][ioprio];
2942         case IOPRIO_CLASS_BE:
2943                 return &cfqd->async_cfqq[1][ioprio];
2944         case IOPRIO_CLASS_IDLE:
2945                 return &cfqd->async_idle_cfqq;
2946         default:
2947                 BUG();
2948         }
2949 }
2950
2951 static struct cfq_queue *
2952 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2953               gfp_t gfp_mask)
2954 {
2955         const int ioprio = task_ioprio(ioc);
2956         const int ioprio_class = task_ioprio_class(ioc);
2957         struct cfq_queue **async_cfqq = NULL;
2958         struct cfq_queue *cfqq = NULL;
2959
2960         if (!is_sync) {
2961                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2962                 cfqq = *async_cfqq;
2963         }
2964
2965         if (!cfqq)
2966                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2967
2968         /*
2969          * pin the queue now that it's allocated, scheduler exit will prune it
2970          */
2971         if (!is_sync && !(*async_cfqq)) {
2972                 cfqq->ref++;
2973                 *async_cfqq = cfqq;
2974         }
2975
2976         cfqq->ref++;
2977         return cfqq;
2978 }
2979
2980 /**
2981  * cfq_create_cic - create and link a cfq_io_cq
2982  * @cfqd: cfqd of interest
2983  * @gfp_mask: allocation mask
2984  *
2985  * Make sure cfq_io_cq linking %current->io_context and @cfqd exists.  If
2986  * ioc and/or cic doesn't exist, they will be created using @gfp_mask.
2987  */
2988 static int cfq_create_cic(struct cfq_data *cfqd, gfp_t gfp_mask)
2989 {
2990         struct request_queue *q = cfqd->queue;
2991         struct io_cq *icq = NULL;
2992         struct cfq_io_cq *cic;
2993         struct io_context *ioc;
2994         int ret = -ENOMEM;
2995
2996         might_sleep_if(gfp_mask & __GFP_WAIT);
2997
2998         /* allocate stuff */
2999         ioc = create_io_context(current, gfp_mask, q->node);
3000         if (!ioc)
3001                 goto out;
3002
3003         cic = cfq_alloc_cic(cfqd, gfp_mask);
3004         if (!cic)
3005                 goto out;
3006         icq = &cic->icq;
3007
3008         ret = radix_tree_preload(gfp_mask);
3009         if (ret)
3010                 goto out;
3011
3012         icq->ioc = ioc;
3013         icq->q = cfqd->queue;
3014
3015         /* lock both q and ioc and try to link @icq */
3016         spin_lock_irq(q->queue_lock);
3017         spin_lock(&ioc->lock);
3018
3019         ret = radix_tree_insert(&ioc->icq_tree, q->id, icq);
3020         if (likely(!ret)) {
3021                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
3022                 list_add(&icq->q_node, &q->icq_list);
3023                 icq = NULL;
3024         } else if (ret == -EEXIST) {
3025                 /* someone else already did it */
3026                 ret = 0;
3027         }
3028
3029         spin_unlock(&ioc->lock);
3030         spin_unlock_irq(q->queue_lock);
3031
3032         radix_tree_preload_end();
3033 out:
3034         if (ret)
3035                 printk(KERN_ERR "cfq: icq link failed!\n");
3036         if (icq)
3037                 cfq_icq_free(icq);
3038         return ret;
3039 }
3040
3041 /**
3042  * cfq_get_cic - acquire cfq_io_cq and bump refcnt on io_context
3043  * @cfqd: cfqd to setup cic for
3044  * @gfp_mask: allocation mask
3045  *
3046  * Return cfq_io_cq associating @cfqd and %current->io_context and
3047  * bump refcnt on io_context.  If ioc or cic doesn't exist, they're created
3048  * using @gfp_mask.
3049  *
3050  * Must be called under queue_lock which may be released and re-acquired.
3051  * This function also may sleep depending on @gfp_mask.
3052  */
3053 static struct cfq_io_cq *cfq_get_cic(struct cfq_data *cfqd, gfp_t gfp_mask)
3054 {
3055         struct request_queue *q = cfqd->queue;
3056         struct cfq_io_cq *cic = NULL;
3057         struct io_context *ioc;
3058         int err;
3059
3060         lockdep_assert_held(q->queue_lock);
3061
3062         while (true) {
3063                 /* fast path */
3064                 ioc = current->io_context;
3065                 if (likely(ioc)) {
3066                         cic = cfq_cic_lookup(cfqd, ioc);
3067                         if (likely(cic))
3068                                 break;
3069                 }
3070
3071                 /* slow path - unlock, create missing ones and retry */
3072                 spin_unlock_irq(q->queue_lock);
3073                 err = cfq_create_cic(cfqd, gfp_mask);
3074                 spin_lock_irq(q->queue_lock);
3075                 if (err)
3076                         return NULL;
3077         }
3078
3079         /* bump @ioc's refcnt and handle changed notifications */
3080         get_io_context(ioc);
3081
3082         if (unlikely(cic->icq.changed)) {
3083                 if (test_and_clear_bit(ICQ_IOPRIO_CHANGED, &cic->icq.changed))
3084                         changed_ioprio(cic);
3085 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3086                 if (test_and_clear_bit(ICQ_CGROUP_CHANGED, &cic->icq.changed))
3087                         changed_cgroup(cic);
3088 #endif
3089         }
3090
3091         return cic;
3092 }
3093
3094 static void
3095 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3096 {
3097         unsigned long elapsed = jiffies - ttime->last_end_request;
3098         elapsed = min(elapsed, 2UL * slice_idle);
3099
3100         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3101         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3102         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3103 }
3104
3105 static void
3106 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3107                         struct cfq_io_cq *cic)
3108 {
3109         if (cfq_cfqq_sync(cfqq)) {
3110                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3111                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3112                         cfqd->cfq_slice_idle);
3113         }
3114 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3115         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3116 #endif
3117 }
3118
3119 static void
3120 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3121                        struct request *rq)
3122 {
3123         sector_t sdist = 0;
3124         sector_t n_sec = blk_rq_sectors(rq);
3125         if (cfqq->last_request_pos) {
3126                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3127                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3128                 else
3129                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3130         }
3131
3132         cfqq->seek_history <<= 1;
3133         if (blk_queue_nonrot(cfqd->queue))
3134                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3135         else
3136                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3137 }
3138
3139 /*
3140  * Disable idle window if the process thinks too long or seeks so much that
3141  * it doesn't matter
3142  */
3143 static void
3144 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3145                        struct cfq_io_cq *cic)
3146 {
3147         int old_idle, enable_idle;
3148
3149         /*
3150          * Don't idle for async or idle io prio class
3151          */
3152         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3153                 return;
3154
3155         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3156
3157         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3158                 cfq_mark_cfqq_deep(cfqq);
3159
3160         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3161                 enable_idle = 0;
3162         else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3163                  !cfqd->cfq_slice_idle ||
3164                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3165                 enable_idle = 0;
3166         else if (sample_valid(cic->ttime.ttime_samples)) {
3167                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3168                         enable_idle = 0;
3169                 else
3170                         enable_idle = 1;
3171         }
3172
3173         if (old_idle != enable_idle) {
3174                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3175                 if (enable_idle)
3176                         cfq_mark_cfqq_idle_window(cfqq);
3177                 else
3178                         cfq_clear_cfqq_idle_window(cfqq);
3179         }
3180 }
3181
3182 /*
3183  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3184  * no or if we aren't sure, a 1 will cause a preempt.
3185  */
3186 static bool
3187 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3188                    struct request *rq)
3189 {
3190         struct cfq_queue *cfqq;
3191
3192         cfqq = cfqd->active_queue;
3193         if (!cfqq)
3194                 return false;
3195
3196         if (cfq_class_idle(new_cfqq))
3197                 return false;
3198
3199         if (cfq_class_idle(cfqq))
3200                 return true;
3201
3202         /*
3203          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3204          */
3205         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3206                 return false;
3207
3208         /*
3209          * if the new request is sync, but the currently running queue is
3210          * not, let the sync request have priority.
3211          */
3212         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3213                 return true;
3214
3215         if (new_cfqq->cfqg != cfqq->cfqg)
3216                 return false;
3217
3218         if (cfq_slice_used(cfqq))
3219                 return true;
3220
3221         /* Allow preemption only if we are idling on sync-noidle tree */
3222         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3223             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3224             new_cfqq->service_tree->count == 2 &&
3225             RB_EMPTY_ROOT(&cfqq->sort_list))
3226                 return true;
3227
3228         /*
3229          * So both queues are sync. Let the new request get disk time if
3230          * it's a metadata request and the current queue is doing regular IO.
3231          */
3232         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3233                 return true;
3234
3235         /*
3236          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3237          */
3238         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3239                 return true;
3240
3241         /* An idle queue should not be idle now for some reason */
3242         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3243                 return true;
3244
3245         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3246                 return false;
3247
3248         /*
3249          * if this request is as-good as one we would expect from the
3250          * current cfqq, let it preempt
3251          */
3252         if (cfq_rq_close(cfqd, cfqq, rq))
3253                 return true;
3254
3255         return false;
3256 }
3257
3258 /*
3259  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3260  * let it have half of its nominal slice.
3261  */
3262 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3263 {
3264         struct cfq_queue *old_cfqq = cfqd->active_queue;
3265
3266         cfq_log_cfqq(cfqd, cfqq, "preempt");
3267         cfq_slice_expired(cfqd, 1);
3268
3269         /*
3270          * workload type is changed, don't save slice, otherwise preempt
3271          * doesn't happen
3272          */
3273         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3274                 cfqq->cfqg->saved_workload_slice = 0;
3275
3276         /*
3277          * Put the new queue at the front of the of the current list,
3278          * so we know that it will be selected next.
3279          */
3280         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3281
3282         cfq_service_tree_add(cfqd, cfqq, 1);
3283
3284         cfqq->slice_end = 0;
3285         cfq_mark_cfqq_slice_new(cfqq);
3286 }
3287
3288 /*
3289  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3290  * something we should do about it
3291  */
3292 static void
3293 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3294                 struct request *rq)
3295 {
3296         struct cfq_io_cq *cic = RQ_CIC(rq);
3297
3298         cfqd->rq_queued++;
3299         if (rq->cmd_flags & REQ_PRIO)
3300                 cfqq->prio_pending++;
3301
3302         cfq_update_io_thinktime(cfqd, cfqq, cic);
3303         cfq_update_io_seektime(cfqd, cfqq, rq);
3304         cfq_update_idle_window(cfqd, cfqq, cic);
3305
3306         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3307
3308         if (cfqq == cfqd->active_queue) {
3309                 /*
3310                  * Remember that we saw a request from this process, but
3311                  * don't start queuing just yet. Otherwise we risk seeing lots
3312                  * of tiny requests, because we disrupt the normal plugging
3313                  * and merging. If the request is already larger than a single
3314                  * page, let it rip immediately. For that case we assume that
3315                  * merging is already done. Ditto for a busy system that
3316                  * has other work pending, don't risk delaying until the
3317                  * idle timer unplug to continue working.
3318                  */
3319                 if (cfq_cfqq_wait_request(cfqq)) {
3320                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3321                             cfqd->busy_queues > 1) {
3322                                 cfq_del_timer(cfqd, cfqq);
3323                                 cfq_clear_cfqq_wait_request(cfqq);
3324                                 __blk_run_queue(cfqd->queue);
3325                         } else {
3326                                 cfq_blkiocg_update_idle_time_stats(
3327                                                 &cfqq->cfqg->blkg);
3328                                 cfq_mark_cfqq_must_dispatch(cfqq);
3329                         }
3330                 }
3331         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3332                 /*
3333                  * not the active queue - expire current slice if it is
3334                  * idle and has expired it's mean thinktime or this new queue
3335                  * has some old slice time left and is of higher priority or
3336                  * this new queue is RT and the current one is BE
3337                  */
3338                 cfq_preempt_queue(cfqd, cfqq);
3339                 __blk_run_queue(cfqd->queue);
3340         }
3341 }
3342
3343 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3344 {
3345         struct cfq_data *cfqd = q->elevator->elevator_data;
3346         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3347
3348         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3349         cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3350
3351         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3352         list_add_tail(&rq->queuelist, &cfqq->fifo);
3353         cfq_add_rq_rb(rq);
3354         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3355                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3356                         rq_is_sync(rq));
3357         cfq_rq_enqueued(cfqd, cfqq, rq);
3358 }
3359
3360 /*
3361  * Update hw_tag based on peak queue depth over 50 samples under
3362  * sufficient load.
3363  */
3364 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3365 {
3366         struct cfq_queue *cfqq = cfqd->active_queue;
3367
3368         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3369                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3370
3371         if (cfqd->hw_tag == 1)
3372                 return;
3373
3374         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3375             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3376                 return;
3377
3378         /*
3379          * If active queue hasn't enough requests and can idle, cfq might not
3380          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3381          * case
3382          */
3383         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3384             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3385             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3386                 return;
3387
3388         if (cfqd->hw_tag_samples++ < 50)
3389                 return;
3390
3391         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3392                 cfqd->hw_tag = 1;
3393         else
3394                 cfqd->hw_tag = 0;
3395 }
3396
3397 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3398 {
3399         struct cfq_io_cq *cic = cfqd->active_cic;
3400
3401         /* If the queue already has requests, don't wait */
3402         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3403                 return false;
3404
3405         /* If there are other queues in the group, don't wait */
3406         if (cfqq->cfqg->nr_cfqq > 1)
3407                 return false;
3408
3409         /* the only queue in the group, but think time is big */
3410         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3411                 return false;
3412
3413         if (cfq_slice_used(cfqq))
3414                 return true;
3415
3416         /* if slice left is less than think time, wait busy */
3417         if (cic && sample_valid(cic->ttime.ttime_samples)
3418             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3419                 return true;
3420
3421         /*
3422          * If think times is less than a jiffy than ttime_mean=0 and above
3423          * will not be true. It might happen that slice has not expired yet
3424          * but will expire soon (4-5 ns) during select_queue(). To cover the
3425          * case where think time is less than a jiffy, mark the queue wait
3426          * busy if only 1 jiffy is left in the slice.
3427          */
3428         if (cfqq->slice_end - jiffies == 1)
3429                 return true;
3430
3431         return false;
3432 }
3433
3434 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3435 {
3436         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3437         struct cfq_data *cfqd = cfqq->cfqd;
3438         const int sync = rq_is_sync(rq);
3439         unsigned long now;
3440
3441         now = jiffies;
3442         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3443                      !!(rq->cmd_flags & REQ_NOIDLE));
3444
3445         cfq_update_hw_tag(cfqd);
3446
3447         WARN_ON(!cfqd->rq_in_driver);
3448         WARN_ON(!cfqq->dispatched);
3449         cfqd->rq_in_driver--;
3450         cfqq->dispatched--;
3451         (RQ_CFQG(rq))->dispatched--;
3452         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3453                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3454                         rq_data_dir(rq), rq_is_sync(rq));
3455
3456         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3457
3458         if (sync) {
3459                 struct cfq_rb_root *service_tree;
3460
3461                 RQ_CIC(rq)->ttime.last_end_request = now;
3462
3463                 if (cfq_cfqq_on_rr(cfqq))
3464                         service_tree = cfqq->service_tree;
3465                 else
3466                         service_tree = service_tree_for(cfqq->cfqg,
3467                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3468                 service_tree->ttime.last_end_request = now;
3469                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3470                         cfqd->last_delayed_sync = now;
3471         }
3472
3473 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3474         cfqq->cfqg->ttime.last_end_request = now;
3475 #endif
3476
3477         /*
3478          * If this is the active queue, check if it needs to be expired,
3479          * or if we want to idle in case it has no pending requests.
3480          */
3481         if (cfqd->active_queue == cfqq) {
3482                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3483
3484                 if (cfq_cfqq_slice_new(cfqq)) {
3485                         cfq_set_prio_slice(cfqd, cfqq);
3486                         cfq_clear_cfqq_slice_new(cfqq);
3487                 }
3488
3489                 /*
3490                  * Should we wait for next request to come in before we expire
3491                  * the queue.
3492                  */
3493                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3494                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3495                         if (!cfqd->cfq_slice_idle)
3496                                 extend_sl = cfqd->cfq_group_idle;
3497                         cfqq->slice_end = jiffies + extend_sl;
3498                         cfq_mark_cfqq_wait_busy(cfqq);
3499                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3500                 }
3501
3502                 /*
3503                  * Idling is not enabled on:
3504                  * - expired queues
3505                  * - idle-priority queues
3506                  * - async queues
3507                  * - queues with still some requests queued
3508                  * - when there is a close cooperator
3509                  */
3510                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3511                         cfq_slice_expired(cfqd, 1);
3512                 else if (sync && cfqq_empty &&
3513                          !cfq_close_cooperator(cfqd, cfqq)) {
3514                         cfq_arm_slice_timer(cfqd);
3515                 }
3516         }
3517
3518         if (!cfqd->rq_in_driver)
3519                 cfq_schedule_dispatch(cfqd);
3520 }
3521
3522 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3523 {
3524         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3525                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3526                 return ELV_MQUEUE_MUST;
3527         }
3528
3529         return ELV_MQUEUE_MAY;
3530 }
3531
3532 static int cfq_may_queue(struct request_queue *q, int rw)
3533 {
3534         struct cfq_data *cfqd = q->elevator->elevator_data;
3535         struct task_struct *tsk = current;
3536         struct cfq_io_cq *cic;
3537         struct cfq_queue *cfqq;
3538
3539         /*
3540          * don't force setup of a queue from here, as a call to may_queue
3541          * does not necessarily imply that a request actually will be queued.
3542          * so just lookup a possibly existing queue, or return 'may queue'
3543          * if that fails
3544          */
3545         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3546         if (!cic)
3547                 return ELV_MQUEUE_MAY;
3548
3549         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3550         if (cfqq) {
3551                 cfq_init_prio_data(cfqq, cic->icq.ioc);
3552
3553                 return __cfq_may_queue(cfqq);
3554         }
3555
3556         return ELV_MQUEUE_MAY;
3557 }
3558
3559 /*
3560  * queue lock held here
3561  */
3562 static void cfq_put_request(struct request *rq)
3563 {
3564         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3565
3566         if (cfqq) {
3567                 const int rw = rq_data_dir(rq);
3568
3569                 BUG_ON(!cfqq->allocated[rw]);
3570                 cfqq->allocated[rw]--;
3571
3572                 put_io_context(RQ_CIC(rq)->icq.ioc, cfqq->cfqd->queue);
3573
3574                 /* Put down rq reference on cfqg */
3575                 cfq_put_cfqg(RQ_CFQG(rq));
3576                 rq->elv.priv[0] = NULL;
3577                 rq->elv.priv[1] = NULL;
3578
3579                 cfq_put_queue(cfqq);
3580         }
3581 }
3582
3583 static struct cfq_queue *
3584 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3585                 struct cfq_queue *cfqq)
3586 {
3587         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3588         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3589         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3590         cfq_put_queue(cfqq);
3591         return cic_to_cfqq(cic, 1);
3592 }
3593
3594 /*
3595  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3596  * was the last process referring to said cfqq.
3597  */
3598 static struct cfq_queue *
3599 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3600 {
3601         if (cfqq_process_refs(cfqq) == 1) {
3602                 cfqq->pid = current->pid;
3603                 cfq_clear_cfqq_coop(cfqq);
3604                 cfq_clear_cfqq_split_coop(cfqq);
3605                 return cfqq;
3606         }
3607
3608         cic_set_cfqq(cic, NULL, 1);
3609
3610         cfq_put_cooperator(cfqq);
3611
3612         cfq_put_queue(cfqq);
3613         return NULL;
3614 }
3615 /*
3616  * Allocate cfq data structures associated with this request.
3617  */
3618 static int
3619 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3620 {
3621         struct cfq_data *cfqd = q->elevator->elevator_data;
3622         struct cfq_io_cq *cic;
3623         const int rw = rq_data_dir(rq);
3624         const bool is_sync = rq_is_sync(rq);
3625         struct cfq_queue *cfqq;
3626
3627         might_sleep_if(gfp_mask & __GFP_WAIT);
3628
3629         spin_lock_irq(q->queue_lock);
3630         cic = cfq_get_cic(cfqd, gfp_mask);
3631         if (!cic)
3632                 goto queue_fail;
3633
3634 new_queue:
3635         cfqq = cic_to_cfqq(cic, is_sync);
3636         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3637                 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3638                 cic_set_cfqq(cic, cfqq, is_sync);
3639         } else {
3640                 /*
3641                  * If the queue was seeky for too long, break it apart.
3642                  */
3643                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3644                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3645                         cfqq = split_cfqq(cic, cfqq);
3646                         if (!cfqq)
3647                                 goto new_queue;
3648                 }
3649
3650                 /*
3651                  * Check to see if this queue is scheduled to merge with
3652                  * another, closely cooperating queue.  The merging of
3653                  * queues happens here as it must be done in process context.
3654                  * The reference on new_cfqq was taken in merge_cfqqs.
3655                  */
3656                 if (cfqq->new_cfqq)
3657                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3658         }
3659
3660         cfqq->allocated[rw]++;
3661
3662         cfqq->ref++;
3663         rq->elv.icq = &cic->icq;
3664         rq->elv.priv[0] = cfqq;
3665         rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
3666         spin_unlock_irq(q->queue_lock);
3667         return 0;
3668
3669 queue_fail:
3670         cfq_schedule_dispatch(cfqd);
3671         spin_unlock_irq(q->queue_lock);
3672         cfq_log(cfqd, "set_request fail");
3673         return 1;
3674 }
3675
3676 static void cfq_kick_queue(struct work_struct *work)
3677 {
3678         struct cfq_data *cfqd =
3679                 container_of(work, struct cfq_data, unplug_work);
3680         struct request_queue *q = cfqd->queue;
3681
3682         spin_lock_irq(q->queue_lock);
3683         __blk_run_queue(cfqd->queue);
3684         spin_unlock_irq(q->queue_lock);
3685 }
3686
3687 /*
3688  * Timer running if the active_queue is currently idling inside its time slice
3689  */
3690 static void cfq_idle_slice_timer(unsigned long data)
3691 {
3692         struct cfq_data *cfqd = (struct cfq_data *) data;
3693         struct cfq_queue *cfqq;
3694         unsigned long flags;
3695         int timed_out = 1;
3696
3697         cfq_log(cfqd, "idle timer fired");
3698
3699         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3700
3701         cfqq = cfqd->active_queue;
3702         if (cfqq) {
3703                 timed_out = 0;
3704
3705                 /*
3706                  * We saw a request before the queue expired, let it through
3707                  */
3708                 if (cfq_cfqq_must_dispatch(cfqq))
3709                         goto out_kick;
3710
3711                 /*
3712                  * expired
3713                  */
3714                 if (cfq_slice_used(cfqq))
3715                         goto expire;
3716
3717                 /*
3718                  * only expire and reinvoke request handler, if there are
3719                  * other queues with pending requests
3720                  */
3721                 if (!cfqd->busy_queues)
3722                         goto out_cont;
3723
3724                 /*
3725                  * not expired and it has a request pending, let it dispatch
3726                  */
3727                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3728                         goto out_kick;
3729
3730                 /*
3731                  * Queue depth flag is reset only when the idle didn't succeed
3732                  */
3733                 cfq_clear_cfqq_deep(cfqq);
3734         }
3735 expire:
3736         cfq_slice_expired(cfqd, timed_out);
3737 out_kick:
3738         cfq_schedule_dispatch(cfqd);
3739 out_cont:
3740         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3741 }
3742
3743 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3744 {
3745         del_timer_sync(&cfqd->idle_slice_timer);
3746         cancel_work_sync(&cfqd->unplug_work);
3747 }
3748
3749 static void cfq_put_async_queues(struct cfq_data *cfqd)
3750 {
3751         int i;
3752
3753         for (i = 0; i < IOPRIO_BE_NR; i++) {
3754                 if (cfqd->async_cfqq[0][i])
3755                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3756                 if (cfqd->async_cfqq[1][i])
3757                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3758         }
3759
3760         if (cfqd->async_idle_cfqq)
3761                 cfq_put_queue(cfqd->async_idle_cfqq);
3762 }
3763
3764 static void cfq_exit_queue(struct elevator_queue *e)
3765 {
3766         struct cfq_data *cfqd = e->elevator_data;
3767         struct request_queue *q = cfqd->queue;
3768         bool wait = false;
3769
3770         cfq_shutdown_timer_wq(cfqd);
3771
3772         spin_lock_irq(q->queue_lock);
3773
3774         if (cfqd->active_queue)
3775                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3776
3777         while (!list_empty(&q->icq_list)) {
3778                 struct io_cq *icq = list_entry(q->icq_list.next,
3779                                                struct io_cq, q_node);
3780                 struct io_context *ioc = icq->ioc;
3781
3782                 spin_lock(&ioc->lock);
3783                 cfq_exit_icq(icq);
3784                 cfq_release_icq(icq);
3785                 spin_unlock(&ioc->lock);
3786         }
3787
3788         cfq_put_async_queues(cfqd);
3789         cfq_release_cfq_groups(cfqd);
3790
3791         /*
3792          * If there are groups which we could not unlink from blkcg list,
3793          * wait for a rcu period for them to be freed.
3794          */
3795         if (cfqd->nr_blkcg_linked_grps)
3796                 wait = true;
3797
3798         spin_unlock_irq(q->queue_lock);
3799
3800         cfq_shutdown_timer_wq(cfqd);
3801
3802         /*
3803          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3804          * Do this wait only if there are other unlinked groups out
3805          * there. This can happen if cgroup deletion path claimed the
3806          * responsibility of cleaning up a group before queue cleanup code
3807          * get to the group.
3808          *
3809          * Do not call synchronize_rcu() unconditionally as there are drivers
3810          * which create/delete request queue hundreds of times during scan/boot
3811          * and synchronize_rcu() can take significant time and slow down boot.
3812          */
3813         if (wait)
3814                 synchronize_rcu();
3815
3816 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3817         /* Free up per cpu stats for root group */
3818         free_percpu(cfqd->root_group.blkg.stats_cpu);
3819 #endif
3820         kfree(cfqd);
3821 }
3822
3823 static void *cfq_init_queue(struct request_queue *q)
3824 {
3825         struct cfq_data *cfqd;
3826         int i, j;
3827         struct cfq_group *cfqg;
3828         struct cfq_rb_root *st;
3829
3830         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3831         if (!cfqd)
3832                 return NULL;
3833
3834         /* Init root service tree */
3835         cfqd->grp_service_tree = CFQ_RB_ROOT;
3836
3837         /* Init root group */
3838         cfqg = &cfqd->root_group;
3839         for_each_cfqg_st(cfqg, i, j, st)
3840                 *st = CFQ_RB_ROOT;
3841         RB_CLEAR_NODE(&cfqg->rb_node);
3842
3843         /* Give preference to root group over other groups */
3844         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3845
3846 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3847         /*
3848          * Set root group reference to 2. One reference will be dropped when
3849          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3850          * Other reference will remain there as we don't want to delete this
3851          * group as it is statically allocated and gets destroyed when
3852          * throtl_data goes away.
3853          */
3854         cfqg->ref = 2;
3855
3856         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3857                 kfree(cfqg);
3858                 kfree(cfqd);
3859                 return NULL;
3860         }
3861
3862         rcu_read_lock();
3863
3864         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3865                                         (void *)cfqd, 0);
3866         rcu_read_unlock();
3867         cfqd->nr_blkcg_linked_grps++;
3868
3869         /* Add group on cfqd->cfqg_list */
3870         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3871 #endif
3872         /*
3873          * Not strictly needed (since RB_ROOT just clears the node and we
3874          * zeroed cfqd on alloc), but better be safe in case someone decides
3875          * to add magic to the rb code
3876          */
3877         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3878                 cfqd->prio_trees[i] = RB_ROOT;
3879
3880         /*
3881          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3882          * Grab a permanent reference to it, so that the normal code flow
3883          * will not attempt to free it.
3884          */
3885         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3886         cfqd->oom_cfqq.ref++;
3887         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3888
3889         cfqd->queue = q;
3890
3891         init_timer(&cfqd->idle_slice_timer);
3892         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3893         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3894
3895         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3896
3897         cfqd->cfq_quantum = cfq_quantum;
3898         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3899         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3900         cfqd->cfq_back_max = cfq_back_max;
3901         cfqd->cfq_back_penalty = cfq_back_penalty;
3902         cfqd->cfq_slice[0] = cfq_slice_async;
3903         cfqd->cfq_slice[1] = cfq_slice_sync;
3904         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3905         cfqd->cfq_slice_idle = cfq_slice_idle;
3906         cfqd->cfq_group_idle = cfq_group_idle;
3907         cfqd->cfq_latency = 1;
3908         cfqd->hw_tag = -1;
3909         /*
3910          * we optimistically start assuming sync ops weren't delayed in last
3911          * second, in order to have larger depth for async operations.
3912          */
3913         cfqd->last_delayed_sync = jiffies - HZ;
3914         return cfqd;
3915 }
3916
3917 /*
3918  * sysfs parts below -->
3919  */
3920 static ssize_t
3921 cfq_var_show(unsigned int var, char *page)
3922 {
3923         return sprintf(page, "%d\n", var);
3924 }
3925
3926 static ssize_t
3927 cfq_var_store(unsigned int *var, const char *page, size_t count)
3928 {
3929         char *p = (char *) page;
3930
3931         *var = simple_strtoul(p, &p, 10);
3932         return count;
3933 }
3934
3935 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3936 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3937 {                                                                       \
3938         struct cfq_data *cfqd = e->elevator_data;                       \
3939         unsigned int __data = __VAR;                                    \
3940         if (__CONV)                                                     \
3941                 __data = jiffies_to_msecs(__data);                      \
3942         return cfq_var_show(__data, (page));                            \
3943 }
3944 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3945 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3946 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3947 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3948 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3949 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3950 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3951 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3952 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3953 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3954 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3955 #undef SHOW_FUNCTION
3956
3957 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3958 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3959 {                                                                       \
3960         struct cfq_data *cfqd = e->elevator_data;                       \
3961         unsigned int __data;                                            \
3962         int ret = cfq_var_store(&__data, (page), count);                \
3963         if (__data < (MIN))                                             \
3964                 __data = (MIN);                                         \
3965         else if (__data > (MAX))                                        \
3966                 __data = (MAX);                                         \
3967         if (__CONV)                                                     \
3968                 *(__PTR) = msecs_to_jiffies(__data);                    \
3969         else                                                            \
3970                 *(__PTR) = __data;                                      \
3971         return ret;                                                     \
3972 }
3973 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3974 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3975                 UINT_MAX, 1);
3976 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3977                 UINT_MAX, 1);
3978 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3979 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3980                 UINT_MAX, 0);
3981 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3982 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
3983 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3984 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3985 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3986                 UINT_MAX, 0);
3987 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3988 #undef STORE_FUNCTION
3989
3990 #define CFQ_ATTR(name) \
3991         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3992
3993 static struct elv_fs_entry cfq_attrs[] = {
3994         CFQ_ATTR(quantum),
3995         CFQ_ATTR(fifo_expire_sync),
3996         CFQ_ATTR(fifo_expire_async),
3997         CFQ_ATTR(back_seek_max),
3998         CFQ_ATTR(back_seek_penalty),
3999         CFQ_ATTR(slice_sync),
4000         CFQ_ATTR(slice_async),
4001         CFQ_ATTR(slice_async_rq),
4002         CFQ_ATTR(slice_idle),
4003         CFQ_ATTR(group_idle),
4004         CFQ_ATTR(low_latency),
4005         __ATTR_NULL
4006 };
4007
4008 static struct elevator_type iosched_cfq = {
4009         .ops = {
4010                 .elevator_merge_fn =            cfq_merge,
4011                 .elevator_merged_fn =           cfq_merged_request,
4012                 .elevator_merge_req_fn =        cfq_merged_requests,
4013                 .elevator_allow_merge_fn =      cfq_allow_merge,
4014                 .elevator_bio_merged_fn =       cfq_bio_merged,
4015                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4016                 .elevator_add_req_fn =          cfq_insert_request,
4017                 .elevator_activate_req_fn =     cfq_activate_request,
4018                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4019                 .elevator_completed_req_fn =    cfq_completed_request,
4020                 .elevator_former_req_fn =       elv_rb_former_request,
4021                 .elevator_latter_req_fn =       elv_rb_latter_request,
4022                 .elevator_set_req_fn =          cfq_set_request,
4023                 .elevator_put_req_fn =          cfq_put_request,
4024                 .elevator_may_queue_fn =        cfq_may_queue,
4025                 .elevator_init_fn =             cfq_init_queue,
4026                 .elevator_exit_fn =             cfq_exit_queue,
4027         },
4028         .icq_size       =       sizeof(struct cfq_io_cq),
4029         .icq_align      =       __alignof__(struct cfq_io_cq),
4030         .elevator_attrs =       cfq_attrs,
4031         .elevator_name  =       "cfq",
4032         .elevator_owner =       THIS_MODULE,
4033 };
4034
4035 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4036 static struct blkio_policy_type blkio_policy_cfq = {
4037         .ops = {
4038                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4039                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4040         },
4041         .plid = BLKIO_POLICY_PROP,
4042 };
4043 #else
4044 static struct blkio_policy_type blkio_policy_cfq;
4045 #endif
4046
4047 static int __init cfq_init(void)
4048 {
4049         int ret;
4050
4051         /*
4052          * could be 0 on HZ < 1000 setups
4053          */
4054         if (!cfq_slice_async)
4055                 cfq_slice_async = 1;
4056         if (!cfq_slice_idle)
4057                 cfq_slice_idle = 1;
4058
4059 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4060         if (!cfq_group_idle)
4061                 cfq_group_idle = 1;
4062 #else
4063                 cfq_group_idle = 0;
4064 #endif
4065         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4066         if (!cfq_pool)
4067                 return -ENOMEM;
4068
4069         ret = elv_register(&iosched_cfq);
4070         if (ret) {
4071                 kmem_cache_destroy(cfq_pool);
4072                 return ret;
4073         }
4074         cfq_icq_pool = iosched_cfq.icq_cache;
4075
4076         blkio_policy_register(&blkio_policy_cfq);
4077
4078         return 0;
4079 }
4080
4081 static void __exit cfq_exit(void)
4082 {
4083         blkio_policy_unregister(&blkio_policy_cfq);
4084         elv_unregister(&iosched_cfq);
4085         kmem_cache_destroy(cfq_pool);
4086 }
4087
4088 module_init(cfq_init);
4089 module_exit(cfq_exit);
4090
4091 MODULE_AUTHOR("Jens Axboe");
4092 MODULE_LICENSE("GPL");
4093 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");