CFQ: add think time check for service tree
[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 "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90         struct cfq_ttime ttime;
91 };
92 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
93                         .ttime = {.last_end_request = jiffies,},}
94
95 /*
96  * Per process-grouping structure
97  */
98 struct cfq_queue {
99         /* reference count */
100         int ref;
101         /* various state flags, see below */
102         unsigned int flags;
103         /* parent cfq_data */
104         struct cfq_data *cfqd;
105         /* service_tree member */
106         struct rb_node rb_node;
107         /* service_tree key */
108         unsigned long rb_key;
109         /* prio tree member */
110         struct rb_node p_node;
111         /* prio tree root we belong to, if any */
112         struct rb_root *p_root;
113         /* sorted list of pending requests */
114         struct rb_root sort_list;
115         /* if fifo isn't expired, next request to serve */
116         struct request *next_rq;
117         /* requests queued in sort_list */
118         int queued[2];
119         /* currently allocated requests */
120         int allocated[2];
121         /* fifo list of requests in sort_list */
122         struct list_head fifo;
123
124         /* time when queue got scheduled in to dispatch first request. */
125         unsigned long dispatch_start;
126         unsigned int allocated_slice;
127         unsigned int slice_dispatch;
128         /* time when first request from queue completed and slice started. */
129         unsigned long slice_start;
130         unsigned long slice_end;
131         long slice_resid;
132
133         /* number of requests that are on the dispatch list or inside driver */
134         int dispatched;
135
136         /* io prio of this group */
137         unsigned short ioprio, org_ioprio;
138         unsigned short ioprio_class;
139
140         pid_t pid;
141
142         u32 seek_history;
143         sector_t last_request_pos;
144
145         struct cfq_rb_root *service_tree;
146         struct cfq_queue *new_cfqq;
147         struct cfq_group *cfqg;
148         /* Number of sectors dispatched from queue in single dispatch round */
149         unsigned long nr_sectors;
150 };
151
152 /*
153  * First index in the service_trees.
154  * IDLE is handled separately, so it has negative index
155  */
156 enum wl_prio_t {
157         BE_WORKLOAD = 0,
158         RT_WORKLOAD = 1,
159         IDLE_WORKLOAD = 2,
160         CFQ_PRIO_NR,
161 };
162
163 /*
164  * Second index in the service_trees.
165  */
166 enum wl_type_t {
167         ASYNC_WORKLOAD = 0,
168         SYNC_NOIDLE_WORKLOAD = 1,
169         SYNC_WORKLOAD = 2
170 };
171
172 /* This is per cgroup per device grouping structure */
173 struct cfq_group {
174         /* group service_tree member */
175         struct rb_node rb_node;
176
177         /* group service_tree key */
178         u64 vdisktime;
179         unsigned int weight;
180         unsigned int new_weight;
181         bool needs_update;
182
183         /* number of cfqq currently on this group */
184         int nr_cfqq;
185
186         /*
187          * Per group busy queues average. Useful for workload slice calc. We
188          * create the array for each prio class but at run time it is used
189          * only for RT and BE class and slot for IDLE class remains unused.
190          * This is primarily done to avoid confusion and a gcc warning.
191          */
192         unsigned int busy_queues_avg[CFQ_PRIO_NR];
193         /*
194          * rr lists of queues with requests. We maintain service trees for
195          * RT and BE classes. These trees are subdivided in subclasses
196          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
197          * class there is no subclassification and all the cfq queues go on
198          * a single tree service_tree_idle.
199          * Counts are embedded in the cfq_rb_root
200          */
201         struct cfq_rb_root service_trees[2][3];
202         struct cfq_rb_root service_tree_idle;
203
204         unsigned long saved_workload_slice;
205         enum wl_type_t saved_workload;
206         enum wl_prio_t saved_serving_prio;
207         struct blkio_group blkg;
208 #ifdef CONFIG_CFQ_GROUP_IOSCHED
209         struct hlist_node cfqd_node;
210         int ref;
211 #endif
212         /* number of requests that are on the dispatch list or inside driver */
213         int dispatched;
214 };
215
216 /*
217  * Per block device queue structure
218  */
219 struct cfq_data {
220         struct request_queue *queue;
221         /* Root service tree for cfq_groups */
222         struct cfq_rb_root grp_service_tree;
223         struct cfq_group root_group;
224
225         /*
226          * The priority currently being served
227          */
228         enum wl_prio_t serving_prio;
229         enum wl_type_t serving_type;
230         unsigned long workload_expires;
231         struct cfq_group *serving_group;
232
233         /*
234          * Each priority tree is sorted by next_request position.  These
235          * trees are used when determining if two or more queues are
236          * interleaving requests (see cfq_close_cooperator).
237          */
238         struct rb_root prio_trees[CFQ_PRIO_LISTS];
239
240         unsigned int busy_queues;
241         unsigned int busy_sync_queues;
242
243         int rq_in_driver;
244         int rq_in_flight[2];
245
246         /*
247          * queue-depth detection
248          */
249         int rq_queued;
250         int hw_tag;
251         /*
252          * hw_tag can be
253          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
254          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
255          *  0 => no NCQ
256          */
257         int hw_tag_est_depth;
258         unsigned int hw_tag_samples;
259
260         /*
261          * idle window management
262          */
263         struct timer_list idle_slice_timer;
264         struct work_struct unplug_work;
265
266         struct cfq_queue *active_queue;
267         struct cfq_io_context *active_cic;
268
269         /*
270          * async queue for each priority case
271          */
272         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
273         struct cfq_queue *async_idle_cfqq;
274
275         sector_t last_position;
276
277         /*
278          * tunables, see top of file
279          */
280         unsigned int cfq_quantum;
281         unsigned int cfq_fifo_expire[2];
282         unsigned int cfq_back_penalty;
283         unsigned int cfq_back_max;
284         unsigned int cfq_slice[2];
285         unsigned int cfq_slice_async_rq;
286         unsigned int cfq_slice_idle;
287         unsigned int cfq_group_idle;
288         unsigned int cfq_latency;
289
290         unsigned int cic_index;
291         struct list_head cic_list;
292
293         /*
294          * Fallback dummy cfqq for extreme OOM conditions
295          */
296         struct cfq_queue oom_cfqq;
297
298         unsigned long last_delayed_sync;
299
300         /* List of cfq groups being managed on this device*/
301         struct hlist_head cfqg_list;
302
303         /* Number of groups which are on blkcg->blkg_list */
304         unsigned int nr_blkcg_linked_grps;
305 };
306
307 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
308
309 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
310                                             enum wl_prio_t prio,
311                                             enum wl_type_t type)
312 {
313         if (!cfqg)
314                 return NULL;
315
316         if (prio == IDLE_WORKLOAD)
317                 return &cfqg->service_tree_idle;
318
319         return &cfqg->service_trees[prio][type];
320 }
321
322 enum cfqq_state_flags {
323         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
324         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
325         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
326         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
327         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
328         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
329         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
330         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
331         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
332         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
333         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
334         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
335         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
336 };
337
338 #define CFQ_CFQQ_FNS(name)                                              \
339 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
340 {                                                                       \
341         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
342 }                                                                       \
343 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
344 {                                                                       \
345         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
346 }                                                                       \
347 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
348 {                                                                       \
349         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
350 }
351
352 CFQ_CFQQ_FNS(on_rr);
353 CFQ_CFQQ_FNS(wait_request);
354 CFQ_CFQQ_FNS(must_dispatch);
355 CFQ_CFQQ_FNS(must_alloc_slice);
356 CFQ_CFQQ_FNS(fifo_expire);
357 CFQ_CFQQ_FNS(idle_window);
358 CFQ_CFQQ_FNS(prio_changed);
359 CFQ_CFQQ_FNS(slice_new);
360 CFQ_CFQQ_FNS(sync);
361 CFQ_CFQQ_FNS(coop);
362 CFQ_CFQQ_FNS(split_coop);
363 CFQ_CFQQ_FNS(deep);
364 CFQ_CFQQ_FNS(wait_busy);
365 #undef CFQ_CFQQ_FNS
366
367 #ifdef CONFIG_CFQ_GROUP_IOSCHED
368 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
369         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
370                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
371                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
372
373 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
374         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
375                                 blkg_path(&(cfqg)->blkg), ##args)       \
376
377 #else
378 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
379         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
380 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
381 #endif
382 #define cfq_log(cfqd, fmt, args...)     \
383         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
384
385 /* Traverses through cfq group service trees */
386 #define for_each_cfqg_st(cfqg, i, j, st) \
387         for (i = 0; i <= IDLE_WORKLOAD; i++) \
388                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
389                         : &cfqg->service_tree_idle; \
390                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
391                         (i == IDLE_WORKLOAD && j == 0); \
392                         j++, st = i < IDLE_WORKLOAD ? \
393                         &cfqg->service_trees[i][j]: NULL) \
394
395 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
396         struct cfq_ttime *ttime, bool group_idle)
397 {
398         unsigned long slice;
399         if (!sample_valid(ttime->ttime_samples))
400                 return false;
401         if (group_idle)
402                 slice = cfqd->cfq_group_idle;
403         else
404                 slice = cfqd->cfq_slice_idle;
405         return ttime->ttime_mean > slice;
406 }
407
408 static inline bool iops_mode(struct cfq_data *cfqd)
409 {
410         /*
411          * If we are not idling on queues and it is a NCQ drive, parallel
412          * execution of requests is on and measuring time is not possible
413          * in most of the cases until and unless we drive shallower queue
414          * depths and that becomes a performance bottleneck. In such cases
415          * switch to start providing fairness in terms of number of IOs.
416          */
417         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
418                 return true;
419         else
420                 return false;
421 }
422
423 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
424 {
425         if (cfq_class_idle(cfqq))
426                 return IDLE_WORKLOAD;
427         if (cfq_class_rt(cfqq))
428                 return RT_WORKLOAD;
429         return BE_WORKLOAD;
430 }
431
432
433 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
434 {
435         if (!cfq_cfqq_sync(cfqq))
436                 return ASYNC_WORKLOAD;
437         if (!cfq_cfqq_idle_window(cfqq))
438                 return SYNC_NOIDLE_WORKLOAD;
439         return SYNC_WORKLOAD;
440 }
441
442 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
443                                         struct cfq_data *cfqd,
444                                         struct cfq_group *cfqg)
445 {
446         if (wl == IDLE_WORKLOAD)
447                 return cfqg->service_tree_idle.count;
448
449         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
450                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
451                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
452 }
453
454 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
455                                         struct cfq_group *cfqg)
456 {
457         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
458                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
459 }
460
461 static void cfq_dispatch_insert(struct request_queue *, struct request *);
462 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
463                                        struct io_context *, gfp_t);
464 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
465                                                 struct io_context *);
466
467 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
468                                             bool is_sync)
469 {
470         return cic->cfqq[is_sync];
471 }
472
473 static inline void cic_set_cfqq(struct cfq_io_context *cic,
474                                 struct cfq_queue *cfqq, bool is_sync)
475 {
476         cic->cfqq[is_sync] = cfqq;
477 }
478
479 #define CIC_DEAD_KEY    1ul
480 #define CIC_DEAD_INDEX_SHIFT    1
481
482 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
483 {
484         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
485 }
486
487 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
488 {
489         struct cfq_data *cfqd = cic->key;
490
491         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
492                 return NULL;
493
494         return cfqd;
495 }
496
497 /*
498  * We regard a request as SYNC, if it's either a read or has the SYNC bit
499  * set (in which case it could also be direct WRITE).
500  */
501 static inline bool cfq_bio_sync(struct bio *bio)
502 {
503         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
504 }
505
506 /*
507  * scheduler run of queue, if there are requests pending and no one in the
508  * driver that will restart queueing
509  */
510 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
511 {
512         if (cfqd->busy_queues) {
513                 cfq_log(cfqd, "schedule dispatch");
514                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
515         }
516 }
517
518 /*
519  * Scale schedule slice based on io priority. Use the sync time slice only
520  * if a queue is marked sync and has sync io queued. A sync queue with async
521  * io only, should not get full sync slice length.
522  */
523 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
524                                  unsigned short prio)
525 {
526         const int base_slice = cfqd->cfq_slice[sync];
527
528         WARN_ON(prio >= IOPRIO_BE_NR);
529
530         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
531 }
532
533 static inline int
534 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
535 {
536         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
537 }
538
539 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
540 {
541         u64 d = delta << CFQ_SERVICE_SHIFT;
542
543         d = d * BLKIO_WEIGHT_DEFAULT;
544         do_div(d, cfqg->weight);
545         return d;
546 }
547
548 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
549 {
550         s64 delta = (s64)(vdisktime - min_vdisktime);
551         if (delta > 0)
552                 min_vdisktime = vdisktime;
553
554         return min_vdisktime;
555 }
556
557 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
558 {
559         s64 delta = (s64)(vdisktime - min_vdisktime);
560         if (delta < 0)
561                 min_vdisktime = vdisktime;
562
563         return min_vdisktime;
564 }
565
566 static void update_min_vdisktime(struct cfq_rb_root *st)
567 {
568         struct cfq_group *cfqg;
569
570         if (st->left) {
571                 cfqg = rb_entry_cfqg(st->left);
572                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
573                                                   cfqg->vdisktime);
574         }
575 }
576
577 /*
578  * get averaged number of queues of RT/BE priority.
579  * average is updated, with a formula that gives more weight to higher numbers,
580  * to quickly follows sudden increases and decrease slowly
581  */
582
583 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
584                                         struct cfq_group *cfqg, bool rt)
585 {
586         unsigned min_q, max_q;
587         unsigned mult  = cfq_hist_divisor - 1;
588         unsigned round = cfq_hist_divisor / 2;
589         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
590
591         min_q = min(cfqg->busy_queues_avg[rt], busy);
592         max_q = max(cfqg->busy_queues_avg[rt], busy);
593         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
594                 cfq_hist_divisor;
595         return cfqg->busy_queues_avg[rt];
596 }
597
598 static inline unsigned
599 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
600 {
601         struct cfq_rb_root *st = &cfqd->grp_service_tree;
602
603         return cfq_target_latency * cfqg->weight / st->total_weight;
604 }
605
606 static inline unsigned
607 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
608 {
609         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
610         if (cfqd->cfq_latency) {
611                 /*
612                  * interested queues (we consider only the ones with the same
613                  * priority class in the cfq group)
614                  */
615                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
616                                                 cfq_class_rt(cfqq));
617                 unsigned sync_slice = cfqd->cfq_slice[1];
618                 unsigned expect_latency = sync_slice * iq;
619                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
620
621                 if (expect_latency > group_slice) {
622                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
623                         /* scale low_slice according to IO priority
624                          * and sync vs async */
625                         unsigned low_slice =
626                                 min(slice, base_low_slice * slice / sync_slice);
627                         /* the adapted slice value is scaled to fit all iqs
628                          * into the target latency */
629                         slice = max(slice * group_slice / expect_latency,
630                                     low_slice);
631                 }
632         }
633         return slice;
634 }
635
636 static inline void
637 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
638 {
639         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
640
641         cfqq->slice_start = jiffies;
642         cfqq->slice_end = jiffies + slice;
643         cfqq->allocated_slice = slice;
644         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
645 }
646
647 /*
648  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
649  * isn't valid until the first request from the dispatch is activated
650  * and the slice time set.
651  */
652 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
653 {
654         if (cfq_cfqq_slice_new(cfqq))
655                 return false;
656         if (time_before(jiffies, cfqq->slice_end))
657                 return false;
658
659         return true;
660 }
661
662 /*
663  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
664  * We choose the request that is closest to the head right now. Distance
665  * behind the head is penalized and only allowed to a certain extent.
666  */
667 static struct request *
668 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
669 {
670         sector_t s1, s2, d1 = 0, d2 = 0;
671         unsigned long back_max;
672 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
673 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
674         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
675
676         if (rq1 == NULL || rq1 == rq2)
677                 return rq2;
678         if (rq2 == NULL)
679                 return rq1;
680
681         if (rq_is_sync(rq1) != rq_is_sync(rq2))
682                 return rq_is_sync(rq1) ? rq1 : rq2;
683
684         s1 = blk_rq_pos(rq1);
685         s2 = blk_rq_pos(rq2);
686
687         /*
688          * by definition, 1KiB is 2 sectors
689          */
690         back_max = cfqd->cfq_back_max * 2;
691
692         /*
693          * Strict one way elevator _except_ in the case where we allow
694          * short backward seeks which are biased as twice the cost of a
695          * similar forward seek.
696          */
697         if (s1 >= last)
698                 d1 = s1 - last;
699         else if (s1 + back_max >= last)
700                 d1 = (last - s1) * cfqd->cfq_back_penalty;
701         else
702                 wrap |= CFQ_RQ1_WRAP;
703
704         if (s2 >= last)
705                 d2 = s2 - last;
706         else if (s2 + back_max >= last)
707                 d2 = (last - s2) * cfqd->cfq_back_penalty;
708         else
709                 wrap |= CFQ_RQ2_WRAP;
710
711         /* Found required data */
712
713         /*
714          * By doing switch() on the bit mask "wrap" we avoid having to
715          * check two variables for all permutations: --> faster!
716          */
717         switch (wrap) {
718         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
719                 if (d1 < d2)
720                         return rq1;
721                 else if (d2 < d1)
722                         return rq2;
723                 else {
724                         if (s1 >= s2)
725                                 return rq1;
726                         else
727                                 return rq2;
728                 }
729
730         case CFQ_RQ2_WRAP:
731                 return rq1;
732         case CFQ_RQ1_WRAP:
733                 return rq2;
734         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
735         default:
736                 /*
737                  * Since both rqs are wrapped,
738                  * start with the one that's further behind head
739                  * (--> only *one* back seek required),
740                  * since back seek takes more time than forward.
741                  */
742                 if (s1 <= s2)
743                         return rq1;
744                 else
745                         return rq2;
746         }
747 }
748
749 /*
750  * The below is leftmost cache rbtree addon
751  */
752 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
753 {
754         /* Service tree is empty */
755         if (!root->count)
756                 return NULL;
757
758         if (!root->left)
759                 root->left = rb_first(&root->rb);
760
761         if (root->left)
762                 return rb_entry(root->left, struct cfq_queue, rb_node);
763
764         return NULL;
765 }
766
767 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
768 {
769         if (!root->left)
770                 root->left = rb_first(&root->rb);
771
772         if (root->left)
773                 return rb_entry_cfqg(root->left);
774
775         return NULL;
776 }
777
778 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
779 {
780         rb_erase(n, root);
781         RB_CLEAR_NODE(n);
782 }
783
784 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
785 {
786         if (root->left == n)
787                 root->left = NULL;
788         rb_erase_init(n, &root->rb);
789         --root->count;
790 }
791
792 /*
793  * would be nice to take fifo expire time into account as well
794  */
795 static struct request *
796 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
797                   struct request *last)
798 {
799         struct rb_node *rbnext = rb_next(&last->rb_node);
800         struct rb_node *rbprev = rb_prev(&last->rb_node);
801         struct request *next = NULL, *prev = NULL;
802
803         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
804
805         if (rbprev)
806                 prev = rb_entry_rq(rbprev);
807
808         if (rbnext)
809                 next = rb_entry_rq(rbnext);
810         else {
811                 rbnext = rb_first(&cfqq->sort_list);
812                 if (rbnext && rbnext != &last->rb_node)
813                         next = rb_entry_rq(rbnext);
814         }
815
816         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
817 }
818
819 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
820                                       struct cfq_queue *cfqq)
821 {
822         /*
823          * just an approximation, should be ok.
824          */
825         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
826                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
827 }
828
829 static inline s64
830 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
831 {
832         return cfqg->vdisktime - st->min_vdisktime;
833 }
834
835 static void
836 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
837 {
838         struct rb_node **node = &st->rb.rb_node;
839         struct rb_node *parent = NULL;
840         struct cfq_group *__cfqg;
841         s64 key = cfqg_key(st, cfqg);
842         int left = 1;
843
844         while (*node != NULL) {
845                 parent = *node;
846                 __cfqg = rb_entry_cfqg(parent);
847
848                 if (key < cfqg_key(st, __cfqg))
849                         node = &parent->rb_left;
850                 else {
851                         node = &parent->rb_right;
852                         left = 0;
853                 }
854         }
855
856         if (left)
857                 st->left = &cfqg->rb_node;
858
859         rb_link_node(&cfqg->rb_node, parent, node);
860         rb_insert_color(&cfqg->rb_node, &st->rb);
861 }
862
863 static void
864 cfq_update_group_weight(struct cfq_group *cfqg)
865 {
866         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
867         if (cfqg->needs_update) {
868                 cfqg->weight = cfqg->new_weight;
869                 cfqg->needs_update = false;
870         }
871 }
872
873 static void
874 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
875 {
876         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
877
878         cfq_update_group_weight(cfqg);
879         __cfq_group_service_tree_add(st, cfqg);
880         st->total_weight += cfqg->weight;
881 }
882
883 static void
884 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
885 {
886         struct cfq_rb_root *st = &cfqd->grp_service_tree;
887         struct cfq_group *__cfqg;
888         struct rb_node *n;
889
890         cfqg->nr_cfqq++;
891         if (!RB_EMPTY_NODE(&cfqg->rb_node))
892                 return;
893
894         /*
895          * Currently put the group at the end. Later implement something
896          * so that groups get lesser vtime based on their weights, so that
897          * if group does not loose all if it was not continuously backlogged.
898          */
899         n = rb_last(&st->rb);
900         if (n) {
901                 __cfqg = rb_entry_cfqg(n);
902                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
903         } else
904                 cfqg->vdisktime = st->min_vdisktime;
905         cfq_group_service_tree_add(st, cfqg);
906 }
907
908 static void
909 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
910 {
911         st->total_weight -= cfqg->weight;
912         if (!RB_EMPTY_NODE(&cfqg->rb_node))
913                 cfq_rb_erase(&cfqg->rb_node, st);
914 }
915
916 static void
917 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
918 {
919         struct cfq_rb_root *st = &cfqd->grp_service_tree;
920
921         BUG_ON(cfqg->nr_cfqq < 1);
922         cfqg->nr_cfqq--;
923
924         /* If there are other cfq queues under this group, don't delete it */
925         if (cfqg->nr_cfqq)
926                 return;
927
928         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
929         cfq_group_service_tree_del(st, cfqg);
930         cfqg->saved_workload_slice = 0;
931         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
932 }
933
934 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
935                                                 unsigned int *unaccounted_time)
936 {
937         unsigned int slice_used;
938
939         /*
940          * Queue got expired before even a single request completed or
941          * got expired immediately after first request completion.
942          */
943         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
944                 /*
945                  * Also charge the seek time incurred to the group, otherwise
946                  * if there are mutiple queues in the group, each can dispatch
947                  * a single request on seeky media and cause lots of seek time
948                  * and group will never know it.
949                  */
950                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
951                                         1);
952         } else {
953                 slice_used = jiffies - cfqq->slice_start;
954                 if (slice_used > cfqq->allocated_slice) {
955                         *unaccounted_time = slice_used - cfqq->allocated_slice;
956                         slice_used = cfqq->allocated_slice;
957                 }
958                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
959                         *unaccounted_time += cfqq->slice_start -
960                                         cfqq->dispatch_start;
961         }
962
963         return slice_used;
964 }
965
966 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
967                                 struct cfq_queue *cfqq)
968 {
969         struct cfq_rb_root *st = &cfqd->grp_service_tree;
970         unsigned int used_sl, charge, unaccounted_sl = 0;
971         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
972                         - cfqg->service_tree_idle.count;
973
974         BUG_ON(nr_sync < 0);
975         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
976
977         if (iops_mode(cfqd))
978                 charge = cfqq->slice_dispatch;
979         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
980                 charge = cfqq->allocated_slice;
981
982         /* Can't update vdisktime while group is on service tree */
983         cfq_group_service_tree_del(st, cfqg);
984         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
985         /* If a new weight was requested, update now, off tree */
986         cfq_group_service_tree_add(st, cfqg);
987
988         /* This group is being expired. Save the context */
989         if (time_after(cfqd->workload_expires, jiffies)) {
990                 cfqg->saved_workload_slice = cfqd->workload_expires
991                                                 - jiffies;
992                 cfqg->saved_workload = cfqd->serving_type;
993                 cfqg->saved_serving_prio = cfqd->serving_prio;
994         } else
995                 cfqg->saved_workload_slice = 0;
996
997         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
998                                         st->min_vdisktime);
999         cfq_log_cfqq(cfqq->cfqd, cfqq,
1000                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1001                      used_sl, cfqq->slice_dispatch, charge,
1002                      iops_mode(cfqd), cfqq->nr_sectors);
1003         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1004                                           unaccounted_sl);
1005         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1006 }
1007
1008 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1009 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1010 {
1011         if (blkg)
1012                 return container_of(blkg, struct cfq_group, blkg);
1013         return NULL;
1014 }
1015
1016 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1017                                           unsigned int weight)
1018 {
1019         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1020         cfqg->new_weight = weight;
1021         cfqg->needs_update = true;
1022 }
1023
1024 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1025                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1026 {
1027         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1028         unsigned int major, minor;
1029
1030         /*
1031          * Add group onto cgroup list. It might happen that bdi->dev is
1032          * not initialized yet. Initialize this new group without major
1033          * and minor info and this info will be filled in once a new thread
1034          * comes for IO.
1035          */
1036         if (bdi->dev) {
1037                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1038                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1039                                         (void *)cfqd, MKDEV(major, minor));
1040         } else
1041                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1042                                         (void *)cfqd, 0);
1043
1044         cfqd->nr_blkcg_linked_grps++;
1045         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1046
1047         /* Add group on cfqd list */
1048         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1049 }
1050
1051 /*
1052  * Should be called from sleepable context. No request queue lock as per
1053  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1054  * from sleepable context.
1055  */
1056 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1057 {
1058         struct cfq_group *cfqg = NULL;
1059         int i, j, ret;
1060         struct cfq_rb_root *st;
1061
1062         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1063         if (!cfqg)
1064                 return NULL;
1065
1066         for_each_cfqg_st(cfqg, i, j, st)
1067                 *st = CFQ_RB_ROOT;
1068         RB_CLEAR_NODE(&cfqg->rb_node);
1069
1070         /*
1071          * Take the initial reference that will be released on destroy
1072          * This can be thought of a joint reference by cgroup and
1073          * elevator which will be dropped by either elevator exit
1074          * or cgroup deletion path depending on who is exiting first.
1075          */
1076         cfqg->ref = 1;
1077
1078         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1079         if (ret) {
1080                 kfree(cfqg);
1081                 return NULL;
1082         }
1083
1084         return cfqg;
1085 }
1086
1087 static struct cfq_group *
1088 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1089 {
1090         struct cfq_group *cfqg = NULL;
1091         void *key = cfqd;
1092         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1093         unsigned int major, minor;
1094
1095         /*
1096          * This is the common case when there are no blkio cgroups.
1097          * Avoid lookup in this case
1098          */
1099         if (blkcg == &blkio_root_cgroup)
1100                 cfqg = &cfqd->root_group;
1101         else
1102                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1103
1104         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1105                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1106                 cfqg->blkg.dev = MKDEV(major, minor);
1107         }
1108
1109         return cfqg;
1110 }
1111
1112 /*
1113  * Search for the cfq group current task belongs to. request_queue lock must
1114  * be held.
1115  */
1116 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1117 {
1118         struct blkio_cgroup *blkcg;
1119         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1120         struct request_queue *q = cfqd->queue;
1121
1122         rcu_read_lock();
1123         blkcg = task_blkio_cgroup(current);
1124         cfqg = cfq_find_cfqg(cfqd, blkcg);
1125         if (cfqg) {
1126                 rcu_read_unlock();
1127                 return cfqg;
1128         }
1129
1130         /*
1131          * Need to allocate a group. Allocation of group also needs allocation
1132          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1133          * we need to drop rcu lock and queue_lock before we call alloc.
1134          *
1135          * Not taking any queue reference here and assuming that queue is
1136          * around by the time we return. CFQ queue allocation code does
1137          * the same. It might be racy though.
1138          */
1139
1140         rcu_read_unlock();
1141         spin_unlock_irq(q->queue_lock);
1142
1143         cfqg = cfq_alloc_cfqg(cfqd);
1144
1145         spin_lock_irq(q->queue_lock);
1146
1147         rcu_read_lock();
1148         blkcg = task_blkio_cgroup(current);
1149
1150         /*
1151          * If some other thread already allocated the group while we were
1152          * not holding queue lock, free up the group
1153          */
1154         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1155
1156         if (__cfqg) {
1157                 kfree(cfqg);
1158                 rcu_read_unlock();
1159                 return __cfqg;
1160         }
1161
1162         if (!cfqg)
1163                 cfqg = &cfqd->root_group;
1164
1165         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1166         rcu_read_unlock();
1167         return cfqg;
1168 }
1169
1170 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1171 {
1172         cfqg->ref++;
1173         return cfqg;
1174 }
1175
1176 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1177 {
1178         /* Currently, all async queues are mapped to root group */
1179         if (!cfq_cfqq_sync(cfqq))
1180                 cfqg = &cfqq->cfqd->root_group;
1181
1182         cfqq->cfqg = cfqg;
1183         /* cfqq reference on cfqg */
1184         cfqq->cfqg->ref++;
1185 }
1186
1187 static void cfq_put_cfqg(struct cfq_group *cfqg)
1188 {
1189         struct cfq_rb_root *st;
1190         int i, j;
1191
1192         BUG_ON(cfqg->ref <= 0);
1193         cfqg->ref--;
1194         if (cfqg->ref)
1195                 return;
1196         for_each_cfqg_st(cfqg, i, j, st)
1197                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1198         free_percpu(cfqg->blkg.stats_cpu);
1199         kfree(cfqg);
1200 }
1201
1202 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1203 {
1204         /* Something wrong if we are trying to remove same group twice */
1205         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1206
1207         hlist_del_init(&cfqg->cfqd_node);
1208
1209         /*
1210          * Put the reference taken at the time of creation so that when all
1211          * queues are gone, group can be destroyed.
1212          */
1213         cfq_put_cfqg(cfqg);
1214 }
1215
1216 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1217 {
1218         struct hlist_node *pos, *n;
1219         struct cfq_group *cfqg;
1220
1221         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1222                 /*
1223                  * If cgroup removal path got to blk_group first and removed
1224                  * it from cgroup list, then it will take care of destroying
1225                  * cfqg also.
1226                  */
1227                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1228                         cfq_destroy_cfqg(cfqd, cfqg);
1229         }
1230 }
1231
1232 /*
1233  * Blk cgroup controller notification saying that blkio_group object is being
1234  * delinked as associated cgroup object is going away. That also means that
1235  * no new IO will come in this group. So get rid of this group as soon as
1236  * any pending IO in the group is finished.
1237  *
1238  * This function is called under rcu_read_lock(). key is the rcu protected
1239  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1240  * read lock.
1241  *
1242  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1243  * it should not be NULL as even if elevator was exiting, cgroup deltion
1244  * path got to it first.
1245  */
1246 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1247 {
1248         unsigned long  flags;
1249         struct cfq_data *cfqd = key;
1250
1251         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1252         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1253         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1254 }
1255
1256 #else /* GROUP_IOSCHED */
1257 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1258 {
1259         return &cfqd->root_group;
1260 }
1261
1262 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1263 {
1264         return cfqg;
1265 }
1266
1267 static inline void
1268 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1269         cfqq->cfqg = cfqg;
1270 }
1271
1272 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1273 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1274
1275 #endif /* GROUP_IOSCHED */
1276
1277 /*
1278  * The cfqd->service_trees holds all pending cfq_queue's that have
1279  * requests waiting to be processed. It is sorted in the order that
1280  * we will service the queues.
1281  */
1282 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1283                                  bool add_front)
1284 {
1285         struct rb_node **p, *parent;
1286         struct cfq_queue *__cfqq;
1287         unsigned long rb_key;
1288         struct cfq_rb_root *service_tree;
1289         int left;
1290         int new_cfqq = 1;
1291
1292         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1293                                                 cfqq_type(cfqq));
1294         if (cfq_class_idle(cfqq)) {
1295                 rb_key = CFQ_IDLE_DELAY;
1296                 parent = rb_last(&service_tree->rb);
1297                 if (parent && parent != &cfqq->rb_node) {
1298                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1299                         rb_key += __cfqq->rb_key;
1300                 } else
1301                         rb_key += jiffies;
1302         } else if (!add_front) {
1303                 /*
1304                  * Get our rb key offset. Subtract any residual slice
1305                  * value carried from last service. A negative resid
1306                  * count indicates slice overrun, and this should position
1307                  * the next service time further away in the tree.
1308                  */
1309                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1310                 rb_key -= cfqq->slice_resid;
1311                 cfqq->slice_resid = 0;
1312         } else {
1313                 rb_key = -HZ;
1314                 __cfqq = cfq_rb_first(service_tree);
1315                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1316         }
1317
1318         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1319                 new_cfqq = 0;
1320                 /*
1321                  * same position, nothing more to do
1322                  */
1323                 if (rb_key == cfqq->rb_key &&
1324                     cfqq->service_tree == service_tree)
1325                         return;
1326
1327                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1328                 cfqq->service_tree = NULL;
1329         }
1330
1331         left = 1;
1332         parent = NULL;
1333         cfqq->service_tree = service_tree;
1334         p = &service_tree->rb.rb_node;
1335         while (*p) {
1336                 struct rb_node **n;
1337
1338                 parent = *p;
1339                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1340
1341                 /*
1342                  * sort by key, that represents service time.
1343                  */
1344                 if (time_before(rb_key, __cfqq->rb_key))
1345                         n = &(*p)->rb_left;
1346                 else {
1347                         n = &(*p)->rb_right;
1348                         left = 0;
1349                 }
1350
1351                 p = n;
1352         }
1353
1354         if (left)
1355                 service_tree->left = &cfqq->rb_node;
1356
1357         cfqq->rb_key = rb_key;
1358         rb_link_node(&cfqq->rb_node, parent, p);
1359         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1360         service_tree->count++;
1361         if (add_front || !new_cfqq)
1362                 return;
1363         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1364 }
1365
1366 static struct cfq_queue *
1367 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1368                      sector_t sector, struct rb_node **ret_parent,
1369                      struct rb_node ***rb_link)
1370 {
1371         struct rb_node **p, *parent;
1372         struct cfq_queue *cfqq = NULL;
1373
1374         parent = NULL;
1375         p = &root->rb_node;
1376         while (*p) {
1377                 struct rb_node **n;
1378
1379                 parent = *p;
1380                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1381
1382                 /*
1383                  * Sort strictly based on sector.  Smallest to the left,
1384                  * largest to the right.
1385                  */
1386                 if (sector > blk_rq_pos(cfqq->next_rq))
1387                         n = &(*p)->rb_right;
1388                 else if (sector < blk_rq_pos(cfqq->next_rq))
1389                         n = &(*p)->rb_left;
1390                 else
1391                         break;
1392                 p = n;
1393                 cfqq = NULL;
1394         }
1395
1396         *ret_parent = parent;
1397         if (rb_link)
1398                 *rb_link = p;
1399         return cfqq;
1400 }
1401
1402 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1403 {
1404         struct rb_node **p, *parent;
1405         struct cfq_queue *__cfqq;
1406
1407         if (cfqq->p_root) {
1408                 rb_erase(&cfqq->p_node, cfqq->p_root);
1409                 cfqq->p_root = NULL;
1410         }
1411
1412         if (cfq_class_idle(cfqq))
1413                 return;
1414         if (!cfqq->next_rq)
1415                 return;
1416
1417         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1418         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1419                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1420         if (!__cfqq) {
1421                 rb_link_node(&cfqq->p_node, parent, p);
1422                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1423         } else
1424                 cfqq->p_root = NULL;
1425 }
1426
1427 /*
1428  * Update cfqq's position in the service tree.
1429  */
1430 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1431 {
1432         /*
1433          * Resorting requires the cfqq to be on the RR list already.
1434          */
1435         if (cfq_cfqq_on_rr(cfqq)) {
1436                 cfq_service_tree_add(cfqd, cfqq, 0);
1437                 cfq_prio_tree_add(cfqd, cfqq);
1438         }
1439 }
1440
1441 /*
1442  * add to busy list of queues for service, trying to be fair in ordering
1443  * the pending list according to last request service
1444  */
1445 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1446 {
1447         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1448         BUG_ON(cfq_cfqq_on_rr(cfqq));
1449         cfq_mark_cfqq_on_rr(cfqq);
1450         cfqd->busy_queues++;
1451         if (cfq_cfqq_sync(cfqq))
1452                 cfqd->busy_sync_queues++;
1453
1454         cfq_resort_rr_list(cfqd, cfqq);
1455 }
1456
1457 /*
1458  * Called when the cfqq no longer has requests pending, remove it from
1459  * the service tree.
1460  */
1461 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1462 {
1463         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1464         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1465         cfq_clear_cfqq_on_rr(cfqq);
1466
1467         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1468                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1469                 cfqq->service_tree = NULL;
1470         }
1471         if (cfqq->p_root) {
1472                 rb_erase(&cfqq->p_node, cfqq->p_root);
1473                 cfqq->p_root = NULL;
1474         }
1475
1476         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1477         BUG_ON(!cfqd->busy_queues);
1478         cfqd->busy_queues--;
1479         if (cfq_cfqq_sync(cfqq))
1480                 cfqd->busy_sync_queues--;
1481 }
1482
1483 /*
1484  * rb tree support functions
1485  */
1486 static void cfq_del_rq_rb(struct request *rq)
1487 {
1488         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1489         const int sync = rq_is_sync(rq);
1490
1491         BUG_ON(!cfqq->queued[sync]);
1492         cfqq->queued[sync]--;
1493
1494         elv_rb_del(&cfqq->sort_list, rq);
1495
1496         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1497                 /*
1498                  * Queue will be deleted from service tree when we actually
1499                  * expire it later. Right now just remove it from prio tree
1500                  * as it is empty.
1501                  */
1502                 if (cfqq->p_root) {
1503                         rb_erase(&cfqq->p_node, cfqq->p_root);
1504                         cfqq->p_root = NULL;
1505                 }
1506         }
1507 }
1508
1509 static void cfq_add_rq_rb(struct request *rq)
1510 {
1511         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1512         struct cfq_data *cfqd = cfqq->cfqd;
1513         struct request *prev;
1514
1515         cfqq->queued[rq_is_sync(rq)]++;
1516
1517         elv_rb_add(&cfqq->sort_list, rq);
1518
1519         if (!cfq_cfqq_on_rr(cfqq))
1520                 cfq_add_cfqq_rr(cfqd, cfqq);
1521
1522         /*
1523          * check if this request is a better next-serve candidate
1524          */
1525         prev = cfqq->next_rq;
1526         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1527
1528         /*
1529          * adjust priority tree position, if ->next_rq changes
1530          */
1531         if (prev != cfqq->next_rq)
1532                 cfq_prio_tree_add(cfqd, cfqq);
1533
1534         BUG_ON(!cfqq->next_rq);
1535 }
1536
1537 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1538 {
1539         elv_rb_del(&cfqq->sort_list, rq);
1540         cfqq->queued[rq_is_sync(rq)]--;
1541         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1542                                         rq_data_dir(rq), rq_is_sync(rq));
1543         cfq_add_rq_rb(rq);
1544         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1545                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1546                         rq_is_sync(rq));
1547 }
1548
1549 static struct request *
1550 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1551 {
1552         struct task_struct *tsk = current;
1553         struct cfq_io_context *cic;
1554         struct cfq_queue *cfqq;
1555
1556         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1557         if (!cic)
1558                 return NULL;
1559
1560         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1561         if (cfqq) {
1562                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1563
1564                 return elv_rb_find(&cfqq->sort_list, sector);
1565         }
1566
1567         return NULL;
1568 }
1569
1570 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1571 {
1572         struct cfq_data *cfqd = q->elevator->elevator_data;
1573
1574         cfqd->rq_in_driver++;
1575         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1576                                                 cfqd->rq_in_driver);
1577
1578         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1579 }
1580
1581 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1582 {
1583         struct cfq_data *cfqd = q->elevator->elevator_data;
1584
1585         WARN_ON(!cfqd->rq_in_driver);
1586         cfqd->rq_in_driver--;
1587         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1588                                                 cfqd->rq_in_driver);
1589 }
1590
1591 static void cfq_remove_request(struct request *rq)
1592 {
1593         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1594
1595         if (cfqq->next_rq == rq)
1596                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1597
1598         list_del_init(&rq->queuelist);
1599         cfq_del_rq_rb(rq);
1600
1601         cfqq->cfqd->rq_queued--;
1602         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1603                                         rq_data_dir(rq), rq_is_sync(rq));
1604 }
1605
1606 static int cfq_merge(struct request_queue *q, struct request **req,
1607                      struct bio *bio)
1608 {
1609         struct cfq_data *cfqd = q->elevator->elevator_data;
1610         struct request *__rq;
1611
1612         __rq = cfq_find_rq_fmerge(cfqd, bio);
1613         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1614                 *req = __rq;
1615                 return ELEVATOR_FRONT_MERGE;
1616         }
1617
1618         return ELEVATOR_NO_MERGE;
1619 }
1620
1621 static void cfq_merged_request(struct request_queue *q, struct request *req,
1622                                int type)
1623 {
1624         if (type == ELEVATOR_FRONT_MERGE) {
1625                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1626
1627                 cfq_reposition_rq_rb(cfqq, req);
1628         }
1629 }
1630
1631 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1632                                 struct bio *bio)
1633 {
1634         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1635                                         bio_data_dir(bio), cfq_bio_sync(bio));
1636 }
1637
1638 static void
1639 cfq_merged_requests(struct request_queue *q, struct request *rq,
1640                     struct request *next)
1641 {
1642         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1643         /*
1644          * reposition in fifo if next is older than rq
1645          */
1646         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1647             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1648                 list_move(&rq->queuelist, &next->queuelist);
1649                 rq_set_fifo_time(rq, rq_fifo_time(next));
1650         }
1651
1652         if (cfqq->next_rq == next)
1653                 cfqq->next_rq = rq;
1654         cfq_remove_request(next);
1655         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1656                                         rq_data_dir(next), rq_is_sync(next));
1657 }
1658
1659 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1660                            struct bio *bio)
1661 {
1662         struct cfq_data *cfqd = q->elevator->elevator_data;
1663         struct cfq_io_context *cic;
1664         struct cfq_queue *cfqq;
1665
1666         /*
1667          * Disallow merge of a sync bio into an async request.
1668          */
1669         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1670                 return false;
1671
1672         /*
1673          * Lookup the cfqq that this bio will be queued with. Allow
1674          * merge only if rq is queued there.
1675          */
1676         cic = cfq_cic_lookup(cfqd, current->io_context);
1677         if (!cic)
1678                 return false;
1679
1680         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1681         return cfqq == RQ_CFQQ(rq);
1682 }
1683
1684 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1685 {
1686         del_timer(&cfqd->idle_slice_timer);
1687         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1688 }
1689
1690 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1691                                    struct cfq_queue *cfqq)
1692 {
1693         if (cfqq) {
1694                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1695                                 cfqd->serving_prio, cfqd->serving_type);
1696                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1697                 cfqq->slice_start = 0;
1698                 cfqq->dispatch_start = jiffies;
1699                 cfqq->allocated_slice = 0;
1700                 cfqq->slice_end = 0;
1701                 cfqq->slice_dispatch = 0;
1702                 cfqq->nr_sectors = 0;
1703
1704                 cfq_clear_cfqq_wait_request(cfqq);
1705                 cfq_clear_cfqq_must_dispatch(cfqq);
1706                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1707                 cfq_clear_cfqq_fifo_expire(cfqq);
1708                 cfq_mark_cfqq_slice_new(cfqq);
1709
1710                 cfq_del_timer(cfqd, cfqq);
1711         }
1712
1713         cfqd->active_queue = cfqq;
1714 }
1715
1716 /*
1717  * current cfqq expired its slice (or was too idle), select new one
1718  */
1719 static void
1720 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1721                     bool timed_out)
1722 {
1723         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1724
1725         if (cfq_cfqq_wait_request(cfqq))
1726                 cfq_del_timer(cfqd, cfqq);
1727
1728         cfq_clear_cfqq_wait_request(cfqq);
1729         cfq_clear_cfqq_wait_busy(cfqq);
1730
1731         /*
1732          * If this cfqq is shared between multiple processes, check to
1733          * make sure that those processes are still issuing I/Os within
1734          * the mean seek distance.  If not, it may be time to break the
1735          * queues apart again.
1736          */
1737         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1738                 cfq_mark_cfqq_split_coop(cfqq);
1739
1740         /*
1741          * store what was left of this slice, if the queue idled/timed out
1742          */
1743         if (timed_out) {
1744                 if (cfq_cfqq_slice_new(cfqq))
1745                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1746                 else
1747                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1748                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1749         }
1750
1751         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1752
1753         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1754                 cfq_del_cfqq_rr(cfqd, cfqq);
1755
1756         cfq_resort_rr_list(cfqd, cfqq);
1757
1758         if (cfqq == cfqd->active_queue)
1759                 cfqd->active_queue = NULL;
1760
1761         if (cfqd->active_cic) {
1762                 put_io_context(cfqd->active_cic->ioc);
1763                 cfqd->active_cic = NULL;
1764         }
1765 }
1766
1767 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1768 {
1769         struct cfq_queue *cfqq = cfqd->active_queue;
1770
1771         if (cfqq)
1772                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1773 }
1774
1775 /*
1776  * Get next queue for service. Unless we have a queue preemption,
1777  * we'll simply select the first cfqq in the service tree.
1778  */
1779 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1780 {
1781         struct cfq_rb_root *service_tree =
1782                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1783                                         cfqd->serving_type);
1784
1785         if (!cfqd->rq_queued)
1786                 return NULL;
1787
1788         /* There is nothing to dispatch */
1789         if (!service_tree)
1790                 return NULL;
1791         if (RB_EMPTY_ROOT(&service_tree->rb))
1792                 return NULL;
1793         return cfq_rb_first(service_tree);
1794 }
1795
1796 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1797 {
1798         struct cfq_group *cfqg;
1799         struct cfq_queue *cfqq;
1800         int i, j;
1801         struct cfq_rb_root *st;
1802
1803         if (!cfqd->rq_queued)
1804                 return NULL;
1805
1806         cfqg = cfq_get_next_cfqg(cfqd);
1807         if (!cfqg)
1808                 return NULL;
1809
1810         for_each_cfqg_st(cfqg, i, j, st)
1811                 if ((cfqq = cfq_rb_first(st)) != NULL)
1812                         return cfqq;
1813         return NULL;
1814 }
1815
1816 /*
1817  * Get and set a new active queue for service.
1818  */
1819 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1820                                               struct cfq_queue *cfqq)
1821 {
1822         if (!cfqq)
1823                 cfqq = cfq_get_next_queue(cfqd);
1824
1825         __cfq_set_active_queue(cfqd, cfqq);
1826         return cfqq;
1827 }
1828
1829 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1830                                           struct request *rq)
1831 {
1832         if (blk_rq_pos(rq) >= cfqd->last_position)
1833                 return blk_rq_pos(rq) - cfqd->last_position;
1834         else
1835                 return cfqd->last_position - blk_rq_pos(rq);
1836 }
1837
1838 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1839                                struct request *rq)
1840 {
1841         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1842 }
1843
1844 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1845                                     struct cfq_queue *cur_cfqq)
1846 {
1847         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1848         struct rb_node *parent, *node;
1849         struct cfq_queue *__cfqq;
1850         sector_t sector = cfqd->last_position;
1851
1852         if (RB_EMPTY_ROOT(root))
1853                 return NULL;
1854
1855         /*
1856          * First, if we find a request starting at the end of the last
1857          * request, choose it.
1858          */
1859         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1860         if (__cfqq)
1861                 return __cfqq;
1862
1863         /*
1864          * If the exact sector wasn't found, the parent of the NULL leaf
1865          * will contain the closest sector.
1866          */
1867         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1868         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1869                 return __cfqq;
1870
1871         if (blk_rq_pos(__cfqq->next_rq) < sector)
1872                 node = rb_next(&__cfqq->p_node);
1873         else
1874                 node = rb_prev(&__cfqq->p_node);
1875         if (!node)
1876                 return NULL;
1877
1878         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1879         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1880                 return __cfqq;
1881
1882         return NULL;
1883 }
1884
1885 /*
1886  * cfqd - obvious
1887  * cur_cfqq - passed in so that we don't decide that the current queue is
1888  *            closely cooperating with itself.
1889  *
1890  * So, basically we're assuming that that cur_cfqq has dispatched at least
1891  * one request, and that cfqd->last_position reflects a position on the disk
1892  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1893  * assumption.
1894  */
1895 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1896                                               struct cfq_queue *cur_cfqq)
1897 {
1898         struct cfq_queue *cfqq;
1899
1900         if (cfq_class_idle(cur_cfqq))
1901                 return NULL;
1902         if (!cfq_cfqq_sync(cur_cfqq))
1903                 return NULL;
1904         if (CFQQ_SEEKY(cur_cfqq))
1905                 return NULL;
1906
1907         /*
1908          * Don't search priority tree if it's the only queue in the group.
1909          */
1910         if (cur_cfqq->cfqg->nr_cfqq == 1)
1911                 return NULL;
1912
1913         /*
1914          * We should notice if some of the queues are cooperating, eg
1915          * working closely on the same area of the disk. In that case,
1916          * we can group them together and don't waste time idling.
1917          */
1918         cfqq = cfqq_close(cfqd, cur_cfqq);
1919         if (!cfqq)
1920                 return NULL;
1921
1922         /* If new queue belongs to different cfq_group, don't choose it */
1923         if (cur_cfqq->cfqg != cfqq->cfqg)
1924                 return NULL;
1925
1926         /*
1927          * It only makes sense to merge sync queues.
1928          */
1929         if (!cfq_cfqq_sync(cfqq))
1930                 return NULL;
1931         if (CFQQ_SEEKY(cfqq))
1932                 return NULL;
1933
1934         /*
1935          * Do not merge queues of different priority classes
1936          */
1937         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1938                 return NULL;
1939
1940         return cfqq;
1941 }
1942
1943 /*
1944  * Determine whether we should enforce idle window for this queue.
1945  */
1946
1947 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1948 {
1949         enum wl_prio_t prio = cfqq_prio(cfqq);
1950         struct cfq_rb_root *service_tree = cfqq->service_tree;
1951
1952         BUG_ON(!service_tree);
1953         BUG_ON(!service_tree->count);
1954
1955         if (!cfqd->cfq_slice_idle)
1956                 return false;
1957
1958         /* We never do for idle class queues. */
1959         if (prio == IDLE_WORKLOAD)
1960                 return false;
1961
1962         /* We do for queues that were marked with idle window flag. */
1963         if (cfq_cfqq_idle_window(cfqq) &&
1964            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1965                 return true;
1966
1967         /*
1968          * Otherwise, we do only if they are the last ones
1969          * in their service tree.
1970          */
1971         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1972            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1973                 return true;
1974         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1975                         service_tree->count);
1976         return false;
1977 }
1978
1979 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1980 {
1981         struct cfq_queue *cfqq = cfqd->active_queue;
1982         struct cfq_io_context *cic;
1983         unsigned long sl, group_idle = 0;
1984
1985         /*
1986          * SSD device without seek penalty, disable idling. But only do so
1987          * for devices that support queuing, otherwise we still have a problem
1988          * with sync vs async workloads.
1989          */
1990         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1991                 return;
1992
1993         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1994         WARN_ON(cfq_cfqq_slice_new(cfqq));
1995
1996         /*
1997          * idle is disabled, either manually or by past process history
1998          */
1999         if (!cfq_should_idle(cfqd, cfqq)) {
2000                 /* no queue idling. Check for group idling */
2001                 if (cfqd->cfq_group_idle)
2002                         group_idle = cfqd->cfq_group_idle;
2003                 else
2004                         return;
2005         }
2006
2007         /*
2008          * still active requests from this queue, don't idle
2009          */
2010         if (cfqq->dispatched)
2011                 return;
2012
2013         /*
2014          * task has exited, don't wait
2015          */
2016         cic = cfqd->active_cic;
2017         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2018                 return;
2019
2020         /*
2021          * If our average think time is larger than the remaining time
2022          * slice, then don't idle. This avoids overrunning the allotted
2023          * time slice.
2024          */
2025         if (sample_valid(cic->ttime.ttime_samples) &&
2026             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2027                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2028                              cic->ttime.ttime_mean);
2029                 return;
2030         }
2031
2032         /* There are other queues in the group, don't do group idle */
2033         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2034                 return;
2035
2036         cfq_mark_cfqq_wait_request(cfqq);
2037
2038         if (group_idle)
2039                 sl = cfqd->cfq_group_idle;
2040         else
2041                 sl = cfqd->cfq_slice_idle;
2042
2043         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2044         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2045         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2046                         group_idle ? 1 : 0);
2047 }
2048
2049 /*
2050  * Move request from internal lists to the request queue dispatch list.
2051  */
2052 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2053 {
2054         struct cfq_data *cfqd = q->elevator->elevator_data;
2055         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2056
2057         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2058
2059         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2060         cfq_remove_request(rq);
2061         cfqq->dispatched++;
2062         (RQ_CFQG(rq))->dispatched++;
2063         elv_dispatch_sort(q, rq);
2064
2065         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2066         cfqq->nr_sectors += blk_rq_sectors(rq);
2067         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2068                                         rq_data_dir(rq), rq_is_sync(rq));
2069 }
2070
2071 /*
2072  * return expired entry, or NULL to just start from scratch in rbtree
2073  */
2074 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2075 {
2076         struct request *rq = NULL;
2077
2078         if (cfq_cfqq_fifo_expire(cfqq))
2079                 return NULL;
2080
2081         cfq_mark_cfqq_fifo_expire(cfqq);
2082
2083         if (list_empty(&cfqq->fifo))
2084                 return NULL;
2085
2086         rq = rq_entry_fifo(cfqq->fifo.next);
2087         if (time_before(jiffies, rq_fifo_time(rq)))
2088                 rq = NULL;
2089
2090         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2091         return rq;
2092 }
2093
2094 static inline int
2095 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2096 {
2097         const int base_rq = cfqd->cfq_slice_async_rq;
2098
2099         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2100
2101         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2102 }
2103
2104 /*
2105  * Must be called with the queue_lock held.
2106  */
2107 static int cfqq_process_refs(struct cfq_queue *cfqq)
2108 {
2109         int process_refs, io_refs;
2110
2111         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2112         process_refs = cfqq->ref - io_refs;
2113         BUG_ON(process_refs < 0);
2114         return process_refs;
2115 }
2116
2117 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2118 {
2119         int process_refs, new_process_refs;
2120         struct cfq_queue *__cfqq;
2121
2122         /*
2123          * If there are no process references on the new_cfqq, then it is
2124          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2125          * chain may have dropped their last reference (not just their
2126          * last process reference).
2127          */
2128         if (!cfqq_process_refs(new_cfqq))
2129                 return;
2130
2131         /* Avoid a circular list and skip interim queue merges */
2132         while ((__cfqq = new_cfqq->new_cfqq)) {
2133                 if (__cfqq == cfqq)
2134                         return;
2135                 new_cfqq = __cfqq;
2136         }
2137
2138         process_refs = cfqq_process_refs(cfqq);
2139         new_process_refs = cfqq_process_refs(new_cfqq);
2140         /*
2141          * If the process for the cfqq has gone away, there is no
2142          * sense in merging the queues.
2143          */
2144         if (process_refs == 0 || new_process_refs == 0)
2145                 return;
2146
2147         /*
2148          * Merge in the direction of the lesser amount of work.
2149          */
2150         if (new_process_refs >= process_refs) {
2151                 cfqq->new_cfqq = new_cfqq;
2152                 new_cfqq->ref += process_refs;
2153         } else {
2154                 new_cfqq->new_cfqq = cfqq;
2155                 cfqq->ref += new_process_refs;
2156         }
2157 }
2158
2159 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2160                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2161 {
2162         struct cfq_queue *queue;
2163         int i;
2164         bool key_valid = false;
2165         unsigned long lowest_key = 0;
2166         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2167
2168         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2169                 /* select the one with lowest rb_key */
2170                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2171                 if (queue &&
2172                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2173                         lowest_key = queue->rb_key;
2174                         cur_best = i;
2175                         key_valid = true;
2176                 }
2177         }
2178
2179         return cur_best;
2180 }
2181
2182 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2183 {
2184         unsigned slice;
2185         unsigned count;
2186         struct cfq_rb_root *st;
2187         unsigned group_slice;
2188         enum wl_prio_t original_prio = cfqd->serving_prio;
2189
2190         /* Choose next priority. RT > BE > IDLE */
2191         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2192                 cfqd->serving_prio = RT_WORKLOAD;
2193         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2194                 cfqd->serving_prio = BE_WORKLOAD;
2195         else {
2196                 cfqd->serving_prio = IDLE_WORKLOAD;
2197                 cfqd->workload_expires = jiffies + 1;
2198                 return;
2199         }
2200
2201         if (original_prio != cfqd->serving_prio)
2202                 goto new_workload;
2203
2204         /*
2205          * For RT and BE, we have to choose also the type
2206          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2207          * expiration time
2208          */
2209         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2210         count = st->count;
2211
2212         /*
2213          * check workload expiration, and that we still have other queues ready
2214          */
2215         if (count && !time_after(jiffies, cfqd->workload_expires))
2216                 return;
2217
2218 new_workload:
2219         /* otherwise select new workload type */
2220         cfqd->serving_type =
2221                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2222         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2223         count = st->count;
2224
2225         /*
2226          * the workload slice is computed as a fraction of target latency
2227          * proportional to the number of queues in that workload, over
2228          * all the queues in the same priority class
2229          */
2230         group_slice = cfq_group_slice(cfqd, cfqg);
2231
2232         slice = group_slice * count /
2233                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2234                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2235
2236         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2237                 unsigned int tmp;
2238
2239                 /*
2240                  * Async queues are currently system wide. Just taking
2241                  * proportion of queues with-in same group will lead to higher
2242                  * async ratio system wide as generally root group is going
2243                  * to have higher weight. A more accurate thing would be to
2244                  * calculate system wide asnc/sync ratio.
2245                  */
2246                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2247                 tmp = tmp/cfqd->busy_queues;
2248                 slice = min_t(unsigned, slice, tmp);
2249
2250                 /* async workload slice is scaled down according to
2251                  * the sync/async slice ratio. */
2252                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2253         } else
2254                 /* sync workload slice is at least 2 * cfq_slice_idle */
2255                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2256
2257         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2258         cfq_log(cfqd, "workload slice:%d", slice);
2259         cfqd->workload_expires = jiffies + slice;
2260 }
2261
2262 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2263 {
2264         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2265         struct cfq_group *cfqg;
2266
2267         if (RB_EMPTY_ROOT(&st->rb))
2268                 return NULL;
2269         cfqg = cfq_rb_first_group(st);
2270         update_min_vdisktime(st);
2271         return cfqg;
2272 }
2273
2274 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2275 {
2276         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2277
2278         cfqd->serving_group = cfqg;
2279
2280         /* Restore the workload type data */
2281         if (cfqg->saved_workload_slice) {
2282                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2283                 cfqd->serving_type = cfqg->saved_workload;
2284                 cfqd->serving_prio = cfqg->saved_serving_prio;
2285         } else
2286                 cfqd->workload_expires = jiffies - 1;
2287
2288         choose_service_tree(cfqd, cfqg);
2289 }
2290
2291 /*
2292  * Select a queue for service. If we have a current active queue,
2293  * check whether to continue servicing it, or retrieve and set a new one.
2294  */
2295 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2296 {
2297         struct cfq_queue *cfqq, *new_cfqq = NULL;
2298
2299         cfqq = cfqd->active_queue;
2300         if (!cfqq)
2301                 goto new_queue;
2302
2303         if (!cfqd->rq_queued)
2304                 return NULL;
2305
2306         /*
2307          * We were waiting for group to get backlogged. Expire the queue
2308          */
2309         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2310                 goto expire;
2311
2312         /*
2313          * The active queue has run out of time, expire it and select new.
2314          */
2315         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2316                 /*
2317                  * If slice had not expired at the completion of last request
2318                  * we might not have turned on wait_busy flag. Don't expire
2319                  * the queue yet. Allow the group to get backlogged.
2320                  *
2321                  * The very fact that we have used the slice, that means we
2322                  * have been idling all along on this queue and it should be
2323                  * ok to wait for this request to complete.
2324                  */
2325                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2326                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2327                         cfqq = NULL;
2328                         goto keep_queue;
2329                 } else
2330                         goto check_group_idle;
2331         }
2332
2333         /*
2334          * The active queue has requests and isn't expired, allow it to
2335          * dispatch.
2336          */
2337         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2338                 goto keep_queue;
2339
2340         /*
2341          * If another queue has a request waiting within our mean seek
2342          * distance, let it run.  The expire code will check for close
2343          * cooperators and put the close queue at the front of the service
2344          * tree.  If possible, merge the expiring queue with the new cfqq.
2345          */
2346         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2347         if (new_cfqq) {
2348                 if (!cfqq->new_cfqq)
2349                         cfq_setup_merge(cfqq, new_cfqq);
2350                 goto expire;
2351         }
2352
2353         /*
2354          * No requests pending. If the active queue still has requests in
2355          * flight or is idling for a new request, allow either of these
2356          * conditions to happen (or time out) before selecting a new queue.
2357          */
2358         if (timer_pending(&cfqd->idle_slice_timer)) {
2359                 cfqq = NULL;
2360                 goto keep_queue;
2361         }
2362
2363         /*
2364          * This is a deep seek queue, but the device is much faster than
2365          * the queue can deliver, don't idle
2366          **/
2367         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2368             (cfq_cfqq_slice_new(cfqq) ||
2369             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2370                 cfq_clear_cfqq_deep(cfqq);
2371                 cfq_clear_cfqq_idle_window(cfqq);
2372         }
2373
2374         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2375                 cfqq = NULL;
2376                 goto keep_queue;
2377         }
2378
2379         /*
2380          * If group idle is enabled and there are requests dispatched from
2381          * this group, wait for requests to complete.
2382          */
2383 check_group_idle:
2384         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2385             && cfqq->cfqg->dispatched) {
2386                 cfqq = NULL;
2387                 goto keep_queue;
2388         }
2389
2390 expire:
2391         cfq_slice_expired(cfqd, 0);
2392 new_queue:
2393         /*
2394          * Current queue expired. Check if we have to switch to a new
2395          * service tree
2396          */
2397         if (!new_cfqq)
2398                 cfq_choose_cfqg(cfqd);
2399
2400         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2401 keep_queue:
2402         return cfqq;
2403 }
2404
2405 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2406 {
2407         int dispatched = 0;
2408
2409         while (cfqq->next_rq) {
2410                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2411                 dispatched++;
2412         }
2413
2414         BUG_ON(!list_empty(&cfqq->fifo));
2415
2416         /* By default cfqq is not expired if it is empty. Do it explicitly */
2417         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2418         return dispatched;
2419 }
2420
2421 /*
2422  * Drain our current requests. Used for barriers and when switching
2423  * io schedulers on-the-fly.
2424  */
2425 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2426 {
2427         struct cfq_queue *cfqq;
2428         int dispatched = 0;
2429
2430         /* Expire the timeslice of the current active queue first */
2431         cfq_slice_expired(cfqd, 0);
2432         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2433                 __cfq_set_active_queue(cfqd, cfqq);
2434                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2435         }
2436
2437         BUG_ON(cfqd->busy_queues);
2438
2439         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2440         return dispatched;
2441 }
2442
2443 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2444         struct cfq_queue *cfqq)
2445 {
2446         /* the queue hasn't finished any request, can't estimate */
2447         if (cfq_cfqq_slice_new(cfqq))
2448                 return true;
2449         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2450                 cfqq->slice_end))
2451                 return true;
2452
2453         return false;
2454 }
2455
2456 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2457 {
2458         unsigned int max_dispatch;
2459
2460         /*
2461          * Drain async requests before we start sync IO
2462          */
2463         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2464                 return false;
2465
2466         /*
2467          * If this is an async queue and we have sync IO in flight, let it wait
2468          */
2469         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2470                 return false;
2471
2472         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2473         if (cfq_class_idle(cfqq))
2474                 max_dispatch = 1;
2475
2476         /*
2477          * Does this cfqq already have too much IO in flight?
2478          */
2479         if (cfqq->dispatched >= max_dispatch) {
2480                 bool promote_sync = false;
2481                 /*
2482                  * idle queue must always only have a single IO in flight
2483                  */
2484                 if (cfq_class_idle(cfqq))
2485                         return false;
2486
2487                 /*
2488                  * If there is only one sync queue
2489                  * we can ignore async queue here and give the sync
2490                  * queue no dispatch limit. The reason is a sync queue can
2491                  * preempt async queue, limiting the sync queue doesn't make
2492                  * sense. This is useful for aiostress test.
2493                  */
2494                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2495                         promote_sync = true;
2496
2497                 /*
2498                  * We have other queues, don't allow more IO from this one
2499                  */
2500                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2501                                 !promote_sync)
2502                         return false;
2503
2504                 /*
2505                  * Sole queue user, no limit
2506                  */
2507                 if (cfqd->busy_queues == 1 || promote_sync)
2508                         max_dispatch = -1;
2509                 else
2510                         /*
2511                          * Normally we start throttling cfqq when cfq_quantum/2
2512                          * requests have been dispatched. But we can drive
2513                          * deeper queue depths at the beginning of slice
2514                          * subjected to upper limit of cfq_quantum.
2515                          * */
2516                         max_dispatch = cfqd->cfq_quantum;
2517         }
2518
2519         /*
2520          * Async queues must wait a bit before being allowed dispatch.
2521          * We also ramp up the dispatch depth gradually for async IO,
2522          * based on the last sync IO we serviced
2523          */
2524         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2525                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2526                 unsigned int depth;
2527
2528                 depth = last_sync / cfqd->cfq_slice[1];
2529                 if (!depth && !cfqq->dispatched)
2530                         depth = 1;
2531                 if (depth < max_dispatch)
2532                         max_dispatch = depth;
2533         }
2534
2535         /*
2536          * If we're below the current max, allow a dispatch
2537          */
2538         return cfqq->dispatched < max_dispatch;
2539 }
2540
2541 /*
2542  * Dispatch a request from cfqq, moving them to the request queue
2543  * dispatch list.
2544  */
2545 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2546 {
2547         struct request *rq;
2548
2549         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2550
2551         if (!cfq_may_dispatch(cfqd, cfqq))
2552                 return false;
2553
2554         /*
2555          * follow expired path, else get first next available
2556          */
2557         rq = cfq_check_fifo(cfqq);
2558         if (!rq)
2559                 rq = cfqq->next_rq;
2560
2561         /*
2562          * insert request into driver dispatch list
2563          */
2564         cfq_dispatch_insert(cfqd->queue, rq);
2565
2566         if (!cfqd->active_cic) {
2567                 struct cfq_io_context *cic = RQ_CIC(rq);
2568
2569                 atomic_long_inc(&cic->ioc->refcount);
2570                 cfqd->active_cic = cic;
2571         }
2572
2573         return true;
2574 }
2575
2576 /*
2577  * Find the cfqq that we need to service and move a request from that to the
2578  * dispatch list
2579  */
2580 static int cfq_dispatch_requests(struct request_queue *q, int force)
2581 {
2582         struct cfq_data *cfqd = q->elevator->elevator_data;
2583         struct cfq_queue *cfqq;
2584
2585         if (!cfqd->busy_queues)
2586                 return 0;
2587
2588         if (unlikely(force))
2589                 return cfq_forced_dispatch(cfqd);
2590
2591         cfqq = cfq_select_queue(cfqd);
2592         if (!cfqq)
2593                 return 0;
2594
2595         /*
2596          * Dispatch a request from this cfqq, if it is allowed
2597          */
2598         if (!cfq_dispatch_request(cfqd, cfqq))
2599                 return 0;
2600
2601         cfqq->slice_dispatch++;
2602         cfq_clear_cfqq_must_dispatch(cfqq);
2603
2604         /*
2605          * expire an async queue immediately if it has used up its slice. idle
2606          * queue always expire after 1 dispatch round.
2607          */
2608         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2609             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2610             cfq_class_idle(cfqq))) {
2611                 cfqq->slice_end = jiffies + 1;
2612                 cfq_slice_expired(cfqd, 0);
2613         }
2614
2615         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2616         return 1;
2617 }
2618
2619 /*
2620  * task holds one reference to the queue, dropped when task exits. each rq
2621  * in-flight on this queue also holds a reference, dropped when rq is freed.
2622  *
2623  * Each cfq queue took a reference on the parent group. Drop it now.
2624  * queue lock must be held here.
2625  */
2626 static void cfq_put_queue(struct cfq_queue *cfqq)
2627 {
2628         struct cfq_data *cfqd = cfqq->cfqd;
2629         struct cfq_group *cfqg;
2630
2631         BUG_ON(cfqq->ref <= 0);
2632
2633         cfqq->ref--;
2634         if (cfqq->ref)
2635                 return;
2636
2637         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2638         BUG_ON(rb_first(&cfqq->sort_list));
2639         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2640         cfqg = cfqq->cfqg;
2641
2642         if (unlikely(cfqd->active_queue == cfqq)) {
2643                 __cfq_slice_expired(cfqd, cfqq, 0);
2644                 cfq_schedule_dispatch(cfqd);
2645         }
2646
2647         BUG_ON(cfq_cfqq_on_rr(cfqq));
2648         kmem_cache_free(cfq_pool, cfqq);
2649         cfq_put_cfqg(cfqg);
2650 }
2651
2652 /*
2653  * Call func for each cic attached to this ioc.
2654  */
2655 static void
2656 call_for_each_cic(struct io_context *ioc,
2657                   void (*func)(struct io_context *, struct cfq_io_context *))
2658 {
2659         struct cfq_io_context *cic;
2660         struct hlist_node *n;
2661
2662         rcu_read_lock();
2663
2664         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2665                 func(ioc, cic);
2666
2667         rcu_read_unlock();
2668 }
2669
2670 static void cfq_cic_free_rcu(struct rcu_head *head)
2671 {
2672         struct cfq_io_context *cic;
2673
2674         cic = container_of(head, struct cfq_io_context, rcu_head);
2675
2676         kmem_cache_free(cfq_ioc_pool, cic);
2677         elv_ioc_count_dec(cfq_ioc_count);
2678
2679         if (ioc_gone) {
2680                 /*
2681                  * CFQ scheduler is exiting, grab exit lock and check
2682                  * the pending io context count. If it hits zero,
2683                  * complete ioc_gone and set it back to NULL
2684                  */
2685                 spin_lock(&ioc_gone_lock);
2686                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2687                         complete(ioc_gone);
2688                         ioc_gone = NULL;
2689                 }
2690                 spin_unlock(&ioc_gone_lock);
2691         }
2692 }
2693
2694 static void cfq_cic_free(struct cfq_io_context *cic)
2695 {
2696         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2697 }
2698
2699 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2700 {
2701         unsigned long flags;
2702         unsigned long dead_key = (unsigned long) cic->key;
2703
2704         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2705
2706         spin_lock_irqsave(&ioc->lock, flags);
2707         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2708         hlist_del_rcu(&cic->cic_list);
2709         spin_unlock_irqrestore(&ioc->lock, flags);
2710
2711         cfq_cic_free(cic);
2712 }
2713
2714 /*
2715  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2716  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2717  * and ->trim() which is called with the task lock held
2718  */
2719 static void cfq_free_io_context(struct io_context *ioc)
2720 {
2721         /*
2722          * ioc->refcount is zero here, or we are called from elv_unregister(),
2723          * so no more cic's are allowed to be linked into this ioc.  So it
2724          * should be ok to iterate over the known list, we will see all cic's
2725          * since no new ones are added.
2726          */
2727         call_for_each_cic(ioc, cic_free_func);
2728 }
2729
2730 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2731 {
2732         struct cfq_queue *__cfqq, *next;
2733
2734         /*
2735          * If this queue was scheduled to merge with another queue, be
2736          * sure to drop the reference taken on that queue (and others in
2737          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2738          */
2739         __cfqq = cfqq->new_cfqq;
2740         while (__cfqq) {
2741                 if (__cfqq == cfqq) {
2742                         WARN(1, "cfqq->new_cfqq loop detected\n");
2743                         break;
2744                 }
2745                 next = __cfqq->new_cfqq;
2746                 cfq_put_queue(__cfqq);
2747                 __cfqq = next;
2748         }
2749 }
2750
2751 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2752 {
2753         if (unlikely(cfqq == cfqd->active_queue)) {
2754                 __cfq_slice_expired(cfqd, cfqq, 0);
2755                 cfq_schedule_dispatch(cfqd);
2756         }
2757
2758         cfq_put_cooperator(cfqq);
2759
2760         cfq_put_queue(cfqq);
2761 }
2762
2763 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2764                                          struct cfq_io_context *cic)
2765 {
2766         struct io_context *ioc = cic->ioc;
2767
2768         list_del_init(&cic->queue_list);
2769
2770         /*
2771          * Make sure dead mark is seen for dead queues
2772          */
2773         smp_wmb();
2774         cic->key = cfqd_dead_key(cfqd);
2775
2776         rcu_read_lock();
2777         if (rcu_dereference(ioc->ioc_data) == cic) {
2778                 rcu_read_unlock();
2779                 spin_lock(&ioc->lock);
2780                 rcu_assign_pointer(ioc->ioc_data, NULL);
2781                 spin_unlock(&ioc->lock);
2782         } else
2783                 rcu_read_unlock();
2784
2785         if (cic->cfqq[BLK_RW_ASYNC]) {
2786                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2787                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2788         }
2789
2790         if (cic->cfqq[BLK_RW_SYNC]) {
2791                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2792                 cic->cfqq[BLK_RW_SYNC] = NULL;
2793         }
2794 }
2795
2796 static void cfq_exit_single_io_context(struct io_context *ioc,
2797                                        struct cfq_io_context *cic)
2798 {
2799         struct cfq_data *cfqd = cic_to_cfqd(cic);
2800
2801         if (cfqd) {
2802                 struct request_queue *q = cfqd->queue;
2803                 unsigned long flags;
2804
2805                 spin_lock_irqsave(q->queue_lock, flags);
2806
2807                 /*
2808                  * Ensure we get a fresh copy of the ->key to prevent
2809                  * race between exiting task and queue
2810                  */
2811                 smp_read_barrier_depends();
2812                 if (cic->key == cfqd)
2813                         __cfq_exit_single_io_context(cfqd, cic);
2814
2815                 spin_unlock_irqrestore(q->queue_lock, flags);
2816         }
2817 }
2818
2819 /*
2820  * The process that ioc belongs to has exited, we need to clean up
2821  * and put the internal structures we have that belongs to that process.
2822  */
2823 static void cfq_exit_io_context(struct io_context *ioc)
2824 {
2825         call_for_each_cic(ioc, cfq_exit_single_io_context);
2826 }
2827
2828 static struct cfq_io_context *
2829 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2830 {
2831         struct cfq_io_context *cic;
2832
2833         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2834                                                         cfqd->queue->node);
2835         if (cic) {
2836                 cic->ttime.last_end_request = jiffies;
2837                 INIT_LIST_HEAD(&cic->queue_list);
2838                 INIT_HLIST_NODE(&cic->cic_list);
2839                 cic->dtor = cfq_free_io_context;
2840                 cic->exit = cfq_exit_io_context;
2841                 elv_ioc_count_inc(cfq_ioc_count);
2842         }
2843
2844         return cic;
2845 }
2846
2847 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2848 {
2849         struct task_struct *tsk = current;
2850         int ioprio_class;
2851
2852         if (!cfq_cfqq_prio_changed(cfqq))
2853                 return;
2854
2855         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2856         switch (ioprio_class) {
2857         default:
2858                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2859         case IOPRIO_CLASS_NONE:
2860                 /*
2861                  * no prio set, inherit CPU scheduling settings
2862                  */
2863                 cfqq->ioprio = task_nice_ioprio(tsk);
2864                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2865                 break;
2866         case IOPRIO_CLASS_RT:
2867                 cfqq->ioprio = task_ioprio(ioc);
2868                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2869                 break;
2870         case IOPRIO_CLASS_BE:
2871                 cfqq->ioprio = task_ioprio(ioc);
2872                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2873                 break;
2874         case IOPRIO_CLASS_IDLE:
2875                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2876                 cfqq->ioprio = 7;
2877                 cfq_clear_cfqq_idle_window(cfqq);
2878                 break;
2879         }
2880
2881         /*
2882          * keep track of original prio settings in case we have to temporarily
2883          * elevate the priority of this queue
2884          */
2885         cfqq->org_ioprio = cfqq->ioprio;
2886         cfq_clear_cfqq_prio_changed(cfqq);
2887 }
2888
2889 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2890 {
2891         struct cfq_data *cfqd = cic_to_cfqd(cic);
2892         struct cfq_queue *cfqq;
2893         unsigned long flags;
2894
2895         if (unlikely(!cfqd))
2896                 return;
2897
2898         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2899
2900         cfqq = cic->cfqq[BLK_RW_ASYNC];
2901         if (cfqq) {
2902                 struct cfq_queue *new_cfqq;
2903                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2904                                                 GFP_ATOMIC);
2905                 if (new_cfqq) {
2906                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2907                         cfq_put_queue(cfqq);
2908                 }
2909         }
2910
2911         cfqq = cic->cfqq[BLK_RW_SYNC];
2912         if (cfqq)
2913                 cfq_mark_cfqq_prio_changed(cfqq);
2914
2915         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2916 }
2917
2918 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2919 {
2920         call_for_each_cic(ioc, changed_ioprio);
2921         ioc->ioprio_changed = 0;
2922 }
2923
2924 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2925                           pid_t pid, bool is_sync)
2926 {
2927         RB_CLEAR_NODE(&cfqq->rb_node);
2928         RB_CLEAR_NODE(&cfqq->p_node);
2929         INIT_LIST_HEAD(&cfqq->fifo);
2930
2931         cfqq->ref = 0;
2932         cfqq->cfqd = cfqd;
2933
2934         cfq_mark_cfqq_prio_changed(cfqq);
2935
2936         if (is_sync) {
2937                 if (!cfq_class_idle(cfqq))
2938                         cfq_mark_cfqq_idle_window(cfqq);
2939                 cfq_mark_cfqq_sync(cfqq);
2940         }
2941         cfqq->pid = pid;
2942 }
2943
2944 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2945 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2946 {
2947         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2948         struct cfq_data *cfqd = cic_to_cfqd(cic);
2949         unsigned long flags;
2950         struct request_queue *q;
2951
2952         if (unlikely(!cfqd))
2953                 return;
2954
2955         q = cfqd->queue;
2956
2957         spin_lock_irqsave(q->queue_lock, flags);
2958
2959         if (sync_cfqq) {
2960                 /*
2961                  * Drop reference to sync queue. A new sync queue will be
2962                  * assigned in new group upon arrival of a fresh request.
2963                  */
2964                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2965                 cic_set_cfqq(cic, NULL, 1);
2966                 cfq_put_queue(sync_cfqq);
2967         }
2968
2969         spin_unlock_irqrestore(q->queue_lock, flags);
2970 }
2971
2972 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2973 {
2974         call_for_each_cic(ioc, changed_cgroup);
2975         ioc->cgroup_changed = 0;
2976 }
2977 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2978
2979 static struct cfq_queue *
2980 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2981                      struct io_context *ioc, gfp_t gfp_mask)
2982 {
2983         struct cfq_queue *cfqq, *new_cfqq = NULL;
2984         struct cfq_io_context *cic;
2985         struct cfq_group *cfqg;
2986
2987 retry:
2988         cfqg = cfq_get_cfqg(cfqd);
2989         cic = cfq_cic_lookup(cfqd, ioc);
2990         /* cic always exists here */
2991         cfqq = cic_to_cfqq(cic, is_sync);
2992
2993         /*
2994          * Always try a new alloc if we fell back to the OOM cfqq
2995          * originally, since it should just be a temporary situation.
2996          */
2997         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2998                 cfqq = NULL;
2999                 if (new_cfqq) {
3000                         cfqq = new_cfqq;
3001                         new_cfqq = NULL;
3002                 } else if (gfp_mask & __GFP_WAIT) {
3003                         spin_unlock_irq(cfqd->queue->queue_lock);
3004                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3005                                         gfp_mask | __GFP_ZERO,
3006                                         cfqd->queue->node);
3007                         spin_lock_irq(cfqd->queue->queue_lock);
3008                         if (new_cfqq)
3009                                 goto retry;
3010                 } else {
3011                         cfqq = kmem_cache_alloc_node(cfq_pool,
3012                                         gfp_mask | __GFP_ZERO,
3013                                         cfqd->queue->node);
3014                 }
3015
3016                 if (cfqq) {
3017                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3018                         cfq_init_prio_data(cfqq, ioc);
3019                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3020                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3021                 } else
3022                         cfqq = &cfqd->oom_cfqq;
3023         }
3024
3025         if (new_cfqq)
3026                 kmem_cache_free(cfq_pool, new_cfqq);
3027
3028         return cfqq;
3029 }
3030
3031 static struct cfq_queue **
3032 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3033 {
3034         switch (ioprio_class) {
3035         case IOPRIO_CLASS_RT:
3036                 return &cfqd->async_cfqq[0][ioprio];
3037         case IOPRIO_CLASS_BE:
3038                 return &cfqd->async_cfqq[1][ioprio];
3039         case IOPRIO_CLASS_IDLE:
3040                 return &cfqd->async_idle_cfqq;
3041         default:
3042                 BUG();
3043         }
3044 }
3045
3046 static struct cfq_queue *
3047 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3048               gfp_t gfp_mask)
3049 {
3050         const int ioprio = task_ioprio(ioc);
3051         const int ioprio_class = task_ioprio_class(ioc);
3052         struct cfq_queue **async_cfqq = NULL;
3053         struct cfq_queue *cfqq = NULL;
3054
3055         if (!is_sync) {
3056                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3057                 cfqq = *async_cfqq;
3058         }
3059
3060         if (!cfqq)
3061                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3062
3063         /*
3064          * pin the queue now that it's allocated, scheduler exit will prune it
3065          */
3066         if (!is_sync && !(*async_cfqq)) {
3067                 cfqq->ref++;
3068                 *async_cfqq = cfqq;
3069         }
3070
3071         cfqq->ref++;
3072         return cfqq;
3073 }
3074
3075 /*
3076  * We drop cfq io contexts lazily, so we may find a dead one.
3077  */
3078 static void
3079 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3080                   struct cfq_io_context *cic)
3081 {
3082         unsigned long flags;
3083
3084         WARN_ON(!list_empty(&cic->queue_list));
3085         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3086
3087         spin_lock_irqsave(&ioc->lock, flags);
3088
3089         BUG_ON(rcu_dereference_check(ioc->ioc_data,
3090                 lockdep_is_held(&ioc->lock)) == cic);
3091
3092         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3093         hlist_del_rcu(&cic->cic_list);
3094         spin_unlock_irqrestore(&ioc->lock, flags);
3095
3096         cfq_cic_free(cic);
3097 }
3098
3099 static struct cfq_io_context *
3100 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3101 {
3102         struct cfq_io_context *cic;
3103         unsigned long flags;
3104
3105         if (unlikely(!ioc))
3106                 return NULL;
3107
3108         rcu_read_lock();
3109
3110         /*
3111          * we maintain a last-hit cache, to avoid browsing over the tree
3112          */
3113         cic = rcu_dereference(ioc->ioc_data);
3114         if (cic && cic->key == cfqd) {
3115                 rcu_read_unlock();
3116                 return cic;
3117         }
3118
3119         do {
3120                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3121                 rcu_read_unlock();
3122                 if (!cic)
3123                         break;
3124                 if (unlikely(cic->key != cfqd)) {
3125                         cfq_drop_dead_cic(cfqd, ioc, cic);
3126                         rcu_read_lock();
3127                         continue;
3128                 }
3129
3130                 spin_lock_irqsave(&ioc->lock, flags);
3131                 rcu_assign_pointer(ioc->ioc_data, cic);
3132                 spin_unlock_irqrestore(&ioc->lock, flags);
3133                 break;
3134         } while (1);
3135
3136         return cic;
3137 }
3138
3139 /*
3140  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3141  * the process specific cfq io context when entered from the block layer.
3142  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3143  */
3144 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3145                         struct cfq_io_context *cic, gfp_t gfp_mask)
3146 {
3147         unsigned long flags;
3148         int ret;
3149
3150         ret = radix_tree_preload(gfp_mask);
3151         if (!ret) {
3152                 cic->ioc = ioc;
3153                 cic->key = cfqd;
3154
3155                 spin_lock_irqsave(&ioc->lock, flags);
3156                 ret = radix_tree_insert(&ioc->radix_root,
3157                                                 cfqd->cic_index, cic);
3158                 if (!ret)
3159                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3160                 spin_unlock_irqrestore(&ioc->lock, flags);
3161
3162                 radix_tree_preload_end();
3163
3164                 if (!ret) {
3165                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3166                         list_add(&cic->queue_list, &cfqd->cic_list);
3167                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3168                 }
3169         }
3170
3171         if (ret)
3172                 printk(KERN_ERR "cfq: cic link failed!\n");
3173
3174         return ret;
3175 }
3176
3177 /*
3178  * Setup general io context and cfq io context. There can be several cfq
3179  * io contexts per general io context, if this process is doing io to more
3180  * than one device managed by cfq.
3181  */
3182 static struct cfq_io_context *
3183 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3184 {
3185         struct io_context *ioc = NULL;
3186         struct cfq_io_context *cic;
3187
3188         might_sleep_if(gfp_mask & __GFP_WAIT);
3189
3190         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3191         if (!ioc)
3192                 return NULL;
3193
3194         cic = cfq_cic_lookup(cfqd, ioc);
3195         if (cic)
3196                 goto out;
3197
3198         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3199         if (cic == NULL)
3200                 goto err;
3201
3202         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3203                 goto err_free;
3204
3205 out:
3206         smp_read_barrier_depends();
3207         if (unlikely(ioc->ioprio_changed))
3208                 cfq_ioc_set_ioprio(ioc);
3209
3210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3211         if (unlikely(ioc->cgroup_changed))
3212                 cfq_ioc_set_cgroup(ioc);
3213 #endif
3214         return cic;
3215 err_free:
3216         cfq_cic_free(cic);
3217 err:
3218         put_io_context(ioc);
3219         return NULL;
3220 }
3221
3222 static void
3223 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3224 {
3225         unsigned long elapsed = jiffies - ttime->last_end_request;
3226         elapsed = min(elapsed, 2UL * slice_idle);
3227
3228         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3229         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3230         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3231 }
3232
3233 static void
3234 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3235         struct cfq_io_context *cic)
3236 {
3237         if (cfq_cfqq_sync(cfqq)) {
3238                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3239                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3240                         cfqd->cfq_slice_idle);
3241         }
3242 }
3243
3244 static void
3245 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3246                        struct request *rq)
3247 {
3248         sector_t sdist = 0;
3249         sector_t n_sec = blk_rq_sectors(rq);
3250         if (cfqq->last_request_pos) {
3251                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3252                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3253                 else
3254                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3255         }
3256
3257         cfqq->seek_history <<= 1;
3258         if (blk_queue_nonrot(cfqd->queue))
3259                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3260         else
3261                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3262 }
3263
3264 /*
3265  * Disable idle window if the process thinks too long or seeks so much that
3266  * it doesn't matter
3267  */
3268 static void
3269 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3270                        struct cfq_io_context *cic)
3271 {
3272         int old_idle, enable_idle;
3273
3274         /*
3275          * Don't idle for async or idle io prio class
3276          */
3277         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3278                 return;
3279
3280         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3281
3282         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3283                 cfq_mark_cfqq_deep(cfqq);
3284
3285         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3286                 enable_idle = 0;
3287         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3288             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3289                 enable_idle = 0;
3290         else if (sample_valid(cic->ttime.ttime_samples)) {
3291                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3292                         enable_idle = 0;
3293                 else
3294                         enable_idle = 1;
3295         }
3296
3297         if (old_idle != enable_idle) {
3298                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3299                 if (enable_idle)
3300                         cfq_mark_cfqq_idle_window(cfqq);
3301                 else
3302                         cfq_clear_cfqq_idle_window(cfqq);
3303         }
3304 }
3305
3306 /*
3307  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3308  * no or if we aren't sure, a 1 will cause a preempt.
3309  */
3310 static bool
3311 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3312                    struct request *rq)
3313 {
3314         struct cfq_queue *cfqq;
3315
3316         cfqq = cfqd->active_queue;
3317         if (!cfqq)
3318                 return false;
3319
3320         if (cfq_class_idle(new_cfqq))
3321                 return false;
3322
3323         if (cfq_class_idle(cfqq))
3324                 return true;
3325
3326         /*
3327          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3328          */
3329         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3330                 return false;
3331
3332         /*
3333          * if the new request is sync, but the currently running queue is
3334          * not, let the sync request have priority.
3335          */
3336         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3337                 return true;
3338
3339         if (new_cfqq->cfqg != cfqq->cfqg)
3340                 return false;
3341
3342         if (cfq_slice_used(cfqq))
3343                 return true;
3344
3345         /* Allow preemption only if we are idling on sync-noidle tree */
3346         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3347             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3348             new_cfqq->service_tree->count == 2 &&
3349             RB_EMPTY_ROOT(&cfqq->sort_list))
3350                 return true;
3351
3352         /*
3353          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3354          */
3355         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3356                 return true;
3357
3358         /* An idle queue should not be idle now for some reason */
3359         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3360                 return true;
3361
3362         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3363                 return false;
3364
3365         /*
3366          * if this request is as-good as one we would expect from the
3367          * current cfqq, let it preempt
3368          */
3369         if (cfq_rq_close(cfqd, cfqq, rq))
3370                 return true;
3371
3372         return false;
3373 }
3374
3375 /*
3376  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3377  * let it have half of its nominal slice.
3378  */
3379 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3380 {
3381         struct cfq_queue *old_cfqq = cfqd->active_queue;
3382
3383         cfq_log_cfqq(cfqd, cfqq, "preempt");
3384         cfq_slice_expired(cfqd, 1);
3385
3386         /*
3387          * workload type is changed, don't save slice, otherwise preempt
3388          * doesn't happen
3389          */
3390         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3391                 cfqq->cfqg->saved_workload_slice = 0;
3392
3393         /*
3394          * Put the new queue at the front of the of the current list,
3395          * so we know that it will be selected next.
3396          */
3397         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3398
3399         cfq_service_tree_add(cfqd, cfqq, 1);
3400
3401         cfqq->slice_end = 0;
3402         cfq_mark_cfqq_slice_new(cfqq);
3403 }
3404
3405 /*
3406  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3407  * something we should do about it
3408  */
3409 static void
3410 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3411                 struct request *rq)
3412 {
3413         struct cfq_io_context *cic = RQ_CIC(rq);
3414
3415         cfqd->rq_queued++;
3416
3417         cfq_update_io_thinktime(cfqd, cfqq, cic);
3418         cfq_update_io_seektime(cfqd, cfqq, rq);
3419         cfq_update_idle_window(cfqd, cfqq, cic);
3420
3421         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3422
3423         if (cfqq == cfqd->active_queue) {
3424                 /*
3425                  * Remember that we saw a request from this process, but
3426                  * don't start queuing just yet. Otherwise we risk seeing lots
3427                  * of tiny requests, because we disrupt the normal plugging
3428                  * and merging. If the request is already larger than a single
3429                  * page, let it rip immediately. For that case we assume that
3430                  * merging is already done. Ditto for a busy system that
3431                  * has other work pending, don't risk delaying until the
3432                  * idle timer unplug to continue working.
3433                  */
3434                 if (cfq_cfqq_wait_request(cfqq)) {
3435                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3436                             cfqd->busy_queues > 1) {
3437                                 cfq_del_timer(cfqd, cfqq);
3438                                 cfq_clear_cfqq_wait_request(cfqq);
3439                                 __blk_run_queue(cfqd->queue);
3440                         } else {
3441                                 cfq_blkiocg_update_idle_time_stats(
3442                                                 &cfqq->cfqg->blkg);
3443                                 cfq_mark_cfqq_must_dispatch(cfqq);
3444                         }
3445                 }
3446         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3447                 /*
3448                  * not the active queue - expire current slice if it is
3449                  * idle and has expired it's mean thinktime or this new queue
3450                  * has some old slice time left and is of higher priority or
3451                  * this new queue is RT and the current one is BE
3452                  */
3453                 cfq_preempt_queue(cfqd, cfqq);
3454                 __blk_run_queue(cfqd->queue);
3455         }
3456 }
3457
3458 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3459 {
3460         struct cfq_data *cfqd = q->elevator->elevator_data;
3461         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3462
3463         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3464         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3465
3466         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3467         list_add_tail(&rq->queuelist, &cfqq->fifo);
3468         cfq_add_rq_rb(rq);
3469         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3470                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3471                         rq_is_sync(rq));
3472         cfq_rq_enqueued(cfqd, cfqq, rq);
3473 }
3474
3475 /*
3476  * Update hw_tag based on peak queue depth over 50 samples under
3477  * sufficient load.
3478  */
3479 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3480 {
3481         struct cfq_queue *cfqq = cfqd->active_queue;
3482
3483         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3484                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3485
3486         if (cfqd->hw_tag == 1)
3487                 return;
3488
3489         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3490             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3491                 return;
3492
3493         /*
3494          * If active queue hasn't enough requests and can idle, cfq might not
3495          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3496          * case
3497          */
3498         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3499             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3500             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3501                 return;
3502
3503         if (cfqd->hw_tag_samples++ < 50)
3504                 return;
3505
3506         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3507                 cfqd->hw_tag = 1;
3508         else
3509                 cfqd->hw_tag = 0;
3510 }
3511
3512 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3513 {
3514         struct cfq_io_context *cic = cfqd->active_cic;
3515
3516         /* If the queue already has requests, don't wait */
3517         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3518                 return false;
3519
3520         /* If there are other queues in the group, don't wait */
3521         if (cfqq->cfqg->nr_cfqq > 1)
3522                 return false;
3523
3524         if (cfq_slice_used(cfqq))
3525                 return true;
3526
3527         /* if slice left is less than think time, wait busy */
3528         if (cic && sample_valid(cic->ttime.ttime_samples)
3529             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3530                 return true;
3531
3532         /*
3533          * If think times is less than a jiffy than ttime_mean=0 and above
3534          * will not be true. It might happen that slice has not expired yet
3535          * but will expire soon (4-5 ns) during select_queue(). To cover the
3536          * case where think time is less than a jiffy, mark the queue wait
3537          * busy if only 1 jiffy is left in the slice.
3538          */
3539         if (cfqq->slice_end - jiffies == 1)
3540                 return true;
3541
3542         return false;
3543 }
3544
3545 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3546 {
3547         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3548         struct cfq_data *cfqd = cfqq->cfqd;
3549         const int sync = rq_is_sync(rq);
3550         unsigned long now;
3551
3552         now = jiffies;
3553         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3554                      !!(rq->cmd_flags & REQ_NOIDLE));
3555
3556         cfq_update_hw_tag(cfqd);
3557
3558         WARN_ON(!cfqd->rq_in_driver);
3559         WARN_ON(!cfqq->dispatched);
3560         cfqd->rq_in_driver--;
3561         cfqq->dispatched--;
3562         (RQ_CFQG(rq))->dispatched--;
3563         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3564                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3565                         rq_data_dir(rq), rq_is_sync(rq));
3566
3567         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3568
3569         if (sync) {
3570                 struct cfq_rb_root *service_tree;
3571
3572                 RQ_CIC(rq)->ttime.last_end_request = now;
3573
3574                 if (cfq_cfqq_on_rr(cfqq))
3575                         service_tree = cfqq->service_tree;
3576                 else
3577                         service_tree = service_tree_for(cfqq->cfqg,
3578                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3579                 service_tree->ttime.last_end_request = now;
3580                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3581                         cfqd->last_delayed_sync = now;
3582         }
3583
3584         /*
3585          * If this is the active queue, check if it needs to be expired,
3586          * or if we want to idle in case it has no pending requests.
3587          */
3588         if (cfqd->active_queue == cfqq) {
3589                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3590
3591                 if (cfq_cfqq_slice_new(cfqq)) {
3592                         cfq_set_prio_slice(cfqd, cfqq);
3593                         cfq_clear_cfqq_slice_new(cfqq);
3594                 }
3595
3596                 /*
3597                  * Should we wait for next request to come in before we expire
3598                  * the queue.
3599                  */
3600                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3601                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3602                         if (!cfqd->cfq_slice_idle)
3603                                 extend_sl = cfqd->cfq_group_idle;
3604                         cfqq->slice_end = jiffies + extend_sl;
3605                         cfq_mark_cfqq_wait_busy(cfqq);
3606                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3607                 }
3608
3609                 /*
3610                  * Idling is not enabled on:
3611                  * - expired queues
3612                  * - idle-priority queues
3613                  * - async queues
3614                  * - queues with still some requests queued
3615                  * - when there is a close cooperator
3616                  */
3617                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3618                         cfq_slice_expired(cfqd, 1);
3619                 else if (sync && cfqq_empty &&
3620                          !cfq_close_cooperator(cfqd, cfqq)) {
3621                         cfq_arm_slice_timer(cfqd);
3622                 }
3623         }
3624
3625         if (!cfqd->rq_in_driver)
3626                 cfq_schedule_dispatch(cfqd);
3627 }
3628
3629 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3630 {
3631         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3632                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3633                 return ELV_MQUEUE_MUST;
3634         }
3635
3636         return ELV_MQUEUE_MAY;
3637 }
3638
3639 static int cfq_may_queue(struct request_queue *q, int rw)
3640 {
3641         struct cfq_data *cfqd = q->elevator->elevator_data;
3642         struct task_struct *tsk = current;
3643         struct cfq_io_context *cic;
3644         struct cfq_queue *cfqq;
3645
3646         /*
3647          * don't force setup of a queue from here, as a call to may_queue
3648          * does not necessarily imply that a request actually will be queued.
3649          * so just lookup a possibly existing queue, or return 'may queue'
3650          * if that fails
3651          */
3652         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3653         if (!cic)
3654                 return ELV_MQUEUE_MAY;
3655
3656         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3657         if (cfqq) {
3658                 cfq_init_prio_data(cfqq, cic->ioc);
3659
3660                 return __cfq_may_queue(cfqq);
3661         }
3662
3663         return ELV_MQUEUE_MAY;
3664 }
3665
3666 /*
3667  * queue lock held here
3668  */
3669 static void cfq_put_request(struct request *rq)
3670 {
3671         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3672
3673         if (cfqq) {
3674                 const int rw = rq_data_dir(rq);
3675
3676                 BUG_ON(!cfqq->allocated[rw]);
3677                 cfqq->allocated[rw]--;
3678
3679                 put_io_context(RQ_CIC(rq)->ioc);
3680
3681                 rq->elevator_private[0] = NULL;
3682                 rq->elevator_private[1] = NULL;
3683
3684                 /* Put down rq reference on cfqg */
3685                 cfq_put_cfqg(RQ_CFQG(rq));
3686                 rq->elevator_private[2] = NULL;
3687
3688                 cfq_put_queue(cfqq);
3689         }
3690 }
3691
3692 static struct cfq_queue *
3693 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3694                 struct cfq_queue *cfqq)
3695 {
3696         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3697         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3698         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3699         cfq_put_queue(cfqq);
3700         return cic_to_cfqq(cic, 1);
3701 }
3702
3703 /*
3704  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3705  * was the last process referring to said cfqq.
3706  */
3707 static struct cfq_queue *
3708 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3709 {
3710         if (cfqq_process_refs(cfqq) == 1) {
3711                 cfqq->pid = current->pid;
3712                 cfq_clear_cfqq_coop(cfqq);
3713                 cfq_clear_cfqq_split_coop(cfqq);
3714                 return cfqq;
3715         }
3716
3717         cic_set_cfqq(cic, NULL, 1);
3718
3719         cfq_put_cooperator(cfqq);
3720
3721         cfq_put_queue(cfqq);
3722         return NULL;
3723 }
3724 /*
3725  * Allocate cfq data structures associated with this request.
3726  */
3727 static int
3728 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3729 {
3730         struct cfq_data *cfqd = q->elevator->elevator_data;
3731         struct cfq_io_context *cic;
3732         const int rw = rq_data_dir(rq);
3733         const bool is_sync = rq_is_sync(rq);
3734         struct cfq_queue *cfqq;
3735         unsigned long flags;
3736
3737         might_sleep_if(gfp_mask & __GFP_WAIT);
3738
3739         cic = cfq_get_io_context(cfqd, gfp_mask);
3740
3741         spin_lock_irqsave(q->queue_lock, flags);
3742
3743         if (!cic)
3744                 goto queue_fail;
3745
3746 new_queue:
3747         cfqq = cic_to_cfqq(cic, is_sync);
3748         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3749                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3750                 cic_set_cfqq(cic, cfqq, is_sync);
3751         } else {
3752                 /*
3753                  * If the queue was seeky for too long, break it apart.
3754                  */
3755                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3756                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3757                         cfqq = split_cfqq(cic, cfqq);
3758                         if (!cfqq)
3759                                 goto new_queue;
3760                 }
3761
3762                 /*
3763                  * Check to see if this queue is scheduled to merge with
3764                  * another, closely cooperating queue.  The merging of
3765                  * queues happens here as it must be done in process context.
3766                  * The reference on new_cfqq was taken in merge_cfqqs.
3767                  */
3768                 if (cfqq->new_cfqq)
3769                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3770         }
3771
3772         cfqq->allocated[rw]++;
3773
3774         cfqq->ref++;
3775         rq->elevator_private[0] = cic;
3776         rq->elevator_private[1] = cfqq;
3777         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3778         spin_unlock_irqrestore(q->queue_lock, flags);
3779         return 0;
3780
3781 queue_fail:
3782         cfq_schedule_dispatch(cfqd);
3783         spin_unlock_irqrestore(q->queue_lock, flags);
3784         cfq_log(cfqd, "set_request fail");
3785         return 1;
3786 }
3787
3788 static void cfq_kick_queue(struct work_struct *work)
3789 {
3790         struct cfq_data *cfqd =
3791                 container_of(work, struct cfq_data, unplug_work);
3792         struct request_queue *q = cfqd->queue;
3793
3794         spin_lock_irq(q->queue_lock);
3795         __blk_run_queue(cfqd->queue);
3796         spin_unlock_irq(q->queue_lock);
3797 }
3798
3799 /*
3800  * Timer running if the active_queue is currently idling inside its time slice
3801  */
3802 static void cfq_idle_slice_timer(unsigned long data)
3803 {
3804         struct cfq_data *cfqd = (struct cfq_data *) data;
3805         struct cfq_queue *cfqq;
3806         unsigned long flags;
3807         int timed_out = 1;
3808
3809         cfq_log(cfqd, "idle timer fired");
3810
3811         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3812
3813         cfqq = cfqd->active_queue;
3814         if (cfqq) {
3815                 timed_out = 0;
3816
3817                 /*
3818                  * We saw a request before the queue expired, let it through
3819                  */
3820                 if (cfq_cfqq_must_dispatch(cfqq))
3821                         goto out_kick;
3822
3823                 /*
3824                  * expired
3825                  */
3826                 if (cfq_slice_used(cfqq))
3827                         goto expire;
3828
3829                 /*
3830                  * only expire and reinvoke request handler, if there are
3831                  * other queues with pending requests
3832                  */
3833                 if (!cfqd->busy_queues)
3834                         goto out_cont;
3835
3836                 /*
3837                  * not expired and it has a request pending, let it dispatch
3838                  */
3839                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3840                         goto out_kick;
3841
3842                 /*
3843                  * Queue depth flag is reset only when the idle didn't succeed
3844                  */
3845                 cfq_clear_cfqq_deep(cfqq);
3846         }
3847 expire:
3848         cfq_slice_expired(cfqd, timed_out);
3849 out_kick:
3850         cfq_schedule_dispatch(cfqd);
3851 out_cont:
3852         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3853 }
3854
3855 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3856 {
3857         del_timer_sync(&cfqd->idle_slice_timer);
3858         cancel_work_sync(&cfqd->unplug_work);
3859 }
3860
3861 static void cfq_put_async_queues(struct cfq_data *cfqd)
3862 {
3863         int i;
3864
3865         for (i = 0; i < IOPRIO_BE_NR; i++) {
3866                 if (cfqd->async_cfqq[0][i])
3867                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3868                 if (cfqd->async_cfqq[1][i])
3869                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3870         }
3871
3872         if (cfqd->async_idle_cfqq)
3873                 cfq_put_queue(cfqd->async_idle_cfqq);
3874 }
3875
3876 static void cfq_exit_queue(struct elevator_queue *e)
3877 {
3878         struct cfq_data *cfqd = e->elevator_data;
3879         struct request_queue *q = cfqd->queue;
3880         bool wait = false;
3881
3882         cfq_shutdown_timer_wq(cfqd);
3883
3884         spin_lock_irq(q->queue_lock);
3885
3886         if (cfqd->active_queue)
3887                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3888
3889         while (!list_empty(&cfqd->cic_list)) {
3890                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3891                                                         struct cfq_io_context,
3892                                                         queue_list);
3893
3894                 __cfq_exit_single_io_context(cfqd, cic);
3895         }
3896
3897         cfq_put_async_queues(cfqd);
3898         cfq_release_cfq_groups(cfqd);
3899
3900         /*
3901          * If there are groups which we could not unlink from blkcg list,
3902          * wait for a rcu period for them to be freed.
3903          */
3904         if (cfqd->nr_blkcg_linked_grps)
3905                 wait = true;
3906
3907         spin_unlock_irq(q->queue_lock);
3908
3909         cfq_shutdown_timer_wq(cfqd);
3910
3911         spin_lock(&cic_index_lock);
3912         ida_remove(&cic_index_ida, cfqd->cic_index);
3913         spin_unlock(&cic_index_lock);
3914
3915         /*
3916          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3917          * Do this wait only if there are other unlinked groups out
3918          * there. This can happen if cgroup deletion path claimed the
3919          * responsibility of cleaning up a group before queue cleanup code
3920          * get to the group.
3921          *
3922          * Do not call synchronize_rcu() unconditionally as there are drivers
3923          * which create/delete request queue hundreds of times during scan/boot
3924          * and synchronize_rcu() can take significant time and slow down boot.
3925          */
3926         if (wait)
3927                 synchronize_rcu();
3928
3929 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3930         /* Free up per cpu stats for root group */
3931         free_percpu(cfqd->root_group.blkg.stats_cpu);
3932 #endif
3933         kfree(cfqd);
3934 }
3935
3936 static int cfq_alloc_cic_index(void)
3937 {
3938         int index, error;
3939
3940         do {
3941                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3942                         return -ENOMEM;
3943
3944                 spin_lock(&cic_index_lock);
3945                 error = ida_get_new(&cic_index_ida, &index);
3946                 spin_unlock(&cic_index_lock);
3947                 if (error && error != -EAGAIN)
3948                         return error;
3949         } while (error);
3950
3951         return index;
3952 }
3953
3954 static void *cfq_init_queue(struct request_queue *q)
3955 {
3956         struct cfq_data *cfqd;
3957         int i, j;
3958         struct cfq_group *cfqg;
3959         struct cfq_rb_root *st;
3960
3961         i = cfq_alloc_cic_index();
3962         if (i < 0)
3963                 return NULL;
3964
3965         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3966         if (!cfqd) {
3967                 spin_lock(&cic_index_lock);
3968                 ida_remove(&cic_index_ida, i);
3969                 spin_unlock(&cic_index_lock);
3970                 return NULL;
3971         }
3972
3973         /*
3974          * Don't need take queue_lock in the routine, since we are
3975          * initializing the ioscheduler, and nobody is using cfqd
3976          */
3977         cfqd->cic_index = i;
3978
3979         /* Init root service tree */
3980         cfqd->grp_service_tree = CFQ_RB_ROOT;
3981
3982         /* Init root group */
3983         cfqg = &cfqd->root_group;
3984         for_each_cfqg_st(cfqg, i, j, st)
3985                 *st = CFQ_RB_ROOT;
3986         RB_CLEAR_NODE(&cfqg->rb_node);
3987
3988         /* Give preference to root group over other groups */
3989         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3990
3991 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3992         /*
3993          * Set root group reference to 2. One reference will be dropped when
3994          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3995          * Other reference will remain there as we don't want to delete this
3996          * group as it is statically allocated and gets destroyed when
3997          * throtl_data goes away.
3998          */
3999         cfqg->ref = 2;
4000
4001         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4002                 kfree(cfqg);
4003                 kfree(cfqd);
4004                 return NULL;
4005         }
4006
4007         rcu_read_lock();
4008
4009         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4010                                         (void *)cfqd, 0);
4011         rcu_read_unlock();
4012         cfqd->nr_blkcg_linked_grps++;
4013
4014         /* Add group on cfqd->cfqg_list */
4015         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4016 #endif
4017         /*
4018          * Not strictly needed (since RB_ROOT just clears the node and we
4019          * zeroed cfqd on alloc), but better be safe in case someone decides
4020          * to add magic to the rb code
4021          */
4022         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4023                 cfqd->prio_trees[i] = RB_ROOT;
4024
4025         /*
4026          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4027          * Grab a permanent reference to it, so that the normal code flow
4028          * will not attempt to free it.
4029          */
4030         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4031         cfqd->oom_cfqq.ref++;
4032         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4033
4034         INIT_LIST_HEAD(&cfqd->cic_list);
4035
4036         cfqd->queue = q;
4037
4038         init_timer(&cfqd->idle_slice_timer);
4039         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4040         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4041
4042         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4043
4044         cfqd->cfq_quantum = cfq_quantum;
4045         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4046         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4047         cfqd->cfq_back_max = cfq_back_max;
4048         cfqd->cfq_back_penalty = cfq_back_penalty;
4049         cfqd->cfq_slice[0] = cfq_slice_async;
4050         cfqd->cfq_slice[1] = cfq_slice_sync;
4051         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4052         cfqd->cfq_slice_idle = cfq_slice_idle;
4053         cfqd->cfq_group_idle = cfq_group_idle;
4054         cfqd->cfq_latency = 1;
4055         cfqd->hw_tag = -1;
4056         /*
4057          * we optimistically start assuming sync ops weren't delayed in last
4058          * second, in order to have larger depth for async operations.
4059          */
4060         cfqd->last_delayed_sync = jiffies - HZ;
4061         return cfqd;
4062 }
4063
4064 static void cfq_slab_kill(void)
4065 {
4066         /*
4067          * Caller already ensured that pending RCU callbacks are completed,
4068          * so we should have no busy allocations at this point.
4069          */
4070         if (cfq_pool)
4071                 kmem_cache_destroy(cfq_pool);
4072         if (cfq_ioc_pool)
4073                 kmem_cache_destroy(cfq_ioc_pool);
4074 }
4075
4076 static int __init cfq_slab_setup(void)
4077 {
4078         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4079         if (!cfq_pool)
4080                 goto fail;
4081
4082         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4083         if (!cfq_ioc_pool)
4084                 goto fail;
4085
4086         return 0;
4087 fail:
4088         cfq_slab_kill();
4089         return -ENOMEM;
4090 }
4091
4092 /*
4093  * sysfs parts below -->
4094  */
4095 static ssize_t
4096 cfq_var_show(unsigned int var, char *page)
4097 {
4098         return sprintf(page, "%d\n", var);
4099 }
4100
4101 static ssize_t
4102 cfq_var_store(unsigned int *var, const char *page, size_t count)
4103 {
4104         char *p = (char *) page;
4105
4106         *var = simple_strtoul(p, &p, 10);
4107         return count;
4108 }
4109
4110 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4111 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4112 {                                                                       \
4113         struct cfq_data *cfqd = e->elevator_data;                       \
4114         unsigned int __data = __VAR;                                    \
4115         if (__CONV)                                                     \
4116                 __data = jiffies_to_msecs(__data);                      \
4117         return cfq_var_show(__data, (page));                            \
4118 }
4119 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4120 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4121 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4122 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4123 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4124 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4125 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4126 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4127 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4128 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4129 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4130 #undef SHOW_FUNCTION
4131
4132 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4133 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4134 {                                                                       \
4135         struct cfq_data *cfqd = e->elevator_data;                       \
4136         unsigned int __data;                                            \
4137         int ret = cfq_var_store(&__data, (page), count);                \
4138         if (__data < (MIN))                                             \
4139                 __data = (MIN);                                         \
4140         else if (__data > (MAX))                                        \
4141                 __data = (MAX);                                         \
4142         if (__CONV)                                                     \
4143                 *(__PTR) = msecs_to_jiffies(__data);                    \
4144         else                                                            \
4145                 *(__PTR) = __data;                                      \
4146         return ret;                                                     \
4147 }
4148 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4149 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4150                 UINT_MAX, 1);
4151 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4152                 UINT_MAX, 1);
4153 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4154 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4155                 UINT_MAX, 0);
4156 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4157 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4158 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4159 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4160 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4161                 UINT_MAX, 0);
4162 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4163 #undef STORE_FUNCTION
4164
4165 #define CFQ_ATTR(name) \
4166         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4167
4168 static struct elv_fs_entry cfq_attrs[] = {
4169         CFQ_ATTR(quantum),
4170         CFQ_ATTR(fifo_expire_sync),
4171         CFQ_ATTR(fifo_expire_async),
4172         CFQ_ATTR(back_seek_max),
4173         CFQ_ATTR(back_seek_penalty),
4174         CFQ_ATTR(slice_sync),
4175         CFQ_ATTR(slice_async),
4176         CFQ_ATTR(slice_async_rq),
4177         CFQ_ATTR(slice_idle),
4178         CFQ_ATTR(group_idle),
4179         CFQ_ATTR(low_latency),
4180         __ATTR_NULL
4181 };
4182
4183 static struct elevator_type iosched_cfq = {
4184         .ops = {
4185                 .elevator_merge_fn =            cfq_merge,
4186                 .elevator_merged_fn =           cfq_merged_request,
4187                 .elevator_merge_req_fn =        cfq_merged_requests,
4188                 .elevator_allow_merge_fn =      cfq_allow_merge,
4189                 .elevator_bio_merged_fn =       cfq_bio_merged,
4190                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4191                 .elevator_add_req_fn =          cfq_insert_request,
4192                 .elevator_activate_req_fn =     cfq_activate_request,
4193                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4194                 .elevator_completed_req_fn =    cfq_completed_request,
4195                 .elevator_former_req_fn =       elv_rb_former_request,
4196                 .elevator_latter_req_fn =       elv_rb_latter_request,
4197                 .elevator_set_req_fn =          cfq_set_request,
4198                 .elevator_put_req_fn =          cfq_put_request,
4199                 .elevator_may_queue_fn =        cfq_may_queue,
4200                 .elevator_init_fn =             cfq_init_queue,
4201                 .elevator_exit_fn =             cfq_exit_queue,
4202                 .trim =                         cfq_free_io_context,
4203         },
4204         .elevator_attrs =       cfq_attrs,
4205         .elevator_name =        "cfq",
4206         .elevator_owner =       THIS_MODULE,
4207 };
4208
4209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4210 static struct blkio_policy_type blkio_policy_cfq = {
4211         .ops = {
4212                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4213                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4214         },
4215         .plid = BLKIO_POLICY_PROP,
4216 };
4217 #else
4218 static struct blkio_policy_type blkio_policy_cfq;
4219 #endif
4220
4221 static int __init cfq_init(void)
4222 {
4223         /*
4224          * could be 0 on HZ < 1000 setups
4225          */
4226         if (!cfq_slice_async)
4227                 cfq_slice_async = 1;
4228         if (!cfq_slice_idle)
4229                 cfq_slice_idle = 1;
4230
4231 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4232         if (!cfq_group_idle)
4233                 cfq_group_idle = 1;
4234 #else
4235                 cfq_group_idle = 0;
4236 #endif
4237         if (cfq_slab_setup())
4238                 return -ENOMEM;
4239
4240         elv_register(&iosched_cfq);
4241         blkio_policy_register(&blkio_policy_cfq);
4242
4243         return 0;
4244 }
4245
4246 static void __exit cfq_exit(void)
4247 {
4248         DECLARE_COMPLETION_ONSTACK(all_gone);
4249         blkio_policy_unregister(&blkio_policy_cfq);
4250         elv_unregister(&iosched_cfq);
4251         ioc_gone = &all_gone;
4252         /* ioc_gone's update must be visible before reading ioc_count */
4253         smp_wmb();
4254
4255         /*
4256          * this also protects us from entering cfq_slab_kill() with
4257          * pending RCU callbacks
4258          */
4259         if (elv_ioc_count_read(cfq_ioc_count))
4260                 wait_for_completion(&all_gone);
4261         ida_destroy(&cic_index_ida);
4262         cfq_slab_kill();
4263 }
4264
4265 module_init(cfq_init);
4266 module_exit(cfq_exit);
4267
4268 MODULE_AUTHOR("Jens Axboe");
4269 MODULE_LICENSE("GPL");
4270 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");