block, cfq: reorganize cfq_io_context into generic and cfq specific parts
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk.h"
18 #include "cfq.h"
19
20 /*
21  * tunables
22  */
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
37
38 /*
39  * offset from end of service tree
40  */
41 #define CFQ_IDLE_DELAY          (HZ / 5)
42
43 /*
44  * below this threshold, we consider thinktime immediate
45  */
46 #define CFQ_MIN_TT              (2)
47
48 #define CFQ_SLICE_SCALE         (5)
49 #define CFQ_HW_QUEUE_MIN        (5)
50 #define CFQ_SERVICE_SHIFT       12
51
52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
56
57 #define RQ_CIC(rq)              icq_to_cic((rq)->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_icq_pool;
63
64 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
65 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
66 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
67
68 #define sample_valid(samples)   ((samples) > 80)
69 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
70
71 struct cfq_ttime {
72         unsigned long last_end_request;
73
74         unsigned long ttime_total;
75         unsigned long ttime_samples;
76         unsigned long ttime_mean;
77 };
78
79 /*
80  * Most of our rbtree usage is for sorting with min extraction, so
81  * if we cache the leftmost node we don't have to walk down the tree
82  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
83  * move this into the elevator for the rq sorting as well.
84  */
85 struct cfq_rb_root {
86         struct rb_root rb;
87         struct rb_node *left;
88         unsigned count;
89         unsigned total_weight;
90         u64 min_vdisktime;
91         struct cfq_ttime ttime;
92 };
93 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
94                         .ttime = {.last_end_request = jiffies,},}
95
96 /*
97  * Per process-grouping structure
98  */
99 struct cfq_queue {
100         /* reference count */
101         int ref;
102         /* various state flags, see below */
103         unsigned int flags;
104         /* parent cfq_data */
105         struct cfq_data *cfqd;
106         /* service_tree member */
107         struct rb_node rb_node;
108         /* service_tree key */
109         unsigned long rb_key;
110         /* prio tree member */
111         struct rb_node p_node;
112         /* prio tree root we belong to, if any */
113         struct rb_root *p_root;
114         /* sorted list of pending requests */
115         struct rb_root sort_list;
116         /* if fifo isn't expired, next request to serve */
117         struct request *next_rq;
118         /* requests queued in sort_list */
119         int queued[2];
120         /* currently allocated requests */
121         int allocated[2];
122         /* fifo list of requests in sort_list */
123         struct list_head fifo;
124
125         /* time when queue got scheduled in to dispatch first request. */
126         unsigned long dispatch_start;
127         unsigned int allocated_slice;
128         unsigned int slice_dispatch;
129         /* time when first request from queue completed and slice started. */
130         unsigned long slice_start;
131         unsigned long slice_end;
132         long slice_resid;
133
134         /* pending priority requests */
135         int prio_pending;
136         /* number of requests that are on the dispatch list or inside driver */
137         int dispatched;
138
139         /* io prio of this group */
140         unsigned short ioprio, org_ioprio;
141         unsigned short ioprio_class;
142
143         pid_t pid;
144
145         u32 seek_history;
146         sector_t last_request_pos;
147
148         struct cfq_rb_root *service_tree;
149         struct cfq_queue *new_cfqq;
150         struct cfq_group *cfqg;
151         /* Number of sectors dispatched from queue in single dispatch round */
152         unsigned long nr_sectors;
153 };
154
155 /*
156  * First index in the service_trees.
157  * IDLE is handled separately, so it has negative index
158  */
159 enum wl_prio_t {
160         BE_WORKLOAD = 0,
161         RT_WORKLOAD = 1,
162         IDLE_WORKLOAD = 2,
163         CFQ_PRIO_NR,
164 };
165
166 /*
167  * Second index in the service_trees.
168  */
169 enum wl_type_t {
170         ASYNC_WORKLOAD = 0,
171         SYNC_NOIDLE_WORKLOAD = 1,
172         SYNC_WORKLOAD = 2
173 };
174
175 /* This is per cgroup per device grouping structure */
176 struct cfq_group {
177         /* group service_tree member */
178         struct rb_node rb_node;
179
180         /* group service_tree key */
181         u64 vdisktime;
182         unsigned int weight;
183         unsigned int new_weight;
184         bool needs_update;
185
186         /* number of cfqq currently on this group */
187         int nr_cfqq;
188
189         /*
190          * Per group busy queues average. Useful for workload slice calc. We
191          * create the array for each prio class but at run time it is used
192          * only for RT and BE class and slot for IDLE class remains unused.
193          * This is primarily done to avoid confusion and a gcc warning.
194          */
195         unsigned int busy_queues_avg[CFQ_PRIO_NR];
196         /*
197          * rr lists of queues with requests. We maintain service trees for
198          * RT and BE classes. These trees are subdivided in subclasses
199          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
200          * class there is no subclassification and all the cfq queues go on
201          * a single tree service_tree_idle.
202          * Counts are embedded in the cfq_rb_root
203          */
204         struct cfq_rb_root service_trees[2][3];
205         struct cfq_rb_root service_tree_idle;
206
207         unsigned long saved_workload_slice;
208         enum wl_type_t saved_workload;
209         enum wl_prio_t saved_serving_prio;
210         struct blkio_group blkg;
211 #ifdef CONFIG_CFQ_GROUP_IOSCHED
212         struct hlist_node cfqd_node;
213         int ref;
214 #endif
215         /* number of requests that are on the dispatch list or inside driver */
216         int dispatched;
217         struct cfq_ttime ttime;
218 };
219
220 struct cfq_io_cq {
221         struct io_cq            icq;            /* must be the first member */
222         struct cfq_queue        *cfqq[2];
223         struct cfq_ttime        ttime;
224 };
225
226 /*
227  * Per block device queue structure
228  */
229 struct cfq_data {
230         struct request_queue *queue;
231         /* Root service tree for cfq_groups */
232         struct cfq_rb_root grp_service_tree;
233         struct cfq_group root_group;
234
235         /*
236          * The priority currently being served
237          */
238         enum wl_prio_t serving_prio;
239         enum wl_type_t serving_type;
240         unsigned long workload_expires;
241         struct cfq_group *serving_group;
242
243         /*
244          * Each priority tree is sorted by next_request position.  These
245          * trees are used when determining if two or more queues are
246          * interleaving requests (see cfq_close_cooperator).
247          */
248         struct rb_root prio_trees[CFQ_PRIO_LISTS];
249
250         unsigned int busy_queues;
251         unsigned int busy_sync_queues;
252
253         int rq_in_driver;
254         int rq_in_flight[2];
255
256         /*
257          * queue-depth detection
258          */
259         int rq_queued;
260         int hw_tag;
261         /*
262          * hw_tag can be
263          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
264          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
265          *  0 => no NCQ
266          */
267         int hw_tag_est_depth;
268         unsigned int hw_tag_samples;
269
270         /*
271          * idle window management
272          */
273         struct timer_list idle_slice_timer;
274         struct work_struct unplug_work;
275
276         struct cfq_queue *active_queue;
277         struct cfq_io_cq *active_cic;
278
279         /*
280          * async queue for each priority case
281          */
282         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
283         struct cfq_queue *async_idle_cfqq;
284
285         sector_t last_position;
286
287         /*
288          * tunables, see top of file
289          */
290         unsigned int cfq_quantum;
291         unsigned int cfq_fifo_expire[2];
292         unsigned int cfq_back_penalty;
293         unsigned int cfq_back_max;
294         unsigned int cfq_slice[2];
295         unsigned int cfq_slice_async_rq;
296         unsigned int cfq_slice_idle;
297         unsigned int cfq_group_idle;
298         unsigned int cfq_latency;
299
300         struct list_head icq_list;
301
302         /*
303          * Fallback dummy cfqq for extreme OOM conditions
304          */
305         struct cfq_queue oom_cfqq;
306
307         unsigned long last_delayed_sync;
308
309         /* List of cfq groups being managed on this device*/
310         struct hlist_head cfqg_list;
311
312         /* Number of groups which are on blkcg->blkg_list */
313         unsigned int nr_blkcg_linked_grps;
314 };
315
316 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
317
318 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
319                                             enum wl_prio_t prio,
320                                             enum wl_type_t type)
321 {
322         if (!cfqg)
323                 return NULL;
324
325         if (prio == IDLE_WORKLOAD)
326                 return &cfqg->service_tree_idle;
327
328         return &cfqg->service_trees[prio][type];
329 }
330
331 enum cfqq_state_flags {
332         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
333         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
334         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
335         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
336         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
337         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
338         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
339         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
340         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
341         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
342         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
343         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
344         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
345 };
346
347 #define CFQ_CFQQ_FNS(name)                                              \
348 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
349 {                                                                       \
350         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
351 }                                                                       \
352 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
353 {                                                                       \
354         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
355 }                                                                       \
356 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
357 {                                                                       \
358         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
359 }
360
361 CFQ_CFQQ_FNS(on_rr);
362 CFQ_CFQQ_FNS(wait_request);
363 CFQ_CFQQ_FNS(must_dispatch);
364 CFQ_CFQQ_FNS(must_alloc_slice);
365 CFQ_CFQQ_FNS(fifo_expire);
366 CFQ_CFQQ_FNS(idle_window);
367 CFQ_CFQQ_FNS(prio_changed);
368 CFQ_CFQQ_FNS(slice_new);
369 CFQ_CFQQ_FNS(sync);
370 CFQ_CFQQ_FNS(coop);
371 CFQ_CFQQ_FNS(split_coop);
372 CFQ_CFQQ_FNS(deep);
373 CFQ_CFQQ_FNS(wait_busy);
374 #undef CFQ_CFQQ_FNS
375
376 #ifdef CONFIG_CFQ_GROUP_IOSCHED
377 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
378         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
379                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
380                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
381
382 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
383         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
384                                 blkg_path(&(cfqg)->blkg), ##args)       \
385
386 #else
387 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
388         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
389 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
390 #endif
391 #define cfq_log(cfqd, fmt, args...)     \
392         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
393
394 /* Traverses through cfq group service trees */
395 #define for_each_cfqg_st(cfqg, i, j, st) \
396         for (i = 0; i <= IDLE_WORKLOAD; i++) \
397                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
398                         : &cfqg->service_tree_idle; \
399                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
400                         (i == IDLE_WORKLOAD && j == 0); \
401                         j++, st = i < IDLE_WORKLOAD ? \
402                         &cfqg->service_trees[i][j]: NULL) \
403
404 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
405         struct cfq_ttime *ttime, bool group_idle)
406 {
407         unsigned long slice;
408         if (!sample_valid(ttime->ttime_samples))
409                 return false;
410         if (group_idle)
411                 slice = cfqd->cfq_group_idle;
412         else
413                 slice = cfqd->cfq_slice_idle;
414         return ttime->ttime_mean > slice;
415 }
416
417 static inline bool iops_mode(struct cfq_data *cfqd)
418 {
419         /*
420          * If we are not idling on queues and it is a NCQ drive, parallel
421          * execution of requests is on and measuring time is not possible
422          * in most of the cases until and unless we drive shallower queue
423          * depths and that becomes a performance bottleneck. In such cases
424          * switch to start providing fairness in terms of number of IOs.
425          */
426         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
427                 return true;
428         else
429                 return false;
430 }
431
432 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
433 {
434         if (cfq_class_idle(cfqq))
435                 return IDLE_WORKLOAD;
436         if (cfq_class_rt(cfqq))
437                 return RT_WORKLOAD;
438         return BE_WORKLOAD;
439 }
440
441
442 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
443 {
444         if (!cfq_cfqq_sync(cfqq))
445                 return ASYNC_WORKLOAD;
446         if (!cfq_cfqq_idle_window(cfqq))
447                 return SYNC_NOIDLE_WORKLOAD;
448         return SYNC_WORKLOAD;
449 }
450
451 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
452                                         struct cfq_data *cfqd,
453                                         struct cfq_group *cfqg)
454 {
455         if (wl == IDLE_WORKLOAD)
456                 return cfqg->service_tree_idle.count;
457
458         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
459                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
460                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
461 }
462
463 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
464                                         struct cfq_group *cfqg)
465 {
466         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
467                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
468 }
469
470 static void cfq_dispatch_insert(struct request_queue *, struct request *);
471 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
472                                        struct io_context *, gfp_t);
473 static struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *, struct io_context *);
474
475 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
476 {
477         /* cic->icq is the first member, %NULL will convert to %NULL */
478         return container_of(icq, struct cfq_io_cq, icq);
479 }
480
481 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
482 {
483         return cic->cfqq[is_sync];
484 }
485
486 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
487                                 bool is_sync)
488 {
489         cic->cfqq[is_sync] = cfqq;
490 }
491
492 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
493 {
494         return cic->icq.q->elevator->elevator_data;
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         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
685                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
686
687         s1 = blk_rq_pos(rq1);
688         s2 = blk_rq_pos(rq2);
689
690         /*
691          * by definition, 1KiB is 2 sectors
692          */
693         back_max = cfqd->cfq_back_max * 2;
694
695         /*
696          * Strict one way elevator _except_ in the case where we allow
697          * short backward seeks which are biased as twice the cost of a
698          * similar forward seek.
699          */
700         if (s1 >= last)
701                 d1 = s1 - last;
702         else if (s1 + back_max >= last)
703                 d1 = (last - s1) * cfqd->cfq_back_penalty;
704         else
705                 wrap |= CFQ_RQ1_WRAP;
706
707         if (s2 >= last)
708                 d2 = s2 - last;
709         else if (s2 + back_max >= last)
710                 d2 = (last - s2) * cfqd->cfq_back_penalty;
711         else
712                 wrap |= CFQ_RQ2_WRAP;
713
714         /* Found required data */
715
716         /*
717          * By doing switch() on the bit mask "wrap" we avoid having to
718          * check two variables for all permutations: --> faster!
719          */
720         switch (wrap) {
721         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
722                 if (d1 < d2)
723                         return rq1;
724                 else if (d2 < d1)
725                         return rq2;
726                 else {
727                         if (s1 >= s2)
728                                 return rq1;
729                         else
730                                 return rq2;
731                 }
732
733         case CFQ_RQ2_WRAP:
734                 return rq1;
735         case CFQ_RQ1_WRAP:
736                 return rq2;
737         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
738         default:
739                 /*
740                  * Since both rqs are wrapped,
741                  * start with the one that's further behind head
742                  * (--> only *one* back seek required),
743                  * since back seek takes more time than forward.
744                  */
745                 if (s1 <= s2)
746                         return rq1;
747                 else
748                         return rq2;
749         }
750 }
751
752 /*
753  * The below is leftmost cache rbtree addon
754  */
755 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
756 {
757         /* Service tree is empty */
758         if (!root->count)
759                 return NULL;
760
761         if (!root->left)
762                 root->left = rb_first(&root->rb);
763
764         if (root->left)
765                 return rb_entry(root->left, struct cfq_queue, rb_node);
766
767         return NULL;
768 }
769
770 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
771 {
772         if (!root->left)
773                 root->left = rb_first(&root->rb);
774
775         if (root->left)
776                 return rb_entry_cfqg(root->left);
777
778         return NULL;
779 }
780
781 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
782 {
783         rb_erase(n, root);
784         RB_CLEAR_NODE(n);
785 }
786
787 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
788 {
789         if (root->left == n)
790                 root->left = NULL;
791         rb_erase_init(n, &root->rb);
792         --root->count;
793 }
794
795 /*
796  * would be nice to take fifo expire time into account as well
797  */
798 static struct request *
799 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
800                   struct request *last)
801 {
802         struct rb_node *rbnext = rb_next(&last->rb_node);
803         struct rb_node *rbprev = rb_prev(&last->rb_node);
804         struct request *next = NULL, *prev = NULL;
805
806         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
807
808         if (rbprev)
809                 prev = rb_entry_rq(rbprev);
810
811         if (rbnext)
812                 next = rb_entry_rq(rbnext);
813         else {
814                 rbnext = rb_first(&cfqq->sort_list);
815                 if (rbnext && rbnext != &last->rb_node)
816                         next = rb_entry_rq(rbnext);
817         }
818
819         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
820 }
821
822 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
823                                       struct cfq_queue *cfqq)
824 {
825         /*
826          * just an approximation, should be ok.
827          */
828         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
829                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
830 }
831
832 static inline s64
833 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
834 {
835         return cfqg->vdisktime - st->min_vdisktime;
836 }
837
838 static void
839 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
840 {
841         struct rb_node **node = &st->rb.rb_node;
842         struct rb_node *parent = NULL;
843         struct cfq_group *__cfqg;
844         s64 key = cfqg_key(st, cfqg);
845         int left = 1;
846
847         while (*node != NULL) {
848                 parent = *node;
849                 __cfqg = rb_entry_cfqg(parent);
850
851                 if (key < cfqg_key(st, __cfqg))
852                         node = &parent->rb_left;
853                 else {
854                         node = &parent->rb_right;
855                         left = 0;
856                 }
857         }
858
859         if (left)
860                 st->left = &cfqg->rb_node;
861
862         rb_link_node(&cfqg->rb_node, parent, node);
863         rb_insert_color(&cfqg->rb_node, &st->rb);
864 }
865
866 static void
867 cfq_update_group_weight(struct cfq_group *cfqg)
868 {
869         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
870         if (cfqg->needs_update) {
871                 cfqg->weight = cfqg->new_weight;
872                 cfqg->needs_update = false;
873         }
874 }
875
876 static void
877 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
878 {
879         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
880
881         cfq_update_group_weight(cfqg);
882         __cfq_group_service_tree_add(st, cfqg);
883         st->total_weight += cfqg->weight;
884 }
885
886 static void
887 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
888 {
889         struct cfq_rb_root *st = &cfqd->grp_service_tree;
890         struct cfq_group *__cfqg;
891         struct rb_node *n;
892
893         cfqg->nr_cfqq++;
894         if (!RB_EMPTY_NODE(&cfqg->rb_node))
895                 return;
896
897         /*
898          * Currently put the group at the end. Later implement something
899          * so that groups get lesser vtime based on their weights, so that
900          * if group does not loose all if it was not continuously backlogged.
901          */
902         n = rb_last(&st->rb);
903         if (n) {
904                 __cfqg = rb_entry_cfqg(n);
905                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
906         } else
907                 cfqg->vdisktime = st->min_vdisktime;
908         cfq_group_service_tree_add(st, cfqg);
909 }
910
911 static void
912 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
913 {
914         st->total_weight -= cfqg->weight;
915         if (!RB_EMPTY_NODE(&cfqg->rb_node))
916                 cfq_rb_erase(&cfqg->rb_node, st);
917 }
918
919 static void
920 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
921 {
922         struct cfq_rb_root *st = &cfqd->grp_service_tree;
923
924         BUG_ON(cfqg->nr_cfqq < 1);
925         cfqg->nr_cfqq--;
926
927         /* If there are other cfq queues under this group, don't delete it */
928         if (cfqg->nr_cfqq)
929                 return;
930
931         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
932         cfq_group_service_tree_del(st, cfqg);
933         cfqg->saved_workload_slice = 0;
934         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
935 }
936
937 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
938                                                 unsigned int *unaccounted_time)
939 {
940         unsigned int slice_used;
941
942         /*
943          * Queue got expired before even a single request completed or
944          * got expired immediately after first request completion.
945          */
946         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
947                 /*
948                  * Also charge the seek time incurred to the group, otherwise
949                  * if there are mutiple queues in the group, each can dispatch
950                  * a single request on seeky media and cause lots of seek time
951                  * and group will never know it.
952                  */
953                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
954                                         1);
955         } else {
956                 slice_used = jiffies - cfqq->slice_start;
957                 if (slice_used > cfqq->allocated_slice) {
958                         *unaccounted_time = slice_used - cfqq->allocated_slice;
959                         slice_used = cfqq->allocated_slice;
960                 }
961                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
962                         *unaccounted_time += cfqq->slice_start -
963                                         cfqq->dispatch_start;
964         }
965
966         return slice_used;
967 }
968
969 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
970                                 struct cfq_queue *cfqq)
971 {
972         struct cfq_rb_root *st = &cfqd->grp_service_tree;
973         unsigned int used_sl, charge, unaccounted_sl = 0;
974         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
975                         - cfqg->service_tree_idle.count;
976
977         BUG_ON(nr_sync < 0);
978         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
979
980         if (iops_mode(cfqd))
981                 charge = cfqq->slice_dispatch;
982         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
983                 charge = cfqq->allocated_slice;
984
985         /* Can't update vdisktime while group is on service tree */
986         cfq_group_service_tree_del(st, cfqg);
987         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
988         /* If a new weight was requested, update now, off tree */
989         cfq_group_service_tree_add(st, cfqg);
990
991         /* This group is being expired. Save the context */
992         if (time_after(cfqd->workload_expires, jiffies)) {
993                 cfqg->saved_workload_slice = cfqd->workload_expires
994                                                 - jiffies;
995                 cfqg->saved_workload = cfqd->serving_type;
996                 cfqg->saved_serving_prio = cfqd->serving_prio;
997         } else
998                 cfqg->saved_workload_slice = 0;
999
1000         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1001                                         st->min_vdisktime);
1002         cfq_log_cfqq(cfqq->cfqd, cfqq,
1003                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1004                      used_sl, cfqq->slice_dispatch, charge,
1005                      iops_mode(cfqd), cfqq->nr_sectors);
1006         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1007                                           unaccounted_sl);
1008         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1009 }
1010
1011 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1012 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1013 {
1014         if (blkg)
1015                 return container_of(blkg, struct cfq_group, blkg);
1016         return NULL;
1017 }
1018
1019 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1020                                           unsigned int weight)
1021 {
1022         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1023         cfqg->new_weight = weight;
1024         cfqg->needs_update = true;
1025 }
1026
1027 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1028                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1029 {
1030         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1031         unsigned int major, minor;
1032
1033         /*
1034          * Add group onto cgroup list. It might happen that bdi->dev is
1035          * not initialized yet. Initialize this new group without major
1036          * and minor info and this info will be filled in once a new thread
1037          * comes for IO.
1038          */
1039         if (bdi->dev) {
1040                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1041                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1042                                         (void *)cfqd, MKDEV(major, minor));
1043         } else
1044                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1045                                         (void *)cfqd, 0);
1046
1047         cfqd->nr_blkcg_linked_grps++;
1048         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1049
1050         /* Add group on cfqd list */
1051         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1052 }
1053
1054 /*
1055  * Should be called from sleepable context. No request queue lock as per
1056  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1057  * from sleepable context.
1058  */
1059 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1060 {
1061         struct cfq_group *cfqg = NULL;
1062         int i, j, ret;
1063         struct cfq_rb_root *st;
1064
1065         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1066         if (!cfqg)
1067                 return NULL;
1068
1069         for_each_cfqg_st(cfqg, i, j, st)
1070                 *st = CFQ_RB_ROOT;
1071         RB_CLEAR_NODE(&cfqg->rb_node);
1072
1073         cfqg->ttime.last_end_request = jiffies;
1074
1075         /*
1076          * Take the initial reference that will be released on destroy
1077          * This can be thought of a joint reference by cgroup and
1078          * elevator which will be dropped by either elevator exit
1079          * or cgroup deletion path depending on who is exiting first.
1080          */
1081         cfqg->ref = 1;
1082
1083         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1084         if (ret) {
1085                 kfree(cfqg);
1086                 return NULL;
1087         }
1088
1089         return cfqg;
1090 }
1091
1092 static struct cfq_group *
1093 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1094 {
1095         struct cfq_group *cfqg = NULL;
1096         void *key = cfqd;
1097         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1098         unsigned int major, minor;
1099
1100         /*
1101          * This is the common case when there are no blkio cgroups.
1102          * Avoid lookup in this case
1103          */
1104         if (blkcg == &blkio_root_cgroup)
1105                 cfqg = &cfqd->root_group;
1106         else
1107                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1108
1109         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1110                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1111                 cfqg->blkg.dev = MKDEV(major, minor);
1112         }
1113
1114         return cfqg;
1115 }
1116
1117 /*
1118  * Search for the cfq group current task belongs to. request_queue lock must
1119  * be held.
1120  */
1121 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1122 {
1123         struct blkio_cgroup *blkcg;
1124         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1125         struct request_queue *q = cfqd->queue;
1126
1127         rcu_read_lock();
1128         blkcg = task_blkio_cgroup(current);
1129         cfqg = cfq_find_cfqg(cfqd, blkcg);
1130         if (cfqg) {
1131                 rcu_read_unlock();
1132                 return cfqg;
1133         }
1134
1135         /*
1136          * Need to allocate a group. Allocation of group also needs allocation
1137          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1138          * we need to drop rcu lock and queue_lock before we call alloc.
1139          *
1140          * Not taking any queue reference here and assuming that queue is
1141          * around by the time we return. CFQ queue allocation code does
1142          * the same. It might be racy though.
1143          */
1144
1145         rcu_read_unlock();
1146         spin_unlock_irq(q->queue_lock);
1147
1148         cfqg = cfq_alloc_cfqg(cfqd);
1149
1150         spin_lock_irq(q->queue_lock);
1151
1152         rcu_read_lock();
1153         blkcg = task_blkio_cgroup(current);
1154
1155         /*
1156          * If some other thread already allocated the group while we were
1157          * not holding queue lock, free up the group
1158          */
1159         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1160
1161         if (__cfqg) {
1162                 kfree(cfqg);
1163                 rcu_read_unlock();
1164                 return __cfqg;
1165         }
1166
1167         if (!cfqg)
1168                 cfqg = &cfqd->root_group;
1169
1170         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1171         rcu_read_unlock();
1172         return cfqg;
1173 }
1174
1175 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1176 {
1177         cfqg->ref++;
1178         return cfqg;
1179 }
1180
1181 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1182 {
1183         /* Currently, all async queues are mapped to root group */
1184         if (!cfq_cfqq_sync(cfqq))
1185                 cfqg = &cfqq->cfqd->root_group;
1186
1187         cfqq->cfqg = cfqg;
1188         /* cfqq reference on cfqg */
1189         cfqq->cfqg->ref++;
1190 }
1191
1192 static void cfq_put_cfqg(struct cfq_group *cfqg)
1193 {
1194         struct cfq_rb_root *st;
1195         int i, j;
1196
1197         BUG_ON(cfqg->ref <= 0);
1198         cfqg->ref--;
1199         if (cfqg->ref)
1200                 return;
1201         for_each_cfqg_st(cfqg, i, j, st)
1202                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1203         free_percpu(cfqg->blkg.stats_cpu);
1204         kfree(cfqg);
1205 }
1206
1207 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1208 {
1209         /* Something wrong if we are trying to remove same group twice */
1210         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1211
1212         hlist_del_init(&cfqg->cfqd_node);
1213
1214         BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1215         cfqd->nr_blkcg_linked_grps--;
1216
1217         /*
1218          * Put the reference taken at the time of creation so that when all
1219          * queues are gone, group can be destroyed.
1220          */
1221         cfq_put_cfqg(cfqg);
1222 }
1223
1224 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1225 {
1226         struct hlist_node *pos, *n;
1227         struct cfq_group *cfqg;
1228
1229         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1230                 /*
1231                  * If cgroup removal path got to blk_group first and removed
1232                  * it from cgroup list, then it will take care of destroying
1233                  * cfqg also.
1234                  */
1235                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1236                         cfq_destroy_cfqg(cfqd, cfqg);
1237         }
1238 }
1239
1240 /*
1241  * Blk cgroup controller notification saying that blkio_group object is being
1242  * delinked as associated cgroup object is going away. That also means that
1243  * no new IO will come in this group. So get rid of this group as soon as
1244  * any pending IO in the group is finished.
1245  *
1246  * This function is called under rcu_read_lock(). key is the rcu protected
1247  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1248  * read lock.
1249  *
1250  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1251  * it should not be NULL as even if elevator was exiting, cgroup deltion
1252  * path got to it first.
1253  */
1254 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1255 {
1256         unsigned long  flags;
1257         struct cfq_data *cfqd = key;
1258
1259         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1260         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1261         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1262 }
1263
1264 #else /* GROUP_IOSCHED */
1265 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1266 {
1267         return &cfqd->root_group;
1268 }
1269
1270 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1271 {
1272         return cfqg;
1273 }
1274
1275 static inline void
1276 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1277         cfqq->cfqg = cfqg;
1278 }
1279
1280 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1281 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1282
1283 #endif /* GROUP_IOSCHED */
1284
1285 /*
1286  * The cfqd->service_trees holds all pending cfq_queue's that have
1287  * requests waiting to be processed. It is sorted in the order that
1288  * we will service the queues.
1289  */
1290 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1291                                  bool add_front)
1292 {
1293         struct rb_node **p, *parent;
1294         struct cfq_queue *__cfqq;
1295         unsigned long rb_key;
1296         struct cfq_rb_root *service_tree;
1297         int left;
1298         int new_cfqq = 1;
1299
1300         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1301                                                 cfqq_type(cfqq));
1302         if (cfq_class_idle(cfqq)) {
1303                 rb_key = CFQ_IDLE_DELAY;
1304                 parent = rb_last(&service_tree->rb);
1305                 if (parent && parent != &cfqq->rb_node) {
1306                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1307                         rb_key += __cfqq->rb_key;
1308                 } else
1309                         rb_key += jiffies;
1310         } else if (!add_front) {
1311                 /*
1312                  * Get our rb key offset. Subtract any residual slice
1313                  * value carried from last service. A negative resid
1314                  * count indicates slice overrun, and this should position
1315                  * the next service time further away in the tree.
1316                  */
1317                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1318                 rb_key -= cfqq->slice_resid;
1319                 cfqq->slice_resid = 0;
1320         } else {
1321                 rb_key = -HZ;
1322                 __cfqq = cfq_rb_first(service_tree);
1323                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1324         }
1325
1326         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1327                 new_cfqq = 0;
1328                 /*
1329                  * same position, nothing more to do
1330                  */
1331                 if (rb_key == cfqq->rb_key &&
1332                     cfqq->service_tree == service_tree)
1333                         return;
1334
1335                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1336                 cfqq->service_tree = NULL;
1337         }
1338
1339         left = 1;
1340         parent = NULL;
1341         cfqq->service_tree = service_tree;
1342         p = &service_tree->rb.rb_node;
1343         while (*p) {
1344                 struct rb_node **n;
1345
1346                 parent = *p;
1347                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1348
1349                 /*
1350                  * sort by key, that represents service time.
1351                  */
1352                 if (time_before(rb_key, __cfqq->rb_key))
1353                         n = &(*p)->rb_left;
1354                 else {
1355                         n = &(*p)->rb_right;
1356                         left = 0;
1357                 }
1358
1359                 p = n;
1360         }
1361
1362         if (left)
1363                 service_tree->left = &cfqq->rb_node;
1364
1365         cfqq->rb_key = rb_key;
1366         rb_link_node(&cfqq->rb_node, parent, p);
1367         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1368         service_tree->count++;
1369         if (add_front || !new_cfqq)
1370                 return;
1371         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1372 }
1373
1374 static struct cfq_queue *
1375 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1376                      sector_t sector, struct rb_node **ret_parent,
1377                      struct rb_node ***rb_link)
1378 {
1379         struct rb_node **p, *parent;
1380         struct cfq_queue *cfqq = NULL;
1381
1382         parent = NULL;
1383         p = &root->rb_node;
1384         while (*p) {
1385                 struct rb_node **n;
1386
1387                 parent = *p;
1388                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1389
1390                 /*
1391                  * Sort strictly based on sector.  Smallest to the left,
1392                  * largest to the right.
1393                  */
1394                 if (sector > blk_rq_pos(cfqq->next_rq))
1395                         n = &(*p)->rb_right;
1396                 else if (sector < blk_rq_pos(cfqq->next_rq))
1397                         n = &(*p)->rb_left;
1398                 else
1399                         break;
1400                 p = n;
1401                 cfqq = NULL;
1402         }
1403
1404         *ret_parent = parent;
1405         if (rb_link)
1406                 *rb_link = p;
1407         return cfqq;
1408 }
1409
1410 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1411 {
1412         struct rb_node **p, *parent;
1413         struct cfq_queue *__cfqq;
1414
1415         if (cfqq->p_root) {
1416                 rb_erase(&cfqq->p_node, cfqq->p_root);
1417                 cfqq->p_root = NULL;
1418         }
1419
1420         if (cfq_class_idle(cfqq))
1421                 return;
1422         if (!cfqq->next_rq)
1423                 return;
1424
1425         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1426         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1427                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1428         if (!__cfqq) {
1429                 rb_link_node(&cfqq->p_node, parent, p);
1430                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1431         } else
1432                 cfqq->p_root = NULL;
1433 }
1434
1435 /*
1436  * Update cfqq's position in the service tree.
1437  */
1438 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1439 {
1440         /*
1441          * Resorting requires the cfqq to be on the RR list already.
1442          */
1443         if (cfq_cfqq_on_rr(cfqq)) {
1444                 cfq_service_tree_add(cfqd, cfqq, 0);
1445                 cfq_prio_tree_add(cfqd, cfqq);
1446         }
1447 }
1448
1449 /*
1450  * add to busy list of queues for service, trying to be fair in ordering
1451  * the pending list according to last request service
1452  */
1453 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1454 {
1455         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1456         BUG_ON(cfq_cfqq_on_rr(cfqq));
1457         cfq_mark_cfqq_on_rr(cfqq);
1458         cfqd->busy_queues++;
1459         if (cfq_cfqq_sync(cfqq))
1460                 cfqd->busy_sync_queues++;
1461
1462         cfq_resort_rr_list(cfqd, cfqq);
1463 }
1464
1465 /*
1466  * Called when the cfqq no longer has requests pending, remove it from
1467  * the service tree.
1468  */
1469 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1470 {
1471         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1472         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1473         cfq_clear_cfqq_on_rr(cfqq);
1474
1475         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1476                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1477                 cfqq->service_tree = NULL;
1478         }
1479         if (cfqq->p_root) {
1480                 rb_erase(&cfqq->p_node, cfqq->p_root);
1481                 cfqq->p_root = NULL;
1482         }
1483
1484         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1485         BUG_ON(!cfqd->busy_queues);
1486         cfqd->busy_queues--;
1487         if (cfq_cfqq_sync(cfqq))
1488                 cfqd->busy_sync_queues--;
1489 }
1490
1491 /*
1492  * rb tree support functions
1493  */
1494 static void cfq_del_rq_rb(struct request *rq)
1495 {
1496         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1497         const int sync = rq_is_sync(rq);
1498
1499         BUG_ON(!cfqq->queued[sync]);
1500         cfqq->queued[sync]--;
1501
1502         elv_rb_del(&cfqq->sort_list, rq);
1503
1504         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1505                 /*
1506                  * Queue will be deleted from service tree when we actually
1507                  * expire it later. Right now just remove it from prio tree
1508                  * as it is empty.
1509                  */
1510                 if (cfqq->p_root) {
1511                         rb_erase(&cfqq->p_node, cfqq->p_root);
1512                         cfqq->p_root = NULL;
1513                 }
1514         }
1515 }
1516
1517 static void cfq_add_rq_rb(struct request *rq)
1518 {
1519         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1520         struct cfq_data *cfqd = cfqq->cfqd;
1521         struct request *prev;
1522
1523         cfqq->queued[rq_is_sync(rq)]++;
1524
1525         elv_rb_add(&cfqq->sort_list, rq);
1526
1527         if (!cfq_cfqq_on_rr(cfqq))
1528                 cfq_add_cfqq_rr(cfqd, cfqq);
1529
1530         /*
1531          * check if this request is a better next-serve candidate
1532          */
1533         prev = cfqq->next_rq;
1534         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1535
1536         /*
1537          * adjust priority tree position, if ->next_rq changes
1538          */
1539         if (prev != cfqq->next_rq)
1540                 cfq_prio_tree_add(cfqd, cfqq);
1541
1542         BUG_ON(!cfqq->next_rq);
1543 }
1544
1545 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1546 {
1547         elv_rb_del(&cfqq->sort_list, rq);
1548         cfqq->queued[rq_is_sync(rq)]--;
1549         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1550                                         rq_data_dir(rq), rq_is_sync(rq));
1551         cfq_add_rq_rb(rq);
1552         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1553                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1554                         rq_is_sync(rq));
1555 }
1556
1557 static struct request *
1558 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1559 {
1560         struct task_struct *tsk = current;
1561         struct cfq_io_cq *cic;
1562         struct cfq_queue *cfqq;
1563
1564         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1565         if (!cic)
1566                 return NULL;
1567
1568         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1569         if (cfqq) {
1570                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1571
1572                 return elv_rb_find(&cfqq->sort_list, sector);
1573         }
1574
1575         return NULL;
1576 }
1577
1578 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1579 {
1580         struct cfq_data *cfqd = q->elevator->elevator_data;
1581
1582         cfqd->rq_in_driver++;
1583         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1584                                                 cfqd->rq_in_driver);
1585
1586         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1587 }
1588
1589 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1590 {
1591         struct cfq_data *cfqd = q->elevator->elevator_data;
1592
1593         WARN_ON(!cfqd->rq_in_driver);
1594         cfqd->rq_in_driver--;
1595         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1596                                                 cfqd->rq_in_driver);
1597 }
1598
1599 static void cfq_remove_request(struct request *rq)
1600 {
1601         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1602
1603         if (cfqq->next_rq == rq)
1604                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1605
1606         list_del_init(&rq->queuelist);
1607         cfq_del_rq_rb(rq);
1608
1609         cfqq->cfqd->rq_queued--;
1610         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1611                                         rq_data_dir(rq), rq_is_sync(rq));
1612         if (rq->cmd_flags & REQ_PRIO) {
1613                 WARN_ON(!cfqq->prio_pending);
1614                 cfqq->prio_pending--;
1615         }
1616 }
1617
1618 static int cfq_merge(struct request_queue *q, struct request **req,
1619                      struct bio *bio)
1620 {
1621         struct cfq_data *cfqd = q->elevator->elevator_data;
1622         struct request *__rq;
1623
1624         __rq = cfq_find_rq_fmerge(cfqd, bio);
1625         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1626                 *req = __rq;
1627                 return ELEVATOR_FRONT_MERGE;
1628         }
1629
1630         return ELEVATOR_NO_MERGE;
1631 }
1632
1633 static void cfq_merged_request(struct request_queue *q, struct request *req,
1634                                int type)
1635 {
1636         if (type == ELEVATOR_FRONT_MERGE) {
1637                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1638
1639                 cfq_reposition_rq_rb(cfqq, req);
1640         }
1641 }
1642
1643 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1644                                 struct bio *bio)
1645 {
1646         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1647                                         bio_data_dir(bio), cfq_bio_sync(bio));
1648 }
1649
1650 static void
1651 cfq_merged_requests(struct request_queue *q, struct request *rq,
1652                     struct request *next)
1653 {
1654         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1655         /*
1656          * reposition in fifo if next is older than rq
1657          */
1658         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1659             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1660                 list_move(&rq->queuelist, &next->queuelist);
1661                 rq_set_fifo_time(rq, rq_fifo_time(next));
1662         }
1663
1664         if (cfqq->next_rq == next)
1665                 cfqq->next_rq = rq;
1666         cfq_remove_request(next);
1667         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1668                                         rq_data_dir(next), rq_is_sync(next));
1669 }
1670
1671 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1672                            struct bio *bio)
1673 {
1674         struct cfq_data *cfqd = q->elevator->elevator_data;
1675         struct cfq_io_cq *cic;
1676         struct cfq_queue *cfqq;
1677
1678         /*
1679          * Disallow merge of a sync bio into an async request.
1680          */
1681         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1682                 return false;
1683
1684         /*
1685          * Lookup the cfqq that this bio will be queued with and allow
1686          * merge only if rq is queued there.  This function can be called
1687          * from plug merge without queue_lock.  In such cases, ioc of @rq
1688          * and %current are guaranteed to be equal.  Avoid lookup which
1689          * requires queue_lock by using @rq's cic.
1690          */
1691         if (current->io_context == RQ_CIC(rq)->icq.ioc) {
1692                 cic = RQ_CIC(rq);
1693         } else {
1694                 cic = cfq_cic_lookup(cfqd, current->io_context);
1695                 if (!cic)
1696                         return false;
1697         }
1698
1699         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1700         return cfqq == RQ_CFQQ(rq);
1701 }
1702
1703 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1704 {
1705         del_timer(&cfqd->idle_slice_timer);
1706         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1707 }
1708
1709 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1710                                    struct cfq_queue *cfqq)
1711 {
1712         if (cfqq) {
1713                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1714                                 cfqd->serving_prio, cfqd->serving_type);
1715                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1716                 cfqq->slice_start = 0;
1717                 cfqq->dispatch_start = jiffies;
1718                 cfqq->allocated_slice = 0;
1719                 cfqq->slice_end = 0;
1720                 cfqq->slice_dispatch = 0;
1721                 cfqq->nr_sectors = 0;
1722
1723                 cfq_clear_cfqq_wait_request(cfqq);
1724                 cfq_clear_cfqq_must_dispatch(cfqq);
1725                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1726                 cfq_clear_cfqq_fifo_expire(cfqq);
1727                 cfq_mark_cfqq_slice_new(cfqq);
1728
1729                 cfq_del_timer(cfqd, cfqq);
1730         }
1731
1732         cfqd->active_queue = cfqq;
1733 }
1734
1735 /*
1736  * current cfqq expired its slice (or was too idle), select new one
1737  */
1738 static void
1739 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1740                     bool timed_out)
1741 {
1742         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1743
1744         if (cfq_cfqq_wait_request(cfqq))
1745                 cfq_del_timer(cfqd, cfqq);
1746
1747         cfq_clear_cfqq_wait_request(cfqq);
1748         cfq_clear_cfqq_wait_busy(cfqq);
1749
1750         /*
1751          * If this cfqq is shared between multiple processes, check to
1752          * make sure that those processes are still issuing I/Os within
1753          * the mean seek distance.  If not, it may be time to break the
1754          * queues apart again.
1755          */
1756         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1757                 cfq_mark_cfqq_split_coop(cfqq);
1758
1759         /*
1760          * store what was left of this slice, if the queue idled/timed out
1761          */
1762         if (timed_out) {
1763                 if (cfq_cfqq_slice_new(cfqq))
1764                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1765                 else
1766                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1767                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1768         }
1769
1770         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1771
1772         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1773                 cfq_del_cfqq_rr(cfqd, cfqq);
1774
1775         cfq_resort_rr_list(cfqd, cfqq);
1776
1777         if (cfqq == cfqd->active_queue)
1778                 cfqd->active_queue = NULL;
1779
1780         if (cfqd->active_cic) {
1781                 put_io_context(cfqd->active_cic->icq.ioc, cfqd->queue);
1782                 cfqd->active_cic = NULL;
1783         }
1784 }
1785
1786 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1787 {
1788         struct cfq_queue *cfqq = cfqd->active_queue;
1789
1790         if (cfqq)
1791                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1792 }
1793
1794 /*
1795  * Get next queue for service. Unless we have a queue preemption,
1796  * we'll simply select the first cfqq in the service tree.
1797  */
1798 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1799 {
1800         struct cfq_rb_root *service_tree =
1801                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1802                                         cfqd->serving_type);
1803
1804         if (!cfqd->rq_queued)
1805                 return NULL;
1806
1807         /* There is nothing to dispatch */
1808         if (!service_tree)
1809                 return NULL;
1810         if (RB_EMPTY_ROOT(&service_tree->rb))
1811                 return NULL;
1812         return cfq_rb_first(service_tree);
1813 }
1814
1815 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1816 {
1817         struct cfq_group *cfqg;
1818         struct cfq_queue *cfqq;
1819         int i, j;
1820         struct cfq_rb_root *st;
1821
1822         if (!cfqd->rq_queued)
1823                 return NULL;
1824
1825         cfqg = cfq_get_next_cfqg(cfqd);
1826         if (!cfqg)
1827                 return NULL;
1828
1829         for_each_cfqg_st(cfqg, i, j, st)
1830                 if ((cfqq = cfq_rb_first(st)) != NULL)
1831                         return cfqq;
1832         return NULL;
1833 }
1834
1835 /*
1836  * Get and set a new active queue for service.
1837  */
1838 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1839                                               struct cfq_queue *cfqq)
1840 {
1841         if (!cfqq)
1842                 cfqq = cfq_get_next_queue(cfqd);
1843
1844         __cfq_set_active_queue(cfqd, cfqq);
1845         return cfqq;
1846 }
1847
1848 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1849                                           struct request *rq)
1850 {
1851         if (blk_rq_pos(rq) >= cfqd->last_position)
1852                 return blk_rq_pos(rq) - cfqd->last_position;
1853         else
1854                 return cfqd->last_position - blk_rq_pos(rq);
1855 }
1856
1857 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1858                                struct request *rq)
1859 {
1860         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1861 }
1862
1863 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1864                                     struct cfq_queue *cur_cfqq)
1865 {
1866         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1867         struct rb_node *parent, *node;
1868         struct cfq_queue *__cfqq;
1869         sector_t sector = cfqd->last_position;
1870
1871         if (RB_EMPTY_ROOT(root))
1872                 return NULL;
1873
1874         /*
1875          * First, if we find a request starting at the end of the last
1876          * request, choose it.
1877          */
1878         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1879         if (__cfqq)
1880                 return __cfqq;
1881
1882         /*
1883          * If the exact sector wasn't found, the parent of the NULL leaf
1884          * will contain the closest sector.
1885          */
1886         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1887         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1888                 return __cfqq;
1889
1890         if (blk_rq_pos(__cfqq->next_rq) < sector)
1891                 node = rb_next(&__cfqq->p_node);
1892         else
1893                 node = rb_prev(&__cfqq->p_node);
1894         if (!node)
1895                 return NULL;
1896
1897         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1898         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1899                 return __cfqq;
1900
1901         return NULL;
1902 }
1903
1904 /*
1905  * cfqd - obvious
1906  * cur_cfqq - passed in so that we don't decide that the current queue is
1907  *            closely cooperating with itself.
1908  *
1909  * So, basically we're assuming that that cur_cfqq has dispatched at least
1910  * one request, and that cfqd->last_position reflects a position on the disk
1911  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1912  * assumption.
1913  */
1914 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1915                                               struct cfq_queue *cur_cfqq)
1916 {
1917         struct cfq_queue *cfqq;
1918
1919         if (cfq_class_idle(cur_cfqq))
1920                 return NULL;
1921         if (!cfq_cfqq_sync(cur_cfqq))
1922                 return NULL;
1923         if (CFQQ_SEEKY(cur_cfqq))
1924                 return NULL;
1925
1926         /*
1927          * Don't search priority tree if it's the only queue in the group.
1928          */
1929         if (cur_cfqq->cfqg->nr_cfqq == 1)
1930                 return NULL;
1931
1932         /*
1933          * We should notice if some of the queues are cooperating, eg
1934          * working closely on the same area of the disk. In that case,
1935          * we can group them together and don't waste time idling.
1936          */
1937         cfqq = cfqq_close(cfqd, cur_cfqq);
1938         if (!cfqq)
1939                 return NULL;
1940
1941         /* If new queue belongs to different cfq_group, don't choose it */
1942         if (cur_cfqq->cfqg != cfqq->cfqg)
1943                 return NULL;
1944
1945         /*
1946          * It only makes sense to merge sync queues.
1947          */
1948         if (!cfq_cfqq_sync(cfqq))
1949                 return NULL;
1950         if (CFQQ_SEEKY(cfqq))
1951                 return NULL;
1952
1953         /*
1954          * Do not merge queues of different priority classes
1955          */
1956         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1957                 return NULL;
1958
1959         return cfqq;
1960 }
1961
1962 /*
1963  * Determine whether we should enforce idle window for this queue.
1964  */
1965
1966 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1967 {
1968         enum wl_prio_t prio = cfqq_prio(cfqq);
1969         struct cfq_rb_root *service_tree = cfqq->service_tree;
1970
1971         BUG_ON(!service_tree);
1972         BUG_ON(!service_tree->count);
1973
1974         if (!cfqd->cfq_slice_idle)
1975                 return false;
1976
1977         /* We never do for idle class queues. */
1978         if (prio == IDLE_WORKLOAD)
1979                 return false;
1980
1981         /* We do for queues that were marked with idle window flag. */
1982         if (cfq_cfqq_idle_window(cfqq) &&
1983            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1984                 return true;
1985
1986         /*
1987          * Otherwise, we do only if they are the last ones
1988          * in their service tree.
1989          */
1990         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1991            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1992                 return true;
1993         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1994                         service_tree->count);
1995         return false;
1996 }
1997
1998 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1999 {
2000         struct cfq_queue *cfqq = cfqd->active_queue;
2001         struct cfq_io_cq *cic;
2002         unsigned long sl, group_idle = 0;
2003
2004         /*
2005          * SSD device without seek penalty, disable idling. But only do so
2006          * for devices that support queuing, otherwise we still have a problem
2007          * with sync vs async workloads.
2008          */
2009         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2010                 return;
2011
2012         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2013         WARN_ON(cfq_cfqq_slice_new(cfqq));
2014
2015         /*
2016          * idle is disabled, either manually or by past process history
2017          */
2018         if (!cfq_should_idle(cfqd, cfqq)) {
2019                 /* no queue idling. Check for group idling */
2020                 if (cfqd->cfq_group_idle)
2021                         group_idle = cfqd->cfq_group_idle;
2022                 else
2023                         return;
2024         }
2025
2026         /*
2027          * still active requests from this queue, don't idle
2028          */
2029         if (cfqq->dispatched)
2030                 return;
2031
2032         /*
2033          * task has exited, don't wait
2034          */
2035         cic = cfqd->active_cic;
2036         if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
2037                 return;
2038
2039         /*
2040          * If our average think time is larger than the remaining time
2041          * slice, then don't idle. This avoids overrunning the allotted
2042          * time slice.
2043          */
2044         if (sample_valid(cic->ttime.ttime_samples) &&
2045             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2046                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2047                              cic->ttime.ttime_mean);
2048                 return;
2049         }
2050
2051         /* There are other queues in the group, don't do group idle */
2052         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2053                 return;
2054
2055         cfq_mark_cfqq_wait_request(cfqq);
2056
2057         if (group_idle)
2058                 sl = cfqd->cfq_group_idle;
2059         else
2060                 sl = cfqd->cfq_slice_idle;
2061
2062         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2063         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2064         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2065                         group_idle ? 1 : 0);
2066 }
2067
2068 /*
2069  * Move request from internal lists to the request queue dispatch list.
2070  */
2071 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2072 {
2073         struct cfq_data *cfqd = q->elevator->elevator_data;
2074         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2075
2076         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2077
2078         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2079         cfq_remove_request(rq);
2080         cfqq->dispatched++;
2081         (RQ_CFQG(rq))->dispatched++;
2082         elv_dispatch_sort(q, rq);
2083
2084         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2085         cfqq->nr_sectors += blk_rq_sectors(rq);
2086         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2087                                         rq_data_dir(rq), rq_is_sync(rq));
2088 }
2089
2090 /*
2091  * return expired entry, or NULL to just start from scratch in rbtree
2092  */
2093 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2094 {
2095         struct request *rq = NULL;
2096
2097         if (cfq_cfqq_fifo_expire(cfqq))
2098                 return NULL;
2099
2100         cfq_mark_cfqq_fifo_expire(cfqq);
2101
2102         if (list_empty(&cfqq->fifo))
2103                 return NULL;
2104
2105         rq = rq_entry_fifo(cfqq->fifo.next);
2106         if (time_before(jiffies, rq_fifo_time(rq)))
2107                 rq = NULL;
2108
2109         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2110         return rq;
2111 }
2112
2113 static inline int
2114 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2115 {
2116         const int base_rq = cfqd->cfq_slice_async_rq;
2117
2118         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2119
2120         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2121 }
2122
2123 /*
2124  * Must be called with the queue_lock held.
2125  */
2126 static int cfqq_process_refs(struct cfq_queue *cfqq)
2127 {
2128         int process_refs, io_refs;
2129
2130         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2131         process_refs = cfqq->ref - io_refs;
2132         BUG_ON(process_refs < 0);
2133         return process_refs;
2134 }
2135
2136 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2137 {
2138         int process_refs, new_process_refs;
2139         struct cfq_queue *__cfqq;
2140
2141         /*
2142          * If there are no process references on the new_cfqq, then it is
2143          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2144          * chain may have dropped their last reference (not just their
2145          * last process reference).
2146          */
2147         if (!cfqq_process_refs(new_cfqq))
2148                 return;
2149
2150         /* Avoid a circular list and skip interim queue merges */
2151         while ((__cfqq = new_cfqq->new_cfqq)) {
2152                 if (__cfqq == cfqq)
2153                         return;
2154                 new_cfqq = __cfqq;
2155         }
2156
2157         process_refs = cfqq_process_refs(cfqq);
2158         new_process_refs = cfqq_process_refs(new_cfqq);
2159         /*
2160          * If the process for the cfqq has gone away, there is no
2161          * sense in merging the queues.
2162          */
2163         if (process_refs == 0 || new_process_refs == 0)
2164                 return;
2165
2166         /*
2167          * Merge in the direction of the lesser amount of work.
2168          */
2169         if (new_process_refs >= process_refs) {
2170                 cfqq->new_cfqq = new_cfqq;
2171                 new_cfqq->ref += process_refs;
2172         } else {
2173                 new_cfqq->new_cfqq = cfqq;
2174                 cfqq->ref += new_process_refs;
2175         }
2176 }
2177
2178 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2179                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2180 {
2181         struct cfq_queue *queue;
2182         int i;
2183         bool key_valid = false;
2184         unsigned long lowest_key = 0;
2185         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2186
2187         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2188                 /* select the one with lowest rb_key */
2189                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2190                 if (queue &&
2191                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2192                         lowest_key = queue->rb_key;
2193                         cur_best = i;
2194                         key_valid = true;
2195                 }
2196         }
2197
2198         return cur_best;
2199 }
2200
2201 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2202 {
2203         unsigned slice;
2204         unsigned count;
2205         struct cfq_rb_root *st;
2206         unsigned group_slice;
2207         enum wl_prio_t original_prio = cfqd->serving_prio;
2208
2209         /* Choose next priority. RT > BE > IDLE */
2210         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2211                 cfqd->serving_prio = RT_WORKLOAD;
2212         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2213                 cfqd->serving_prio = BE_WORKLOAD;
2214         else {
2215                 cfqd->serving_prio = IDLE_WORKLOAD;
2216                 cfqd->workload_expires = jiffies + 1;
2217                 return;
2218         }
2219
2220         if (original_prio != cfqd->serving_prio)
2221                 goto new_workload;
2222
2223         /*
2224          * For RT and BE, we have to choose also the type
2225          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2226          * expiration time
2227          */
2228         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2229         count = st->count;
2230
2231         /*
2232          * check workload expiration, and that we still have other queues ready
2233          */
2234         if (count && !time_after(jiffies, cfqd->workload_expires))
2235                 return;
2236
2237 new_workload:
2238         /* otherwise select new workload type */
2239         cfqd->serving_type =
2240                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2241         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2242         count = st->count;
2243
2244         /*
2245          * the workload slice is computed as a fraction of target latency
2246          * proportional to the number of queues in that workload, over
2247          * all the queues in the same priority class
2248          */
2249         group_slice = cfq_group_slice(cfqd, cfqg);
2250
2251         slice = group_slice * count /
2252                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2253                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2254
2255         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2256                 unsigned int tmp;
2257
2258                 /*
2259                  * Async queues are currently system wide. Just taking
2260                  * proportion of queues with-in same group will lead to higher
2261                  * async ratio system wide as generally root group is going
2262                  * to have higher weight. A more accurate thing would be to
2263                  * calculate system wide asnc/sync ratio.
2264                  */
2265                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2266                 tmp = tmp/cfqd->busy_queues;
2267                 slice = min_t(unsigned, slice, tmp);
2268
2269                 /* async workload slice is scaled down according to
2270                  * the sync/async slice ratio. */
2271                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2272         } else
2273                 /* sync workload slice is at least 2 * cfq_slice_idle */
2274                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2275
2276         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2277         cfq_log(cfqd, "workload slice:%d", slice);
2278         cfqd->workload_expires = jiffies + slice;
2279 }
2280
2281 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2282 {
2283         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2284         struct cfq_group *cfqg;
2285
2286         if (RB_EMPTY_ROOT(&st->rb))
2287                 return NULL;
2288         cfqg = cfq_rb_first_group(st);
2289         update_min_vdisktime(st);
2290         return cfqg;
2291 }
2292
2293 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2294 {
2295         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2296
2297         cfqd->serving_group = cfqg;
2298
2299         /* Restore the workload type data */
2300         if (cfqg->saved_workload_slice) {
2301                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2302                 cfqd->serving_type = cfqg->saved_workload;
2303                 cfqd->serving_prio = cfqg->saved_serving_prio;
2304         } else
2305                 cfqd->workload_expires = jiffies - 1;
2306
2307         choose_service_tree(cfqd, cfqg);
2308 }
2309
2310 /*
2311  * Select a queue for service. If we have a current active queue,
2312  * check whether to continue servicing it, or retrieve and set a new one.
2313  */
2314 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2315 {
2316         struct cfq_queue *cfqq, *new_cfqq = NULL;
2317
2318         cfqq = cfqd->active_queue;
2319         if (!cfqq)
2320                 goto new_queue;
2321
2322         if (!cfqd->rq_queued)
2323                 return NULL;
2324
2325         /*
2326          * We were waiting for group to get backlogged. Expire the queue
2327          */
2328         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2329                 goto expire;
2330
2331         /*
2332          * The active queue has run out of time, expire it and select new.
2333          */
2334         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2335                 /*
2336                  * If slice had not expired at the completion of last request
2337                  * we might not have turned on wait_busy flag. Don't expire
2338                  * the queue yet. Allow the group to get backlogged.
2339                  *
2340                  * The very fact that we have used the slice, that means we
2341                  * have been idling all along on this queue and it should be
2342                  * ok to wait for this request to complete.
2343                  */
2344                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2345                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2346                         cfqq = NULL;
2347                         goto keep_queue;
2348                 } else
2349                         goto check_group_idle;
2350         }
2351
2352         /*
2353          * The active queue has requests and isn't expired, allow it to
2354          * dispatch.
2355          */
2356         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2357                 goto keep_queue;
2358
2359         /*
2360          * If another queue has a request waiting within our mean seek
2361          * distance, let it run.  The expire code will check for close
2362          * cooperators and put the close queue at the front of the service
2363          * tree.  If possible, merge the expiring queue with the new cfqq.
2364          */
2365         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2366         if (new_cfqq) {
2367                 if (!cfqq->new_cfqq)
2368                         cfq_setup_merge(cfqq, new_cfqq);
2369                 goto expire;
2370         }
2371
2372         /*
2373          * No requests pending. If the active queue still has requests in
2374          * flight or is idling for a new request, allow either of these
2375          * conditions to happen (or time out) before selecting a new queue.
2376          */
2377         if (timer_pending(&cfqd->idle_slice_timer)) {
2378                 cfqq = NULL;
2379                 goto keep_queue;
2380         }
2381
2382         /*
2383          * This is a deep seek queue, but the device is much faster than
2384          * the queue can deliver, don't idle
2385          **/
2386         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2387             (cfq_cfqq_slice_new(cfqq) ||
2388             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2389                 cfq_clear_cfqq_deep(cfqq);
2390                 cfq_clear_cfqq_idle_window(cfqq);
2391         }
2392
2393         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2394                 cfqq = NULL;
2395                 goto keep_queue;
2396         }
2397
2398         /*
2399          * If group idle is enabled and there are requests dispatched from
2400          * this group, wait for requests to complete.
2401          */
2402 check_group_idle:
2403         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2404             cfqq->cfqg->dispatched &&
2405             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2406                 cfqq = NULL;
2407                 goto keep_queue;
2408         }
2409
2410 expire:
2411         cfq_slice_expired(cfqd, 0);
2412 new_queue:
2413         /*
2414          * Current queue expired. Check if we have to switch to a new
2415          * service tree
2416          */
2417         if (!new_cfqq)
2418                 cfq_choose_cfqg(cfqd);
2419
2420         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2421 keep_queue:
2422         return cfqq;
2423 }
2424
2425 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2426 {
2427         int dispatched = 0;
2428
2429         while (cfqq->next_rq) {
2430                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2431                 dispatched++;
2432         }
2433
2434         BUG_ON(!list_empty(&cfqq->fifo));
2435
2436         /* By default cfqq is not expired if it is empty. Do it explicitly */
2437         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2438         return dispatched;
2439 }
2440
2441 /*
2442  * Drain our current requests. Used for barriers and when switching
2443  * io schedulers on-the-fly.
2444  */
2445 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2446 {
2447         struct cfq_queue *cfqq;
2448         int dispatched = 0;
2449
2450         /* Expire the timeslice of the current active queue first */
2451         cfq_slice_expired(cfqd, 0);
2452         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2453                 __cfq_set_active_queue(cfqd, cfqq);
2454                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2455         }
2456
2457         BUG_ON(cfqd->busy_queues);
2458
2459         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2460         return dispatched;
2461 }
2462
2463 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2464         struct cfq_queue *cfqq)
2465 {
2466         /* the queue hasn't finished any request, can't estimate */
2467         if (cfq_cfqq_slice_new(cfqq))
2468                 return true;
2469         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2470                 cfqq->slice_end))
2471                 return true;
2472
2473         return false;
2474 }
2475
2476 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2477 {
2478         unsigned int max_dispatch;
2479
2480         /*
2481          * Drain async requests before we start sync IO
2482          */
2483         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2484                 return false;
2485
2486         /*
2487          * If this is an async queue and we have sync IO in flight, let it wait
2488          */
2489         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2490                 return false;
2491
2492         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2493         if (cfq_class_idle(cfqq))
2494                 max_dispatch = 1;
2495
2496         /*
2497          * Does this cfqq already have too much IO in flight?
2498          */
2499         if (cfqq->dispatched >= max_dispatch) {
2500                 bool promote_sync = false;
2501                 /*
2502                  * idle queue must always only have a single IO in flight
2503                  */
2504                 if (cfq_class_idle(cfqq))
2505                         return false;
2506
2507                 /*
2508                  * If there is only one sync queue
2509                  * we can ignore async queue here and give the sync
2510                  * queue no dispatch limit. The reason is a sync queue can
2511                  * preempt async queue, limiting the sync queue doesn't make
2512                  * sense. This is useful for aiostress test.
2513                  */
2514                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2515                         promote_sync = true;
2516
2517                 /*
2518                  * We have other queues, don't allow more IO from this one
2519                  */
2520                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2521                                 !promote_sync)
2522                         return false;
2523
2524                 /*
2525                  * Sole queue user, no limit
2526                  */
2527                 if (cfqd->busy_queues == 1 || promote_sync)
2528                         max_dispatch = -1;
2529                 else
2530                         /*
2531                          * Normally we start throttling cfqq when cfq_quantum/2
2532                          * requests have been dispatched. But we can drive
2533                          * deeper queue depths at the beginning of slice
2534                          * subjected to upper limit of cfq_quantum.
2535                          * */
2536                         max_dispatch = cfqd->cfq_quantum;
2537         }
2538
2539         /*
2540          * Async queues must wait a bit before being allowed dispatch.
2541          * We also ramp up the dispatch depth gradually for async IO,
2542          * based on the last sync IO we serviced
2543          */
2544         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2545                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2546                 unsigned int depth;
2547
2548                 depth = last_sync / cfqd->cfq_slice[1];
2549                 if (!depth && !cfqq->dispatched)
2550                         depth = 1;
2551                 if (depth < max_dispatch)
2552                         max_dispatch = depth;
2553         }
2554
2555         /*
2556          * If we're below the current max, allow a dispatch
2557          */
2558         return cfqq->dispatched < max_dispatch;
2559 }
2560
2561 /*
2562  * Dispatch a request from cfqq, moving them to the request queue
2563  * dispatch list.
2564  */
2565 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2566 {
2567         struct request *rq;
2568
2569         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2570
2571         if (!cfq_may_dispatch(cfqd, cfqq))
2572                 return false;
2573
2574         /*
2575          * follow expired path, else get first next available
2576          */
2577         rq = cfq_check_fifo(cfqq);
2578         if (!rq)
2579                 rq = cfqq->next_rq;
2580
2581         /*
2582          * insert request into driver dispatch list
2583          */
2584         cfq_dispatch_insert(cfqd->queue, rq);
2585
2586         if (!cfqd->active_cic) {
2587                 struct cfq_io_cq *cic = RQ_CIC(rq);
2588
2589                 atomic_long_inc(&cic->icq.ioc->refcount);
2590                 cfqd->active_cic = cic;
2591         }
2592
2593         return true;
2594 }
2595
2596 /*
2597  * Find the cfqq that we need to service and move a request from that to the
2598  * dispatch list
2599  */
2600 static int cfq_dispatch_requests(struct request_queue *q, int force)
2601 {
2602         struct cfq_data *cfqd = q->elevator->elevator_data;
2603         struct cfq_queue *cfqq;
2604
2605         if (!cfqd->busy_queues)
2606                 return 0;
2607
2608         if (unlikely(force))
2609                 return cfq_forced_dispatch(cfqd);
2610
2611         cfqq = cfq_select_queue(cfqd);
2612         if (!cfqq)
2613                 return 0;
2614
2615         /*
2616          * Dispatch a request from this cfqq, if it is allowed
2617          */
2618         if (!cfq_dispatch_request(cfqd, cfqq))
2619                 return 0;
2620
2621         cfqq->slice_dispatch++;
2622         cfq_clear_cfqq_must_dispatch(cfqq);
2623
2624         /*
2625          * expire an async queue immediately if it has used up its slice. idle
2626          * queue always expire after 1 dispatch round.
2627          */
2628         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2629             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2630             cfq_class_idle(cfqq))) {
2631                 cfqq->slice_end = jiffies + 1;
2632                 cfq_slice_expired(cfqd, 0);
2633         }
2634
2635         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2636         return 1;
2637 }
2638
2639 /*
2640  * task holds one reference to the queue, dropped when task exits. each rq
2641  * in-flight on this queue also holds a reference, dropped when rq is freed.
2642  *
2643  * Each cfq queue took a reference on the parent group. Drop it now.
2644  * queue lock must be held here.
2645  */
2646 static void cfq_put_queue(struct cfq_queue *cfqq)
2647 {
2648         struct cfq_data *cfqd = cfqq->cfqd;
2649         struct cfq_group *cfqg;
2650
2651         BUG_ON(cfqq->ref <= 0);
2652
2653         cfqq->ref--;
2654         if (cfqq->ref)
2655                 return;
2656
2657         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2658         BUG_ON(rb_first(&cfqq->sort_list));
2659         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2660         cfqg = cfqq->cfqg;
2661
2662         if (unlikely(cfqd->active_queue == cfqq)) {
2663                 __cfq_slice_expired(cfqd, cfqq, 0);
2664                 cfq_schedule_dispatch(cfqd);
2665         }
2666
2667         BUG_ON(cfq_cfqq_on_rr(cfqq));
2668         kmem_cache_free(cfq_pool, cfqq);
2669         cfq_put_cfqg(cfqg);
2670 }
2671
2672 static void cfq_icq_free_rcu(struct rcu_head *head)
2673 {
2674         kmem_cache_free(cfq_icq_pool,
2675                         icq_to_cic(container_of(head, struct io_cq, rcu_head)));
2676 }
2677
2678 static void cfq_icq_free(struct io_cq *icq)
2679 {
2680         call_rcu(&icq->rcu_head, cfq_icq_free_rcu);
2681 }
2682
2683 static void cfq_release_icq(struct io_cq *icq)
2684 {
2685         struct io_context *ioc = icq->ioc;
2686
2687         radix_tree_delete(&ioc->icq_tree, icq->q->id);
2688         hlist_del(&icq->ioc_node);
2689         cfq_icq_free(icq);
2690 }
2691
2692 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2693 {
2694         struct cfq_queue *__cfqq, *next;
2695
2696         /*
2697          * If this queue was scheduled to merge with another queue, be
2698          * sure to drop the reference taken on that queue (and others in
2699          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2700          */
2701         __cfqq = cfqq->new_cfqq;
2702         while (__cfqq) {
2703                 if (__cfqq == cfqq) {
2704                         WARN(1, "cfqq->new_cfqq loop detected\n");
2705                         break;
2706                 }
2707                 next = __cfqq->new_cfqq;
2708                 cfq_put_queue(__cfqq);
2709                 __cfqq = next;
2710         }
2711 }
2712
2713 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2714 {
2715         if (unlikely(cfqq == cfqd->active_queue)) {
2716                 __cfq_slice_expired(cfqd, cfqq, 0);
2717                 cfq_schedule_dispatch(cfqd);
2718         }
2719
2720         cfq_put_cooperator(cfqq);
2721
2722         cfq_put_queue(cfqq);
2723 }
2724
2725 static void cfq_exit_icq(struct io_cq *icq)
2726 {
2727         struct cfq_io_cq *cic = icq_to_cic(icq);
2728         struct cfq_data *cfqd = cic_to_cfqd(cic);
2729         struct io_context *ioc = icq->ioc;
2730
2731         list_del_init(&icq->q_node);
2732
2733         /*
2734          * Both setting lookup hint to and clearing it from @icq are done
2735          * under queue_lock.  If it's not pointing to @icq now, it never
2736          * will.  Hint assignment itself can race safely.
2737          */
2738         if (rcu_dereference_raw(ioc->icq_hint) == icq)
2739                 rcu_assign_pointer(ioc->icq_hint, NULL);
2740
2741         if (cic->cfqq[BLK_RW_ASYNC]) {
2742                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2743                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2744         }
2745
2746         if (cic->cfqq[BLK_RW_SYNC]) {
2747                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2748                 cic->cfqq[BLK_RW_SYNC] = NULL;
2749         }
2750 }
2751
2752 static struct cfq_io_cq *cfq_alloc_cic(struct cfq_data *cfqd, gfp_t gfp_mask)
2753 {
2754         struct cfq_io_cq *cic;
2755
2756         cic = kmem_cache_alloc_node(cfq_icq_pool, gfp_mask | __GFP_ZERO,
2757                                                         cfqd->queue->node);
2758         if (cic) {
2759                 cic->ttime.last_end_request = jiffies;
2760                 INIT_LIST_HEAD(&cic->icq.q_node);
2761                 INIT_HLIST_NODE(&cic->icq.ioc_node);
2762                 cic->icq.exit = cfq_exit_icq;
2763                 cic->icq.release = cfq_release_icq;
2764         }
2765
2766         return cic;
2767 }
2768
2769 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2770 {
2771         struct task_struct *tsk = current;
2772         int ioprio_class;
2773
2774         if (!cfq_cfqq_prio_changed(cfqq))
2775                 return;
2776
2777         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2778         switch (ioprio_class) {
2779         default:
2780                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2781         case IOPRIO_CLASS_NONE:
2782                 /*
2783                  * no prio set, inherit CPU scheduling settings
2784                  */
2785                 cfqq->ioprio = task_nice_ioprio(tsk);
2786                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2787                 break;
2788         case IOPRIO_CLASS_RT:
2789                 cfqq->ioprio = task_ioprio(ioc);
2790                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2791                 break;
2792         case IOPRIO_CLASS_BE:
2793                 cfqq->ioprio = task_ioprio(ioc);
2794                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2795                 break;
2796         case IOPRIO_CLASS_IDLE:
2797                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2798                 cfqq->ioprio = 7;
2799                 cfq_clear_cfqq_idle_window(cfqq);
2800                 break;
2801         }
2802
2803         /*
2804          * keep track of original prio settings in case we have to temporarily
2805          * elevate the priority of this queue
2806          */
2807         cfqq->org_ioprio = cfqq->ioprio;
2808         cfq_clear_cfqq_prio_changed(cfqq);
2809 }
2810
2811 static void changed_ioprio(struct cfq_io_cq *cic)
2812 {
2813         struct cfq_data *cfqd = cic_to_cfqd(cic);
2814         struct cfq_queue *cfqq;
2815
2816         if (unlikely(!cfqd))
2817                 return;
2818
2819         cfqq = cic->cfqq[BLK_RW_ASYNC];
2820         if (cfqq) {
2821                 struct cfq_queue *new_cfqq;
2822                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2823                                                 GFP_ATOMIC);
2824                 if (new_cfqq) {
2825                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2826                         cfq_put_queue(cfqq);
2827                 }
2828         }
2829
2830         cfqq = cic->cfqq[BLK_RW_SYNC];
2831         if (cfqq)
2832                 cfq_mark_cfqq_prio_changed(cfqq);
2833 }
2834
2835 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2836                           pid_t pid, bool is_sync)
2837 {
2838         RB_CLEAR_NODE(&cfqq->rb_node);
2839         RB_CLEAR_NODE(&cfqq->p_node);
2840         INIT_LIST_HEAD(&cfqq->fifo);
2841
2842         cfqq->ref = 0;
2843         cfqq->cfqd = cfqd;
2844
2845         cfq_mark_cfqq_prio_changed(cfqq);
2846
2847         if (is_sync) {
2848                 if (!cfq_class_idle(cfqq))
2849                         cfq_mark_cfqq_idle_window(cfqq);
2850                 cfq_mark_cfqq_sync(cfqq);
2851         }
2852         cfqq->pid = pid;
2853 }
2854
2855 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2856 static void changed_cgroup(struct cfq_io_cq *cic)
2857 {
2858         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2859         struct cfq_data *cfqd = cic_to_cfqd(cic);
2860         struct request_queue *q;
2861
2862         if (unlikely(!cfqd))
2863                 return;
2864
2865         q = cfqd->queue;
2866
2867         if (sync_cfqq) {
2868                 /*
2869                  * Drop reference to sync queue. A new sync queue will be
2870                  * assigned in new group upon arrival of a fresh request.
2871                  */
2872                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2873                 cic_set_cfqq(cic, NULL, 1);
2874                 cfq_put_queue(sync_cfqq);
2875         }
2876 }
2877 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2878
2879 static struct cfq_queue *
2880 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2881                      struct io_context *ioc, gfp_t gfp_mask)
2882 {
2883         struct cfq_queue *cfqq, *new_cfqq = NULL;
2884         struct cfq_io_cq *cic;
2885         struct cfq_group *cfqg;
2886
2887 retry:
2888         cfqg = cfq_get_cfqg(cfqd);
2889         cic = cfq_cic_lookup(cfqd, ioc);
2890         /* cic always exists here */
2891         cfqq = cic_to_cfqq(cic, is_sync);
2892
2893         /*
2894          * Always try a new alloc if we fell back to the OOM cfqq
2895          * originally, since it should just be a temporary situation.
2896          */
2897         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2898                 cfqq = NULL;
2899                 if (new_cfqq) {
2900                         cfqq = new_cfqq;
2901                         new_cfqq = NULL;
2902                 } else if (gfp_mask & __GFP_WAIT) {
2903                         spin_unlock_irq(cfqd->queue->queue_lock);
2904                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2905                                         gfp_mask | __GFP_ZERO,
2906                                         cfqd->queue->node);
2907                         spin_lock_irq(cfqd->queue->queue_lock);
2908                         if (new_cfqq)
2909                                 goto retry;
2910                 } else {
2911                         cfqq = kmem_cache_alloc_node(cfq_pool,
2912                                         gfp_mask | __GFP_ZERO,
2913                                         cfqd->queue->node);
2914                 }
2915
2916                 if (cfqq) {
2917                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2918                         cfq_init_prio_data(cfqq, ioc);
2919                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2920                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2921                 } else
2922                         cfqq = &cfqd->oom_cfqq;
2923         }
2924
2925         if (new_cfqq)
2926                 kmem_cache_free(cfq_pool, new_cfqq);
2927
2928         return cfqq;
2929 }
2930
2931 static struct cfq_queue **
2932 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2933 {
2934         switch (ioprio_class) {
2935         case IOPRIO_CLASS_RT:
2936                 return &cfqd->async_cfqq[0][ioprio];
2937         case IOPRIO_CLASS_BE:
2938                 return &cfqd->async_cfqq[1][ioprio];
2939         case IOPRIO_CLASS_IDLE:
2940                 return &cfqd->async_idle_cfqq;
2941         default:
2942                 BUG();
2943         }
2944 }
2945
2946 static struct cfq_queue *
2947 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2948               gfp_t gfp_mask)
2949 {
2950         const int ioprio = task_ioprio(ioc);
2951         const int ioprio_class = task_ioprio_class(ioc);
2952         struct cfq_queue **async_cfqq = NULL;
2953         struct cfq_queue *cfqq = NULL;
2954
2955         if (!is_sync) {
2956                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2957                 cfqq = *async_cfqq;
2958         }
2959
2960         if (!cfqq)
2961                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2962
2963         /*
2964          * pin the queue now that it's allocated, scheduler exit will prune it
2965          */
2966         if (!is_sync && !(*async_cfqq)) {
2967                 cfqq->ref++;
2968                 *async_cfqq = cfqq;
2969         }
2970
2971         cfqq->ref++;
2972         return cfqq;
2973 }
2974
2975 /**
2976  * cfq_cic_lookup - lookup cfq_io_cq
2977  * @cfqd: the associated cfq_data
2978  * @ioc: the associated io_context
2979  *
2980  * Look up cfq_io_cq associated with @cfqd - @ioc pair.  Must be called
2981  * with queue_lock held.
2982  */
2983 static struct cfq_io_cq *
2984 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2985 {
2986         struct request_queue *q = cfqd->queue;
2987         struct io_cq *icq;
2988
2989         lockdep_assert_held(cfqd->queue->queue_lock);
2990         if (unlikely(!ioc))
2991                 return NULL;
2992
2993         /*
2994          * icq's are indexed from @ioc using radix tree and hint pointer,
2995          * both of which are protected with RCU.  All removals are done
2996          * holding both q and ioc locks, and we're holding q lock - if we
2997          * find a icq which points to us, it's guaranteed to be valid.
2998          */
2999         rcu_read_lock();
3000         icq = rcu_dereference(ioc->icq_hint);
3001         if (icq && icq->q == q)
3002                 goto out;
3003
3004         icq = radix_tree_lookup(&ioc->icq_tree, cfqd->queue->id);
3005         if (icq && icq->q == q)
3006                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
3007         else
3008                 icq = NULL;
3009 out:
3010         rcu_read_unlock();
3011         return icq_to_cic(icq);
3012 }
3013
3014 /**
3015  * cfq_create_cic - create and link a cfq_io_cq
3016  * @cfqd: cfqd of interest
3017  * @gfp_mask: allocation mask
3018  *
3019  * Make sure cfq_io_cq linking %current->io_context and @cfqd exists.  If
3020  * ioc and/or cic doesn't exist, they will be created using @gfp_mask.
3021  */
3022 static int cfq_create_cic(struct cfq_data *cfqd, gfp_t gfp_mask)
3023 {
3024         struct request_queue *q = cfqd->queue;
3025         struct io_cq *icq = NULL;
3026         struct cfq_io_cq *cic;
3027         struct io_context *ioc;
3028         int ret = -ENOMEM;
3029
3030         might_sleep_if(gfp_mask & __GFP_WAIT);
3031
3032         /* allocate stuff */
3033         ioc = create_io_context(current, gfp_mask, q->node);
3034         if (!ioc)
3035                 goto out;
3036
3037         cic = cfq_alloc_cic(cfqd, gfp_mask);
3038         if (!cic)
3039                 goto out;
3040         icq = &cic->icq;
3041
3042         ret = radix_tree_preload(gfp_mask);
3043         if (ret)
3044                 goto out;
3045
3046         icq->ioc = ioc;
3047         icq->q = cfqd->queue;
3048
3049         /* lock both q and ioc and try to link @icq */
3050         spin_lock_irq(q->queue_lock);
3051         spin_lock(&ioc->lock);
3052
3053         ret = radix_tree_insert(&ioc->icq_tree, q->id, icq);
3054         if (likely(!ret)) {
3055                 hlist_add_head(&icq->ioc_node, &ioc->icq_list);
3056                 list_add(&icq->q_node, &cfqd->icq_list);
3057                 icq = NULL;
3058         } else if (ret == -EEXIST) {
3059                 /* someone else already did it */
3060                 ret = 0;
3061         }
3062
3063         spin_unlock(&ioc->lock);
3064         spin_unlock_irq(q->queue_lock);
3065
3066         radix_tree_preload_end();
3067 out:
3068         if (ret)
3069                 printk(KERN_ERR "cfq: icq link failed!\n");
3070         if (icq)
3071                 cfq_icq_free(icq);
3072         return ret;
3073 }
3074
3075 /**
3076  * cfq_get_cic - acquire cfq_io_cq and bump refcnt on io_context
3077  * @cfqd: cfqd to setup cic for
3078  * @gfp_mask: allocation mask
3079  *
3080  * Return cfq_io_cq associating @cfqd and %current->io_context and
3081  * bump refcnt on io_context.  If ioc or cic doesn't exist, they're created
3082  * using @gfp_mask.
3083  *
3084  * Must be called under queue_lock which may be released and re-acquired.
3085  * This function also may sleep depending on @gfp_mask.
3086  */
3087 static struct cfq_io_cq *cfq_get_cic(struct cfq_data *cfqd, gfp_t gfp_mask)
3088 {
3089         struct request_queue *q = cfqd->queue;
3090         struct cfq_io_cq *cic = NULL;
3091         struct io_context *ioc;
3092         int err;
3093
3094         lockdep_assert_held(q->queue_lock);
3095
3096         while (true) {
3097                 /* fast path */
3098                 ioc = current->io_context;
3099                 if (likely(ioc)) {
3100                         cic = cfq_cic_lookup(cfqd, ioc);
3101                         if (likely(cic))
3102                                 break;
3103                 }
3104
3105                 /* slow path - unlock, create missing ones and retry */
3106                 spin_unlock_irq(q->queue_lock);
3107                 err = cfq_create_cic(cfqd, gfp_mask);
3108                 spin_lock_irq(q->queue_lock);
3109                 if (err)
3110                         return NULL;
3111         }
3112
3113         /* bump @ioc's refcnt and handle changed notifications */
3114         get_io_context(ioc);
3115
3116         if (unlikely(cic->icq.changed)) {
3117                 if (test_and_clear_bit(ICQ_IOPRIO_CHANGED, &cic->icq.changed))
3118                         changed_ioprio(cic);
3119 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3120                 if (test_and_clear_bit(ICQ_CGROUP_CHANGED, &cic->icq.changed))
3121                         changed_cgroup(cic);
3122 #endif
3123         }
3124
3125         return cic;
3126 }
3127
3128 static void
3129 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3130 {
3131         unsigned long elapsed = jiffies - ttime->last_end_request;
3132         elapsed = min(elapsed, 2UL * slice_idle);
3133
3134         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3135         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3136         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3137 }
3138
3139 static void
3140 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3141                         struct cfq_io_cq *cic)
3142 {
3143         if (cfq_cfqq_sync(cfqq)) {
3144                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3145                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3146                         cfqd->cfq_slice_idle);
3147         }
3148 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3149         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3150 #endif
3151 }
3152
3153 static void
3154 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3155                        struct request *rq)
3156 {
3157         sector_t sdist = 0;
3158         sector_t n_sec = blk_rq_sectors(rq);
3159         if (cfqq->last_request_pos) {
3160                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3161                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3162                 else
3163                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3164         }
3165
3166         cfqq->seek_history <<= 1;
3167         if (blk_queue_nonrot(cfqd->queue))
3168                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3169         else
3170                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3171 }
3172
3173 /*
3174  * Disable idle window if the process thinks too long or seeks so much that
3175  * it doesn't matter
3176  */
3177 static void
3178 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3179                        struct cfq_io_cq *cic)
3180 {
3181         int old_idle, enable_idle;
3182
3183         /*
3184          * Don't idle for async or idle io prio class
3185          */
3186         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3187                 return;
3188
3189         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3190
3191         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3192                 cfq_mark_cfqq_deep(cfqq);
3193
3194         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3195                 enable_idle = 0;
3196         else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3197                  !cfqd->cfq_slice_idle ||
3198                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3199                 enable_idle = 0;
3200         else if (sample_valid(cic->ttime.ttime_samples)) {
3201                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3202                         enable_idle = 0;
3203                 else
3204                         enable_idle = 1;
3205         }
3206
3207         if (old_idle != enable_idle) {
3208                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3209                 if (enable_idle)
3210                         cfq_mark_cfqq_idle_window(cfqq);
3211                 else
3212                         cfq_clear_cfqq_idle_window(cfqq);
3213         }
3214 }
3215
3216 /*
3217  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3218  * no or if we aren't sure, a 1 will cause a preempt.
3219  */
3220 static bool
3221 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3222                    struct request *rq)
3223 {
3224         struct cfq_queue *cfqq;
3225
3226         cfqq = cfqd->active_queue;
3227         if (!cfqq)
3228                 return false;
3229
3230         if (cfq_class_idle(new_cfqq))
3231                 return false;
3232
3233         if (cfq_class_idle(cfqq))
3234                 return true;
3235
3236         /*
3237          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3238          */
3239         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3240                 return false;
3241
3242         /*
3243          * if the new request is sync, but the currently running queue is
3244          * not, let the sync request have priority.
3245          */
3246         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3247                 return true;
3248
3249         if (new_cfqq->cfqg != cfqq->cfqg)
3250                 return false;
3251
3252         if (cfq_slice_used(cfqq))
3253                 return true;
3254
3255         /* Allow preemption only if we are idling on sync-noidle tree */
3256         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3257             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3258             new_cfqq->service_tree->count == 2 &&
3259             RB_EMPTY_ROOT(&cfqq->sort_list))
3260                 return true;
3261
3262         /*
3263          * So both queues are sync. Let the new request get disk time if
3264          * it's a metadata request and the current queue is doing regular IO.
3265          */
3266         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3267                 return true;
3268
3269         /*
3270          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3271          */
3272         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3273                 return true;
3274
3275         /* An idle queue should not be idle now for some reason */
3276         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3277                 return true;
3278
3279         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3280                 return false;
3281
3282         /*
3283          * if this request is as-good as one we would expect from the
3284          * current cfqq, let it preempt
3285          */
3286         if (cfq_rq_close(cfqd, cfqq, rq))
3287                 return true;
3288
3289         return false;
3290 }
3291
3292 /*
3293  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3294  * let it have half of its nominal slice.
3295  */
3296 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3297 {
3298         struct cfq_queue *old_cfqq = cfqd->active_queue;
3299
3300         cfq_log_cfqq(cfqd, cfqq, "preempt");
3301         cfq_slice_expired(cfqd, 1);
3302
3303         /*
3304          * workload type is changed, don't save slice, otherwise preempt
3305          * doesn't happen
3306          */
3307         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3308                 cfqq->cfqg->saved_workload_slice = 0;
3309
3310         /*
3311          * Put the new queue at the front of the of the current list,
3312          * so we know that it will be selected next.
3313          */
3314         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3315
3316         cfq_service_tree_add(cfqd, cfqq, 1);
3317
3318         cfqq->slice_end = 0;
3319         cfq_mark_cfqq_slice_new(cfqq);
3320 }
3321
3322 /*
3323  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3324  * something we should do about it
3325  */
3326 static void
3327 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3328                 struct request *rq)
3329 {
3330         struct cfq_io_cq *cic = RQ_CIC(rq);
3331
3332         cfqd->rq_queued++;
3333         if (rq->cmd_flags & REQ_PRIO)
3334                 cfqq->prio_pending++;
3335
3336         cfq_update_io_thinktime(cfqd, cfqq, cic);
3337         cfq_update_io_seektime(cfqd, cfqq, rq);
3338         cfq_update_idle_window(cfqd, cfqq, cic);
3339
3340         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3341
3342         if (cfqq == cfqd->active_queue) {
3343                 /*
3344                  * Remember that we saw a request from this process, but
3345                  * don't start queuing just yet. Otherwise we risk seeing lots
3346                  * of tiny requests, because we disrupt the normal plugging
3347                  * and merging. If the request is already larger than a single
3348                  * page, let it rip immediately. For that case we assume that
3349                  * merging is already done. Ditto for a busy system that
3350                  * has other work pending, don't risk delaying until the
3351                  * idle timer unplug to continue working.
3352                  */
3353                 if (cfq_cfqq_wait_request(cfqq)) {
3354                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3355                             cfqd->busy_queues > 1) {
3356                                 cfq_del_timer(cfqd, cfqq);
3357                                 cfq_clear_cfqq_wait_request(cfqq);
3358                                 __blk_run_queue(cfqd->queue);
3359                         } else {
3360                                 cfq_blkiocg_update_idle_time_stats(
3361                                                 &cfqq->cfqg->blkg);
3362                                 cfq_mark_cfqq_must_dispatch(cfqq);
3363                         }
3364                 }
3365         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3366                 /*
3367                  * not the active queue - expire current slice if it is
3368                  * idle and has expired it's mean thinktime or this new queue
3369                  * has some old slice time left and is of higher priority or
3370                  * this new queue is RT and the current one is BE
3371                  */
3372                 cfq_preempt_queue(cfqd, cfqq);
3373                 __blk_run_queue(cfqd->queue);
3374         }
3375 }
3376
3377 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3378 {
3379         struct cfq_data *cfqd = q->elevator->elevator_data;
3380         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3381
3382         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3383         cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3384
3385         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3386         list_add_tail(&rq->queuelist, &cfqq->fifo);
3387         cfq_add_rq_rb(rq);
3388         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3389                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3390                         rq_is_sync(rq));
3391         cfq_rq_enqueued(cfqd, cfqq, rq);
3392 }
3393
3394 /*
3395  * Update hw_tag based on peak queue depth over 50 samples under
3396  * sufficient load.
3397  */
3398 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3399 {
3400         struct cfq_queue *cfqq = cfqd->active_queue;
3401
3402         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3403                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3404
3405         if (cfqd->hw_tag == 1)
3406                 return;
3407
3408         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3409             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3410                 return;
3411
3412         /*
3413          * If active queue hasn't enough requests and can idle, cfq might not
3414          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3415          * case
3416          */
3417         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3418             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3419             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3420                 return;
3421
3422         if (cfqd->hw_tag_samples++ < 50)
3423                 return;
3424
3425         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3426                 cfqd->hw_tag = 1;
3427         else
3428                 cfqd->hw_tag = 0;
3429 }
3430
3431 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3432 {
3433         struct cfq_io_cq *cic = cfqd->active_cic;
3434
3435         /* If the queue already has requests, don't wait */
3436         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3437                 return false;
3438
3439         /* If there are other queues in the group, don't wait */
3440         if (cfqq->cfqg->nr_cfqq > 1)
3441                 return false;
3442
3443         /* the only queue in the group, but think time is big */
3444         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3445                 return false;
3446
3447         if (cfq_slice_used(cfqq))
3448                 return true;
3449
3450         /* if slice left is less than think time, wait busy */
3451         if (cic && sample_valid(cic->ttime.ttime_samples)
3452             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3453                 return true;
3454
3455         /*
3456          * If think times is less than a jiffy than ttime_mean=0 and above
3457          * will not be true. It might happen that slice has not expired yet
3458          * but will expire soon (4-5 ns) during select_queue(). To cover the
3459          * case where think time is less than a jiffy, mark the queue wait
3460          * busy if only 1 jiffy is left in the slice.
3461          */
3462         if (cfqq->slice_end - jiffies == 1)
3463                 return true;
3464
3465         return false;
3466 }
3467
3468 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3469 {
3470         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3471         struct cfq_data *cfqd = cfqq->cfqd;
3472         const int sync = rq_is_sync(rq);
3473         unsigned long now;
3474
3475         now = jiffies;
3476         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3477                      !!(rq->cmd_flags & REQ_NOIDLE));
3478
3479         cfq_update_hw_tag(cfqd);
3480
3481         WARN_ON(!cfqd->rq_in_driver);
3482         WARN_ON(!cfqq->dispatched);
3483         cfqd->rq_in_driver--;
3484         cfqq->dispatched--;
3485         (RQ_CFQG(rq))->dispatched--;
3486         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3487                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3488                         rq_data_dir(rq), rq_is_sync(rq));
3489
3490         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3491
3492         if (sync) {
3493                 struct cfq_rb_root *service_tree;
3494
3495                 RQ_CIC(rq)->ttime.last_end_request = now;
3496
3497                 if (cfq_cfqq_on_rr(cfqq))
3498                         service_tree = cfqq->service_tree;
3499                 else
3500                         service_tree = service_tree_for(cfqq->cfqg,
3501                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3502                 service_tree->ttime.last_end_request = now;
3503                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3504                         cfqd->last_delayed_sync = now;
3505         }
3506
3507 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3508         cfqq->cfqg->ttime.last_end_request = now;
3509 #endif
3510
3511         /*
3512          * If this is the active queue, check if it needs to be expired,
3513          * or if we want to idle in case it has no pending requests.
3514          */
3515         if (cfqd->active_queue == cfqq) {
3516                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3517
3518                 if (cfq_cfqq_slice_new(cfqq)) {
3519                         cfq_set_prio_slice(cfqd, cfqq);
3520                         cfq_clear_cfqq_slice_new(cfqq);
3521                 }
3522
3523                 /*
3524                  * Should we wait for next request to come in before we expire
3525                  * the queue.
3526                  */
3527                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3528                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3529                         if (!cfqd->cfq_slice_idle)
3530                                 extend_sl = cfqd->cfq_group_idle;
3531                         cfqq->slice_end = jiffies + extend_sl;
3532                         cfq_mark_cfqq_wait_busy(cfqq);
3533                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3534                 }
3535
3536                 /*
3537                  * Idling is not enabled on:
3538                  * - expired queues
3539                  * - idle-priority queues
3540                  * - async queues
3541                  * - queues with still some requests queued
3542                  * - when there is a close cooperator
3543                  */
3544                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3545                         cfq_slice_expired(cfqd, 1);
3546                 else if (sync && cfqq_empty &&
3547                          !cfq_close_cooperator(cfqd, cfqq)) {
3548                         cfq_arm_slice_timer(cfqd);
3549                 }
3550         }
3551
3552         if (!cfqd->rq_in_driver)
3553                 cfq_schedule_dispatch(cfqd);
3554 }
3555
3556 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3557 {
3558         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3559                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3560                 return ELV_MQUEUE_MUST;
3561         }
3562
3563         return ELV_MQUEUE_MAY;
3564 }
3565
3566 static int cfq_may_queue(struct request_queue *q, int rw)
3567 {
3568         struct cfq_data *cfqd = q->elevator->elevator_data;
3569         struct task_struct *tsk = current;
3570         struct cfq_io_cq *cic;
3571         struct cfq_queue *cfqq;
3572
3573         /*
3574          * don't force setup of a queue from here, as a call to may_queue
3575          * does not necessarily imply that a request actually will be queued.
3576          * so just lookup a possibly existing queue, or return 'may queue'
3577          * if that fails
3578          */
3579         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3580         if (!cic)
3581                 return ELV_MQUEUE_MAY;
3582
3583         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3584         if (cfqq) {
3585                 cfq_init_prio_data(cfqq, cic->icq.ioc);
3586
3587                 return __cfq_may_queue(cfqq);
3588         }
3589
3590         return ELV_MQUEUE_MAY;
3591 }
3592
3593 /*
3594  * queue lock held here
3595  */
3596 static void cfq_put_request(struct request *rq)
3597 {
3598         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3599
3600         if (cfqq) {
3601                 const int rw = rq_data_dir(rq);
3602
3603                 BUG_ON(!cfqq->allocated[rw]);
3604                 cfqq->allocated[rw]--;
3605
3606                 put_io_context(RQ_CIC(rq)->icq.ioc, cfqq->cfqd->queue);
3607
3608                 rq->elevator_private[0] = NULL;
3609                 rq->elevator_private[1] = NULL;
3610
3611                 /* Put down rq reference on cfqg */
3612                 cfq_put_cfqg(RQ_CFQG(rq));
3613                 rq->elevator_private[2] = NULL;
3614
3615                 cfq_put_queue(cfqq);
3616         }
3617 }
3618
3619 static struct cfq_queue *
3620 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3621                 struct cfq_queue *cfqq)
3622 {
3623         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3624         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3625         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3626         cfq_put_queue(cfqq);
3627         return cic_to_cfqq(cic, 1);
3628 }
3629
3630 /*
3631  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3632  * was the last process referring to said cfqq.
3633  */
3634 static struct cfq_queue *
3635 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3636 {
3637         if (cfqq_process_refs(cfqq) == 1) {
3638                 cfqq->pid = current->pid;
3639                 cfq_clear_cfqq_coop(cfqq);
3640                 cfq_clear_cfqq_split_coop(cfqq);
3641                 return cfqq;
3642         }
3643
3644         cic_set_cfqq(cic, NULL, 1);
3645
3646         cfq_put_cooperator(cfqq);
3647
3648         cfq_put_queue(cfqq);
3649         return NULL;
3650 }
3651 /*
3652  * Allocate cfq data structures associated with this request.
3653  */
3654 static int
3655 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3656 {
3657         struct cfq_data *cfqd = q->elevator->elevator_data;
3658         struct cfq_io_cq *cic;
3659         const int rw = rq_data_dir(rq);
3660         const bool is_sync = rq_is_sync(rq);
3661         struct cfq_queue *cfqq;
3662
3663         might_sleep_if(gfp_mask & __GFP_WAIT);
3664
3665         spin_lock_irq(q->queue_lock);
3666         cic = cfq_get_cic(cfqd, gfp_mask);
3667         if (!cic)
3668                 goto queue_fail;
3669
3670 new_queue:
3671         cfqq = cic_to_cfqq(cic, is_sync);
3672         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3673                 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3674                 cic_set_cfqq(cic, cfqq, is_sync);
3675         } else {
3676                 /*
3677                  * If the queue was seeky for too long, break it apart.
3678                  */
3679                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3680                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3681                         cfqq = split_cfqq(cic, cfqq);
3682                         if (!cfqq)
3683                                 goto new_queue;
3684                 }
3685
3686                 /*
3687                  * Check to see if this queue is scheduled to merge with
3688                  * another, closely cooperating queue.  The merging of
3689                  * queues happens here as it must be done in process context.
3690                  * The reference on new_cfqq was taken in merge_cfqqs.
3691                  */
3692                 if (cfqq->new_cfqq)
3693                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3694         }
3695
3696         cfqq->allocated[rw]++;
3697
3698         cfqq->ref++;
3699         rq->elevator_private[0] = &cic->icq;
3700         rq->elevator_private[1] = cfqq;
3701         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3702         spin_unlock_irq(q->queue_lock);
3703         return 0;
3704
3705 queue_fail:
3706         cfq_schedule_dispatch(cfqd);
3707         spin_unlock_irq(q->queue_lock);
3708         cfq_log(cfqd, "set_request fail");
3709         return 1;
3710 }
3711
3712 static void cfq_kick_queue(struct work_struct *work)
3713 {
3714         struct cfq_data *cfqd =
3715                 container_of(work, struct cfq_data, unplug_work);
3716         struct request_queue *q = cfqd->queue;
3717
3718         spin_lock_irq(q->queue_lock);
3719         __blk_run_queue(cfqd->queue);
3720         spin_unlock_irq(q->queue_lock);
3721 }
3722
3723 /*
3724  * Timer running if the active_queue is currently idling inside its time slice
3725  */
3726 static void cfq_idle_slice_timer(unsigned long data)
3727 {
3728         struct cfq_data *cfqd = (struct cfq_data *) data;
3729         struct cfq_queue *cfqq;
3730         unsigned long flags;
3731         int timed_out = 1;
3732
3733         cfq_log(cfqd, "idle timer fired");
3734
3735         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3736
3737         cfqq = cfqd->active_queue;
3738         if (cfqq) {
3739                 timed_out = 0;
3740
3741                 /*
3742                  * We saw a request before the queue expired, let it through
3743                  */
3744                 if (cfq_cfqq_must_dispatch(cfqq))
3745                         goto out_kick;
3746
3747                 /*
3748                  * expired
3749                  */
3750                 if (cfq_slice_used(cfqq))
3751                         goto expire;
3752
3753                 /*
3754                  * only expire and reinvoke request handler, if there are
3755                  * other queues with pending requests
3756                  */
3757                 if (!cfqd->busy_queues)
3758                         goto out_cont;
3759
3760                 /*
3761                  * not expired and it has a request pending, let it dispatch
3762                  */
3763                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3764                         goto out_kick;
3765
3766                 /*
3767                  * Queue depth flag is reset only when the idle didn't succeed
3768                  */
3769                 cfq_clear_cfqq_deep(cfqq);
3770         }
3771 expire:
3772         cfq_slice_expired(cfqd, timed_out);
3773 out_kick:
3774         cfq_schedule_dispatch(cfqd);
3775 out_cont:
3776         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3777 }
3778
3779 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3780 {
3781         del_timer_sync(&cfqd->idle_slice_timer);
3782         cancel_work_sync(&cfqd->unplug_work);
3783 }
3784
3785 static void cfq_put_async_queues(struct cfq_data *cfqd)
3786 {
3787         int i;
3788
3789         for (i = 0; i < IOPRIO_BE_NR; i++) {
3790                 if (cfqd->async_cfqq[0][i])
3791                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3792                 if (cfqd->async_cfqq[1][i])
3793                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3794         }
3795
3796         if (cfqd->async_idle_cfqq)
3797                 cfq_put_queue(cfqd->async_idle_cfqq);
3798 }
3799
3800 static void cfq_exit_queue(struct elevator_queue *e)
3801 {
3802         struct cfq_data *cfqd = e->elevator_data;
3803         struct request_queue *q = cfqd->queue;
3804         bool wait = false;
3805
3806         cfq_shutdown_timer_wq(cfqd);
3807
3808         spin_lock_irq(q->queue_lock);
3809
3810         if (cfqd->active_queue)
3811                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3812
3813         while (!list_empty(&cfqd->icq_list)) {
3814                 struct io_cq *icq = list_entry(cfqd->icq_list.next,
3815                                                struct io_cq, q_node);
3816                 struct io_context *ioc = icq->ioc;
3817
3818                 spin_lock(&ioc->lock);
3819                 cfq_exit_icq(icq);
3820                 cfq_release_icq(icq);
3821                 spin_unlock(&ioc->lock);
3822         }
3823
3824         cfq_put_async_queues(cfqd);
3825         cfq_release_cfq_groups(cfqd);
3826
3827         /*
3828          * If there are groups which we could not unlink from blkcg list,
3829          * wait for a rcu period for them to be freed.
3830          */
3831         if (cfqd->nr_blkcg_linked_grps)
3832                 wait = true;
3833
3834         spin_unlock_irq(q->queue_lock);
3835
3836         cfq_shutdown_timer_wq(cfqd);
3837
3838         /*
3839          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3840          * Do this wait only if there are other unlinked groups out
3841          * there. This can happen if cgroup deletion path claimed the
3842          * responsibility of cleaning up a group before queue cleanup code
3843          * get to the group.
3844          *
3845          * Do not call synchronize_rcu() unconditionally as there are drivers
3846          * which create/delete request queue hundreds of times during scan/boot
3847          * and synchronize_rcu() can take significant time and slow down boot.
3848          */
3849         if (wait)
3850                 synchronize_rcu();
3851
3852 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3853         /* Free up per cpu stats for root group */
3854         free_percpu(cfqd->root_group.blkg.stats_cpu);
3855 #endif
3856         kfree(cfqd);
3857 }
3858
3859 static void *cfq_init_queue(struct request_queue *q)
3860 {
3861         struct cfq_data *cfqd;
3862         int i, j;
3863         struct cfq_group *cfqg;
3864         struct cfq_rb_root *st;
3865
3866         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3867         if (!cfqd)
3868                 return NULL;
3869
3870         /* Init root service tree */
3871         cfqd->grp_service_tree = CFQ_RB_ROOT;
3872
3873         /* Init root group */
3874         cfqg = &cfqd->root_group;
3875         for_each_cfqg_st(cfqg, i, j, st)
3876                 *st = CFQ_RB_ROOT;
3877         RB_CLEAR_NODE(&cfqg->rb_node);
3878
3879         /* Give preference to root group over other groups */
3880         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3881
3882 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3883         /*
3884          * Set root group reference to 2. One reference will be dropped when
3885          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3886          * Other reference will remain there as we don't want to delete this
3887          * group as it is statically allocated and gets destroyed when
3888          * throtl_data goes away.
3889          */
3890         cfqg->ref = 2;
3891
3892         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3893                 kfree(cfqg);
3894                 kfree(cfqd);
3895                 return NULL;
3896         }
3897
3898         rcu_read_lock();
3899
3900         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3901                                         (void *)cfqd, 0);
3902         rcu_read_unlock();
3903         cfqd->nr_blkcg_linked_grps++;
3904
3905         /* Add group on cfqd->cfqg_list */
3906         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3907 #endif
3908         /*
3909          * Not strictly needed (since RB_ROOT just clears the node and we
3910          * zeroed cfqd on alloc), but better be safe in case someone decides
3911          * to add magic to the rb code
3912          */
3913         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3914                 cfqd->prio_trees[i] = RB_ROOT;
3915
3916         /*
3917          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3918          * Grab a permanent reference to it, so that the normal code flow
3919          * will not attempt to free it.
3920          */
3921         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3922         cfqd->oom_cfqq.ref++;
3923         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3924
3925         INIT_LIST_HEAD(&cfqd->icq_list);
3926
3927         cfqd->queue = q;
3928
3929         init_timer(&cfqd->idle_slice_timer);
3930         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3931         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3932
3933         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3934
3935         cfqd->cfq_quantum = cfq_quantum;
3936         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3937         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3938         cfqd->cfq_back_max = cfq_back_max;
3939         cfqd->cfq_back_penalty = cfq_back_penalty;
3940         cfqd->cfq_slice[0] = cfq_slice_async;
3941         cfqd->cfq_slice[1] = cfq_slice_sync;
3942         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3943         cfqd->cfq_slice_idle = cfq_slice_idle;
3944         cfqd->cfq_group_idle = cfq_group_idle;
3945         cfqd->cfq_latency = 1;
3946         cfqd->hw_tag = -1;
3947         /*
3948          * we optimistically start assuming sync ops weren't delayed in last
3949          * second, in order to have larger depth for async operations.
3950          */
3951         cfqd->last_delayed_sync = jiffies - HZ;
3952         return cfqd;
3953 }
3954
3955 static void cfq_slab_kill(void)
3956 {
3957         /*
3958          * Caller already ensured that pending RCU callbacks are completed,
3959          * so we should have no busy allocations at this point.
3960          */
3961         if (cfq_pool)
3962                 kmem_cache_destroy(cfq_pool);
3963         if (cfq_icq_pool)
3964                 kmem_cache_destroy(cfq_icq_pool);
3965 }
3966
3967 static int __init cfq_slab_setup(void)
3968 {
3969         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3970         if (!cfq_pool)
3971                 goto fail;
3972
3973         cfq_icq_pool = KMEM_CACHE(cfq_io_cq, 0);
3974         if (!cfq_icq_pool)
3975                 goto fail;
3976
3977         return 0;
3978 fail:
3979         cfq_slab_kill();
3980         return -ENOMEM;
3981 }
3982
3983 /*
3984  * sysfs parts below -->
3985  */
3986 static ssize_t
3987 cfq_var_show(unsigned int var, char *page)
3988 {
3989         return sprintf(page, "%d\n", var);
3990 }
3991
3992 static ssize_t
3993 cfq_var_store(unsigned int *var, const char *page, size_t count)
3994 {
3995         char *p = (char *) page;
3996
3997         *var = simple_strtoul(p, &p, 10);
3998         return count;
3999 }
4000
4001 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4002 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4003 {                                                                       \
4004         struct cfq_data *cfqd = e->elevator_data;                       \
4005         unsigned int __data = __VAR;                                    \
4006         if (__CONV)                                                     \
4007                 __data = jiffies_to_msecs(__data);                      \
4008         return cfq_var_show(__data, (page));                            \
4009 }
4010 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4011 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4012 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4013 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4014 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4015 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4016 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4017 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4018 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4019 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4020 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4021 #undef SHOW_FUNCTION
4022
4023 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4024 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4025 {                                                                       \
4026         struct cfq_data *cfqd = e->elevator_data;                       \
4027         unsigned int __data;                                            \
4028         int ret = cfq_var_store(&__data, (page), count);                \
4029         if (__data < (MIN))                                             \
4030                 __data = (MIN);                                         \
4031         else if (__data > (MAX))                                        \
4032                 __data = (MAX);                                         \
4033         if (__CONV)                                                     \
4034                 *(__PTR) = msecs_to_jiffies(__data);                    \
4035         else                                                            \
4036                 *(__PTR) = __data;                                      \
4037         return ret;                                                     \
4038 }
4039 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4040 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4041                 UINT_MAX, 1);
4042 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4043                 UINT_MAX, 1);
4044 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4045 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4046                 UINT_MAX, 0);
4047 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4048 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4049 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4050 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4051 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4052                 UINT_MAX, 0);
4053 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4054 #undef STORE_FUNCTION
4055
4056 #define CFQ_ATTR(name) \
4057         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4058
4059 static struct elv_fs_entry cfq_attrs[] = {
4060         CFQ_ATTR(quantum),
4061         CFQ_ATTR(fifo_expire_sync),
4062         CFQ_ATTR(fifo_expire_async),
4063         CFQ_ATTR(back_seek_max),
4064         CFQ_ATTR(back_seek_penalty),
4065         CFQ_ATTR(slice_sync),
4066         CFQ_ATTR(slice_async),
4067         CFQ_ATTR(slice_async_rq),
4068         CFQ_ATTR(slice_idle),
4069         CFQ_ATTR(group_idle),
4070         CFQ_ATTR(low_latency),
4071         __ATTR_NULL
4072 };
4073
4074 static struct elevator_type iosched_cfq = {
4075         .ops = {
4076                 .elevator_merge_fn =            cfq_merge,
4077                 .elevator_merged_fn =           cfq_merged_request,
4078                 .elevator_merge_req_fn =        cfq_merged_requests,
4079                 .elevator_allow_merge_fn =      cfq_allow_merge,
4080                 .elevator_bio_merged_fn =       cfq_bio_merged,
4081                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4082                 .elevator_add_req_fn =          cfq_insert_request,
4083                 .elevator_activate_req_fn =     cfq_activate_request,
4084                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4085                 .elevator_completed_req_fn =    cfq_completed_request,
4086                 .elevator_former_req_fn =       elv_rb_former_request,
4087                 .elevator_latter_req_fn =       elv_rb_latter_request,
4088                 .elevator_set_req_fn =          cfq_set_request,
4089                 .elevator_put_req_fn =          cfq_put_request,
4090                 .elevator_may_queue_fn =        cfq_may_queue,
4091                 .elevator_init_fn =             cfq_init_queue,
4092                 .elevator_exit_fn =             cfq_exit_queue,
4093         },
4094         .elevator_attrs =       cfq_attrs,
4095         .elevator_name =        "cfq",
4096         .elevator_owner =       THIS_MODULE,
4097 };
4098
4099 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4100 static struct blkio_policy_type blkio_policy_cfq = {
4101         .ops = {
4102                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4103                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4104         },
4105         .plid = BLKIO_POLICY_PROP,
4106 };
4107 #else
4108 static struct blkio_policy_type blkio_policy_cfq;
4109 #endif
4110
4111 static int __init cfq_init(void)
4112 {
4113         /*
4114          * could be 0 on HZ < 1000 setups
4115          */
4116         if (!cfq_slice_async)
4117                 cfq_slice_async = 1;
4118         if (!cfq_slice_idle)
4119                 cfq_slice_idle = 1;
4120
4121 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4122         if (!cfq_group_idle)
4123                 cfq_group_idle = 1;
4124 #else
4125                 cfq_group_idle = 0;
4126 #endif
4127         if (cfq_slab_setup())
4128                 return -ENOMEM;
4129
4130         elv_register(&iosched_cfq);
4131         blkio_policy_register(&blkio_policy_cfq);
4132
4133         return 0;
4134 }
4135
4136 static void __exit cfq_exit(void)
4137 {
4138         blkio_policy_unregister(&blkio_policy_cfq);
4139         elv_unregister(&iosched_cfq);
4140         rcu_barrier();  /* make sure all cic RCU frees are complete */
4141         cfq_slab_kill();
4142 }
4143
4144 module_init(cfq_init);
4145 module_exit(cfq_exit);
4146
4147 MODULE_AUTHOR("Jens Axboe");
4148 MODULE_LICENSE("GPL");
4149 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");