block, cfq: move icq creation and rq->elv.icq association to block core
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk.h"
18 #include "cfq.h"
19
20 /*
21  * tunables
22  */
23 /* max queue in one round of service */
24 static const int cfq_quantum = 8;
25 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
26 /* maximum backwards seek, in KiB */
27 static const int cfq_back_max = 16 * 1024;
28 /* penalty of a backwards seek */
29 static const int cfq_back_penalty = 2;
30 static const int cfq_slice_sync = HZ / 10;
31 static int cfq_slice_async = HZ / 25;
32 static const int cfq_slice_async_rq = 2;
33 static int cfq_slice_idle = HZ / 125;
34 static int cfq_group_idle = HZ / 125;
35 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
36 static const int cfq_hist_divisor = 4;
37
38 /*
39  * offset from end of service tree
40  */
41 #define CFQ_IDLE_DELAY          (HZ / 5)
42
43 /*
44  * below this threshold, we consider thinktime immediate
45  */
46 #define CFQ_MIN_TT              (2)
47
48 #define CFQ_SLICE_SCALE         (5)
49 #define CFQ_HW_QUEUE_MIN        (5)
50 #define CFQ_SERVICE_SHIFT       12
51
52 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
53 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
54 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
55 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
56
57 #define RQ_CIC(rq)              icq_to_cic((rq)->elv.icq)
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elv.priv[0])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elv.priv[1])
60
61 static struct kmem_cache *cfq_pool;
62
63 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
64 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
65 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
66
67 #define sample_valid(samples)   ((samples) > 80)
68 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
69
70 struct cfq_ttime {
71         unsigned long last_end_request;
72
73         unsigned long ttime_total;
74         unsigned long ttime_samples;
75         unsigned long ttime_mean;
76 };
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90         struct cfq_ttime ttime;
91 };
92 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
93                         .ttime = {.last_end_request = jiffies,},}
94
95 /*
96  * Per process-grouping structure
97  */
98 struct cfq_queue {
99         /* reference count */
100         int ref;
101         /* various state flags, see below */
102         unsigned int flags;
103         /* parent cfq_data */
104         struct cfq_data *cfqd;
105         /* service_tree member */
106         struct rb_node rb_node;
107         /* service_tree key */
108         unsigned long rb_key;
109         /* prio tree member */
110         struct rb_node p_node;
111         /* prio tree root we belong to, if any */
112         struct rb_root *p_root;
113         /* sorted list of pending requests */
114         struct rb_root sort_list;
115         /* if fifo isn't expired, next request to serve */
116         struct request *next_rq;
117         /* requests queued in sort_list */
118         int queued[2];
119         /* currently allocated requests */
120         int allocated[2];
121         /* fifo list of requests in sort_list */
122         struct list_head fifo;
123
124         /* time when queue got scheduled in to dispatch first request. */
125         unsigned long dispatch_start;
126         unsigned int allocated_slice;
127         unsigned int slice_dispatch;
128         /* time when first request from queue completed and slice started. */
129         unsigned long slice_start;
130         unsigned long slice_end;
131         long slice_resid;
132
133         /* pending priority requests */
134         int prio_pending;
135         /* number of requests that are on the dispatch list or inside driver */
136         int dispatched;
137
138         /* io prio of this group */
139         unsigned short ioprio, org_ioprio;
140         unsigned short ioprio_class;
141
142         pid_t pid;
143
144         u32 seek_history;
145         sector_t last_request_pos;
146
147         struct cfq_rb_root *service_tree;
148         struct cfq_queue *new_cfqq;
149         struct cfq_group *cfqg;
150         /* Number of sectors dispatched from queue in single dispatch round */
151         unsigned long nr_sectors;
152 };
153
154 /*
155  * First index in the service_trees.
156  * IDLE is handled separately, so it has negative index
157  */
158 enum wl_prio_t {
159         BE_WORKLOAD = 0,
160         RT_WORKLOAD = 1,
161         IDLE_WORKLOAD = 2,
162         CFQ_PRIO_NR,
163 };
164
165 /*
166  * Second index in the service_trees.
167  */
168 enum wl_type_t {
169         ASYNC_WORKLOAD = 0,
170         SYNC_NOIDLE_WORKLOAD = 1,
171         SYNC_WORKLOAD = 2
172 };
173
174 /* This is per cgroup per device grouping structure */
175 struct cfq_group {
176         /* group service_tree member */
177         struct rb_node rb_node;
178
179         /* group service_tree key */
180         u64 vdisktime;
181         unsigned int weight;
182         unsigned int new_weight;
183         bool needs_update;
184
185         /* number of cfqq currently on this group */
186         int nr_cfqq;
187
188         /*
189          * Per group busy queues average. Useful for workload slice calc. We
190          * create the array for each prio class but at run time it is used
191          * only for RT and BE class and slot for IDLE class remains unused.
192          * This is primarily done to avoid confusion and a gcc warning.
193          */
194         unsigned int busy_queues_avg[CFQ_PRIO_NR];
195         /*
196          * rr lists of queues with requests. We maintain service trees for
197          * RT and BE classes. These trees are subdivided in subclasses
198          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199          * class there is no subclassification and all the cfq queues go on
200          * a single tree service_tree_idle.
201          * Counts are embedded in the cfq_rb_root
202          */
203         struct cfq_rb_root service_trees[2][3];
204         struct cfq_rb_root service_tree_idle;
205
206         unsigned long saved_workload_slice;
207         enum wl_type_t saved_workload;
208         enum wl_prio_t saved_serving_prio;
209         struct blkio_group blkg;
210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
211         struct hlist_node cfqd_node;
212         int ref;
213 #endif
214         /* number of requests that are on the dispatch list or inside driver */
215         int dispatched;
216         struct cfq_ttime ttime;
217 };
218
219 struct cfq_io_cq {
220         struct io_cq            icq;            /* must be the first member */
221         struct cfq_queue        *cfqq[2];
222         struct cfq_ttime        ttime;
223 };
224
225 /*
226  * Per block device queue structure
227  */
228 struct cfq_data {
229         struct request_queue *queue;
230         /* Root service tree for cfq_groups */
231         struct cfq_rb_root grp_service_tree;
232         struct cfq_group root_group;
233
234         /*
235          * The priority currently being served
236          */
237         enum wl_prio_t serving_prio;
238         enum wl_type_t serving_type;
239         unsigned long workload_expires;
240         struct cfq_group *serving_group;
241
242         /*
243          * Each priority tree is sorted by next_request position.  These
244          * trees are used when determining if two or more queues are
245          * interleaving requests (see cfq_close_cooperator).
246          */
247         struct rb_root prio_trees[CFQ_PRIO_LISTS];
248
249         unsigned int busy_queues;
250         unsigned int busy_sync_queues;
251
252         int rq_in_driver;
253         int rq_in_flight[2];
254
255         /*
256          * queue-depth detection
257          */
258         int rq_queued;
259         int hw_tag;
260         /*
261          * hw_tag can be
262          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
263          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
264          *  0 => no NCQ
265          */
266         int hw_tag_est_depth;
267         unsigned int hw_tag_samples;
268
269         /*
270          * idle window management
271          */
272         struct timer_list idle_slice_timer;
273         struct work_struct unplug_work;
274
275         struct cfq_queue *active_queue;
276         struct cfq_io_cq *active_cic;
277
278         /*
279          * async queue for each priority case
280          */
281         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
282         struct cfq_queue *async_idle_cfqq;
283
284         sector_t last_position;
285
286         /*
287          * tunables, see top of file
288          */
289         unsigned int cfq_quantum;
290         unsigned int cfq_fifo_expire[2];
291         unsigned int cfq_back_penalty;
292         unsigned int cfq_back_max;
293         unsigned int cfq_slice[2];
294         unsigned int cfq_slice_async_rq;
295         unsigned int cfq_slice_idle;
296         unsigned int cfq_group_idle;
297         unsigned int cfq_latency;
298
299         /*
300          * Fallback dummy cfqq for extreme OOM conditions
301          */
302         struct cfq_queue oom_cfqq;
303
304         unsigned long last_delayed_sync;
305
306         /* List of cfq groups being managed on this device*/
307         struct hlist_head cfqg_list;
308
309         /* Number of groups which are on blkcg->blkg_list */
310         unsigned int nr_blkcg_linked_grps;
311 };
312
313 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
314
315 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
316                                             enum wl_prio_t prio,
317                                             enum wl_type_t type)
318 {
319         if (!cfqg)
320                 return NULL;
321
322         if (prio == IDLE_WORKLOAD)
323                 return &cfqg->service_tree_idle;
324
325         return &cfqg->service_trees[prio][type];
326 }
327
328 enum cfqq_state_flags {
329         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
330         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
331         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
332         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
333         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
334         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
335         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
336         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
337         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
338         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
339         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
340         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
341         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
342 };
343
344 #define CFQ_CFQQ_FNS(name)                                              \
345 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
346 {                                                                       \
347         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
348 }                                                                       \
349 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
350 {                                                                       \
351         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
352 }                                                                       \
353 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
354 {                                                                       \
355         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
356 }
357
358 CFQ_CFQQ_FNS(on_rr);
359 CFQ_CFQQ_FNS(wait_request);
360 CFQ_CFQQ_FNS(must_dispatch);
361 CFQ_CFQQ_FNS(must_alloc_slice);
362 CFQ_CFQQ_FNS(fifo_expire);
363 CFQ_CFQQ_FNS(idle_window);
364 CFQ_CFQQ_FNS(prio_changed);
365 CFQ_CFQQ_FNS(slice_new);
366 CFQ_CFQQ_FNS(sync);
367 CFQ_CFQQ_FNS(coop);
368 CFQ_CFQQ_FNS(split_coop);
369 CFQ_CFQQ_FNS(deep);
370 CFQ_CFQQ_FNS(wait_busy);
371 #undef CFQ_CFQQ_FNS
372
373 #ifdef CONFIG_CFQ_GROUP_IOSCHED
374 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
375         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
376                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
377                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
378
379 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
380         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
381                                 blkg_path(&(cfqg)->blkg), ##args)       \
382
383 #else
384 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
385         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
386 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
387 #endif
388 #define cfq_log(cfqd, fmt, args...)     \
389         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
390
391 /* Traverses through cfq group service trees */
392 #define for_each_cfqg_st(cfqg, i, j, st) \
393         for (i = 0; i <= IDLE_WORKLOAD; i++) \
394                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
395                         : &cfqg->service_tree_idle; \
396                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
397                         (i == IDLE_WORKLOAD && j == 0); \
398                         j++, st = i < IDLE_WORKLOAD ? \
399                         &cfqg->service_trees[i][j]: NULL) \
400
401 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
402         struct cfq_ttime *ttime, bool group_idle)
403 {
404         unsigned long slice;
405         if (!sample_valid(ttime->ttime_samples))
406                 return false;
407         if (group_idle)
408                 slice = cfqd->cfq_group_idle;
409         else
410                 slice = cfqd->cfq_slice_idle;
411         return ttime->ttime_mean > slice;
412 }
413
414 static inline bool iops_mode(struct cfq_data *cfqd)
415 {
416         /*
417          * If we are not idling on queues and it is a NCQ drive, parallel
418          * execution of requests is on and measuring time is not possible
419          * in most of the cases until and unless we drive shallower queue
420          * depths and that becomes a performance bottleneck. In such cases
421          * switch to start providing fairness in terms of number of IOs.
422          */
423         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
424                 return true;
425         else
426                 return false;
427 }
428
429 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
430 {
431         if (cfq_class_idle(cfqq))
432                 return IDLE_WORKLOAD;
433         if (cfq_class_rt(cfqq))
434                 return RT_WORKLOAD;
435         return BE_WORKLOAD;
436 }
437
438
439 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
440 {
441         if (!cfq_cfqq_sync(cfqq))
442                 return ASYNC_WORKLOAD;
443         if (!cfq_cfqq_idle_window(cfqq))
444                 return SYNC_NOIDLE_WORKLOAD;
445         return SYNC_WORKLOAD;
446 }
447
448 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
449                                         struct cfq_data *cfqd,
450                                         struct cfq_group *cfqg)
451 {
452         if (wl == IDLE_WORKLOAD)
453                 return cfqg->service_tree_idle.count;
454
455         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
456                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
457                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
458 }
459
460 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
461                                         struct cfq_group *cfqg)
462 {
463         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
464                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
465 }
466
467 static void cfq_dispatch_insert(struct request_queue *, struct request *);
468 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
469                                        struct io_context *, gfp_t);
470
471 static inline struct cfq_io_cq *icq_to_cic(struct io_cq *icq)
472 {
473         /* cic->icq is the first member, %NULL will convert to %NULL */
474         return container_of(icq, struct cfq_io_cq, icq);
475 }
476
477 static inline struct cfq_io_cq *cfq_cic_lookup(struct cfq_data *cfqd,
478                                                struct io_context *ioc)
479 {
480         if (ioc)
481                 return icq_to_cic(ioc_lookup_icq(ioc, cfqd->queue));
482         return NULL;
483 }
484
485 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_cq *cic, bool is_sync)
486 {
487         return cic->cfqq[is_sync];
488 }
489
490 static inline void cic_set_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq,
491                                 bool is_sync)
492 {
493         cic->cfqq[is_sync] = cfqq;
494 }
495
496 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_cq *cic)
497 {
498         return cic->icq.q->elevator->elevator_data;
499 }
500
501 /*
502  * We regard a request as SYNC, if it's either a read or has the SYNC bit
503  * set (in which case it could also be direct WRITE).
504  */
505 static inline bool cfq_bio_sync(struct bio *bio)
506 {
507         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
508 }
509
510 /*
511  * scheduler run of queue, if there are requests pending and no one in the
512  * driver that will restart queueing
513  */
514 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
515 {
516         if (cfqd->busy_queues) {
517                 cfq_log(cfqd, "schedule dispatch");
518                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
519         }
520 }
521
522 /*
523  * Scale schedule slice based on io priority. Use the sync time slice only
524  * if a queue is marked sync and has sync io queued. A sync queue with async
525  * io only, should not get full sync slice length.
526  */
527 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
528                                  unsigned short prio)
529 {
530         const int base_slice = cfqd->cfq_slice[sync];
531
532         WARN_ON(prio >= IOPRIO_BE_NR);
533
534         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
535 }
536
537 static inline int
538 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
539 {
540         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
541 }
542
543 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
544 {
545         u64 d = delta << CFQ_SERVICE_SHIFT;
546
547         d = d * BLKIO_WEIGHT_DEFAULT;
548         do_div(d, cfqg->weight);
549         return d;
550 }
551
552 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
553 {
554         s64 delta = (s64)(vdisktime - min_vdisktime);
555         if (delta > 0)
556                 min_vdisktime = vdisktime;
557
558         return min_vdisktime;
559 }
560
561 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
562 {
563         s64 delta = (s64)(vdisktime - min_vdisktime);
564         if (delta < 0)
565                 min_vdisktime = vdisktime;
566
567         return min_vdisktime;
568 }
569
570 static void update_min_vdisktime(struct cfq_rb_root *st)
571 {
572         struct cfq_group *cfqg;
573
574         if (st->left) {
575                 cfqg = rb_entry_cfqg(st->left);
576                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
577                                                   cfqg->vdisktime);
578         }
579 }
580
581 /*
582  * get averaged number of queues of RT/BE priority.
583  * average is updated, with a formula that gives more weight to higher numbers,
584  * to quickly follows sudden increases and decrease slowly
585  */
586
587 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
588                                         struct cfq_group *cfqg, bool rt)
589 {
590         unsigned min_q, max_q;
591         unsigned mult  = cfq_hist_divisor - 1;
592         unsigned round = cfq_hist_divisor / 2;
593         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
594
595         min_q = min(cfqg->busy_queues_avg[rt], busy);
596         max_q = max(cfqg->busy_queues_avg[rt], busy);
597         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
598                 cfq_hist_divisor;
599         return cfqg->busy_queues_avg[rt];
600 }
601
602 static inline unsigned
603 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
604 {
605         struct cfq_rb_root *st = &cfqd->grp_service_tree;
606
607         return cfq_target_latency * cfqg->weight / st->total_weight;
608 }
609
610 static inline unsigned
611 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
612 {
613         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
614         if (cfqd->cfq_latency) {
615                 /*
616                  * interested queues (we consider only the ones with the same
617                  * priority class in the cfq group)
618                  */
619                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
620                                                 cfq_class_rt(cfqq));
621                 unsigned sync_slice = cfqd->cfq_slice[1];
622                 unsigned expect_latency = sync_slice * iq;
623                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
624
625                 if (expect_latency > group_slice) {
626                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
627                         /* scale low_slice according to IO priority
628                          * and sync vs async */
629                         unsigned low_slice =
630                                 min(slice, base_low_slice * slice / sync_slice);
631                         /* the adapted slice value is scaled to fit all iqs
632                          * into the target latency */
633                         slice = max(slice * group_slice / expect_latency,
634                                     low_slice);
635                 }
636         }
637         return slice;
638 }
639
640 static inline void
641 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
642 {
643         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
644
645         cfqq->slice_start = jiffies;
646         cfqq->slice_end = jiffies + slice;
647         cfqq->allocated_slice = slice;
648         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
649 }
650
651 /*
652  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
653  * isn't valid until the first request from the dispatch is activated
654  * and the slice time set.
655  */
656 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
657 {
658         if (cfq_cfqq_slice_new(cfqq))
659                 return false;
660         if (time_before(jiffies, cfqq->slice_end))
661                 return false;
662
663         return true;
664 }
665
666 /*
667  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
668  * We choose the request that is closest to the head right now. Distance
669  * behind the head is penalized and only allowed to a certain extent.
670  */
671 static struct request *
672 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
673 {
674         sector_t s1, s2, d1 = 0, d2 = 0;
675         unsigned long back_max;
676 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
677 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
678         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
679
680         if (rq1 == NULL || rq1 == rq2)
681                 return rq2;
682         if (rq2 == NULL)
683                 return rq1;
684
685         if (rq_is_sync(rq1) != rq_is_sync(rq2))
686                 return rq_is_sync(rq1) ? rq1 : rq2;
687
688         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_PRIO)
689                 return rq1->cmd_flags & REQ_PRIO ? rq1 : rq2;
690
691         s1 = blk_rq_pos(rq1);
692         s2 = blk_rq_pos(rq2);
693
694         /*
695          * by definition, 1KiB is 2 sectors
696          */
697         back_max = cfqd->cfq_back_max * 2;
698
699         /*
700          * Strict one way elevator _except_ in the case where we allow
701          * short backward seeks which are biased as twice the cost of a
702          * similar forward seek.
703          */
704         if (s1 >= last)
705                 d1 = s1 - last;
706         else if (s1 + back_max >= last)
707                 d1 = (last - s1) * cfqd->cfq_back_penalty;
708         else
709                 wrap |= CFQ_RQ1_WRAP;
710
711         if (s2 >= last)
712                 d2 = s2 - last;
713         else if (s2 + back_max >= last)
714                 d2 = (last - s2) * cfqd->cfq_back_penalty;
715         else
716                 wrap |= CFQ_RQ2_WRAP;
717
718         /* Found required data */
719
720         /*
721          * By doing switch() on the bit mask "wrap" we avoid having to
722          * check two variables for all permutations: --> faster!
723          */
724         switch (wrap) {
725         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
726                 if (d1 < d2)
727                         return rq1;
728                 else if (d2 < d1)
729                         return rq2;
730                 else {
731                         if (s1 >= s2)
732                                 return rq1;
733                         else
734                                 return rq2;
735                 }
736
737         case CFQ_RQ2_WRAP:
738                 return rq1;
739         case CFQ_RQ1_WRAP:
740                 return rq2;
741         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
742         default:
743                 /*
744                  * Since both rqs are wrapped,
745                  * start with the one that's further behind head
746                  * (--> only *one* back seek required),
747                  * since back seek takes more time than forward.
748                  */
749                 if (s1 <= s2)
750                         return rq1;
751                 else
752                         return rq2;
753         }
754 }
755
756 /*
757  * The below is leftmost cache rbtree addon
758  */
759 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
760 {
761         /* Service tree is empty */
762         if (!root->count)
763                 return NULL;
764
765         if (!root->left)
766                 root->left = rb_first(&root->rb);
767
768         if (root->left)
769                 return rb_entry(root->left, struct cfq_queue, rb_node);
770
771         return NULL;
772 }
773
774 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
775 {
776         if (!root->left)
777                 root->left = rb_first(&root->rb);
778
779         if (root->left)
780                 return rb_entry_cfqg(root->left);
781
782         return NULL;
783 }
784
785 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
786 {
787         rb_erase(n, root);
788         RB_CLEAR_NODE(n);
789 }
790
791 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
792 {
793         if (root->left == n)
794                 root->left = NULL;
795         rb_erase_init(n, &root->rb);
796         --root->count;
797 }
798
799 /*
800  * would be nice to take fifo expire time into account as well
801  */
802 static struct request *
803 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
804                   struct request *last)
805 {
806         struct rb_node *rbnext = rb_next(&last->rb_node);
807         struct rb_node *rbprev = rb_prev(&last->rb_node);
808         struct request *next = NULL, *prev = NULL;
809
810         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
811
812         if (rbprev)
813                 prev = rb_entry_rq(rbprev);
814
815         if (rbnext)
816                 next = rb_entry_rq(rbnext);
817         else {
818                 rbnext = rb_first(&cfqq->sort_list);
819                 if (rbnext && rbnext != &last->rb_node)
820                         next = rb_entry_rq(rbnext);
821         }
822
823         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
824 }
825
826 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
827                                       struct cfq_queue *cfqq)
828 {
829         /*
830          * just an approximation, should be ok.
831          */
832         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
833                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
834 }
835
836 static inline s64
837 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
838 {
839         return cfqg->vdisktime - st->min_vdisktime;
840 }
841
842 static void
843 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
844 {
845         struct rb_node **node = &st->rb.rb_node;
846         struct rb_node *parent = NULL;
847         struct cfq_group *__cfqg;
848         s64 key = cfqg_key(st, cfqg);
849         int left = 1;
850
851         while (*node != NULL) {
852                 parent = *node;
853                 __cfqg = rb_entry_cfqg(parent);
854
855                 if (key < cfqg_key(st, __cfqg))
856                         node = &parent->rb_left;
857                 else {
858                         node = &parent->rb_right;
859                         left = 0;
860                 }
861         }
862
863         if (left)
864                 st->left = &cfqg->rb_node;
865
866         rb_link_node(&cfqg->rb_node, parent, node);
867         rb_insert_color(&cfqg->rb_node, &st->rb);
868 }
869
870 static void
871 cfq_update_group_weight(struct cfq_group *cfqg)
872 {
873         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
874         if (cfqg->needs_update) {
875                 cfqg->weight = cfqg->new_weight;
876                 cfqg->needs_update = false;
877         }
878 }
879
880 static void
881 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
882 {
883         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
884
885         cfq_update_group_weight(cfqg);
886         __cfq_group_service_tree_add(st, cfqg);
887         st->total_weight += cfqg->weight;
888 }
889
890 static void
891 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
892 {
893         struct cfq_rb_root *st = &cfqd->grp_service_tree;
894         struct cfq_group *__cfqg;
895         struct rb_node *n;
896
897         cfqg->nr_cfqq++;
898         if (!RB_EMPTY_NODE(&cfqg->rb_node))
899                 return;
900
901         /*
902          * Currently put the group at the end. Later implement something
903          * so that groups get lesser vtime based on their weights, so that
904          * if group does not loose all if it was not continuously backlogged.
905          */
906         n = rb_last(&st->rb);
907         if (n) {
908                 __cfqg = rb_entry_cfqg(n);
909                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
910         } else
911                 cfqg->vdisktime = st->min_vdisktime;
912         cfq_group_service_tree_add(st, cfqg);
913 }
914
915 static void
916 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
917 {
918         st->total_weight -= cfqg->weight;
919         if (!RB_EMPTY_NODE(&cfqg->rb_node))
920                 cfq_rb_erase(&cfqg->rb_node, st);
921 }
922
923 static void
924 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
925 {
926         struct cfq_rb_root *st = &cfqd->grp_service_tree;
927
928         BUG_ON(cfqg->nr_cfqq < 1);
929         cfqg->nr_cfqq--;
930
931         /* If there are other cfq queues under this group, don't delete it */
932         if (cfqg->nr_cfqq)
933                 return;
934
935         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
936         cfq_group_service_tree_del(st, cfqg);
937         cfqg->saved_workload_slice = 0;
938         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
939 }
940
941 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
942                                                 unsigned int *unaccounted_time)
943 {
944         unsigned int slice_used;
945
946         /*
947          * Queue got expired before even a single request completed or
948          * got expired immediately after first request completion.
949          */
950         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
951                 /*
952                  * Also charge the seek time incurred to the group, otherwise
953                  * if there are mutiple queues in the group, each can dispatch
954                  * a single request on seeky media and cause lots of seek time
955                  * and group will never know it.
956                  */
957                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
958                                         1);
959         } else {
960                 slice_used = jiffies - cfqq->slice_start;
961                 if (slice_used > cfqq->allocated_slice) {
962                         *unaccounted_time = slice_used - cfqq->allocated_slice;
963                         slice_used = cfqq->allocated_slice;
964                 }
965                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
966                         *unaccounted_time += cfqq->slice_start -
967                                         cfqq->dispatch_start;
968         }
969
970         return slice_used;
971 }
972
973 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
974                                 struct cfq_queue *cfqq)
975 {
976         struct cfq_rb_root *st = &cfqd->grp_service_tree;
977         unsigned int used_sl, charge, unaccounted_sl = 0;
978         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
979                         - cfqg->service_tree_idle.count;
980
981         BUG_ON(nr_sync < 0);
982         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
983
984         if (iops_mode(cfqd))
985                 charge = cfqq->slice_dispatch;
986         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
987                 charge = cfqq->allocated_slice;
988
989         /* Can't update vdisktime while group is on service tree */
990         cfq_group_service_tree_del(st, cfqg);
991         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
992         /* If a new weight was requested, update now, off tree */
993         cfq_group_service_tree_add(st, cfqg);
994
995         /* This group is being expired. Save the context */
996         if (time_after(cfqd->workload_expires, jiffies)) {
997                 cfqg->saved_workload_slice = cfqd->workload_expires
998                                                 - jiffies;
999                 cfqg->saved_workload = cfqd->serving_type;
1000                 cfqg->saved_serving_prio = cfqd->serving_prio;
1001         } else
1002                 cfqg->saved_workload_slice = 0;
1003
1004         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1005                                         st->min_vdisktime);
1006         cfq_log_cfqq(cfqq->cfqd, cfqq,
1007                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1008                      used_sl, cfqq->slice_dispatch, charge,
1009                      iops_mode(cfqd), cfqq->nr_sectors);
1010         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1011                                           unaccounted_sl);
1012         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1013 }
1014
1015 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1016 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1017 {
1018         if (blkg)
1019                 return container_of(blkg, struct cfq_group, blkg);
1020         return NULL;
1021 }
1022
1023 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1024                                           unsigned int weight)
1025 {
1026         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1027         cfqg->new_weight = weight;
1028         cfqg->needs_update = true;
1029 }
1030
1031 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1032                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1033 {
1034         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1035         unsigned int major, minor;
1036
1037         /*
1038          * Add group onto cgroup list. It might happen that bdi->dev is
1039          * not initialized yet. Initialize this new group without major
1040          * and minor info and this info will be filled in once a new thread
1041          * comes for IO.
1042          */
1043         if (bdi->dev) {
1044                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1045                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1046                                         (void *)cfqd, MKDEV(major, minor));
1047         } else
1048                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1049                                         (void *)cfqd, 0);
1050
1051         cfqd->nr_blkcg_linked_grps++;
1052         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1053
1054         /* Add group on cfqd list */
1055         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1056 }
1057
1058 /*
1059  * Should be called from sleepable context. No request queue lock as per
1060  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1061  * from sleepable context.
1062  */
1063 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1064 {
1065         struct cfq_group *cfqg = NULL;
1066         int i, j, ret;
1067         struct cfq_rb_root *st;
1068
1069         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1070         if (!cfqg)
1071                 return NULL;
1072
1073         for_each_cfqg_st(cfqg, i, j, st)
1074                 *st = CFQ_RB_ROOT;
1075         RB_CLEAR_NODE(&cfqg->rb_node);
1076
1077         cfqg->ttime.last_end_request = jiffies;
1078
1079         /*
1080          * Take the initial reference that will be released on destroy
1081          * This can be thought of a joint reference by cgroup and
1082          * elevator which will be dropped by either elevator exit
1083          * or cgroup deletion path depending on who is exiting first.
1084          */
1085         cfqg->ref = 1;
1086
1087         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1088         if (ret) {
1089                 kfree(cfqg);
1090                 return NULL;
1091         }
1092
1093         return cfqg;
1094 }
1095
1096 static struct cfq_group *
1097 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1098 {
1099         struct cfq_group *cfqg = NULL;
1100         void *key = cfqd;
1101         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1102         unsigned int major, minor;
1103
1104         /*
1105          * This is the common case when there are no blkio cgroups.
1106          * Avoid lookup in this case
1107          */
1108         if (blkcg == &blkio_root_cgroup)
1109                 cfqg = &cfqd->root_group;
1110         else
1111                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1112
1113         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1114                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1115                 cfqg->blkg.dev = MKDEV(major, minor);
1116         }
1117
1118         return cfqg;
1119 }
1120
1121 /*
1122  * Search for the cfq group current task belongs to. request_queue lock must
1123  * be held.
1124  */
1125 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1126 {
1127         struct blkio_cgroup *blkcg;
1128         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1129         struct request_queue *q = cfqd->queue;
1130
1131         rcu_read_lock();
1132         blkcg = task_blkio_cgroup(current);
1133         cfqg = cfq_find_cfqg(cfqd, blkcg);
1134         if (cfqg) {
1135                 rcu_read_unlock();
1136                 return cfqg;
1137         }
1138
1139         /*
1140          * Need to allocate a group. Allocation of group also needs allocation
1141          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1142          * we need to drop rcu lock and queue_lock before we call alloc.
1143          *
1144          * Not taking any queue reference here and assuming that queue is
1145          * around by the time we return. CFQ queue allocation code does
1146          * the same. It might be racy though.
1147          */
1148
1149         rcu_read_unlock();
1150         spin_unlock_irq(q->queue_lock);
1151
1152         cfqg = cfq_alloc_cfqg(cfqd);
1153
1154         spin_lock_irq(q->queue_lock);
1155
1156         rcu_read_lock();
1157         blkcg = task_blkio_cgroup(current);
1158
1159         /*
1160          * If some other thread already allocated the group while we were
1161          * not holding queue lock, free up the group
1162          */
1163         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1164
1165         if (__cfqg) {
1166                 kfree(cfqg);
1167                 rcu_read_unlock();
1168                 return __cfqg;
1169         }
1170
1171         if (!cfqg)
1172                 cfqg = &cfqd->root_group;
1173
1174         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1175         rcu_read_unlock();
1176         return cfqg;
1177 }
1178
1179 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1180 {
1181         cfqg->ref++;
1182         return cfqg;
1183 }
1184
1185 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1186 {
1187         /* Currently, all async queues are mapped to root group */
1188         if (!cfq_cfqq_sync(cfqq))
1189                 cfqg = &cfqq->cfqd->root_group;
1190
1191         cfqq->cfqg = cfqg;
1192         /* cfqq reference on cfqg */
1193         cfqq->cfqg->ref++;
1194 }
1195
1196 static void cfq_put_cfqg(struct cfq_group *cfqg)
1197 {
1198         struct cfq_rb_root *st;
1199         int i, j;
1200
1201         BUG_ON(cfqg->ref <= 0);
1202         cfqg->ref--;
1203         if (cfqg->ref)
1204                 return;
1205         for_each_cfqg_st(cfqg, i, j, st)
1206                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1207         free_percpu(cfqg->blkg.stats_cpu);
1208         kfree(cfqg);
1209 }
1210
1211 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1212 {
1213         /* Something wrong if we are trying to remove same group twice */
1214         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1215
1216         hlist_del_init(&cfqg->cfqd_node);
1217
1218         BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1219         cfqd->nr_blkcg_linked_grps--;
1220
1221         /*
1222          * Put the reference taken at the time of creation so that when all
1223          * queues are gone, group can be destroyed.
1224          */
1225         cfq_put_cfqg(cfqg);
1226 }
1227
1228 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1229 {
1230         struct hlist_node *pos, *n;
1231         struct cfq_group *cfqg;
1232
1233         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1234                 /*
1235                  * If cgroup removal path got to blk_group first and removed
1236                  * it from cgroup list, then it will take care of destroying
1237                  * cfqg also.
1238                  */
1239                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1240                         cfq_destroy_cfqg(cfqd, cfqg);
1241         }
1242 }
1243
1244 /*
1245  * Blk cgroup controller notification saying that blkio_group object is being
1246  * delinked as associated cgroup object is going away. That also means that
1247  * no new IO will come in this group. So get rid of this group as soon as
1248  * any pending IO in the group is finished.
1249  *
1250  * This function is called under rcu_read_lock(). key is the rcu protected
1251  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1252  * read lock.
1253  *
1254  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1255  * it should not be NULL as even if elevator was exiting, cgroup deltion
1256  * path got to it first.
1257  */
1258 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1259 {
1260         unsigned long  flags;
1261         struct cfq_data *cfqd = key;
1262
1263         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1264         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1265         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1266 }
1267
1268 #else /* GROUP_IOSCHED */
1269 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1270 {
1271         return &cfqd->root_group;
1272 }
1273
1274 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1275 {
1276         return cfqg;
1277 }
1278
1279 static inline void
1280 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1281         cfqq->cfqg = cfqg;
1282 }
1283
1284 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1285 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1286
1287 #endif /* GROUP_IOSCHED */
1288
1289 /*
1290  * The cfqd->service_trees holds all pending cfq_queue's that have
1291  * requests waiting to be processed. It is sorted in the order that
1292  * we will service the queues.
1293  */
1294 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1295                                  bool add_front)
1296 {
1297         struct rb_node **p, *parent;
1298         struct cfq_queue *__cfqq;
1299         unsigned long rb_key;
1300         struct cfq_rb_root *service_tree;
1301         int left;
1302         int new_cfqq = 1;
1303
1304         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1305                                                 cfqq_type(cfqq));
1306         if (cfq_class_idle(cfqq)) {
1307                 rb_key = CFQ_IDLE_DELAY;
1308                 parent = rb_last(&service_tree->rb);
1309                 if (parent && parent != &cfqq->rb_node) {
1310                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1311                         rb_key += __cfqq->rb_key;
1312                 } else
1313                         rb_key += jiffies;
1314         } else if (!add_front) {
1315                 /*
1316                  * Get our rb key offset. Subtract any residual slice
1317                  * value carried from last service. A negative resid
1318                  * count indicates slice overrun, and this should position
1319                  * the next service time further away in the tree.
1320                  */
1321                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1322                 rb_key -= cfqq->slice_resid;
1323                 cfqq->slice_resid = 0;
1324         } else {
1325                 rb_key = -HZ;
1326                 __cfqq = cfq_rb_first(service_tree);
1327                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1328         }
1329
1330         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1331                 new_cfqq = 0;
1332                 /*
1333                  * same position, nothing more to do
1334                  */
1335                 if (rb_key == cfqq->rb_key &&
1336                     cfqq->service_tree == service_tree)
1337                         return;
1338
1339                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1340                 cfqq->service_tree = NULL;
1341         }
1342
1343         left = 1;
1344         parent = NULL;
1345         cfqq->service_tree = service_tree;
1346         p = &service_tree->rb.rb_node;
1347         while (*p) {
1348                 struct rb_node **n;
1349
1350                 parent = *p;
1351                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1352
1353                 /*
1354                  * sort by key, that represents service time.
1355                  */
1356                 if (time_before(rb_key, __cfqq->rb_key))
1357                         n = &(*p)->rb_left;
1358                 else {
1359                         n = &(*p)->rb_right;
1360                         left = 0;
1361                 }
1362
1363                 p = n;
1364         }
1365
1366         if (left)
1367                 service_tree->left = &cfqq->rb_node;
1368
1369         cfqq->rb_key = rb_key;
1370         rb_link_node(&cfqq->rb_node, parent, p);
1371         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1372         service_tree->count++;
1373         if (add_front || !new_cfqq)
1374                 return;
1375         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1376 }
1377
1378 static struct cfq_queue *
1379 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1380                      sector_t sector, struct rb_node **ret_parent,
1381                      struct rb_node ***rb_link)
1382 {
1383         struct rb_node **p, *parent;
1384         struct cfq_queue *cfqq = NULL;
1385
1386         parent = NULL;
1387         p = &root->rb_node;
1388         while (*p) {
1389                 struct rb_node **n;
1390
1391                 parent = *p;
1392                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1393
1394                 /*
1395                  * Sort strictly based on sector.  Smallest to the left,
1396                  * largest to the right.
1397                  */
1398                 if (sector > blk_rq_pos(cfqq->next_rq))
1399                         n = &(*p)->rb_right;
1400                 else if (sector < blk_rq_pos(cfqq->next_rq))
1401                         n = &(*p)->rb_left;
1402                 else
1403                         break;
1404                 p = n;
1405                 cfqq = NULL;
1406         }
1407
1408         *ret_parent = parent;
1409         if (rb_link)
1410                 *rb_link = p;
1411         return cfqq;
1412 }
1413
1414 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1415 {
1416         struct rb_node **p, *parent;
1417         struct cfq_queue *__cfqq;
1418
1419         if (cfqq->p_root) {
1420                 rb_erase(&cfqq->p_node, cfqq->p_root);
1421                 cfqq->p_root = NULL;
1422         }
1423
1424         if (cfq_class_idle(cfqq))
1425                 return;
1426         if (!cfqq->next_rq)
1427                 return;
1428
1429         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1430         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1431                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1432         if (!__cfqq) {
1433                 rb_link_node(&cfqq->p_node, parent, p);
1434                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1435         } else
1436                 cfqq->p_root = NULL;
1437 }
1438
1439 /*
1440  * Update cfqq's position in the service tree.
1441  */
1442 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1443 {
1444         /*
1445          * Resorting requires the cfqq to be on the RR list already.
1446          */
1447         if (cfq_cfqq_on_rr(cfqq)) {
1448                 cfq_service_tree_add(cfqd, cfqq, 0);
1449                 cfq_prio_tree_add(cfqd, cfqq);
1450         }
1451 }
1452
1453 /*
1454  * add to busy list of queues for service, trying to be fair in ordering
1455  * the pending list according to last request service
1456  */
1457 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1458 {
1459         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1460         BUG_ON(cfq_cfqq_on_rr(cfqq));
1461         cfq_mark_cfqq_on_rr(cfqq);
1462         cfqd->busy_queues++;
1463         if (cfq_cfqq_sync(cfqq))
1464                 cfqd->busy_sync_queues++;
1465
1466         cfq_resort_rr_list(cfqd, cfqq);
1467 }
1468
1469 /*
1470  * Called when the cfqq no longer has requests pending, remove it from
1471  * the service tree.
1472  */
1473 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1474 {
1475         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1476         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1477         cfq_clear_cfqq_on_rr(cfqq);
1478
1479         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1480                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1481                 cfqq->service_tree = NULL;
1482         }
1483         if (cfqq->p_root) {
1484                 rb_erase(&cfqq->p_node, cfqq->p_root);
1485                 cfqq->p_root = NULL;
1486         }
1487
1488         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1489         BUG_ON(!cfqd->busy_queues);
1490         cfqd->busy_queues--;
1491         if (cfq_cfqq_sync(cfqq))
1492                 cfqd->busy_sync_queues--;
1493 }
1494
1495 /*
1496  * rb tree support functions
1497  */
1498 static void cfq_del_rq_rb(struct request *rq)
1499 {
1500         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1501         const int sync = rq_is_sync(rq);
1502
1503         BUG_ON(!cfqq->queued[sync]);
1504         cfqq->queued[sync]--;
1505
1506         elv_rb_del(&cfqq->sort_list, rq);
1507
1508         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1509                 /*
1510                  * Queue will be deleted from service tree when we actually
1511                  * expire it later. Right now just remove it from prio tree
1512                  * as it is empty.
1513                  */
1514                 if (cfqq->p_root) {
1515                         rb_erase(&cfqq->p_node, cfqq->p_root);
1516                         cfqq->p_root = NULL;
1517                 }
1518         }
1519 }
1520
1521 static void cfq_add_rq_rb(struct request *rq)
1522 {
1523         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1524         struct cfq_data *cfqd = cfqq->cfqd;
1525         struct request *prev;
1526
1527         cfqq->queued[rq_is_sync(rq)]++;
1528
1529         elv_rb_add(&cfqq->sort_list, rq);
1530
1531         if (!cfq_cfqq_on_rr(cfqq))
1532                 cfq_add_cfqq_rr(cfqd, cfqq);
1533
1534         /*
1535          * check if this request is a better next-serve candidate
1536          */
1537         prev = cfqq->next_rq;
1538         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1539
1540         /*
1541          * adjust priority tree position, if ->next_rq changes
1542          */
1543         if (prev != cfqq->next_rq)
1544                 cfq_prio_tree_add(cfqd, cfqq);
1545
1546         BUG_ON(!cfqq->next_rq);
1547 }
1548
1549 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1550 {
1551         elv_rb_del(&cfqq->sort_list, rq);
1552         cfqq->queued[rq_is_sync(rq)]--;
1553         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1554                                         rq_data_dir(rq), rq_is_sync(rq));
1555         cfq_add_rq_rb(rq);
1556         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1557                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1558                         rq_is_sync(rq));
1559 }
1560
1561 static struct request *
1562 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1563 {
1564         struct task_struct *tsk = current;
1565         struct cfq_io_cq *cic;
1566         struct cfq_queue *cfqq;
1567
1568         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1569         if (!cic)
1570                 return NULL;
1571
1572         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1573         if (cfqq) {
1574                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1575
1576                 return elv_rb_find(&cfqq->sort_list, sector);
1577         }
1578
1579         return NULL;
1580 }
1581
1582 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1583 {
1584         struct cfq_data *cfqd = q->elevator->elevator_data;
1585
1586         cfqd->rq_in_driver++;
1587         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1588                                                 cfqd->rq_in_driver);
1589
1590         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1591 }
1592
1593 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1594 {
1595         struct cfq_data *cfqd = q->elevator->elevator_data;
1596
1597         WARN_ON(!cfqd->rq_in_driver);
1598         cfqd->rq_in_driver--;
1599         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1600                                                 cfqd->rq_in_driver);
1601 }
1602
1603 static void cfq_remove_request(struct request *rq)
1604 {
1605         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1606
1607         if (cfqq->next_rq == rq)
1608                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1609
1610         list_del_init(&rq->queuelist);
1611         cfq_del_rq_rb(rq);
1612
1613         cfqq->cfqd->rq_queued--;
1614         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1615                                         rq_data_dir(rq), rq_is_sync(rq));
1616         if (rq->cmd_flags & REQ_PRIO) {
1617                 WARN_ON(!cfqq->prio_pending);
1618                 cfqq->prio_pending--;
1619         }
1620 }
1621
1622 static int cfq_merge(struct request_queue *q, struct request **req,
1623                      struct bio *bio)
1624 {
1625         struct cfq_data *cfqd = q->elevator->elevator_data;
1626         struct request *__rq;
1627
1628         __rq = cfq_find_rq_fmerge(cfqd, bio);
1629         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1630                 *req = __rq;
1631                 return ELEVATOR_FRONT_MERGE;
1632         }
1633
1634         return ELEVATOR_NO_MERGE;
1635 }
1636
1637 static void cfq_merged_request(struct request_queue *q, struct request *req,
1638                                int type)
1639 {
1640         if (type == ELEVATOR_FRONT_MERGE) {
1641                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1642
1643                 cfq_reposition_rq_rb(cfqq, req);
1644         }
1645 }
1646
1647 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1648                                 struct bio *bio)
1649 {
1650         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1651                                         bio_data_dir(bio), cfq_bio_sync(bio));
1652 }
1653
1654 static void
1655 cfq_merged_requests(struct request_queue *q, struct request *rq,
1656                     struct request *next)
1657 {
1658         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1659         /*
1660          * reposition in fifo if next is older than rq
1661          */
1662         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1663             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1664                 list_move(&rq->queuelist, &next->queuelist);
1665                 rq_set_fifo_time(rq, rq_fifo_time(next));
1666         }
1667
1668         if (cfqq->next_rq == next)
1669                 cfqq->next_rq = rq;
1670         cfq_remove_request(next);
1671         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1672                                         rq_data_dir(next), rq_is_sync(next));
1673 }
1674
1675 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1676                            struct bio *bio)
1677 {
1678         struct cfq_data *cfqd = q->elevator->elevator_data;
1679         struct cfq_io_cq *cic;
1680         struct cfq_queue *cfqq;
1681
1682         /*
1683          * Disallow merge of a sync bio into an async request.
1684          */
1685         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1686                 return false;
1687
1688         /*
1689          * Lookup the cfqq that this bio will be queued with and allow
1690          * merge only if rq is queued there.  This function can be called
1691          * from plug merge without queue_lock.  In such cases, ioc of @rq
1692          * and %current are guaranteed to be equal.  Avoid lookup which
1693          * requires queue_lock by using @rq's cic.
1694          */
1695         if (current->io_context == RQ_CIC(rq)->icq.ioc) {
1696                 cic = RQ_CIC(rq);
1697         } else {
1698                 cic = cfq_cic_lookup(cfqd, current->io_context);
1699                 if (!cic)
1700                         return false;
1701         }
1702
1703         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1704         return cfqq == RQ_CFQQ(rq);
1705 }
1706
1707 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1708 {
1709         del_timer(&cfqd->idle_slice_timer);
1710         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1711 }
1712
1713 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1714                                    struct cfq_queue *cfqq)
1715 {
1716         if (cfqq) {
1717                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1718                                 cfqd->serving_prio, cfqd->serving_type);
1719                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1720                 cfqq->slice_start = 0;
1721                 cfqq->dispatch_start = jiffies;
1722                 cfqq->allocated_slice = 0;
1723                 cfqq->slice_end = 0;
1724                 cfqq->slice_dispatch = 0;
1725                 cfqq->nr_sectors = 0;
1726
1727                 cfq_clear_cfqq_wait_request(cfqq);
1728                 cfq_clear_cfqq_must_dispatch(cfqq);
1729                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1730                 cfq_clear_cfqq_fifo_expire(cfqq);
1731                 cfq_mark_cfqq_slice_new(cfqq);
1732
1733                 cfq_del_timer(cfqd, cfqq);
1734         }
1735
1736         cfqd->active_queue = cfqq;
1737 }
1738
1739 /*
1740  * current cfqq expired its slice (or was too idle), select new one
1741  */
1742 static void
1743 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1744                     bool timed_out)
1745 {
1746         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1747
1748         if (cfq_cfqq_wait_request(cfqq))
1749                 cfq_del_timer(cfqd, cfqq);
1750
1751         cfq_clear_cfqq_wait_request(cfqq);
1752         cfq_clear_cfqq_wait_busy(cfqq);
1753
1754         /*
1755          * If this cfqq is shared between multiple processes, check to
1756          * make sure that those processes are still issuing I/Os within
1757          * the mean seek distance.  If not, it may be time to break the
1758          * queues apart again.
1759          */
1760         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1761                 cfq_mark_cfqq_split_coop(cfqq);
1762
1763         /*
1764          * store what was left of this slice, if the queue idled/timed out
1765          */
1766         if (timed_out) {
1767                 if (cfq_cfqq_slice_new(cfqq))
1768                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1769                 else
1770                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1771                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1772         }
1773
1774         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1775
1776         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1777                 cfq_del_cfqq_rr(cfqd, cfqq);
1778
1779         cfq_resort_rr_list(cfqd, cfqq);
1780
1781         if (cfqq == cfqd->active_queue)
1782                 cfqd->active_queue = NULL;
1783
1784         if (cfqd->active_cic) {
1785                 put_io_context(cfqd->active_cic->icq.ioc, cfqd->queue);
1786                 cfqd->active_cic = NULL;
1787         }
1788 }
1789
1790 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1791 {
1792         struct cfq_queue *cfqq = cfqd->active_queue;
1793
1794         if (cfqq)
1795                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1796 }
1797
1798 /*
1799  * Get next queue for service. Unless we have a queue preemption,
1800  * we'll simply select the first cfqq in the service tree.
1801  */
1802 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1803 {
1804         struct cfq_rb_root *service_tree =
1805                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1806                                         cfqd->serving_type);
1807
1808         if (!cfqd->rq_queued)
1809                 return NULL;
1810
1811         /* There is nothing to dispatch */
1812         if (!service_tree)
1813                 return NULL;
1814         if (RB_EMPTY_ROOT(&service_tree->rb))
1815                 return NULL;
1816         return cfq_rb_first(service_tree);
1817 }
1818
1819 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1820 {
1821         struct cfq_group *cfqg;
1822         struct cfq_queue *cfqq;
1823         int i, j;
1824         struct cfq_rb_root *st;
1825
1826         if (!cfqd->rq_queued)
1827                 return NULL;
1828
1829         cfqg = cfq_get_next_cfqg(cfqd);
1830         if (!cfqg)
1831                 return NULL;
1832
1833         for_each_cfqg_st(cfqg, i, j, st)
1834                 if ((cfqq = cfq_rb_first(st)) != NULL)
1835                         return cfqq;
1836         return NULL;
1837 }
1838
1839 /*
1840  * Get and set a new active queue for service.
1841  */
1842 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1843                                               struct cfq_queue *cfqq)
1844 {
1845         if (!cfqq)
1846                 cfqq = cfq_get_next_queue(cfqd);
1847
1848         __cfq_set_active_queue(cfqd, cfqq);
1849         return cfqq;
1850 }
1851
1852 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1853                                           struct request *rq)
1854 {
1855         if (blk_rq_pos(rq) >= cfqd->last_position)
1856                 return blk_rq_pos(rq) - cfqd->last_position;
1857         else
1858                 return cfqd->last_position - blk_rq_pos(rq);
1859 }
1860
1861 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1862                                struct request *rq)
1863 {
1864         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1865 }
1866
1867 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1868                                     struct cfq_queue *cur_cfqq)
1869 {
1870         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1871         struct rb_node *parent, *node;
1872         struct cfq_queue *__cfqq;
1873         sector_t sector = cfqd->last_position;
1874
1875         if (RB_EMPTY_ROOT(root))
1876                 return NULL;
1877
1878         /*
1879          * First, if we find a request starting at the end of the last
1880          * request, choose it.
1881          */
1882         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1883         if (__cfqq)
1884                 return __cfqq;
1885
1886         /*
1887          * If the exact sector wasn't found, the parent of the NULL leaf
1888          * will contain the closest sector.
1889          */
1890         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1891         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1892                 return __cfqq;
1893
1894         if (blk_rq_pos(__cfqq->next_rq) < sector)
1895                 node = rb_next(&__cfqq->p_node);
1896         else
1897                 node = rb_prev(&__cfqq->p_node);
1898         if (!node)
1899                 return NULL;
1900
1901         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1902         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1903                 return __cfqq;
1904
1905         return NULL;
1906 }
1907
1908 /*
1909  * cfqd - obvious
1910  * cur_cfqq - passed in so that we don't decide that the current queue is
1911  *            closely cooperating with itself.
1912  *
1913  * So, basically we're assuming that that cur_cfqq has dispatched at least
1914  * one request, and that cfqd->last_position reflects a position on the disk
1915  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1916  * assumption.
1917  */
1918 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1919                                               struct cfq_queue *cur_cfqq)
1920 {
1921         struct cfq_queue *cfqq;
1922
1923         if (cfq_class_idle(cur_cfqq))
1924                 return NULL;
1925         if (!cfq_cfqq_sync(cur_cfqq))
1926                 return NULL;
1927         if (CFQQ_SEEKY(cur_cfqq))
1928                 return NULL;
1929
1930         /*
1931          * Don't search priority tree if it's the only queue in the group.
1932          */
1933         if (cur_cfqq->cfqg->nr_cfqq == 1)
1934                 return NULL;
1935
1936         /*
1937          * We should notice if some of the queues are cooperating, eg
1938          * working closely on the same area of the disk. In that case,
1939          * we can group them together and don't waste time idling.
1940          */
1941         cfqq = cfqq_close(cfqd, cur_cfqq);
1942         if (!cfqq)
1943                 return NULL;
1944
1945         /* If new queue belongs to different cfq_group, don't choose it */
1946         if (cur_cfqq->cfqg != cfqq->cfqg)
1947                 return NULL;
1948
1949         /*
1950          * It only makes sense to merge sync queues.
1951          */
1952         if (!cfq_cfqq_sync(cfqq))
1953                 return NULL;
1954         if (CFQQ_SEEKY(cfqq))
1955                 return NULL;
1956
1957         /*
1958          * Do not merge queues of different priority classes
1959          */
1960         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1961                 return NULL;
1962
1963         return cfqq;
1964 }
1965
1966 /*
1967  * Determine whether we should enforce idle window for this queue.
1968  */
1969
1970 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1971 {
1972         enum wl_prio_t prio = cfqq_prio(cfqq);
1973         struct cfq_rb_root *service_tree = cfqq->service_tree;
1974
1975         BUG_ON(!service_tree);
1976         BUG_ON(!service_tree->count);
1977
1978         if (!cfqd->cfq_slice_idle)
1979                 return false;
1980
1981         /* We never do for idle class queues. */
1982         if (prio == IDLE_WORKLOAD)
1983                 return false;
1984
1985         /* We do for queues that were marked with idle window flag. */
1986         if (cfq_cfqq_idle_window(cfqq) &&
1987            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1988                 return true;
1989
1990         /*
1991          * Otherwise, we do only if they are the last ones
1992          * in their service tree.
1993          */
1994         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1995            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1996                 return true;
1997         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1998                         service_tree->count);
1999         return false;
2000 }
2001
2002 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2003 {
2004         struct cfq_queue *cfqq = cfqd->active_queue;
2005         struct cfq_io_cq *cic;
2006         unsigned long sl, group_idle = 0;
2007
2008         /*
2009          * SSD device without seek penalty, disable idling. But only do so
2010          * for devices that support queuing, otherwise we still have a problem
2011          * with sync vs async workloads.
2012          */
2013         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2014                 return;
2015
2016         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2017         WARN_ON(cfq_cfqq_slice_new(cfqq));
2018
2019         /*
2020          * idle is disabled, either manually or by past process history
2021          */
2022         if (!cfq_should_idle(cfqd, cfqq)) {
2023                 /* no queue idling. Check for group idling */
2024                 if (cfqd->cfq_group_idle)
2025                         group_idle = cfqd->cfq_group_idle;
2026                 else
2027                         return;
2028         }
2029
2030         /*
2031          * still active requests from this queue, don't idle
2032          */
2033         if (cfqq->dispatched)
2034                 return;
2035
2036         /*
2037          * task has exited, don't wait
2038          */
2039         cic = cfqd->active_cic;
2040         if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
2041                 return;
2042
2043         /*
2044          * If our average think time is larger than the remaining time
2045          * slice, then don't idle. This avoids overrunning the allotted
2046          * time slice.
2047          */
2048         if (sample_valid(cic->ttime.ttime_samples) &&
2049             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2050                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2051                              cic->ttime.ttime_mean);
2052                 return;
2053         }
2054
2055         /* There are other queues in the group, don't do group idle */
2056         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2057                 return;
2058
2059         cfq_mark_cfqq_wait_request(cfqq);
2060
2061         if (group_idle)
2062                 sl = cfqd->cfq_group_idle;
2063         else
2064                 sl = cfqd->cfq_slice_idle;
2065
2066         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2067         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2068         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2069                         group_idle ? 1 : 0);
2070 }
2071
2072 /*
2073  * Move request from internal lists to the request queue dispatch list.
2074  */
2075 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2076 {
2077         struct cfq_data *cfqd = q->elevator->elevator_data;
2078         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2079
2080         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2081
2082         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2083         cfq_remove_request(rq);
2084         cfqq->dispatched++;
2085         (RQ_CFQG(rq))->dispatched++;
2086         elv_dispatch_sort(q, rq);
2087
2088         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2089         cfqq->nr_sectors += blk_rq_sectors(rq);
2090         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2091                                         rq_data_dir(rq), rq_is_sync(rq));
2092 }
2093
2094 /*
2095  * return expired entry, or NULL to just start from scratch in rbtree
2096  */
2097 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2098 {
2099         struct request *rq = NULL;
2100
2101         if (cfq_cfqq_fifo_expire(cfqq))
2102                 return NULL;
2103
2104         cfq_mark_cfqq_fifo_expire(cfqq);
2105
2106         if (list_empty(&cfqq->fifo))
2107                 return NULL;
2108
2109         rq = rq_entry_fifo(cfqq->fifo.next);
2110         if (time_before(jiffies, rq_fifo_time(rq)))
2111                 rq = NULL;
2112
2113         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2114         return rq;
2115 }
2116
2117 static inline int
2118 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2119 {
2120         const int base_rq = cfqd->cfq_slice_async_rq;
2121
2122         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2123
2124         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2125 }
2126
2127 /*
2128  * Must be called with the queue_lock held.
2129  */
2130 static int cfqq_process_refs(struct cfq_queue *cfqq)
2131 {
2132         int process_refs, io_refs;
2133
2134         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2135         process_refs = cfqq->ref - io_refs;
2136         BUG_ON(process_refs < 0);
2137         return process_refs;
2138 }
2139
2140 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2141 {
2142         int process_refs, new_process_refs;
2143         struct cfq_queue *__cfqq;
2144
2145         /*
2146          * If there are no process references on the new_cfqq, then it is
2147          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2148          * chain may have dropped their last reference (not just their
2149          * last process reference).
2150          */
2151         if (!cfqq_process_refs(new_cfqq))
2152                 return;
2153
2154         /* Avoid a circular list and skip interim queue merges */
2155         while ((__cfqq = new_cfqq->new_cfqq)) {
2156                 if (__cfqq == cfqq)
2157                         return;
2158                 new_cfqq = __cfqq;
2159         }
2160
2161         process_refs = cfqq_process_refs(cfqq);
2162         new_process_refs = cfqq_process_refs(new_cfqq);
2163         /*
2164          * If the process for the cfqq has gone away, there is no
2165          * sense in merging the queues.
2166          */
2167         if (process_refs == 0 || new_process_refs == 0)
2168                 return;
2169
2170         /*
2171          * Merge in the direction of the lesser amount of work.
2172          */
2173         if (new_process_refs >= process_refs) {
2174                 cfqq->new_cfqq = new_cfqq;
2175                 new_cfqq->ref += process_refs;
2176         } else {
2177                 new_cfqq->new_cfqq = cfqq;
2178                 cfqq->ref += new_process_refs;
2179         }
2180 }
2181
2182 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2183                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2184 {
2185         struct cfq_queue *queue;
2186         int i;
2187         bool key_valid = false;
2188         unsigned long lowest_key = 0;
2189         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2190
2191         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2192                 /* select the one with lowest rb_key */
2193                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2194                 if (queue &&
2195                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2196                         lowest_key = queue->rb_key;
2197                         cur_best = i;
2198                         key_valid = true;
2199                 }
2200         }
2201
2202         return cur_best;
2203 }
2204
2205 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2206 {
2207         unsigned slice;
2208         unsigned count;
2209         struct cfq_rb_root *st;
2210         unsigned group_slice;
2211         enum wl_prio_t original_prio = cfqd->serving_prio;
2212
2213         /* Choose next priority. RT > BE > IDLE */
2214         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2215                 cfqd->serving_prio = RT_WORKLOAD;
2216         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2217                 cfqd->serving_prio = BE_WORKLOAD;
2218         else {
2219                 cfqd->serving_prio = IDLE_WORKLOAD;
2220                 cfqd->workload_expires = jiffies + 1;
2221                 return;
2222         }
2223
2224         if (original_prio != cfqd->serving_prio)
2225                 goto new_workload;
2226
2227         /*
2228          * For RT and BE, we have to choose also the type
2229          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2230          * expiration time
2231          */
2232         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2233         count = st->count;
2234
2235         /*
2236          * check workload expiration, and that we still have other queues ready
2237          */
2238         if (count && !time_after(jiffies, cfqd->workload_expires))
2239                 return;
2240
2241 new_workload:
2242         /* otherwise select new workload type */
2243         cfqd->serving_type =
2244                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2245         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2246         count = st->count;
2247
2248         /*
2249          * the workload slice is computed as a fraction of target latency
2250          * proportional to the number of queues in that workload, over
2251          * all the queues in the same priority class
2252          */
2253         group_slice = cfq_group_slice(cfqd, cfqg);
2254
2255         slice = group_slice * count /
2256                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2257                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2258
2259         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2260                 unsigned int tmp;
2261
2262                 /*
2263                  * Async queues are currently system wide. Just taking
2264                  * proportion of queues with-in same group will lead to higher
2265                  * async ratio system wide as generally root group is going
2266                  * to have higher weight. A more accurate thing would be to
2267                  * calculate system wide asnc/sync ratio.
2268                  */
2269                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2270                 tmp = tmp/cfqd->busy_queues;
2271                 slice = min_t(unsigned, slice, tmp);
2272
2273                 /* async workload slice is scaled down according to
2274                  * the sync/async slice ratio. */
2275                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2276         } else
2277                 /* sync workload slice is at least 2 * cfq_slice_idle */
2278                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2279
2280         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2281         cfq_log(cfqd, "workload slice:%d", slice);
2282         cfqd->workload_expires = jiffies + slice;
2283 }
2284
2285 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2286 {
2287         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2288         struct cfq_group *cfqg;
2289
2290         if (RB_EMPTY_ROOT(&st->rb))
2291                 return NULL;
2292         cfqg = cfq_rb_first_group(st);
2293         update_min_vdisktime(st);
2294         return cfqg;
2295 }
2296
2297 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2298 {
2299         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2300
2301         cfqd->serving_group = cfqg;
2302
2303         /* Restore the workload type data */
2304         if (cfqg->saved_workload_slice) {
2305                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2306                 cfqd->serving_type = cfqg->saved_workload;
2307                 cfqd->serving_prio = cfqg->saved_serving_prio;
2308         } else
2309                 cfqd->workload_expires = jiffies - 1;
2310
2311         choose_service_tree(cfqd, cfqg);
2312 }
2313
2314 /*
2315  * Select a queue for service. If we have a current active queue,
2316  * check whether to continue servicing it, or retrieve and set a new one.
2317  */
2318 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2319 {
2320         struct cfq_queue *cfqq, *new_cfqq = NULL;
2321
2322         cfqq = cfqd->active_queue;
2323         if (!cfqq)
2324                 goto new_queue;
2325
2326         if (!cfqd->rq_queued)
2327                 return NULL;
2328
2329         /*
2330          * We were waiting for group to get backlogged. Expire the queue
2331          */
2332         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2333                 goto expire;
2334
2335         /*
2336          * The active queue has run out of time, expire it and select new.
2337          */
2338         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2339                 /*
2340                  * If slice had not expired at the completion of last request
2341                  * we might not have turned on wait_busy flag. Don't expire
2342                  * the queue yet. Allow the group to get backlogged.
2343                  *
2344                  * The very fact that we have used the slice, that means we
2345                  * have been idling all along on this queue and it should be
2346                  * ok to wait for this request to complete.
2347                  */
2348                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2349                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2350                         cfqq = NULL;
2351                         goto keep_queue;
2352                 } else
2353                         goto check_group_idle;
2354         }
2355
2356         /*
2357          * The active queue has requests and isn't expired, allow it to
2358          * dispatch.
2359          */
2360         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2361                 goto keep_queue;
2362
2363         /*
2364          * If another queue has a request waiting within our mean seek
2365          * distance, let it run.  The expire code will check for close
2366          * cooperators and put the close queue at the front of the service
2367          * tree.  If possible, merge the expiring queue with the new cfqq.
2368          */
2369         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2370         if (new_cfqq) {
2371                 if (!cfqq->new_cfqq)
2372                         cfq_setup_merge(cfqq, new_cfqq);
2373                 goto expire;
2374         }
2375
2376         /*
2377          * No requests pending. If the active queue still has requests in
2378          * flight or is idling for a new request, allow either of these
2379          * conditions to happen (or time out) before selecting a new queue.
2380          */
2381         if (timer_pending(&cfqd->idle_slice_timer)) {
2382                 cfqq = NULL;
2383                 goto keep_queue;
2384         }
2385
2386         /*
2387          * This is a deep seek queue, but the device is much faster than
2388          * the queue can deliver, don't idle
2389          **/
2390         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2391             (cfq_cfqq_slice_new(cfqq) ||
2392             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2393                 cfq_clear_cfqq_deep(cfqq);
2394                 cfq_clear_cfqq_idle_window(cfqq);
2395         }
2396
2397         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2398                 cfqq = NULL;
2399                 goto keep_queue;
2400         }
2401
2402         /*
2403          * If group idle is enabled and there are requests dispatched from
2404          * this group, wait for requests to complete.
2405          */
2406 check_group_idle:
2407         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2408             cfqq->cfqg->dispatched &&
2409             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2410                 cfqq = NULL;
2411                 goto keep_queue;
2412         }
2413
2414 expire:
2415         cfq_slice_expired(cfqd, 0);
2416 new_queue:
2417         /*
2418          * Current queue expired. Check if we have to switch to a new
2419          * service tree
2420          */
2421         if (!new_cfqq)
2422                 cfq_choose_cfqg(cfqd);
2423
2424         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2425 keep_queue:
2426         return cfqq;
2427 }
2428
2429 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2430 {
2431         int dispatched = 0;
2432
2433         while (cfqq->next_rq) {
2434                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2435                 dispatched++;
2436         }
2437
2438         BUG_ON(!list_empty(&cfqq->fifo));
2439
2440         /* By default cfqq is not expired if it is empty. Do it explicitly */
2441         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2442         return dispatched;
2443 }
2444
2445 /*
2446  * Drain our current requests. Used for barriers and when switching
2447  * io schedulers on-the-fly.
2448  */
2449 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2450 {
2451         struct cfq_queue *cfqq;
2452         int dispatched = 0;
2453
2454         /* Expire the timeslice of the current active queue first */
2455         cfq_slice_expired(cfqd, 0);
2456         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2457                 __cfq_set_active_queue(cfqd, cfqq);
2458                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2459         }
2460
2461         BUG_ON(cfqd->busy_queues);
2462
2463         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2464         return dispatched;
2465 }
2466
2467 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2468         struct cfq_queue *cfqq)
2469 {
2470         /* the queue hasn't finished any request, can't estimate */
2471         if (cfq_cfqq_slice_new(cfqq))
2472                 return true;
2473         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2474                 cfqq->slice_end))
2475                 return true;
2476
2477         return false;
2478 }
2479
2480 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2481 {
2482         unsigned int max_dispatch;
2483
2484         /*
2485          * Drain async requests before we start sync IO
2486          */
2487         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2488                 return false;
2489
2490         /*
2491          * If this is an async queue and we have sync IO in flight, let it wait
2492          */
2493         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2494                 return false;
2495
2496         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2497         if (cfq_class_idle(cfqq))
2498                 max_dispatch = 1;
2499
2500         /*
2501          * Does this cfqq already have too much IO in flight?
2502          */
2503         if (cfqq->dispatched >= max_dispatch) {
2504                 bool promote_sync = false;
2505                 /*
2506                  * idle queue must always only have a single IO in flight
2507                  */
2508                 if (cfq_class_idle(cfqq))
2509                         return false;
2510
2511                 /*
2512                  * If there is only one sync queue
2513                  * we can ignore async queue here and give the sync
2514                  * queue no dispatch limit. The reason is a sync queue can
2515                  * preempt async queue, limiting the sync queue doesn't make
2516                  * sense. This is useful for aiostress test.
2517                  */
2518                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2519                         promote_sync = true;
2520
2521                 /*
2522                  * We have other queues, don't allow more IO from this one
2523                  */
2524                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2525                                 !promote_sync)
2526                         return false;
2527
2528                 /*
2529                  * Sole queue user, no limit
2530                  */
2531                 if (cfqd->busy_queues == 1 || promote_sync)
2532                         max_dispatch = -1;
2533                 else
2534                         /*
2535                          * Normally we start throttling cfqq when cfq_quantum/2
2536                          * requests have been dispatched. But we can drive
2537                          * deeper queue depths at the beginning of slice
2538                          * subjected to upper limit of cfq_quantum.
2539                          * */
2540                         max_dispatch = cfqd->cfq_quantum;
2541         }
2542
2543         /*
2544          * Async queues must wait a bit before being allowed dispatch.
2545          * We also ramp up the dispatch depth gradually for async IO,
2546          * based on the last sync IO we serviced
2547          */
2548         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2549                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2550                 unsigned int depth;
2551
2552                 depth = last_sync / cfqd->cfq_slice[1];
2553                 if (!depth && !cfqq->dispatched)
2554                         depth = 1;
2555                 if (depth < max_dispatch)
2556                         max_dispatch = depth;
2557         }
2558
2559         /*
2560          * If we're below the current max, allow a dispatch
2561          */
2562         return cfqq->dispatched < max_dispatch;
2563 }
2564
2565 /*
2566  * Dispatch a request from cfqq, moving them to the request queue
2567  * dispatch list.
2568  */
2569 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2570 {
2571         struct request *rq;
2572
2573         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2574
2575         if (!cfq_may_dispatch(cfqd, cfqq))
2576                 return false;
2577
2578         /*
2579          * follow expired path, else get first next available
2580          */
2581         rq = cfq_check_fifo(cfqq);
2582         if (!rq)
2583                 rq = cfqq->next_rq;
2584
2585         /*
2586          * insert request into driver dispatch list
2587          */
2588         cfq_dispatch_insert(cfqd->queue, rq);
2589
2590         if (!cfqd->active_cic) {
2591                 struct cfq_io_cq *cic = RQ_CIC(rq);
2592
2593                 atomic_long_inc(&cic->icq.ioc->refcount);
2594                 cfqd->active_cic = cic;
2595         }
2596
2597         return true;
2598 }
2599
2600 /*
2601  * Find the cfqq that we need to service and move a request from that to the
2602  * dispatch list
2603  */
2604 static int cfq_dispatch_requests(struct request_queue *q, int force)
2605 {
2606         struct cfq_data *cfqd = q->elevator->elevator_data;
2607         struct cfq_queue *cfqq;
2608
2609         if (!cfqd->busy_queues)
2610                 return 0;
2611
2612         if (unlikely(force))
2613                 return cfq_forced_dispatch(cfqd);
2614
2615         cfqq = cfq_select_queue(cfqd);
2616         if (!cfqq)
2617                 return 0;
2618
2619         /*
2620          * Dispatch a request from this cfqq, if it is allowed
2621          */
2622         if (!cfq_dispatch_request(cfqd, cfqq))
2623                 return 0;
2624
2625         cfqq->slice_dispatch++;
2626         cfq_clear_cfqq_must_dispatch(cfqq);
2627
2628         /*
2629          * expire an async queue immediately if it has used up its slice. idle
2630          * queue always expire after 1 dispatch round.
2631          */
2632         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2633             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2634             cfq_class_idle(cfqq))) {
2635                 cfqq->slice_end = jiffies + 1;
2636                 cfq_slice_expired(cfqd, 0);
2637         }
2638
2639         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2640         return 1;
2641 }
2642
2643 /*
2644  * task holds one reference to the queue, dropped when task exits. each rq
2645  * in-flight on this queue also holds a reference, dropped when rq is freed.
2646  *
2647  * Each cfq queue took a reference on the parent group. Drop it now.
2648  * queue lock must be held here.
2649  */
2650 static void cfq_put_queue(struct cfq_queue *cfqq)
2651 {
2652         struct cfq_data *cfqd = cfqq->cfqd;
2653         struct cfq_group *cfqg;
2654
2655         BUG_ON(cfqq->ref <= 0);
2656
2657         cfqq->ref--;
2658         if (cfqq->ref)
2659                 return;
2660
2661         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2662         BUG_ON(rb_first(&cfqq->sort_list));
2663         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2664         cfqg = cfqq->cfqg;
2665
2666         if (unlikely(cfqd->active_queue == cfqq)) {
2667                 __cfq_slice_expired(cfqd, cfqq, 0);
2668                 cfq_schedule_dispatch(cfqd);
2669         }
2670
2671         BUG_ON(cfq_cfqq_on_rr(cfqq));
2672         kmem_cache_free(cfq_pool, cfqq);
2673         cfq_put_cfqg(cfqg);
2674 }
2675
2676 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2677 {
2678         struct cfq_queue *__cfqq, *next;
2679
2680         /*
2681          * If this queue was scheduled to merge with another queue, be
2682          * sure to drop the reference taken on that queue (and others in
2683          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2684          */
2685         __cfqq = cfqq->new_cfqq;
2686         while (__cfqq) {
2687                 if (__cfqq == cfqq) {
2688                         WARN(1, "cfqq->new_cfqq loop detected\n");
2689                         break;
2690                 }
2691                 next = __cfqq->new_cfqq;
2692                 cfq_put_queue(__cfqq);
2693                 __cfqq = next;
2694         }
2695 }
2696
2697 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2698 {
2699         if (unlikely(cfqq == cfqd->active_queue)) {
2700                 __cfq_slice_expired(cfqd, cfqq, 0);
2701                 cfq_schedule_dispatch(cfqd);
2702         }
2703
2704         cfq_put_cooperator(cfqq);
2705
2706         cfq_put_queue(cfqq);
2707 }
2708
2709 static void cfq_init_icq(struct io_cq *icq)
2710 {
2711         struct cfq_io_cq *cic = icq_to_cic(icq);
2712
2713         cic->ttime.last_end_request = jiffies;
2714 }
2715
2716 static void cfq_exit_icq(struct io_cq *icq)
2717 {
2718         struct cfq_io_cq *cic = icq_to_cic(icq);
2719         struct cfq_data *cfqd = cic_to_cfqd(cic);
2720
2721         if (cic->cfqq[BLK_RW_ASYNC]) {
2722                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2723                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2724         }
2725
2726         if (cic->cfqq[BLK_RW_SYNC]) {
2727                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2728                 cic->cfqq[BLK_RW_SYNC] = NULL;
2729         }
2730 }
2731
2732 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2733 {
2734         struct task_struct *tsk = current;
2735         int ioprio_class;
2736
2737         if (!cfq_cfqq_prio_changed(cfqq))
2738                 return;
2739
2740         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2741         switch (ioprio_class) {
2742         default:
2743                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2744         case IOPRIO_CLASS_NONE:
2745                 /*
2746                  * no prio set, inherit CPU scheduling settings
2747                  */
2748                 cfqq->ioprio = task_nice_ioprio(tsk);
2749                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2750                 break;
2751         case IOPRIO_CLASS_RT:
2752                 cfqq->ioprio = task_ioprio(ioc);
2753                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2754                 break;
2755         case IOPRIO_CLASS_BE:
2756                 cfqq->ioprio = task_ioprio(ioc);
2757                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2758                 break;
2759         case IOPRIO_CLASS_IDLE:
2760                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2761                 cfqq->ioprio = 7;
2762                 cfq_clear_cfqq_idle_window(cfqq);
2763                 break;
2764         }
2765
2766         /*
2767          * keep track of original prio settings in case we have to temporarily
2768          * elevate the priority of this queue
2769          */
2770         cfqq->org_ioprio = cfqq->ioprio;
2771         cfq_clear_cfqq_prio_changed(cfqq);
2772 }
2773
2774 static void changed_ioprio(struct cfq_io_cq *cic)
2775 {
2776         struct cfq_data *cfqd = cic_to_cfqd(cic);
2777         struct cfq_queue *cfqq;
2778
2779         if (unlikely(!cfqd))
2780                 return;
2781
2782         cfqq = cic->cfqq[BLK_RW_ASYNC];
2783         if (cfqq) {
2784                 struct cfq_queue *new_cfqq;
2785                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2786                                                 GFP_ATOMIC);
2787                 if (new_cfqq) {
2788                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2789                         cfq_put_queue(cfqq);
2790                 }
2791         }
2792
2793         cfqq = cic->cfqq[BLK_RW_SYNC];
2794         if (cfqq)
2795                 cfq_mark_cfqq_prio_changed(cfqq);
2796 }
2797
2798 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2799                           pid_t pid, bool is_sync)
2800 {
2801         RB_CLEAR_NODE(&cfqq->rb_node);
2802         RB_CLEAR_NODE(&cfqq->p_node);
2803         INIT_LIST_HEAD(&cfqq->fifo);
2804
2805         cfqq->ref = 0;
2806         cfqq->cfqd = cfqd;
2807
2808         cfq_mark_cfqq_prio_changed(cfqq);
2809
2810         if (is_sync) {
2811                 if (!cfq_class_idle(cfqq))
2812                         cfq_mark_cfqq_idle_window(cfqq);
2813                 cfq_mark_cfqq_sync(cfqq);
2814         }
2815         cfqq->pid = pid;
2816 }
2817
2818 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2819 static void changed_cgroup(struct cfq_io_cq *cic)
2820 {
2821         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2822         struct cfq_data *cfqd = cic_to_cfqd(cic);
2823         struct request_queue *q;
2824
2825         if (unlikely(!cfqd))
2826                 return;
2827
2828         q = cfqd->queue;
2829
2830         if (sync_cfqq) {
2831                 /*
2832                  * Drop reference to sync queue. A new sync queue will be
2833                  * assigned in new group upon arrival of a fresh request.
2834                  */
2835                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2836                 cic_set_cfqq(cic, NULL, 1);
2837                 cfq_put_queue(sync_cfqq);
2838         }
2839 }
2840 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2841
2842 static struct cfq_queue *
2843 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2844                      struct io_context *ioc, gfp_t gfp_mask)
2845 {
2846         struct cfq_queue *cfqq, *new_cfqq = NULL;
2847         struct cfq_io_cq *cic;
2848         struct cfq_group *cfqg;
2849
2850 retry:
2851         cfqg = cfq_get_cfqg(cfqd);
2852         cic = cfq_cic_lookup(cfqd, ioc);
2853         /* cic always exists here */
2854         cfqq = cic_to_cfqq(cic, is_sync);
2855
2856         /*
2857          * Always try a new alloc if we fell back to the OOM cfqq
2858          * originally, since it should just be a temporary situation.
2859          */
2860         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2861                 cfqq = NULL;
2862                 if (new_cfqq) {
2863                         cfqq = new_cfqq;
2864                         new_cfqq = NULL;
2865                 } else if (gfp_mask & __GFP_WAIT) {
2866                         spin_unlock_irq(cfqd->queue->queue_lock);
2867                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2868                                         gfp_mask | __GFP_ZERO,
2869                                         cfqd->queue->node);
2870                         spin_lock_irq(cfqd->queue->queue_lock);
2871                         if (new_cfqq)
2872                                 goto retry;
2873                 } else {
2874                         cfqq = kmem_cache_alloc_node(cfq_pool,
2875                                         gfp_mask | __GFP_ZERO,
2876                                         cfqd->queue->node);
2877                 }
2878
2879                 if (cfqq) {
2880                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2881                         cfq_init_prio_data(cfqq, ioc);
2882                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2883                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2884                 } else
2885                         cfqq = &cfqd->oom_cfqq;
2886         }
2887
2888         if (new_cfqq)
2889                 kmem_cache_free(cfq_pool, new_cfqq);
2890
2891         return cfqq;
2892 }
2893
2894 static struct cfq_queue **
2895 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2896 {
2897         switch (ioprio_class) {
2898         case IOPRIO_CLASS_RT:
2899                 return &cfqd->async_cfqq[0][ioprio];
2900         case IOPRIO_CLASS_BE:
2901                 return &cfqd->async_cfqq[1][ioprio];
2902         case IOPRIO_CLASS_IDLE:
2903                 return &cfqd->async_idle_cfqq;
2904         default:
2905                 BUG();
2906         }
2907 }
2908
2909 static struct cfq_queue *
2910 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2911               gfp_t gfp_mask)
2912 {
2913         const int ioprio = task_ioprio(ioc);
2914         const int ioprio_class = task_ioprio_class(ioc);
2915         struct cfq_queue **async_cfqq = NULL;
2916         struct cfq_queue *cfqq = NULL;
2917
2918         if (!is_sync) {
2919                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2920                 cfqq = *async_cfqq;
2921         }
2922
2923         if (!cfqq)
2924                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2925
2926         /*
2927          * pin the queue now that it's allocated, scheduler exit will prune it
2928          */
2929         if (!is_sync && !(*async_cfqq)) {
2930                 cfqq->ref++;
2931                 *async_cfqq = cfqq;
2932         }
2933
2934         cfqq->ref++;
2935         return cfqq;
2936 }
2937
2938 static void
2939 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
2940 {
2941         unsigned long elapsed = jiffies - ttime->last_end_request;
2942         elapsed = min(elapsed, 2UL * slice_idle);
2943
2944         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2945         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2946         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2947 }
2948
2949 static void
2950 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2951                         struct cfq_io_cq *cic)
2952 {
2953         if (cfq_cfqq_sync(cfqq)) {
2954                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
2955                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2956                         cfqd->cfq_slice_idle);
2957         }
2958 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2959         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2960 #endif
2961 }
2962
2963 static void
2964 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2965                        struct request *rq)
2966 {
2967         sector_t sdist = 0;
2968         sector_t n_sec = blk_rq_sectors(rq);
2969         if (cfqq->last_request_pos) {
2970                 if (cfqq->last_request_pos < blk_rq_pos(rq))
2971                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2972                 else
2973                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2974         }
2975
2976         cfqq->seek_history <<= 1;
2977         if (blk_queue_nonrot(cfqd->queue))
2978                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2979         else
2980                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2981 }
2982
2983 /*
2984  * Disable idle window if the process thinks too long or seeks so much that
2985  * it doesn't matter
2986  */
2987 static void
2988 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2989                        struct cfq_io_cq *cic)
2990 {
2991         int old_idle, enable_idle;
2992
2993         /*
2994          * Don't idle for async or idle io prio class
2995          */
2996         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
2997                 return;
2998
2999         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3000
3001         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3002                 cfq_mark_cfqq_deep(cfqq);
3003
3004         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3005                 enable_idle = 0;
3006         else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3007                  !cfqd->cfq_slice_idle ||
3008                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3009                 enable_idle = 0;
3010         else if (sample_valid(cic->ttime.ttime_samples)) {
3011                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3012                         enable_idle = 0;
3013                 else
3014                         enable_idle = 1;
3015         }
3016
3017         if (old_idle != enable_idle) {
3018                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3019                 if (enable_idle)
3020                         cfq_mark_cfqq_idle_window(cfqq);
3021                 else
3022                         cfq_clear_cfqq_idle_window(cfqq);
3023         }
3024 }
3025
3026 /*
3027  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3028  * no or if we aren't sure, a 1 will cause a preempt.
3029  */
3030 static bool
3031 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3032                    struct request *rq)
3033 {
3034         struct cfq_queue *cfqq;
3035
3036         cfqq = cfqd->active_queue;
3037         if (!cfqq)
3038                 return false;
3039
3040         if (cfq_class_idle(new_cfqq))
3041                 return false;
3042
3043         if (cfq_class_idle(cfqq))
3044                 return true;
3045
3046         /*
3047          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3048          */
3049         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3050                 return false;
3051
3052         /*
3053          * if the new request is sync, but the currently running queue is
3054          * not, let the sync request have priority.
3055          */
3056         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3057                 return true;
3058
3059         if (new_cfqq->cfqg != cfqq->cfqg)
3060                 return false;
3061
3062         if (cfq_slice_used(cfqq))
3063                 return true;
3064
3065         /* Allow preemption only if we are idling on sync-noidle tree */
3066         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3067             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3068             new_cfqq->service_tree->count == 2 &&
3069             RB_EMPTY_ROOT(&cfqq->sort_list))
3070                 return true;
3071
3072         /*
3073          * So both queues are sync. Let the new request get disk time if
3074          * it's a metadata request and the current queue is doing regular IO.
3075          */
3076         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3077                 return true;
3078
3079         /*
3080          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3081          */
3082         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3083                 return true;
3084
3085         /* An idle queue should not be idle now for some reason */
3086         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3087                 return true;
3088
3089         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3090                 return false;
3091
3092         /*
3093          * if this request is as-good as one we would expect from the
3094          * current cfqq, let it preempt
3095          */
3096         if (cfq_rq_close(cfqd, cfqq, rq))
3097                 return true;
3098
3099         return false;
3100 }
3101
3102 /*
3103  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3104  * let it have half of its nominal slice.
3105  */
3106 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3107 {
3108         struct cfq_queue *old_cfqq = cfqd->active_queue;
3109
3110         cfq_log_cfqq(cfqd, cfqq, "preempt");
3111         cfq_slice_expired(cfqd, 1);
3112
3113         /*
3114          * workload type is changed, don't save slice, otherwise preempt
3115          * doesn't happen
3116          */
3117         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3118                 cfqq->cfqg->saved_workload_slice = 0;
3119
3120         /*
3121          * Put the new queue at the front of the of the current list,
3122          * so we know that it will be selected next.
3123          */
3124         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3125
3126         cfq_service_tree_add(cfqd, cfqq, 1);
3127
3128         cfqq->slice_end = 0;
3129         cfq_mark_cfqq_slice_new(cfqq);
3130 }
3131
3132 /*
3133  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3134  * something we should do about it
3135  */
3136 static void
3137 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3138                 struct request *rq)
3139 {
3140         struct cfq_io_cq *cic = RQ_CIC(rq);
3141
3142         cfqd->rq_queued++;
3143         if (rq->cmd_flags & REQ_PRIO)
3144                 cfqq->prio_pending++;
3145
3146         cfq_update_io_thinktime(cfqd, cfqq, cic);
3147         cfq_update_io_seektime(cfqd, cfqq, rq);
3148         cfq_update_idle_window(cfqd, cfqq, cic);
3149
3150         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3151
3152         if (cfqq == cfqd->active_queue) {
3153                 /*
3154                  * Remember that we saw a request from this process, but
3155                  * don't start queuing just yet. Otherwise we risk seeing lots
3156                  * of tiny requests, because we disrupt the normal plugging
3157                  * and merging. If the request is already larger than a single
3158                  * page, let it rip immediately. For that case we assume that
3159                  * merging is already done. Ditto for a busy system that
3160                  * has other work pending, don't risk delaying until the
3161                  * idle timer unplug to continue working.
3162                  */
3163                 if (cfq_cfqq_wait_request(cfqq)) {
3164                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3165                             cfqd->busy_queues > 1) {
3166                                 cfq_del_timer(cfqd, cfqq);
3167                                 cfq_clear_cfqq_wait_request(cfqq);
3168                                 __blk_run_queue(cfqd->queue);
3169                         } else {
3170                                 cfq_blkiocg_update_idle_time_stats(
3171                                                 &cfqq->cfqg->blkg);
3172                                 cfq_mark_cfqq_must_dispatch(cfqq);
3173                         }
3174                 }
3175         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3176                 /*
3177                  * not the active queue - expire current slice if it is
3178                  * idle and has expired it's mean thinktime or this new queue
3179                  * has some old slice time left and is of higher priority or
3180                  * this new queue is RT and the current one is BE
3181                  */
3182                 cfq_preempt_queue(cfqd, cfqq);
3183                 __blk_run_queue(cfqd->queue);
3184         }
3185 }
3186
3187 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3188 {
3189         struct cfq_data *cfqd = q->elevator->elevator_data;
3190         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3191
3192         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3193         cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3194
3195         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3196         list_add_tail(&rq->queuelist, &cfqq->fifo);
3197         cfq_add_rq_rb(rq);
3198         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3199                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3200                         rq_is_sync(rq));
3201         cfq_rq_enqueued(cfqd, cfqq, rq);
3202 }
3203
3204 /*
3205  * Update hw_tag based on peak queue depth over 50 samples under
3206  * sufficient load.
3207  */
3208 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3209 {
3210         struct cfq_queue *cfqq = cfqd->active_queue;
3211
3212         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3213                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3214
3215         if (cfqd->hw_tag == 1)
3216                 return;
3217
3218         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3219             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3220                 return;
3221
3222         /*
3223          * If active queue hasn't enough requests and can idle, cfq might not
3224          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3225          * case
3226          */
3227         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3228             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3229             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3230                 return;
3231
3232         if (cfqd->hw_tag_samples++ < 50)
3233                 return;
3234
3235         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3236                 cfqd->hw_tag = 1;
3237         else
3238                 cfqd->hw_tag = 0;
3239 }
3240
3241 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3242 {
3243         struct cfq_io_cq *cic = cfqd->active_cic;
3244
3245         /* If the queue already has requests, don't wait */
3246         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3247                 return false;
3248
3249         /* If there are other queues in the group, don't wait */
3250         if (cfqq->cfqg->nr_cfqq > 1)
3251                 return false;
3252
3253         /* the only queue in the group, but think time is big */
3254         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3255                 return false;
3256
3257         if (cfq_slice_used(cfqq))
3258                 return true;
3259
3260         /* if slice left is less than think time, wait busy */
3261         if (cic && sample_valid(cic->ttime.ttime_samples)
3262             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3263                 return true;
3264
3265         /*
3266          * If think times is less than a jiffy than ttime_mean=0 and above
3267          * will not be true. It might happen that slice has not expired yet
3268          * but will expire soon (4-5 ns) during select_queue(). To cover the
3269          * case where think time is less than a jiffy, mark the queue wait
3270          * busy if only 1 jiffy is left in the slice.
3271          */
3272         if (cfqq->slice_end - jiffies == 1)
3273                 return true;
3274
3275         return false;
3276 }
3277
3278 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3279 {
3280         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3281         struct cfq_data *cfqd = cfqq->cfqd;
3282         const int sync = rq_is_sync(rq);
3283         unsigned long now;
3284
3285         now = jiffies;
3286         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3287                      !!(rq->cmd_flags & REQ_NOIDLE));
3288
3289         cfq_update_hw_tag(cfqd);
3290
3291         WARN_ON(!cfqd->rq_in_driver);
3292         WARN_ON(!cfqq->dispatched);
3293         cfqd->rq_in_driver--;
3294         cfqq->dispatched--;
3295         (RQ_CFQG(rq))->dispatched--;
3296         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3297                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3298                         rq_data_dir(rq), rq_is_sync(rq));
3299
3300         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3301
3302         if (sync) {
3303                 struct cfq_rb_root *service_tree;
3304
3305                 RQ_CIC(rq)->ttime.last_end_request = now;
3306
3307                 if (cfq_cfqq_on_rr(cfqq))
3308                         service_tree = cfqq->service_tree;
3309                 else
3310                         service_tree = service_tree_for(cfqq->cfqg,
3311                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3312                 service_tree->ttime.last_end_request = now;
3313                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3314                         cfqd->last_delayed_sync = now;
3315         }
3316
3317 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3318         cfqq->cfqg->ttime.last_end_request = now;
3319 #endif
3320
3321         /*
3322          * If this is the active queue, check if it needs to be expired,
3323          * or if we want to idle in case it has no pending requests.
3324          */
3325         if (cfqd->active_queue == cfqq) {
3326                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3327
3328                 if (cfq_cfqq_slice_new(cfqq)) {
3329                         cfq_set_prio_slice(cfqd, cfqq);
3330                         cfq_clear_cfqq_slice_new(cfqq);
3331                 }
3332
3333                 /*
3334                  * Should we wait for next request to come in before we expire
3335                  * the queue.
3336                  */
3337                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3338                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3339                         if (!cfqd->cfq_slice_idle)
3340                                 extend_sl = cfqd->cfq_group_idle;
3341                         cfqq->slice_end = jiffies + extend_sl;
3342                         cfq_mark_cfqq_wait_busy(cfqq);
3343                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3344                 }
3345
3346                 /*
3347                  * Idling is not enabled on:
3348                  * - expired queues
3349                  * - idle-priority queues
3350                  * - async queues
3351                  * - queues with still some requests queued
3352                  * - when there is a close cooperator
3353                  */
3354                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3355                         cfq_slice_expired(cfqd, 1);
3356                 else if (sync && cfqq_empty &&
3357                          !cfq_close_cooperator(cfqd, cfqq)) {
3358                         cfq_arm_slice_timer(cfqd);
3359                 }
3360         }
3361
3362         if (!cfqd->rq_in_driver)
3363                 cfq_schedule_dispatch(cfqd);
3364 }
3365
3366 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3367 {
3368         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3369                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3370                 return ELV_MQUEUE_MUST;
3371         }
3372
3373         return ELV_MQUEUE_MAY;
3374 }
3375
3376 static int cfq_may_queue(struct request_queue *q, int rw)
3377 {
3378         struct cfq_data *cfqd = q->elevator->elevator_data;
3379         struct task_struct *tsk = current;
3380         struct cfq_io_cq *cic;
3381         struct cfq_queue *cfqq;
3382
3383         /*
3384          * don't force setup of a queue from here, as a call to may_queue
3385          * does not necessarily imply that a request actually will be queued.
3386          * so just lookup a possibly existing queue, or return 'may queue'
3387          * if that fails
3388          */
3389         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3390         if (!cic)
3391                 return ELV_MQUEUE_MAY;
3392
3393         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3394         if (cfqq) {
3395                 cfq_init_prio_data(cfqq, cic->icq.ioc);
3396
3397                 return __cfq_may_queue(cfqq);
3398         }
3399
3400         return ELV_MQUEUE_MAY;
3401 }
3402
3403 /*
3404  * queue lock held here
3405  */
3406 static void cfq_put_request(struct request *rq)
3407 {
3408         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3409
3410         if (cfqq) {
3411                 const int rw = rq_data_dir(rq);
3412
3413                 BUG_ON(!cfqq->allocated[rw]);
3414                 cfqq->allocated[rw]--;
3415
3416                 /* Put down rq reference on cfqg */
3417                 cfq_put_cfqg(RQ_CFQG(rq));
3418                 rq->elv.priv[0] = NULL;
3419                 rq->elv.priv[1] = NULL;
3420
3421                 cfq_put_queue(cfqq);
3422         }
3423 }
3424
3425 static struct cfq_queue *
3426 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3427                 struct cfq_queue *cfqq)
3428 {
3429         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3430         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3431         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3432         cfq_put_queue(cfqq);
3433         return cic_to_cfqq(cic, 1);
3434 }
3435
3436 /*
3437  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3438  * was the last process referring to said cfqq.
3439  */
3440 static struct cfq_queue *
3441 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3442 {
3443         if (cfqq_process_refs(cfqq) == 1) {
3444                 cfqq->pid = current->pid;
3445                 cfq_clear_cfqq_coop(cfqq);
3446                 cfq_clear_cfqq_split_coop(cfqq);
3447                 return cfqq;
3448         }
3449
3450         cic_set_cfqq(cic, NULL, 1);
3451
3452         cfq_put_cooperator(cfqq);
3453
3454         cfq_put_queue(cfqq);
3455         return NULL;
3456 }
3457 /*
3458  * Allocate cfq data structures associated with this request.
3459  */
3460 static int
3461 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3462 {
3463         struct cfq_data *cfqd = q->elevator->elevator_data;
3464         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3465         const int rw = rq_data_dir(rq);
3466         const bool is_sync = rq_is_sync(rq);
3467         struct cfq_queue *cfqq;
3468
3469         might_sleep_if(gfp_mask & __GFP_WAIT);
3470
3471         spin_lock_irq(q->queue_lock);
3472
3473         /* handle changed notifications */
3474         if (unlikely(cic->icq.changed)) {
3475                 if (test_and_clear_bit(ICQ_IOPRIO_CHANGED, &cic->icq.changed))
3476                         changed_ioprio(cic);
3477 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3478                 if (test_and_clear_bit(ICQ_CGROUP_CHANGED, &cic->icq.changed))
3479                         changed_cgroup(cic);
3480 #endif
3481         }
3482
3483 new_queue:
3484         cfqq = cic_to_cfqq(cic, is_sync);
3485         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3486                 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3487                 cic_set_cfqq(cic, cfqq, is_sync);
3488         } else {
3489                 /*
3490                  * If the queue was seeky for too long, break it apart.
3491                  */
3492                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3493                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3494                         cfqq = split_cfqq(cic, cfqq);
3495                         if (!cfqq)
3496                                 goto new_queue;
3497                 }
3498
3499                 /*
3500                  * Check to see if this queue is scheduled to merge with
3501                  * another, closely cooperating queue.  The merging of
3502                  * queues happens here as it must be done in process context.
3503                  * The reference on new_cfqq was taken in merge_cfqqs.
3504                  */
3505                 if (cfqq->new_cfqq)
3506                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3507         }
3508
3509         cfqq->allocated[rw]++;
3510
3511         cfqq->ref++;
3512         rq->elv.priv[0] = cfqq;
3513         rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
3514         spin_unlock_irq(q->queue_lock);
3515         return 0;
3516 }
3517
3518 static void cfq_kick_queue(struct work_struct *work)
3519 {
3520         struct cfq_data *cfqd =
3521                 container_of(work, struct cfq_data, unplug_work);
3522         struct request_queue *q = cfqd->queue;
3523
3524         spin_lock_irq(q->queue_lock);
3525         __blk_run_queue(cfqd->queue);
3526         spin_unlock_irq(q->queue_lock);
3527 }
3528
3529 /*
3530  * Timer running if the active_queue is currently idling inside its time slice
3531  */
3532 static void cfq_idle_slice_timer(unsigned long data)
3533 {
3534         struct cfq_data *cfqd = (struct cfq_data *) data;
3535         struct cfq_queue *cfqq;
3536         unsigned long flags;
3537         int timed_out = 1;
3538
3539         cfq_log(cfqd, "idle timer fired");
3540
3541         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3542
3543         cfqq = cfqd->active_queue;
3544         if (cfqq) {
3545                 timed_out = 0;
3546
3547                 /*
3548                  * We saw a request before the queue expired, let it through
3549                  */
3550                 if (cfq_cfqq_must_dispatch(cfqq))
3551                         goto out_kick;
3552
3553                 /*
3554                  * expired
3555                  */
3556                 if (cfq_slice_used(cfqq))
3557                         goto expire;
3558
3559                 /*
3560                  * only expire and reinvoke request handler, if there are
3561                  * other queues with pending requests
3562                  */
3563                 if (!cfqd->busy_queues)
3564                         goto out_cont;
3565
3566                 /*
3567                  * not expired and it has a request pending, let it dispatch
3568                  */
3569                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3570                         goto out_kick;
3571
3572                 /*
3573                  * Queue depth flag is reset only when the idle didn't succeed
3574                  */
3575                 cfq_clear_cfqq_deep(cfqq);
3576         }
3577 expire:
3578         cfq_slice_expired(cfqd, timed_out);
3579 out_kick:
3580         cfq_schedule_dispatch(cfqd);
3581 out_cont:
3582         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3583 }
3584
3585 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3586 {
3587         del_timer_sync(&cfqd->idle_slice_timer);
3588         cancel_work_sync(&cfqd->unplug_work);
3589 }
3590
3591 static void cfq_put_async_queues(struct cfq_data *cfqd)
3592 {
3593         int i;
3594
3595         for (i = 0; i < IOPRIO_BE_NR; i++) {
3596                 if (cfqd->async_cfqq[0][i])
3597                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3598                 if (cfqd->async_cfqq[1][i])
3599                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3600         }
3601
3602         if (cfqd->async_idle_cfqq)
3603                 cfq_put_queue(cfqd->async_idle_cfqq);
3604 }
3605
3606 static void cfq_exit_queue(struct elevator_queue *e)
3607 {
3608         struct cfq_data *cfqd = e->elevator_data;
3609         struct request_queue *q = cfqd->queue;
3610         bool wait = false;
3611
3612         cfq_shutdown_timer_wq(cfqd);
3613
3614         spin_lock_irq(q->queue_lock);
3615
3616         if (cfqd->active_queue)
3617                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3618
3619         cfq_put_async_queues(cfqd);
3620         cfq_release_cfq_groups(cfqd);
3621
3622         /*
3623          * If there are groups which we could not unlink from blkcg list,
3624          * wait for a rcu period for them to be freed.
3625          */
3626         if (cfqd->nr_blkcg_linked_grps)
3627                 wait = true;
3628
3629         spin_unlock_irq(q->queue_lock);
3630
3631         cfq_shutdown_timer_wq(cfqd);
3632
3633         /*
3634          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3635          * Do this wait only if there are other unlinked groups out
3636          * there. This can happen if cgroup deletion path claimed the
3637          * responsibility of cleaning up a group before queue cleanup code
3638          * get to the group.
3639          *
3640          * Do not call synchronize_rcu() unconditionally as there are drivers
3641          * which create/delete request queue hundreds of times during scan/boot
3642          * and synchronize_rcu() can take significant time and slow down boot.
3643          */
3644         if (wait)
3645                 synchronize_rcu();
3646
3647 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3648         /* Free up per cpu stats for root group */
3649         free_percpu(cfqd->root_group.blkg.stats_cpu);
3650 #endif
3651         kfree(cfqd);
3652 }
3653
3654 static void *cfq_init_queue(struct request_queue *q)
3655 {
3656         struct cfq_data *cfqd;
3657         int i, j;
3658         struct cfq_group *cfqg;
3659         struct cfq_rb_root *st;
3660
3661         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3662         if (!cfqd)
3663                 return NULL;
3664
3665         /* Init root service tree */
3666         cfqd->grp_service_tree = CFQ_RB_ROOT;
3667
3668         /* Init root group */
3669         cfqg = &cfqd->root_group;
3670         for_each_cfqg_st(cfqg, i, j, st)
3671                 *st = CFQ_RB_ROOT;
3672         RB_CLEAR_NODE(&cfqg->rb_node);
3673
3674         /* Give preference to root group over other groups */
3675         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3676
3677 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3678         /*
3679          * Set root group reference to 2. One reference will be dropped when
3680          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3681          * Other reference will remain there as we don't want to delete this
3682          * group as it is statically allocated and gets destroyed when
3683          * throtl_data goes away.
3684          */
3685         cfqg->ref = 2;
3686
3687         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3688                 kfree(cfqg);
3689                 kfree(cfqd);
3690                 return NULL;
3691         }
3692
3693         rcu_read_lock();
3694
3695         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3696                                         (void *)cfqd, 0);
3697         rcu_read_unlock();
3698         cfqd->nr_blkcg_linked_grps++;
3699
3700         /* Add group on cfqd->cfqg_list */
3701         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3702 #endif
3703         /*
3704          * Not strictly needed (since RB_ROOT just clears the node and we
3705          * zeroed cfqd on alloc), but better be safe in case someone decides
3706          * to add magic to the rb code
3707          */
3708         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3709                 cfqd->prio_trees[i] = RB_ROOT;
3710
3711         /*
3712          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3713          * Grab a permanent reference to it, so that the normal code flow
3714          * will not attempt to free it.
3715          */
3716         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3717         cfqd->oom_cfqq.ref++;
3718         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3719
3720         cfqd->queue = q;
3721
3722         init_timer(&cfqd->idle_slice_timer);
3723         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3724         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3725
3726         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3727
3728         cfqd->cfq_quantum = cfq_quantum;
3729         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3730         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3731         cfqd->cfq_back_max = cfq_back_max;
3732         cfqd->cfq_back_penalty = cfq_back_penalty;
3733         cfqd->cfq_slice[0] = cfq_slice_async;
3734         cfqd->cfq_slice[1] = cfq_slice_sync;
3735         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3736         cfqd->cfq_slice_idle = cfq_slice_idle;
3737         cfqd->cfq_group_idle = cfq_group_idle;
3738         cfqd->cfq_latency = 1;
3739         cfqd->hw_tag = -1;
3740         /*
3741          * we optimistically start assuming sync ops weren't delayed in last
3742          * second, in order to have larger depth for async operations.
3743          */
3744         cfqd->last_delayed_sync = jiffies - HZ;
3745         return cfqd;
3746 }
3747
3748 /*
3749  * sysfs parts below -->
3750  */
3751 static ssize_t
3752 cfq_var_show(unsigned int var, char *page)
3753 {
3754         return sprintf(page, "%d\n", var);
3755 }
3756
3757 static ssize_t
3758 cfq_var_store(unsigned int *var, const char *page, size_t count)
3759 {
3760         char *p = (char *) page;
3761
3762         *var = simple_strtoul(p, &p, 10);
3763         return count;
3764 }
3765
3766 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3767 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3768 {                                                                       \
3769         struct cfq_data *cfqd = e->elevator_data;                       \
3770         unsigned int __data = __VAR;                                    \
3771         if (__CONV)                                                     \
3772                 __data = jiffies_to_msecs(__data);                      \
3773         return cfq_var_show(__data, (page));                            \
3774 }
3775 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3776 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3777 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3778 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3779 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3780 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3781 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3782 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3783 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3784 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3785 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3786 #undef SHOW_FUNCTION
3787
3788 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3789 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3790 {                                                                       \
3791         struct cfq_data *cfqd = e->elevator_data;                       \
3792         unsigned int __data;                                            \
3793         int ret = cfq_var_store(&__data, (page), count);                \
3794         if (__data < (MIN))                                             \
3795                 __data = (MIN);                                         \
3796         else if (__data > (MAX))                                        \
3797                 __data = (MAX);                                         \
3798         if (__CONV)                                                     \
3799                 *(__PTR) = msecs_to_jiffies(__data);                    \
3800         else                                                            \
3801                 *(__PTR) = __data;                                      \
3802         return ret;                                                     \
3803 }
3804 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3805 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3806                 UINT_MAX, 1);
3807 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3808                 UINT_MAX, 1);
3809 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3810 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3811                 UINT_MAX, 0);
3812 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3813 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
3814 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3815 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3816 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3817                 UINT_MAX, 0);
3818 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3819 #undef STORE_FUNCTION
3820
3821 #define CFQ_ATTR(name) \
3822         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3823
3824 static struct elv_fs_entry cfq_attrs[] = {
3825         CFQ_ATTR(quantum),
3826         CFQ_ATTR(fifo_expire_sync),
3827         CFQ_ATTR(fifo_expire_async),
3828         CFQ_ATTR(back_seek_max),
3829         CFQ_ATTR(back_seek_penalty),
3830         CFQ_ATTR(slice_sync),
3831         CFQ_ATTR(slice_async),
3832         CFQ_ATTR(slice_async_rq),
3833         CFQ_ATTR(slice_idle),
3834         CFQ_ATTR(group_idle),
3835         CFQ_ATTR(low_latency),
3836         __ATTR_NULL
3837 };
3838
3839 static struct elevator_type iosched_cfq = {
3840         .ops = {
3841                 .elevator_merge_fn =            cfq_merge,
3842                 .elevator_merged_fn =           cfq_merged_request,
3843                 .elevator_merge_req_fn =        cfq_merged_requests,
3844                 .elevator_allow_merge_fn =      cfq_allow_merge,
3845                 .elevator_bio_merged_fn =       cfq_bio_merged,
3846                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3847                 .elevator_add_req_fn =          cfq_insert_request,
3848                 .elevator_activate_req_fn =     cfq_activate_request,
3849                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3850                 .elevator_completed_req_fn =    cfq_completed_request,
3851                 .elevator_former_req_fn =       elv_rb_former_request,
3852                 .elevator_latter_req_fn =       elv_rb_latter_request,
3853                 .elevator_init_icq_fn =         cfq_init_icq,
3854                 .elevator_exit_icq_fn =         cfq_exit_icq,
3855                 .elevator_set_req_fn =          cfq_set_request,
3856                 .elevator_put_req_fn =          cfq_put_request,
3857                 .elevator_may_queue_fn =        cfq_may_queue,
3858                 .elevator_init_fn =             cfq_init_queue,
3859                 .elevator_exit_fn =             cfq_exit_queue,
3860         },
3861         .icq_size       =       sizeof(struct cfq_io_cq),
3862         .icq_align      =       __alignof__(struct cfq_io_cq),
3863         .elevator_attrs =       cfq_attrs,
3864         .elevator_name  =       "cfq",
3865         .elevator_owner =       THIS_MODULE,
3866 };
3867
3868 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3869 static struct blkio_policy_type blkio_policy_cfq = {
3870         .ops = {
3871                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3872                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3873         },
3874         .plid = BLKIO_POLICY_PROP,
3875 };
3876 #else
3877 static struct blkio_policy_type blkio_policy_cfq;
3878 #endif
3879
3880 static int __init cfq_init(void)
3881 {
3882         int ret;
3883
3884         /*
3885          * could be 0 on HZ < 1000 setups
3886          */
3887         if (!cfq_slice_async)
3888                 cfq_slice_async = 1;
3889         if (!cfq_slice_idle)
3890                 cfq_slice_idle = 1;
3891
3892 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3893         if (!cfq_group_idle)
3894                 cfq_group_idle = 1;
3895 #else
3896                 cfq_group_idle = 0;
3897 #endif
3898         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3899         if (!cfq_pool)
3900                 return -ENOMEM;
3901
3902         ret = elv_register(&iosched_cfq);
3903         if (ret) {
3904                 kmem_cache_destroy(cfq_pool);
3905                 return ret;
3906         }
3907
3908         blkio_policy_register(&blkio_policy_cfq);
3909
3910         return 0;
3911 }
3912
3913 static void __exit cfq_exit(void)
3914 {
3915         blkio_policy_unregister(&blkio_policy_cfq);
3916         elv_unregister(&iosched_cfq);
3917         kmem_cache_destroy(cfq_pool);
3918 }
3919
3920 module_init(cfq_init);
3921 module_exit(cfq_exit);
3922
3923 MODULE_AUTHOR("Jens Axboe");
3924 MODULE_LICENSE("GPL");
3925 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");