block: strip out locking optimization in put_io_context()
[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         struct cfq_data *cfqd = q->elevator->elevator_data;
1660
1661         /*
1662          * reposition in fifo if next is older than rq
1663          */
1664         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1665             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1666                 list_move(&rq->queuelist, &next->queuelist);
1667                 rq_set_fifo_time(rq, rq_fifo_time(next));
1668         }
1669
1670         if (cfqq->next_rq == next)
1671                 cfqq->next_rq = rq;
1672         cfq_remove_request(next);
1673         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1674                                         rq_data_dir(next), rq_is_sync(next));
1675
1676         cfqq = RQ_CFQQ(next);
1677         /*
1678          * all requests of this queue are merged to other queues, delete it
1679          * from the service tree. If it's the active_queue,
1680          * cfq_dispatch_requests() will choose to expire it or do idle
1681          */
1682         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list) &&
1683             cfqq != cfqd->active_queue)
1684                 cfq_del_cfqq_rr(cfqd, cfqq);
1685 }
1686
1687 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1688                            struct bio *bio)
1689 {
1690         struct cfq_data *cfqd = q->elevator->elevator_data;
1691         struct cfq_io_cq *cic;
1692         struct cfq_queue *cfqq;
1693
1694         /*
1695          * Disallow merge of a sync bio into an async request.
1696          */
1697         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1698                 return false;
1699
1700         /*
1701          * Lookup the cfqq that this bio will be queued with and allow
1702          * merge only if rq is queued there.  This function can be called
1703          * from plug merge without queue_lock.  In such cases, ioc of @rq
1704          * and %current are guaranteed to be equal.  Avoid lookup which
1705          * requires queue_lock by using @rq's cic.
1706          */
1707         if (current->io_context == RQ_CIC(rq)->icq.ioc) {
1708                 cic = RQ_CIC(rq);
1709         } else {
1710                 cic = cfq_cic_lookup(cfqd, current->io_context);
1711                 if (!cic)
1712                         return false;
1713         }
1714
1715         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1716         return cfqq == RQ_CFQQ(rq);
1717 }
1718
1719 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1720 {
1721         del_timer(&cfqd->idle_slice_timer);
1722         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1723 }
1724
1725 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1726                                    struct cfq_queue *cfqq)
1727 {
1728         if (cfqq) {
1729                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1730                                 cfqd->serving_prio, cfqd->serving_type);
1731                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1732                 cfqq->slice_start = 0;
1733                 cfqq->dispatch_start = jiffies;
1734                 cfqq->allocated_slice = 0;
1735                 cfqq->slice_end = 0;
1736                 cfqq->slice_dispatch = 0;
1737                 cfqq->nr_sectors = 0;
1738
1739                 cfq_clear_cfqq_wait_request(cfqq);
1740                 cfq_clear_cfqq_must_dispatch(cfqq);
1741                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1742                 cfq_clear_cfqq_fifo_expire(cfqq);
1743                 cfq_mark_cfqq_slice_new(cfqq);
1744
1745                 cfq_del_timer(cfqd, cfqq);
1746         }
1747
1748         cfqd->active_queue = cfqq;
1749 }
1750
1751 /*
1752  * current cfqq expired its slice (or was too idle), select new one
1753  */
1754 static void
1755 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1756                     bool timed_out)
1757 {
1758         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1759
1760         if (cfq_cfqq_wait_request(cfqq))
1761                 cfq_del_timer(cfqd, cfqq);
1762
1763         cfq_clear_cfqq_wait_request(cfqq);
1764         cfq_clear_cfqq_wait_busy(cfqq);
1765
1766         /*
1767          * If this cfqq is shared between multiple processes, check to
1768          * make sure that those processes are still issuing I/Os within
1769          * the mean seek distance.  If not, it may be time to break the
1770          * queues apart again.
1771          */
1772         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1773                 cfq_mark_cfqq_split_coop(cfqq);
1774
1775         /*
1776          * store what was left of this slice, if the queue idled/timed out
1777          */
1778         if (timed_out) {
1779                 if (cfq_cfqq_slice_new(cfqq))
1780                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1781                 else
1782                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1783                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1784         }
1785
1786         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1787
1788         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1789                 cfq_del_cfqq_rr(cfqd, cfqq);
1790
1791         cfq_resort_rr_list(cfqd, cfqq);
1792
1793         if (cfqq == cfqd->active_queue)
1794                 cfqd->active_queue = NULL;
1795
1796         if (cfqd->active_cic) {
1797                 put_io_context(cfqd->active_cic->icq.ioc);
1798                 cfqd->active_cic = NULL;
1799         }
1800 }
1801
1802 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1803 {
1804         struct cfq_queue *cfqq = cfqd->active_queue;
1805
1806         if (cfqq)
1807                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1808 }
1809
1810 /*
1811  * Get next queue for service. Unless we have a queue preemption,
1812  * we'll simply select the first cfqq in the service tree.
1813  */
1814 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1815 {
1816         struct cfq_rb_root *service_tree =
1817                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1818                                         cfqd->serving_type);
1819
1820         if (!cfqd->rq_queued)
1821                 return NULL;
1822
1823         /* There is nothing to dispatch */
1824         if (!service_tree)
1825                 return NULL;
1826         if (RB_EMPTY_ROOT(&service_tree->rb))
1827                 return NULL;
1828         return cfq_rb_first(service_tree);
1829 }
1830
1831 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1832 {
1833         struct cfq_group *cfqg;
1834         struct cfq_queue *cfqq;
1835         int i, j;
1836         struct cfq_rb_root *st;
1837
1838         if (!cfqd->rq_queued)
1839                 return NULL;
1840
1841         cfqg = cfq_get_next_cfqg(cfqd);
1842         if (!cfqg)
1843                 return NULL;
1844
1845         for_each_cfqg_st(cfqg, i, j, st)
1846                 if ((cfqq = cfq_rb_first(st)) != NULL)
1847                         return cfqq;
1848         return NULL;
1849 }
1850
1851 /*
1852  * Get and set a new active queue for service.
1853  */
1854 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1855                                               struct cfq_queue *cfqq)
1856 {
1857         if (!cfqq)
1858                 cfqq = cfq_get_next_queue(cfqd);
1859
1860         __cfq_set_active_queue(cfqd, cfqq);
1861         return cfqq;
1862 }
1863
1864 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1865                                           struct request *rq)
1866 {
1867         if (blk_rq_pos(rq) >= cfqd->last_position)
1868                 return blk_rq_pos(rq) - cfqd->last_position;
1869         else
1870                 return cfqd->last_position - blk_rq_pos(rq);
1871 }
1872
1873 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1874                                struct request *rq)
1875 {
1876         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1877 }
1878
1879 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1880                                     struct cfq_queue *cur_cfqq)
1881 {
1882         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1883         struct rb_node *parent, *node;
1884         struct cfq_queue *__cfqq;
1885         sector_t sector = cfqd->last_position;
1886
1887         if (RB_EMPTY_ROOT(root))
1888                 return NULL;
1889
1890         /*
1891          * First, if we find a request starting at the end of the last
1892          * request, choose it.
1893          */
1894         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1895         if (__cfqq)
1896                 return __cfqq;
1897
1898         /*
1899          * If the exact sector wasn't found, the parent of the NULL leaf
1900          * will contain the closest sector.
1901          */
1902         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1903         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1904                 return __cfqq;
1905
1906         if (blk_rq_pos(__cfqq->next_rq) < sector)
1907                 node = rb_next(&__cfqq->p_node);
1908         else
1909                 node = rb_prev(&__cfqq->p_node);
1910         if (!node)
1911                 return NULL;
1912
1913         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1914         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1915                 return __cfqq;
1916
1917         return NULL;
1918 }
1919
1920 /*
1921  * cfqd - obvious
1922  * cur_cfqq - passed in so that we don't decide that the current queue is
1923  *            closely cooperating with itself.
1924  *
1925  * So, basically we're assuming that that cur_cfqq has dispatched at least
1926  * one request, and that cfqd->last_position reflects a position on the disk
1927  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1928  * assumption.
1929  */
1930 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1931                                               struct cfq_queue *cur_cfqq)
1932 {
1933         struct cfq_queue *cfqq;
1934
1935         if (cfq_class_idle(cur_cfqq))
1936                 return NULL;
1937         if (!cfq_cfqq_sync(cur_cfqq))
1938                 return NULL;
1939         if (CFQQ_SEEKY(cur_cfqq))
1940                 return NULL;
1941
1942         /*
1943          * Don't search priority tree if it's the only queue in the group.
1944          */
1945         if (cur_cfqq->cfqg->nr_cfqq == 1)
1946                 return NULL;
1947
1948         /*
1949          * We should notice if some of the queues are cooperating, eg
1950          * working closely on the same area of the disk. In that case,
1951          * we can group them together and don't waste time idling.
1952          */
1953         cfqq = cfqq_close(cfqd, cur_cfqq);
1954         if (!cfqq)
1955                 return NULL;
1956
1957         /* If new queue belongs to different cfq_group, don't choose it */
1958         if (cur_cfqq->cfqg != cfqq->cfqg)
1959                 return NULL;
1960
1961         /*
1962          * It only makes sense to merge sync queues.
1963          */
1964         if (!cfq_cfqq_sync(cfqq))
1965                 return NULL;
1966         if (CFQQ_SEEKY(cfqq))
1967                 return NULL;
1968
1969         /*
1970          * Do not merge queues of different priority classes
1971          */
1972         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1973                 return NULL;
1974
1975         return cfqq;
1976 }
1977
1978 /*
1979  * Determine whether we should enforce idle window for this queue.
1980  */
1981
1982 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1983 {
1984         enum wl_prio_t prio = cfqq_prio(cfqq);
1985         struct cfq_rb_root *service_tree = cfqq->service_tree;
1986
1987         BUG_ON(!service_tree);
1988         BUG_ON(!service_tree->count);
1989
1990         if (!cfqd->cfq_slice_idle)
1991                 return false;
1992
1993         /* We never do for idle class queues. */
1994         if (prio == IDLE_WORKLOAD)
1995                 return false;
1996
1997         /* We do for queues that were marked with idle window flag. */
1998         if (cfq_cfqq_idle_window(cfqq) &&
1999            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
2000                 return true;
2001
2002         /*
2003          * Otherwise, we do only if they are the last ones
2004          * in their service tree.
2005          */
2006         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
2007            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
2008                 return true;
2009         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
2010                         service_tree->count);
2011         return false;
2012 }
2013
2014 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
2015 {
2016         struct cfq_queue *cfqq = cfqd->active_queue;
2017         struct cfq_io_cq *cic;
2018         unsigned long sl, group_idle = 0;
2019
2020         /*
2021          * SSD device without seek penalty, disable idling. But only do so
2022          * for devices that support queuing, otherwise we still have a problem
2023          * with sync vs async workloads.
2024          */
2025         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2026                 return;
2027
2028         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2029         WARN_ON(cfq_cfqq_slice_new(cfqq));
2030
2031         /*
2032          * idle is disabled, either manually or by past process history
2033          */
2034         if (!cfq_should_idle(cfqd, cfqq)) {
2035                 /* no queue idling. Check for group idling */
2036                 if (cfqd->cfq_group_idle)
2037                         group_idle = cfqd->cfq_group_idle;
2038                 else
2039                         return;
2040         }
2041
2042         /*
2043          * still active requests from this queue, don't idle
2044          */
2045         if (cfqq->dispatched)
2046                 return;
2047
2048         /*
2049          * task has exited, don't wait
2050          */
2051         cic = cfqd->active_cic;
2052         if (!cic || !atomic_read(&cic->icq.ioc->nr_tasks))
2053                 return;
2054
2055         /*
2056          * If our average think time is larger than the remaining time
2057          * slice, then don't idle. This avoids overrunning the allotted
2058          * time slice.
2059          */
2060         if (sample_valid(cic->ttime.ttime_samples) &&
2061             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2062                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2063                              cic->ttime.ttime_mean);
2064                 return;
2065         }
2066
2067         /* There are other queues in the group, don't do group idle */
2068         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2069                 return;
2070
2071         cfq_mark_cfqq_wait_request(cfqq);
2072
2073         if (group_idle)
2074                 sl = cfqd->cfq_group_idle;
2075         else
2076                 sl = cfqd->cfq_slice_idle;
2077
2078         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2079         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2080         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2081                         group_idle ? 1 : 0);
2082 }
2083
2084 /*
2085  * Move request from internal lists to the request queue dispatch list.
2086  */
2087 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2088 {
2089         struct cfq_data *cfqd = q->elevator->elevator_data;
2090         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2091
2092         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2093
2094         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2095         cfq_remove_request(rq);
2096         cfqq->dispatched++;
2097         (RQ_CFQG(rq))->dispatched++;
2098         elv_dispatch_sort(q, rq);
2099
2100         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2101         cfqq->nr_sectors += blk_rq_sectors(rq);
2102         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2103                                         rq_data_dir(rq), rq_is_sync(rq));
2104 }
2105
2106 /*
2107  * return expired entry, or NULL to just start from scratch in rbtree
2108  */
2109 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2110 {
2111         struct request *rq = NULL;
2112
2113         if (cfq_cfqq_fifo_expire(cfqq))
2114                 return NULL;
2115
2116         cfq_mark_cfqq_fifo_expire(cfqq);
2117
2118         if (list_empty(&cfqq->fifo))
2119                 return NULL;
2120
2121         rq = rq_entry_fifo(cfqq->fifo.next);
2122         if (time_before(jiffies, rq_fifo_time(rq)))
2123                 rq = NULL;
2124
2125         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2126         return rq;
2127 }
2128
2129 static inline int
2130 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2131 {
2132         const int base_rq = cfqd->cfq_slice_async_rq;
2133
2134         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2135
2136         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2137 }
2138
2139 /*
2140  * Must be called with the queue_lock held.
2141  */
2142 static int cfqq_process_refs(struct cfq_queue *cfqq)
2143 {
2144         int process_refs, io_refs;
2145
2146         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2147         process_refs = cfqq->ref - io_refs;
2148         BUG_ON(process_refs < 0);
2149         return process_refs;
2150 }
2151
2152 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2153 {
2154         int process_refs, new_process_refs;
2155         struct cfq_queue *__cfqq;
2156
2157         /*
2158          * If there are no process references on the new_cfqq, then it is
2159          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2160          * chain may have dropped their last reference (not just their
2161          * last process reference).
2162          */
2163         if (!cfqq_process_refs(new_cfqq))
2164                 return;
2165
2166         /* Avoid a circular list and skip interim queue merges */
2167         while ((__cfqq = new_cfqq->new_cfqq)) {
2168                 if (__cfqq == cfqq)
2169                         return;
2170                 new_cfqq = __cfqq;
2171         }
2172
2173         process_refs = cfqq_process_refs(cfqq);
2174         new_process_refs = cfqq_process_refs(new_cfqq);
2175         /*
2176          * If the process for the cfqq has gone away, there is no
2177          * sense in merging the queues.
2178          */
2179         if (process_refs == 0 || new_process_refs == 0)
2180                 return;
2181
2182         /*
2183          * Merge in the direction of the lesser amount of work.
2184          */
2185         if (new_process_refs >= process_refs) {
2186                 cfqq->new_cfqq = new_cfqq;
2187                 new_cfqq->ref += process_refs;
2188         } else {
2189                 new_cfqq->new_cfqq = cfqq;
2190                 cfqq->ref += new_process_refs;
2191         }
2192 }
2193
2194 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2195                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2196 {
2197         struct cfq_queue *queue;
2198         int i;
2199         bool key_valid = false;
2200         unsigned long lowest_key = 0;
2201         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2202
2203         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2204                 /* select the one with lowest rb_key */
2205                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2206                 if (queue &&
2207                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2208                         lowest_key = queue->rb_key;
2209                         cur_best = i;
2210                         key_valid = true;
2211                 }
2212         }
2213
2214         return cur_best;
2215 }
2216
2217 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2218 {
2219         unsigned slice;
2220         unsigned count;
2221         struct cfq_rb_root *st;
2222         unsigned group_slice;
2223         enum wl_prio_t original_prio = cfqd->serving_prio;
2224
2225         /* Choose next priority. RT > BE > IDLE */
2226         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2227                 cfqd->serving_prio = RT_WORKLOAD;
2228         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2229                 cfqd->serving_prio = BE_WORKLOAD;
2230         else {
2231                 cfqd->serving_prio = IDLE_WORKLOAD;
2232                 cfqd->workload_expires = jiffies + 1;
2233                 return;
2234         }
2235
2236         if (original_prio != cfqd->serving_prio)
2237                 goto new_workload;
2238
2239         /*
2240          * For RT and BE, we have to choose also the type
2241          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2242          * expiration time
2243          */
2244         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2245         count = st->count;
2246
2247         /*
2248          * check workload expiration, and that we still have other queues ready
2249          */
2250         if (count && !time_after(jiffies, cfqd->workload_expires))
2251                 return;
2252
2253 new_workload:
2254         /* otherwise select new workload type */
2255         cfqd->serving_type =
2256                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2257         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2258         count = st->count;
2259
2260         /*
2261          * the workload slice is computed as a fraction of target latency
2262          * proportional to the number of queues in that workload, over
2263          * all the queues in the same priority class
2264          */
2265         group_slice = cfq_group_slice(cfqd, cfqg);
2266
2267         slice = group_slice * count /
2268                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2269                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2270
2271         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2272                 unsigned int tmp;
2273
2274                 /*
2275                  * Async queues are currently system wide. Just taking
2276                  * proportion of queues with-in same group will lead to higher
2277                  * async ratio system wide as generally root group is going
2278                  * to have higher weight. A more accurate thing would be to
2279                  * calculate system wide asnc/sync ratio.
2280                  */
2281                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2282                 tmp = tmp/cfqd->busy_queues;
2283                 slice = min_t(unsigned, slice, tmp);
2284
2285                 /* async workload slice is scaled down according to
2286                  * the sync/async slice ratio. */
2287                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2288         } else
2289                 /* sync workload slice is at least 2 * cfq_slice_idle */
2290                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2291
2292         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2293         cfq_log(cfqd, "workload slice:%d", slice);
2294         cfqd->workload_expires = jiffies + slice;
2295 }
2296
2297 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2298 {
2299         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2300         struct cfq_group *cfqg;
2301
2302         if (RB_EMPTY_ROOT(&st->rb))
2303                 return NULL;
2304         cfqg = cfq_rb_first_group(st);
2305         update_min_vdisktime(st);
2306         return cfqg;
2307 }
2308
2309 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2310 {
2311         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2312
2313         cfqd->serving_group = cfqg;
2314
2315         /* Restore the workload type data */
2316         if (cfqg->saved_workload_slice) {
2317                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2318                 cfqd->serving_type = cfqg->saved_workload;
2319                 cfqd->serving_prio = cfqg->saved_serving_prio;
2320         } else
2321                 cfqd->workload_expires = jiffies - 1;
2322
2323         choose_service_tree(cfqd, cfqg);
2324 }
2325
2326 /*
2327  * Select a queue for service. If we have a current active queue,
2328  * check whether to continue servicing it, or retrieve and set a new one.
2329  */
2330 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2331 {
2332         struct cfq_queue *cfqq, *new_cfqq = NULL;
2333
2334         cfqq = cfqd->active_queue;
2335         if (!cfqq)
2336                 goto new_queue;
2337
2338         if (!cfqd->rq_queued)
2339                 return NULL;
2340
2341         /*
2342          * We were waiting for group to get backlogged. Expire the queue
2343          */
2344         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2345                 goto expire;
2346
2347         /*
2348          * The active queue has run out of time, expire it and select new.
2349          */
2350         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2351                 /*
2352                  * If slice had not expired at the completion of last request
2353                  * we might not have turned on wait_busy flag. Don't expire
2354                  * the queue yet. Allow the group to get backlogged.
2355                  *
2356                  * The very fact that we have used the slice, that means we
2357                  * have been idling all along on this queue and it should be
2358                  * ok to wait for this request to complete.
2359                  */
2360                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2361                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2362                         cfqq = NULL;
2363                         goto keep_queue;
2364                 } else
2365                         goto check_group_idle;
2366         }
2367
2368         /*
2369          * The active queue has requests and isn't expired, allow it to
2370          * dispatch.
2371          */
2372         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2373                 goto keep_queue;
2374
2375         /*
2376          * If another queue has a request waiting within our mean seek
2377          * distance, let it run.  The expire code will check for close
2378          * cooperators and put the close queue at the front of the service
2379          * tree.  If possible, merge the expiring queue with the new cfqq.
2380          */
2381         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2382         if (new_cfqq) {
2383                 if (!cfqq->new_cfqq)
2384                         cfq_setup_merge(cfqq, new_cfqq);
2385                 goto expire;
2386         }
2387
2388         /*
2389          * No requests pending. If the active queue still has requests in
2390          * flight or is idling for a new request, allow either of these
2391          * conditions to happen (or time out) before selecting a new queue.
2392          */
2393         if (timer_pending(&cfqd->idle_slice_timer)) {
2394                 cfqq = NULL;
2395                 goto keep_queue;
2396         }
2397
2398         /*
2399          * This is a deep seek queue, but the device is much faster than
2400          * the queue can deliver, don't idle
2401          **/
2402         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2403             (cfq_cfqq_slice_new(cfqq) ||
2404             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2405                 cfq_clear_cfqq_deep(cfqq);
2406                 cfq_clear_cfqq_idle_window(cfqq);
2407         }
2408
2409         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2410                 cfqq = NULL;
2411                 goto keep_queue;
2412         }
2413
2414         /*
2415          * If group idle is enabled and there are requests dispatched from
2416          * this group, wait for requests to complete.
2417          */
2418 check_group_idle:
2419         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2420             cfqq->cfqg->dispatched &&
2421             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2422                 cfqq = NULL;
2423                 goto keep_queue;
2424         }
2425
2426 expire:
2427         cfq_slice_expired(cfqd, 0);
2428 new_queue:
2429         /*
2430          * Current queue expired. Check if we have to switch to a new
2431          * service tree
2432          */
2433         if (!new_cfqq)
2434                 cfq_choose_cfqg(cfqd);
2435
2436         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2437 keep_queue:
2438         return cfqq;
2439 }
2440
2441 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2442 {
2443         int dispatched = 0;
2444
2445         while (cfqq->next_rq) {
2446                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2447                 dispatched++;
2448         }
2449
2450         BUG_ON(!list_empty(&cfqq->fifo));
2451
2452         /* By default cfqq is not expired if it is empty. Do it explicitly */
2453         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2454         return dispatched;
2455 }
2456
2457 /*
2458  * Drain our current requests. Used for barriers and when switching
2459  * io schedulers on-the-fly.
2460  */
2461 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2462 {
2463         struct cfq_queue *cfqq;
2464         int dispatched = 0;
2465
2466         /* Expire the timeslice of the current active queue first */
2467         cfq_slice_expired(cfqd, 0);
2468         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2469                 __cfq_set_active_queue(cfqd, cfqq);
2470                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2471         }
2472
2473         BUG_ON(cfqd->busy_queues);
2474
2475         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2476         return dispatched;
2477 }
2478
2479 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2480         struct cfq_queue *cfqq)
2481 {
2482         /* the queue hasn't finished any request, can't estimate */
2483         if (cfq_cfqq_slice_new(cfqq))
2484                 return true;
2485         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2486                 cfqq->slice_end))
2487                 return true;
2488
2489         return false;
2490 }
2491
2492 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2493 {
2494         unsigned int max_dispatch;
2495
2496         /*
2497          * Drain async requests before we start sync IO
2498          */
2499         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2500                 return false;
2501
2502         /*
2503          * If this is an async queue and we have sync IO in flight, let it wait
2504          */
2505         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2506                 return false;
2507
2508         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2509         if (cfq_class_idle(cfqq))
2510                 max_dispatch = 1;
2511
2512         /*
2513          * Does this cfqq already have too much IO in flight?
2514          */
2515         if (cfqq->dispatched >= max_dispatch) {
2516                 bool promote_sync = false;
2517                 /*
2518                  * idle queue must always only have a single IO in flight
2519                  */
2520                 if (cfq_class_idle(cfqq))
2521                         return false;
2522
2523                 /*
2524                  * If there is only one sync queue
2525                  * we can ignore async queue here and give the sync
2526                  * queue no dispatch limit. The reason is a sync queue can
2527                  * preempt async queue, limiting the sync queue doesn't make
2528                  * sense. This is useful for aiostress test.
2529                  */
2530                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2531                         promote_sync = true;
2532
2533                 /*
2534                  * We have other queues, don't allow more IO from this one
2535                  */
2536                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2537                                 !promote_sync)
2538                         return false;
2539
2540                 /*
2541                  * Sole queue user, no limit
2542                  */
2543                 if (cfqd->busy_queues == 1 || promote_sync)
2544                         max_dispatch = -1;
2545                 else
2546                         /*
2547                          * Normally we start throttling cfqq when cfq_quantum/2
2548                          * requests have been dispatched. But we can drive
2549                          * deeper queue depths at the beginning of slice
2550                          * subjected to upper limit of cfq_quantum.
2551                          * */
2552                         max_dispatch = cfqd->cfq_quantum;
2553         }
2554
2555         /*
2556          * Async queues must wait a bit before being allowed dispatch.
2557          * We also ramp up the dispatch depth gradually for async IO,
2558          * based on the last sync IO we serviced
2559          */
2560         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2561                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2562                 unsigned int depth;
2563
2564                 depth = last_sync / cfqd->cfq_slice[1];
2565                 if (!depth && !cfqq->dispatched)
2566                         depth = 1;
2567                 if (depth < max_dispatch)
2568                         max_dispatch = depth;
2569         }
2570
2571         /*
2572          * If we're below the current max, allow a dispatch
2573          */
2574         return cfqq->dispatched < max_dispatch;
2575 }
2576
2577 /*
2578  * Dispatch a request from cfqq, moving them to the request queue
2579  * dispatch list.
2580  */
2581 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2582 {
2583         struct request *rq;
2584
2585         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2586
2587         if (!cfq_may_dispatch(cfqd, cfqq))
2588                 return false;
2589
2590         /*
2591          * follow expired path, else get first next available
2592          */
2593         rq = cfq_check_fifo(cfqq);
2594         if (!rq)
2595                 rq = cfqq->next_rq;
2596
2597         /*
2598          * insert request into driver dispatch list
2599          */
2600         cfq_dispatch_insert(cfqd->queue, rq);
2601
2602         if (!cfqd->active_cic) {
2603                 struct cfq_io_cq *cic = RQ_CIC(rq);
2604
2605                 atomic_long_inc(&cic->icq.ioc->refcount);
2606                 cfqd->active_cic = cic;
2607         }
2608
2609         return true;
2610 }
2611
2612 /*
2613  * Find the cfqq that we need to service and move a request from that to the
2614  * dispatch list
2615  */
2616 static int cfq_dispatch_requests(struct request_queue *q, int force)
2617 {
2618         struct cfq_data *cfqd = q->elevator->elevator_data;
2619         struct cfq_queue *cfqq;
2620
2621         if (!cfqd->busy_queues)
2622                 return 0;
2623
2624         if (unlikely(force))
2625                 return cfq_forced_dispatch(cfqd);
2626
2627         cfqq = cfq_select_queue(cfqd);
2628         if (!cfqq)
2629                 return 0;
2630
2631         /*
2632          * Dispatch a request from this cfqq, if it is allowed
2633          */
2634         if (!cfq_dispatch_request(cfqd, cfqq))
2635                 return 0;
2636
2637         cfqq->slice_dispatch++;
2638         cfq_clear_cfqq_must_dispatch(cfqq);
2639
2640         /*
2641          * expire an async queue immediately if it has used up its slice. idle
2642          * queue always expire after 1 dispatch round.
2643          */
2644         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2645             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2646             cfq_class_idle(cfqq))) {
2647                 cfqq->slice_end = jiffies + 1;
2648                 cfq_slice_expired(cfqd, 0);
2649         }
2650
2651         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2652         return 1;
2653 }
2654
2655 /*
2656  * task holds one reference to the queue, dropped when task exits. each rq
2657  * in-flight on this queue also holds a reference, dropped when rq is freed.
2658  *
2659  * Each cfq queue took a reference on the parent group. Drop it now.
2660  * queue lock must be held here.
2661  */
2662 static void cfq_put_queue(struct cfq_queue *cfqq)
2663 {
2664         struct cfq_data *cfqd = cfqq->cfqd;
2665         struct cfq_group *cfqg;
2666
2667         BUG_ON(cfqq->ref <= 0);
2668
2669         cfqq->ref--;
2670         if (cfqq->ref)
2671                 return;
2672
2673         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2674         BUG_ON(rb_first(&cfqq->sort_list));
2675         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2676         cfqg = cfqq->cfqg;
2677
2678         if (unlikely(cfqd->active_queue == cfqq)) {
2679                 __cfq_slice_expired(cfqd, cfqq, 0);
2680                 cfq_schedule_dispatch(cfqd);
2681         }
2682
2683         BUG_ON(cfq_cfqq_on_rr(cfqq));
2684         kmem_cache_free(cfq_pool, cfqq);
2685         cfq_put_cfqg(cfqg);
2686 }
2687
2688 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2689 {
2690         struct cfq_queue *__cfqq, *next;
2691
2692         /*
2693          * If this queue was scheduled to merge with another queue, be
2694          * sure to drop the reference taken on that queue (and others in
2695          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2696          */
2697         __cfqq = cfqq->new_cfqq;
2698         while (__cfqq) {
2699                 if (__cfqq == cfqq) {
2700                         WARN(1, "cfqq->new_cfqq loop detected\n");
2701                         break;
2702                 }
2703                 next = __cfqq->new_cfqq;
2704                 cfq_put_queue(__cfqq);
2705                 __cfqq = next;
2706         }
2707 }
2708
2709 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2710 {
2711         if (unlikely(cfqq == cfqd->active_queue)) {
2712                 __cfq_slice_expired(cfqd, cfqq, 0);
2713                 cfq_schedule_dispatch(cfqd);
2714         }
2715
2716         cfq_put_cooperator(cfqq);
2717
2718         cfq_put_queue(cfqq);
2719 }
2720
2721 static void cfq_init_icq(struct io_cq *icq)
2722 {
2723         struct cfq_io_cq *cic = icq_to_cic(icq);
2724
2725         cic->ttime.last_end_request = jiffies;
2726 }
2727
2728 static void cfq_exit_icq(struct io_cq *icq)
2729 {
2730         struct cfq_io_cq *cic = icq_to_cic(icq);
2731         struct cfq_data *cfqd = cic_to_cfqd(cic);
2732
2733         if (cic->cfqq[BLK_RW_ASYNC]) {
2734                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2735                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2736         }
2737
2738         if (cic->cfqq[BLK_RW_SYNC]) {
2739                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2740                 cic->cfqq[BLK_RW_SYNC] = NULL;
2741         }
2742 }
2743
2744 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2745 {
2746         struct task_struct *tsk = current;
2747         int ioprio_class;
2748
2749         if (!cfq_cfqq_prio_changed(cfqq))
2750                 return;
2751
2752         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2753         switch (ioprio_class) {
2754         default:
2755                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2756         case IOPRIO_CLASS_NONE:
2757                 /*
2758                  * no prio set, inherit CPU scheduling settings
2759                  */
2760                 cfqq->ioprio = task_nice_ioprio(tsk);
2761                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2762                 break;
2763         case IOPRIO_CLASS_RT:
2764                 cfqq->ioprio = task_ioprio(ioc);
2765                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2766                 break;
2767         case IOPRIO_CLASS_BE:
2768                 cfqq->ioprio = task_ioprio(ioc);
2769                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2770                 break;
2771         case IOPRIO_CLASS_IDLE:
2772                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2773                 cfqq->ioprio = 7;
2774                 cfq_clear_cfqq_idle_window(cfqq);
2775                 break;
2776         }
2777
2778         /*
2779          * keep track of original prio settings in case we have to temporarily
2780          * elevate the priority of this queue
2781          */
2782         cfqq->org_ioprio = cfqq->ioprio;
2783         cfq_clear_cfqq_prio_changed(cfqq);
2784 }
2785
2786 static void changed_ioprio(struct cfq_io_cq *cic)
2787 {
2788         struct cfq_data *cfqd = cic_to_cfqd(cic);
2789         struct cfq_queue *cfqq;
2790
2791         if (unlikely(!cfqd))
2792                 return;
2793
2794         cfqq = cic->cfqq[BLK_RW_ASYNC];
2795         if (cfqq) {
2796                 struct cfq_queue *new_cfqq;
2797                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->icq.ioc,
2798                                                 GFP_ATOMIC);
2799                 if (new_cfqq) {
2800                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2801                         cfq_put_queue(cfqq);
2802                 }
2803         }
2804
2805         cfqq = cic->cfqq[BLK_RW_SYNC];
2806         if (cfqq)
2807                 cfq_mark_cfqq_prio_changed(cfqq);
2808 }
2809
2810 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2811                           pid_t pid, bool is_sync)
2812 {
2813         RB_CLEAR_NODE(&cfqq->rb_node);
2814         RB_CLEAR_NODE(&cfqq->p_node);
2815         INIT_LIST_HEAD(&cfqq->fifo);
2816
2817         cfqq->ref = 0;
2818         cfqq->cfqd = cfqd;
2819
2820         cfq_mark_cfqq_prio_changed(cfqq);
2821
2822         if (is_sync) {
2823                 if (!cfq_class_idle(cfqq))
2824                         cfq_mark_cfqq_idle_window(cfqq);
2825                 cfq_mark_cfqq_sync(cfqq);
2826         }
2827         cfqq->pid = pid;
2828 }
2829
2830 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2831 static void changed_cgroup(struct cfq_io_cq *cic)
2832 {
2833         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2834         struct cfq_data *cfqd = cic_to_cfqd(cic);
2835         struct request_queue *q;
2836
2837         if (unlikely(!cfqd))
2838                 return;
2839
2840         q = cfqd->queue;
2841
2842         if (sync_cfqq) {
2843                 /*
2844                  * Drop reference to sync queue. A new sync queue will be
2845                  * assigned in new group upon arrival of a fresh request.
2846                  */
2847                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2848                 cic_set_cfqq(cic, NULL, 1);
2849                 cfq_put_queue(sync_cfqq);
2850         }
2851 }
2852 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2853
2854 static struct cfq_queue *
2855 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2856                      struct io_context *ioc, gfp_t gfp_mask)
2857 {
2858         struct cfq_queue *cfqq, *new_cfqq = NULL;
2859         struct cfq_io_cq *cic;
2860         struct cfq_group *cfqg;
2861
2862 retry:
2863         cfqg = cfq_get_cfqg(cfqd);
2864         cic = cfq_cic_lookup(cfqd, ioc);
2865         /* cic always exists here */
2866         cfqq = cic_to_cfqq(cic, is_sync);
2867
2868         /*
2869          * Always try a new alloc if we fell back to the OOM cfqq
2870          * originally, since it should just be a temporary situation.
2871          */
2872         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2873                 cfqq = NULL;
2874                 if (new_cfqq) {
2875                         cfqq = new_cfqq;
2876                         new_cfqq = NULL;
2877                 } else if (gfp_mask & __GFP_WAIT) {
2878                         spin_unlock_irq(cfqd->queue->queue_lock);
2879                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2880                                         gfp_mask | __GFP_ZERO,
2881                                         cfqd->queue->node);
2882                         spin_lock_irq(cfqd->queue->queue_lock);
2883                         if (new_cfqq)
2884                                 goto retry;
2885                 } else {
2886                         cfqq = kmem_cache_alloc_node(cfq_pool,
2887                                         gfp_mask | __GFP_ZERO,
2888                                         cfqd->queue->node);
2889                 }
2890
2891                 if (cfqq) {
2892                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2893                         cfq_init_prio_data(cfqq, ioc);
2894                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2895                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2896                 } else
2897                         cfqq = &cfqd->oom_cfqq;
2898         }
2899
2900         if (new_cfqq)
2901                 kmem_cache_free(cfq_pool, new_cfqq);
2902
2903         return cfqq;
2904 }
2905
2906 static struct cfq_queue **
2907 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2908 {
2909         switch (ioprio_class) {
2910         case IOPRIO_CLASS_RT:
2911                 return &cfqd->async_cfqq[0][ioprio];
2912         case IOPRIO_CLASS_BE:
2913                 return &cfqd->async_cfqq[1][ioprio];
2914         case IOPRIO_CLASS_IDLE:
2915                 return &cfqd->async_idle_cfqq;
2916         default:
2917                 BUG();
2918         }
2919 }
2920
2921 static struct cfq_queue *
2922 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2923               gfp_t gfp_mask)
2924 {
2925         const int ioprio = task_ioprio(ioc);
2926         const int ioprio_class = task_ioprio_class(ioc);
2927         struct cfq_queue **async_cfqq = NULL;
2928         struct cfq_queue *cfqq = NULL;
2929
2930         if (!is_sync) {
2931                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2932                 cfqq = *async_cfqq;
2933         }
2934
2935         if (!cfqq)
2936                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2937
2938         /*
2939          * pin the queue now that it's allocated, scheduler exit will prune it
2940          */
2941         if (!is_sync && !(*async_cfqq)) {
2942                 cfqq->ref++;
2943                 *async_cfqq = cfqq;
2944         }
2945
2946         cfqq->ref++;
2947         return cfqq;
2948 }
2949
2950 static void
2951 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
2952 {
2953         unsigned long elapsed = jiffies - ttime->last_end_request;
2954         elapsed = min(elapsed, 2UL * slice_idle);
2955
2956         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
2957         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
2958         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
2959 }
2960
2961 static void
2962 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2963                         struct cfq_io_cq *cic)
2964 {
2965         if (cfq_cfqq_sync(cfqq)) {
2966                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
2967                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
2968                         cfqd->cfq_slice_idle);
2969         }
2970 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2971         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
2972 #endif
2973 }
2974
2975 static void
2976 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2977                        struct request *rq)
2978 {
2979         sector_t sdist = 0;
2980         sector_t n_sec = blk_rq_sectors(rq);
2981         if (cfqq->last_request_pos) {
2982                 if (cfqq->last_request_pos < blk_rq_pos(rq))
2983                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
2984                 else
2985                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
2986         }
2987
2988         cfqq->seek_history <<= 1;
2989         if (blk_queue_nonrot(cfqd->queue))
2990                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
2991         else
2992                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
2993 }
2994
2995 /*
2996  * Disable idle window if the process thinks too long or seeks so much that
2997  * it doesn't matter
2998  */
2999 static void
3000 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3001                        struct cfq_io_cq *cic)
3002 {
3003         int old_idle, enable_idle;
3004
3005         /*
3006          * Don't idle for async or idle io prio class
3007          */
3008         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3009                 return;
3010
3011         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3012
3013         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3014                 cfq_mark_cfqq_deep(cfqq);
3015
3016         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3017                 enable_idle = 0;
3018         else if (!atomic_read(&cic->icq.ioc->nr_tasks) ||
3019                  !cfqd->cfq_slice_idle ||
3020                  (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3021                 enable_idle = 0;
3022         else if (sample_valid(cic->ttime.ttime_samples)) {
3023                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3024                         enable_idle = 0;
3025                 else
3026                         enable_idle = 1;
3027         }
3028
3029         if (old_idle != enable_idle) {
3030                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3031                 if (enable_idle)
3032                         cfq_mark_cfqq_idle_window(cfqq);
3033                 else
3034                         cfq_clear_cfqq_idle_window(cfqq);
3035         }
3036 }
3037
3038 /*
3039  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3040  * no or if we aren't sure, a 1 will cause a preempt.
3041  */
3042 static bool
3043 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3044                    struct request *rq)
3045 {
3046         struct cfq_queue *cfqq;
3047
3048         cfqq = cfqd->active_queue;
3049         if (!cfqq)
3050                 return false;
3051
3052         if (cfq_class_idle(new_cfqq))
3053                 return false;
3054
3055         if (cfq_class_idle(cfqq))
3056                 return true;
3057
3058         /*
3059          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3060          */
3061         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3062                 return false;
3063
3064         /*
3065          * if the new request is sync, but the currently running queue is
3066          * not, let the sync request have priority.
3067          */
3068         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3069                 return true;
3070
3071         if (new_cfqq->cfqg != cfqq->cfqg)
3072                 return false;
3073
3074         if (cfq_slice_used(cfqq))
3075                 return true;
3076
3077         /* Allow preemption only if we are idling on sync-noidle tree */
3078         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3079             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3080             new_cfqq->service_tree->count == 2 &&
3081             RB_EMPTY_ROOT(&cfqq->sort_list))
3082                 return true;
3083
3084         /*
3085          * So both queues are sync. Let the new request get disk time if
3086          * it's a metadata request and the current queue is doing regular IO.
3087          */
3088         if ((rq->cmd_flags & REQ_PRIO) && !cfqq->prio_pending)
3089                 return true;
3090
3091         /*
3092          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3093          */
3094         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3095                 return true;
3096
3097         /* An idle queue should not be idle now for some reason */
3098         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3099                 return true;
3100
3101         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3102                 return false;
3103
3104         /*
3105          * if this request is as-good as one we would expect from the
3106          * current cfqq, let it preempt
3107          */
3108         if (cfq_rq_close(cfqd, cfqq, rq))
3109                 return true;
3110
3111         return false;
3112 }
3113
3114 /*
3115  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3116  * let it have half of its nominal slice.
3117  */
3118 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3119 {
3120         enum wl_type_t old_type = cfqq_type(cfqd->active_queue);
3121
3122         cfq_log_cfqq(cfqd, cfqq, "preempt");
3123         cfq_slice_expired(cfqd, 1);
3124
3125         /*
3126          * workload type is changed, don't save slice, otherwise preempt
3127          * doesn't happen
3128          */
3129         if (old_type != cfqq_type(cfqq))
3130                 cfqq->cfqg->saved_workload_slice = 0;
3131
3132         /*
3133          * Put the new queue at the front of the of the current list,
3134          * so we know that it will be selected next.
3135          */
3136         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3137
3138         cfq_service_tree_add(cfqd, cfqq, 1);
3139
3140         cfqq->slice_end = 0;
3141         cfq_mark_cfqq_slice_new(cfqq);
3142 }
3143
3144 /*
3145  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3146  * something we should do about it
3147  */
3148 static void
3149 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3150                 struct request *rq)
3151 {
3152         struct cfq_io_cq *cic = RQ_CIC(rq);
3153
3154         cfqd->rq_queued++;
3155         if (rq->cmd_flags & REQ_PRIO)
3156                 cfqq->prio_pending++;
3157
3158         cfq_update_io_thinktime(cfqd, cfqq, cic);
3159         cfq_update_io_seektime(cfqd, cfqq, rq);
3160         cfq_update_idle_window(cfqd, cfqq, cic);
3161
3162         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3163
3164         if (cfqq == cfqd->active_queue) {
3165                 /*
3166                  * Remember that we saw a request from this process, but
3167                  * don't start queuing just yet. Otherwise we risk seeing lots
3168                  * of tiny requests, because we disrupt the normal plugging
3169                  * and merging. If the request is already larger than a single
3170                  * page, let it rip immediately. For that case we assume that
3171                  * merging is already done. Ditto for a busy system that
3172                  * has other work pending, don't risk delaying until the
3173                  * idle timer unplug to continue working.
3174                  */
3175                 if (cfq_cfqq_wait_request(cfqq)) {
3176                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3177                             cfqd->busy_queues > 1) {
3178                                 cfq_del_timer(cfqd, cfqq);
3179                                 cfq_clear_cfqq_wait_request(cfqq);
3180                                 __blk_run_queue(cfqd->queue);
3181                         } else {
3182                                 cfq_blkiocg_update_idle_time_stats(
3183                                                 &cfqq->cfqg->blkg);
3184                                 cfq_mark_cfqq_must_dispatch(cfqq);
3185                         }
3186                 }
3187         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3188                 /*
3189                  * not the active queue - expire current slice if it is
3190                  * idle and has expired it's mean thinktime or this new queue
3191                  * has some old slice time left and is of higher priority or
3192                  * this new queue is RT and the current one is BE
3193                  */
3194                 cfq_preempt_queue(cfqd, cfqq);
3195                 __blk_run_queue(cfqd->queue);
3196         }
3197 }
3198
3199 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3200 {
3201         struct cfq_data *cfqd = q->elevator->elevator_data;
3202         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3203
3204         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3205         cfq_init_prio_data(cfqq, RQ_CIC(rq)->icq.ioc);
3206
3207         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3208         list_add_tail(&rq->queuelist, &cfqq->fifo);
3209         cfq_add_rq_rb(rq);
3210         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3211                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3212                         rq_is_sync(rq));
3213         cfq_rq_enqueued(cfqd, cfqq, rq);
3214 }
3215
3216 /*
3217  * Update hw_tag based on peak queue depth over 50 samples under
3218  * sufficient load.
3219  */
3220 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3221 {
3222         struct cfq_queue *cfqq = cfqd->active_queue;
3223
3224         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3225                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3226
3227         if (cfqd->hw_tag == 1)
3228                 return;
3229
3230         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3231             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3232                 return;
3233
3234         /*
3235          * If active queue hasn't enough requests and can idle, cfq might not
3236          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3237          * case
3238          */
3239         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3240             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3241             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3242                 return;
3243
3244         if (cfqd->hw_tag_samples++ < 50)
3245                 return;
3246
3247         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3248                 cfqd->hw_tag = 1;
3249         else
3250                 cfqd->hw_tag = 0;
3251 }
3252
3253 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3254 {
3255         struct cfq_io_cq *cic = cfqd->active_cic;
3256
3257         /* If the queue already has requests, don't wait */
3258         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3259                 return false;
3260
3261         /* If there are other queues in the group, don't wait */
3262         if (cfqq->cfqg->nr_cfqq > 1)
3263                 return false;
3264
3265         /* the only queue in the group, but think time is big */
3266         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3267                 return false;
3268
3269         if (cfq_slice_used(cfqq))
3270                 return true;
3271
3272         /* if slice left is less than think time, wait busy */
3273         if (cic && sample_valid(cic->ttime.ttime_samples)
3274             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3275                 return true;
3276
3277         /*
3278          * If think times is less than a jiffy than ttime_mean=0 and above
3279          * will not be true. It might happen that slice has not expired yet
3280          * but will expire soon (4-5 ns) during select_queue(). To cover the
3281          * case where think time is less than a jiffy, mark the queue wait
3282          * busy if only 1 jiffy is left in the slice.
3283          */
3284         if (cfqq->slice_end - jiffies == 1)
3285                 return true;
3286
3287         return false;
3288 }
3289
3290 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3291 {
3292         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3293         struct cfq_data *cfqd = cfqq->cfqd;
3294         const int sync = rq_is_sync(rq);
3295         unsigned long now;
3296
3297         now = jiffies;
3298         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3299                      !!(rq->cmd_flags & REQ_NOIDLE));
3300
3301         cfq_update_hw_tag(cfqd);
3302
3303         WARN_ON(!cfqd->rq_in_driver);
3304         WARN_ON(!cfqq->dispatched);
3305         cfqd->rq_in_driver--;
3306         cfqq->dispatched--;
3307         (RQ_CFQG(rq))->dispatched--;
3308         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3309                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3310                         rq_data_dir(rq), rq_is_sync(rq));
3311
3312         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3313
3314         if (sync) {
3315                 struct cfq_rb_root *service_tree;
3316
3317                 RQ_CIC(rq)->ttime.last_end_request = now;
3318
3319                 if (cfq_cfqq_on_rr(cfqq))
3320                         service_tree = cfqq->service_tree;
3321                 else
3322                         service_tree = service_tree_for(cfqq->cfqg,
3323                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3324                 service_tree->ttime.last_end_request = now;
3325                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3326                         cfqd->last_delayed_sync = now;
3327         }
3328
3329 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3330         cfqq->cfqg->ttime.last_end_request = now;
3331 #endif
3332
3333         /*
3334          * If this is the active queue, check if it needs to be expired,
3335          * or if we want to idle in case it has no pending requests.
3336          */
3337         if (cfqd->active_queue == cfqq) {
3338                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3339
3340                 if (cfq_cfqq_slice_new(cfqq)) {
3341                         cfq_set_prio_slice(cfqd, cfqq);
3342                         cfq_clear_cfqq_slice_new(cfqq);
3343                 }
3344
3345                 /*
3346                  * Should we wait for next request to come in before we expire
3347                  * the queue.
3348                  */
3349                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3350                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3351                         if (!cfqd->cfq_slice_idle)
3352                                 extend_sl = cfqd->cfq_group_idle;
3353                         cfqq->slice_end = jiffies + extend_sl;
3354                         cfq_mark_cfqq_wait_busy(cfqq);
3355                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3356                 }
3357
3358                 /*
3359                  * Idling is not enabled on:
3360                  * - expired queues
3361                  * - idle-priority queues
3362                  * - async queues
3363                  * - queues with still some requests queued
3364                  * - when there is a close cooperator
3365                  */
3366                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3367                         cfq_slice_expired(cfqd, 1);
3368                 else if (sync && cfqq_empty &&
3369                          !cfq_close_cooperator(cfqd, cfqq)) {
3370                         cfq_arm_slice_timer(cfqd);
3371                 }
3372         }
3373
3374         if (!cfqd->rq_in_driver)
3375                 cfq_schedule_dispatch(cfqd);
3376 }
3377
3378 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3379 {
3380         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3381                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3382                 return ELV_MQUEUE_MUST;
3383         }
3384
3385         return ELV_MQUEUE_MAY;
3386 }
3387
3388 static int cfq_may_queue(struct request_queue *q, int rw)
3389 {
3390         struct cfq_data *cfqd = q->elevator->elevator_data;
3391         struct task_struct *tsk = current;
3392         struct cfq_io_cq *cic;
3393         struct cfq_queue *cfqq;
3394
3395         /*
3396          * don't force setup of a queue from here, as a call to may_queue
3397          * does not necessarily imply that a request actually will be queued.
3398          * so just lookup a possibly existing queue, or return 'may queue'
3399          * if that fails
3400          */
3401         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3402         if (!cic)
3403                 return ELV_MQUEUE_MAY;
3404
3405         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3406         if (cfqq) {
3407                 cfq_init_prio_data(cfqq, cic->icq.ioc);
3408
3409                 return __cfq_may_queue(cfqq);
3410         }
3411
3412         return ELV_MQUEUE_MAY;
3413 }
3414
3415 /*
3416  * queue lock held here
3417  */
3418 static void cfq_put_request(struct request *rq)
3419 {
3420         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3421
3422         if (cfqq) {
3423                 const int rw = rq_data_dir(rq);
3424
3425                 BUG_ON(!cfqq->allocated[rw]);
3426                 cfqq->allocated[rw]--;
3427
3428                 /* Put down rq reference on cfqg */
3429                 cfq_put_cfqg(RQ_CFQG(rq));
3430                 rq->elv.priv[0] = NULL;
3431                 rq->elv.priv[1] = NULL;
3432
3433                 cfq_put_queue(cfqq);
3434         }
3435 }
3436
3437 static struct cfq_queue *
3438 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_cq *cic,
3439                 struct cfq_queue *cfqq)
3440 {
3441         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3442         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3443         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3444         cfq_put_queue(cfqq);
3445         return cic_to_cfqq(cic, 1);
3446 }
3447
3448 /*
3449  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3450  * was the last process referring to said cfqq.
3451  */
3452 static struct cfq_queue *
3453 split_cfqq(struct cfq_io_cq *cic, struct cfq_queue *cfqq)
3454 {
3455         if (cfqq_process_refs(cfqq) == 1) {
3456                 cfqq->pid = current->pid;
3457                 cfq_clear_cfqq_coop(cfqq);
3458                 cfq_clear_cfqq_split_coop(cfqq);
3459                 return cfqq;
3460         }
3461
3462         cic_set_cfqq(cic, NULL, 1);
3463
3464         cfq_put_cooperator(cfqq);
3465
3466         cfq_put_queue(cfqq);
3467         return NULL;
3468 }
3469 /*
3470  * Allocate cfq data structures associated with this request.
3471  */
3472 static int
3473 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3474 {
3475         struct cfq_data *cfqd = q->elevator->elevator_data;
3476         struct cfq_io_cq *cic = icq_to_cic(rq->elv.icq);
3477         const int rw = rq_data_dir(rq);
3478         const bool is_sync = rq_is_sync(rq);
3479         struct cfq_queue *cfqq;
3480
3481         might_sleep_if(gfp_mask & __GFP_WAIT);
3482
3483         spin_lock_irq(q->queue_lock);
3484
3485         /* handle changed notifications */
3486         if (unlikely(cic->icq.changed)) {
3487                 if (test_and_clear_bit(ICQ_IOPRIO_CHANGED, &cic->icq.changed))
3488                         changed_ioprio(cic);
3489 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3490                 if (test_and_clear_bit(ICQ_CGROUP_CHANGED, &cic->icq.changed))
3491                         changed_cgroup(cic);
3492 #endif
3493         }
3494
3495 new_queue:
3496         cfqq = cic_to_cfqq(cic, is_sync);
3497         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3498                 cfqq = cfq_get_queue(cfqd, is_sync, cic->icq.ioc, gfp_mask);
3499                 cic_set_cfqq(cic, cfqq, is_sync);
3500         } else {
3501                 /*
3502                  * If the queue was seeky for too long, break it apart.
3503                  */
3504                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3505                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3506                         cfqq = split_cfqq(cic, cfqq);
3507                         if (!cfqq)
3508                                 goto new_queue;
3509                 }
3510
3511                 /*
3512                  * Check to see if this queue is scheduled to merge with
3513                  * another, closely cooperating queue.  The merging of
3514                  * queues happens here as it must be done in process context.
3515                  * The reference on new_cfqq was taken in merge_cfqqs.
3516                  */
3517                 if (cfqq->new_cfqq)
3518                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3519         }
3520
3521         cfqq->allocated[rw]++;
3522
3523         cfqq->ref++;
3524         rq->elv.priv[0] = cfqq;
3525         rq->elv.priv[1] = cfq_ref_get_cfqg(cfqq->cfqg);
3526         spin_unlock_irq(q->queue_lock);
3527         return 0;
3528 }
3529
3530 static void cfq_kick_queue(struct work_struct *work)
3531 {
3532         struct cfq_data *cfqd =
3533                 container_of(work, struct cfq_data, unplug_work);
3534         struct request_queue *q = cfqd->queue;
3535
3536         spin_lock_irq(q->queue_lock);
3537         __blk_run_queue(cfqd->queue);
3538         spin_unlock_irq(q->queue_lock);
3539 }
3540
3541 /*
3542  * Timer running if the active_queue is currently idling inside its time slice
3543  */
3544 static void cfq_idle_slice_timer(unsigned long data)
3545 {
3546         struct cfq_data *cfqd = (struct cfq_data *) data;
3547         struct cfq_queue *cfqq;
3548         unsigned long flags;
3549         int timed_out = 1;
3550
3551         cfq_log(cfqd, "idle timer fired");
3552
3553         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3554
3555         cfqq = cfqd->active_queue;
3556         if (cfqq) {
3557                 timed_out = 0;
3558
3559                 /*
3560                  * We saw a request before the queue expired, let it through
3561                  */
3562                 if (cfq_cfqq_must_dispatch(cfqq))
3563                         goto out_kick;
3564
3565                 /*
3566                  * expired
3567                  */
3568                 if (cfq_slice_used(cfqq))
3569                         goto expire;
3570
3571                 /*
3572                  * only expire and reinvoke request handler, if there are
3573                  * other queues with pending requests
3574                  */
3575                 if (!cfqd->busy_queues)
3576                         goto out_cont;
3577
3578                 /*
3579                  * not expired and it has a request pending, let it dispatch
3580                  */
3581                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3582                         goto out_kick;
3583
3584                 /*
3585                  * Queue depth flag is reset only when the idle didn't succeed
3586                  */
3587                 cfq_clear_cfqq_deep(cfqq);
3588         }
3589 expire:
3590         cfq_slice_expired(cfqd, timed_out);
3591 out_kick:
3592         cfq_schedule_dispatch(cfqd);
3593 out_cont:
3594         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3595 }
3596
3597 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3598 {
3599         del_timer_sync(&cfqd->idle_slice_timer);
3600         cancel_work_sync(&cfqd->unplug_work);
3601 }
3602
3603 static void cfq_put_async_queues(struct cfq_data *cfqd)
3604 {
3605         int i;
3606
3607         for (i = 0; i < IOPRIO_BE_NR; i++) {
3608                 if (cfqd->async_cfqq[0][i])
3609                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3610                 if (cfqd->async_cfqq[1][i])
3611                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3612         }
3613
3614         if (cfqd->async_idle_cfqq)
3615                 cfq_put_queue(cfqd->async_idle_cfqq);
3616 }
3617
3618 static void cfq_exit_queue(struct elevator_queue *e)
3619 {
3620         struct cfq_data *cfqd = e->elevator_data;
3621         struct request_queue *q = cfqd->queue;
3622         bool wait = false;
3623
3624         cfq_shutdown_timer_wq(cfqd);
3625
3626         spin_lock_irq(q->queue_lock);
3627
3628         if (cfqd->active_queue)
3629                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3630
3631         cfq_put_async_queues(cfqd);
3632         cfq_release_cfq_groups(cfqd);
3633
3634         /*
3635          * If there are groups which we could not unlink from blkcg list,
3636          * wait for a rcu period for them to be freed.
3637          */
3638         if (cfqd->nr_blkcg_linked_grps)
3639                 wait = true;
3640
3641         spin_unlock_irq(q->queue_lock);
3642
3643         cfq_shutdown_timer_wq(cfqd);
3644
3645         /*
3646          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3647          * Do this wait only if there are other unlinked groups out
3648          * there. This can happen if cgroup deletion path claimed the
3649          * responsibility of cleaning up a group before queue cleanup code
3650          * get to the group.
3651          *
3652          * Do not call synchronize_rcu() unconditionally as there are drivers
3653          * which create/delete request queue hundreds of times during scan/boot
3654          * and synchronize_rcu() can take significant time and slow down boot.
3655          */
3656         if (wait)
3657                 synchronize_rcu();
3658
3659 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3660         /* Free up per cpu stats for root group */
3661         free_percpu(cfqd->root_group.blkg.stats_cpu);
3662 #endif
3663         kfree(cfqd);
3664 }
3665
3666 static void *cfq_init_queue(struct request_queue *q)
3667 {
3668         struct cfq_data *cfqd;
3669         int i, j;
3670         struct cfq_group *cfqg;
3671         struct cfq_rb_root *st;
3672
3673         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3674         if (!cfqd)
3675                 return NULL;
3676
3677         /* Init root service tree */
3678         cfqd->grp_service_tree = CFQ_RB_ROOT;
3679
3680         /* Init root group */
3681         cfqg = &cfqd->root_group;
3682         for_each_cfqg_st(cfqg, i, j, st)
3683                 *st = CFQ_RB_ROOT;
3684         RB_CLEAR_NODE(&cfqg->rb_node);
3685
3686         /* Give preference to root group over other groups */
3687         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3688
3689 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3690         /*
3691          * Set root group reference to 2. One reference will be dropped when
3692          * all groups on cfqd->cfqg_list are being deleted during queue exit.
3693          * Other reference will remain there as we don't want to delete this
3694          * group as it is statically allocated and gets destroyed when
3695          * throtl_data goes away.
3696          */
3697         cfqg->ref = 2;
3698
3699         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
3700                 kfree(cfqg);
3701                 kfree(cfqd);
3702                 return NULL;
3703         }
3704
3705         rcu_read_lock();
3706
3707         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3708                                         (void *)cfqd, 0);
3709         rcu_read_unlock();
3710         cfqd->nr_blkcg_linked_grps++;
3711
3712         /* Add group on cfqd->cfqg_list */
3713         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
3714 #endif
3715         /*
3716          * Not strictly needed (since RB_ROOT just clears the node and we
3717          * zeroed cfqd on alloc), but better be safe in case someone decides
3718          * to add magic to the rb code
3719          */
3720         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3721                 cfqd->prio_trees[i] = RB_ROOT;
3722
3723         /*
3724          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3725          * Grab a permanent reference to it, so that the normal code flow
3726          * will not attempt to free it.
3727          */
3728         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3729         cfqd->oom_cfqq.ref++;
3730         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3731
3732         cfqd->queue = q;
3733
3734         init_timer(&cfqd->idle_slice_timer);
3735         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3736         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3737
3738         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3739
3740         cfqd->cfq_quantum = cfq_quantum;
3741         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3742         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3743         cfqd->cfq_back_max = cfq_back_max;
3744         cfqd->cfq_back_penalty = cfq_back_penalty;
3745         cfqd->cfq_slice[0] = cfq_slice_async;
3746         cfqd->cfq_slice[1] = cfq_slice_sync;
3747         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3748         cfqd->cfq_slice_idle = cfq_slice_idle;
3749         cfqd->cfq_group_idle = cfq_group_idle;
3750         cfqd->cfq_latency = 1;
3751         cfqd->hw_tag = -1;
3752         /*
3753          * we optimistically start assuming sync ops weren't delayed in last
3754          * second, in order to have larger depth for async operations.
3755          */
3756         cfqd->last_delayed_sync = jiffies - HZ;
3757         return cfqd;
3758 }
3759
3760 /*
3761  * sysfs parts below -->
3762  */
3763 static ssize_t
3764 cfq_var_show(unsigned int var, char *page)
3765 {
3766         return sprintf(page, "%d\n", var);
3767 }
3768
3769 static ssize_t
3770 cfq_var_store(unsigned int *var, const char *page, size_t count)
3771 {
3772         char *p = (char *) page;
3773
3774         *var = simple_strtoul(p, &p, 10);
3775         return count;
3776 }
3777
3778 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3779 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3780 {                                                                       \
3781         struct cfq_data *cfqd = e->elevator_data;                       \
3782         unsigned int __data = __VAR;                                    \
3783         if (__CONV)                                                     \
3784                 __data = jiffies_to_msecs(__data);                      \
3785         return cfq_var_show(__data, (page));                            \
3786 }
3787 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3788 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3789 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3790 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3791 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3792 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3793 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
3794 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3795 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3796 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3797 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3798 #undef SHOW_FUNCTION
3799
3800 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3801 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3802 {                                                                       \
3803         struct cfq_data *cfqd = e->elevator_data;                       \
3804         unsigned int __data;                                            \
3805         int ret = cfq_var_store(&__data, (page), count);                \
3806         if (__data < (MIN))                                             \
3807                 __data = (MIN);                                         \
3808         else if (__data > (MAX))                                        \
3809                 __data = (MAX);                                         \
3810         if (__CONV)                                                     \
3811                 *(__PTR) = msecs_to_jiffies(__data);                    \
3812         else                                                            \
3813                 *(__PTR) = __data;                                      \
3814         return ret;                                                     \
3815 }
3816 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3817 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3818                 UINT_MAX, 1);
3819 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3820                 UINT_MAX, 1);
3821 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3822 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3823                 UINT_MAX, 0);
3824 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3825 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
3826 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3827 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3828 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3829                 UINT_MAX, 0);
3830 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3831 #undef STORE_FUNCTION
3832
3833 #define CFQ_ATTR(name) \
3834         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3835
3836 static struct elv_fs_entry cfq_attrs[] = {
3837         CFQ_ATTR(quantum),
3838         CFQ_ATTR(fifo_expire_sync),
3839         CFQ_ATTR(fifo_expire_async),
3840         CFQ_ATTR(back_seek_max),
3841         CFQ_ATTR(back_seek_penalty),
3842         CFQ_ATTR(slice_sync),
3843         CFQ_ATTR(slice_async),
3844         CFQ_ATTR(slice_async_rq),
3845         CFQ_ATTR(slice_idle),
3846         CFQ_ATTR(group_idle),
3847         CFQ_ATTR(low_latency),
3848         __ATTR_NULL
3849 };
3850
3851 static struct elevator_type iosched_cfq = {
3852         .ops = {
3853                 .elevator_merge_fn =            cfq_merge,
3854                 .elevator_merged_fn =           cfq_merged_request,
3855                 .elevator_merge_req_fn =        cfq_merged_requests,
3856                 .elevator_allow_merge_fn =      cfq_allow_merge,
3857                 .elevator_bio_merged_fn =       cfq_bio_merged,
3858                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3859                 .elevator_add_req_fn =          cfq_insert_request,
3860                 .elevator_activate_req_fn =     cfq_activate_request,
3861                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3862                 .elevator_completed_req_fn =    cfq_completed_request,
3863                 .elevator_former_req_fn =       elv_rb_former_request,
3864                 .elevator_latter_req_fn =       elv_rb_latter_request,
3865                 .elevator_init_icq_fn =         cfq_init_icq,
3866                 .elevator_exit_icq_fn =         cfq_exit_icq,
3867                 .elevator_set_req_fn =          cfq_set_request,
3868                 .elevator_put_req_fn =          cfq_put_request,
3869                 .elevator_may_queue_fn =        cfq_may_queue,
3870                 .elevator_init_fn =             cfq_init_queue,
3871                 .elevator_exit_fn =             cfq_exit_queue,
3872         },
3873         .icq_size       =       sizeof(struct cfq_io_cq),
3874         .icq_align      =       __alignof__(struct cfq_io_cq),
3875         .elevator_attrs =       cfq_attrs,
3876         .elevator_name  =       "cfq",
3877         .elevator_owner =       THIS_MODULE,
3878 };
3879
3880 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3881 static struct blkio_policy_type blkio_policy_cfq = {
3882         .ops = {
3883                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3884                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3885         },
3886         .plid = BLKIO_POLICY_PROP,
3887 };
3888 #else
3889 static struct blkio_policy_type blkio_policy_cfq;
3890 #endif
3891
3892 static int __init cfq_init(void)
3893 {
3894         int ret;
3895
3896         /*
3897          * could be 0 on HZ < 1000 setups
3898          */
3899         if (!cfq_slice_async)
3900                 cfq_slice_async = 1;
3901         if (!cfq_slice_idle)
3902                 cfq_slice_idle = 1;
3903
3904 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3905         if (!cfq_group_idle)
3906                 cfq_group_idle = 1;
3907 #else
3908                 cfq_group_idle = 0;
3909 #endif
3910         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3911         if (!cfq_pool)
3912                 return -ENOMEM;
3913
3914         ret = elv_register(&iosched_cfq);
3915         if (ret) {
3916                 kmem_cache_destroy(cfq_pool);
3917                 return ret;
3918         }
3919
3920         blkio_policy_register(&blkio_policy_cfq);
3921
3922         return 0;
3923 }
3924
3925 static void __exit cfq_exit(void)
3926 {
3927         blkio_policy_unregister(&blkio_policy_cfq);
3928         elv_unregister(&iosched_cfq);
3929         kmem_cache_destroy(cfq_pool);
3930 }
3931
3932 module_init(cfq_init);
3933 module_exit(cfq_exit);
3934
3935 MODULE_AUTHOR("Jens Axboe");
3936 MODULE_LICENSE("GPL");
3937 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");