block: flush MEDIA_CHANGE from drivers on close(2)
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         int ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         /* Number of sectors dispatched from queue in single dispatch round */
150         unsigned long nr_sectors;
151 };
152
153 /*
154  * First index in the service_trees.
155  * IDLE is handled separately, so it has negative index
156  */
157 enum wl_prio_t {
158         BE_WORKLOAD = 0,
159         RT_WORKLOAD = 1,
160         IDLE_WORKLOAD = 2,
161         CFQ_PRIO_NR,
162 };
163
164 /*
165  * Second index in the service_trees.
166  */
167 enum wl_type_t {
168         ASYNC_WORKLOAD = 0,
169         SYNC_NOIDLE_WORKLOAD = 1,
170         SYNC_WORKLOAD = 2
171 };
172
173 /* This is per cgroup per device grouping structure */
174 struct cfq_group {
175         /* group service_tree member */
176         struct rb_node rb_node;
177
178         /* group service_tree key */
179         u64 vdisktime;
180         unsigned int weight;
181         unsigned int new_weight;
182         bool needs_update;
183
184         /* number of cfqq currently on this group */
185         int nr_cfqq;
186
187         /*
188          * Per group busy queues average. Useful for workload slice calc. We
189          * create the array for each prio class but at run time it is used
190          * only for RT and BE class and slot for IDLE class remains unused.
191          * This is primarily done to avoid confusion and a gcc warning.
192          */
193         unsigned int busy_queues_avg[CFQ_PRIO_NR];
194         /*
195          * rr lists of queues with requests. We maintain service trees for
196          * RT and BE classes. These trees are subdivided in subclasses
197          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198          * class there is no subclassification and all the cfq queues go on
199          * a single tree service_tree_idle.
200          * Counts are embedded in the cfq_rb_root
201          */
202         struct cfq_rb_root service_trees[2][3];
203         struct cfq_rb_root service_tree_idle;
204
205         unsigned long saved_workload_slice;
206         enum wl_type_t saved_workload;
207         enum wl_prio_t saved_serving_prio;
208         struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210         struct hlist_node cfqd_node;
211         int ref;
212 #endif
213         /* number of requests that are on the dispatch list or inside driver */
214         int dispatched;
215 };
216
217 /*
218  * Per block device queue structure
219  */
220 struct cfq_data {
221         struct request_queue *queue;
222         /* Root service tree for cfq_groups */
223         struct cfq_rb_root grp_service_tree;
224         struct cfq_group root_group;
225
226         /*
227          * The priority currently being served
228          */
229         enum wl_prio_t serving_prio;
230         enum wl_type_t serving_type;
231         unsigned long workload_expires;
232         struct cfq_group *serving_group;
233
234         /*
235          * Each priority tree is sorted by next_request position.  These
236          * trees are used when determining if two or more queues are
237          * interleaving requests (see cfq_close_cooperator).
238          */
239         struct rb_root prio_trees[CFQ_PRIO_LISTS];
240
241         unsigned int busy_queues;
242         unsigned int busy_sync_queues;
243
244         int rq_in_driver;
245         int rq_in_flight[2];
246
247         /*
248          * queue-depth detection
249          */
250         int rq_queued;
251         int hw_tag;
252         /*
253          * hw_tag can be
254          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256          *  0 => no NCQ
257          */
258         int hw_tag_est_depth;
259         unsigned int hw_tag_samples;
260
261         /*
262          * idle window management
263          */
264         struct timer_list idle_slice_timer;
265         struct work_struct unplug_work;
266
267         struct cfq_queue *active_queue;
268         struct cfq_io_context *active_cic;
269
270         /*
271          * async queue for each priority case
272          */
273         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274         struct cfq_queue *async_idle_cfqq;
275
276         sector_t last_position;
277
278         /*
279          * tunables, see top of file
280          */
281         unsigned int cfq_quantum;
282         unsigned int cfq_fifo_expire[2];
283         unsigned int cfq_back_penalty;
284         unsigned int cfq_back_max;
285         unsigned int cfq_slice[2];
286         unsigned int cfq_slice_async_rq;
287         unsigned int cfq_slice_idle;
288         unsigned int cfq_group_idle;
289         unsigned int cfq_latency;
290
291         unsigned int cic_index;
292         struct list_head cic_list;
293
294         /*
295          * Fallback dummy cfqq for extreme OOM conditions
296          */
297         struct cfq_queue oom_cfqq;
298
299         unsigned long last_delayed_sync;
300
301         /* List of cfq groups being managed on this device*/
302         struct hlist_head cfqg_list;
303
304         /* Number of groups which are on blkcg->blkg_list */
305         unsigned int nr_blkcg_linked_grps;
306 };
307
308 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
309
310 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
311                                             enum wl_prio_t prio,
312                                             enum wl_type_t type)
313 {
314         if (!cfqg)
315                 return NULL;
316
317         if (prio == IDLE_WORKLOAD)
318                 return &cfqg->service_tree_idle;
319
320         return &cfqg->service_trees[prio][type];
321 }
322
323 enum cfqq_state_flags {
324         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
325         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
326         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
327         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
328         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
329         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
330         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
331         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
332         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
333         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
334         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
335         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
336         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
337 };
338
339 #define CFQ_CFQQ_FNS(name)                                              \
340 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
341 {                                                                       \
342         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
343 }                                                                       \
344 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
345 {                                                                       \
346         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
347 }                                                                       \
348 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
349 {                                                                       \
350         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
351 }
352
353 CFQ_CFQQ_FNS(on_rr);
354 CFQ_CFQQ_FNS(wait_request);
355 CFQ_CFQQ_FNS(must_dispatch);
356 CFQ_CFQQ_FNS(must_alloc_slice);
357 CFQ_CFQQ_FNS(fifo_expire);
358 CFQ_CFQQ_FNS(idle_window);
359 CFQ_CFQQ_FNS(prio_changed);
360 CFQ_CFQQ_FNS(slice_new);
361 CFQ_CFQQ_FNS(sync);
362 CFQ_CFQQ_FNS(coop);
363 CFQ_CFQQ_FNS(split_coop);
364 CFQ_CFQQ_FNS(deep);
365 CFQ_CFQQ_FNS(wait_busy);
366 #undef CFQ_CFQQ_FNS
367
368 #ifdef CONFIG_CFQ_GROUP_IOSCHED
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
370         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
371                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
372                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
373
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
375         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
376                                 blkg_path(&(cfqg)->blkg), ##args)       \
377
378 #else
379 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
380         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
382 #endif
383 #define cfq_log(cfqd, fmt, args...)     \
384         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
385
386 /* Traverses through cfq group service trees */
387 #define for_each_cfqg_st(cfqg, i, j, st) \
388         for (i = 0; i <= IDLE_WORKLOAD; i++) \
389                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
390                         : &cfqg->service_tree_idle; \
391                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
392                         (i == IDLE_WORKLOAD && j == 0); \
393                         j++, st = i < IDLE_WORKLOAD ? \
394                         &cfqg->service_trees[i][j]: NULL) \
395
396
397 static inline bool iops_mode(struct cfq_data *cfqd)
398 {
399         /*
400          * If we are not idling on queues and it is a NCQ drive, parallel
401          * execution of requests is on and measuring time is not possible
402          * in most of the cases until and unless we drive shallower queue
403          * depths and that becomes a performance bottleneck. In such cases
404          * switch to start providing fairness in terms of number of IOs.
405          */
406         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
407                 return true;
408         else
409                 return false;
410 }
411
412 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
413 {
414         if (cfq_class_idle(cfqq))
415                 return IDLE_WORKLOAD;
416         if (cfq_class_rt(cfqq))
417                 return RT_WORKLOAD;
418         return BE_WORKLOAD;
419 }
420
421
422 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
423 {
424         if (!cfq_cfqq_sync(cfqq))
425                 return ASYNC_WORKLOAD;
426         if (!cfq_cfqq_idle_window(cfqq))
427                 return SYNC_NOIDLE_WORKLOAD;
428         return SYNC_WORKLOAD;
429 }
430
431 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
432                                         struct cfq_data *cfqd,
433                                         struct cfq_group *cfqg)
434 {
435         if (wl == IDLE_WORKLOAD)
436                 return cfqg->service_tree_idle.count;
437
438         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
439                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
440                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
441 }
442
443 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
444                                         struct cfq_group *cfqg)
445 {
446         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
447                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
448 }
449
450 static void cfq_dispatch_insert(struct request_queue *, struct request *);
451 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
452                                        struct io_context *, gfp_t);
453 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
454                                                 struct io_context *);
455
456 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
457                                             bool is_sync)
458 {
459         return cic->cfqq[is_sync];
460 }
461
462 static inline void cic_set_cfqq(struct cfq_io_context *cic,
463                                 struct cfq_queue *cfqq, bool is_sync)
464 {
465         cic->cfqq[is_sync] = cfqq;
466 }
467
468 #define CIC_DEAD_KEY    1ul
469 #define CIC_DEAD_INDEX_SHIFT    1
470
471 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
472 {
473         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
474 }
475
476 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
477 {
478         struct cfq_data *cfqd = cic->key;
479
480         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
481                 return NULL;
482
483         return cfqd;
484 }
485
486 /*
487  * We regard a request as SYNC, if it's either a read or has the SYNC bit
488  * set (in which case it could also be direct WRITE).
489  */
490 static inline bool cfq_bio_sync(struct bio *bio)
491 {
492         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
493 }
494
495 /*
496  * scheduler run of queue, if there are requests pending and no one in the
497  * driver that will restart queueing
498  */
499 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
500 {
501         if (cfqd->busy_queues) {
502                 cfq_log(cfqd, "schedule dispatch");
503                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
504         }
505 }
506
507 /*
508  * Scale schedule slice based on io priority. Use the sync time slice only
509  * if a queue is marked sync and has sync io queued. A sync queue with async
510  * io only, should not get full sync slice length.
511  */
512 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
513                                  unsigned short prio)
514 {
515         const int base_slice = cfqd->cfq_slice[sync];
516
517         WARN_ON(prio >= IOPRIO_BE_NR);
518
519         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
520 }
521
522 static inline int
523 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
524 {
525         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
526 }
527
528 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
529 {
530         u64 d = delta << CFQ_SERVICE_SHIFT;
531
532         d = d * BLKIO_WEIGHT_DEFAULT;
533         do_div(d, cfqg->weight);
534         return d;
535 }
536
537 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
538 {
539         s64 delta = (s64)(vdisktime - min_vdisktime);
540         if (delta > 0)
541                 min_vdisktime = vdisktime;
542
543         return min_vdisktime;
544 }
545
546 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
547 {
548         s64 delta = (s64)(vdisktime - min_vdisktime);
549         if (delta < 0)
550                 min_vdisktime = vdisktime;
551
552         return min_vdisktime;
553 }
554
555 static void update_min_vdisktime(struct cfq_rb_root *st)
556 {
557         struct cfq_group *cfqg;
558
559         if (st->left) {
560                 cfqg = rb_entry_cfqg(st->left);
561                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
562                                                   cfqg->vdisktime);
563         }
564 }
565
566 /*
567  * get averaged number of queues of RT/BE priority.
568  * average is updated, with a formula that gives more weight to higher numbers,
569  * to quickly follows sudden increases and decrease slowly
570  */
571
572 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
573                                         struct cfq_group *cfqg, bool rt)
574 {
575         unsigned min_q, max_q;
576         unsigned mult  = cfq_hist_divisor - 1;
577         unsigned round = cfq_hist_divisor / 2;
578         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
579
580         min_q = min(cfqg->busy_queues_avg[rt], busy);
581         max_q = max(cfqg->busy_queues_avg[rt], busy);
582         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
583                 cfq_hist_divisor;
584         return cfqg->busy_queues_avg[rt];
585 }
586
587 static inline unsigned
588 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
589 {
590         struct cfq_rb_root *st = &cfqd->grp_service_tree;
591
592         return cfq_target_latency * cfqg->weight / st->total_weight;
593 }
594
595 static inline unsigned
596 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
597 {
598         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
599         if (cfqd->cfq_latency) {
600                 /*
601                  * interested queues (we consider only the ones with the same
602                  * priority class in the cfq group)
603                  */
604                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
605                                                 cfq_class_rt(cfqq));
606                 unsigned sync_slice = cfqd->cfq_slice[1];
607                 unsigned expect_latency = sync_slice * iq;
608                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
609
610                 if (expect_latency > group_slice) {
611                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
612                         /* scale low_slice according to IO priority
613                          * and sync vs async */
614                         unsigned low_slice =
615                                 min(slice, base_low_slice * slice / sync_slice);
616                         /* the adapted slice value is scaled to fit all iqs
617                          * into the target latency */
618                         slice = max(slice * group_slice / expect_latency,
619                                     low_slice);
620                 }
621         }
622         return slice;
623 }
624
625 static inline void
626 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
627 {
628         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
629
630         cfqq->slice_start = jiffies;
631         cfqq->slice_end = jiffies + slice;
632         cfqq->allocated_slice = slice;
633         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
634 }
635
636 /*
637  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
638  * isn't valid until the first request from the dispatch is activated
639  * and the slice time set.
640  */
641 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
642 {
643         if (cfq_cfqq_slice_new(cfqq))
644                 return false;
645         if (time_before(jiffies, cfqq->slice_end))
646                 return false;
647
648         return true;
649 }
650
651 /*
652  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
653  * We choose the request that is closest to the head right now. Distance
654  * behind the head is penalized and only allowed to a certain extent.
655  */
656 static struct request *
657 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
658 {
659         sector_t s1, s2, d1 = 0, d2 = 0;
660         unsigned long back_max;
661 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
662 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
663         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
664
665         if (rq1 == NULL || rq1 == rq2)
666                 return rq2;
667         if (rq2 == NULL)
668                 return rq1;
669
670         if (rq_is_sync(rq1) != rq_is_sync(rq2))
671                 return rq_is_sync(rq1) ? rq1 : rq2;
672
673         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META)
674                 return rq1->cmd_flags & REQ_META ? rq1 : rq2;
675
676         s1 = blk_rq_pos(rq1);
677         s2 = blk_rq_pos(rq2);
678
679         /*
680          * by definition, 1KiB is 2 sectors
681          */
682         back_max = cfqd->cfq_back_max * 2;
683
684         /*
685          * Strict one way elevator _except_ in the case where we allow
686          * short backward seeks which are biased as twice the cost of a
687          * similar forward seek.
688          */
689         if (s1 >= last)
690                 d1 = s1 - last;
691         else if (s1 + back_max >= last)
692                 d1 = (last - s1) * cfqd->cfq_back_penalty;
693         else
694                 wrap |= CFQ_RQ1_WRAP;
695
696         if (s2 >= last)
697                 d2 = s2 - last;
698         else if (s2 + back_max >= last)
699                 d2 = (last - s2) * cfqd->cfq_back_penalty;
700         else
701                 wrap |= CFQ_RQ2_WRAP;
702
703         /* Found required data */
704
705         /*
706          * By doing switch() on the bit mask "wrap" we avoid having to
707          * check two variables for all permutations: --> faster!
708          */
709         switch (wrap) {
710         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
711                 if (d1 < d2)
712                         return rq1;
713                 else if (d2 < d1)
714                         return rq2;
715                 else {
716                         if (s1 >= s2)
717                                 return rq1;
718                         else
719                                 return rq2;
720                 }
721
722         case CFQ_RQ2_WRAP:
723                 return rq1;
724         case CFQ_RQ1_WRAP:
725                 return rq2;
726         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
727         default:
728                 /*
729                  * Since both rqs are wrapped,
730                  * start with the one that's further behind head
731                  * (--> only *one* back seek required),
732                  * since back seek takes more time than forward.
733                  */
734                 if (s1 <= s2)
735                         return rq1;
736                 else
737                         return rq2;
738         }
739 }
740
741 /*
742  * The below is leftmost cache rbtree addon
743  */
744 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
745 {
746         /* Service tree is empty */
747         if (!root->count)
748                 return NULL;
749
750         if (!root->left)
751                 root->left = rb_first(&root->rb);
752
753         if (root->left)
754                 return rb_entry(root->left, struct cfq_queue, rb_node);
755
756         return NULL;
757 }
758
759 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
760 {
761         if (!root->left)
762                 root->left = rb_first(&root->rb);
763
764         if (root->left)
765                 return rb_entry_cfqg(root->left);
766
767         return NULL;
768 }
769
770 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
771 {
772         rb_erase(n, root);
773         RB_CLEAR_NODE(n);
774 }
775
776 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
777 {
778         if (root->left == n)
779                 root->left = NULL;
780         rb_erase_init(n, &root->rb);
781         --root->count;
782 }
783
784 /*
785  * would be nice to take fifo expire time into account as well
786  */
787 static struct request *
788 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789                   struct request *last)
790 {
791         struct rb_node *rbnext = rb_next(&last->rb_node);
792         struct rb_node *rbprev = rb_prev(&last->rb_node);
793         struct request *next = NULL, *prev = NULL;
794
795         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
796
797         if (rbprev)
798                 prev = rb_entry_rq(rbprev);
799
800         if (rbnext)
801                 next = rb_entry_rq(rbnext);
802         else {
803                 rbnext = rb_first(&cfqq->sort_list);
804                 if (rbnext && rbnext != &last->rb_node)
805                         next = rb_entry_rq(rbnext);
806         }
807
808         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
809 }
810
811 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812                                       struct cfq_queue *cfqq)
813 {
814         /*
815          * just an approximation, should be ok.
816          */
817         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
818                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
819 }
820
821 static inline s64
822 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
823 {
824         return cfqg->vdisktime - st->min_vdisktime;
825 }
826
827 static void
828 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
829 {
830         struct rb_node **node = &st->rb.rb_node;
831         struct rb_node *parent = NULL;
832         struct cfq_group *__cfqg;
833         s64 key = cfqg_key(st, cfqg);
834         int left = 1;
835
836         while (*node != NULL) {
837                 parent = *node;
838                 __cfqg = rb_entry_cfqg(parent);
839
840                 if (key < cfqg_key(st, __cfqg))
841                         node = &parent->rb_left;
842                 else {
843                         node = &parent->rb_right;
844                         left = 0;
845                 }
846         }
847
848         if (left)
849                 st->left = &cfqg->rb_node;
850
851         rb_link_node(&cfqg->rb_node, parent, node);
852         rb_insert_color(&cfqg->rb_node, &st->rb);
853 }
854
855 static void
856 cfq_update_group_weight(struct cfq_group *cfqg)
857 {
858         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
859         if (cfqg->needs_update) {
860                 cfqg->weight = cfqg->new_weight;
861                 cfqg->needs_update = false;
862         }
863 }
864
865 static void
866 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
867 {
868         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
869
870         cfq_update_group_weight(cfqg);
871         __cfq_group_service_tree_add(st, cfqg);
872         st->total_weight += cfqg->weight;
873 }
874
875 static void
876 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
877 {
878         struct cfq_rb_root *st = &cfqd->grp_service_tree;
879         struct cfq_group *__cfqg;
880         struct rb_node *n;
881
882         cfqg->nr_cfqq++;
883         if (!RB_EMPTY_NODE(&cfqg->rb_node))
884                 return;
885
886         /*
887          * Currently put the group at the end. Later implement something
888          * so that groups get lesser vtime based on their weights, so that
889          * if group does not loose all if it was not continuously backlogged.
890          */
891         n = rb_last(&st->rb);
892         if (n) {
893                 __cfqg = rb_entry_cfqg(n);
894                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
895         } else
896                 cfqg->vdisktime = st->min_vdisktime;
897         cfq_group_service_tree_add(st, cfqg);
898 }
899
900 static void
901 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
902 {
903         st->total_weight -= cfqg->weight;
904         if (!RB_EMPTY_NODE(&cfqg->rb_node))
905                 cfq_rb_erase(&cfqg->rb_node, st);
906 }
907
908 static void
909 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
910 {
911         struct cfq_rb_root *st = &cfqd->grp_service_tree;
912
913         BUG_ON(cfqg->nr_cfqq < 1);
914         cfqg->nr_cfqq--;
915
916         /* If there are other cfq queues under this group, don't delete it */
917         if (cfqg->nr_cfqq)
918                 return;
919
920         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
921         cfq_group_service_tree_del(st, cfqg);
922         cfqg->saved_workload_slice = 0;
923         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
924 }
925
926 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
927                                                 unsigned int *unaccounted_time)
928 {
929         unsigned int slice_used;
930
931         /*
932          * Queue got expired before even a single request completed or
933          * got expired immediately after first request completion.
934          */
935         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
936                 /*
937                  * Also charge the seek time incurred to the group, otherwise
938                  * if there are mutiple queues in the group, each can dispatch
939                  * a single request on seeky media and cause lots of seek time
940                  * and group will never know it.
941                  */
942                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
943                                         1);
944         } else {
945                 slice_used = jiffies - cfqq->slice_start;
946                 if (slice_used > cfqq->allocated_slice) {
947                         *unaccounted_time = slice_used - cfqq->allocated_slice;
948                         slice_used = cfqq->allocated_slice;
949                 }
950                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
951                         *unaccounted_time += cfqq->slice_start -
952                                         cfqq->dispatch_start;
953         }
954
955         return slice_used;
956 }
957
958 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
959                                 struct cfq_queue *cfqq)
960 {
961         struct cfq_rb_root *st = &cfqd->grp_service_tree;
962         unsigned int used_sl, charge, unaccounted_sl = 0;
963         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
964                         - cfqg->service_tree_idle.count;
965
966         BUG_ON(nr_sync < 0);
967         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
968
969         if (iops_mode(cfqd))
970                 charge = cfqq->slice_dispatch;
971         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
972                 charge = cfqq->allocated_slice;
973
974         /* Can't update vdisktime while group is on service tree */
975         cfq_group_service_tree_del(st, cfqg);
976         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
977         /* If a new weight was requested, update now, off tree */
978         cfq_group_service_tree_add(st, cfqg);
979
980         /* This group is being expired. Save the context */
981         if (time_after(cfqd->workload_expires, jiffies)) {
982                 cfqg->saved_workload_slice = cfqd->workload_expires
983                                                 - jiffies;
984                 cfqg->saved_workload = cfqd->serving_type;
985                 cfqg->saved_serving_prio = cfqd->serving_prio;
986         } else
987                 cfqg->saved_workload_slice = 0;
988
989         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
990                                         st->min_vdisktime);
991         cfq_log_cfqq(cfqq->cfqd, cfqq,
992                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
993                      used_sl, cfqq->slice_dispatch, charge,
994                      iops_mode(cfqd), cfqq->nr_sectors);
995         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
996                                           unaccounted_sl);
997         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
998 }
999
1000 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1001 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1002 {
1003         if (blkg)
1004                 return container_of(blkg, struct cfq_group, blkg);
1005         return NULL;
1006 }
1007
1008 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1009                                           unsigned int weight)
1010 {
1011         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1012         cfqg->new_weight = weight;
1013         cfqg->needs_update = true;
1014 }
1015
1016 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1017                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1018 {
1019         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1020         unsigned int major, minor;
1021
1022         /*
1023          * Add group onto cgroup list. It might happen that bdi->dev is
1024          * not initialized yet. Initialize this new group without major
1025          * and minor info and this info will be filled in once a new thread
1026          * comes for IO.
1027          */
1028         if (bdi->dev) {
1029                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1030                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1031                                         (void *)cfqd, MKDEV(major, minor));
1032         } else
1033                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1034                                         (void *)cfqd, 0);
1035
1036         cfqd->nr_blkcg_linked_grps++;
1037         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1038
1039         /* Add group on cfqd list */
1040         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1041 }
1042
1043 /*
1044  * Should be called from sleepable context. No request queue lock as per
1045  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1046  * from sleepable context.
1047  */
1048 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1049 {
1050         struct cfq_group *cfqg = NULL;
1051         int i, j, ret;
1052         struct cfq_rb_root *st;
1053
1054         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1055         if (!cfqg)
1056                 return NULL;
1057
1058         for_each_cfqg_st(cfqg, i, j, st)
1059                 *st = CFQ_RB_ROOT;
1060         RB_CLEAR_NODE(&cfqg->rb_node);
1061
1062         /*
1063          * Take the initial reference that will be released on destroy
1064          * This can be thought of a joint reference by cgroup and
1065          * elevator which will be dropped by either elevator exit
1066          * or cgroup deletion path depending on who is exiting first.
1067          */
1068         cfqg->ref = 1;
1069
1070         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1071         if (ret) {
1072                 kfree(cfqg);
1073                 return NULL;
1074         }
1075
1076         return cfqg;
1077 }
1078
1079 static struct cfq_group *
1080 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1081 {
1082         struct cfq_group *cfqg = NULL;
1083         void *key = cfqd;
1084         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1085         unsigned int major, minor;
1086
1087         /*
1088          * This is the common case when there are no blkio cgroups.
1089          * Avoid lookup in this case
1090          */
1091         if (blkcg == &blkio_root_cgroup)
1092                 cfqg = &cfqd->root_group;
1093         else
1094                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1095
1096         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1097                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1098                 cfqg->blkg.dev = MKDEV(major, minor);
1099         }
1100
1101         return cfqg;
1102 }
1103
1104 /*
1105  * Search for the cfq group current task belongs to. request_queue lock must
1106  * be held.
1107  */
1108 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1109 {
1110         struct blkio_cgroup *blkcg;
1111         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1112         struct request_queue *q = cfqd->queue;
1113
1114         rcu_read_lock();
1115         blkcg = task_blkio_cgroup(current);
1116         cfqg = cfq_find_cfqg(cfqd, blkcg);
1117         if (cfqg) {
1118                 rcu_read_unlock();
1119                 return cfqg;
1120         }
1121
1122         /*
1123          * Need to allocate a group. Allocation of group also needs allocation
1124          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1125          * we need to drop rcu lock and queue_lock before we call alloc.
1126          *
1127          * Not taking any queue reference here and assuming that queue is
1128          * around by the time we return. CFQ queue allocation code does
1129          * the same. It might be racy though.
1130          */
1131
1132         rcu_read_unlock();
1133         spin_unlock_irq(q->queue_lock);
1134
1135         cfqg = cfq_alloc_cfqg(cfqd);
1136
1137         spin_lock_irq(q->queue_lock);
1138
1139         rcu_read_lock();
1140         blkcg = task_blkio_cgroup(current);
1141
1142         /*
1143          * If some other thread already allocated the group while we were
1144          * not holding queue lock, free up the group
1145          */
1146         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1147
1148         if (__cfqg) {
1149                 kfree(cfqg);
1150                 rcu_read_unlock();
1151                 return __cfqg;
1152         }
1153
1154         if (!cfqg)
1155                 cfqg = &cfqd->root_group;
1156
1157         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1158         rcu_read_unlock();
1159         return cfqg;
1160 }
1161
1162 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1163 {
1164         cfqg->ref++;
1165         return cfqg;
1166 }
1167
1168 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1169 {
1170         /* Currently, all async queues are mapped to root group */
1171         if (!cfq_cfqq_sync(cfqq))
1172                 cfqg = &cfqq->cfqd->root_group;
1173
1174         cfqq->cfqg = cfqg;
1175         /* cfqq reference on cfqg */
1176         cfqq->cfqg->ref++;
1177 }
1178
1179 static void cfq_put_cfqg(struct cfq_group *cfqg)
1180 {
1181         struct cfq_rb_root *st;
1182         int i, j;
1183
1184         BUG_ON(cfqg->ref <= 0);
1185         cfqg->ref--;
1186         if (cfqg->ref)
1187                 return;
1188         for_each_cfqg_st(cfqg, i, j, st)
1189                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1190         free_percpu(cfqg->blkg.stats_cpu);
1191         kfree(cfqg);
1192 }
1193
1194 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1195 {
1196         /* Something wrong if we are trying to remove same group twice */
1197         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1198
1199         hlist_del_init(&cfqg->cfqd_node);
1200
1201         /*
1202          * Put the reference taken at the time of creation so that when all
1203          * queues are gone, group can be destroyed.
1204          */
1205         cfq_put_cfqg(cfqg);
1206 }
1207
1208 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1209 {
1210         struct hlist_node *pos, *n;
1211         struct cfq_group *cfqg;
1212
1213         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1214                 /*
1215                  * If cgroup removal path got to blk_group first and removed
1216                  * it from cgroup list, then it will take care of destroying
1217                  * cfqg also.
1218                  */
1219                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1220                         cfq_destroy_cfqg(cfqd, cfqg);
1221         }
1222 }
1223
1224 /*
1225  * Blk cgroup controller notification saying that blkio_group object is being
1226  * delinked as associated cgroup object is going away. That also means that
1227  * no new IO will come in this group. So get rid of this group as soon as
1228  * any pending IO in the group is finished.
1229  *
1230  * This function is called under rcu_read_lock(). key is the rcu protected
1231  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1232  * read lock.
1233  *
1234  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1235  * it should not be NULL as even if elevator was exiting, cgroup deltion
1236  * path got to it first.
1237  */
1238 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1239 {
1240         unsigned long  flags;
1241         struct cfq_data *cfqd = key;
1242
1243         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1244         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1245         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1246 }
1247
1248 #else /* GROUP_IOSCHED */
1249 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1250 {
1251         return &cfqd->root_group;
1252 }
1253
1254 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1255 {
1256         return cfqg;
1257 }
1258
1259 static inline void
1260 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1261         cfqq->cfqg = cfqg;
1262 }
1263
1264 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1265 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1266
1267 #endif /* GROUP_IOSCHED */
1268
1269 /*
1270  * The cfqd->service_trees holds all pending cfq_queue's that have
1271  * requests waiting to be processed. It is sorted in the order that
1272  * we will service the queues.
1273  */
1274 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1275                                  bool add_front)
1276 {
1277         struct rb_node **p, *parent;
1278         struct cfq_queue *__cfqq;
1279         unsigned long rb_key;
1280         struct cfq_rb_root *service_tree;
1281         int left;
1282         int new_cfqq = 1;
1283
1284         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1285                                                 cfqq_type(cfqq));
1286         if (cfq_class_idle(cfqq)) {
1287                 rb_key = CFQ_IDLE_DELAY;
1288                 parent = rb_last(&service_tree->rb);
1289                 if (parent && parent != &cfqq->rb_node) {
1290                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1291                         rb_key += __cfqq->rb_key;
1292                 } else
1293                         rb_key += jiffies;
1294         } else if (!add_front) {
1295                 /*
1296                  * Get our rb key offset. Subtract any residual slice
1297                  * value carried from last service. A negative resid
1298                  * count indicates slice overrun, and this should position
1299                  * the next service time further away in the tree.
1300                  */
1301                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1302                 rb_key -= cfqq->slice_resid;
1303                 cfqq->slice_resid = 0;
1304         } else {
1305                 rb_key = -HZ;
1306                 __cfqq = cfq_rb_first(service_tree);
1307                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1308         }
1309
1310         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1311                 new_cfqq = 0;
1312                 /*
1313                  * same position, nothing more to do
1314                  */
1315                 if (rb_key == cfqq->rb_key &&
1316                     cfqq->service_tree == service_tree)
1317                         return;
1318
1319                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1320                 cfqq->service_tree = NULL;
1321         }
1322
1323         left = 1;
1324         parent = NULL;
1325         cfqq->service_tree = service_tree;
1326         p = &service_tree->rb.rb_node;
1327         while (*p) {
1328                 struct rb_node **n;
1329
1330                 parent = *p;
1331                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1332
1333                 /*
1334                  * sort by key, that represents service time.
1335                  */
1336                 if (time_before(rb_key, __cfqq->rb_key))
1337                         n = &(*p)->rb_left;
1338                 else {
1339                         n = &(*p)->rb_right;
1340                         left = 0;
1341                 }
1342
1343                 p = n;
1344         }
1345
1346         if (left)
1347                 service_tree->left = &cfqq->rb_node;
1348
1349         cfqq->rb_key = rb_key;
1350         rb_link_node(&cfqq->rb_node, parent, p);
1351         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1352         service_tree->count++;
1353         if (add_front || !new_cfqq)
1354                 return;
1355         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1356 }
1357
1358 static struct cfq_queue *
1359 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1360                      sector_t sector, struct rb_node **ret_parent,
1361                      struct rb_node ***rb_link)
1362 {
1363         struct rb_node **p, *parent;
1364         struct cfq_queue *cfqq = NULL;
1365
1366         parent = NULL;
1367         p = &root->rb_node;
1368         while (*p) {
1369                 struct rb_node **n;
1370
1371                 parent = *p;
1372                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1373
1374                 /*
1375                  * Sort strictly based on sector.  Smallest to the left,
1376                  * largest to the right.
1377                  */
1378                 if (sector > blk_rq_pos(cfqq->next_rq))
1379                         n = &(*p)->rb_right;
1380                 else if (sector < blk_rq_pos(cfqq->next_rq))
1381                         n = &(*p)->rb_left;
1382                 else
1383                         break;
1384                 p = n;
1385                 cfqq = NULL;
1386         }
1387
1388         *ret_parent = parent;
1389         if (rb_link)
1390                 *rb_link = p;
1391         return cfqq;
1392 }
1393
1394 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1395 {
1396         struct rb_node **p, *parent;
1397         struct cfq_queue *__cfqq;
1398
1399         if (cfqq->p_root) {
1400                 rb_erase(&cfqq->p_node, cfqq->p_root);
1401                 cfqq->p_root = NULL;
1402         }
1403
1404         if (cfq_class_idle(cfqq))
1405                 return;
1406         if (!cfqq->next_rq)
1407                 return;
1408
1409         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1410         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1411                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1412         if (!__cfqq) {
1413                 rb_link_node(&cfqq->p_node, parent, p);
1414                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1415         } else
1416                 cfqq->p_root = NULL;
1417 }
1418
1419 /*
1420  * Update cfqq's position in the service tree.
1421  */
1422 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1423 {
1424         /*
1425          * Resorting requires the cfqq to be on the RR list already.
1426          */
1427         if (cfq_cfqq_on_rr(cfqq)) {
1428                 cfq_service_tree_add(cfqd, cfqq, 0);
1429                 cfq_prio_tree_add(cfqd, cfqq);
1430         }
1431 }
1432
1433 /*
1434  * add to busy list of queues for service, trying to be fair in ordering
1435  * the pending list according to last request service
1436  */
1437 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1438 {
1439         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1440         BUG_ON(cfq_cfqq_on_rr(cfqq));
1441         cfq_mark_cfqq_on_rr(cfqq);
1442         cfqd->busy_queues++;
1443         if (cfq_cfqq_sync(cfqq))
1444                 cfqd->busy_sync_queues++;
1445
1446         cfq_resort_rr_list(cfqd, cfqq);
1447 }
1448
1449 /*
1450  * Called when the cfqq no longer has requests pending, remove it from
1451  * the service tree.
1452  */
1453 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1454 {
1455         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1456         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1457         cfq_clear_cfqq_on_rr(cfqq);
1458
1459         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1460                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1461                 cfqq->service_tree = NULL;
1462         }
1463         if (cfqq->p_root) {
1464                 rb_erase(&cfqq->p_node, cfqq->p_root);
1465                 cfqq->p_root = NULL;
1466         }
1467
1468         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1469         BUG_ON(!cfqd->busy_queues);
1470         cfqd->busy_queues--;
1471         if (cfq_cfqq_sync(cfqq))
1472                 cfqd->busy_sync_queues--;
1473 }
1474
1475 /*
1476  * rb tree support functions
1477  */
1478 static void cfq_del_rq_rb(struct request *rq)
1479 {
1480         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1481         const int sync = rq_is_sync(rq);
1482
1483         BUG_ON(!cfqq->queued[sync]);
1484         cfqq->queued[sync]--;
1485
1486         elv_rb_del(&cfqq->sort_list, rq);
1487
1488         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1489                 /*
1490                  * Queue will be deleted from service tree when we actually
1491                  * expire it later. Right now just remove it from prio tree
1492                  * as it is empty.
1493                  */
1494                 if (cfqq->p_root) {
1495                         rb_erase(&cfqq->p_node, cfqq->p_root);
1496                         cfqq->p_root = NULL;
1497                 }
1498         }
1499 }
1500
1501 static void cfq_add_rq_rb(struct request *rq)
1502 {
1503         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1504         struct cfq_data *cfqd = cfqq->cfqd;
1505         struct request *prev;
1506
1507         cfqq->queued[rq_is_sync(rq)]++;
1508
1509         elv_rb_add(&cfqq->sort_list, rq);
1510
1511         if (!cfq_cfqq_on_rr(cfqq))
1512                 cfq_add_cfqq_rr(cfqd, cfqq);
1513
1514         /*
1515          * check if this request is a better next-serve candidate
1516          */
1517         prev = cfqq->next_rq;
1518         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1519
1520         /*
1521          * adjust priority tree position, if ->next_rq changes
1522          */
1523         if (prev != cfqq->next_rq)
1524                 cfq_prio_tree_add(cfqd, cfqq);
1525
1526         BUG_ON(!cfqq->next_rq);
1527 }
1528
1529 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1530 {
1531         elv_rb_del(&cfqq->sort_list, rq);
1532         cfqq->queued[rq_is_sync(rq)]--;
1533         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1534                                         rq_data_dir(rq), rq_is_sync(rq));
1535         cfq_add_rq_rb(rq);
1536         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1537                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1538                         rq_is_sync(rq));
1539 }
1540
1541 static struct request *
1542 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1543 {
1544         struct task_struct *tsk = current;
1545         struct cfq_io_context *cic;
1546         struct cfq_queue *cfqq;
1547
1548         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1549         if (!cic)
1550                 return NULL;
1551
1552         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1553         if (cfqq) {
1554                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1555
1556                 return elv_rb_find(&cfqq->sort_list, sector);
1557         }
1558
1559         return NULL;
1560 }
1561
1562 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1563 {
1564         struct cfq_data *cfqd = q->elevator->elevator_data;
1565
1566         cfqd->rq_in_driver++;
1567         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1568                                                 cfqd->rq_in_driver);
1569
1570         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1571 }
1572
1573 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1574 {
1575         struct cfq_data *cfqd = q->elevator->elevator_data;
1576
1577         WARN_ON(!cfqd->rq_in_driver);
1578         cfqd->rq_in_driver--;
1579         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1580                                                 cfqd->rq_in_driver);
1581 }
1582
1583 static void cfq_remove_request(struct request *rq)
1584 {
1585         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1586
1587         if (cfqq->next_rq == rq)
1588                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1589
1590         list_del_init(&rq->queuelist);
1591         cfq_del_rq_rb(rq);
1592
1593         cfqq->cfqd->rq_queued--;
1594         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1595                                         rq_data_dir(rq), rq_is_sync(rq));
1596         if (rq->cmd_flags & REQ_META) {
1597                 WARN_ON(!cfqq->meta_pending);
1598                 cfqq->meta_pending--;
1599         }
1600 }
1601
1602 static int cfq_merge(struct request_queue *q, struct request **req,
1603                      struct bio *bio)
1604 {
1605         struct cfq_data *cfqd = q->elevator->elevator_data;
1606         struct request *__rq;
1607
1608         __rq = cfq_find_rq_fmerge(cfqd, bio);
1609         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1610                 *req = __rq;
1611                 return ELEVATOR_FRONT_MERGE;
1612         }
1613
1614         return ELEVATOR_NO_MERGE;
1615 }
1616
1617 static void cfq_merged_request(struct request_queue *q, struct request *req,
1618                                int type)
1619 {
1620         if (type == ELEVATOR_FRONT_MERGE) {
1621                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1622
1623                 cfq_reposition_rq_rb(cfqq, req);
1624         }
1625 }
1626
1627 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1628                                 struct bio *bio)
1629 {
1630         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1631                                         bio_data_dir(bio), cfq_bio_sync(bio));
1632 }
1633
1634 static void
1635 cfq_merged_requests(struct request_queue *q, struct request *rq,
1636                     struct request *next)
1637 {
1638         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1639         /*
1640          * reposition in fifo if next is older than rq
1641          */
1642         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1643             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1644                 list_move(&rq->queuelist, &next->queuelist);
1645                 rq_set_fifo_time(rq, rq_fifo_time(next));
1646         }
1647
1648         if (cfqq->next_rq == next)
1649                 cfqq->next_rq = rq;
1650         cfq_remove_request(next);
1651         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1652                                         rq_data_dir(next), rq_is_sync(next));
1653 }
1654
1655 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1656                            struct bio *bio)
1657 {
1658         struct cfq_data *cfqd = q->elevator->elevator_data;
1659         struct cfq_io_context *cic;
1660         struct cfq_queue *cfqq;
1661
1662         /*
1663          * Disallow merge of a sync bio into an async request.
1664          */
1665         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1666                 return false;
1667
1668         /*
1669          * Lookup the cfqq that this bio will be queued with. Allow
1670          * merge only if rq is queued there.
1671          */
1672         cic = cfq_cic_lookup(cfqd, current->io_context);
1673         if (!cic)
1674                 return false;
1675
1676         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1677         return cfqq == RQ_CFQQ(rq);
1678 }
1679
1680 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1681 {
1682         del_timer(&cfqd->idle_slice_timer);
1683         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1684 }
1685
1686 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1687                                    struct cfq_queue *cfqq)
1688 {
1689         if (cfqq) {
1690                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1691                                 cfqd->serving_prio, cfqd->serving_type);
1692                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1693                 cfqq->slice_start = 0;
1694                 cfqq->dispatch_start = jiffies;
1695                 cfqq->allocated_slice = 0;
1696                 cfqq->slice_end = 0;
1697                 cfqq->slice_dispatch = 0;
1698                 cfqq->nr_sectors = 0;
1699
1700                 cfq_clear_cfqq_wait_request(cfqq);
1701                 cfq_clear_cfqq_must_dispatch(cfqq);
1702                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1703                 cfq_clear_cfqq_fifo_expire(cfqq);
1704                 cfq_mark_cfqq_slice_new(cfqq);
1705
1706                 cfq_del_timer(cfqd, cfqq);
1707         }
1708
1709         cfqd->active_queue = cfqq;
1710 }
1711
1712 /*
1713  * current cfqq expired its slice (or was too idle), select new one
1714  */
1715 static void
1716 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1717                     bool timed_out)
1718 {
1719         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1720
1721         if (cfq_cfqq_wait_request(cfqq))
1722                 cfq_del_timer(cfqd, cfqq);
1723
1724         cfq_clear_cfqq_wait_request(cfqq);
1725         cfq_clear_cfqq_wait_busy(cfqq);
1726
1727         /*
1728          * If this cfqq is shared between multiple processes, check to
1729          * make sure that those processes are still issuing I/Os within
1730          * the mean seek distance.  If not, it may be time to break the
1731          * queues apart again.
1732          */
1733         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1734                 cfq_mark_cfqq_split_coop(cfqq);
1735
1736         /*
1737          * store what was left of this slice, if the queue idled/timed out
1738          */
1739         if (timed_out) {
1740                 if (cfq_cfqq_slice_new(cfqq))
1741                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1742                 else
1743                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1744                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1745         }
1746
1747         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1748
1749         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1750                 cfq_del_cfqq_rr(cfqd, cfqq);
1751
1752         cfq_resort_rr_list(cfqd, cfqq);
1753
1754         if (cfqq == cfqd->active_queue)
1755                 cfqd->active_queue = NULL;
1756
1757         if (cfqd->active_cic) {
1758                 put_io_context(cfqd->active_cic->ioc);
1759                 cfqd->active_cic = NULL;
1760         }
1761 }
1762
1763 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1764 {
1765         struct cfq_queue *cfqq = cfqd->active_queue;
1766
1767         if (cfqq)
1768                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1769 }
1770
1771 /*
1772  * Get next queue for service. Unless we have a queue preemption,
1773  * we'll simply select the first cfqq in the service tree.
1774  */
1775 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1776 {
1777         struct cfq_rb_root *service_tree =
1778                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1779                                         cfqd->serving_type);
1780
1781         if (!cfqd->rq_queued)
1782                 return NULL;
1783
1784         /* There is nothing to dispatch */
1785         if (!service_tree)
1786                 return NULL;
1787         if (RB_EMPTY_ROOT(&service_tree->rb))
1788                 return NULL;
1789         return cfq_rb_first(service_tree);
1790 }
1791
1792 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1793 {
1794         struct cfq_group *cfqg;
1795         struct cfq_queue *cfqq;
1796         int i, j;
1797         struct cfq_rb_root *st;
1798
1799         if (!cfqd->rq_queued)
1800                 return NULL;
1801
1802         cfqg = cfq_get_next_cfqg(cfqd);
1803         if (!cfqg)
1804                 return NULL;
1805
1806         for_each_cfqg_st(cfqg, i, j, st)
1807                 if ((cfqq = cfq_rb_first(st)) != NULL)
1808                         return cfqq;
1809         return NULL;
1810 }
1811
1812 /*
1813  * Get and set a new active queue for service.
1814  */
1815 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1816                                               struct cfq_queue *cfqq)
1817 {
1818         if (!cfqq)
1819                 cfqq = cfq_get_next_queue(cfqd);
1820
1821         __cfq_set_active_queue(cfqd, cfqq);
1822         return cfqq;
1823 }
1824
1825 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1826                                           struct request *rq)
1827 {
1828         if (blk_rq_pos(rq) >= cfqd->last_position)
1829                 return blk_rq_pos(rq) - cfqd->last_position;
1830         else
1831                 return cfqd->last_position - blk_rq_pos(rq);
1832 }
1833
1834 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1835                                struct request *rq)
1836 {
1837         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1838 }
1839
1840 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1841                                     struct cfq_queue *cur_cfqq)
1842 {
1843         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1844         struct rb_node *parent, *node;
1845         struct cfq_queue *__cfqq;
1846         sector_t sector = cfqd->last_position;
1847
1848         if (RB_EMPTY_ROOT(root))
1849                 return NULL;
1850
1851         /*
1852          * First, if we find a request starting at the end of the last
1853          * request, choose it.
1854          */
1855         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1856         if (__cfqq)
1857                 return __cfqq;
1858
1859         /*
1860          * If the exact sector wasn't found, the parent of the NULL leaf
1861          * will contain the closest sector.
1862          */
1863         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1864         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1865                 return __cfqq;
1866
1867         if (blk_rq_pos(__cfqq->next_rq) < sector)
1868                 node = rb_next(&__cfqq->p_node);
1869         else
1870                 node = rb_prev(&__cfqq->p_node);
1871         if (!node)
1872                 return NULL;
1873
1874         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1875         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1876                 return __cfqq;
1877
1878         return NULL;
1879 }
1880
1881 /*
1882  * cfqd - obvious
1883  * cur_cfqq - passed in so that we don't decide that the current queue is
1884  *            closely cooperating with itself.
1885  *
1886  * So, basically we're assuming that that cur_cfqq has dispatched at least
1887  * one request, and that cfqd->last_position reflects a position on the disk
1888  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1889  * assumption.
1890  */
1891 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1892                                               struct cfq_queue *cur_cfqq)
1893 {
1894         struct cfq_queue *cfqq;
1895
1896         if (cfq_class_idle(cur_cfqq))
1897                 return NULL;
1898         if (!cfq_cfqq_sync(cur_cfqq))
1899                 return NULL;
1900         if (CFQQ_SEEKY(cur_cfqq))
1901                 return NULL;
1902
1903         /*
1904          * Don't search priority tree if it's the only queue in the group.
1905          */
1906         if (cur_cfqq->cfqg->nr_cfqq == 1)
1907                 return NULL;
1908
1909         /*
1910          * We should notice if some of the queues are cooperating, eg
1911          * working closely on the same area of the disk. In that case,
1912          * we can group them together and don't waste time idling.
1913          */
1914         cfqq = cfqq_close(cfqd, cur_cfqq);
1915         if (!cfqq)
1916                 return NULL;
1917
1918         /* If new queue belongs to different cfq_group, don't choose it */
1919         if (cur_cfqq->cfqg != cfqq->cfqg)
1920                 return NULL;
1921
1922         /*
1923          * It only makes sense to merge sync queues.
1924          */
1925         if (!cfq_cfqq_sync(cfqq))
1926                 return NULL;
1927         if (CFQQ_SEEKY(cfqq))
1928                 return NULL;
1929
1930         /*
1931          * Do not merge queues of different priority classes
1932          */
1933         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1934                 return NULL;
1935
1936         return cfqq;
1937 }
1938
1939 /*
1940  * Determine whether we should enforce idle window for this queue.
1941  */
1942
1943 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1944 {
1945         enum wl_prio_t prio = cfqq_prio(cfqq);
1946         struct cfq_rb_root *service_tree = cfqq->service_tree;
1947
1948         BUG_ON(!service_tree);
1949         BUG_ON(!service_tree->count);
1950
1951         if (!cfqd->cfq_slice_idle)
1952                 return false;
1953
1954         /* We never do for idle class queues. */
1955         if (prio == IDLE_WORKLOAD)
1956                 return false;
1957
1958         /* We do for queues that were marked with idle window flag. */
1959         if (cfq_cfqq_idle_window(cfqq) &&
1960            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1961                 return true;
1962
1963         /*
1964          * Otherwise, we do only if they are the last ones
1965          * in their service tree.
1966          */
1967         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1968                 return true;
1969         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1970                         service_tree->count);
1971         return false;
1972 }
1973
1974 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1975 {
1976         struct cfq_queue *cfqq = cfqd->active_queue;
1977         struct cfq_io_context *cic;
1978         unsigned long sl, group_idle = 0;
1979
1980         /*
1981          * SSD device without seek penalty, disable idling. But only do so
1982          * for devices that support queuing, otherwise we still have a problem
1983          * with sync vs async workloads.
1984          */
1985         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1986                 return;
1987
1988         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1989         WARN_ON(cfq_cfqq_slice_new(cfqq));
1990
1991         /*
1992          * idle is disabled, either manually or by past process history
1993          */
1994         if (!cfq_should_idle(cfqd, cfqq)) {
1995                 /* no queue idling. Check for group idling */
1996                 if (cfqd->cfq_group_idle)
1997                         group_idle = cfqd->cfq_group_idle;
1998                 else
1999                         return;
2000         }
2001
2002         /*
2003          * still active requests from this queue, don't idle
2004          */
2005         if (cfqq->dispatched)
2006                 return;
2007
2008         /*
2009          * task has exited, don't wait
2010          */
2011         cic = cfqd->active_cic;
2012         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2013                 return;
2014
2015         /*
2016          * If our average think time is larger than the remaining time
2017          * slice, then don't idle. This avoids overrunning the allotted
2018          * time slice.
2019          */
2020         if (sample_valid(cic->ttime_samples) &&
2021             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
2022                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2023                              cic->ttime_mean);
2024                 return;
2025         }
2026
2027         /* There are other queues in the group, don't do group idle */
2028         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2029                 return;
2030
2031         cfq_mark_cfqq_wait_request(cfqq);
2032
2033         if (group_idle)
2034                 sl = cfqd->cfq_group_idle;
2035         else
2036                 sl = cfqd->cfq_slice_idle;
2037
2038         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2039         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2040         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2041                         group_idle ? 1 : 0);
2042 }
2043
2044 /*
2045  * Move request from internal lists to the request queue dispatch list.
2046  */
2047 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2048 {
2049         struct cfq_data *cfqd = q->elevator->elevator_data;
2050         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2051
2052         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2053
2054         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2055         cfq_remove_request(rq);
2056         cfqq->dispatched++;
2057         (RQ_CFQG(rq))->dispatched++;
2058         elv_dispatch_sort(q, rq);
2059
2060         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2061         cfqq->nr_sectors += blk_rq_sectors(rq);
2062         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2063                                         rq_data_dir(rq), rq_is_sync(rq));
2064 }
2065
2066 /*
2067  * return expired entry, or NULL to just start from scratch in rbtree
2068  */
2069 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2070 {
2071         struct request *rq = NULL;
2072
2073         if (cfq_cfqq_fifo_expire(cfqq))
2074                 return NULL;
2075
2076         cfq_mark_cfqq_fifo_expire(cfqq);
2077
2078         if (list_empty(&cfqq->fifo))
2079                 return NULL;
2080
2081         rq = rq_entry_fifo(cfqq->fifo.next);
2082         if (time_before(jiffies, rq_fifo_time(rq)))
2083                 rq = NULL;
2084
2085         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2086         return rq;
2087 }
2088
2089 static inline int
2090 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2091 {
2092         const int base_rq = cfqd->cfq_slice_async_rq;
2093
2094         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2095
2096         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2097 }
2098
2099 /*
2100  * Must be called with the queue_lock held.
2101  */
2102 static int cfqq_process_refs(struct cfq_queue *cfqq)
2103 {
2104         int process_refs, io_refs;
2105
2106         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2107         process_refs = cfqq->ref - io_refs;
2108         BUG_ON(process_refs < 0);
2109         return process_refs;
2110 }
2111
2112 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2113 {
2114         int process_refs, new_process_refs;
2115         struct cfq_queue *__cfqq;
2116
2117         /*
2118          * If there are no process references on the new_cfqq, then it is
2119          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2120          * chain may have dropped their last reference (not just their
2121          * last process reference).
2122          */
2123         if (!cfqq_process_refs(new_cfqq))
2124                 return;
2125
2126         /* Avoid a circular list and skip interim queue merges */
2127         while ((__cfqq = new_cfqq->new_cfqq)) {
2128                 if (__cfqq == cfqq)
2129                         return;
2130                 new_cfqq = __cfqq;
2131         }
2132
2133         process_refs = cfqq_process_refs(cfqq);
2134         new_process_refs = cfqq_process_refs(new_cfqq);
2135         /*
2136          * If the process for the cfqq has gone away, there is no
2137          * sense in merging the queues.
2138          */
2139         if (process_refs == 0 || new_process_refs == 0)
2140                 return;
2141
2142         /*
2143          * Merge in the direction of the lesser amount of work.
2144          */
2145         if (new_process_refs >= process_refs) {
2146                 cfqq->new_cfqq = new_cfqq;
2147                 new_cfqq->ref += process_refs;
2148         } else {
2149                 new_cfqq->new_cfqq = cfqq;
2150                 cfqq->ref += new_process_refs;
2151         }
2152 }
2153
2154 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2155                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2156 {
2157         struct cfq_queue *queue;
2158         int i;
2159         bool key_valid = false;
2160         unsigned long lowest_key = 0;
2161         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2162
2163         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2164                 /* select the one with lowest rb_key */
2165                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2166                 if (queue &&
2167                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2168                         lowest_key = queue->rb_key;
2169                         cur_best = i;
2170                         key_valid = true;
2171                 }
2172         }
2173
2174         return cur_best;
2175 }
2176
2177 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2178 {
2179         unsigned slice;
2180         unsigned count;
2181         struct cfq_rb_root *st;
2182         unsigned group_slice;
2183         enum wl_prio_t original_prio = cfqd->serving_prio;
2184
2185         /* Choose next priority. RT > BE > IDLE */
2186         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2187                 cfqd->serving_prio = RT_WORKLOAD;
2188         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2189                 cfqd->serving_prio = BE_WORKLOAD;
2190         else {
2191                 cfqd->serving_prio = IDLE_WORKLOAD;
2192                 cfqd->workload_expires = jiffies + 1;
2193                 return;
2194         }
2195
2196         if (original_prio != cfqd->serving_prio)
2197                 goto new_workload;
2198
2199         /*
2200          * For RT and BE, we have to choose also the type
2201          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2202          * expiration time
2203          */
2204         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2205         count = st->count;
2206
2207         /*
2208          * check workload expiration, and that we still have other queues ready
2209          */
2210         if (count && !time_after(jiffies, cfqd->workload_expires))
2211                 return;
2212
2213 new_workload:
2214         /* otherwise select new workload type */
2215         cfqd->serving_type =
2216                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2217         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2218         count = st->count;
2219
2220         /*
2221          * the workload slice is computed as a fraction of target latency
2222          * proportional to the number of queues in that workload, over
2223          * all the queues in the same priority class
2224          */
2225         group_slice = cfq_group_slice(cfqd, cfqg);
2226
2227         slice = group_slice * count /
2228                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2229                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2230
2231         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2232                 unsigned int tmp;
2233
2234                 /*
2235                  * Async queues are currently system wide. Just taking
2236                  * proportion of queues with-in same group will lead to higher
2237                  * async ratio system wide as generally root group is going
2238                  * to have higher weight. A more accurate thing would be to
2239                  * calculate system wide asnc/sync ratio.
2240                  */
2241                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2242                 tmp = tmp/cfqd->busy_queues;
2243                 slice = min_t(unsigned, slice, tmp);
2244
2245                 /* async workload slice is scaled down according to
2246                  * the sync/async slice ratio. */
2247                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2248         } else
2249                 /* sync workload slice is at least 2 * cfq_slice_idle */
2250                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2251
2252         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2253         cfq_log(cfqd, "workload slice:%d", slice);
2254         cfqd->workload_expires = jiffies + slice;
2255 }
2256
2257 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2258 {
2259         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2260         struct cfq_group *cfqg;
2261
2262         if (RB_EMPTY_ROOT(&st->rb))
2263                 return NULL;
2264         cfqg = cfq_rb_first_group(st);
2265         update_min_vdisktime(st);
2266         return cfqg;
2267 }
2268
2269 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2270 {
2271         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2272
2273         cfqd->serving_group = cfqg;
2274
2275         /* Restore the workload type data */
2276         if (cfqg->saved_workload_slice) {
2277                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2278                 cfqd->serving_type = cfqg->saved_workload;
2279                 cfqd->serving_prio = cfqg->saved_serving_prio;
2280         } else
2281                 cfqd->workload_expires = jiffies - 1;
2282
2283         choose_service_tree(cfqd, cfqg);
2284 }
2285
2286 /*
2287  * Select a queue for service. If we have a current active queue,
2288  * check whether to continue servicing it, or retrieve and set a new one.
2289  */
2290 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2291 {
2292         struct cfq_queue *cfqq, *new_cfqq = NULL;
2293
2294         cfqq = cfqd->active_queue;
2295         if (!cfqq)
2296                 goto new_queue;
2297
2298         if (!cfqd->rq_queued)
2299                 return NULL;
2300
2301         /*
2302          * We were waiting for group to get backlogged. Expire the queue
2303          */
2304         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2305                 goto expire;
2306
2307         /*
2308          * The active queue has run out of time, expire it and select new.
2309          */
2310         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2311                 /*
2312                  * If slice had not expired at the completion of last request
2313                  * we might not have turned on wait_busy flag. Don't expire
2314                  * the queue yet. Allow the group to get backlogged.
2315                  *
2316                  * The very fact that we have used the slice, that means we
2317                  * have been idling all along on this queue and it should be
2318                  * ok to wait for this request to complete.
2319                  */
2320                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2321                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2322                         cfqq = NULL;
2323                         goto keep_queue;
2324                 } else
2325                         goto check_group_idle;
2326         }
2327
2328         /*
2329          * The active queue has requests and isn't expired, allow it to
2330          * dispatch.
2331          */
2332         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2333                 goto keep_queue;
2334
2335         /*
2336          * If another queue has a request waiting within our mean seek
2337          * distance, let it run.  The expire code will check for close
2338          * cooperators and put the close queue at the front of the service
2339          * tree.  If possible, merge the expiring queue with the new cfqq.
2340          */
2341         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2342         if (new_cfqq) {
2343                 if (!cfqq->new_cfqq)
2344                         cfq_setup_merge(cfqq, new_cfqq);
2345                 goto expire;
2346         }
2347
2348         /*
2349          * No requests pending. If the active queue still has requests in
2350          * flight or is idling for a new request, allow either of these
2351          * conditions to happen (or time out) before selecting a new queue.
2352          */
2353         if (timer_pending(&cfqd->idle_slice_timer)) {
2354                 cfqq = NULL;
2355                 goto keep_queue;
2356         }
2357
2358         /*
2359          * This is a deep seek queue, but the device is much faster than
2360          * the queue can deliver, don't idle
2361          **/
2362         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2363             (cfq_cfqq_slice_new(cfqq) ||
2364             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2365                 cfq_clear_cfqq_deep(cfqq);
2366                 cfq_clear_cfqq_idle_window(cfqq);
2367         }
2368
2369         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2370                 cfqq = NULL;
2371                 goto keep_queue;
2372         }
2373
2374         /*
2375          * If group idle is enabled and there are requests dispatched from
2376          * this group, wait for requests to complete.
2377          */
2378 check_group_idle:
2379         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2380             && cfqq->cfqg->dispatched) {
2381                 cfqq = NULL;
2382                 goto keep_queue;
2383         }
2384
2385 expire:
2386         cfq_slice_expired(cfqd, 0);
2387 new_queue:
2388         /*
2389          * Current queue expired. Check if we have to switch to a new
2390          * service tree
2391          */
2392         if (!new_cfqq)
2393                 cfq_choose_cfqg(cfqd);
2394
2395         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2396 keep_queue:
2397         return cfqq;
2398 }
2399
2400 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2401 {
2402         int dispatched = 0;
2403
2404         while (cfqq->next_rq) {
2405                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2406                 dispatched++;
2407         }
2408
2409         BUG_ON(!list_empty(&cfqq->fifo));
2410
2411         /* By default cfqq is not expired if it is empty. Do it explicitly */
2412         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2413         return dispatched;
2414 }
2415
2416 /*
2417  * Drain our current requests. Used for barriers and when switching
2418  * io schedulers on-the-fly.
2419  */
2420 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2421 {
2422         struct cfq_queue *cfqq;
2423         int dispatched = 0;
2424
2425         /* Expire the timeslice of the current active queue first */
2426         cfq_slice_expired(cfqd, 0);
2427         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2428                 __cfq_set_active_queue(cfqd, cfqq);
2429                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2430         }
2431
2432         BUG_ON(cfqd->busy_queues);
2433
2434         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2435         return dispatched;
2436 }
2437
2438 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2439         struct cfq_queue *cfqq)
2440 {
2441         /* the queue hasn't finished any request, can't estimate */
2442         if (cfq_cfqq_slice_new(cfqq))
2443                 return true;
2444         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2445                 cfqq->slice_end))
2446                 return true;
2447
2448         return false;
2449 }
2450
2451 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2452 {
2453         unsigned int max_dispatch;
2454
2455         /*
2456          * Drain async requests before we start sync IO
2457          */
2458         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2459                 return false;
2460
2461         /*
2462          * If this is an async queue and we have sync IO in flight, let it wait
2463          */
2464         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2465                 return false;
2466
2467         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2468         if (cfq_class_idle(cfqq))
2469                 max_dispatch = 1;
2470
2471         /*
2472          * Does this cfqq already have too much IO in flight?
2473          */
2474         if (cfqq->dispatched >= max_dispatch) {
2475                 bool promote_sync = false;
2476                 /*
2477                  * idle queue must always only have a single IO in flight
2478                  */
2479                 if (cfq_class_idle(cfqq))
2480                         return false;
2481
2482                 /*
2483                  * If there is only one sync queue
2484                  * we can ignore async queue here and give the sync
2485                  * queue no dispatch limit. The reason is a sync queue can
2486                  * preempt async queue, limiting the sync queue doesn't make
2487                  * sense. This is useful for aiostress test.
2488                  */
2489                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2490                         promote_sync = true;
2491
2492                 /*
2493                  * We have other queues, don't allow more IO from this one
2494                  */
2495                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2496                                 !promote_sync)
2497                         return false;
2498
2499                 /*
2500                  * Sole queue user, no limit
2501                  */
2502                 if (cfqd->busy_queues == 1 || promote_sync)
2503                         max_dispatch = -1;
2504                 else
2505                         /*
2506                          * Normally we start throttling cfqq when cfq_quantum/2
2507                          * requests have been dispatched. But we can drive
2508                          * deeper queue depths at the beginning of slice
2509                          * subjected to upper limit of cfq_quantum.
2510                          * */
2511                         max_dispatch = cfqd->cfq_quantum;
2512         }
2513
2514         /*
2515          * Async queues must wait a bit before being allowed dispatch.
2516          * We also ramp up the dispatch depth gradually for async IO,
2517          * based on the last sync IO we serviced
2518          */
2519         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2520                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2521                 unsigned int depth;
2522
2523                 depth = last_sync / cfqd->cfq_slice[1];
2524                 if (!depth && !cfqq->dispatched)
2525                         depth = 1;
2526                 if (depth < max_dispatch)
2527                         max_dispatch = depth;
2528         }
2529
2530         /*
2531          * If we're below the current max, allow a dispatch
2532          */
2533         return cfqq->dispatched < max_dispatch;
2534 }
2535
2536 /*
2537  * Dispatch a request from cfqq, moving them to the request queue
2538  * dispatch list.
2539  */
2540 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2541 {
2542         struct request *rq;
2543
2544         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2545
2546         if (!cfq_may_dispatch(cfqd, cfqq))
2547                 return false;
2548
2549         /*
2550          * follow expired path, else get first next available
2551          */
2552         rq = cfq_check_fifo(cfqq);
2553         if (!rq)
2554                 rq = cfqq->next_rq;
2555
2556         /*
2557          * insert request into driver dispatch list
2558          */
2559         cfq_dispatch_insert(cfqd->queue, rq);
2560
2561         if (!cfqd->active_cic) {
2562                 struct cfq_io_context *cic = RQ_CIC(rq);
2563
2564                 atomic_long_inc(&cic->ioc->refcount);
2565                 cfqd->active_cic = cic;
2566         }
2567
2568         return true;
2569 }
2570
2571 /*
2572  * Find the cfqq that we need to service and move a request from that to the
2573  * dispatch list
2574  */
2575 static int cfq_dispatch_requests(struct request_queue *q, int force)
2576 {
2577         struct cfq_data *cfqd = q->elevator->elevator_data;
2578         struct cfq_queue *cfqq;
2579
2580         if (!cfqd->busy_queues)
2581                 return 0;
2582
2583         if (unlikely(force))
2584                 return cfq_forced_dispatch(cfqd);
2585
2586         cfqq = cfq_select_queue(cfqd);
2587         if (!cfqq)
2588                 return 0;
2589
2590         /*
2591          * Dispatch a request from this cfqq, if it is allowed
2592          */
2593         if (!cfq_dispatch_request(cfqd, cfqq))
2594                 return 0;
2595
2596         cfqq->slice_dispatch++;
2597         cfq_clear_cfqq_must_dispatch(cfqq);
2598
2599         /*
2600          * expire an async queue immediately if it has used up its slice. idle
2601          * queue always expire after 1 dispatch round.
2602          */
2603         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2604             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2605             cfq_class_idle(cfqq))) {
2606                 cfqq->slice_end = jiffies + 1;
2607                 cfq_slice_expired(cfqd, 0);
2608         }
2609
2610         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2611         return 1;
2612 }
2613
2614 /*
2615  * task holds one reference to the queue, dropped when task exits. each rq
2616  * in-flight on this queue also holds a reference, dropped when rq is freed.
2617  *
2618  * Each cfq queue took a reference on the parent group. Drop it now.
2619  * queue lock must be held here.
2620  */
2621 static void cfq_put_queue(struct cfq_queue *cfqq)
2622 {
2623         struct cfq_data *cfqd = cfqq->cfqd;
2624         struct cfq_group *cfqg;
2625
2626         BUG_ON(cfqq->ref <= 0);
2627
2628         cfqq->ref--;
2629         if (cfqq->ref)
2630                 return;
2631
2632         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2633         BUG_ON(rb_first(&cfqq->sort_list));
2634         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2635         cfqg = cfqq->cfqg;
2636
2637         if (unlikely(cfqd->active_queue == cfqq)) {
2638                 __cfq_slice_expired(cfqd, cfqq, 0);
2639                 cfq_schedule_dispatch(cfqd);
2640         }
2641
2642         BUG_ON(cfq_cfqq_on_rr(cfqq));
2643         kmem_cache_free(cfq_pool, cfqq);
2644         cfq_put_cfqg(cfqg);
2645 }
2646
2647 /*
2648  * Call func for each cic attached to this ioc.
2649  */
2650 static void
2651 call_for_each_cic(struct io_context *ioc,
2652                   void (*func)(struct io_context *, struct cfq_io_context *))
2653 {
2654         struct cfq_io_context *cic;
2655         struct hlist_node *n;
2656
2657         rcu_read_lock();
2658
2659         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2660                 func(ioc, cic);
2661
2662         rcu_read_unlock();
2663 }
2664
2665 static void cfq_cic_free_rcu(struct rcu_head *head)
2666 {
2667         struct cfq_io_context *cic;
2668
2669         cic = container_of(head, struct cfq_io_context, rcu_head);
2670
2671         kmem_cache_free(cfq_ioc_pool, cic);
2672         elv_ioc_count_dec(cfq_ioc_count);
2673
2674         if (ioc_gone) {
2675                 /*
2676                  * CFQ scheduler is exiting, grab exit lock and check
2677                  * the pending io context count. If it hits zero,
2678                  * complete ioc_gone and set it back to NULL
2679                  */
2680                 spin_lock(&ioc_gone_lock);
2681                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2682                         complete(ioc_gone);
2683                         ioc_gone = NULL;
2684                 }
2685                 spin_unlock(&ioc_gone_lock);
2686         }
2687 }
2688
2689 static void cfq_cic_free(struct cfq_io_context *cic)
2690 {
2691         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2692 }
2693
2694 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2695 {
2696         unsigned long flags;
2697         unsigned long dead_key = (unsigned long) cic->key;
2698
2699         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2700
2701         spin_lock_irqsave(&ioc->lock, flags);
2702         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2703         hlist_del_rcu(&cic->cic_list);
2704         spin_unlock_irqrestore(&ioc->lock, flags);
2705
2706         cfq_cic_free(cic);
2707 }
2708
2709 /*
2710  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2711  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2712  * and ->trim() which is called with the task lock held
2713  */
2714 static void cfq_free_io_context(struct io_context *ioc)
2715 {
2716         /*
2717          * ioc->refcount is zero here, or we are called from elv_unregister(),
2718          * so no more cic's are allowed to be linked into this ioc.  So it
2719          * should be ok to iterate over the known list, we will see all cic's
2720          * since no new ones are added.
2721          */
2722         call_for_each_cic(ioc, cic_free_func);
2723 }
2724
2725 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2726 {
2727         struct cfq_queue *__cfqq, *next;
2728
2729         /*
2730          * If this queue was scheduled to merge with another queue, be
2731          * sure to drop the reference taken on that queue (and others in
2732          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2733          */
2734         __cfqq = cfqq->new_cfqq;
2735         while (__cfqq) {
2736                 if (__cfqq == cfqq) {
2737                         WARN(1, "cfqq->new_cfqq loop detected\n");
2738                         break;
2739                 }
2740                 next = __cfqq->new_cfqq;
2741                 cfq_put_queue(__cfqq);
2742                 __cfqq = next;
2743         }
2744 }
2745
2746 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2747 {
2748         if (unlikely(cfqq == cfqd->active_queue)) {
2749                 __cfq_slice_expired(cfqd, cfqq, 0);
2750                 cfq_schedule_dispatch(cfqd);
2751         }
2752
2753         cfq_put_cooperator(cfqq);
2754
2755         cfq_put_queue(cfqq);
2756 }
2757
2758 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2759                                          struct cfq_io_context *cic)
2760 {
2761         struct io_context *ioc = cic->ioc;
2762
2763         list_del_init(&cic->queue_list);
2764
2765         /*
2766          * Make sure dead mark is seen for dead queues
2767          */
2768         smp_wmb();
2769         cic->key = cfqd_dead_key(cfqd);
2770
2771         rcu_read_lock();
2772         if (rcu_dereference(ioc->ioc_data) == cic) {
2773                 rcu_read_unlock();
2774                 spin_lock(&ioc->lock);
2775                 rcu_assign_pointer(ioc->ioc_data, NULL);
2776                 spin_unlock(&ioc->lock);
2777         } else
2778                 rcu_read_unlock();
2779
2780         if (cic->cfqq[BLK_RW_ASYNC]) {
2781                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2782                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2783         }
2784
2785         if (cic->cfqq[BLK_RW_SYNC]) {
2786                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2787                 cic->cfqq[BLK_RW_SYNC] = NULL;
2788         }
2789 }
2790
2791 static void cfq_exit_single_io_context(struct io_context *ioc,
2792                                        struct cfq_io_context *cic)
2793 {
2794         struct cfq_data *cfqd = cic_to_cfqd(cic);
2795
2796         if (cfqd) {
2797                 struct request_queue *q = cfqd->queue;
2798                 unsigned long flags;
2799
2800                 spin_lock_irqsave(q->queue_lock, flags);
2801
2802                 /*
2803                  * Ensure we get a fresh copy of the ->key to prevent
2804                  * race between exiting task and queue
2805                  */
2806                 smp_read_barrier_depends();
2807                 if (cic->key == cfqd)
2808                         __cfq_exit_single_io_context(cfqd, cic);
2809
2810                 spin_unlock_irqrestore(q->queue_lock, flags);
2811         }
2812 }
2813
2814 /*
2815  * The process that ioc belongs to has exited, we need to clean up
2816  * and put the internal structures we have that belongs to that process.
2817  */
2818 static void cfq_exit_io_context(struct io_context *ioc)
2819 {
2820         call_for_each_cic(ioc, cfq_exit_single_io_context);
2821 }
2822
2823 static struct cfq_io_context *
2824 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2825 {
2826         struct cfq_io_context *cic;
2827
2828         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2829                                                         cfqd->queue->node);
2830         if (cic) {
2831                 cic->last_end_request = jiffies;
2832                 INIT_LIST_HEAD(&cic->queue_list);
2833                 INIT_HLIST_NODE(&cic->cic_list);
2834                 cic->dtor = cfq_free_io_context;
2835                 cic->exit = cfq_exit_io_context;
2836                 elv_ioc_count_inc(cfq_ioc_count);
2837         }
2838
2839         return cic;
2840 }
2841
2842 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2843 {
2844         struct task_struct *tsk = current;
2845         int ioprio_class;
2846
2847         if (!cfq_cfqq_prio_changed(cfqq))
2848                 return;
2849
2850         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2851         switch (ioprio_class) {
2852         default:
2853                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2854         case IOPRIO_CLASS_NONE:
2855                 /*
2856                  * no prio set, inherit CPU scheduling settings
2857                  */
2858                 cfqq->ioprio = task_nice_ioprio(tsk);
2859                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2860                 break;
2861         case IOPRIO_CLASS_RT:
2862                 cfqq->ioprio = task_ioprio(ioc);
2863                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2864                 break;
2865         case IOPRIO_CLASS_BE:
2866                 cfqq->ioprio = task_ioprio(ioc);
2867                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2868                 break;
2869         case IOPRIO_CLASS_IDLE:
2870                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2871                 cfqq->ioprio = 7;
2872                 cfq_clear_cfqq_idle_window(cfqq);
2873                 break;
2874         }
2875
2876         /*
2877          * keep track of original prio settings in case we have to temporarily
2878          * elevate the priority of this queue
2879          */
2880         cfqq->org_ioprio = cfqq->ioprio;
2881         cfqq->org_ioprio_class = cfqq->ioprio_class;
2882         cfq_clear_cfqq_prio_changed(cfqq);
2883 }
2884
2885 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2886 {
2887         struct cfq_data *cfqd = cic_to_cfqd(cic);
2888         struct cfq_queue *cfqq;
2889         unsigned long flags;
2890
2891         if (unlikely(!cfqd))
2892                 return;
2893
2894         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2895
2896         cfqq = cic->cfqq[BLK_RW_ASYNC];
2897         if (cfqq) {
2898                 struct cfq_queue *new_cfqq;
2899                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2900                                                 GFP_ATOMIC);
2901                 if (new_cfqq) {
2902                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2903                         cfq_put_queue(cfqq);
2904                 }
2905         }
2906
2907         cfqq = cic->cfqq[BLK_RW_SYNC];
2908         if (cfqq)
2909                 cfq_mark_cfqq_prio_changed(cfqq);
2910
2911         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2912 }
2913
2914 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2915 {
2916         call_for_each_cic(ioc, changed_ioprio);
2917         ioc->ioprio_changed = 0;
2918 }
2919
2920 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2921                           pid_t pid, bool is_sync)
2922 {
2923         RB_CLEAR_NODE(&cfqq->rb_node);
2924         RB_CLEAR_NODE(&cfqq->p_node);
2925         INIT_LIST_HEAD(&cfqq->fifo);
2926
2927         cfqq->ref = 0;
2928         cfqq->cfqd = cfqd;
2929
2930         cfq_mark_cfqq_prio_changed(cfqq);
2931
2932         if (is_sync) {
2933                 if (!cfq_class_idle(cfqq))
2934                         cfq_mark_cfqq_idle_window(cfqq);
2935                 cfq_mark_cfqq_sync(cfqq);
2936         }
2937         cfqq->pid = pid;
2938 }
2939
2940 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2941 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2942 {
2943         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2944         struct cfq_data *cfqd = cic_to_cfqd(cic);
2945         unsigned long flags;
2946         struct request_queue *q;
2947
2948         if (unlikely(!cfqd))
2949                 return;
2950
2951         q = cfqd->queue;
2952
2953         spin_lock_irqsave(q->queue_lock, flags);
2954
2955         if (sync_cfqq) {
2956                 /*
2957                  * Drop reference to sync queue. A new sync queue will be
2958                  * assigned in new group upon arrival of a fresh request.
2959                  */
2960                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2961                 cic_set_cfqq(cic, NULL, 1);
2962                 cfq_put_queue(sync_cfqq);
2963         }
2964
2965         spin_unlock_irqrestore(q->queue_lock, flags);
2966 }
2967
2968 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2969 {
2970         call_for_each_cic(ioc, changed_cgroup);
2971         ioc->cgroup_changed = 0;
2972 }
2973 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2974
2975 static struct cfq_queue *
2976 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2977                      struct io_context *ioc, gfp_t gfp_mask)
2978 {
2979         struct cfq_queue *cfqq, *new_cfqq = NULL;
2980         struct cfq_io_context *cic;
2981         struct cfq_group *cfqg;
2982
2983 retry:
2984         cfqg = cfq_get_cfqg(cfqd);
2985         cic = cfq_cic_lookup(cfqd, ioc);
2986         /* cic always exists here */
2987         cfqq = cic_to_cfqq(cic, is_sync);
2988
2989         /*
2990          * Always try a new alloc if we fell back to the OOM cfqq
2991          * originally, since it should just be a temporary situation.
2992          */
2993         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2994                 cfqq = NULL;
2995                 if (new_cfqq) {
2996                         cfqq = new_cfqq;
2997                         new_cfqq = NULL;
2998                 } else if (gfp_mask & __GFP_WAIT) {
2999                         spin_unlock_irq(cfqd->queue->queue_lock);
3000                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3001                                         gfp_mask | __GFP_ZERO,
3002                                         cfqd->queue->node);
3003                         spin_lock_irq(cfqd->queue->queue_lock);
3004                         if (new_cfqq)
3005                                 goto retry;
3006                 } else {
3007                         cfqq = kmem_cache_alloc_node(cfq_pool,
3008                                         gfp_mask | __GFP_ZERO,
3009                                         cfqd->queue->node);
3010                 }
3011
3012                 if (cfqq) {
3013                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3014                         cfq_init_prio_data(cfqq, ioc);
3015                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3016                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3017                 } else
3018                         cfqq = &cfqd->oom_cfqq;
3019         }
3020
3021         if (new_cfqq)
3022                 kmem_cache_free(cfq_pool, new_cfqq);
3023
3024         return cfqq;
3025 }
3026
3027 static struct cfq_queue **
3028 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3029 {
3030         switch (ioprio_class) {
3031         case IOPRIO_CLASS_RT:
3032                 return &cfqd->async_cfqq[0][ioprio];
3033         case IOPRIO_CLASS_BE:
3034                 return &cfqd->async_cfqq[1][ioprio];
3035         case IOPRIO_CLASS_IDLE:
3036                 return &cfqd->async_idle_cfqq;
3037         default:
3038                 BUG();
3039         }
3040 }
3041
3042 static struct cfq_queue *
3043 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3044               gfp_t gfp_mask)
3045 {
3046         const int ioprio = task_ioprio(ioc);
3047         const int ioprio_class = task_ioprio_class(ioc);
3048         struct cfq_queue **async_cfqq = NULL;
3049         struct cfq_queue *cfqq = NULL;
3050
3051         if (!is_sync) {
3052                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3053                 cfqq = *async_cfqq;
3054         }
3055
3056         if (!cfqq)
3057                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3058
3059         /*
3060          * pin the queue now that it's allocated, scheduler exit will prune it
3061          */
3062         if (!is_sync && !(*async_cfqq)) {
3063                 cfqq->ref++;
3064                 *async_cfqq = cfqq;
3065         }
3066
3067         cfqq->ref++;
3068         return cfqq;
3069 }
3070
3071 /*
3072  * We drop cfq io contexts lazily, so we may find a dead one.
3073  */
3074 static void
3075 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3076                   struct cfq_io_context *cic)
3077 {
3078         unsigned long flags;
3079
3080         WARN_ON(!list_empty(&cic->queue_list));
3081         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3082
3083         spin_lock_irqsave(&ioc->lock, flags);
3084
3085         BUG_ON(rcu_dereference_check(ioc->ioc_data,
3086                 lockdep_is_held(&ioc->lock)) == cic);
3087
3088         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3089         hlist_del_rcu(&cic->cic_list);
3090         spin_unlock_irqrestore(&ioc->lock, flags);
3091
3092         cfq_cic_free(cic);
3093 }
3094
3095 static struct cfq_io_context *
3096 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3097 {
3098         struct cfq_io_context *cic;
3099         unsigned long flags;
3100
3101         if (unlikely(!ioc))
3102                 return NULL;
3103
3104         rcu_read_lock();
3105
3106         /*
3107          * we maintain a last-hit cache, to avoid browsing over the tree
3108          */
3109         cic = rcu_dereference(ioc->ioc_data);
3110         if (cic && cic->key == cfqd) {
3111                 rcu_read_unlock();
3112                 return cic;
3113         }
3114
3115         do {
3116                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3117                 rcu_read_unlock();
3118                 if (!cic)
3119                         break;
3120                 if (unlikely(cic->key != cfqd)) {
3121                         cfq_drop_dead_cic(cfqd, ioc, cic);
3122                         rcu_read_lock();
3123                         continue;
3124                 }
3125
3126                 spin_lock_irqsave(&ioc->lock, flags);
3127                 rcu_assign_pointer(ioc->ioc_data, cic);
3128                 spin_unlock_irqrestore(&ioc->lock, flags);
3129                 break;
3130         } while (1);
3131
3132         return cic;
3133 }
3134
3135 /*
3136  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3137  * the process specific cfq io context when entered from the block layer.
3138  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3139  */
3140 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3141                         struct cfq_io_context *cic, gfp_t gfp_mask)
3142 {
3143         unsigned long flags;
3144         int ret;
3145
3146         ret = radix_tree_preload(gfp_mask);
3147         if (!ret) {
3148                 cic->ioc = ioc;
3149                 cic->key = cfqd;
3150
3151                 spin_lock_irqsave(&ioc->lock, flags);
3152                 ret = radix_tree_insert(&ioc->radix_root,
3153                                                 cfqd->cic_index, cic);
3154                 if (!ret)
3155                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3156                 spin_unlock_irqrestore(&ioc->lock, flags);
3157
3158                 radix_tree_preload_end();
3159
3160                 if (!ret) {
3161                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3162                         list_add(&cic->queue_list, &cfqd->cic_list);
3163                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3164                 }
3165         }
3166
3167         if (ret)
3168                 printk(KERN_ERR "cfq: cic link failed!\n");
3169
3170         return ret;
3171 }
3172
3173 /*
3174  * Setup general io context and cfq io context. There can be several cfq
3175  * io contexts per general io context, if this process is doing io to more
3176  * than one device managed by cfq.
3177  */
3178 static struct cfq_io_context *
3179 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3180 {
3181         struct io_context *ioc = NULL;
3182         struct cfq_io_context *cic;
3183
3184         might_sleep_if(gfp_mask & __GFP_WAIT);
3185
3186         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3187         if (!ioc)
3188                 return NULL;
3189
3190         cic = cfq_cic_lookup(cfqd, ioc);
3191         if (cic)
3192                 goto out;
3193
3194         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3195         if (cic == NULL)
3196                 goto err;
3197
3198         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3199                 goto err_free;
3200
3201 out:
3202         smp_read_barrier_depends();
3203         if (unlikely(ioc->ioprio_changed))
3204                 cfq_ioc_set_ioprio(ioc);
3205
3206 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3207         if (unlikely(ioc->cgroup_changed))
3208                 cfq_ioc_set_cgroup(ioc);
3209 #endif
3210         return cic;
3211 err_free:
3212         cfq_cic_free(cic);
3213 err:
3214         put_io_context(ioc);
3215         return NULL;
3216 }
3217
3218 static void
3219 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3220 {
3221         unsigned long elapsed = jiffies - cic->last_end_request;
3222         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3223
3224         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3225         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3226         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3227 }
3228
3229 static void
3230 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3231                        struct request *rq)
3232 {
3233         sector_t sdist = 0;
3234         sector_t n_sec = blk_rq_sectors(rq);
3235         if (cfqq->last_request_pos) {
3236                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3237                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3238                 else
3239                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3240         }
3241
3242         cfqq->seek_history <<= 1;
3243         if (blk_queue_nonrot(cfqd->queue))
3244                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3245         else
3246                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3247 }
3248
3249 /*
3250  * Disable idle window if the process thinks too long or seeks so much that
3251  * it doesn't matter
3252  */
3253 static void
3254 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3255                        struct cfq_io_context *cic)
3256 {
3257         int old_idle, enable_idle;
3258
3259         /*
3260          * Don't idle for async or idle io prio class
3261          */
3262         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3263                 return;
3264
3265         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3266
3267         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3268                 cfq_mark_cfqq_deep(cfqq);
3269
3270         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3271                 enable_idle = 0;
3272         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3273             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3274                 enable_idle = 0;
3275         else if (sample_valid(cic->ttime_samples)) {
3276                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3277                         enable_idle = 0;
3278                 else
3279                         enable_idle = 1;
3280         }
3281
3282         if (old_idle != enable_idle) {
3283                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3284                 if (enable_idle)
3285                         cfq_mark_cfqq_idle_window(cfqq);
3286                 else
3287                         cfq_clear_cfqq_idle_window(cfqq);
3288         }
3289 }
3290
3291 /*
3292  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3293  * no or if we aren't sure, a 1 will cause a preempt.
3294  */
3295 static bool
3296 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3297                    struct request *rq)
3298 {
3299         struct cfq_queue *cfqq;
3300
3301         cfqq = cfqd->active_queue;
3302         if (!cfqq)
3303                 return false;
3304
3305         if (cfq_class_idle(new_cfqq))
3306                 return false;
3307
3308         if (cfq_class_idle(cfqq))
3309                 return true;
3310
3311         /*
3312          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3313          */
3314         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3315                 return false;
3316
3317         /*
3318          * if the new request is sync, but the currently running queue is
3319          * not, let the sync request have priority.
3320          */
3321         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3322                 return true;
3323
3324         if (new_cfqq->cfqg != cfqq->cfqg)
3325                 return false;
3326
3327         if (cfq_slice_used(cfqq))
3328                 return true;
3329
3330         /* Allow preemption only if we are idling on sync-noidle tree */
3331         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3332             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3333             new_cfqq->service_tree->count == 2 &&
3334             RB_EMPTY_ROOT(&cfqq->sort_list))
3335                 return true;
3336
3337         /*
3338          * So both queues are sync. Let the new request get disk time if
3339          * it's a metadata request and the current queue is doing regular IO.
3340          */
3341         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3342                 return true;
3343
3344         /*
3345          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3346          */
3347         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3348                 return true;
3349
3350         /* An idle queue should not be idle now for some reason */
3351         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3352                 return true;
3353
3354         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3355                 return false;
3356
3357         /*
3358          * if this request is as-good as one we would expect from the
3359          * current cfqq, let it preempt
3360          */
3361         if (cfq_rq_close(cfqd, cfqq, rq))
3362                 return true;
3363
3364         return false;
3365 }
3366
3367 /*
3368  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3369  * let it have half of its nominal slice.
3370  */
3371 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3372 {
3373         struct cfq_queue *old_cfqq = cfqd->active_queue;
3374
3375         cfq_log_cfqq(cfqd, cfqq, "preempt");
3376         cfq_slice_expired(cfqd, 1);
3377
3378         /*
3379          * workload type is changed, don't save slice, otherwise preempt
3380          * doesn't happen
3381          */
3382         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3383                 cfqq->cfqg->saved_workload_slice = 0;
3384
3385         /*
3386          * Put the new queue at the front of the of the current list,
3387          * so we know that it will be selected next.
3388          */
3389         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3390
3391         cfq_service_tree_add(cfqd, cfqq, 1);
3392
3393         cfqq->slice_end = 0;
3394         cfq_mark_cfqq_slice_new(cfqq);
3395 }
3396
3397 /*
3398  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3399  * something we should do about it
3400  */
3401 static void
3402 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3403                 struct request *rq)
3404 {
3405         struct cfq_io_context *cic = RQ_CIC(rq);
3406
3407         cfqd->rq_queued++;
3408         if (rq->cmd_flags & REQ_META)
3409                 cfqq->meta_pending++;
3410
3411         cfq_update_io_thinktime(cfqd, cic);
3412         cfq_update_io_seektime(cfqd, cfqq, rq);
3413         cfq_update_idle_window(cfqd, cfqq, cic);
3414
3415         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3416
3417         if (cfqq == cfqd->active_queue) {
3418                 /*
3419                  * Remember that we saw a request from this process, but
3420                  * don't start queuing just yet. Otherwise we risk seeing lots
3421                  * of tiny requests, because we disrupt the normal plugging
3422                  * and merging. If the request is already larger than a single
3423                  * page, let it rip immediately. For that case we assume that
3424                  * merging is already done. Ditto for a busy system that
3425                  * has other work pending, don't risk delaying until the
3426                  * idle timer unplug to continue working.
3427                  */
3428                 if (cfq_cfqq_wait_request(cfqq)) {
3429                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3430                             cfqd->busy_queues > 1) {
3431                                 cfq_del_timer(cfqd, cfqq);
3432                                 cfq_clear_cfqq_wait_request(cfqq);
3433                                 __blk_run_queue(cfqd->queue);
3434                         } else {
3435                                 cfq_blkiocg_update_idle_time_stats(
3436                                                 &cfqq->cfqg->blkg);
3437                                 cfq_mark_cfqq_must_dispatch(cfqq);
3438                         }
3439                 }
3440         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3441                 /*
3442                  * not the active queue - expire current slice if it is
3443                  * idle and has expired it's mean thinktime or this new queue
3444                  * has some old slice time left and is of higher priority or
3445                  * this new queue is RT and the current one is BE
3446                  */
3447                 cfq_preempt_queue(cfqd, cfqq);
3448                 __blk_run_queue(cfqd->queue);
3449         }
3450 }
3451
3452 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3453 {
3454         struct cfq_data *cfqd = q->elevator->elevator_data;
3455         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3456
3457         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3458         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3459
3460         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3461         list_add_tail(&rq->queuelist, &cfqq->fifo);
3462         cfq_add_rq_rb(rq);
3463         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3464                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3465                         rq_is_sync(rq));
3466         cfq_rq_enqueued(cfqd, cfqq, rq);
3467 }
3468
3469 /*
3470  * Update hw_tag based on peak queue depth over 50 samples under
3471  * sufficient load.
3472  */
3473 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3474 {
3475         struct cfq_queue *cfqq = cfqd->active_queue;
3476
3477         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3478                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3479
3480         if (cfqd->hw_tag == 1)
3481                 return;
3482
3483         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3484             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3485                 return;
3486
3487         /*
3488          * If active queue hasn't enough requests and can idle, cfq might not
3489          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3490          * case
3491          */
3492         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3493             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3494             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3495                 return;
3496
3497         if (cfqd->hw_tag_samples++ < 50)
3498                 return;
3499
3500         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3501                 cfqd->hw_tag = 1;
3502         else
3503                 cfqd->hw_tag = 0;
3504 }
3505
3506 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3507 {
3508         struct cfq_io_context *cic = cfqd->active_cic;
3509
3510         /* If the queue already has requests, don't wait */
3511         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3512                 return false;
3513
3514         /* If there are other queues in the group, don't wait */
3515         if (cfqq->cfqg->nr_cfqq > 1)
3516                 return false;
3517
3518         if (cfq_slice_used(cfqq))
3519                 return true;
3520
3521         /* if slice left is less than think time, wait busy */
3522         if (cic && sample_valid(cic->ttime_samples)
3523             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3524                 return true;
3525
3526         /*
3527          * If think times is less than a jiffy than ttime_mean=0 and above
3528          * will not be true. It might happen that slice has not expired yet
3529          * but will expire soon (4-5 ns) during select_queue(). To cover the
3530          * case where think time is less than a jiffy, mark the queue wait
3531          * busy if only 1 jiffy is left in the slice.
3532          */
3533         if (cfqq->slice_end - jiffies == 1)
3534                 return true;
3535
3536         return false;
3537 }
3538
3539 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3540 {
3541         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3542         struct cfq_data *cfqd = cfqq->cfqd;
3543         const int sync = rq_is_sync(rq);
3544         unsigned long now;
3545
3546         now = jiffies;
3547         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3548                      !!(rq->cmd_flags & REQ_NOIDLE));
3549
3550         cfq_update_hw_tag(cfqd);
3551
3552         WARN_ON(!cfqd->rq_in_driver);
3553         WARN_ON(!cfqq->dispatched);
3554         cfqd->rq_in_driver--;
3555         cfqq->dispatched--;
3556         (RQ_CFQG(rq))->dispatched--;
3557         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3558                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3559                         rq_data_dir(rq), rq_is_sync(rq));
3560
3561         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3562
3563         if (sync) {
3564                 RQ_CIC(rq)->last_end_request = now;
3565                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3566                         cfqd->last_delayed_sync = now;
3567         }
3568
3569         /*
3570          * If this is the active queue, check if it needs to be expired,
3571          * or if we want to idle in case it has no pending requests.
3572          */
3573         if (cfqd->active_queue == cfqq) {
3574                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3575
3576                 if (cfq_cfqq_slice_new(cfqq)) {
3577                         cfq_set_prio_slice(cfqd, cfqq);
3578                         cfq_clear_cfqq_slice_new(cfqq);
3579                 }
3580
3581                 /*
3582                  * Should we wait for next request to come in before we expire
3583                  * the queue.
3584                  */
3585                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3586                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3587                         if (!cfqd->cfq_slice_idle)
3588                                 extend_sl = cfqd->cfq_group_idle;
3589                         cfqq->slice_end = jiffies + extend_sl;
3590                         cfq_mark_cfqq_wait_busy(cfqq);
3591                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3592                 }
3593
3594                 /*
3595                  * Idling is not enabled on:
3596                  * - expired queues
3597                  * - idle-priority queues
3598                  * - async queues
3599                  * - queues with still some requests queued
3600                  * - when there is a close cooperator
3601                  */
3602                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3603                         cfq_slice_expired(cfqd, 1);
3604                 else if (sync && cfqq_empty &&
3605                          !cfq_close_cooperator(cfqd, cfqq)) {
3606                         cfq_arm_slice_timer(cfqd);
3607                 }
3608         }
3609
3610         if (!cfqd->rq_in_driver)
3611                 cfq_schedule_dispatch(cfqd);
3612 }
3613
3614 /*
3615  * we temporarily boost lower priority queues if they are holding fs exclusive
3616  * resources. they are boosted to normal prio (CLASS_BE/4)
3617  */
3618 static void cfq_prio_boost(struct cfq_queue *cfqq)
3619 {
3620         if (has_fs_excl()) {
3621                 /*
3622                  * boost idle prio on transactions that would lock out other
3623                  * users of the filesystem
3624                  */
3625                 if (cfq_class_idle(cfqq))
3626                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3627                 if (cfqq->ioprio > IOPRIO_NORM)
3628                         cfqq->ioprio = IOPRIO_NORM;
3629         } else {
3630                 /*
3631                  * unboost the queue (if needed)
3632                  */
3633                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3634                 cfqq->ioprio = cfqq->org_ioprio;
3635         }
3636 }
3637
3638 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3639 {
3640         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3641                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3642                 return ELV_MQUEUE_MUST;
3643         }
3644
3645         return ELV_MQUEUE_MAY;
3646 }
3647
3648 static int cfq_may_queue(struct request_queue *q, int rw)
3649 {
3650         struct cfq_data *cfqd = q->elevator->elevator_data;
3651         struct task_struct *tsk = current;
3652         struct cfq_io_context *cic;
3653         struct cfq_queue *cfqq;
3654
3655         /*
3656          * don't force setup of a queue from here, as a call to may_queue
3657          * does not necessarily imply that a request actually will be queued.
3658          * so just lookup a possibly existing queue, or return 'may queue'
3659          * if that fails
3660          */
3661         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3662         if (!cic)
3663                 return ELV_MQUEUE_MAY;
3664
3665         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3666         if (cfqq) {
3667                 cfq_init_prio_data(cfqq, cic->ioc);
3668                 cfq_prio_boost(cfqq);
3669
3670                 return __cfq_may_queue(cfqq);
3671         }
3672
3673         return ELV_MQUEUE_MAY;
3674 }
3675
3676 /*
3677  * queue lock held here
3678  */
3679 static void cfq_put_request(struct request *rq)
3680 {
3681         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3682
3683         if (cfqq) {
3684                 const int rw = rq_data_dir(rq);
3685
3686                 BUG_ON(!cfqq->allocated[rw]);
3687                 cfqq->allocated[rw]--;
3688
3689                 put_io_context(RQ_CIC(rq)->ioc);
3690
3691                 rq->elevator_private[0] = NULL;
3692                 rq->elevator_private[1] = NULL;
3693
3694                 /* Put down rq reference on cfqg */
3695                 cfq_put_cfqg(RQ_CFQG(rq));
3696                 rq->elevator_private[2] = NULL;
3697
3698                 cfq_put_queue(cfqq);
3699         }
3700 }
3701
3702 static struct cfq_queue *
3703 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3704                 struct cfq_queue *cfqq)
3705 {
3706         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3707         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3708         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3709         cfq_put_queue(cfqq);
3710         return cic_to_cfqq(cic, 1);
3711 }
3712
3713 /*
3714  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3715  * was the last process referring to said cfqq.
3716  */
3717 static struct cfq_queue *
3718 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3719 {
3720         if (cfqq_process_refs(cfqq) == 1) {
3721                 cfqq->pid = current->pid;
3722                 cfq_clear_cfqq_coop(cfqq);
3723                 cfq_clear_cfqq_split_coop(cfqq);
3724                 return cfqq;
3725         }
3726
3727         cic_set_cfqq(cic, NULL, 1);
3728
3729         cfq_put_cooperator(cfqq);
3730
3731         cfq_put_queue(cfqq);
3732         return NULL;
3733 }
3734 /*
3735  * Allocate cfq data structures associated with this request.
3736  */
3737 static int
3738 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3739 {
3740         struct cfq_data *cfqd = q->elevator->elevator_data;
3741         struct cfq_io_context *cic;
3742         const int rw = rq_data_dir(rq);
3743         const bool is_sync = rq_is_sync(rq);
3744         struct cfq_queue *cfqq;
3745         unsigned long flags;
3746
3747         might_sleep_if(gfp_mask & __GFP_WAIT);
3748
3749         cic = cfq_get_io_context(cfqd, gfp_mask);
3750
3751         spin_lock_irqsave(q->queue_lock, flags);
3752
3753         if (!cic)
3754                 goto queue_fail;
3755
3756 new_queue:
3757         cfqq = cic_to_cfqq(cic, is_sync);
3758         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3759                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3760                 cic_set_cfqq(cic, cfqq, is_sync);
3761         } else {
3762                 /*
3763                  * If the queue was seeky for too long, break it apart.
3764                  */
3765                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3766                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3767                         cfqq = split_cfqq(cic, cfqq);
3768                         if (!cfqq)
3769                                 goto new_queue;
3770                 }
3771
3772                 /*
3773                  * Check to see if this queue is scheduled to merge with
3774                  * another, closely cooperating queue.  The merging of
3775                  * queues happens here as it must be done in process context.
3776                  * The reference on new_cfqq was taken in merge_cfqqs.
3777                  */
3778                 if (cfqq->new_cfqq)
3779                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3780         }
3781
3782         cfqq->allocated[rw]++;
3783
3784         cfqq->ref++;
3785         rq->elevator_private[0] = cic;
3786         rq->elevator_private[1] = cfqq;
3787         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3788         spin_unlock_irqrestore(q->queue_lock, flags);
3789         return 0;
3790
3791 queue_fail:
3792         cfq_schedule_dispatch(cfqd);
3793         spin_unlock_irqrestore(q->queue_lock, flags);
3794         cfq_log(cfqd, "set_request fail");
3795         return 1;
3796 }
3797
3798 static void cfq_kick_queue(struct work_struct *work)
3799 {
3800         struct cfq_data *cfqd =
3801                 container_of(work, struct cfq_data, unplug_work);
3802         struct request_queue *q = cfqd->queue;
3803
3804         spin_lock_irq(q->queue_lock);
3805         __blk_run_queue(cfqd->queue);
3806         spin_unlock_irq(q->queue_lock);
3807 }
3808
3809 /*
3810  * Timer running if the active_queue is currently idling inside its time slice
3811  */
3812 static void cfq_idle_slice_timer(unsigned long data)
3813 {
3814         struct cfq_data *cfqd = (struct cfq_data *) data;
3815         struct cfq_queue *cfqq;
3816         unsigned long flags;
3817         int timed_out = 1;
3818
3819         cfq_log(cfqd, "idle timer fired");
3820
3821         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3822
3823         cfqq = cfqd->active_queue;
3824         if (cfqq) {
3825                 timed_out = 0;
3826
3827                 /*
3828                  * We saw a request before the queue expired, let it through
3829                  */
3830                 if (cfq_cfqq_must_dispatch(cfqq))
3831                         goto out_kick;
3832
3833                 /*
3834                  * expired
3835                  */
3836                 if (cfq_slice_used(cfqq))
3837                         goto expire;
3838
3839                 /*
3840                  * only expire and reinvoke request handler, if there are
3841                  * other queues with pending requests
3842                  */
3843                 if (!cfqd->busy_queues)
3844                         goto out_cont;
3845
3846                 /*
3847                  * not expired and it has a request pending, let it dispatch
3848                  */
3849                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3850                         goto out_kick;
3851
3852                 /*
3853                  * Queue depth flag is reset only when the idle didn't succeed
3854                  */
3855                 cfq_clear_cfqq_deep(cfqq);
3856         }
3857 expire:
3858         cfq_slice_expired(cfqd, timed_out);
3859 out_kick:
3860         cfq_schedule_dispatch(cfqd);
3861 out_cont:
3862         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3863 }
3864
3865 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3866 {
3867         del_timer_sync(&cfqd->idle_slice_timer);
3868         cancel_work_sync(&cfqd->unplug_work);
3869 }
3870
3871 static void cfq_put_async_queues(struct cfq_data *cfqd)
3872 {
3873         int i;
3874
3875         for (i = 0; i < IOPRIO_BE_NR; i++) {
3876                 if (cfqd->async_cfqq[0][i])
3877                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3878                 if (cfqd->async_cfqq[1][i])
3879                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3880         }
3881
3882         if (cfqd->async_idle_cfqq)
3883                 cfq_put_queue(cfqd->async_idle_cfqq);
3884 }
3885
3886 static void cfq_exit_queue(struct elevator_queue *e)
3887 {
3888         struct cfq_data *cfqd = e->elevator_data;
3889         struct request_queue *q = cfqd->queue;
3890         bool wait = false;
3891
3892         cfq_shutdown_timer_wq(cfqd);
3893
3894         spin_lock_irq(q->queue_lock);
3895
3896         if (cfqd->active_queue)
3897                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3898
3899         while (!list_empty(&cfqd->cic_list)) {
3900                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3901                                                         struct cfq_io_context,
3902                                                         queue_list);
3903
3904                 __cfq_exit_single_io_context(cfqd, cic);
3905         }
3906
3907         cfq_put_async_queues(cfqd);
3908         cfq_release_cfq_groups(cfqd);
3909
3910         /*
3911          * If there are groups which we could not unlink from blkcg list,
3912          * wait for a rcu period for them to be freed.
3913          */
3914         if (cfqd->nr_blkcg_linked_grps)
3915                 wait = true;
3916
3917         spin_unlock_irq(q->queue_lock);
3918
3919         cfq_shutdown_timer_wq(cfqd);
3920
3921         spin_lock(&cic_index_lock);
3922         ida_remove(&cic_index_ida, cfqd->cic_index);
3923         spin_unlock(&cic_index_lock);
3924
3925         /*
3926          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3927          * Do this wait only if there are other unlinked groups out
3928          * there. This can happen if cgroup deletion path claimed the
3929          * responsibility of cleaning up a group before queue cleanup code
3930          * get to the group.
3931          *
3932          * Do not call synchronize_rcu() unconditionally as there are drivers
3933          * which create/delete request queue hundreds of times during scan/boot
3934          * and synchronize_rcu() can take significant time and slow down boot.
3935          */
3936         if (wait)
3937                 synchronize_rcu();
3938
3939 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3940         /* Free up per cpu stats for root group */
3941         free_percpu(cfqd->root_group.blkg.stats_cpu);
3942 #endif
3943         kfree(cfqd);
3944 }
3945
3946 static int cfq_alloc_cic_index(void)
3947 {
3948         int index, error;
3949
3950         do {
3951                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3952                         return -ENOMEM;
3953
3954                 spin_lock(&cic_index_lock);
3955                 error = ida_get_new(&cic_index_ida, &index);
3956                 spin_unlock(&cic_index_lock);
3957                 if (error && error != -EAGAIN)
3958                         return error;
3959         } while (error);
3960
3961         return index;
3962 }
3963
3964 static void *cfq_init_queue(struct request_queue *q)
3965 {
3966         struct cfq_data *cfqd;
3967         int i, j;
3968         struct cfq_group *cfqg;
3969         struct cfq_rb_root *st;
3970
3971         i = cfq_alloc_cic_index();
3972         if (i < 0)
3973                 return NULL;
3974
3975         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3976         if (!cfqd) {
3977                 spin_lock(&cic_index_lock);
3978                 ida_remove(&cic_index_ida, i);
3979                 spin_unlock(&cic_index_lock);
3980                 return NULL;
3981         }
3982
3983         /*
3984          * Don't need take queue_lock in the routine, since we are
3985          * initializing the ioscheduler, and nobody is using cfqd
3986          */
3987         cfqd->cic_index = i;
3988
3989         /* Init root service tree */
3990         cfqd->grp_service_tree = CFQ_RB_ROOT;
3991
3992         /* Init root group */
3993         cfqg = &cfqd->root_group;
3994         for_each_cfqg_st(cfqg, i, j, st)
3995                 *st = CFQ_RB_ROOT;
3996         RB_CLEAR_NODE(&cfqg->rb_node);
3997
3998         /* Give preference to root group over other groups */
3999         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
4000
4001 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4002         /*
4003          * Set root group reference to 2. One reference will be dropped when
4004          * all groups on cfqd->cfqg_list are being deleted during queue exit.
4005          * Other reference will remain there as we don't want to delete this
4006          * group as it is statically allocated and gets destroyed when
4007          * throtl_data goes away.
4008          */
4009         cfqg->ref = 2;
4010
4011         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4012                 kfree(cfqg);
4013                 kfree(cfqd);
4014                 return NULL;
4015         }
4016
4017         rcu_read_lock();
4018
4019         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4020                                         (void *)cfqd, 0);
4021         rcu_read_unlock();
4022         cfqd->nr_blkcg_linked_grps++;
4023
4024         /* Add group on cfqd->cfqg_list */
4025         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4026 #endif
4027         /*
4028          * Not strictly needed (since RB_ROOT just clears the node and we
4029          * zeroed cfqd on alloc), but better be safe in case someone decides
4030          * to add magic to the rb code
4031          */
4032         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4033                 cfqd->prio_trees[i] = RB_ROOT;
4034
4035         /*
4036          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4037          * Grab a permanent reference to it, so that the normal code flow
4038          * will not attempt to free it.
4039          */
4040         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4041         cfqd->oom_cfqq.ref++;
4042         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4043
4044         INIT_LIST_HEAD(&cfqd->cic_list);
4045
4046         cfqd->queue = q;
4047
4048         init_timer(&cfqd->idle_slice_timer);
4049         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4050         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4051
4052         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4053
4054         cfqd->cfq_quantum = cfq_quantum;
4055         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4056         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4057         cfqd->cfq_back_max = cfq_back_max;
4058         cfqd->cfq_back_penalty = cfq_back_penalty;
4059         cfqd->cfq_slice[0] = cfq_slice_async;
4060         cfqd->cfq_slice[1] = cfq_slice_sync;
4061         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4062         cfqd->cfq_slice_idle = cfq_slice_idle;
4063         cfqd->cfq_group_idle = cfq_group_idle;
4064         cfqd->cfq_latency = 1;
4065         cfqd->hw_tag = -1;
4066         /*
4067          * we optimistically start assuming sync ops weren't delayed in last
4068          * second, in order to have larger depth for async operations.
4069          */
4070         cfqd->last_delayed_sync = jiffies - HZ;
4071         return cfqd;
4072 }
4073
4074 static void cfq_slab_kill(void)
4075 {
4076         /*
4077          * Caller already ensured that pending RCU callbacks are completed,
4078          * so we should have no busy allocations at this point.
4079          */
4080         if (cfq_pool)
4081                 kmem_cache_destroy(cfq_pool);
4082         if (cfq_ioc_pool)
4083                 kmem_cache_destroy(cfq_ioc_pool);
4084 }
4085
4086 static int __init cfq_slab_setup(void)
4087 {
4088         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4089         if (!cfq_pool)
4090                 goto fail;
4091
4092         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4093         if (!cfq_ioc_pool)
4094                 goto fail;
4095
4096         return 0;
4097 fail:
4098         cfq_slab_kill();
4099         return -ENOMEM;
4100 }
4101
4102 /*
4103  * sysfs parts below -->
4104  */
4105 static ssize_t
4106 cfq_var_show(unsigned int var, char *page)
4107 {
4108         return sprintf(page, "%d\n", var);
4109 }
4110
4111 static ssize_t
4112 cfq_var_store(unsigned int *var, const char *page, size_t count)
4113 {
4114         char *p = (char *) page;
4115
4116         *var = simple_strtoul(p, &p, 10);
4117         return count;
4118 }
4119
4120 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4121 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4122 {                                                                       \
4123         struct cfq_data *cfqd = e->elevator_data;                       \
4124         unsigned int __data = __VAR;                                    \
4125         if (__CONV)                                                     \
4126                 __data = jiffies_to_msecs(__data);                      \
4127         return cfq_var_show(__data, (page));                            \
4128 }
4129 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4130 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4131 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4132 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4133 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4134 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4135 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4136 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4137 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4138 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4139 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4140 #undef SHOW_FUNCTION
4141
4142 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4143 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4144 {                                                                       \
4145         struct cfq_data *cfqd = e->elevator_data;                       \
4146         unsigned int __data;                                            \
4147         int ret = cfq_var_store(&__data, (page), count);                \
4148         if (__data < (MIN))                                             \
4149                 __data = (MIN);                                         \
4150         else if (__data > (MAX))                                        \
4151                 __data = (MAX);                                         \
4152         if (__CONV)                                                     \
4153                 *(__PTR) = msecs_to_jiffies(__data);                    \
4154         else                                                            \
4155                 *(__PTR) = __data;                                      \
4156         return ret;                                                     \
4157 }
4158 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4159 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4160                 UINT_MAX, 1);
4161 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4162                 UINT_MAX, 1);
4163 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4164 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4165                 UINT_MAX, 0);
4166 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4167 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4168 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4169 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4170 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4171                 UINT_MAX, 0);
4172 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4173 #undef STORE_FUNCTION
4174
4175 #define CFQ_ATTR(name) \
4176         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4177
4178 static struct elv_fs_entry cfq_attrs[] = {
4179         CFQ_ATTR(quantum),
4180         CFQ_ATTR(fifo_expire_sync),
4181         CFQ_ATTR(fifo_expire_async),
4182         CFQ_ATTR(back_seek_max),
4183         CFQ_ATTR(back_seek_penalty),
4184         CFQ_ATTR(slice_sync),
4185         CFQ_ATTR(slice_async),
4186         CFQ_ATTR(slice_async_rq),
4187         CFQ_ATTR(slice_idle),
4188         CFQ_ATTR(group_idle),
4189         CFQ_ATTR(low_latency),
4190         __ATTR_NULL
4191 };
4192
4193 static struct elevator_type iosched_cfq = {
4194         .ops = {
4195                 .elevator_merge_fn =            cfq_merge,
4196                 .elevator_merged_fn =           cfq_merged_request,
4197                 .elevator_merge_req_fn =        cfq_merged_requests,
4198                 .elevator_allow_merge_fn =      cfq_allow_merge,
4199                 .elevator_bio_merged_fn =       cfq_bio_merged,
4200                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4201                 .elevator_add_req_fn =          cfq_insert_request,
4202                 .elevator_activate_req_fn =     cfq_activate_request,
4203                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4204                 .elevator_completed_req_fn =    cfq_completed_request,
4205                 .elevator_former_req_fn =       elv_rb_former_request,
4206                 .elevator_latter_req_fn =       elv_rb_latter_request,
4207                 .elevator_set_req_fn =          cfq_set_request,
4208                 .elevator_put_req_fn =          cfq_put_request,
4209                 .elevator_may_queue_fn =        cfq_may_queue,
4210                 .elevator_init_fn =             cfq_init_queue,
4211                 .elevator_exit_fn =             cfq_exit_queue,
4212                 .trim =                         cfq_free_io_context,
4213         },
4214         .elevator_attrs =       cfq_attrs,
4215         .elevator_name =        "cfq",
4216         .elevator_owner =       THIS_MODULE,
4217 };
4218
4219 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4220 static struct blkio_policy_type blkio_policy_cfq = {
4221         .ops = {
4222                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4223                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4224         },
4225         .plid = BLKIO_POLICY_PROP,
4226 };
4227 #else
4228 static struct blkio_policy_type blkio_policy_cfq;
4229 #endif
4230
4231 static int __init cfq_init(void)
4232 {
4233         /*
4234          * could be 0 on HZ < 1000 setups
4235          */
4236         if (!cfq_slice_async)
4237                 cfq_slice_async = 1;
4238         if (!cfq_slice_idle)
4239                 cfq_slice_idle = 1;
4240
4241 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4242         if (!cfq_group_idle)
4243                 cfq_group_idle = 1;
4244 #else
4245                 cfq_group_idle = 0;
4246 #endif
4247         if (cfq_slab_setup())
4248                 return -ENOMEM;
4249
4250         elv_register(&iosched_cfq);
4251         blkio_policy_register(&blkio_policy_cfq);
4252
4253         return 0;
4254 }
4255
4256 static void __exit cfq_exit(void)
4257 {
4258         DECLARE_COMPLETION_ONSTACK(all_gone);
4259         blkio_policy_unregister(&blkio_policy_cfq);
4260         elv_unregister(&iosched_cfq);
4261         ioc_gone = &all_gone;
4262         /* ioc_gone's update must be visible before reading ioc_count */
4263         smp_wmb();
4264
4265         /*
4266          * this also protects us from entering cfq_slab_kill() with
4267          * pending RCU callbacks
4268          */
4269         if (elv_ioc_count_read(cfq_ioc_count))
4270                 wait_for_completion(&all_gone);
4271         ida_destroy(&cic_index_ida);
4272         cfq_slab_kill();
4273 }
4274
4275 module_init(cfq_init);
4276 module_exit(cfq_exit);
4277
4278 MODULE_AUTHOR("Jens Axboe");
4279 MODULE_LICENSE("GPL");
4280 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");