Merge branch 'for-linus' of git://git.kernel.dk/linux-block
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90         struct cfq_ttime ttime;
91 };
92 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, \
93                         .ttime = {.last_end_request = jiffies,},}
94
95 /*
96  * Per process-grouping structure
97  */
98 struct cfq_queue {
99         /* reference count */
100         int ref;
101         /* various state flags, see below */
102         unsigned int flags;
103         /* parent cfq_data */
104         struct cfq_data *cfqd;
105         /* service_tree member */
106         struct rb_node rb_node;
107         /* service_tree key */
108         unsigned long rb_key;
109         /* prio tree member */
110         struct rb_node p_node;
111         /* prio tree root we belong to, if any */
112         struct rb_root *p_root;
113         /* sorted list of pending requests */
114         struct rb_root sort_list;
115         /* if fifo isn't expired, next request to serve */
116         struct request *next_rq;
117         /* requests queued in sort_list */
118         int queued[2];
119         /* currently allocated requests */
120         int allocated[2];
121         /* fifo list of requests in sort_list */
122         struct list_head fifo;
123
124         /* time when queue got scheduled in to dispatch first request. */
125         unsigned long dispatch_start;
126         unsigned int allocated_slice;
127         unsigned int slice_dispatch;
128         /* time when first request from queue completed and slice started. */
129         unsigned long slice_start;
130         unsigned long slice_end;
131         long slice_resid;
132
133         /* pending metadata requests */
134         int meta_pending;
135         /* number of requests that are on the dispatch list or inside driver */
136         int dispatched;
137
138         /* io prio of this group */
139         unsigned short ioprio, org_ioprio;
140         unsigned short ioprio_class;
141
142         pid_t pid;
143
144         u32 seek_history;
145         sector_t last_request_pos;
146
147         struct cfq_rb_root *service_tree;
148         struct cfq_queue *new_cfqq;
149         struct cfq_group *cfqg;
150         /* Number of sectors dispatched from queue in single dispatch round */
151         unsigned long nr_sectors;
152 };
153
154 /*
155  * First index in the service_trees.
156  * IDLE is handled separately, so it has negative index
157  */
158 enum wl_prio_t {
159         BE_WORKLOAD = 0,
160         RT_WORKLOAD = 1,
161         IDLE_WORKLOAD = 2,
162         CFQ_PRIO_NR,
163 };
164
165 /*
166  * Second index in the service_trees.
167  */
168 enum wl_type_t {
169         ASYNC_WORKLOAD = 0,
170         SYNC_NOIDLE_WORKLOAD = 1,
171         SYNC_WORKLOAD = 2
172 };
173
174 /* This is per cgroup per device grouping structure */
175 struct cfq_group {
176         /* group service_tree member */
177         struct rb_node rb_node;
178
179         /* group service_tree key */
180         u64 vdisktime;
181         unsigned int weight;
182         unsigned int new_weight;
183         bool needs_update;
184
185         /* number of cfqq currently on this group */
186         int nr_cfqq;
187
188         /*
189          * Per group busy queues average. Useful for workload slice calc. We
190          * create the array for each prio class but at run time it is used
191          * only for RT and BE class and slot for IDLE class remains unused.
192          * This is primarily done to avoid confusion and a gcc warning.
193          */
194         unsigned int busy_queues_avg[CFQ_PRIO_NR];
195         /*
196          * rr lists of queues with requests. We maintain service trees for
197          * RT and BE classes. These trees are subdivided in subclasses
198          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
199          * class there is no subclassification and all the cfq queues go on
200          * a single tree service_tree_idle.
201          * Counts are embedded in the cfq_rb_root
202          */
203         struct cfq_rb_root service_trees[2][3];
204         struct cfq_rb_root service_tree_idle;
205
206         unsigned long saved_workload_slice;
207         enum wl_type_t saved_workload;
208         enum wl_prio_t saved_serving_prio;
209         struct blkio_group blkg;
210 #ifdef CONFIG_CFQ_GROUP_IOSCHED
211         struct hlist_node cfqd_node;
212         int ref;
213 #endif
214         /* number of requests that are on the dispatch list or inside driver */
215         int dispatched;
216         struct cfq_ttime ttime;
217 };
218
219 /*
220  * Per block device queue structure
221  */
222 struct cfq_data {
223         struct request_queue *queue;
224         /* Root service tree for cfq_groups */
225         struct cfq_rb_root grp_service_tree;
226         struct cfq_group root_group;
227
228         /*
229          * The priority currently being served
230          */
231         enum wl_prio_t serving_prio;
232         enum wl_type_t serving_type;
233         unsigned long workload_expires;
234         struct cfq_group *serving_group;
235
236         /*
237          * Each priority tree is sorted by next_request position.  These
238          * trees are used when determining if two or more queues are
239          * interleaving requests (see cfq_close_cooperator).
240          */
241         struct rb_root prio_trees[CFQ_PRIO_LISTS];
242
243         unsigned int busy_queues;
244         unsigned int busy_sync_queues;
245
246         int rq_in_driver;
247         int rq_in_flight[2];
248
249         /*
250          * queue-depth detection
251          */
252         int rq_queued;
253         int hw_tag;
254         /*
255          * hw_tag can be
256          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
257          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
258          *  0 => no NCQ
259          */
260         int hw_tag_est_depth;
261         unsigned int hw_tag_samples;
262
263         /*
264          * idle window management
265          */
266         struct timer_list idle_slice_timer;
267         struct work_struct unplug_work;
268
269         struct cfq_queue *active_queue;
270         struct cfq_io_context *active_cic;
271
272         /*
273          * async queue for each priority case
274          */
275         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
276         struct cfq_queue *async_idle_cfqq;
277
278         sector_t last_position;
279
280         /*
281          * tunables, see top of file
282          */
283         unsigned int cfq_quantum;
284         unsigned int cfq_fifo_expire[2];
285         unsigned int cfq_back_penalty;
286         unsigned int cfq_back_max;
287         unsigned int cfq_slice[2];
288         unsigned int cfq_slice_async_rq;
289         unsigned int cfq_slice_idle;
290         unsigned int cfq_group_idle;
291         unsigned int cfq_latency;
292
293         unsigned int cic_index;
294         struct list_head cic_list;
295
296         /*
297          * Fallback dummy cfqq for extreme OOM conditions
298          */
299         struct cfq_queue oom_cfqq;
300
301         unsigned long last_delayed_sync;
302
303         /* List of cfq groups being managed on this device*/
304         struct hlist_head cfqg_list;
305
306         /* Number of groups which are on blkcg->blkg_list */
307         unsigned int nr_blkcg_linked_grps;
308 };
309
310 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
311
312 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
313                                             enum wl_prio_t prio,
314                                             enum wl_type_t type)
315 {
316         if (!cfqg)
317                 return NULL;
318
319         if (prio == IDLE_WORKLOAD)
320                 return &cfqg->service_tree_idle;
321
322         return &cfqg->service_trees[prio][type];
323 }
324
325 enum cfqq_state_flags {
326         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
327         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
328         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
329         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
330         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
331         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
332         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
333         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
334         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
335         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
336         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
337         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
338         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
339 };
340
341 #define CFQ_CFQQ_FNS(name)                                              \
342 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
343 {                                                                       \
344         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
345 }                                                                       \
346 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
347 {                                                                       \
348         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
349 }                                                                       \
350 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
351 {                                                                       \
352         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
353 }
354
355 CFQ_CFQQ_FNS(on_rr);
356 CFQ_CFQQ_FNS(wait_request);
357 CFQ_CFQQ_FNS(must_dispatch);
358 CFQ_CFQQ_FNS(must_alloc_slice);
359 CFQ_CFQQ_FNS(fifo_expire);
360 CFQ_CFQQ_FNS(idle_window);
361 CFQ_CFQQ_FNS(prio_changed);
362 CFQ_CFQQ_FNS(slice_new);
363 CFQ_CFQQ_FNS(sync);
364 CFQ_CFQQ_FNS(coop);
365 CFQ_CFQQ_FNS(split_coop);
366 CFQ_CFQQ_FNS(deep);
367 CFQ_CFQQ_FNS(wait_busy);
368 #undef CFQ_CFQQ_FNS
369
370 #ifdef CONFIG_CFQ_GROUP_IOSCHED
371 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
372         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
373                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
374                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
375
376 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
377         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
378                                 blkg_path(&(cfqg)->blkg), ##args)       \
379
380 #else
381 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
382         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
383 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
384 #endif
385 #define cfq_log(cfqd, fmt, args...)     \
386         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
387
388 /* Traverses through cfq group service trees */
389 #define for_each_cfqg_st(cfqg, i, j, st) \
390         for (i = 0; i <= IDLE_WORKLOAD; i++) \
391                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
392                         : &cfqg->service_tree_idle; \
393                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
394                         (i == IDLE_WORKLOAD && j == 0); \
395                         j++, st = i < IDLE_WORKLOAD ? \
396                         &cfqg->service_trees[i][j]: NULL) \
397
398 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
399         struct cfq_ttime *ttime, bool group_idle)
400 {
401         unsigned long slice;
402         if (!sample_valid(ttime->ttime_samples))
403                 return false;
404         if (group_idle)
405                 slice = cfqd->cfq_group_idle;
406         else
407                 slice = cfqd->cfq_slice_idle;
408         return ttime->ttime_mean > slice;
409 }
410
411 static inline bool iops_mode(struct cfq_data *cfqd)
412 {
413         /*
414          * If we are not idling on queues and it is a NCQ drive, parallel
415          * execution of requests is on and measuring time is not possible
416          * in most of the cases until and unless we drive shallower queue
417          * depths and that becomes a performance bottleneck. In such cases
418          * switch to start providing fairness in terms of number of IOs.
419          */
420         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
421                 return true;
422         else
423                 return false;
424 }
425
426 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
427 {
428         if (cfq_class_idle(cfqq))
429                 return IDLE_WORKLOAD;
430         if (cfq_class_rt(cfqq))
431                 return RT_WORKLOAD;
432         return BE_WORKLOAD;
433 }
434
435
436 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
437 {
438         if (!cfq_cfqq_sync(cfqq))
439                 return ASYNC_WORKLOAD;
440         if (!cfq_cfqq_idle_window(cfqq))
441                 return SYNC_NOIDLE_WORKLOAD;
442         return SYNC_WORKLOAD;
443 }
444
445 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
446                                         struct cfq_data *cfqd,
447                                         struct cfq_group *cfqg)
448 {
449         if (wl == IDLE_WORKLOAD)
450                 return cfqg->service_tree_idle.count;
451
452         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
453                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
454                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
455 }
456
457 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
458                                         struct cfq_group *cfqg)
459 {
460         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
461                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
462 }
463
464 static void cfq_dispatch_insert(struct request_queue *, struct request *);
465 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
466                                        struct io_context *, gfp_t);
467 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
468                                                 struct io_context *);
469
470 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
471                                             bool is_sync)
472 {
473         return cic->cfqq[is_sync];
474 }
475
476 static inline void cic_set_cfqq(struct cfq_io_context *cic,
477                                 struct cfq_queue *cfqq, bool is_sync)
478 {
479         cic->cfqq[is_sync] = cfqq;
480 }
481
482 #define CIC_DEAD_KEY    1ul
483 #define CIC_DEAD_INDEX_SHIFT    1
484
485 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
486 {
487         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
488 }
489
490 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
491 {
492         struct cfq_data *cfqd = cic->key;
493
494         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
495                 return NULL;
496
497         return cfqd;
498 }
499
500 /*
501  * We regard a request as SYNC, if it's either a read or has the SYNC bit
502  * set (in which case it could also be direct WRITE).
503  */
504 static inline bool cfq_bio_sync(struct bio *bio)
505 {
506         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
507 }
508
509 /*
510  * scheduler run of queue, if there are requests pending and no one in the
511  * driver that will restart queueing
512  */
513 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
514 {
515         if (cfqd->busy_queues) {
516                 cfq_log(cfqd, "schedule dispatch");
517                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
518         }
519 }
520
521 /*
522  * Scale schedule slice based on io priority. Use the sync time slice only
523  * if a queue is marked sync and has sync io queued. A sync queue with async
524  * io only, should not get full sync slice length.
525  */
526 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
527                                  unsigned short prio)
528 {
529         const int base_slice = cfqd->cfq_slice[sync];
530
531         WARN_ON(prio >= IOPRIO_BE_NR);
532
533         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
534 }
535
536 static inline int
537 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
538 {
539         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
540 }
541
542 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
543 {
544         u64 d = delta << CFQ_SERVICE_SHIFT;
545
546         d = d * BLKIO_WEIGHT_DEFAULT;
547         do_div(d, cfqg->weight);
548         return d;
549 }
550
551 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
552 {
553         s64 delta = (s64)(vdisktime - min_vdisktime);
554         if (delta > 0)
555                 min_vdisktime = vdisktime;
556
557         return min_vdisktime;
558 }
559
560 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
561 {
562         s64 delta = (s64)(vdisktime - min_vdisktime);
563         if (delta < 0)
564                 min_vdisktime = vdisktime;
565
566         return min_vdisktime;
567 }
568
569 static void update_min_vdisktime(struct cfq_rb_root *st)
570 {
571         struct cfq_group *cfqg;
572
573         if (st->left) {
574                 cfqg = rb_entry_cfqg(st->left);
575                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
576                                                   cfqg->vdisktime);
577         }
578 }
579
580 /*
581  * get averaged number of queues of RT/BE priority.
582  * average is updated, with a formula that gives more weight to higher numbers,
583  * to quickly follows sudden increases and decrease slowly
584  */
585
586 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
587                                         struct cfq_group *cfqg, bool rt)
588 {
589         unsigned min_q, max_q;
590         unsigned mult  = cfq_hist_divisor - 1;
591         unsigned round = cfq_hist_divisor / 2;
592         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
593
594         min_q = min(cfqg->busy_queues_avg[rt], busy);
595         max_q = max(cfqg->busy_queues_avg[rt], busy);
596         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
597                 cfq_hist_divisor;
598         return cfqg->busy_queues_avg[rt];
599 }
600
601 static inline unsigned
602 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
603 {
604         struct cfq_rb_root *st = &cfqd->grp_service_tree;
605
606         return cfq_target_latency * cfqg->weight / st->total_weight;
607 }
608
609 static inline unsigned
610 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
611 {
612         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
613         if (cfqd->cfq_latency) {
614                 /*
615                  * interested queues (we consider only the ones with the same
616                  * priority class in the cfq group)
617                  */
618                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
619                                                 cfq_class_rt(cfqq));
620                 unsigned sync_slice = cfqd->cfq_slice[1];
621                 unsigned expect_latency = sync_slice * iq;
622                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
623
624                 if (expect_latency > group_slice) {
625                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
626                         /* scale low_slice according to IO priority
627                          * and sync vs async */
628                         unsigned low_slice =
629                                 min(slice, base_low_slice * slice / sync_slice);
630                         /* the adapted slice value is scaled to fit all iqs
631                          * into the target latency */
632                         slice = max(slice * group_slice / expect_latency,
633                                     low_slice);
634                 }
635         }
636         return slice;
637 }
638
639 static inline void
640 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
641 {
642         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
643
644         cfqq->slice_start = jiffies;
645         cfqq->slice_end = jiffies + slice;
646         cfqq->allocated_slice = slice;
647         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
648 }
649
650 /*
651  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
652  * isn't valid until the first request from the dispatch is activated
653  * and the slice time set.
654  */
655 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
656 {
657         if (cfq_cfqq_slice_new(cfqq))
658                 return false;
659         if (time_before(jiffies, cfqq->slice_end))
660                 return false;
661
662         return true;
663 }
664
665 /*
666  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
667  * We choose the request that is closest to the head right now. Distance
668  * behind the head is penalized and only allowed to a certain extent.
669  */
670 static struct request *
671 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
672 {
673         sector_t s1, s2, d1 = 0, d2 = 0;
674         unsigned long back_max;
675 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
676 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
677         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
678
679         if (rq1 == NULL || rq1 == rq2)
680                 return rq2;
681         if (rq2 == NULL)
682                 return rq1;
683
684         if (rq_is_sync(rq1) != rq_is_sync(rq2))
685                 return rq_is_sync(rq1) ? rq1 : rq2;
686
687         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META)
688                 return rq1->cmd_flags & REQ_META ? rq1 : rq2;
689
690         s1 = blk_rq_pos(rq1);
691         s2 = blk_rq_pos(rq2);
692
693         /*
694          * by definition, 1KiB is 2 sectors
695          */
696         back_max = cfqd->cfq_back_max * 2;
697
698         /*
699          * Strict one way elevator _except_ in the case where we allow
700          * short backward seeks which are biased as twice the cost of a
701          * similar forward seek.
702          */
703         if (s1 >= last)
704                 d1 = s1 - last;
705         else if (s1 + back_max >= last)
706                 d1 = (last - s1) * cfqd->cfq_back_penalty;
707         else
708                 wrap |= CFQ_RQ1_WRAP;
709
710         if (s2 >= last)
711                 d2 = s2 - last;
712         else if (s2 + back_max >= last)
713                 d2 = (last - s2) * cfqd->cfq_back_penalty;
714         else
715                 wrap |= CFQ_RQ2_WRAP;
716
717         /* Found required data */
718
719         /*
720          * By doing switch() on the bit mask "wrap" we avoid having to
721          * check two variables for all permutations: --> faster!
722          */
723         switch (wrap) {
724         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
725                 if (d1 < d2)
726                         return rq1;
727                 else if (d2 < d1)
728                         return rq2;
729                 else {
730                         if (s1 >= s2)
731                                 return rq1;
732                         else
733                                 return rq2;
734                 }
735
736         case CFQ_RQ2_WRAP:
737                 return rq1;
738         case CFQ_RQ1_WRAP:
739                 return rq2;
740         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
741         default:
742                 /*
743                  * Since both rqs are wrapped,
744                  * start with the one that's further behind head
745                  * (--> only *one* back seek required),
746                  * since back seek takes more time than forward.
747                  */
748                 if (s1 <= s2)
749                         return rq1;
750                 else
751                         return rq2;
752         }
753 }
754
755 /*
756  * The below is leftmost cache rbtree addon
757  */
758 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
759 {
760         /* Service tree is empty */
761         if (!root->count)
762                 return NULL;
763
764         if (!root->left)
765                 root->left = rb_first(&root->rb);
766
767         if (root->left)
768                 return rb_entry(root->left, struct cfq_queue, rb_node);
769
770         return NULL;
771 }
772
773 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
774 {
775         if (!root->left)
776                 root->left = rb_first(&root->rb);
777
778         if (root->left)
779                 return rb_entry_cfqg(root->left);
780
781         return NULL;
782 }
783
784 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
785 {
786         rb_erase(n, root);
787         RB_CLEAR_NODE(n);
788 }
789
790 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
791 {
792         if (root->left == n)
793                 root->left = NULL;
794         rb_erase_init(n, &root->rb);
795         --root->count;
796 }
797
798 /*
799  * would be nice to take fifo expire time into account as well
800  */
801 static struct request *
802 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
803                   struct request *last)
804 {
805         struct rb_node *rbnext = rb_next(&last->rb_node);
806         struct rb_node *rbprev = rb_prev(&last->rb_node);
807         struct request *next = NULL, *prev = NULL;
808
809         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
810
811         if (rbprev)
812                 prev = rb_entry_rq(rbprev);
813
814         if (rbnext)
815                 next = rb_entry_rq(rbnext);
816         else {
817                 rbnext = rb_first(&cfqq->sort_list);
818                 if (rbnext && rbnext != &last->rb_node)
819                         next = rb_entry_rq(rbnext);
820         }
821
822         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
823 }
824
825 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
826                                       struct cfq_queue *cfqq)
827 {
828         /*
829          * just an approximation, should be ok.
830          */
831         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
832                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
833 }
834
835 static inline s64
836 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
837 {
838         return cfqg->vdisktime - st->min_vdisktime;
839 }
840
841 static void
842 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
843 {
844         struct rb_node **node = &st->rb.rb_node;
845         struct rb_node *parent = NULL;
846         struct cfq_group *__cfqg;
847         s64 key = cfqg_key(st, cfqg);
848         int left = 1;
849
850         while (*node != NULL) {
851                 parent = *node;
852                 __cfqg = rb_entry_cfqg(parent);
853
854                 if (key < cfqg_key(st, __cfqg))
855                         node = &parent->rb_left;
856                 else {
857                         node = &parent->rb_right;
858                         left = 0;
859                 }
860         }
861
862         if (left)
863                 st->left = &cfqg->rb_node;
864
865         rb_link_node(&cfqg->rb_node, parent, node);
866         rb_insert_color(&cfqg->rb_node, &st->rb);
867 }
868
869 static void
870 cfq_update_group_weight(struct cfq_group *cfqg)
871 {
872         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
873         if (cfqg->needs_update) {
874                 cfqg->weight = cfqg->new_weight;
875                 cfqg->needs_update = false;
876         }
877 }
878
879 static void
880 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
881 {
882         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
883
884         cfq_update_group_weight(cfqg);
885         __cfq_group_service_tree_add(st, cfqg);
886         st->total_weight += cfqg->weight;
887 }
888
889 static void
890 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
891 {
892         struct cfq_rb_root *st = &cfqd->grp_service_tree;
893         struct cfq_group *__cfqg;
894         struct rb_node *n;
895
896         cfqg->nr_cfqq++;
897         if (!RB_EMPTY_NODE(&cfqg->rb_node))
898                 return;
899
900         /*
901          * Currently put the group at the end. Later implement something
902          * so that groups get lesser vtime based on their weights, so that
903          * if group does not loose all if it was not continuously backlogged.
904          */
905         n = rb_last(&st->rb);
906         if (n) {
907                 __cfqg = rb_entry_cfqg(n);
908                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
909         } else
910                 cfqg->vdisktime = st->min_vdisktime;
911         cfq_group_service_tree_add(st, cfqg);
912 }
913
914 static void
915 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
916 {
917         st->total_weight -= cfqg->weight;
918         if (!RB_EMPTY_NODE(&cfqg->rb_node))
919                 cfq_rb_erase(&cfqg->rb_node, st);
920 }
921
922 static void
923 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
924 {
925         struct cfq_rb_root *st = &cfqd->grp_service_tree;
926
927         BUG_ON(cfqg->nr_cfqq < 1);
928         cfqg->nr_cfqq--;
929
930         /* If there are other cfq queues under this group, don't delete it */
931         if (cfqg->nr_cfqq)
932                 return;
933
934         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
935         cfq_group_service_tree_del(st, cfqg);
936         cfqg->saved_workload_slice = 0;
937         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
938 }
939
940 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
941                                                 unsigned int *unaccounted_time)
942 {
943         unsigned int slice_used;
944
945         /*
946          * Queue got expired before even a single request completed or
947          * got expired immediately after first request completion.
948          */
949         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
950                 /*
951                  * Also charge the seek time incurred to the group, otherwise
952                  * if there are mutiple queues in the group, each can dispatch
953                  * a single request on seeky media and cause lots of seek time
954                  * and group will never know it.
955                  */
956                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
957                                         1);
958         } else {
959                 slice_used = jiffies - cfqq->slice_start;
960                 if (slice_used > cfqq->allocated_slice) {
961                         *unaccounted_time = slice_used - cfqq->allocated_slice;
962                         slice_used = cfqq->allocated_slice;
963                 }
964                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
965                         *unaccounted_time += cfqq->slice_start -
966                                         cfqq->dispatch_start;
967         }
968
969         return slice_used;
970 }
971
972 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
973                                 struct cfq_queue *cfqq)
974 {
975         struct cfq_rb_root *st = &cfqd->grp_service_tree;
976         unsigned int used_sl, charge, unaccounted_sl = 0;
977         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
978                         - cfqg->service_tree_idle.count;
979
980         BUG_ON(nr_sync < 0);
981         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
982
983         if (iops_mode(cfqd))
984                 charge = cfqq->slice_dispatch;
985         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
986                 charge = cfqq->allocated_slice;
987
988         /* Can't update vdisktime while group is on service tree */
989         cfq_group_service_tree_del(st, cfqg);
990         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
991         /* If a new weight was requested, update now, off tree */
992         cfq_group_service_tree_add(st, cfqg);
993
994         /* This group is being expired. Save the context */
995         if (time_after(cfqd->workload_expires, jiffies)) {
996                 cfqg->saved_workload_slice = cfqd->workload_expires
997                                                 - jiffies;
998                 cfqg->saved_workload = cfqd->serving_type;
999                 cfqg->saved_serving_prio = cfqd->serving_prio;
1000         } else
1001                 cfqg->saved_workload_slice = 0;
1002
1003         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
1004                                         st->min_vdisktime);
1005         cfq_log_cfqq(cfqq->cfqd, cfqq,
1006                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1007                      used_sl, cfqq->slice_dispatch, charge,
1008                      iops_mode(cfqd), cfqq->nr_sectors);
1009         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1010                                           unaccounted_sl);
1011         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1012 }
1013
1014 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1015 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1016 {
1017         if (blkg)
1018                 return container_of(blkg, struct cfq_group, blkg);
1019         return NULL;
1020 }
1021
1022 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1023                                           unsigned int weight)
1024 {
1025         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1026         cfqg->new_weight = weight;
1027         cfqg->needs_update = true;
1028 }
1029
1030 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1031                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1032 {
1033         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1034         unsigned int major, minor;
1035
1036         /*
1037          * Add group onto cgroup list. It might happen that bdi->dev is
1038          * not initialized yet. Initialize this new group without major
1039          * and minor info and this info will be filled in once a new thread
1040          * comes for IO.
1041          */
1042         if (bdi->dev) {
1043                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1044                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1045                                         (void *)cfqd, MKDEV(major, minor));
1046         } else
1047                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1048                                         (void *)cfqd, 0);
1049
1050         cfqd->nr_blkcg_linked_grps++;
1051         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1052
1053         /* Add group on cfqd list */
1054         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1055 }
1056
1057 /*
1058  * Should be called from sleepable context. No request queue lock as per
1059  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1060  * from sleepable context.
1061  */
1062 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1063 {
1064         struct cfq_group *cfqg = NULL;
1065         int i, j, ret;
1066         struct cfq_rb_root *st;
1067
1068         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1069         if (!cfqg)
1070                 return NULL;
1071
1072         for_each_cfqg_st(cfqg, i, j, st)
1073                 *st = CFQ_RB_ROOT;
1074         RB_CLEAR_NODE(&cfqg->rb_node);
1075
1076         cfqg->ttime.last_end_request = jiffies;
1077
1078         /*
1079          * Take the initial reference that will be released on destroy
1080          * This can be thought of a joint reference by cgroup and
1081          * elevator which will be dropped by either elevator exit
1082          * or cgroup deletion path depending on who is exiting first.
1083          */
1084         cfqg->ref = 1;
1085
1086         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1087         if (ret) {
1088                 kfree(cfqg);
1089                 return NULL;
1090         }
1091
1092         return cfqg;
1093 }
1094
1095 static struct cfq_group *
1096 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1097 {
1098         struct cfq_group *cfqg = NULL;
1099         void *key = cfqd;
1100         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1101         unsigned int major, minor;
1102
1103         /*
1104          * This is the common case when there are no blkio cgroups.
1105          * Avoid lookup in this case
1106          */
1107         if (blkcg == &blkio_root_cgroup)
1108                 cfqg = &cfqd->root_group;
1109         else
1110                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1111
1112         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1113                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1114                 cfqg->blkg.dev = MKDEV(major, minor);
1115         }
1116
1117         return cfqg;
1118 }
1119
1120 /*
1121  * Search for the cfq group current task belongs to. request_queue lock must
1122  * be held.
1123  */
1124 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1125 {
1126         struct blkio_cgroup *blkcg;
1127         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1128         struct request_queue *q = cfqd->queue;
1129
1130         rcu_read_lock();
1131         blkcg = task_blkio_cgroup(current);
1132         cfqg = cfq_find_cfqg(cfqd, blkcg);
1133         if (cfqg) {
1134                 rcu_read_unlock();
1135                 return cfqg;
1136         }
1137
1138         /*
1139          * Need to allocate a group. Allocation of group also needs allocation
1140          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1141          * we need to drop rcu lock and queue_lock before we call alloc.
1142          *
1143          * Not taking any queue reference here and assuming that queue is
1144          * around by the time we return. CFQ queue allocation code does
1145          * the same. It might be racy though.
1146          */
1147
1148         rcu_read_unlock();
1149         spin_unlock_irq(q->queue_lock);
1150
1151         cfqg = cfq_alloc_cfqg(cfqd);
1152
1153         spin_lock_irq(q->queue_lock);
1154
1155         rcu_read_lock();
1156         blkcg = task_blkio_cgroup(current);
1157
1158         /*
1159          * If some other thread already allocated the group while we were
1160          * not holding queue lock, free up the group
1161          */
1162         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1163
1164         if (__cfqg) {
1165                 kfree(cfqg);
1166                 rcu_read_unlock();
1167                 return __cfqg;
1168         }
1169
1170         if (!cfqg)
1171                 cfqg = &cfqd->root_group;
1172
1173         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1174         rcu_read_unlock();
1175         return cfqg;
1176 }
1177
1178 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1179 {
1180         cfqg->ref++;
1181         return cfqg;
1182 }
1183
1184 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1185 {
1186         /* Currently, all async queues are mapped to root group */
1187         if (!cfq_cfqq_sync(cfqq))
1188                 cfqg = &cfqq->cfqd->root_group;
1189
1190         cfqq->cfqg = cfqg;
1191         /* cfqq reference on cfqg */
1192         cfqq->cfqg->ref++;
1193 }
1194
1195 static void cfq_put_cfqg(struct cfq_group *cfqg)
1196 {
1197         struct cfq_rb_root *st;
1198         int i, j;
1199
1200         BUG_ON(cfqg->ref <= 0);
1201         cfqg->ref--;
1202         if (cfqg->ref)
1203                 return;
1204         for_each_cfqg_st(cfqg, i, j, st)
1205                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1206         free_percpu(cfqg->blkg.stats_cpu);
1207         kfree(cfqg);
1208 }
1209
1210 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1211 {
1212         /* Something wrong if we are trying to remove same group twice */
1213         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1214
1215         hlist_del_init(&cfqg->cfqd_node);
1216
1217         BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1218         cfqd->nr_blkcg_linked_grps--;
1219
1220         /*
1221          * Put the reference taken at the time of creation so that when all
1222          * queues are gone, group can be destroyed.
1223          */
1224         cfq_put_cfqg(cfqg);
1225 }
1226
1227 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1228 {
1229         struct hlist_node *pos, *n;
1230         struct cfq_group *cfqg;
1231
1232         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1233                 /*
1234                  * If cgroup removal path got to blk_group first and removed
1235                  * it from cgroup list, then it will take care of destroying
1236                  * cfqg also.
1237                  */
1238                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1239                         cfq_destroy_cfqg(cfqd, cfqg);
1240         }
1241 }
1242
1243 /*
1244  * Blk cgroup controller notification saying that blkio_group object is being
1245  * delinked as associated cgroup object is going away. That also means that
1246  * no new IO will come in this group. So get rid of this group as soon as
1247  * any pending IO in the group is finished.
1248  *
1249  * This function is called under rcu_read_lock(). key is the rcu protected
1250  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1251  * read lock.
1252  *
1253  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1254  * it should not be NULL as even if elevator was exiting, cgroup deltion
1255  * path got to it first.
1256  */
1257 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1258 {
1259         unsigned long  flags;
1260         struct cfq_data *cfqd = key;
1261
1262         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1263         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1264         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1265 }
1266
1267 #else /* GROUP_IOSCHED */
1268 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1269 {
1270         return &cfqd->root_group;
1271 }
1272
1273 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1274 {
1275         return cfqg;
1276 }
1277
1278 static inline void
1279 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1280         cfqq->cfqg = cfqg;
1281 }
1282
1283 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1284 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1285
1286 #endif /* GROUP_IOSCHED */
1287
1288 /*
1289  * The cfqd->service_trees holds all pending cfq_queue's that have
1290  * requests waiting to be processed. It is sorted in the order that
1291  * we will service the queues.
1292  */
1293 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1294                                  bool add_front)
1295 {
1296         struct rb_node **p, *parent;
1297         struct cfq_queue *__cfqq;
1298         unsigned long rb_key;
1299         struct cfq_rb_root *service_tree;
1300         int left;
1301         int new_cfqq = 1;
1302
1303         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1304                                                 cfqq_type(cfqq));
1305         if (cfq_class_idle(cfqq)) {
1306                 rb_key = CFQ_IDLE_DELAY;
1307                 parent = rb_last(&service_tree->rb);
1308                 if (parent && parent != &cfqq->rb_node) {
1309                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1310                         rb_key += __cfqq->rb_key;
1311                 } else
1312                         rb_key += jiffies;
1313         } else if (!add_front) {
1314                 /*
1315                  * Get our rb key offset. Subtract any residual slice
1316                  * value carried from last service. A negative resid
1317                  * count indicates slice overrun, and this should position
1318                  * the next service time further away in the tree.
1319                  */
1320                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1321                 rb_key -= cfqq->slice_resid;
1322                 cfqq->slice_resid = 0;
1323         } else {
1324                 rb_key = -HZ;
1325                 __cfqq = cfq_rb_first(service_tree);
1326                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1327         }
1328
1329         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1330                 new_cfqq = 0;
1331                 /*
1332                  * same position, nothing more to do
1333                  */
1334                 if (rb_key == cfqq->rb_key &&
1335                     cfqq->service_tree == service_tree)
1336                         return;
1337
1338                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1339                 cfqq->service_tree = NULL;
1340         }
1341
1342         left = 1;
1343         parent = NULL;
1344         cfqq->service_tree = service_tree;
1345         p = &service_tree->rb.rb_node;
1346         while (*p) {
1347                 struct rb_node **n;
1348
1349                 parent = *p;
1350                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1351
1352                 /*
1353                  * sort by key, that represents service time.
1354                  */
1355                 if (time_before(rb_key, __cfqq->rb_key))
1356                         n = &(*p)->rb_left;
1357                 else {
1358                         n = &(*p)->rb_right;
1359                         left = 0;
1360                 }
1361
1362                 p = n;
1363         }
1364
1365         if (left)
1366                 service_tree->left = &cfqq->rb_node;
1367
1368         cfqq->rb_key = rb_key;
1369         rb_link_node(&cfqq->rb_node, parent, p);
1370         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1371         service_tree->count++;
1372         if (add_front || !new_cfqq)
1373                 return;
1374         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1375 }
1376
1377 static struct cfq_queue *
1378 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1379                      sector_t sector, struct rb_node **ret_parent,
1380                      struct rb_node ***rb_link)
1381 {
1382         struct rb_node **p, *parent;
1383         struct cfq_queue *cfqq = NULL;
1384
1385         parent = NULL;
1386         p = &root->rb_node;
1387         while (*p) {
1388                 struct rb_node **n;
1389
1390                 parent = *p;
1391                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1392
1393                 /*
1394                  * Sort strictly based on sector.  Smallest to the left,
1395                  * largest to the right.
1396                  */
1397                 if (sector > blk_rq_pos(cfqq->next_rq))
1398                         n = &(*p)->rb_right;
1399                 else if (sector < blk_rq_pos(cfqq->next_rq))
1400                         n = &(*p)->rb_left;
1401                 else
1402                         break;
1403                 p = n;
1404                 cfqq = NULL;
1405         }
1406
1407         *ret_parent = parent;
1408         if (rb_link)
1409                 *rb_link = p;
1410         return cfqq;
1411 }
1412
1413 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1414 {
1415         struct rb_node **p, *parent;
1416         struct cfq_queue *__cfqq;
1417
1418         if (cfqq->p_root) {
1419                 rb_erase(&cfqq->p_node, cfqq->p_root);
1420                 cfqq->p_root = NULL;
1421         }
1422
1423         if (cfq_class_idle(cfqq))
1424                 return;
1425         if (!cfqq->next_rq)
1426                 return;
1427
1428         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1429         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1430                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1431         if (!__cfqq) {
1432                 rb_link_node(&cfqq->p_node, parent, p);
1433                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1434         } else
1435                 cfqq->p_root = NULL;
1436 }
1437
1438 /*
1439  * Update cfqq's position in the service tree.
1440  */
1441 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1442 {
1443         /*
1444          * Resorting requires the cfqq to be on the RR list already.
1445          */
1446         if (cfq_cfqq_on_rr(cfqq)) {
1447                 cfq_service_tree_add(cfqd, cfqq, 0);
1448                 cfq_prio_tree_add(cfqd, cfqq);
1449         }
1450 }
1451
1452 /*
1453  * add to busy list of queues for service, trying to be fair in ordering
1454  * the pending list according to last request service
1455  */
1456 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1457 {
1458         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1459         BUG_ON(cfq_cfqq_on_rr(cfqq));
1460         cfq_mark_cfqq_on_rr(cfqq);
1461         cfqd->busy_queues++;
1462         if (cfq_cfqq_sync(cfqq))
1463                 cfqd->busy_sync_queues++;
1464
1465         cfq_resort_rr_list(cfqd, cfqq);
1466 }
1467
1468 /*
1469  * Called when the cfqq no longer has requests pending, remove it from
1470  * the service tree.
1471  */
1472 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1473 {
1474         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1475         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1476         cfq_clear_cfqq_on_rr(cfqq);
1477
1478         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1479                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1480                 cfqq->service_tree = NULL;
1481         }
1482         if (cfqq->p_root) {
1483                 rb_erase(&cfqq->p_node, cfqq->p_root);
1484                 cfqq->p_root = NULL;
1485         }
1486
1487         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1488         BUG_ON(!cfqd->busy_queues);
1489         cfqd->busy_queues--;
1490         if (cfq_cfqq_sync(cfqq))
1491                 cfqd->busy_sync_queues--;
1492 }
1493
1494 /*
1495  * rb tree support functions
1496  */
1497 static void cfq_del_rq_rb(struct request *rq)
1498 {
1499         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1500         const int sync = rq_is_sync(rq);
1501
1502         BUG_ON(!cfqq->queued[sync]);
1503         cfqq->queued[sync]--;
1504
1505         elv_rb_del(&cfqq->sort_list, rq);
1506
1507         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1508                 /*
1509                  * Queue will be deleted from service tree when we actually
1510                  * expire it later. Right now just remove it from prio tree
1511                  * as it is empty.
1512                  */
1513                 if (cfqq->p_root) {
1514                         rb_erase(&cfqq->p_node, cfqq->p_root);
1515                         cfqq->p_root = NULL;
1516                 }
1517         }
1518 }
1519
1520 static void cfq_add_rq_rb(struct request *rq)
1521 {
1522         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1523         struct cfq_data *cfqd = cfqq->cfqd;
1524         struct request *prev;
1525
1526         cfqq->queued[rq_is_sync(rq)]++;
1527
1528         elv_rb_add(&cfqq->sort_list, rq);
1529
1530         if (!cfq_cfqq_on_rr(cfqq))
1531                 cfq_add_cfqq_rr(cfqd, cfqq);
1532
1533         /*
1534          * check if this request is a better next-serve candidate
1535          */
1536         prev = cfqq->next_rq;
1537         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1538
1539         /*
1540          * adjust priority tree position, if ->next_rq changes
1541          */
1542         if (prev != cfqq->next_rq)
1543                 cfq_prio_tree_add(cfqd, cfqq);
1544
1545         BUG_ON(!cfqq->next_rq);
1546 }
1547
1548 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1549 {
1550         elv_rb_del(&cfqq->sort_list, rq);
1551         cfqq->queued[rq_is_sync(rq)]--;
1552         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1553                                         rq_data_dir(rq), rq_is_sync(rq));
1554         cfq_add_rq_rb(rq);
1555         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1556                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1557                         rq_is_sync(rq));
1558 }
1559
1560 static struct request *
1561 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1562 {
1563         struct task_struct *tsk = current;
1564         struct cfq_io_context *cic;
1565         struct cfq_queue *cfqq;
1566
1567         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1568         if (!cic)
1569                 return NULL;
1570
1571         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1572         if (cfqq) {
1573                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1574
1575                 return elv_rb_find(&cfqq->sort_list, sector);
1576         }
1577
1578         return NULL;
1579 }
1580
1581 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1582 {
1583         struct cfq_data *cfqd = q->elevator->elevator_data;
1584
1585         cfqd->rq_in_driver++;
1586         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1587                                                 cfqd->rq_in_driver);
1588
1589         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1590 }
1591
1592 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1593 {
1594         struct cfq_data *cfqd = q->elevator->elevator_data;
1595
1596         WARN_ON(!cfqd->rq_in_driver);
1597         cfqd->rq_in_driver--;
1598         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1599                                                 cfqd->rq_in_driver);
1600 }
1601
1602 static void cfq_remove_request(struct request *rq)
1603 {
1604         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1605
1606         if (cfqq->next_rq == rq)
1607                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1608
1609         list_del_init(&rq->queuelist);
1610         cfq_del_rq_rb(rq);
1611
1612         cfqq->cfqd->rq_queued--;
1613         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1614                                         rq_data_dir(rq), rq_is_sync(rq));
1615         if (rq->cmd_flags & REQ_META) {
1616                 WARN_ON(!cfqq->meta_pending);
1617                 cfqq->meta_pending--;
1618         }
1619 }
1620
1621 static int cfq_merge(struct request_queue *q, struct request **req,
1622                      struct bio *bio)
1623 {
1624         struct cfq_data *cfqd = q->elevator->elevator_data;
1625         struct request *__rq;
1626
1627         __rq = cfq_find_rq_fmerge(cfqd, bio);
1628         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1629                 *req = __rq;
1630                 return ELEVATOR_FRONT_MERGE;
1631         }
1632
1633         return ELEVATOR_NO_MERGE;
1634 }
1635
1636 static void cfq_merged_request(struct request_queue *q, struct request *req,
1637                                int type)
1638 {
1639         if (type == ELEVATOR_FRONT_MERGE) {
1640                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1641
1642                 cfq_reposition_rq_rb(cfqq, req);
1643         }
1644 }
1645
1646 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1647                                 struct bio *bio)
1648 {
1649         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1650                                         bio_data_dir(bio), cfq_bio_sync(bio));
1651 }
1652
1653 static void
1654 cfq_merged_requests(struct request_queue *q, struct request *rq,
1655                     struct request *next)
1656 {
1657         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1658         /*
1659          * reposition in fifo if next is older than rq
1660          */
1661         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1662             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1663                 list_move(&rq->queuelist, &next->queuelist);
1664                 rq_set_fifo_time(rq, rq_fifo_time(next));
1665         }
1666
1667         if (cfqq->next_rq == next)
1668                 cfqq->next_rq = rq;
1669         cfq_remove_request(next);
1670         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1671                                         rq_data_dir(next), rq_is_sync(next));
1672 }
1673
1674 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1675                            struct bio *bio)
1676 {
1677         struct cfq_data *cfqd = q->elevator->elevator_data;
1678         struct cfq_io_context *cic;
1679         struct cfq_queue *cfqq;
1680
1681         /*
1682          * Disallow merge of a sync bio into an async request.
1683          */
1684         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1685                 return false;
1686
1687         /*
1688          * Lookup the cfqq that this bio will be queued with. Allow
1689          * merge only if rq is queued there.
1690          */
1691         cic = cfq_cic_lookup(cfqd, current->io_context);
1692         if (!cic)
1693                 return false;
1694
1695         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1696         return cfqq == RQ_CFQQ(rq);
1697 }
1698
1699 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1700 {
1701         del_timer(&cfqd->idle_slice_timer);
1702         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1703 }
1704
1705 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1706                                    struct cfq_queue *cfqq)
1707 {
1708         if (cfqq) {
1709                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1710                                 cfqd->serving_prio, cfqd->serving_type);
1711                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1712                 cfqq->slice_start = 0;
1713                 cfqq->dispatch_start = jiffies;
1714                 cfqq->allocated_slice = 0;
1715                 cfqq->slice_end = 0;
1716                 cfqq->slice_dispatch = 0;
1717                 cfqq->nr_sectors = 0;
1718
1719                 cfq_clear_cfqq_wait_request(cfqq);
1720                 cfq_clear_cfqq_must_dispatch(cfqq);
1721                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1722                 cfq_clear_cfqq_fifo_expire(cfqq);
1723                 cfq_mark_cfqq_slice_new(cfqq);
1724
1725                 cfq_del_timer(cfqd, cfqq);
1726         }
1727
1728         cfqd->active_queue = cfqq;
1729 }
1730
1731 /*
1732  * current cfqq expired its slice (or was too idle), select new one
1733  */
1734 static void
1735 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1736                     bool timed_out)
1737 {
1738         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1739
1740         if (cfq_cfqq_wait_request(cfqq))
1741                 cfq_del_timer(cfqd, cfqq);
1742
1743         cfq_clear_cfqq_wait_request(cfqq);
1744         cfq_clear_cfqq_wait_busy(cfqq);
1745
1746         /*
1747          * If this cfqq is shared between multiple processes, check to
1748          * make sure that those processes are still issuing I/Os within
1749          * the mean seek distance.  If not, it may be time to break the
1750          * queues apart again.
1751          */
1752         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1753                 cfq_mark_cfqq_split_coop(cfqq);
1754
1755         /*
1756          * store what was left of this slice, if the queue idled/timed out
1757          */
1758         if (timed_out) {
1759                 if (cfq_cfqq_slice_new(cfqq))
1760                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1761                 else
1762                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1763                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1764         }
1765
1766         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1767
1768         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1769                 cfq_del_cfqq_rr(cfqd, cfqq);
1770
1771         cfq_resort_rr_list(cfqd, cfqq);
1772
1773         if (cfqq == cfqd->active_queue)
1774                 cfqd->active_queue = NULL;
1775
1776         if (cfqd->active_cic) {
1777                 put_io_context(cfqd->active_cic->ioc);
1778                 cfqd->active_cic = NULL;
1779         }
1780 }
1781
1782 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1783 {
1784         struct cfq_queue *cfqq = cfqd->active_queue;
1785
1786         if (cfqq)
1787                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1788 }
1789
1790 /*
1791  * Get next queue for service. Unless we have a queue preemption,
1792  * we'll simply select the first cfqq in the service tree.
1793  */
1794 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1795 {
1796         struct cfq_rb_root *service_tree =
1797                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1798                                         cfqd->serving_type);
1799
1800         if (!cfqd->rq_queued)
1801                 return NULL;
1802
1803         /* There is nothing to dispatch */
1804         if (!service_tree)
1805                 return NULL;
1806         if (RB_EMPTY_ROOT(&service_tree->rb))
1807                 return NULL;
1808         return cfq_rb_first(service_tree);
1809 }
1810
1811 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1812 {
1813         struct cfq_group *cfqg;
1814         struct cfq_queue *cfqq;
1815         int i, j;
1816         struct cfq_rb_root *st;
1817
1818         if (!cfqd->rq_queued)
1819                 return NULL;
1820
1821         cfqg = cfq_get_next_cfqg(cfqd);
1822         if (!cfqg)
1823                 return NULL;
1824
1825         for_each_cfqg_st(cfqg, i, j, st)
1826                 if ((cfqq = cfq_rb_first(st)) != NULL)
1827                         return cfqq;
1828         return NULL;
1829 }
1830
1831 /*
1832  * Get and set a new active queue for service.
1833  */
1834 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1835                                               struct cfq_queue *cfqq)
1836 {
1837         if (!cfqq)
1838                 cfqq = cfq_get_next_queue(cfqd);
1839
1840         __cfq_set_active_queue(cfqd, cfqq);
1841         return cfqq;
1842 }
1843
1844 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1845                                           struct request *rq)
1846 {
1847         if (blk_rq_pos(rq) >= cfqd->last_position)
1848                 return blk_rq_pos(rq) - cfqd->last_position;
1849         else
1850                 return cfqd->last_position - blk_rq_pos(rq);
1851 }
1852
1853 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1854                                struct request *rq)
1855 {
1856         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1857 }
1858
1859 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1860                                     struct cfq_queue *cur_cfqq)
1861 {
1862         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1863         struct rb_node *parent, *node;
1864         struct cfq_queue *__cfqq;
1865         sector_t sector = cfqd->last_position;
1866
1867         if (RB_EMPTY_ROOT(root))
1868                 return NULL;
1869
1870         /*
1871          * First, if we find a request starting at the end of the last
1872          * request, choose it.
1873          */
1874         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1875         if (__cfqq)
1876                 return __cfqq;
1877
1878         /*
1879          * If the exact sector wasn't found, the parent of the NULL leaf
1880          * will contain the closest sector.
1881          */
1882         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1883         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1884                 return __cfqq;
1885
1886         if (blk_rq_pos(__cfqq->next_rq) < sector)
1887                 node = rb_next(&__cfqq->p_node);
1888         else
1889                 node = rb_prev(&__cfqq->p_node);
1890         if (!node)
1891                 return NULL;
1892
1893         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1894         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1895                 return __cfqq;
1896
1897         return NULL;
1898 }
1899
1900 /*
1901  * cfqd - obvious
1902  * cur_cfqq - passed in so that we don't decide that the current queue is
1903  *            closely cooperating with itself.
1904  *
1905  * So, basically we're assuming that that cur_cfqq has dispatched at least
1906  * one request, and that cfqd->last_position reflects a position on the disk
1907  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1908  * assumption.
1909  */
1910 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1911                                               struct cfq_queue *cur_cfqq)
1912 {
1913         struct cfq_queue *cfqq;
1914
1915         if (cfq_class_idle(cur_cfqq))
1916                 return NULL;
1917         if (!cfq_cfqq_sync(cur_cfqq))
1918                 return NULL;
1919         if (CFQQ_SEEKY(cur_cfqq))
1920                 return NULL;
1921
1922         /*
1923          * Don't search priority tree if it's the only queue in the group.
1924          */
1925         if (cur_cfqq->cfqg->nr_cfqq == 1)
1926                 return NULL;
1927
1928         /*
1929          * We should notice if some of the queues are cooperating, eg
1930          * working closely on the same area of the disk. In that case,
1931          * we can group them together and don't waste time idling.
1932          */
1933         cfqq = cfqq_close(cfqd, cur_cfqq);
1934         if (!cfqq)
1935                 return NULL;
1936
1937         /* If new queue belongs to different cfq_group, don't choose it */
1938         if (cur_cfqq->cfqg != cfqq->cfqg)
1939                 return NULL;
1940
1941         /*
1942          * It only makes sense to merge sync queues.
1943          */
1944         if (!cfq_cfqq_sync(cfqq))
1945                 return NULL;
1946         if (CFQQ_SEEKY(cfqq))
1947                 return NULL;
1948
1949         /*
1950          * Do not merge queues of different priority classes
1951          */
1952         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1953                 return NULL;
1954
1955         return cfqq;
1956 }
1957
1958 /*
1959  * Determine whether we should enforce idle window for this queue.
1960  */
1961
1962 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1963 {
1964         enum wl_prio_t prio = cfqq_prio(cfqq);
1965         struct cfq_rb_root *service_tree = cfqq->service_tree;
1966
1967         BUG_ON(!service_tree);
1968         BUG_ON(!service_tree->count);
1969
1970         if (!cfqd->cfq_slice_idle)
1971                 return false;
1972
1973         /* We never do for idle class queues. */
1974         if (prio == IDLE_WORKLOAD)
1975                 return false;
1976
1977         /* We do for queues that were marked with idle window flag. */
1978         if (cfq_cfqq_idle_window(cfqq) &&
1979            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1980                 return true;
1981
1982         /*
1983          * Otherwise, we do only if they are the last ones
1984          * in their service tree.
1985          */
1986         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1987            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1988                 return true;
1989         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1990                         service_tree->count);
1991         return false;
1992 }
1993
1994 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1995 {
1996         struct cfq_queue *cfqq = cfqd->active_queue;
1997         struct cfq_io_context *cic;
1998         unsigned long sl, group_idle = 0;
1999
2000         /*
2001          * SSD device without seek penalty, disable idling. But only do so
2002          * for devices that support queuing, otherwise we still have a problem
2003          * with sync vs async workloads.
2004          */
2005         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
2006                 return;
2007
2008         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2009         WARN_ON(cfq_cfqq_slice_new(cfqq));
2010
2011         /*
2012          * idle is disabled, either manually or by past process history
2013          */
2014         if (!cfq_should_idle(cfqd, cfqq)) {
2015                 /* no queue idling. Check for group idling */
2016                 if (cfqd->cfq_group_idle)
2017                         group_idle = cfqd->cfq_group_idle;
2018                 else
2019                         return;
2020         }
2021
2022         /*
2023          * still active requests from this queue, don't idle
2024          */
2025         if (cfqq->dispatched)
2026                 return;
2027
2028         /*
2029          * task has exited, don't wait
2030          */
2031         cic = cfqd->active_cic;
2032         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2033                 return;
2034
2035         /*
2036          * If our average think time is larger than the remaining time
2037          * slice, then don't idle. This avoids overrunning the allotted
2038          * time slice.
2039          */
2040         if (sample_valid(cic->ttime.ttime_samples) &&
2041             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2042                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2043                              cic->ttime.ttime_mean);
2044                 return;
2045         }
2046
2047         /* There are other queues in the group, don't do group idle */
2048         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2049                 return;
2050
2051         cfq_mark_cfqq_wait_request(cfqq);
2052
2053         if (group_idle)
2054                 sl = cfqd->cfq_group_idle;
2055         else
2056                 sl = cfqd->cfq_slice_idle;
2057
2058         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2059         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2060         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2061                         group_idle ? 1 : 0);
2062 }
2063
2064 /*
2065  * Move request from internal lists to the request queue dispatch list.
2066  */
2067 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2068 {
2069         struct cfq_data *cfqd = q->elevator->elevator_data;
2070         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2071
2072         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2073
2074         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2075         cfq_remove_request(rq);
2076         cfqq->dispatched++;
2077         (RQ_CFQG(rq))->dispatched++;
2078         elv_dispatch_sort(q, rq);
2079
2080         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2081         cfqq->nr_sectors += blk_rq_sectors(rq);
2082         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2083                                         rq_data_dir(rq), rq_is_sync(rq));
2084 }
2085
2086 /*
2087  * return expired entry, or NULL to just start from scratch in rbtree
2088  */
2089 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2090 {
2091         struct request *rq = NULL;
2092
2093         if (cfq_cfqq_fifo_expire(cfqq))
2094                 return NULL;
2095
2096         cfq_mark_cfqq_fifo_expire(cfqq);
2097
2098         if (list_empty(&cfqq->fifo))
2099                 return NULL;
2100
2101         rq = rq_entry_fifo(cfqq->fifo.next);
2102         if (time_before(jiffies, rq_fifo_time(rq)))
2103                 rq = NULL;
2104
2105         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2106         return rq;
2107 }
2108
2109 static inline int
2110 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2111 {
2112         const int base_rq = cfqd->cfq_slice_async_rq;
2113
2114         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2115
2116         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2117 }
2118
2119 /*
2120  * Must be called with the queue_lock held.
2121  */
2122 static int cfqq_process_refs(struct cfq_queue *cfqq)
2123 {
2124         int process_refs, io_refs;
2125
2126         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2127         process_refs = cfqq->ref - io_refs;
2128         BUG_ON(process_refs < 0);
2129         return process_refs;
2130 }
2131
2132 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2133 {
2134         int process_refs, new_process_refs;
2135         struct cfq_queue *__cfqq;
2136
2137         /*
2138          * If there are no process references on the new_cfqq, then it is
2139          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2140          * chain may have dropped their last reference (not just their
2141          * last process reference).
2142          */
2143         if (!cfqq_process_refs(new_cfqq))
2144                 return;
2145
2146         /* Avoid a circular list and skip interim queue merges */
2147         while ((__cfqq = new_cfqq->new_cfqq)) {
2148                 if (__cfqq == cfqq)
2149                         return;
2150                 new_cfqq = __cfqq;
2151         }
2152
2153         process_refs = cfqq_process_refs(cfqq);
2154         new_process_refs = cfqq_process_refs(new_cfqq);
2155         /*
2156          * If the process for the cfqq has gone away, there is no
2157          * sense in merging the queues.
2158          */
2159         if (process_refs == 0 || new_process_refs == 0)
2160                 return;
2161
2162         /*
2163          * Merge in the direction of the lesser amount of work.
2164          */
2165         if (new_process_refs >= process_refs) {
2166                 cfqq->new_cfqq = new_cfqq;
2167                 new_cfqq->ref += process_refs;
2168         } else {
2169                 new_cfqq->new_cfqq = cfqq;
2170                 cfqq->ref += new_process_refs;
2171         }
2172 }
2173
2174 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2175                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2176 {
2177         struct cfq_queue *queue;
2178         int i;
2179         bool key_valid = false;
2180         unsigned long lowest_key = 0;
2181         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2182
2183         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2184                 /* select the one with lowest rb_key */
2185                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2186                 if (queue &&
2187                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2188                         lowest_key = queue->rb_key;
2189                         cur_best = i;
2190                         key_valid = true;
2191                 }
2192         }
2193
2194         return cur_best;
2195 }
2196
2197 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2198 {
2199         unsigned slice;
2200         unsigned count;
2201         struct cfq_rb_root *st;
2202         unsigned group_slice;
2203         enum wl_prio_t original_prio = cfqd->serving_prio;
2204
2205         /* Choose next priority. RT > BE > IDLE */
2206         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2207                 cfqd->serving_prio = RT_WORKLOAD;
2208         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2209                 cfqd->serving_prio = BE_WORKLOAD;
2210         else {
2211                 cfqd->serving_prio = IDLE_WORKLOAD;
2212                 cfqd->workload_expires = jiffies + 1;
2213                 return;
2214         }
2215
2216         if (original_prio != cfqd->serving_prio)
2217                 goto new_workload;
2218
2219         /*
2220          * For RT and BE, we have to choose also the type
2221          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2222          * expiration time
2223          */
2224         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2225         count = st->count;
2226
2227         /*
2228          * check workload expiration, and that we still have other queues ready
2229          */
2230         if (count && !time_after(jiffies, cfqd->workload_expires))
2231                 return;
2232
2233 new_workload:
2234         /* otherwise select new workload type */
2235         cfqd->serving_type =
2236                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2237         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2238         count = st->count;
2239
2240         /*
2241          * the workload slice is computed as a fraction of target latency
2242          * proportional to the number of queues in that workload, over
2243          * all the queues in the same priority class
2244          */
2245         group_slice = cfq_group_slice(cfqd, cfqg);
2246
2247         slice = group_slice * count /
2248                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2249                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2250
2251         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2252                 unsigned int tmp;
2253
2254                 /*
2255                  * Async queues are currently system wide. Just taking
2256                  * proportion of queues with-in same group will lead to higher
2257                  * async ratio system wide as generally root group is going
2258                  * to have higher weight. A more accurate thing would be to
2259                  * calculate system wide asnc/sync ratio.
2260                  */
2261                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2262                 tmp = tmp/cfqd->busy_queues;
2263                 slice = min_t(unsigned, slice, tmp);
2264
2265                 /* async workload slice is scaled down according to
2266                  * the sync/async slice ratio. */
2267                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2268         } else
2269                 /* sync workload slice is at least 2 * cfq_slice_idle */
2270                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2271
2272         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2273         cfq_log(cfqd, "workload slice:%d", slice);
2274         cfqd->workload_expires = jiffies + slice;
2275 }
2276
2277 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2278 {
2279         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2280         struct cfq_group *cfqg;
2281
2282         if (RB_EMPTY_ROOT(&st->rb))
2283                 return NULL;
2284         cfqg = cfq_rb_first_group(st);
2285         update_min_vdisktime(st);
2286         return cfqg;
2287 }
2288
2289 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2290 {
2291         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2292
2293         cfqd->serving_group = cfqg;
2294
2295         /* Restore the workload type data */
2296         if (cfqg->saved_workload_slice) {
2297                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2298                 cfqd->serving_type = cfqg->saved_workload;
2299                 cfqd->serving_prio = cfqg->saved_serving_prio;
2300         } else
2301                 cfqd->workload_expires = jiffies - 1;
2302
2303         choose_service_tree(cfqd, cfqg);
2304 }
2305
2306 /*
2307  * Select a queue for service. If we have a current active queue,
2308  * check whether to continue servicing it, or retrieve and set a new one.
2309  */
2310 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2311 {
2312         struct cfq_queue *cfqq, *new_cfqq = NULL;
2313
2314         cfqq = cfqd->active_queue;
2315         if (!cfqq)
2316                 goto new_queue;
2317
2318         if (!cfqd->rq_queued)
2319                 return NULL;
2320
2321         /*
2322          * We were waiting for group to get backlogged. Expire the queue
2323          */
2324         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2325                 goto expire;
2326
2327         /*
2328          * The active queue has run out of time, expire it and select new.
2329          */
2330         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2331                 /*
2332                  * If slice had not expired at the completion of last request
2333                  * we might not have turned on wait_busy flag. Don't expire
2334                  * the queue yet. Allow the group to get backlogged.
2335                  *
2336                  * The very fact that we have used the slice, that means we
2337                  * have been idling all along on this queue and it should be
2338                  * ok to wait for this request to complete.
2339                  */
2340                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2341                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2342                         cfqq = NULL;
2343                         goto keep_queue;
2344                 } else
2345                         goto check_group_idle;
2346         }
2347
2348         /*
2349          * The active queue has requests and isn't expired, allow it to
2350          * dispatch.
2351          */
2352         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2353                 goto keep_queue;
2354
2355         /*
2356          * If another queue has a request waiting within our mean seek
2357          * distance, let it run.  The expire code will check for close
2358          * cooperators and put the close queue at the front of the service
2359          * tree.  If possible, merge the expiring queue with the new cfqq.
2360          */
2361         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2362         if (new_cfqq) {
2363                 if (!cfqq->new_cfqq)
2364                         cfq_setup_merge(cfqq, new_cfqq);
2365                 goto expire;
2366         }
2367
2368         /*
2369          * No requests pending. If the active queue still has requests in
2370          * flight or is idling for a new request, allow either of these
2371          * conditions to happen (or time out) before selecting a new queue.
2372          */
2373         if (timer_pending(&cfqd->idle_slice_timer)) {
2374                 cfqq = NULL;
2375                 goto keep_queue;
2376         }
2377
2378         /*
2379          * This is a deep seek queue, but the device is much faster than
2380          * the queue can deliver, don't idle
2381          **/
2382         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2383             (cfq_cfqq_slice_new(cfqq) ||
2384             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2385                 cfq_clear_cfqq_deep(cfqq);
2386                 cfq_clear_cfqq_idle_window(cfqq);
2387         }
2388
2389         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2390                 cfqq = NULL;
2391                 goto keep_queue;
2392         }
2393
2394         /*
2395          * If group idle is enabled and there are requests dispatched from
2396          * this group, wait for requests to complete.
2397          */
2398 check_group_idle:
2399         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2400             cfqq->cfqg->dispatched &&
2401             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2402                 cfqq = NULL;
2403                 goto keep_queue;
2404         }
2405
2406 expire:
2407         cfq_slice_expired(cfqd, 0);
2408 new_queue:
2409         /*
2410          * Current queue expired. Check if we have to switch to a new
2411          * service tree
2412          */
2413         if (!new_cfqq)
2414                 cfq_choose_cfqg(cfqd);
2415
2416         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2417 keep_queue:
2418         return cfqq;
2419 }
2420
2421 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2422 {
2423         int dispatched = 0;
2424
2425         while (cfqq->next_rq) {
2426                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2427                 dispatched++;
2428         }
2429
2430         BUG_ON(!list_empty(&cfqq->fifo));
2431
2432         /* By default cfqq is not expired if it is empty. Do it explicitly */
2433         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2434         return dispatched;
2435 }
2436
2437 /*
2438  * Drain our current requests. Used for barriers and when switching
2439  * io schedulers on-the-fly.
2440  */
2441 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2442 {
2443         struct cfq_queue *cfqq;
2444         int dispatched = 0;
2445
2446         /* Expire the timeslice of the current active queue first */
2447         cfq_slice_expired(cfqd, 0);
2448         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2449                 __cfq_set_active_queue(cfqd, cfqq);
2450                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2451         }
2452
2453         BUG_ON(cfqd->busy_queues);
2454
2455         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2456         return dispatched;
2457 }
2458
2459 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2460         struct cfq_queue *cfqq)
2461 {
2462         /* the queue hasn't finished any request, can't estimate */
2463         if (cfq_cfqq_slice_new(cfqq))
2464                 return true;
2465         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2466                 cfqq->slice_end))
2467                 return true;
2468
2469         return false;
2470 }
2471
2472 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2473 {
2474         unsigned int max_dispatch;
2475
2476         /*
2477          * Drain async requests before we start sync IO
2478          */
2479         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2480                 return false;
2481
2482         /*
2483          * If this is an async queue and we have sync IO in flight, let it wait
2484          */
2485         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2486                 return false;
2487
2488         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2489         if (cfq_class_idle(cfqq))
2490                 max_dispatch = 1;
2491
2492         /*
2493          * Does this cfqq already have too much IO in flight?
2494          */
2495         if (cfqq->dispatched >= max_dispatch) {
2496                 bool promote_sync = false;
2497                 /*
2498                  * idle queue must always only have a single IO in flight
2499                  */
2500                 if (cfq_class_idle(cfqq))
2501                         return false;
2502
2503                 /*
2504                  * If there is only one sync queue
2505                  * we can ignore async queue here and give the sync
2506                  * queue no dispatch limit. The reason is a sync queue can
2507                  * preempt async queue, limiting the sync queue doesn't make
2508                  * sense. This is useful for aiostress test.
2509                  */
2510                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2511                         promote_sync = true;
2512
2513                 /*
2514                  * We have other queues, don't allow more IO from this one
2515                  */
2516                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2517                                 !promote_sync)
2518                         return false;
2519
2520                 /*
2521                  * Sole queue user, no limit
2522                  */
2523                 if (cfqd->busy_queues == 1 || promote_sync)
2524                         max_dispatch = -1;
2525                 else
2526                         /*
2527                          * Normally we start throttling cfqq when cfq_quantum/2
2528                          * requests have been dispatched. But we can drive
2529                          * deeper queue depths at the beginning of slice
2530                          * subjected to upper limit of cfq_quantum.
2531                          * */
2532                         max_dispatch = cfqd->cfq_quantum;
2533         }
2534
2535         /*
2536          * Async queues must wait a bit before being allowed dispatch.
2537          * We also ramp up the dispatch depth gradually for async IO,
2538          * based on the last sync IO we serviced
2539          */
2540         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2541                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2542                 unsigned int depth;
2543
2544                 depth = last_sync / cfqd->cfq_slice[1];
2545                 if (!depth && !cfqq->dispatched)
2546                         depth = 1;
2547                 if (depth < max_dispatch)
2548                         max_dispatch = depth;
2549         }
2550
2551         /*
2552          * If we're below the current max, allow a dispatch
2553          */
2554         return cfqq->dispatched < max_dispatch;
2555 }
2556
2557 /*
2558  * Dispatch a request from cfqq, moving them to the request queue
2559  * dispatch list.
2560  */
2561 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2562 {
2563         struct request *rq;
2564
2565         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2566
2567         if (!cfq_may_dispatch(cfqd, cfqq))
2568                 return false;
2569
2570         /*
2571          * follow expired path, else get first next available
2572          */
2573         rq = cfq_check_fifo(cfqq);
2574         if (!rq)
2575                 rq = cfqq->next_rq;
2576
2577         /*
2578          * insert request into driver dispatch list
2579          */
2580         cfq_dispatch_insert(cfqd->queue, rq);
2581
2582         if (!cfqd->active_cic) {
2583                 struct cfq_io_context *cic = RQ_CIC(rq);
2584
2585                 atomic_long_inc(&cic->ioc->refcount);
2586                 cfqd->active_cic = cic;
2587         }
2588
2589         return true;
2590 }
2591
2592 /*
2593  * Find the cfqq that we need to service and move a request from that to the
2594  * dispatch list
2595  */
2596 static int cfq_dispatch_requests(struct request_queue *q, int force)
2597 {
2598         struct cfq_data *cfqd = q->elevator->elevator_data;
2599         struct cfq_queue *cfqq;
2600
2601         if (!cfqd->busy_queues)
2602                 return 0;
2603
2604         if (unlikely(force))
2605                 return cfq_forced_dispatch(cfqd);
2606
2607         cfqq = cfq_select_queue(cfqd);
2608         if (!cfqq)
2609                 return 0;
2610
2611         /*
2612          * Dispatch a request from this cfqq, if it is allowed
2613          */
2614         if (!cfq_dispatch_request(cfqd, cfqq))
2615                 return 0;
2616
2617         cfqq->slice_dispatch++;
2618         cfq_clear_cfqq_must_dispatch(cfqq);
2619
2620         /*
2621          * expire an async queue immediately if it has used up its slice. idle
2622          * queue always expire after 1 dispatch round.
2623          */
2624         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2625             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2626             cfq_class_idle(cfqq))) {
2627                 cfqq->slice_end = jiffies + 1;
2628                 cfq_slice_expired(cfqd, 0);
2629         }
2630
2631         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2632         return 1;
2633 }
2634
2635 /*
2636  * task holds one reference to the queue, dropped when task exits. each rq
2637  * in-flight on this queue also holds a reference, dropped when rq is freed.
2638  *
2639  * Each cfq queue took a reference on the parent group. Drop it now.
2640  * queue lock must be held here.
2641  */
2642 static void cfq_put_queue(struct cfq_queue *cfqq)
2643 {
2644         struct cfq_data *cfqd = cfqq->cfqd;
2645         struct cfq_group *cfqg;
2646
2647         BUG_ON(cfqq->ref <= 0);
2648
2649         cfqq->ref--;
2650         if (cfqq->ref)
2651                 return;
2652
2653         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2654         BUG_ON(rb_first(&cfqq->sort_list));
2655         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2656         cfqg = cfqq->cfqg;
2657
2658         if (unlikely(cfqd->active_queue == cfqq)) {
2659                 __cfq_slice_expired(cfqd, cfqq, 0);
2660                 cfq_schedule_dispatch(cfqd);
2661         }
2662
2663         BUG_ON(cfq_cfqq_on_rr(cfqq));
2664         kmem_cache_free(cfq_pool, cfqq);
2665         cfq_put_cfqg(cfqg);
2666 }
2667
2668 /*
2669  * Call func for each cic attached to this ioc.
2670  */
2671 static void
2672 call_for_each_cic(struct io_context *ioc,
2673                   void (*func)(struct io_context *, struct cfq_io_context *))
2674 {
2675         struct cfq_io_context *cic;
2676         struct hlist_node *n;
2677
2678         rcu_read_lock();
2679
2680         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2681                 func(ioc, cic);
2682
2683         rcu_read_unlock();
2684 }
2685
2686 static void cfq_cic_free_rcu(struct rcu_head *head)
2687 {
2688         struct cfq_io_context *cic;
2689
2690         cic = container_of(head, struct cfq_io_context, rcu_head);
2691
2692         kmem_cache_free(cfq_ioc_pool, cic);
2693         elv_ioc_count_dec(cfq_ioc_count);
2694
2695         if (ioc_gone) {
2696                 /*
2697                  * CFQ scheduler is exiting, grab exit lock and check
2698                  * the pending io context count. If it hits zero,
2699                  * complete ioc_gone and set it back to NULL
2700                  */
2701                 spin_lock(&ioc_gone_lock);
2702                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2703                         complete(ioc_gone);
2704                         ioc_gone = NULL;
2705                 }
2706                 spin_unlock(&ioc_gone_lock);
2707         }
2708 }
2709
2710 static void cfq_cic_free(struct cfq_io_context *cic)
2711 {
2712         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2713 }
2714
2715 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2716 {
2717         unsigned long flags;
2718         unsigned long dead_key = (unsigned long) cic->key;
2719
2720         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2721
2722         spin_lock_irqsave(&ioc->lock, flags);
2723         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2724         hlist_del_rcu(&cic->cic_list);
2725         spin_unlock_irqrestore(&ioc->lock, flags);
2726
2727         cfq_cic_free(cic);
2728 }
2729
2730 /*
2731  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2732  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2733  * and ->trim() which is called with the task lock held
2734  */
2735 static void cfq_free_io_context(struct io_context *ioc)
2736 {
2737         /*
2738          * ioc->refcount is zero here, or we are called from elv_unregister(),
2739          * so no more cic's are allowed to be linked into this ioc.  So it
2740          * should be ok to iterate over the known list, we will see all cic's
2741          * since no new ones are added.
2742          */
2743         call_for_each_cic(ioc, cic_free_func);
2744 }
2745
2746 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2747 {
2748         struct cfq_queue *__cfqq, *next;
2749
2750         /*
2751          * If this queue was scheduled to merge with another queue, be
2752          * sure to drop the reference taken on that queue (and others in
2753          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2754          */
2755         __cfqq = cfqq->new_cfqq;
2756         while (__cfqq) {
2757                 if (__cfqq == cfqq) {
2758                         WARN(1, "cfqq->new_cfqq loop detected\n");
2759                         break;
2760                 }
2761                 next = __cfqq->new_cfqq;
2762                 cfq_put_queue(__cfqq);
2763                 __cfqq = next;
2764         }
2765 }
2766
2767 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2768 {
2769         if (unlikely(cfqq == cfqd->active_queue)) {
2770                 __cfq_slice_expired(cfqd, cfqq, 0);
2771                 cfq_schedule_dispatch(cfqd);
2772         }
2773
2774         cfq_put_cooperator(cfqq);
2775
2776         cfq_put_queue(cfqq);
2777 }
2778
2779 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2780                                          struct cfq_io_context *cic)
2781 {
2782         struct io_context *ioc = cic->ioc;
2783
2784         list_del_init(&cic->queue_list);
2785
2786         /*
2787          * Make sure dead mark is seen for dead queues
2788          */
2789         smp_wmb();
2790         cic->key = cfqd_dead_key(cfqd);
2791
2792         rcu_read_lock();
2793         if (rcu_dereference(ioc->ioc_data) == cic) {
2794                 rcu_read_unlock();
2795                 spin_lock(&ioc->lock);
2796                 rcu_assign_pointer(ioc->ioc_data, NULL);
2797                 spin_unlock(&ioc->lock);
2798         } else
2799                 rcu_read_unlock();
2800
2801         if (cic->cfqq[BLK_RW_ASYNC]) {
2802                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2803                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2804         }
2805
2806         if (cic->cfqq[BLK_RW_SYNC]) {
2807                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2808                 cic->cfqq[BLK_RW_SYNC] = NULL;
2809         }
2810 }
2811
2812 static void cfq_exit_single_io_context(struct io_context *ioc,
2813                                        struct cfq_io_context *cic)
2814 {
2815         struct cfq_data *cfqd = cic_to_cfqd(cic);
2816
2817         if (cfqd) {
2818                 struct request_queue *q = cfqd->queue;
2819                 unsigned long flags;
2820
2821                 spin_lock_irqsave(q->queue_lock, flags);
2822
2823                 /*
2824                  * Ensure we get a fresh copy of the ->key to prevent
2825                  * race between exiting task and queue
2826                  */
2827                 smp_read_barrier_depends();
2828                 if (cic->key == cfqd)
2829                         __cfq_exit_single_io_context(cfqd, cic);
2830
2831                 spin_unlock_irqrestore(q->queue_lock, flags);
2832         }
2833 }
2834
2835 /*
2836  * The process that ioc belongs to has exited, we need to clean up
2837  * and put the internal structures we have that belongs to that process.
2838  */
2839 static void cfq_exit_io_context(struct io_context *ioc)
2840 {
2841         call_for_each_cic(ioc, cfq_exit_single_io_context);
2842 }
2843
2844 static struct cfq_io_context *
2845 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2846 {
2847         struct cfq_io_context *cic;
2848
2849         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2850                                                         cfqd->queue->node);
2851         if (cic) {
2852                 cic->ttime.last_end_request = jiffies;
2853                 INIT_LIST_HEAD(&cic->queue_list);
2854                 INIT_HLIST_NODE(&cic->cic_list);
2855                 cic->dtor = cfq_free_io_context;
2856                 cic->exit = cfq_exit_io_context;
2857                 elv_ioc_count_inc(cfq_ioc_count);
2858         }
2859
2860         return cic;
2861 }
2862
2863 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2864 {
2865         struct task_struct *tsk = current;
2866         int ioprio_class;
2867
2868         if (!cfq_cfqq_prio_changed(cfqq))
2869                 return;
2870
2871         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2872         switch (ioprio_class) {
2873         default:
2874                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2875         case IOPRIO_CLASS_NONE:
2876                 /*
2877                  * no prio set, inherit CPU scheduling settings
2878                  */
2879                 cfqq->ioprio = task_nice_ioprio(tsk);
2880                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2881                 break;
2882         case IOPRIO_CLASS_RT:
2883                 cfqq->ioprio = task_ioprio(ioc);
2884                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2885                 break;
2886         case IOPRIO_CLASS_BE:
2887                 cfqq->ioprio = task_ioprio(ioc);
2888                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2889                 break;
2890         case IOPRIO_CLASS_IDLE:
2891                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2892                 cfqq->ioprio = 7;
2893                 cfq_clear_cfqq_idle_window(cfqq);
2894                 break;
2895         }
2896
2897         /*
2898          * keep track of original prio settings in case we have to temporarily
2899          * elevate the priority of this queue
2900          */
2901         cfqq->org_ioprio = cfqq->ioprio;
2902         cfq_clear_cfqq_prio_changed(cfqq);
2903 }
2904
2905 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2906 {
2907         struct cfq_data *cfqd = cic_to_cfqd(cic);
2908         struct cfq_queue *cfqq;
2909         unsigned long flags;
2910
2911         if (unlikely(!cfqd))
2912                 return;
2913
2914         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2915
2916         cfqq = cic->cfqq[BLK_RW_ASYNC];
2917         if (cfqq) {
2918                 struct cfq_queue *new_cfqq;
2919                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2920                                                 GFP_ATOMIC);
2921                 if (new_cfqq) {
2922                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2923                         cfq_put_queue(cfqq);
2924                 }
2925         }
2926
2927         cfqq = cic->cfqq[BLK_RW_SYNC];
2928         if (cfqq)
2929                 cfq_mark_cfqq_prio_changed(cfqq);
2930
2931         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2932 }
2933
2934 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2935 {
2936         call_for_each_cic(ioc, changed_ioprio);
2937         ioc->ioprio_changed = 0;
2938 }
2939
2940 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2941                           pid_t pid, bool is_sync)
2942 {
2943         RB_CLEAR_NODE(&cfqq->rb_node);
2944         RB_CLEAR_NODE(&cfqq->p_node);
2945         INIT_LIST_HEAD(&cfqq->fifo);
2946
2947         cfqq->ref = 0;
2948         cfqq->cfqd = cfqd;
2949
2950         cfq_mark_cfqq_prio_changed(cfqq);
2951
2952         if (is_sync) {
2953                 if (!cfq_class_idle(cfqq))
2954                         cfq_mark_cfqq_idle_window(cfqq);
2955                 cfq_mark_cfqq_sync(cfqq);
2956         }
2957         cfqq->pid = pid;
2958 }
2959
2960 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2961 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2962 {
2963         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2964         struct cfq_data *cfqd = cic_to_cfqd(cic);
2965         unsigned long flags;
2966         struct request_queue *q;
2967
2968         if (unlikely(!cfqd))
2969                 return;
2970
2971         q = cfqd->queue;
2972
2973         spin_lock_irqsave(q->queue_lock, flags);
2974
2975         if (sync_cfqq) {
2976                 /*
2977                  * Drop reference to sync queue. A new sync queue will be
2978                  * assigned in new group upon arrival of a fresh request.
2979                  */
2980                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2981                 cic_set_cfqq(cic, NULL, 1);
2982                 cfq_put_queue(sync_cfqq);
2983         }
2984
2985         spin_unlock_irqrestore(q->queue_lock, flags);
2986 }
2987
2988 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2989 {
2990         call_for_each_cic(ioc, changed_cgroup);
2991         ioc->cgroup_changed = 0;
2992 }
2993 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2994
2995 static struct cfq_queue *
2996 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2997                      struct io_context *ioc, gfp_t gfp_mask)
2998 {
2999         struct cfq_queue *cfqq, *new_cfqq = NULL;
3000         struct cfq_io_context *cic;
3001         struct cfq_group *cfqg;
3002
3003 retry:
3004         cfqg = cfq_get_cfqg(cfqd);
3005         cic = cfq_cic_lookup(cfqd, ioc);
3006         /* cic always exists here */
3007         cfqq = cic_to_cfqq(cic, is_sync);
3008
3009         /*
3010          * Always try a new alloc if we fell back to the OOM cfqq
3011          * originally, since it should just be a temporary situation.
3012          */
3013         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3014                 cfqq = NULL;
3015                 if (new_cfqq) {
3016                         cfqq = new_cfqq;
3017                         new_cfqq = NULL;
3018                 } else if (gfp_mask & __GFP_WAIT) {
3019                         spin_unlock_irq(cfqd->queue->queue_lock);
3020                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3021                                         gfp_mask | __GFP_ZERO,
3022                                         cfqd->queue->node);
3023                         spin_lock_irq(cfqd->queue->queue_lock);
3024                         if (new_cfqq)
3025                                 goto retry;
3026                 } else {
3027                         cfqq = kmem_cache_alloc_node(cfq_pool,
3028                                         gfp_mask | __GFP_ZERO,
3029                                         cfqd->queue->node);
3030                 }
3031
3032                 if (cfqq) {
3033                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3034                         cfq_init_prio_data(cfqq, ioc);
3035                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3036                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3037                 } else
3038                         cfqq = &cfqd->oom_cfqq;
3039         }
3040
3041         if (new_cfqq)
3042                 kmem_cache_free(cfq_pool, new_cfqq);
3043
3044         return cfqq;
3045 }
3046
3047 static struct cfq_queue **
3048 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3049 {
3050         switch (ioprio_class) {
3051         case IOPRIO_CLASS_RT:
3052                 return &cfqd->async_cfqq[0][ioprio];
3053         case IOPRIO_CLASS_BE:
3054                 return &cfqd->async_cfqq[1][ioprio];
3055         case IOPRIO_CLASS_IDLE:
3056                 return &cfqd->async_idle_cfqq;
3057         default:
3058                 BUG();
3059         }
3060 }
3061
3062 static struct cfq_queue *
3063 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3064               gfp_t gfp_mask)
3065 {
3066         const int ioprio = task_ioprio(ioc);
3067         const int ioprio_class = task_ioprio_class(ioc);
3068         struct cfq_queue **async_cfqq = NULL;
3069         struct cfq_queue *cfqq = NULL;
3070
3071         if (!is_sync) {
3072                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3073                 cfqq = *async_cfqq;
3074         }
3075
3076         if (!cfqq)
3077                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3078
3079         /*
3080          * pin the queue now that it's allocated, scheduler exit will prune it
3081          */
3082         if (!is_sync && !(*async_cfqq)) {
3083                 cfqq->ref++;
3084                 *async_cfqq = cfqq;
3085         }
3086
3087         cfqq->ref++;
3088         return cfqq;
3089 }
3090
3091 /*
3092  * We drop cfq io contexts lazily, so we may find a dead one.
3093  */
3094 static void
3095 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3096                   struct cfq_io_context *cic)
3097 {
3098         unsigned long flags;
3099
3100         WARN_ON(!list_empty(&cic->queue_list));
3101         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3102
3103         spin_lock_irqsave(&ioc->lock, flags);
3104
3105         BUG_ON(rcu_dereference_check(ioc->ioc_data,
3106                 lockdep_is_held(&ioc->lock)) == cic);
3107
3108         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3109         hlist_del_rcu(&cic->cic_list);
3110         spin_unlock_irqrestore(&ioc->lock, flags);
3111
3112         cfq_cic_free(cic);
3113 }
3114
3115 static struct cfq_io_context *
3116 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3117 {
3118         struct cfq_io_context *cic;
3119         unsigned long flags;
3120
3121         if (unlikely(!ioc))
3122                 return NULL;
3123
3124         rcu_read_lock();
3125
3126         /*
3127          * we maintain a last-hit cache, to avoid browsing over the tree
3128          */
3129         cic = rcu_dereference(ioc->ioc_data);
3130         if (cic && cic->key == cfqd) {
3131                 rcu_read_unlock();
3132                 return cic;
3133         }
3134
3135         do {
3136                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3137                 rcu_read_unlock();
3138                 if (!cic)
3139                         break;
3140                 if (unlikely(cic->key != cfqd)) {
3141                         cfq_drop_dead_cic(cfqd, ioc, cic);
3142                         rcu_read_lock();
3143                         continue;
3144                 }
3145
3146                 spin_lock_irqsave(&ioc->lock, flags);
3147                 rcu_assign_pointer(ioc->ioc_data, cic);
3148                 spin_unlock_irqrestore(&ioc->lock, flags);
3149                 break;
3150         } while (1);
3151
3152         return cic;
3153 }
3154
3155 /*
3156  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3157  * the process specific cfq io context when entered from the block layer.
3158  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3159  */
3160 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3161                         struct cfq_io_context *cic, gfp_t gfp_mask)
3162 {
3163         unsigned long flags;
3164         int ret;
3165
3166         ret = radix_tree_preload(gfp_mask);
3167         if (!ret) {
3168                 cic->ioc = ioc;
3169                 cic->key = cfqd;
3170
3171                 spin_lock_irqsave(&ioc->lock, flags);
3172                 ret = radix_tree_insert(&ioc->radix_root,
3173                                                 cfqd->cic_index, cic);
3174                 if (!ret)
3175                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3176                 spin_unlock_irqrestore(&ioc->lock, flags);
3177
3178                 radix_tree_preload_end();
3179
3180                 if (!ret) {
3181                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3182                         list_add(&cic->queue_list, &cfqd->cic_list);
3183                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3184                 }
3185         }
3186
3187         if (ret)
3188                 printk(KERN_ERR "cfq: cic link failed!\n");
3189
3190         return ret;
3191 }
3192
3193 /*
3194  * Setup general io context and cfq io context. There can be several cfq
3195  * io contexts per general io context, if this process is doing io to more
3196  * than one device managed by cfq.
3197  */
3198 static struct cfq_io_context *
3199 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3200 {
3201         struct io_context *ioc = NULL;
3202         struct cfq_io_context *cic;
3203
3204         might_sleep_if(gfp_mask & __GFP_WAIT);
3205
3206         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3207         if (!ioc)
3208                 return NULL;
3209
3210         cic = cfq_cic_lookup(cfqd, ioc);
3211         if (cic)
3212                 goto out;
3213
3214         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3215         if (cic == NULL)
3216                 goto err;
3217
3218         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3219                 goto err_free;
3220
3221 out:
3222         smp_read_barrier_depends();
3223         if (unlikely(ioc->ioprio_changed))
3224                 cfq_ioc_set_ioprio(ioc);
3225
3226 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3227         if (unlikely(ioc->cgroup_changed))
3228                 cfq_ioc_set_cgroup(ioc);
3229 #endif
3230         return cic;
3231 err_free:
3232         cfq_cic_free(cic);
3233 err:
3234         put_io_context(ioc);
3235         return NULL;
3236 }
3237
3238 static void
3239 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3240 {
3241         unsigned long elapsed = jiffies - ttime->last_end_request;
3242         elapsed = min(elapsed, 2UL * slice_idle);
3243
3244         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3245         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3246         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3247 }
3248
3249 static void
3250 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3251         struct cfq_io_context *cic)
3252 {
3253         if (cfq_cfqq_sync(cfqq)) {
3254                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3255                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3256                         cfqd->cfq_slice_idle);
3257         }
3258 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3259         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3260 #endif
3261 }
3262
3263 static void
3264 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3265                        struct request *rq)
3266 {
3267         sector_t sdist = 0;
3268         sector_t n_sec = blk_rq_sectors(rq);
3269         if (cfqq->last_request_pos) {
3270                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3271                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3272                 else
3273                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3274         }
3275
3276         cfqq->seek_history <<= 1;
3277         if (blk_queue_nonrot(cfqd->queue))
3278                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3279         else
3280                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3281 }
3282
3283 /*
3284  * Disable idle window if the process thinks too long or seeks so much that
3285  * it doesn't matter
3286  */
3287 static void
3288 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3289                        struct cfq_io_context *cic)
3290 {
3291         int old_idle, enable_idle;
3292
3293         /*
3294          * Don't idle for async or idle io prio class
3295          */
3296         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3297                 return;
3298
3299         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3300
3301         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3302                 cfq_mark_cfqq_deep(cfqq);
3303
3304         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3305                 enable_idle = 0;
3306         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3307             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3308                 enable_idle = 0;
3309         else if (sample_valid(cic->ttime.ttime_samples)) {
3310                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3311                         enable_idle = 0;
3312                 else
3313                         enable_idle = 1;
3314         }
3315
3316         if (old_idle != enable_idle) {
3317                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3318                 if (enable_idle)
3319                         cfq_mark_cfqq_idle_window(cfqq);
3320                 else
3321                         cfq_clear_cfqq_idle_window(cfqq);
3322         }
3323 }
3324
3325 /*
3326  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3327  * no or if we aren't sure, a 1 will cause a preempt.
3328  */
3329 static bool
3330 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3331                    struct request *rq)
3332 {
3333         struct cfq_queue *cfqq;
3334
3335         cfqq = cfqd->active_queue;
3336         if (!cfqq)
3337                 return false;
3338
3339         if (cfq_class_idle(new_cfqq))
3340                 return false;
3341
3342         if (cfq_class_idle(cfqq))
3343                 return true;
3344
3345         /*
3346          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3347          */
3348         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3349                 return false;
3350
3351         /*
3352          * if the new request is sync, but the currently running queue is
3353          * not, let the sync request have priority.
3354          */
3355         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3356                 return true;
3357
3358         if (new_cfqq->cfqg != cfqq->cfqg)
3359                 return false;
3360
3361         if (cfq_slice_used(cfqq))
3362                 return true;
3363
3364         /* Allow preemption only if we are idling on sync-noidle tree */
3365         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3366             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3367             new_cfqq->service_tree->count == 2 &&
3368             RB_EMPTY_ROOT(&cfqq->sort_list))
3369                 return true;
3370
3371         /*
3372          * So both queues are sync. Let the new request get disk time if
3373          * it's a metadata request and the current queue is doing regular IO.
3374          */
3375         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3376                 return true;
3377
3378         /*
3379          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3380          */
3381         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3382                 return true;
3383
3384         /* An idle queue should not be idle now for some reason */
3385         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3386                 return true;
3387
3388         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3389                 return false;
3390
3391         /*
3392          * if this request is as-good as one we would expect from the
3393          * current cfqq, let it preempt
3394          */
3395         if (cfq_rq_close(cfqd, cfqq, rq))
3396                 return true;
3397
3398         return false;
3399 }
3400
3401 /*
3402  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3403  * let it have half of its nominal slice.
3404  */
3405 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3406 {
3407         struct cfq_queue *old_cfqq = cfqd->active_queue;
3408
3409         cfq_log_cfqq(cfqd, cfqq, "preempt");
3410         cfq_slice_expired(cfqd, 1);
3411
3412         /*
3413          * workload type is changed, don't save slice, otherwise preempt
3414          * doesn't happen
3415          */
3416         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3417                 cfqq->cfqg->saved_workload_slice = 0;
3418
3419         /*
3420          * Put the new queue at the front of the of the current list,
3421          * so we know that it will be selected next.
3422          */
3423         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3424
3425         cfq_service_tree_add(cfqd, cfqq, 1);
3426
3427         cfqq->slice_end = 0;
3428         cfq_mark_cfqq_slice_new(cfqq);
3429 }
3430
3431 /*
3432  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3433  * something we should do about it
3434  */
3435 static void
3436 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3437                 struct request *rq)
3438 {
3439         struct cfq_io_context *cic = RQ_CIC(rq);
3440
3441         cfqd->rq_queued++;
3442         if (rq->cmd_flags & REQ_META)
3443                 cfqq->meta_pending++;
3444
3445         cfq_update_io_thinktime(cfqd, cfqq, cic);
3446         cfq_update_io_seektime(cfqd, cfqq, rq);
3447         cfq_update_idle_window(cfqd, cfqq, cic);
3448
3449         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3450
3451         if (cfqq == cfqd->active_queue) {
3452                 /*
3453                  * Remember that we saw a request from this process, but
3454                  * don't start queuing just yet. Otherwise we risk seeing lots
3455                  * of tiny requests, because we disrupt the normal plugging
3456                  * and merging. If the request is already larger than a single
3457                  * page, let it rip immediately. For that case we assume that
3458                  * merging is already done. Ditto for a busy system that
3459                  * has other work pending, don't risk delaying until the
3460                  * idle timer unplug to continue working.
3461                  */
3462                 if (cfq_cfqq_wait_request(cfqq)) {
3463                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3464                             cfqd->busy_queues > 1) {
3465                                 cfq_del_timer(cfqd, cfqq);
3466                                 cfq_clear_cfqq_wait_request(cfqq);
3467                                 __blk_run_queue(cfqd->queue);
3468                         } else {
3469                                 cfq_blkiocg_update_idle_time_stats(
3470                                                 &cfqq->cfqg->blkg);
3471                                 cfq_mark_cfqq_must_dispatch(cfqq);
3472                         }
3473                 }
3474         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3475                 /*
3476                  * not the active queue - expire current slice if it is
3477                  * idle and has expired it's mean thinktime or this new queue
3478                  * has some old slice time left and is of higher priority or
3479                  * this new queue is RT and the current one is BE
3480                  */
3481                 cfq_preempt_queue(cfqd, cfqq);
3482                 __blk_run_queue(cfqd->queue);
3483         }
3484 }
3485
3486 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3487 {
3488         struct cfq_data *cfqd = q->elevator->elevator_data;
3489         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3490
3491         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3492         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3493
3494         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3495         list_add_tail(&rq->queuelist, &cfqq->fifo);
3496         cfq_add_rq_rb(rq);
3497         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3498                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3499                         rq_is_sync(rq));
3500         cfq_rq_enqueued(cfqd, cfqq, rq);
3501 }
3502
3503 /*
3504  * Update hw_tag based on peak queue depth over 50 samples under
3505  * sufficient load.
3506  */
3507 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3508 {
3509         struct cfq_queue *cfqq = cfqd->active_queue;
3510
3511         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3512                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3513
3514         if (cfqd->hw_tag == 1)
3515                 return;
3516
3517         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3518             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3519                 return;
3520
3521         /*
3522          * If active queue hasn't enough requests and can idle, cfq might not
3523          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3524          * case
3525          */
3526         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3527             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3528             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3529                 return;
3530
3531         if (cfqd->hw_tag_samples++ < 50)
3532                 return;
3533
3534         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3535                 cfqd->hw_tag = 1;
3536         else
3537                 cfqd->hw_tag = 0;
3538 }
3539
3540 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3541 {
3542         struct cfq_io_context *cic = cfqd->active_cic;
3543
3544         /* If the queue already has requests, don't wait */
3545         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3546                 return false;
3547
3548         /* If there are other queues in the group, don't wait */
3549         if (cfqq->cfqg->nr_cfqq > 1)
3550                 return false;
3551
3552         /* the only queue in the group, but think time is big */
3553         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3554                 return false;
3555
3556         if (cfq_slice_used(cfqq))
3557                 return true;
3558
3559         /* if slice left is less than think time, wait busy */
3560         if (cic && sample_valid(cic->ttime.ttime_samples)
3561             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3562                 return true;
3563
3564         /*
3565          * If think times is less than a jiffy than ttime_mean=0 and above
3566          * will not be true. It might happen that slice has not expired yet
3567          * but will expire soon (4-5 ns) during select_queue(). To cover the
3568          * case where think time is less than a jiffy, mark the queue wait
3569          * busy if only 1 jiffy is left in the slice.
3570          */
3571         if (cfqq->slice_end - jiffies == 1)
3572                 return true;
3573
3574         return false;
3575 }
3576
3577 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3578 {
3579         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3580         struct cfq_data *cfqd = cfqq->cfqd;
3581         const int sync = rq_is_sync(rq);
3582         unsigned long now;
3583
3584         now = jiffies;
3585         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3586                      !!(rq->cmd_flags & REQ_NOIDLE));
3587
3588         cfq_update_hw_tag(cfqd);
3589
3590         WARN_ON(!cfqd->rq_in_driver);
3591         WARN_ON(!cfqq->dispatched);
3592         cfqd->rq_in_driver--;
3593         cfqq->dispatched--;
3594         (RQ_CFQG(rq))->dispatched--;
3595         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3596                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3597                         rq_data_dir(rq), rq_is_sync(rq));
3598
3599         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3600
3601         if (sync) {
3602                 struct cfq_rb_root *service_tree;
3603
3604                 RQ_CIC(rq)->ttime.last_end_request = now;
3605
3606                 if (cfq_cfqq_on_rr(cfqq))
3607                         service_tree = cfqq->service_tree;
3608                 else
3609                         service_tree = service_tree_for(cfqq->cfqg,
3610                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3611                 service_tree->ttime.last_end_request = now;
3612                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3613                         cfqd->last_delayed_sync = now;
3614         }
3615
3616 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3617         cfqq->cfqg->ttime.last_end_request = now;
3618 #endif
3619
3620         /*
3621          * If this is the active queue, check if it needs to be expired,
3622          * or if we want to idle in case it has no pending requests.
3623          */
3624         if (cfqd->active_queue == cfqq) {
3625                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3626
3627                 if (cfq_cfqq_slice_new(cfqq)) {
3628                         cfq_set_prio_slice(cfqd, cfqq);
3629                         cfq_clear_cfqq_slice_new(cfqq);
3630                 }
3631
3632                 /*
3633                  * Should we wait for next request to come in before we expire
3634                  * the queue.
3635                  */
3636                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3637                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3638                         if (!cfqd->cfq_slice_idle)
3639                                 extend_sl = cfqd->cfq_group_idle;
3640                         cfqq->slice_end = jiffies + extend_sl;
3641                         cfq_mark_cfqq_wait_busy(cfqq);
3642                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3643                 }
3644
3645                 /*
3646                  * Idling is not enabled on:
3647                  * - expired queues
3648                  * - idle-priority queues
3649                  * - async queues
3650                  * - queues with still some requests queued
3651                  * - when there is a close cooperator
3652                  */
3653                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3654                         cfq_slice_expired(cfqd, 1);
3655                 else if (sync && cfqq_empty &&
3656                          !cfq_close_cooperator(cfqd, cfqq)) {
3657                         cfq_arm_slice_timer(cfqd);
3658                 }
3659         }
3660
3661         if (!cfqd->rq_in_driver)
3662                 cfq_schedule_dispatch(cfqd);
3663 }
3664
3665 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3666 {
3667         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3668                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3669                 return ELV_MQUEUE_MUST;
3670         }
3671
3672         return ELV_MQUEUE_MAY;
3673 }
3674
3675 static int cfq_may_queue(struct request_queue *q, int rw)
3676 {
3677         struct cfq_data *cfqd = q->elevator->elevator_data;
3678         struct task_struct *tsk = current;
3679         struct cfq_io_context *cic;
3680         struct cfq_queue *cfqq;
3681
3682         /*
3683          * don't force setup of a queue from here, as a call to may_queue
3684          * does not necessarily imply that a request actually will be queued.
3685          * so just lookup a possibly existing queue, or return 'may queue'
3686          * if that fails
3687          */
3688         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3689         if (!cic)
3690                 return ELV_MQUEUE_MAY;
3691
3692         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3693         if (cfqq) {
3694                 cfq_init_prio_data(cfqq, cic->ioc);
3695
3696                 return __cfq_may_queue(cfqq);
3697         }
3698
3699         return ELV_MQUEUE_MAY;
3700 }
3701
3702 /*
3703  * queue lock held here
3704  */
3705 static void cfq_put_request(struct request *rq)
3706 {
3707         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3708
3709         if (cfqq) {
3710                 const int rw = rq_data_dir(rq);
3711
3712                 BUG_ON(!cfqq->allocated[rw]);
3713                 cfqq->allocated[rw]--;
3714
3715                 put_io_context(RQ_CIC(rq)->ioc);
3716
3717                 rq->elevator_private[0] = NULL;
3718                 rq->elevator_private[1] = NULL;
3719
3720                 /* Put down rq reference on cfqg */
3721                 cfq_put_cfqg(RQ_CFQG(rq));
3722                 rq->elevator_private[2] = NULL;
3723
3724                 cfq_put_queue(cfqq);
3725         }
3726 }
3727
3728 static struct cfq_queue *
3729 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3730                 struct cfq_queue *cfqq)
3731 {
3732         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3733         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3734         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3735         cfq_put_queue(cfqq);
3736         return cic_to_cfqq(cic, 1);
3737 }
3738
3739 /*
3740  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3741  * was the last process referring to said cfqq.
3742  */
3743 static struct cfq_queue *
3744 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3745 {
3746         if (cfqq_process_refs(cfqq) == 1) {
3747                 cfqq->pid = current->pid;
3748                 cfq_clear_cfqq_coop(cfqq);
3749                 cfq_clear_cfqq_split_coop(cfqq);
3750                 return cfqq;
3751         }
3752
3753         cic_set_cfqq(cic, NULL, 1);
3754
3755         cfq_put_cooperator(cfqq);
3756
3757         cfq_put_queue(cfqq);
3758         return NULL;
3759 }
3760 /*
3761  * Allocate cfq data structures associated with this request.
3762  */
3763 static int
3764 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3765 {
3766         struct cfq_data *cfqd = q->elevator->elevator_data;
3767         struct cfq_io_context *cic;
3768         const int rw = rq_data_dir(rq);
3769         const bool is_sync = rq_is_sync(rq);
3770         struct cfq_queue *cfqq;
3771         unsigned long flags;
3772
3773         might_sleep_if(gfp_mask & __GFP_WAIT);
3774
3775         cic = cfq_get_io_context(cfqd, gfp_mask);
3776
3777         spin_lock_irqsave(q->queue_lock, flags);
3778
3779         if (!cic)
3780                 goto queue_fail;
3781
3782 new_queue:
3783         cfqq = cic_to_cfqq(cic, is_sync);
3784         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3785                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3786                 cic_set_cfqq(cic, cfqq, is_sync);
3787         } else {
3788                 /*
3789                  * If the queue was seeky for too long, break it apart.
3790                  */
3791                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3792                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3793                         cfqq = split_cfqq(cic, cfqq);
3794                         if (!cfqq)
3795                                 goto new_queue;
3796                 }
3797
3798                 /*
3799                  * Check to see if this queue is scheduled to merge with
3800                  * another, closely cooperating queue.  The merging of
3801                  * queues happens here as it must be done in process context.
3802                  * The reference on new_cfqq was taken in merge_cfqqs.
3803                  */
3804                 if (cfqq->new_cfqq)
3805                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3806         }
3807
3808         cfqq->allocated[rw]++;
3809
3810         cfqq->ref++;
3811         rq->elevator_private[0] = cic;
3812         rq->elevator_private[1] = cfqq;
3813         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3814         spin_unlock_irqrestore(q->queue_lock, flags);
3815         return 0;
3816
3817 queue_fail:
3818         cfq_schedule_dispatch(cfqd);
3819         spin_unlock_irqrestore(q->queue_lock, flags);
3820         cfq_log(cfqd, "set_request fail");
3821         return 1;
3822 }
3823
3824 static void cfq_kick_queue(struct work_struct *work)
3825 {
3826         struct cfq_data *cfqd =
3827                 container_of(work, struct cfq_data, unplug_work);
3828         struct request_queue *q = cfqd->queue;
3829
3830         spin_lock_irq(q->queue_lock);
3831         __blk_run_queue(cfqd->queue);
3832         spin_unlock_irq(q->queue_lock);
3833 }
3834
3835 /*
3836  * Timer running if the active_queue is currently idling inside its time slice
3837  */
3838 static void cfq_idle_slice_timer(unsigned long data)
3839 {
3840         struct cfq_data *cfqd = (struct cfq_data *) data;
3841         struct cfq_queue *cfqq;
3842         unsigned long flags;
3843         int timed_out = 1;
3844
3845         cfq_log(cfqd, "idle timer fired");
3846
3847         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3848
3849         cfqq = cfqd->active_queue;
3850         if (cfqq) {
3851                 timed_out = 0;
3852
3853                 /*
3854                  * We saw a request before the queue expired, let it through
3855                  */
3856                 if (cfq_cfqq_must_dispatch(cfqq))
3857                         goto out_kick;
3858
3859                 /*
3860                  * expired
3861                  */
3862                 if (cfq_slice_used(cfqq))
3863                         goto expire;
3864
3865                 /*
3866                  * only expire and reinvoke request handler, if there are
3867                  * other queues with pending requests
3868                  */
3869                 if (!cfqd->busy_queues)
3870                         goto out_cont;
3871
3872                 /*
3873                  * not expired and it has a request pending, let it dispatch
3874                  */
3875                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3876                         goto out_kick;
3877
3878                 /*
3879                  * Queue depth flag is reset only when the idle didn't succeed
3880                  */
3881                 cfq_clear_cfqq_deep(cfqq);
3882         }
3883 expire:
3884         cfq_slice_expired(cfqd, timed_out);
3885 out_kick:
3886         cfq_schedule_dispatch(cfqd);
3887 out_cont:
3888         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3889 }
3890
3891 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3892 {
3893         del_timer_sync(&cfqd->idle_slice_timer);
3894         cancel_work_sync(&cfqd->unplug_work);
3895 }
3896
3897 static void cfq_put_async_queues(struct cfq_data *cfqd)
3898 {
3899         int i;
3900
3901         for (i = 0; i < IOPRIO_BE_NR; i++) {
3902                 if (cfqd->async_cfqq[0][i])
3903                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3904                 if (cfqd->async_cfqq[1][i])
3905                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3906         }
3907
3908         if (cfqd->async_idle_cfqq)
3909                 cfq_put_queue(cfqd->async_idle_cfqq);
3910 }
3911
3912 static void cfq_exit_queue(struct elevator_queue *e)
3913 {
3914         struct cfq_data *cfqd = e->elevator_data;
3915         struct request_queue *q = cfqd->queue;
3916         bool wait = false;
3917
3918         cfq_shutdown_timer_wq(cfqd);
3919
3920         spin_lock_irq(q->queue_lock);
3921
3922         if (cfqd->active_queue)
3923                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3924
3925         while (!list_empty(&cfqd->cic_list)) {
3926                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3927                                                         struct cfq_io_context,
3928                                                         queue_list);
3929
3930                 __cfq_exit_single_io_context(cfqd, cic);
3931         }
3932
3933         cfq_put_async_queues(cfqd);
3934         cfq_release_cfq_groups(cfqd);
3935
3936         /*
3937          * If there are groups which we could not unlink from blkcg list,
3938          * wait for a rcu period for them to be freed.
3939          */
3940         if (cfqd->nr_blkcg_linked_grps)
3941                 wait = true;
3942
3943         spin_unlock_irq(q->queue_lock);
3944
3945         cfq_shutdown_timer_wq(cfqd);
3946
3947         spin_lock(&cic_index_lock);
3948         ida_remove(&cic_index_ida, cfqd->cic_index);
3949         spin_unlock(&cic_index_lock);
3950
3951         /*
3952          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3953          * Do this wait only if there are other unlinked groups out
3954          * there. This can happen if cgroup deletion path claimed the
3955          * responsibility of cleaning up a group before queue cleanup code
3956          * get to the group.
3957          *
3958          * Do not call synchronize_rcu() unconditionally as there are drivers
3959          * which create/delete request queue hundreds of times during scan/boot
3960          * and synchronize_rcu() can take significant time and slow down boot.
3961          */
3962         if (wait)
3963                 synchronize_rcu();
3964
3965 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3966         /* Free up per cpu stats for root group */
3967         free_percpu(cfqd->root_group.blkg.stats_cpu);
3968 #endif
3969         kfree(cfqd);
3970 }
3971
3972 static int cfq_alloc_cic_index(void)
3973 {
3974         int index, error;
3975
3976         do {
3977                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3978                         return -ENOMEM;
3979
3980                 spin_lock(&cic_index_lock);
3981                 error = ida_get_new(&cic_index_ida, &index);
3982                 spin_unlock(&cic_index_lock);
3983                 if (error && error != -EAGAIN)
3984                         return error;
3985         } while (error);
3986
3987         return index;
3988 }
3989
3990 static void *cfq_init_queue(struct request_queue *q)
3991 {
3992         struct cfq_data *cfqd;
3993         int i, j;
3994         struct cfq_group *cfqg;
3995         struct cfq_rb_root *st;
3996
3997         i = cfq_alloc_cic_index();
3998         if (i < 0)
3999                 return NULL;
4000
4001         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
4002         if (!cfqd) {
4003                 spin_lock(&cic_index_lock);
4004                 ida_remove(&cic_index_ida, i);
4005                 spin_unlock(&cic_index_lock);
4006                 return NULL;
4007         }
4008
4009         /*
4010          * Don't need take queue_lock in the routine, since we are
4011          * initializing the ioscheduler, and nobody is using cfqd
4012          */
4013         cfqd->cic_index = i;
4014
4015         /* Init root service tree */
4016         cfqd->grp_service_tree = CFQ_RB_ROOT;
4017
4018         /* Init root group */
4019         cfqg = &cfqd->root_group;
4020         for_each_cfqg_st(cfqg, i, j, st)
4021                 *st = CFQ_RB_ROOT;
4022         RB_CLEAR_NODE(&cfqg->rb_node);
4023
4024         /* Give preference to root group over other groups */
4025         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
4026
4027 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4028         /*
4029          * Set root group reference to 2. One reference will be dropped when
4030          * all groups on cfqd->cfqg_list are being deleted during queue exit.
4031          * Other reference will remain there as we don't want to delete this
4032          * group as it is statically allocated and gets destroyed when
4033          * throtl_data goes away.
4034          */
4035         cfqg->ref = 2;
4036
4037         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4038                 kfree(cfqg);
4039                 kfree(cfqd);
4040                 return NULL;
4041         }
4042
4043         rcu_read_lock();
4044
4045         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4046                                         (void *)cfqd, 0);
4047         rcu_read_unlock();
4048         cfqd->nr_blkcg_linked_grps++;
4049
4050         /* Add group on cfqd->cfqg_list */
4051         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4052 #endif
4053         /*
4054          * Not strictly needed (since RB_ROOT just clears the node and we
4055          * zeroed cfqd on alloc), but better be safe in case someone decides
4056          * to add magic to the rb code
4057          */
4058         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4059                 cfqd->prio_trees[i] = RB_ROOT;
4060
4061         /*
4062          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4063          * Grab a permanent reference to it, so that the normal code flow
4064          * will not attempt to free it.
4065          */
4066         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4067         cfqd->oom_cfqq.ref++;
4068         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4069
4070         INIT_LIST_HEAD(&cfqd->cic_list);
4071
4072         cfqd->queue = q;
4073
4074         init_timer(&cfqd->idle_slice_timer);
4075         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4076         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4077
4078         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4079
4080         cfqd->cfq_quantum = cfq_quantum;
4081         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4082         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4083         cfqd->cfq_back_max = cfq_back_max;
4084         cfqd->cfq_back_penalty = cfq_back_penalty;
4085         cfqd->cfq_slice[0] = cfq_slice_async;
4086         cfqd->cfq_slice[1] = cfq_slice_sync;
4087         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4088         cfqd->cfq_slice_idle = cfq_slice_idle;
4089         cfqd->cfq_group_idle = cfq_group_idle;
4090         cfqd->cfq_latency = 1;
4091         cfqd->hw_tag = -1;
4092         /*
4093          * we optimistically start assuming sync ops weren't delayed in last
4094          * second, in order to have larger depth for async operations.
4095          */
4096         cfqd->last_delayed_sync = jiffies - HZ;
4097         return cfqd;
4098 }
4099
4100 static void cfq_slab_kill(void)
4101 {
4102         /*
4103          * Caller already ensured that pending RCU callbacks are completed,
4104          * so we should have no busy allocations at this point.
4105          */
4106         if (cfq_pool)
4107                 kmem_cache_destroy(cfq_pool);
4108         if (cfq_ioc_pool)
4109                 kmem_cache_destroy(cfq_ioc_pool);
4110 }
4111
4112 static int __init cfq_slab_setup(void)
4113 {
4114         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4115         if (!cfq_pool)
4116                 goto fail;
4117
4118         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4119         if (!cfq_ioc_pool)
4120                 goto fail;
4121
4122         return 0;
4123 fail:
4124         cfq_slab_kill();
4125         return -ENOMEM;
4126 }
4127
4128 /*
4129  * sysfs parts below -->
4130  */
4131 static ssize_t
4132 cfq_var_show(unsigned int var, char *page)
4133 {
4134         return sprintf(page, "%d\n", var);
4135 }
4136
4137 static ssize_t
4138 cfq_var_store(unsigned int *var, const char *page, size_t count)
4139 {
4140         char *p = (char *) page;
4141
4142         *var = simple_strtoul(p, &p, 10);
4143         return count;
4144 }
4145
4146 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4147 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4148 {                                                                       \
4149         struct cfq_data *cfqd = e->elevator_data;                       \
4150         unsigned int __data = __VAR;                                    \
4151         if (__CONV)                                                     \
4152                 __data = jiffies_to_msecs(__data);                      \
4153         return cfq_var_show(__data, (page));                            \
4154 }
4155 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4156 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4157 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4158 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4159 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4160 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4161 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4162 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4163 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4164 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4165 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4166 #undef SHOW_FUNCTION
4167
4168 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4169 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4170 {                                                                       \
4171         struct cfq_data *cfqd = e->elevator_data;                       \
4172         unsigned int __data;                                            \
4173         int ret = cfq_var_store(&__data, (page), count);                \
4174         if (__data < (MIN))                                             \
4175                 __data = (MIN);                                         \
4176         else if (__data > (MAX))                                        \
4177                 __data = (MAX);                                         \
4178         if (__CONV)                                                     \
4179                 *(__PTR) = msecs_to_jiffies(__data);                    \
4180         else                                                            \
4181                 *(__PTR) = __data;                                      \
4182         return ret;                                                     \
4183 }
4184 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4185 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4186                 UINT_MAX, 1);
4187 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4188                 UINT_MAX, 1);
4189 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4190 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4191                 UINT_MAX, 0);
4192 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4193 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4194 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4195 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4196 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4197                 UINT_MAX, 0);
4198 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4199 #undef STORE_FUNCTION
4200
4201 #define CFQ_ATTR(name) \
4202         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4203
4204 static struct elv_fs_entry cfq_attrs[] = {
4205         CFQ_ATTR(quantum),
4206         CFQ_ATTR(fifo_expire_sync),
4207         CFQ_ATTR(fifo_expire_async),
4208         CFQ_ATTR(back_seek_max),
4209         CFQ_ATTR(back_seek_penalty),
4210         CFQ_ATTR(slice_sync),
4211         CFQ_ATTR(slice_async),
4212         CFQ_ATTR(slice_async_rq),
4213         CFQ_ATTR(slice_idle),
4214         CFQ_ATTR(group_idle),
4215         CFQ_ATTR(low_latency),
4216         __ATTR_NULL
4217 };
4218
4219 static struct elevator_type iosched_cfq = {
4220         .ops = {
4221                 .elevator_merge_fn =            cfq_merge,
4222                 .elevator_merged_fn =           cfq_merged_request,
4223                 .elevator_merge_req_fn =        cfq_merged_requests,
4224                 .elevator_allow_merge_fn =      cfq_allow_merge,
4225                 .elevator_bio_merged_fn =       cfq_bio_merged,
4226                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4227                 .elevator_add_req_fn =          cfq_insert_request,
4228                 .elevator_activate_req_fn =     cfq_activate_request,
4229                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4230                 .elevator_completed_req_fn =    cfq_completed_request,
4231                 .elevator_former_req_fn =       elv_rb_former_request,
4232                 .elevator_latter_req_fn =       elv_rb_latter_request,
4233                 .elevator_set_req_fn =          cfq_set_request,
4234                 .elevator_put_req_fn =          cfq_put_request,
4235                 .elevator_may_queue_fn =        cfq_may_queue,
4236                 .elevator_init_fn =             cfq_init_queue,
4237                 .elevator_exit_fn =             cfq_exit_queue,
4238                 .trim =                         cfq_free_io_context,
4239         },
4240         .elevator_attrs =       cfq_attrs,
4241         .elevator_name =        "cfq",
4242         .elevator_owner =       THIS_MODULE,
4243 };
4244
4245 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4246 static struct blkio_policy_type blkio_policy_cfq = {
4247         .ops = {
4248                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4249                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4250         },
4251         .plid = BLKIO_POLICY_PROP,
4252 };
4253 #else
4254 static struct blkio_policy_type blkio_policy_cfq;
4255 #endif
4256
4257 static int __init cfq_init(void)
4258 {
4259         /*
4260          * could be 0 on HZ < 1000 setups
4261          */
4262         if (!cfq_slice_async)
4263                 cfq_slice_async = 1;
4264         if (!cfq_slice_idle)
4265                 cfq_slice_idle = 1;
4266
4267 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4268         if (!cfq_group_idle)
4269                 cfq_group_idle = 1;
4270 #else
4271                 cfq_group_idle = 0;
4272 #endif
4273         if (cfq_slab_setup())
4274                 return -ENOMEM;
4275
4276         elv_register(&iosched_cfq);
4277         blkio_policy_register(&blkio_policy_cfq);
4278
4279         return 0;
4280 }
4281
4282 static void __exit cfq_exit(void)
4283 {
4284         DECLARE_COMPLETION_ONSTACK(all_gone);
4285         blkio_policy_unregister(&blkio_policy_cfq);
4286         elv_unregister(&iosched_cfq);
4287         ioc_gone = &all_gone;
4288         /* ioc_gone's update must be visible before reading ioc_count */
4289         smp_wmb();
4290
4291         /*
4292          * this also protects us from entering cfq_slab_kill() with
4293          * pending RCU callbacks
4294          */
4295         if (elv_ioc_count_read(cfq_ioc_count))
4296                 wait_for_completion(&all_gone);
4297         ida_destroy(&cic_index_ida);
4298         cfq_slab_kill();
4299 }
4300
4301 module_init(cfq_init);
4302 module_exit(cfq_exit);
4303
4304 MODULE_AUTHOR("Jens Axboe");
4305 MODULE_LICENSE("GPL");
4306 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");