6508345376061e73f15c6b8a670dcf051a3a0fc3
[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         /* number of requests that are on the dispatch list or inside driver */
134         int dispatched;
135
136         /* io prio of this group */
137         unsigned short ioprio, org_ioprio;
138         unsigned short ioprio_class;
139
140         pid_t pid;
141
142         u32 seek_history;
143         sector_t last_request_pos;
144
145         struct cfq_rb_root *service_tree;
146         struct cfq_queue *new_cfqq;
147         struct cfq_group *cfqg;
148         /* Number of sectors dispatched from queue in single dispatch round */
149         unsigned long nr_sectors;
150 };
151
152 /*
153  * First index in the service_trees.
154  * IDLE is handled separately, so it has negative index
155  */
156 enum wl_prio_t {
157         BE_WORKLOAD = 0,
158         RT_WORKLOAD = 1,
159         IDLE_WORKLOAD = 2,
160         CFQ_PRIO_NR,
161 };
162
163 /*
164  * Second index in the service_trees.
165  */
166 enum wl_type_t {
167         ASYNC_WORKLOAD = 0,
168         SYNC_NOIDLE_WORKLOAD = 1,
169         SYNC_WORKLOAD = 2
170 };
171
172 /* This is per cgroup per device grouping structure */
173 struct cfq_group {
174         /* group service_tree member */
175         struct rb_node rb_node;
176
177         /* group service_tree key */
178         u64 vdisktime;
179         unsigned int weight;
180         unsigned int new_weight;
181         bool needs_update;
182
183         /* number of cfqq currently on this group */
184         int nr_cfqq;
185
186         /*
187          * Per group busy queues average. Useful for workload slice calc. We
188          * create the array for each prio class but at run time it is used
189          * only for RT and BE class and slot for IDLE class remains unused.
190          * This is primarily done to avoid confusion and a gcc warning.
191          */
192         unsigned int busy_queues_avg[CFQ_PRIO_NR];
193         /*
194          * rr lists of queues with requests. We maintain service trees for
195          * RT and BE classes. These trees are subdivided in subclasses
196          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
197          * class there is no subclassification and all the cfq queues go on
198          * a single tree service_tree_idle.
199          * Counts are embedded in the cfq_rb_root
200          */
201         struct cfq_rb_root service_trees[2][3];
202         struct cfq_rb_root service_tree_idle;
203
204         unsigned long saved_workload_slice;
205         enum wl_type_t saved_workload;
206         enum wl_prio_t saved_serving_prio;
207         struct blkio_group blkg;
208 #ifdef CONFIG_CFQ_GROUP_IOSCHED
209         struct hlist_node cfqd_node;
210         int ref;
211 #endif
212         /* number of requests that are on the dispatch list or inside driver */
213         int dispatched;
214         struct cfq_ttime ttime;
215 };
216
217 /*
218  * Per block device queue structure
219  */
220 struct cfq_data {
221         struct request_queue *queue;
222         /* Root service tree for cfq_groups */
223         struct cfq_rb_root grp_service_tree;
224         struct cfq_group root_group;
225
226         /*
227          * The priority currently being served
228          */
229         enum wl_prio_t serving_prio;
230         enum wl_type_t serving_type;
231         unsigned long workload_expires;
232         struct cfq_group *serving_group;
233
234         /*
235          * Each priority tree is sorted by next_request position.  These
236          * trees are used when determining if two or more queues are
237          * interleaving requests (see cfq_close_cooperator).
238          */
239         struct rb_root prio_trees[CFQ_PRIO_LISTS];
240
241         unsigned int busy_queues;
242         unsigned int busy_sync_queues;
243
244         int rq_in_driver;
245         int rq_in_flight[2];
246
247         /*
248          * queue-depth detection
249          */
250         int rq_queued;
251         int hw_tag;
252         /*
253          * hw_tag can be
254          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256          *  0 => no NCQ
257          */
258         int hw_tag_est_depth;
259         unsigned int hw_tag_samples;
260
261         /*
262          * idle window management
263          */
264         struct timer_list idle_slice_timer;
265         struct work_struct unplug_work;
266
267         struct cfq_queue *active_queue;
268         struct cfq_io_context *active_cic;
269
270         /*
271          * async queue for each priority case
272          */
273         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274         struct cfq_queue *async_idle_cfqq;
275
276         sector_t last_position;
277
278         /*
279          * tunables, see top of file
280          */
281         unsigned int cfq_quantum;
282         unsigned int cfq_fifo_expire[2];
283         unsigned int cfq_back_penalty;
284         unsigned int cfq_back_max;
285         unsigned int cfq_slice[2];
286         unsigned int cfq_slice_async_rq;
287         unsigned int cfq_slice_idle;
288         unsigned int cfq_group_idle;
289         unsigned int cfq_latency;
290
291         unsigned int cic_index;
292         struct list_head cic_list;
293
294         /*
295          * Fallback dummy cfqq for extreme OOM conditions
296          */
297         struct cfq_queue oom_cfqq;
298
299         unsigned long last_delayed_sync;
300
301         /* List of cfq groups being managed on this device*/
302         struct hlist_head cfqg_list;
303
304         /* Number of groups which are on blkcg->blkg_list */
305         unsigned int nr_blkcg_linked_grps;
306 };
307
308 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
309
310 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
311                                             enum wl_prio_t prio,
312                                             enum wl_type_t type)
313 {
314         if (!cfqg)
315                 return NULL;
316
317         if (prio == IDLE_WORKLOAD)
318                 return &cfqg->service_tree_idle;
319
320         return &cfqg->service_trees[prio][type];
321 }
322
323 enum cfqq_state_flags {
324         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
325         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
326         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
327         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
328         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
329         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
330         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
331         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
332         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
333         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
334         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
335         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
336         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
337 };
338
339 #define CFQ_CFQQ_FNS(name)                                              \
340 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
341 {                                                                       \
342         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
343 }                                                                       \
344 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
345 {                                                                       \
346         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
347 }                                                                       \
348 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
349 {                                                                       \
350         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
351 }
352
353 CFQ_CFQQ_FNS(on_rr);
354 CFQ_CFQQ_FNS(wait_request);
355 CFQ_CFQQ_FNS(must_dispatch);
356 CFQ_CFQQ_FNS(must_alloc_slice);
357 CFQ_CFQQ_FNS(fifo_expire);
358 CFQ_CFQQ_FNS(idle_window);
359 CFQ_CFQQ_FNS(prio_changed);
360 CFQ_CFQQ_FNS(slice_new);
361 CFQ_CFQQ_FNS(sync);
362 CFQ_CFQQ_FNS(coop);
363 CFQ_CFQQ_FNS(split_coop);
364 CFQ_CFQQ_FNS(deep);
365 CFQ_CFQQ_FNS(wait_busy);
366 #undef CFQ_CFQQ_FNS
367
368 #ifdef CONFIG_CFQ_GROUP_IOSCHED
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
370         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
371                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
372                         blkg_path(&(cfqq)->cfqg->blkg), ##args)
373
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
375         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
376                                 blkg_path(&(cfqg)->blkg), ##args)       \
377
378 #else
379 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
380         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0)
382 #endif
383 #define cfq_log(cfqd, fmt, args...)     \
384         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
385
386 /* Traverses through cfq group service trees */
387 #define for_each_cfqg_st(cfqg, i, j, st) \
388         for (i = 0; i <= IDLE_WORKLOAD; i++) \
389                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
390                         : &cfqg->service_tree_idle; \
391                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
392                         (i == IDLE_WORKLOAD && j == 0); \
393                         j++, st = i < IDLE_WORKLOAD ? \
394                         &cfqg->service_trees[i][j]: NULL) \
395
396 static inline bool cfq_io_thinktime_big(struct cfq_data *cfqd,
397         struct cfq_ttime *ttime, bool group_idle)
398 {
399         unsigned long slice;
400         if (!sample_valid(ttime->ttime_samples))
401                 return false;
402         if (group_idle)
403                 slice = cfqd->cfq_group_idle;
404         else
405                 slice = cfqd->cfq_slice_idle;
406         return ttime->ttime_mean > slice;
407 }
408
409 static inline bool iops_mode(struct cfq_data *cfqd)
410 {
411         /*
412          * If we are not idling on queues and it is a NCQ drive, parallel
413          * execution of requests is on and measuring time is not possible
414          * in most of the cases until and unless we drive shallower queue
415          * depths and that becomes a performance bottleneck. In such cases
416          * switch to start providing fairness in terms of number of IOs.
417          */
418         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
419                 return true;
420         else
421                 return false;
422 }
423
424 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
425 {
426         if (cfq_class_idle(cfqq))
427                 return IDLE_WORKLOAD;
428         if (cfq_class_rt(cfqq))
429                 return RT_WORKLOAD;
430         return BE_WORKLOAD;
431 }
432
433
434 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
435 {
436         if (!cfq_cfqq_sync(cfqq))
437                 return ASYNC_WORKLOAD;
438         if (!cfq_cfqq_idle_window(cfqq))
439                 return SYNC_NOIDLE_WORKLOAD;
440         return SYNC_WORKLOAD;
441 }
442
443 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
444                                         struct cfq_data *cfqd,
445                                         struct cfq_group *cfqg)
446 {
447         if (wl == IDLE_WORKLOAD)
448                 return cfqg->service_tree_idle.count;
449
450         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
451                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
452                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
453 }
454
455 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
456                                         struct cfq_group *cfqg)
457 {
458         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
459                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
460 }
461
462 static void cfq_dispatch_insert(struct request_queue *, struct request *);
463 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
464                                        struct io_context *, gfp_t);
465 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
466                                                 struct io_context *);
467
468 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
469                                             bool is_sync)
470 {
471         return cic->cfqq[is_sync];
472 }
473
474 static inline void cic_set_cfqq(struct cfq_io_context *cic,
475                                 struct cfq_queue *cfqq, bool is_sync)
476 {
477         cic->cfqq[is_sync] = cfqq;
478 }
479
480 #define CIC_DEAD_KEY    1ul
481 #define CIC_DEAD_INDEX_SHIFT    1
482
483 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
484 {
485         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
486 }
487
488 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
489 {
490         struct cfq_data *cfqd = cic->key;
491
492         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
493                 return NULL;
494
495         return cfqd;
496 }
497
498 /*
499  * We regard a request as SYNC, if it's either a read or has the SYNC bit
500  * set (in which case it could also be direct WRITE).
501  */
502 static inline bool cfq_bio_sync(struct bio *bio)
503 {
504         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
505 }
506
507 /*
508  * scheduler run of queue, if there are requests pending and no one in the
509  * driver that will restart queueing
510  */
511 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
512 {
513         if (cfqd->busy_queues) {
514                 cfq_log(cfqd, "schedule dispatch");
515                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
516         }
517 }
518
519 /*
520  * Scale schedule slice based on io priority. Use the sync time slice only
521  * if a queue is marked sync and has sync io queued. A sync queue with async
522  * io only, should not get full sync slice length.
523  */
524 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
525                                  unsigned short prio)
526 {
527         const int base_slice = cfqd->cfq_slice[sync];
528
529         WARN_ON(prio >= IOPRIO_BE_NR);
530
531         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
532 }
533
534 static inline int
535 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
536 {
537         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
538 }
539
540 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
541 {
542         u64 d = delta << CFQ_SERVICE_SHIFT;
543
544         d = d * BLKIO_WEIGHT_DEFAULT;
545         do_div(d, cfqg->weight);
546         return d;
547 }
548
549 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
550 {
551         s64 delta = (s64)(vdisktime - min_vdisktime);
552         if (delta > 0)
553                 min_vdisktime = vdisktime;
554
555         return min_vdisktime;
556 }
557
558 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
559 {
560         s64 delta = (s64)(vdisktime - min_vdisktime);
561         if (delta < 0)
562                 min_vdisktime = vdisktime;
563
564         return min_vdisktime;
565 }
566
567 static void update_min_vdisktime(struct cfq_rb_root *st)
568 {
569         struct cfq_group *cfqg;
570
571         if (st->left) {
572                 cfqg = rb_entry_cfqg(st->left);
573                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
574                                                   cfqg->vdisktime);
575         }
576 }
577
578 /*
579  * get averaged number of queues of RT/BE priority.
580  * average is updated, with a formula that gives more weight to higher numbers,
581  * to quickly follows sudden increases and decrease slowly
582  */
583
584 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
585                                         struct cfq_group *cfqg, bool rt)
586 {
587         unsigned min_q, max_q;
588         unsigned mult  = cfq_hist_divisor - 1;
589         unsigned round = cfq_hist_divisor / 2;
590         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
591
592         min_q = min(cfqg->busy_queues_avg[rt], busy);
593         max_q = max(cfqg->busy_queues_avg[rt], busy);
594         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
595                 cfq_hist_divisor;
596         return cfqg->busy_queues_avg[rt];
597 }
598
599 static inline unsigned
600 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
601 {
602         struct cfq_rb_root *st = &cfqd->grp_service_tree;
603
604         return cfq_target_latency * cfqg->weight / st->total_weight;
605 }
606
607 static inline unsigned
608 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
609 {
610         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
611         if (cfqd->cfq_latency) {
612                 /*
613                  * interested queues (we consider only the ones with the same
614                  * priority class in the cfq group)
615                  */
616                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
617                                                 cfq_class_rt(cfqq));
618                 unsigned sync_slice = cfqd->cfq_slice[1];
619                 unsigned expect_latency = sync_slice * iq;
620                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
621
622                 if (expect_latency > group_slice) {
623                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
624                         /* scale low_slice according to IO priority
625                          * and sync vs async */
626                         unsigned low_slice =
627                                 min(slice, base_low_slice * slice / sync_slice);
628                         /* the adapted slice value is scaled to fit all iqs
629                          * into the target latency */
630                         slice = max(slice * group_slice / expect_latency,
631                                     low_slice);
632                 }
633         }
634         return slice;
635 }
636
637 static inline void
638 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
639 {
640         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
641
642         cfqq->slice_start = jiffies;
643         cfqq->slice_end = jiffies + slice;
644         cfqq->allocated_slice = slice;
645         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
646 }
647
648 /*
649  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
650  * isn't valid until the first request from the dispatch is activated
651  * and the slice time set.
652  */
653 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
654 {
655         if (cfq_cfqq_slice_new(cfqq))
656                 return false;
657         if (time_before(jiffies, cfqq->slice_end))
658                 return false;
659
660         return true;
661 }
662
663 /*
664  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
665  * We choose the request that is closest to the head right now. Distance
666  * behind the head is penalized and only allowed to a certain extent.
667  */
668 static struct request *
669 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
670 {
671         sector_t s1, s2, d1 = 0, d2 = 0;
672         unsigned long back_max;
673 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
674 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
675         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
676
677         if (rq1 == NULL || rq1 == rq2)
678                 return rq2;
679         if (rq2 == NULL)
680                 return rq1;
681
682         if (rq_is_sync(rq1) != rq_is_sync(rq2))
683                 return rq_is_sync(rq1) ? rq1 : rq2;
684
685         s1 = blk_rq_pos(rq1);
686         s2 = blk_rq_pos(rq2);
687
688         /*
689          * by definition, 1KiB is 2 sectors
690          */
691         back_max = cfqd->cfq_back_max * 2;
692
693         /*
694          * Strict one way elevator _except_ in the case where we allow
695          * short backward seeks which are biased as twice the cost of a
696          * similar forward seek.
697          */
698         if (s1 >= last)
699                 d1 = s1 - last;
700         else if (s1 + back_max >= last)
701                 d1 = (last - s1) * cfqd->cfq_back_penalty;
702         else
703                 wrap |= CFQ_RQ1_WRAP;
704
705         if (s2 >= last)
706                 d2 = s2 - last;
707         else if (s2 + back_max >= last)
708                 d2 = (last - s2) * cfqd->cfq_back_penalty;
709         else
710                 wrap |= CFQ_RQ2_WRAP;
711
712         /* Found required data */
713
714         /*
715          * By doing switch() on the bit mask "wrap" we avoid having to
716          * check two variables for all permutations: --> faster!
717          */
718         switch (wrap) {
719         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
720                 if (d1 < d2)
721                         return rq1;
722                 else if (d2 < d1)
723                         return rq2;
724                 else {
725                         if (s1 >= s2)
726                                 return rq1;
727                         else
728                                 return rq2;
729                 }
730
731         case CFQ_RQ2_WRAP:
732                 return rq1;
733         case CFQ_RQ1_WRAP:
734                 return rq2;
735         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
736         default:
737                 /*
738                  * Since both rqs are wrapped,
739                  * start with the one that's further behind head
740                  * (--> only *one* back seek required),
741                  * since back seek takes more time than forward.
742                  */
743                 if (s1 <= s2)
744                         return rq1;
745                 else
746                         return rq2;
747         }
748 }
749
750 /*
751  * The below is leftmost cache rbtree addon
752  */
753 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
754 {
755         /* Service tree is empty */
756         if (!root->count)
757                 return NULL;
758
759         if (!root->left)
760                 root->left = rb_first(&root->rb);
761
762         if (root->left)
763                 return rb_entry(root->left, struct cfq_queue, rb_node);
764
765         return NULL;
766 }
767
768 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
769 {
770         if (!root->left)
771                 root->left = rb_first(&root->rb);
772
773         if (root->left)
774                 return rb_entry_cfqg(root->left);
775
776         return NULL;
777 }
778
779 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
780 {
781         rb_erase(n, root);
782         RB_CLEAR_NODE(n);
783 }
784
785 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
786 {
787         if (root->left == n)
788                 root->left = NULL;
789         rb_erase_init(n, &root->rb);
790         --root->count;
791 }
792
793 /*
794  * would be nice to take fifo expire time into account as well
795  */
796 static struct request *
797 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
798                   struct request *last)
799 {
800         struct rb_node *rbnext = rb_next(&last->rb_node);
801         struct rb_node *rbprev = rb_prev(&last->rb_node);
802         struct request *next = NULL, *prev = NULL;
803
804         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
805
806         if (rbprev)
807                 prev = rb_entry_rq(rbprev);
808
809         if (rbnext)
810                 next = rb_entry_rq(rbnext);
811         else {
812                 rbnext = rb_first(&cfqq->sort_list);
813                 if (rbnext && rbnext != &last->rb_node)
814                         next = rb_entry_rq(rbnext);
815         }
816
817         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
818 }
819
820 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
821                                       struct cfq_queue *cfqq)
822 {
823         /*
824          * just an approximation, should be ok.
825          */
826         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
827                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
828 }
829
830 static inline s64
831 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
832 {
833         return cfqg->vdisktime - st->min_vdisktime;
834 }
835
836 static void
837 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
838 {
839         struct rb_node **node = &st->rb.rb_node;
840         struct rb_node *parent = NULL;
841         struct cfq_group *__cfqg;
842         s64 key = cfqg_key(st, cfqg);
843         int left = 1;
844
845         while (*node != NULL) {
846                 parent = *node;
847                 __cfqg = rb_entry_cfqg(parent);
848
849                 if (key < cfqg_key(st, __cfqg))
850                         node = &parent->rb_left;
851                 else {
852                         node = &parent->rb_right;
853                         left = 0;
854                 }
855         }
856
857         if (left)
858                 st->left = &cfqg->rb_node;
859
860         rb_link_node(&cfqg->rb_node, parent, node);
861         rb_insert_color(&cfqg->rb_node, &st->rb);
862 }
863
864 static void
865 cfq_update_group_weight(struct cfq_group *cfqg)
866 {
867         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
868         if (cfqg->needs_update) {
869                 cfqg->weight = cfqg->new_weight;
870                 cfqg->needs_update = false;
871         }
872 }
873
874 static void
875 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
876 {
877         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
878
879         cfq_update_group_weight(cfqg);
880         __cfq_group_service_tree_add(st, cfqg);
881         st->total_weight += cfqg->weight;
882 }
883
884 static void
885 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
886 {
887         struct cfq_rb_root *st = &cfqd->grp_service_tree;
888         struct cfq_group *__cfqg;
889         struct rb_node *n;
890
891         cfqg->nr_cfqq++;
892         if (!RB_EMPTY_NODE(&cfqg->rb_node))
893                 return;
894
895         /*
896          * Currently put the group at the end. Later implement something
897          * so that groups get lesser vtime based on their weights, so that
898          * if group does not loose all if it was not continuously backlogged.
899          */
900         n = rb_last(&st->rb);
901         if (n) {
902                 __cfqg = rb_entry_cfqg(n);
903                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
904         } else
905                 cfqg->vdisktime = st->min_vdisktime;
906         cfq_group_service_tree_add(st, cfqg);
907 }
908
909 static void
910 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
911 {
912         st->total_weight -= cfqg->weight;
913         if (!RB_EMPTY_NODE(&cfqg->rb_node))
914                 cfq_rb_erase(&cfqg->rb_node, st);
915 }
916
917 static void
918 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
919 {
920         struct cfq_rb_root *st = &cfqd->grp_service_tree;
921
922         BUG_ON(cfqg->nr_cfqq < 1);
923         cfqg->nr_cfqq--;
924
925         /* If there are other cfq queues under this group, don't delete it */
926         if (cfqg->nr_cfqq)
927                 return;
928
929         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
930         cfq_group_service_tree_del(st, cfqg);
931         cfqg->saved_workload_slice = 0;
932         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
933 }
934
935 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
936                                                 unsigned int *unaccounted_time)
937 {
938         unsigned int slice_used;
939
940         /*
941          * Queue got expired before even a single request completed or
942          * got expired immediately after first request completion.
943          */
944         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
945                 /*
946                  * Also charge the seek time incurred to the group, otherwise
947                  * if there are mutiple queues in the group, each can dispatch
948                  * a single request on seeky media and cause lots of seek time
949                  * and group will never know it.
950                  */
951                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
952                                         1);
953         } else {
954                 slice_used = jiffies - cfqq->slice_start;
955                 if (slice_used > cfqq->allocated_slice) {
956                         *unaccounted_time = slice_used - cfqq->allocated_slice;
957                         slice_used = cfqq->allocated_slice;
958                 }
959                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
960                         *unaccounted_time += cfqq->slice_start -
961                                         cfqq->dispatch_start;
962         }
963
964         return slice_used;
965 }
966
967 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
968                                 struct cfq_queue *cfqq)
969 {
970         struct cfq_rb_root *st = &cfqd->grp_service_tree;
971         unsigned int used_sl, charge, unaccounted_sl = 0;
972         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
973                         - cfqg->service_tree_idle.count;
974
975         BUG_ON(nr_sync < 0);
976         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
977
978         if (iops_mode(cfqd))
979                 charge = cfqq->slice_dispatch;
980         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
981                 charge = cfqq->allocated_slice;
982
983         /* Can't update vdisktime while group is on service tree */
984         cfq_group_service_tree_del(st, cfqg);
985         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
986         /* If a new weight was requested, update now, off tree */
987         cfq_group_service_tree_add(st, cfqg);
988
989         /* This group is being expired. Save the context */
990         if (time_after(cfqd->workload_expires, jiffies)) {
991                 cfqg->saved_workload_slice = cfqd->workload_expires
992                                                 - jiffies;
993                 cfqg->saved_workload = cfqd->serving_type;
994                 cfqg->saved_serving_prio = cfqd->serving_prio;
995         } else
996                 cfqg->saved_workload_slice = 0;
997
998         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
999                                         st->min_vdisktime);
1000         cfq_log_cfqq(cfqq->cfqd, cfqq,
1001                      "sl_used=%u disp=%u charge=%u iops=%u sect=%lu",
1002                      used_sl, cfqq->slice_dispatch, charge,
1003                      iops_mode(cfqd), cfqq->nr_sectors);
1004         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
1005                                           unaccounted_sl);
1006         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
1007 }
1008
1009 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1010 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1011 {
1012         if (blkg)
1013                 return container_of(blkg, struct cfq_group, blkg);
1014         return NULL;
1015 }
1016
1017 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1018                                           unsigned int weight)
1019 {
1020         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1021         cfqg->new_weight = weight;
1022         cfqg->needs_update = true;
1023 }
1024
1025 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1026                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1027 {
1028         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1029         unsigned int major, minor;
1030
1031         /*
1032          * Add group onto cgroup list. It might happen that bdi->dev is
1033          * not initialized yet. Initialize this new group without major
1034          * and minor info and this info will be filled in once a new thread
1035          * comes for IO.
1036          */
1037         if (bdi->dev) {
1038                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1039                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1040                                         (void *)cfqd, MKDEV(major, minor));
1041         } else
1042                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1043                                         (void *)cfqd, 0);
1044
1045         cfqd->nr_blkcg_linked_grps++;
1046         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1047
1048         /* Add group on cfqd list */
1049         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1050 }
1051
1052 /*
1053  * Should be called from sleepable context. No request queue lock as per
1054  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1055  * from sleepable context.
1056  */
1057 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1058 {
1059         struct cfq_group *cfqg = NULL;
1060         int i, j, ret;
1061         struct cfq_rb_root *st;
1062
1063         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1064         if (!cfqg)
1065                 return NULL;
1066
1067         for_each_cfqg_st(cfqg, i, j, st)
1068                 *st = CFQ_RB_ROOT;
1069         RB_CLEAR_NODE(&cfqg->rb_node);
1070
1071         cfqg->ttime.last_end_request = jiffies;
1072
1073         /*
1074          * Take the initial reference that will be released on destroy
1075          * This can be thought of a joint reference by cgroup and
1076          * elevator which will be dropped by either elevator exit
1077          * or cgroup deletion path depending on who is exiting first.
1078          */
1079         cfqg->ref = 1;
1080
1081         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1082         if (ret) {
1083                 kfree(cfqg);
1084                 return NULL;
1085         }
1086
1087         return cfqg;
1088 }
1089
1090 static struct cfq_group *
1091 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1092 {
1093         struct cfq_group *cfqg = NULL;
1094         void *key = cfqd;
1095         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1096         unsigned int major, minor;
1097
1098         /*
1099          * This is the common case when there are no blkio cgroups.
1100          * Avoid lookup in this case
1101          */
1102         if (blkcg == &blkio_root_cgroup)
1103                 cfqg = &cfqd->root_group;
1104         else
1105                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1106
1107         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1108                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1109                 cfqg->blkg.dev = MKDEV(major, minor);
1110         }
1111
1112         return cfqg;
1113 }
1114
1115 /*
1116  * Search for the cfq group current task belongs to. request_queue lock must
1117  * be held.
1118  */
1119 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1120 {
1121         struct blkio_cgroup *blkcg;
1122         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1123         struct request_queue *q = cfqd->queue;
1124
1125         rcu_read_lock();
1126         blkcg = task_blkio_cgroup(current);
1127         cfqg = cfq_find_cfqg(cfqd, blkcg);
1128         if (cfqg) {
1129                 rcu_read_unlock();
1130                 return cfqg;
1131         }
1132
1133         /*
1134          * Need to allocate a group. Allocation of group also needs allocation
1135          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1136          * we need to drop rcu lock and queue_lock before we call alloc.
1137          *
1138          * Not taking any queue reference here and assuming that queue is
1139          * around by the time we return. CFQ queue allocation code does
1140          * the same. It might be racy though.
1141          */
1142
1143         rcu_read_unlock();
1144         spin_unlock_irq(q->queue_lock);
1145
1146         cfqg = cfq_alloc_cfqg(cfqd);
1147
1148         spin_lock_irq(q->queue_lock);
1149
1150         rcu_read_lock();
1151         blkcg = task_blkio_cgroup(current);
1152
1153         /*
1154          * If some other thread already allocated the group while we were
1155          * not holding queue lock, free up the group
1156          */
1157         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1158
1159         if (__cfqg) {
1160                 kfree(cfqg);
1161                 rcu_read_unlock();
1162                 return __cfqg;
1163         }
1164
1165         if (!cfqg)
1166                 cfqg = &cfqd->root_group;
1167
1168         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1169         rcu_read_unlock();
1170         return cfqg;
1171 }
1172
1173 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1174 {
1175         cfqg->ref++;
1176         return cfqg;
1177 }
1178
1179 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1180 {
1181         /* Currently, all async queues are mapped to root group */
1182         if (!cfq_cfqq_sync(cfqq))
1183                 cfqg = &cfqq->cfqd->root_group;
1184
1185         cfqq->cfqg = cfqg;
1186         /* cfqq reference on cfqg */
1187         cfqq->cfqg->ref++;
1188 }
1189
1190 static void cfq_put_cfqg(struct cfq_group *cfqg)
1191 {
1192         struct cfq_rb_root *st;
1193         int i, j;
1194
1195         BUG_ON(cfqg->ref <= 0);
1196         cfqg->ref--;
1197         if (cfqg->ref)
1198                 return;
1199         for_each_cfqg_st(cfqg, i, j, st)
1200                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1201         free_percpu(cfqg->blkg.stats_cpu);
1202         kfree(cfqg);
1203 }
1204
1205 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1206 {
1207         /* Something wrong if we are trying to remove same group twice */
1208         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1209
1210         hlist_del_init(&cfqg->cfqd_node);
1211
1212         BUG_ON(cfqd->nr_blkcg_linked_grps <= 0);
1213         cfqd->nr_blkcg_linked_grps--;
1214
1215         /*
1216          * Put the reference taken at the time of creation so that when all
1217          * queues are gone, group can be destroyed.
1218          */
1219         cfq_put_cfqg(cfqg);
1220 }
1221
1222 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1223 {
1224         struct hlist_node *pos, *n;
1225         struct cfq_group *cfqg;
1226
1227         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1228                 /*
1229                  * If cgroup removal path got to blk_group first and removed
1230                  * it from cgroup list, then it will take care of destroying
1231                  * cfqg also.
1232                  */
1233                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1234                         cfq_destroy_cfqg(cfqd, cfqg);
1235         }
1236 }
1237
1238 /*
1239  * Blk cgroup controller notification saying that blkio_group object is being
1240  * delinked as associated cgroup object is going away. That also means that
1241  * no new IO will come in this group. So get rid of this group as soon as
1242  * any pending IO in the group is finished.
1243  *
1244  * This function is called under rcu_read_lock(). key is the rcu protected
1245  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1246  * read lock.
1247  *
1248  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1249  * it should not be NULL as even if elevator was exiting, cgroup deltion
1250  * path got to it first.
1251  */
1252 static void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1253 {
1254         unsigned long  flags;
1255         struct cfq_data *cfqd = key;
1256
1257         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1258         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1259         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1260 }
1261
1262 #else /* GROUP_IOSCHED */
1263 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1264 {
1265         return &cfqd->root_group;
1266 }
1267
1268 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1269 {
1270         return cfqg;
1271 }
1272
1273 static inline void
1274 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1275         cfqq->cfqg = cfqg;
1276 }
1277
1278 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1279 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1280
1281 #endif /* GROUP_IOSCHED */
1282
1283 /*
1284  * The cfqd->service_trees holds all pending cfq_queue's that have
1285  * requests waiting to be processed. It is sorted in the order that
1286  * we will service the queues.
1287  */
1288 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1289                                  bool add_front)
1290 {
1291         struct rb_node **p, *parent;
1292         struct cfq_queue *__cfqq;
1293         unsigned long rb_key;
1294         struct cfq_rb_root *service_tree;
1295         int left;
1296         int new_cfqq = 1;
1297
1298         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1299                                                 cfqq_type(cfqq));
1300         if (cfq_class_idle(cfqq)) {
1301                 rb_key = CFQ_IDLE_DELAY;
1302                 parent = rb_last(&service_tree->rb);
1303                 if (parent && parent != &cfqq->rb_node) {
1304                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1305                         rb_key += __cfqq->rb_key;
1306                 } else
1307                         rb_key += jiffies;
1308         } else if (!add_front) {
1309                 /*
1310                  * Get our rb key offset. Subtract any residual slice
1311                  * value carried from last service. A negative resid
1312                  * count indicates slice overrun, and this should position
1313                  * the next service time further away in the tree.
1314                  */
1315                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1316                 rb_key -= cfqq->slice_resid;
1317                 cfqq->slice_resid = 0;
1318         } else {
1319                 rb_key = -HZ;
1320                 __cfqq = cfq_rb_first(service_tree);
1321                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1322         }
1323
1324         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1325                 new_cfqq = 0;
1326                 /*
1327                  * same position, nothing more to do
1328                  */
1329                 if (rb_key == cfqq->rb_key &&
1330                     cfqq->service_tree == service_tree)
1331                         return;
1332
1333                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1334                 cfqq->service_tree = NULL;
1335         }
1336
1337         left = 1;
1338         parent = NULL;
1339         cfqq->service_tree = service_tree;
1340         p = &service_tree->rb.rb_node;
1341         while (*p) {
1342                 struct rb_node **n;
1343
1344                 parent = *p;
1345                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1346
1347                 /*
1348                  * sort by key, that represents service time.
1349                  */
1350                 if (time_before(rb_key, __cfqq->rb_key))
1351                         n = &(*p)->rb_left;
1352                 else {
1353                         n = &(*p)->rb_right;
1354                         left = 0;
1355                 }
1356
1357                 p = n;
1358         }
1359
1360         if (left)
1361                 service_tree->left = &cfqq->rb_node;
1362
1363         cfqq->rb_key = rb_key;
1364         rb_link_node(&cfqq->rb_node, parent, p);
1365         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1366         service_tree->count++;
1367         if (add_front || !new_cfqq)
1368                 return;
1369         cfq_group_notify_queue_add(cfqd, cfqq->cfqg);
1370 }
1371
1372 static struct cfq_queue *
1373 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1374                      sector_t sector, struct rb_node **ret_parent,
1375                      struct rb_node ***rb_link)
1376 {
1377         struct rb_node **p, *parent;
1378         struct cfq_queue *cfqq = NULL;
1379
1380         parent = NULL;
1381         p = &root->rb_node;
1382         while (*p) {
1383                 struct rb_node **n;
1384
1385                 parent = *p;
1386                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1387
1388                 /*
1389                  * Sort strictly based on sector.  Smallest to the left,
1390                  * largest to the right.
1391                  */
1392                 if (sector > blk_rq_pos(cfqq->next_rq))
1393                         n = &(*p)->rb_right;
1394                 else if (sector < blk_rq_pos(cfqq->next_rq))
1395                         n = &(*p)->rb_left;
1396                 else
1397                         break;
1398                 p = n;
1399                 cfqq = NULL;
1400         }
1401
1402         *ret_parent = parent;
1403         if (rb_link)
1404                 *rb_link = p;
1405         return cfqq;
1406 }
1407
1408 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1409 {
1410         struct rb_node **p, *parent;
1411         struct cfq_queue *__cfqq;
1412
1413         if (cfqq->p_root) {
1414                 rb_erase(&cfqq->p_node, cfqq->p_root);
1415                 cfqq->p_root = NULL;
1416         }
1417
1418         if (cfq_class_idle(cfqq))
1419                 return;
1420         if (!cfqq->next_rq)
1421                 return;
1422
1423         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1424         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1425                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1426         if (!__cfqq) {
1427                 rb_link_node(&cfqq->p_node, parent, p);
1428                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1429         } else
1430                 cfqq->p_root = NULL;
1431 }
1432
1433 /*
1434  * Update cfqq's position in the service tree.
1435  */
1436 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1437 {
1438         /*
1439          * Resorting requires the cfqq to be on the RR list already.
1440          */
1441         if (cfq_cfqq_on_rr(cfqq)) {
1442                 cfq_service_tree_add(cfqd, cfqq, 0);
1443                 cfq_prio_tree_add(cfqd, cfqq);
1444         }
1445 }
1446
1447 /*
1448  * add to busy list of queues for service, trying to be fair in ordering
1449  * the pending list according to last request service
1450  */
1451 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1452 {
1453         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1454         BUG_ON(cfq_cfqq_on_rr(cfqq));
1455         cfq_mark_cfqq_on_rr(cfqq);
1456         cfqd->busy_queues++;
1457         if (cfq_cfqq_sync(cfqq))
1458                 cfqd->busy_sync_queues++;
1459
1460         cfq_resort_rr_list(cfqd, cfqq);
1461 }
1462
1463 /*
1464  * Called when the cfqq no longer has requests pending, remove it from
1465  * the service tree.
1466  */
1467 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1468 {
1469         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1470         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1471         cfq_clear_cfqq_on_rr(cfqq);
1472
1473         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1474                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1475                 cfqq->service_tree = NULL;
1476         }
1477         if (cfqq->p_root) {
1478                 rb_erase(&cfqq->p_node, cfqq->p_root);
1479                 cfqq->p_root = NULL;
1480         }
1481
1482         cfq_group_notify_queue_del(cfqd, cfqq->cfqg);
1483         BUG_ON(!cfqd->busy_queues);
1484         cfqd->busy_queues--;
1485         if (cfq_cfqq_sync(cfqq))
1486                 cfqd->busy_sync_queues--;
1487 }
1488
1489 /*
1490  * rb tree support functions
1491  */
1492 static void cfq_del_rq_rb(struct request *rq)
1493 {
1494         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1495         const int sync = rq_is_sync(rq);
1496
1497         BUG_ON(!cfqq->queued[sync]);
1498         cfqq->queued[sync]--;
1499
1500         elv_rb_del(&cfqq->sort_list, rq);
1501
1502         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1503                 /*
1504                  * Queue will be deleted from service tree when we actually
1505                  * expire it later. Right now just remove it from prio tree
1506                  * as it is empty.
1507                  */
1508                 if (cfqq->p_root) {
1509                         rb_erase(&cfqq->p_node, cfqq->p_root);
1510                         cfqq->p_root = NULL;
1511                 }
1512         }
1513 }
1514
1515 static void cfq_add_rq_rb(struct request *rq)
1516 {
1517         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1518         struct cfq_data *cfqd = cfqq->cfqd;
1519         struct request *prev;
1520
1521         cfqq->queued[rq_is_sync(rq)]++;
1522
1523         elv_rb_add(&cfqq->sort_list, rq);
1524
1525         if (!cfq_cfqq_on_rr(cfqq))
1526                 cfq_add_cfqq_rr(cfqd, cfqq);
1527
1528         /*
1529          * check if this request is a better next-serve candidate
1530          */
1531         prev = cfqq->next_rq;
1532         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1533
1534         /*
1535          * adjust priority tree position, if ->next_rq changes
1536          */
1537         if (prev != cfqq->next_rq)
1538                 cfq_prio_tree_add(cfqd, cfqq);
1539
1540         BUG_ON(!cfqq->next_rq);
1541 }
1542
1543 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1544 {
1545         elv_rb_del(&cfqq->sort_list, rq);
1546         cfqq->queued[rq_is_sync(rq)]--;
1547         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1548                                         rq_data_dir(rq), rq_is_sync(rq));
1549         cfq_add_rq_rb(rq);
1550         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1551                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1552                         rq_is_sync(rq));
1553 }
1554
1555 static struct request *
1556 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1557 {
1558         struct task_struct *tsk = current;
1559         struct cfq_io_context *cic;
1560         struct cfq_queue *cfqq;
1561
1562         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1563         if (!cic)
1564                 return NULL;
1565
1566         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1567         if (cfqq) {
1568                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1569
1570                 return elv_rb_find(&cfqq->sort_list, sector);
1571         }
1572
1573         return NULL;
1574 }
1575
1576 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1577 {
1578         struct cfq_data *cfqd = q->elevator->elevator_data;
1579
1580         cfqd->rq_in_driver++;
1581         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1582                                                 cfqd->rq_in_driver);
1583
1584         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1585 }
1586
1587 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1588 {
1589         struct cfq_data *cfqd = q->elevator->elevator_data;
1590
1591         WARN_ON(!cfqd->rq_in_driver);
1592         cfqd->rq_in_driver--;
1593         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1594                                                 cfqd->rq_in_driver);
1595 }
1596
1597 static void cfq_remove_request(struct request *rq)
1598 {
1599         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1600
1601         if (cfqq->next_rq == rq)
1602                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1603
1604         list_del_init(&rq->queuelist);
1605         cfq_del_rq_rb(rq);
1606
1607         cfqq->cfqd->rq_queued--;
1608         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1609                                         rq_data_dir(rq), rq_is_sync(rq));
1610 }
1611
1612 static int cfq_merge(struct request_queue *q, struct request **req,
1613                      struct bio *bio)
1614 {
1615         struct cfq_data *cfqd = q->elevator->elevator_data;
1616         struct request *__rq;
1617
1618         __rq = cfq_find_rq_fmerge(cfqd, bio);
1619         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1620                 *req = __rq;
1621                 return ELEVATOR_FRONT_MERGE;
1622         }
1623
1624         return ELEVATOR_NO_MERGE;
1625 }
1626
1627 static void cfq_merged_request(struct request_queue *q, struct request *req,
1628                                int type)
1629 {
1630         if (type == ELEVATOR_FRONT_MERGE) {
1631                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1632
1633                 cfq_reposition_rq_rb(cfqq, req);
1634         }
1635 }
1636
1637 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1638                                 struct bio *bio)
1639 {
1640         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1641                                         bio_data_dir(bio), cfq_bio_sync(bio));
1642 }
1643
1644 static void
1645 cfq_merged_requests(struct request_queue *q, struct request *rq,
1646                     struct request *next)
1647 {
1648         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1649         /*
1650          * reposition in fifo if next is older than rq
1651          */
1652         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1653             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1654                 list_move(&rq->queuelist, &next->queuelist);
1655                 rq_set_fifo_time(rq, rq_fifo_time(next));
1656         }
1657
1658         if (cfqq->next_rq == next)
1659                 cfqq->next_rq = rq;
1660         cfq_remove_request(next);
1661         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1662                                         rq_data_dir(next), rq_is_sync(next));
1663 }
1664
1665 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1666                            struct bio *bio)
1667 {
1668         struct cfq_data *cfqd = q->elevator->elevator_data;
1669         struct cfq_io_context *cic;
1670         struct cfq_queue *cfqq;
1671
1672         /*
1673          * Disallow merge of a sync bio into an async request.
1674          */
1675         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1676                 return false;
1677
1678         /*
1679          * Lookup the cfqq that this bio will be queued with. Allow
1680          * merge only if rq is queued there.
1681          */
1682         cic = cfq_cic_lookup(cfqd, current->io_context);
1683         if (!cic)
1684                 return false;
1685
1686         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1687         return cfqq == RQ_CFQQ(rq);
1688 }
1689
1690 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1691 {
1692         del_timer(&cfqd->idle_slice_timer);
1693         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1694 }
1695
1696 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1697                                    struct cfq_queue *cfqq)
1698 {
1699         if (cfqq) {
1700                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1701                                 cfqd->serving_prio, cfqd->serving_type);
1702                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1703                 cfqq->slice_start = 0;
1704                 cfqq->dispatch_start = jiffies;
1705                 cfqq->allocated_slice = 0;
1706                 cfqq->slice_end = 0;
1707                 cfqq->slice_dispatch = 0;
1708                 cfqq->nr_sectors = 0;
1709
1710                 cfq_clear_cfqq_wait_request(cfqq);
1711                 cfq_clear_cfqq_must_dispatch(cfqq);
1712                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1713                 cfq_clear_cfqq_fifo_expire(cfqq);
1714                 cfq_mark_cfqq_slice_new(cfqq);
1715
1716                 cfq_del_timer(cfqd, cfqq);
1717         }
1718
1719         cfqd->active_queue = cfqq;
1720 }
1721
1722 /*
1723  * current cfqq expired its slice (or was too idle), select new one
1724  */
1725 static void
1726 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1727                     bool timed_out)
1728 {
1729         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1730
1731         if (cfq_cfqq_wait_request(cfqq))
1732                 cfq_del_timer(cfqd, cfqq);
1733
1734         cfq_clear_cfqq_wait_request(cfqq);
1735         cfq_clear_cfqq_wait_busy(cfqq);
1736
1737         /*
1738          * If this cfqq is shared between multiple processes, check to
1739          * make sure that those processes are still issuing I/Os within
1740          * the mean seek distance.  If not, it may be time to break the
1741          * queues apart again.
1742          */
1743         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1744                 cfq_mark_cfqq_split_coop(cfqq);
1745
1746         /*
1747          * store what was left of this slice, if the queue idled/timed out
1748          */
1749         if (timed_out) {
1750                 if (cfq_cfqq_slice_new(cfqq))
1751                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1752                 else
1753                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1754                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1755         }
1756
1757         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1758
1759         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1760                 cfq_del_cfqq_rr(cfqd, cfqq);
1761
1762         cfq_resort_rr_list(cfqd, cfqq);
1763
1764         if (cfqq == cfqd->active_queue)
1765                 cfqd->active_queue = NULL;
1766
1767         if (cfqd->active_cic) {
1768                 put_io_context(cfqd->active_cic->ioc);
1769                 cfqd->active_cic = NULL;
1770         }
1771 }
1772
1773 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1774 {
1775         struct cfq_queue *cfqq = cfqd->active_queue;
1776
1777         if (cfqq)
1778                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1779 }
1780
1781 /*
1782  * Get next queue for service. Unless we have a queue preemption,
1783  * we'll simply select the first cfqq in the service tree.
1784  */
1785 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1786 {
1787         struct cfq_rb_root *service_tree =
1788                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1789                                         cfqd->serving_type);
1790
1791         if (!cfqd->rq_queued)
1792                 return NULL;
1793
1794         /* There is nothing to dispatch */
1795         if (!service_tree)
1796                 return NULL;
1797         if (RB_EMPTY_ROOT(&service_tree->rb))
1798                 return NULL;
1799         return cfq_rb_first(service_tree);
1800 }
1801
1802 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1803 {
1804         struct cfq_group *cfqg;
1805         struct cfq_queue *cfqq;
1806         int i, j;
1807         struct cfq_rb_root *st;
1808
1809         if (!cfqd->rq_queued)
1810                 return NULL;
1811
1812         cfqg = cfq_get_next_cfqg(cfqd);
1813         if (!cfqg)
1814                 return NULL;
1815
1816         for_each_cfqg_st(cfqg, i, j, st)
1817                 if ((cfqq = cfq_rb_first(st)) != NULL)
1818                         return cfqq;
1819         return NULL;
1820 }
1821
1822 /*
1823  * Get and set a new active queue for service.
1824  */
1825 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1826                                               struct cfq_queue *cfqq)
1827 {
1828         if (!cfqq)
1829                 cfqq = cfq_get_next_queue(cfqd);
1830
1831         __cfq_set_active_queue(cfqd, cfqq);
1832         return cfqq;
1833 }
1834
1835 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1836                                           struct request *rq)
1837 {
1838         if (blk_rq_pos(rq) >= cfqd->last_position)
1839                 return blk_rq_pos(rq) - cfqd->last_position;
1840         else
1841                 return cfqd->last_position - blk_rq_pos(rq);
1842 }
1843
1844 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1845                                struct request *rq)
1846 {
1847         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1848 }
1849
1850 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1851                                     struct cfq_queue *cur_cfqq)
1852 {
1853         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1854         struct rb_node *parent, *node;
1855         struct cfq_queue *__cfqq;
1856         sector_t sector = cfqd->last_position;
1857
1858         if (RB_EMPTY_ROOT(root))
1859                 return NULL;
1860
1861         /*
1862          * First, if we find a request starting at the end of the last
1863          * request, choose it.
1864          */
1865         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1866         if (__cfqq)
1867                 return __cfqq;
1868
1869         /*
1870          * If the exact sector wasn't found, the parent of the NULL leaf
1871          * will contain the closest sector.
1872          */
1873         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1874         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1875                 return __cfqq;
1876
1877         if (blk_rq_pos(__cfqq->next_rq) < sector)
1878                 node = rb_next(&__cfqq->p_node);
1879         else
1880                 node = rb_prev(&__cfqq->p_node);
1881         if (!node)
1882                 return NULL;
1883
1884         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1885         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1886                 return __cfqq;
1887
1888         return NULL;
1889 }
1890
1891 /*
1892  * cfqd - obvious
1893  * cur_cfqq - passed in so that we don't decide that the current queue is
1894  *            closely cooperating with itself.
1895  *
1896  * So, basically we're assuming that that cur_cfqq has dispatched at least
1897  * one request, and that cfqd->last_position reflects a position on the disk
1898  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1899  * assumption.
1900  */
1901 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1902                                               struct cfq_queue *cur_cfqq)
1903 {
1904         struct cfq_queue *cfqq;
1905
1906         if (cfq_class_idle(cur_cfqq))
1907                 return NULL;
1908         if (!cfq_cfqq_sync(cur_cfqq))
1909                 return NULL;
1910         if (CFQQ_SEEKY(cur_cfqq))
1911                 return NULL;
1912
1913         /*
1914          * Don't search priority tree if it's the only queue in the group.
1915          */
1916         if (cur_cfqq->cfqg->nr_cfqq == 1)
1917                 return NULL;
1918
1919         /*
1920          * We should notice if some of the queues are cooperating, eg
1921          * working closely on the same area of the disk. In that case,
1922          * we can group them together and don't waste time idling.
1923          */
1924         cfqq = cfqq_close(cfqd, cur_cfqq);
1925         if (!cfqq)
1926                 return NULL;
1927
1928         /* If new queue belongs to different cfq_group, don't choose it */
1929         if (cur_cfqq->cfqg != cfqq->cfqg)
1930                 return NULL;
1931
1932         /*
1933          * It only makes sense to merge sync queues.
1934          */
1935         if (!cfq_cfqq_sync(cfqq))
1936                 return NULL;
1937         if (CFQQ_SEEKY(cfqq))
1938                 return NULL;
1939
1940         /*
1941          * Do not merge queues of different priority classes
1942          */
1943         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1944                 return NULL;
1945
1946         return cfqq;
1947 }
1948
1949 /*
1950  * Determine whether we should enforce idle window for this queue.
1951  */
1952
1953 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1954 {
1955         enum wl_prio_t prio = cfqq_prio(cfqq);
1956         struct cfq_rb_root *service_tree = cfqq->service_tree;
1957
1958         BUG_ON(!service_tree);
1959         BUG_ON(!service_tree->count);
1960
1961         if (!cfqd->cfq_slice_idle)
1962                 return false;
1963
1964         /* We never do for idle class queues. */
1965         if (prio == IDLE_WORKLOAD)
1966                 return false;
1967
1968         /* We do for queues that were marked with idle window flag. */
1969         if (cfq_cfqq_idle_window(cfqq) &&
1970            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1971                 return true;
1972
1973         /*
1974          * Otherwise, we do only if they are the last ones
1975          * in their service tree.
1976          */
1977         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq) &&
1978            !cfq_io_thinktime_big(cfqd, &service_tree->ttime, false))
1979                 return true;
1980         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1981                         service_tree->count);
1982         return false;
1983 }
1984
1985 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1986 {
1987         struct cfq_queue *cfqq = cfqd->active_queue;
1988         struct cfq_io_context *cic;
1989         unsigned long sl, group_idle = 0;
1990
1991         /*
1992          * SSD device without seek penalty, disable idling. But only do so
1993          * for devices that support queuing, otherwise we still have a problem
1994          * with sync vs async workloads.
1995          */
1996         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1997                 return;
1998
1999         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
2000         WARN_ON(cfq_cfqq_slice_new(cfqq));
2001
2002         /*
2003          * idle is disabled, either manually or by past process history
2004          */
2005         if (!cfq_should_idle(cfqd, cfqq)) {
2006                 /* no queue idling. Check for group idling */
2007                 if (cfqd->cfq_group_idle)
2008                         group_idle = cfqd->cfq_group_idle;
2009                 else
2010                         return;
2011         }
2012
2013         /*
2014          * still active requests from this queue, don't idle
2015          */
2016         if (cfqq->dispatched)
2017                 return;
2018
2019         /*
2020          * task has exited, don't wait
2021          */
2022         cic = cfqd->active_cic;
2023         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2024                 return;
2025
2026         /*
2027          * If our average think time is larger than the remaining time
2028          * slice, then don't idle. This avoids overrunning the allotted
2029          * time slice.
2030          */
2031         if (sample_valid(cic->ttime.ttime_samples) &&
2032             (cfqq->slice_end - jiffies < cic->ttime.ttime_mean)) {
2033                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%lu",
2034                              cic->ttime.ttime_mean);
2035                 return;
2036         }
2037
2038         /* There are other queues in the group, don't do group idle */
2039         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2040                 return;
2041
2042         cfq_mark_cfqq_wait_request(cfqq);
2043
2044         if (group_idle)
2045                 sl = cfqd->cfq_group_idle;
2046         else
2047                 sl = cfqd->cfq_slice_idle;
2048
2049         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2050         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2051         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2052                         group_idle ? 1 : 0);
2053 }
2054
2055 /*
2056  * Move request from internal lists to the request queue dispatch list.
2057  */
2058 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2059 {
2060         struct cfq_data *cfqd = q->elevator->elevator_data;
2061         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2062
2063         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2064
2065         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2066         cfq_remove_request(rq);
2067         cfqq->dispatched++;
2068         (RQ_CFQG(rq))->dispatched++;
2069         elv_dispatch_sort(q, rq);
2070
2071         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2072         cfqq->nr_sectors += blk_rq_sectors(rq);
2073         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2074                                         rq_data_dir(rq), rq_is_sync(rq));
2075 }
2076
2077 /*
2078  * return expired entry, or NULL to just start from scratch in rbtree
2079  */
2080 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2081 {
2082         struct request *rq = NULL;
2083
2084         if (cfq_cfqq_fifo_expire(cfqq))
2085                 return NULL;
2086
2087         cfq_mark_cfqq_fifo_expire(cfqq);
2088
2089         if (list_empty(&cfqq->fifo))
2090                 return NULL;
2091
2092         rq = rq_entry_fifo(cfqq->fifo.next);
2093         if (time_before(jiffies, rq_fifo_time(rq)))
2094                 rq = NULL;
2095
2096         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2097         return rq;
2098 }
2099
2100 static inline int
2101 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2102 {
2103         const int base_rq = cfqd->cfq_slice_async_rq;
2104
2105         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2106
2107         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2108 }
2109
2110 /*
2111  * Must be called with the queue_lock held.
2112  */
2113 static int cfqq_process_refs(struct cfq_queue *cfqq)
2114 {
2115         int process_refs, io_refs;
2116
2117         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2118         process_refs = cfqq->ref - io_refs;
2119         BUG_ON(process_refs < 0);
2120         return process_refs;
2121 }
2122
2123 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2124 {
2125         int process_refs, new_process_refs;
2126         struct cfq_queue *__cfqq;
2127
2128         /*
2129          * If there are no process references on the new_cfqq, then it is
2130          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2131          * chain may have dropped their last reference (not just their
2132          * last process reference).
2133          */
2134         if (!cfqq_process_refs(new_cfqq))
2135                 return;
2136
2137         /* Avoid a circular list and skip interim queue merges */
2138         while ((__cfqq = new_cfqq->new_cfqq)) {
2139                 if (__cfqq == cfqq)
2140                         return;
2141                 new_cfqq = __cfqq;
2142         }
2143
2144         process_refs = cfqq_process_refs(cfqq);
2145         new_process_refs = cfqq_process_refs(new_cfqq);
2146         /*
2147          * If the process for the cfqq has gone away, there is no
2148          * sense in merging the queues.
2149          */
2150         if (process_refs == 0 || new_process_refs == 0)
2151                 return;
2152
2153         /*
2154          * Merge in the direction of the lesser amount of work.
2155          */
2156         if (new_process_refs >= process_refs) {
2157                 cfqq->new_cfqq = new_cfqq;
2158                 new_cfqq->ref += process_refs;
2159         } else {
2160                 new_cfqq->new_cfqq = cfqq;
2161                 cfqq->ref += new_process_refs;
2162         }
2163 }
2164
2165 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2166                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2167 {
2168         struct cfq_queue *queue;
2169         int i;
2170         bool key_valid = false;
2171         unsigned long lowest_key = 0;
2172         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2173
2174         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2175                 /* select the one with lowest rb_key */
2176                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2177                 if (queue &&
2178                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2179                         lowest_key = queue->rb_key;
2180                         cur_best = i;
2181                         key_valid = true;
2182                 }
2183         }
2184
2185         return cur_best;
2186 }
2187
2188 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2189 {
2190         unsigned slice;
2191         unsigned count;
2192         struct cfq_rb_root *st;
2193         unsigned group_slice;
2194         enum wl_prio_t original_prio = cfqd->serving_prio;
2195
2196         /* Choose next priority. RT > BE > IDLE */
2197         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2198                 cfqd->serving_prio = RT_WORKLOAD;
2199         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2200                 cfqd->serving_prio = BE_WORKLOAD;
2201         else {
2202                 cfqd->serving_prio = IDLE_WORKLOAD;
2203                 cfqd->workload_expires = jiffies + 1;
2204                 return;
2205         }
2206
2207         if (original_prio != cfqd->serving_prio)
2208                 goto new_workload;
2209
2210         /*
2211          * For RT and BE, we have to choose also the type
2212          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2213          * expiration time
2214          */
2215         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2216         count = st->count;
2217
2218         /*
2219          * check workload expiration, and that we still have other queues ready
2220          */
2221         if (count && !time_after(jiffies, cfqd->workload_expires))
2222                 return;
2223
2224 new_workload:
2225         /* otherwise select new workload type */
2226         cfqd->serving_type =
2227                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2228         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2229         count = st->count;
2230
2231         /*
2232          * the workload slice is computed as a fraction of target latency
2233          * proportional to the number of queues in that workload, over
2234          * all the queues in the same priority class
2235          */
2236         group_slice = cfq_group_slice(cfqd, cfqg);
2237
2238         slice = group_slice * count /
2239                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2240                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2241
2242         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2243                 unsigned int tmp;
2244
2245                 /*
2246                  * Async queues are currently system wide. Just taking
2247                  * proportion of queues with-in same group will lead to higher
2248                  * async ratio system wide as generally root group is going
2249                  * to have higher weight. A more accurate thing would be to
2250                  * calculate system wide asnc/sync ratio.
2251                  */
2252                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2253                 tmp = tmp/cfqd->busy_queues;
2254                 slice = min_t(unsigned, slice, tmp);
2255
2256                 /* async workload slice is scaled down according to
2257                  * the sync/async slice ratio. */
2258                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2259         } else
2260                 /* sync workload slice is at least 2 * cfq_slice_idle */
2261                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2262
2263         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2264         cfq_log(cfqd, "workload slice:%d", slice);
2265         cfqd->workload_expires = jiffies + slice;
2266 }
2267
2268 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2269 {
2270         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2271         struct cfq_group *cfqg;
2272
2273         if (RB_EMPTY_ROOT(&st->rb))
2274                 return NULL;
2275         cfqg = cfq_rb_first_group(st);
2276         update_min_vdisktime(st);
2277         return cfqg;
2278 }
2279
2280 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2281 {
2282         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2283
2284         cfqd->serving_group = cfqg;
2285
2286         /* Restore the workload type data */
2287         if (cfqg->saved_workload_slice) {
2288                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2289                 cfqd->serving_type = cfqg->saved_workload;
2290                 cfqd->serving_prio = cfqg->saved_serving_prio;
2291         } else
2292                 cfqd->workload_expires = jiffies - 1;
2293
2294         choose_service_tree(cfqd, cfqg);
2295 }
2296
2297 /*
2298  * Select a queue for service. If we have a current active queue,
2299  * check whether to continue servicing it, or retrieve and set a new one.
2300  */
2301 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2302 {
2303         struct cfq_queue *cfqq, *new_cfqq = NULL;
2304
2305         cfqq = cfqd->active_queue;
2306         if (!cfqq)
2307                 goto new_queue;
2308
2309         if (!cfqd->rq_queued)
2310                 return NULL;
2311
2312         /*
2313          * We were waiting for group to get backlogged. Expire the queue
2314          */
2315         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2316                 goto expire;
2317
2318         /*
2319          * The active queue has run out of time, expire it and select new.
2320          */
2321         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2322                 /*
2323                  * If slice had not expired at the completion of last request
2324                  * we might not have turned on wait_busy flag. Don't expire
2325                  * the queue yet. Allow the group to get backlogged.
2326                  *
2327                  * The very fact that we have used the slice, that means we
2328                  * have been idling all along on this queue and it should be
2329                  * ok to wait for this request to complete.
2330                  */
2331                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2332                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2333                         cfqq = NULL;
2334                         goto keep_queue;
2335                 } else
2336                         goto check_group_idle;
2337         }
2338
2339         /*
2340          * The active queue has requests and isn't expired, allow it to
2341          * dispatch.
2342          */
2343         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2344                 goto keep_queue;
2345
2346         /*
2347          * If another queue has a request waiting within our mean seek
2348          * distance, let it run.  The expire code will check for close
2349          * cooperators and put the close queue at the front of the service
2350          * tree.  If possible, merge the expiring queue with the new cfqq.
2351          */
2352         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2353         if (new_cfqq) {
2354                 if (!cfqq->new_cfqq)
2355                         cfq_setup_merge(cfqq, new_cfqq);
2356                 goto expire;
2357         }
2358
2359         /*
2360          * No requests pending. If the active queue still has requests in
2361          * flight or is idling for a new request, allow either of these
2362          * conditions to happen (or time out) before selecting a new queue.
2363          */
2364         if (timer_pending(&cfqd->idle_slice_timer)) {
2365                 cfqq = NULL;
2366                 goto keep_queue;
2367         }
2368
2369         /*
2370          * This is a deep seek queue, but the device is much faster than
2371          * the queue can deliver, don't idle
2372          **/
2373         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2374             (cfq_cfqq_slice_new(cfqq) ||
2375             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2376                 cfq_clear_cfqq_deep(cfqq);
2377                 cfq_clear_cfqq_idle_window(cfqq);
2378         }
2379
2380         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2381                 cfqq = NULL;
2382                 goto keep_queue;
2383         }
2384
2385         /*
2386          * If group idle is enabled and there are requests dispatched from
2387          * this group, wait for requests to complete.
2388          */
2389 check_group_idle:
2390         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1 &&
2391             cfqq->cfqg->dispatched &&
2392             !cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true)) {
2393                 cfqq = NULL;
2394                 goto keep_queue;
2395         }
2396
2397 expire:
2398         cfq_slice_expired(cfqd, 0);
2399 new_queue:
2400         /*
2401          * Current queue expired. Check if we have to switch to a new
2402          * service tree
2403          */
2404         if (!new_cfqq)
2405                 cfq_choose_cfqg(cfqd);
2406
2407         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2408 keep_queue:
2409         return cfqq;
2410 }
2411
2412 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2413 {
2414         int dispatched = 0;
2415
2416         while (cfqq->next_rq) {
2417                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2418                 dispatched++;
2419         }
2420
2421         BUG_ON(!list_empty(&cfqq->fifo));
2422
2423         /* By default cfqq is not expired if it is empty. Do it explicitly */
2424         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2425         return dispatched;
2426 }
2427
2428 /*
2429  * Drain our current requests. Used for barriers and when switching
2430  * io schedulers on-the-fly.
2431  */
2432 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2433 {
2434         struct cfq_queue *cfqq;
2435         int dispatched = 0;
2436
2437         /* Expire the timeslice of the current active queue first */
2438         cfq_slice_expired(cfqd, 0);
2439         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2440                 __cfq_set_active_queue(cfqd, cfqq);
2441                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2442         }
2443
2444         BUG_ON(cfqd->busy_queues);
2445
2446         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2447         return dispatched;
2448 }
2449
2450 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2451         struct cfq_queue *cfqq)
2452 {
2453         /* the queue hasn't finished any request, can't estimate */
2454         if (cfq_cfqq_slice_new(cfqq))
2455                 return true;
2456         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2457                 cfqq->slice_end))
2458                 return true;
2459
2460         return false;
2461 }
2462
2463 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2464 {
2465         unsigned int max_dispatch;
2466
2467         /*
2468          * Drain async requests before we start sync IO
2469          */
2470         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2471                 return false;
2472
2473         /*
2474          * If this is an async queue and we have sync IO in flight, let it wait
2475          */
2476         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2477                 return false;
2478
2479         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2480         if (cfq_class_idle(cfqq))
2481                 max_dispatch = 1;
2482
2483         /*
2484          * Does this cfqq already have too much IO in flight?
2485          */
2486         if (cfqq->dispatched >= max_dispatch) {
2487                 bool promote_sync = false;
2488                 /*
2489                  * idle queue must always only have a single IO in flight
2490                  */
2491                 if (cfq_class_idle(cfqq))
2492                         return false;
2493
2494                 /*
2495                  * If there is only one sync queue
2496                  * we can ignore async queue here and give the sync
2497                  * queue no dispatch limit. The reason is a sync queue can
2498                  * preempt async queue, limiting the sync queue doesn't make
2499                  * sense. This is useful for aiostress test.
2500                  */
2501                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2502                         promote_sync = true;
2503
2504                 /*
2505                  * We have other queues, don't allow more IO from this one
2506                  */
2507                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2508                                 !promote_sync)
2509                         return false;
2510
2511                 /*
2512                  * Sole queue user, no limit
2513                  */
2514                 if (cfqd->busy_queues == 1 || promote_sync)
2515                         max_dispatch = -1;
2516                 else
2517                         /*
2518                          * Normally we start throttling cfqq when cfq_quantum/2
2519                          * requests have been dispatched. But we can drive
2520                          * deeper queue depths at the beginning of slice
2521                          * subjected to upper limit of cfq_quantum.
2522                          * */
2523                         max_dispatch = cfqd->cfq_quantum;
2524         }
2525
2526         /*
2527          * Async queues must wait a bit before being allowed dispatch.
2528          * We also ramp up the dispatch depth gradually for async IO,
2529          * based on the last sync IO we serviced
2530          */
2531         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2532                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2533                 unsigned int depth;
2534
2535                 depth = last_sync / cfqd->cfq_slice[1];
2536                 if (!depth && !cfqq->dispatched)
2537                         depth = 1;
2538                 if (depth < max_dispatch)
2539                         max_dispatch = depth;
2540         }
2541
2542         /*
2543          * If we're below the current max, allow a dispatch
2544          */
2545         return cfqq->dispatched < max_dispatch;
2546 }
2547
2548 /*
2549  * Dispatch a request from cfqq, moving them to the request queue
2550  * dispatch list.
2551  */
2552 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2553 {
2554         struct request *rq;
2555
2556         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2557
2558         if (!cfq_may_dispatch(cfqd, cfqq))
2559                 return false;
2560
2561         /*
2562          * follow expired path, else get first next available
2563          */
2564         rq = cfq_check_fifo(cfqq);
2565         if (!rq)
2566                 rq = cfqq->next_rq;
2567
2568         /*
2569          * insert request into driver dispatch list
2570          */
2571         cfq_dispatch_insert(cfqd->queue, rq);
2572
2573         if (!cfqd->active_cic) {
2574                 struct cfq_io_context *cic = RQ_CIC(rq);
2575
2576                 atomic_long_inc(&cic->ioc->refcount);
2577                 cfqd->active_cic = cic;
2578         }
2579
2580         return true;
2581 }
2582
2583 /*
2584  * Find the cfqq that we need to service and move a request from that to the
2585  * dispatch list
2586  */
2587 static int cfq_dispatch_requests(struct request_queue *q, int force)
2588 {
2589         struct cfq_data *cfqd = q->elevator->elevator_data;
2590         struct cfq_queue *cfqq;
2591
2592         if (!cfqd->busy_queues)
2593                 return 0;
2594
2595         if (unlikely(force))
2596                 return cfq_forced_dispatch(cfqd);
2597
2598         cfqq = cfq_select_queue(cfqd);
2599         if (!cfqq)
2600                 return 0;
2601
2602         /*
2603          * Dispatch a request from this cfqq, if it is allowed
2604          */
2605         if (!cfq_dispatch_request(cfqd, cfqq))
2606                 return 0;
2607
2608         cfqq->slice_dispatch++;
2609         cfq_clear_cfqq_must_dispatch(cfqq);
2610
2611         /*
2612          * expire an async queue immediately if it has used up its slice. idle
2613          * queue always expire after 1 dispatch round.
2614          */
2615         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2616             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2617             cfq_class_idle(cfqq))) {
2618                 cfqq->slice_end = jiffies + 1;
2619                 cfq_slice_expired(cfqd, 0);
2620         }
2621
2622         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2623         return 1;
2624 }
2625
2626 /*
2627  * task holds one reference to the queue, dropped when task exits. each rq
2628  * in-flight on this queue also holds a reference, dropped when rq is freed.
2629  *
2630  * Each cfq queue took a reference on the parent group. Drop it now.
2631  * queue lock must be held here.
2632  */
2633 static void cfq_put_queue(struct cfq_queue *cfqq)
2634 {
2635         struct cfq_data *cfqd = cfqq->cfqd;
2636         struct cfq_group *cfqg;
2637
2638         BUG_ON(cfqq->ref <= 0);
2639
2640         cfqq->ref--;
2641         if (cfqq->ref)
2642                 return;
2643
2644         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2645         BUG_ON(rb_first(&cfqq->sort_list));
2646         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2647         cfqg = cfqq->cfqg;
2648
2649         if (unlikely(cfqd->active_queue == cfqq)) {
2650                 __cfq_slice_expired(cfqd, cfqq, 0);
2651                 cfq_schedule_dispatch(cfqd);
2652         }
2653
2654         BUG_ON(cfq_cfqq_on_rr(cfqq));
2655         kmem_cache_free(cfq_pool, cfqq);
2656         cfq_put_cfqg(cfqg);
2657 }
2658
2659 /*
2660  * Call func for each cic attached to this ioc.
2661  */
2662 static void
2663 call_for_each_cic(struct io_context *ioc,
2664                   void (*func)(struct io_context *, struct cfq_io_context *))
2665 {
2666         struct cfq_io_context *cic;
2667         struct hlist_node *n;
2668
2669         rcu_read_lock();
2670
2671         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2672                 func(ioc, cic);
2673
2674         rcu_read_unlock();
2675 }
2676
2677 static void cfq_cic_free_rcu(struct rcu_head *head)
2678 {
2679         struct cfq_io_context *cic;
2680
2681         cic = container_of(head, struct cfq_io_context, rcu_head);
2682
2683         kmem_cache_free(cfq_ioc_pool, cic);
2684         elv_ioc_count_dec(cfq_ioc_count);
2685
2686         if (ioc_gone) {
2687                 /*
2688                  * CFQ scheduler is exiting, grab exit lock and check
2689                  * the pending io context count. If it hits zero,
2690                  * complete ioc_gone and set it back to NULL
2691                  */
2692                 spin_lock(&ioc_gone_lock);
2693                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2694                         complete(ioc_gone);
2695                         ioc_gone = NULL;
2696                 }
2697                 spin_unlock(&ioc_gone_lock);
2698         }
2699 }
2700
2701 static void cfq_cic_free(struct cfq_io_context *cic)
2702 {
2703         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2704 }
2705
2706 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2707 {
2708         unsigned long flags;
2709         unsigned long dead_key = (unsigned long) cic->key;
2710
2711         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2712
2713         spin_lock_irqsave(&ioc->lock, flags);
2714         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2715         hlist_del_rcu(&cic->cic_list);
2716         spin_unlock_irqrestore(&ioc->lock, flags);
2717
2718         cfq_cic_free(cic);
2719 }
2720
2721 /*
2722  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2723  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2724  * and ->trim() which is called with the task lock held
2725  */
2726 static void cfq_free_io_context(struct io_context *ioc)
2727 {
2728         /*
2729          * ioc->refcount is zero here, or we are called from elv_unregister(),
2730          * so no more cic's are allowed to be linked into this ioc.  So it
2731          * should be ok to iterate over the known list, we will see all cic's
2732          * since no new ones are added.
2733          */
2734         call_for_each_cic(ioc, cic_free_func);
2735 }
2736
2737 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2738 {
2739         struct cfq_queue *__cfqq, *next;
2740
2741         /*
2742          * If this queue was scheduled to merge with another queue, be
2743          * sure to drop the reference taken on that queue (and others in
2744          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2745          */
2746         __cfqq = cfqq->new_cfqq;
2747         while (__cfqq) {
2748                 if (__cfqq == cfqq) {
2749                         WARN(1, "cfqq->new_cfqq loop detected\n");
2750                         break;
2751                 }
2752                 next = __cfqq->new_cfqq;
2753                 cfq_put_queue(__cfqq);
2754                 __cfqq = next;
2755         }
2756 }
2757
2758 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2759 {
2760         if (unlikely(cfqq == cfqd->active_queue)) {
2761                 __cfq_slice_expired(cfqd, cfqq, 0);
2762                 cfq_schedule_dispatch(cfqd);
2763         }
2764
2765         cfq_put_cooperator(cfqq);
2766
2767         cfq_put_queue(cfqq);
2768 }
2769
2770 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2771                                          struct cfq_io_context *cic)
2772 {
2773         struct io_context *ioc = cic->ioc;
2774
2775         list_del_init(&cic->queue_list);
2776
2777         /*
2778          * Make sure dead mark is seen for dead queues
2779          */
2780         smp_wmb();
2781         cic->key = cfqd_dead_key(cfqd);
2782
2783         rcu_read_lock();
2784         if (rcu_dereference(ioc->ioc_data) == cic) {
2785                 rcu_read_unlock();
2786                 spin_lock(&ioc->lock);
2787                 rcu_assign_pointer(ioc->ioc_data, NULL);
2788                 spin_unlock(&ioc->lock);
2789         } else
2790                 rcu_read_unlock();
2791
2792         if (cic->cfqq[BLK_RW_ASYNC]) {
2793                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2794                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2795         }
2796
2797         if (cic->cfqq[BLK_RW_SYNC]) {
2798                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2799                 cic->cfqq[BLK_RW_SYNC] = NULL;
2800         }
2801 }
2802
2803 static void cfq_exit_single_io_context(struct io_context *ioc,
2804                                        struct cfq_io_context *cic)
2805 {
2806         struct cfq_data *cfqd = cic_to_cfqd(cic);
2807
2808         if (cfqd) {
2809                 struct request_queue *q = cfqd->queue;
2810                 unsigned long flags;
2811
2812                 spin_lock_irqsave(q->queue_lock, flags);
2813
2814                 /*
2815                  * Ensure we get a fresh copy of the ->key to prevent
2816                  * race between exiting task and queue
2817                  */
2818                 smp_read_barrier_depends();
2819                 if (cic->key == cfqd)
2820                         __cfq_exit_single_io_context(cfqd, cic);
2821
2822                 spin_unlock_irqrestore(q->queue_lock, flags);
2823         }
2824 }
2825
2826 /*
2827  * The process that ioc belongs to has exited, we need to clean up
2828  * and put the internal structures we have that belongs to that process.
2829  */
2830 static void cfq_exit_io_context(struct io_context *ioc)
2831 {
2832         call_for_each_cic(ioc, cfq_exit_single_io_context);
2833 }
2834
2835 static struct cfq_io_context *
2836 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2837 {
2838         struct cfq_io_context *cic;
2839
2840         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2841                                                         cfqd->queue->node);
2842         if (cic) {
2843                 cic->ttime.last_end_request = jiffies;
2844                 INIT_LIST_HEAD(&cic->queue_list);
2845                 INIT_HLIST_NODE(&cic->cic_list);
2846                 cic->dtor = cfq_free_io_context;
2847                 cic->exit = cfq_exit_io_context;
2848                 elv_ioc_count_inc(cfq_ioc_count);
2849         }
2850
2851         return cic;
2852 }
2853
2854 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2855 {
2856         struct task_struct *tsk = current;
2857         int ioprio_class;
2858
2859         if (!cfq_cfqq_prio_changed(cfqq))
2860                 return;
2861
2862         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2863         switch (ioprio_class) {
2864         default:
2865                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2866         case IOPRIO_CLASS_NONE:
2867                 /*
2868                  * no prio set, inherit CPU scheduling settings
2869                  */
2870                 cfqq->ioprio = task_nice_ioprio(tsk);
2871                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2872                 break;
2873         case IOPRIO_CLASS_RT:
2874                 cfqq->ioprio = task_ioprio(ioc);
2875                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2876                 break;
2877         case IOPRIO_CLASS_BE:
2878                 cfqq->ioprio = task_ioprio(ioc);
2879                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2880                 break;
2881         case IOPRIO_CLASS_IDLE:
2882                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2883                 cfqq->ioprio = 7;
2884                 cfq_clear_cfqq_idle_window(cfqq);
2885                 break;
2886         }
2887
2888         /*
2889          * keep track of original prio settings in case we have to temporarily
2890          * elevate the priority of this queue
2891          */
2892         cfqq->org_ioprio = cfqq->ioprio;
2893         cfq_clear_cfqq_prio_changed(cfqq);
2894 }
2895
2896 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2897 {
2898         struct cfq_data *cfqd = cic_to_cfqd(cic);
2899         struct cfq_queue *cfqq;
2900         unsigned long flags;
2901
2902         if (unlikely(!cfqd))
2903                 return;
2904
2905         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2906
2907         cfqq = cic->cfqq[BLK_RW_ASYNC];
2908         if (cfqq) {
2909                 struct cfq_queue *new_cfqq;
2910                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2911                                                 GFP_ATOMIC);
2912                 if (new_cfqq) {
2913                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2914                         cfq_put_queue(cfqq);
2915                 }
2916         }
2917
2918         cfqq = cic->cfqq[BLK_RW_SYNC];
2919         if (cfqq)
2920                 cfq_mark_cfqq_prio_changed(cfqq);
2921
2922         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2923 }
2924
2925 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2926 {
2927         call_for_each_cic(ioc, changed_ioprio);
2928         ioc->ioprio_changed = 0;
2929 }
2930
2931 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2932                           pid_t pid, bool is_sync)
2933 {
2934         RB_CLEAR_NODE(&cfqq->rb_node);
2935         RB_CLEAR_NODE(&cfqq->p_node);
2936         INIT_LIST_HEAD(&cfqq->fifo);
2937
2938         cfqq->ref = 0;
2939         cfqq->cfqd = cfqd;
2940
2941         cfq_mark_cfqq_prio_changed(cfqq);
2942
2943         if (is_sync) {
2944                 if (!cfq_class_idle(cfqq))
2945                         cfq_mark_cfqq_idle_window(cfqq);
2946                 cfq_mark_cfqq_sync(cfqq);
2947         }
2948         cfqq->pid = pid;
2949 }
2950
2951 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2952 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2953 {
2954         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2955         struct cfq_data *cfqd = cic_to_cfqd(cic);
2956         unsigned long flags;
2957         struct request_queue *q;
2958
2959         if (unlikely(!cfqd))
2960                 return;
2961
2962         q = cfqd->queue;
2963
2964         spin_lock_irqsave(q->queue_lock, flags);
2965
2966         if (sync_cfqq) {
2967                 /*
2968                  * Drop reference to sync queue. A new sync queue will be
2969                  * assigned in new group upon arrival of a fresh request.
2970                  */
2971                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2972                 cic_set_cfqq(cic, NULL, 1);
2973                 cfq_put_queue(sync_cfqq);
2974         }
2975
2976         spin_unlock_irqrestore(q->queue_lock, flags);
2977 }
2978
2979 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2980 {
2981         call_for_each_cic(ioc, changed_cgroup);
2982         ioc->cgroup_changed = 0;
2983 }
2984 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2985
2986 static struct cfq_queue *
2987 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2988                      struct io_context *ioc, gfp_t gfp_mask)
2989 {
2990         struct cfq_queue *cfqq, *new_cfqq = NULL;
2991         struct cfq_io_context *cic;
2992         struct cfq_group *cfqg;
2993
2994 retry:
2995         cfqg = cfq_get_cfqg(cfqd);
2996         cic = cfq_cic_lookup(cfqd, ioc);
2997         /* cic always exists here */
2998         cfqq = cic_to_cfqq(cic, is_sync);
2999
3000         /*
3001          * Always try a new alloc if we fell back to the OOM cfqq
3002          * originally, since it should just be a temporary situation.
3003          */
3004         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3005                 cfqq = NULL;
3006                 if (new_cfqq) {
3007                         cfqq = new_cfqq;
3008                         new_cfqq = NULL;
3009                 } else if (gfp_mask & __GFP_WAIT) {
3010                         spin_unlock_irq(cfqd->queue->queue_lock);
3011                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
3012                                         gfp_mask | __GFP_ZERO,
3013                                         cfqd->queue->node);
3014                         spin_lock_irq(cfqd->queue->queue_lock);
3015                         if (new_cfqq)
3016                                 goto retry;
3017                 } else {
3018                         cfqq = kmem_cache_alloc_node(cfq_pool,
3019                                         gfp_mask | __GFP_ZERO,
3020                                         cfqd->queue->node);
3021                 }
3022
3023                 if (cfqq) {
3024                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3025                         cfq_init_prio_data(cfqq, ioc);
3026                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3027                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3028                 } else
3029                         cfqq = &cfqd->oom_cfqq;
3030         }
3031
3032         if (new_cfqq)
3033                 kmem_cache_free(cfq_pool, new_cfqq);
3034
3035         return cfqq;
3036 }
3037
3038 static struct cfq_queue **
3039 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3040 {
3041         switch (ioprio_class) {
3042         case IOPRIO_CLASS_RT:
3043                 return &cfqd->async_cfqq[0][ioprio];
3044         case IOPRIO_CLASS_BE:
3045                 return &cfqd->async_cfqq[1][ioprio];
3046         case IOPRIO_CLASS_IDLE:
3047                 return &cfqd->async_idle_cfqq;
3048         default:
3049                 BUG();
3050         }
3051 }
3052
3053 static struct cfq_queue *
3054 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3055               gfp_t gfp_mask)
3056 {
3057         const int ioprio = task_ioprio(ioc);
3058         const int ioprio_class = task_ioprio_class(ioc);
3059         struct cfq_queue **async_cfqq = NULL;
3060         struct cfq_queue *cfqq = NULL;
3061
3062         if (!is_sync) {
3063                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3064                 cfqq = *async_cfqq;
3065         }
3066
3067         if (!cfqq)
3068                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3069
3070         /*
3071          * pin the queue now that it's allocated, scheduler exit will prune it
3072          */
3073         if (!is_sync && !(*async_cfqq)) {
3074                 cfqq->ref++;
3075                 *async_cfqq = cfqq;
3076         }
3077
3078         cfqq->ref++;
3079         return cfqq;
3080 }
3081
3082 /*
3083  * We drop cfq io contexts lazily, so we may find a dead one.
3084  */
3085 static void
3086 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3087                   struct cfq_io_context *cic)
3088 {
3089         unsigned long flags;
3090
3091         WARN_ON(!list_empty(&cic->queue_list));
3092         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3093
3094         spin_lock_irqsave(&ioc->lock, flags);
3095
3096         BUG_ON(rcu_dereference_check(ioc->ioc_data,
3097                 lockdep_is_held(&ioc->lock)) == cic);
3098
3099         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3100         hlist_del_rcu(&cic->cic_list);
3101         spin_unlock_irqrestore(&ioc->lock, flags);
3102
3103         cfq_cic_free(cic);
3104 }
3105
3106 static struct cfq_io_context *
3107 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3108 {
3109         struct cfq_io_context *cic;
3110         unsigned long flags;
3111
3112         if (unlikely(!ioc))
3113                 return NULL;
3114
3115         rcu_read_lock();
3116
3117         /*
3118          * we maintain a last-hit cache, to avoid browsing over the tree
3119          */
3120         cic = rcu_dereference(ioc->ioc_data);
3121         if (cic && cic->key == cfqd) {
3122                 rcu_read_unlock();
3123                 return cic;
3124         }
3125
3126         do {
3127                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3128                 rcu_read_unlock();
3129                 if (!cic)
3130                         break;
3131                 if (unlikely(cic->key != cfqd)) {
3132                         cfq_drop_dead_cic(cfqd, ioc, cic);
3133                         rcu_read_lock();
3134                         continue;
3135                 }
3136
3137                 spin_lock_irqsave(&ioc->lock, flags);
3138                 rcu_assign_pointer(ioc->ioc_data, cic);
3139                 spin_unlock_irqrestore(&ioc->lock, flags);
3140                 break;
3141         } while (1);
3142
3143         return cic;
3144 }
3145
3146 /*
3147  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3148  * the process specific cfq io context when entered from the block layer.
3149  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3150  */
3151 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3152                         struct cfq_io_context *cic, gfp_t gfp_mask)
3153 {
3154         unsigned long flags;
3155         int ret;
3156
3157         ret = radix_tree_preload(gfp_mask);
3158         if (!ret) {
3159                 cic->ioc = ioc;
3160                 cic->key = cfqd;
3161
3162                 spin_lock_irqsave(&ioc->lock, flags);
3163                 ret = radix_tree_insert(&ioc->radix_root,
3164                                                 cfqd->cic_index, cic);
3165                 if (!ret)
3166                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3167                 spin_unlock_irqrestore(&ioc->lock, flags);
3168
3169                 radix_tree_preload_end();
3170
3171                 if (!ret) {
3172                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3173                         list_add(&cic->queue_list, &cfqd->cic_list);
3174                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3175                 }
3176         }
3177
3178         if (ret)
3179                 printk(KERN_ERR "cfq: cic link failed!\n");
3180
3181         return ret;
3182 }
3183
3184 /*
3185  * Setup general io context and cfq io context. There can be several cfq
3186  * io contexts per general io context, if this process is doing io to more
3187  * than one device managed by cfq.
3188  */
3189 static struct cfq_io_context *
3190 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3191 {
3192         struct io_context *ioc = NULL;
3193         struct cfq_io_context *cic;
3194
3195         might_sleep_if(gfp_mask & __GFP_WAIT);
3196
3197         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3198         if (!ioc)
3199                 return NULL;
3200
3201         cic = cfq_cic_lookup(cfqd, ioc);
3202         if (cic)
3203                 goto out;
3204
3205         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3206         if (cic == NULL)
3207                 goto err;
3208
3209         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3210                 goto err_free;
3211
3212 out:
3213         smp_read_barrier_depends();
3214         if (unlikely(ioc->ioprio_changed))
3215                 cfq_ioc_set_ioprio(ioc);
3216
3217 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3218         if (unlikely(ioc->cgroup_changed))
3219                 cfq_ioc_set_cgroup(ioc);
3220 #endif
3221         return cic;
3222 err_free:
3223         cfq_cic_free(cic);
3224 err:
3225         put_io_context(ioc);
3226         return NULL;
3227 }
3228
3229 static void
3230 __cfq_update_io_thinktime(struct cfq_ttime *ttime, unsigned long slice_idle)
3231 {
3232         unsigned long elapsed = jiffies - ttime->last_end_request;
3233         elapsed = min(elapsed, 2UL * slice_idle);
3234
3235         ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
3236         ttime->ttime_total = (7*ttime->ttime_total + 256*elapsed) / 8;
3237         ttime->ttime_mean = (ttime->ttime_total + 128) / ttime->ttime_samples;
3238 }
3239
3240 static void
3241 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3242         struct cfq_io_context *cic)
3243 {
3244         if (cfq_cfqq_sync(cfqq)) {
3245                 __cfq_update_io_thinktime(&cic->ttime, cfqd->cfq_slice_idle);
3246                 __cfq_update_io_thinktime(&cfqq->service_tree->ttime,
3247                         cfqd->cfq_slice_idle);
3248         }
3249 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3250         __cfq_update_io_thinktime(&cfqq->cfqg->ttime, cfqd->cfq_group_idle);
3251 #endif
3252 }
3253
3254 static void
3255 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3256                        struct request *rq)
3257 {
3258         sector_t sdist = 0;
3259         sector_t n_sec = blk_rq_sectors(rq);
3260         if (cfqq->last_request_pos) {
3261                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3262                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3263                 else
3264                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3265         }
3266
3267         cfqq->seek_history <<= 1;
3268         if (blk_queue_nonrot(cfqd->queue))
3269                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3270         else
3271                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3272 }
3273
3274 /*
3275  * Disable idle window if the process thinks too long or seeks so much that
3276  * it doesn't matter
3277  */
3278 static void
3279 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3280                        struct cfq_io_context *cic)
3281 {
3282         int old_idle, enable_idle;
3283
3284         /*
3285          * Don't idle for async or idle io prio class
3286          */
3287         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3288                 return;
3289
3290         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3291
3292         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3293                 cfq_mark_cfqq_deep(cfqq);
3294
3295         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3296                 enable_idle = 0;
3297         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3298             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3299                 enable_idle = 0;
3300         else if (sample_valid(cic->ttime.ttime_samples)) {
3301                 if (cic->ttime.ttime_mean > cfqd->cfq_slice_idle)
3302                         enable_idle = 0;
3303                 else
3304                         enable_idle = 1;
3305         }
3306
3307         if (old_idle != enable_idle) {
3308                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3309                 if (enable_idle)
3310                         cfq_mark_cfqq_idle_window(cfqq);
3311                 else
3312                         cfq_clear_cfqq_idle_window(cfqq);
3313         }
3314 }
3315
3316 /*
3317  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3318  * no or if we aren't sure, a 1 will cause a preempt.
3319  */
3320 static bool
3321 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3322                    struct request *rq)
3323 {
3324         struct cfq_queue *cfqq;
3325
3326         cfqq = cfqd->active_queue;
3327         if (!cfqq)
3328                 return false;
3329
3330         if (cfq_class_idle(new_cfqq))
3331                 return false;
3332
3333         if (cfq_class_idle(cfqq))
3334                 return true;
3335
3336         /*
3337          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3338          */
3339         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3340                 return false;
3341
3342         /*
3343          * if the new request is sync, but the currently running queue is
3344          * not, let the sync request have priority.
3345          */
3346         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3347                 return true;
3348
3349         if (new_cfqq->cfqg != cfqq->cfqg)
3350                 return false;
3351
3352         if (cfq_slice_used(cfqq))
3353                 return true;
3354
3355         /* Allow preemption only if we are idling on sync-noidle tree */
3356         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3357             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3358             new_cfqq->service_tree->count == 2 &&
3359             RB_EMPTY_ROOT(&cfqq->sort_list))
3360                 return true;
3361
3362         /*
3363          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3364          */
3365         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3366                 return true;
3367
3368         /* An idle queue should not be idle now for some reason */
3369         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3370                 return true;
3371
3372         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3373                 return false;
3374
3375         /*
3376          * if this request is as-good as one we would expect from the
3377          * current cfqq, let it preempt
3378          */
3379         if (cfq_rq_close(cfqd, cfqq, rq))
3380                 return true;
3381
3382         return false;
3383 }
3384
3385 /*
3386  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3387  * let it have half of its nominal slice.
3388  */
3389 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3390 {
3391         struct cfq_queue *old_cfqq = cfqd->active_queue;
3392
3393         cfq_log_cfqq(cfqd, cfqq, "preempt");
3394         cfq_slice_expired(cfqd, 1);
3395
3396         /*
3397          * workload type is changed, don't save slice, otherwise preempt
3398          * doesn't happen
3399          */
3400         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3401                 cfqq->cfqg->saved_workload_slice = 0;
3402
3403         /*
3404          * Put the new queue at the front of the of the current list,
3405          * so we know that it will be selected next.
3406          */
3407         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3408
3409         cfq_service_tree_add(cfqd, cfqq, 1);
3410
3411         cfqq->slice_end = 0;
3412         cfq_mark_cfqq_slice_new(cfqq);
3413 }
3414
3415 /*
3416  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3417  * something we should do about it
3418  */
3419 static void
3420 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3421                 struct request *rq)
3422 {
3423         struct cfq_io_context *cic = RQ_CIC(rq);
3424
3425         cfqd->rq_queued++;
3426
3427         cfq_update_io_thinktime(cfqd, cfqq, cic);
3428         cfq_update_io_seektime(cfqd, cfqq, rq);
3429         cfq_update_idle_window(cfqd, cfqq, cic);
3430
3431         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3432
3433         if (cfqq == cfqd->active_queue) {
3434                 /*
3435                  * Remember that we saw a request from this process, but
3436                  * don't start queuing just yet. Otherwise we risk seeing lots
3437                  * of tiny requests, because we disrupt the normal plugging
3438                  * and merging. If the request is already larger than a single
3439                  * page, let it rip immediately. For that case we assume that
3440                  * merging is already done. Ditto for a busy system that
3441                  * has other work pending, don't risk delaying until the
3442                  * idle timer unplug to continue working.
3443                  */
3444                 if (cfq_cfqq_wait_request(cfqq)) {
3445                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3446                             cfqd->busy_queues > 1) {
3447                                 cfq_del_timer(cfqd, cfqq);
3448                                 cfq_clear_cfqq_wait_request(cfqq);
3449                                 __blk_run_queue(cfqd->queue);
3450                         } else {
3451                                 cfq_blkiocg_update_idle_time_stats(
3452                                                 &cfqq->cfqg->blkg);
3453                                 cfq_mark_cfqq_must_dispatch(cfqq);
3454                         }
3455                 }
3456         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3457                 /*
3458                  * not the active queue - expire current slice if it is
3459                  * idle and has expired it's mean thinktime or this new queue
3460                  * has some old slice time left and is of higher priority or
3461                  * this new queue is RT and the current one is BE
3462                  */
3463                 cfq_preempt_queue(cfqd, cfqq);
3464                 __blk_run_queue(cfqd->queue);
3465         }
3466 }
3467
3468 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3469 {
3470         struct cfq_data *cfqd = q->elevator->elevator_data;
3471         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3472
3473         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3474         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3475
3476         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3477         list_add_tail(&rq->queuelist, &cfqq->fifo);
3478         cfq_add_rq_rb(rq);
3479         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3480                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3481                         rq_is_sync(rq));
3482         cfq_rq_enqueued(cfqd, cfqq, rq);
3483 }
3484
3485 /*
3486  * Update hw_tag based on peak queue depth over 50 samples under
3487  * sufficient load.
3488  */
3489 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3490 {
3491         struct cfq_queue *cfqq = cfqd->active_queue;
3492
3493         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3494                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3495
3496         if (cfqd->hw_tag == 1)
3497                 return;
3498
3499         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3500             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3501                 return;
3502
3503         /*
3504          * If active queue hasn't enough requests and can idle, cfq might not
3505          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3506          * case
3507          */
3508         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3509             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3510             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3511                 return;
3512
3513         if (cfqd->hw_tag_samples++ < 50)
3514                 return;
3515
3516         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3517                 cfqd->hw_tag = 1;
3518         else
3519                 cfqd->hw_tag = 0;
3520 }
3521
3522 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3523 {
3524         struct cfq_io_context *cic = cfqd->active_cic;
3525
3526         /* If the queue already has requests, don't wait */
3527         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3528                 return false;
3529
3530         /* If there are other queues in the group, don't wait */
3531         if (cfqq->cfqg->nr_cfqq > 1)
3532                 return false;
3533
3534         /* the only queue in the group, but think time is big */
3535         if (cfq_io_thinktime_big(cfqd, &cfqq->cfqg->ttime, true))
3536                 return false;
3537
3538         if (cfq_slice_used(cfqq))
3539                 return true;
3540
3541         /* if slice left is less than think time, wait busy */
3542         if (cic && sample_valid(cic->ttime.ttime_samples)
3543             && (cfqq->slice_end - jiffies < cic->ttime.ttime_mean))
3544                 return true;
3545
3546         /*
3547          * If think times is less than a jiffy than ttime_mean=0 and above
3548          * will not be true. It might happen that slice has not expired yet
3549          * but will expire soon (4-5 ns) during select_queue(). To cover the
3550          * case where think time is less than a jiffy, mark the queue wait
3551          * busy if only 1 jiffy is left in the slice.
3552          */
3553         if (cfqq->slice_end - jiffies == 1)
3554                 return true;
3555
3556         return false;
3557 }
3558
3559 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3560 {
3561         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3562         struct cfq_data *cfqd = cfqq->cfqd;
3563         const int sync = rq_is_sync(rq);
3564         unsigned long now;
3565
3566         now = jiffies;
3567         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3568                      !!(rq->cmd_flags & REQ_NOIDLE));
3569
3570         cfq_update_hw_tag(cfqd);
3571
3572         WARN_ON(!cfqd->rq_in_driver);
3573         WARN_ON(!cfqq->dispatched);
3574         cfqd->rq_in_driver--;
3575         cfqq->dispatched--;
3576         (RQ_CFQG(rq))->dispatched--;
3577         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3578                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3579                         rq_data_dir(rq), rq_is_sync(rq));
3580
3581         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3582
3583         if (sync) {
3584                 struct cfq_rb_root *service_tree;
3585
3586                 RQ_CIC(rq)->ttime.last_end_request = now;
3587
3588                 if (cfq_cfqq_on_rr(cfqq))
3589                         service_tree = cfqq->service_tree;
3590                 else
3591                         service_tree = service_tree_for(cfqq->cfqg,
3592                                 cfqq_prio(cfqq), cfqq_type(cfqq));
3593                 service_tree->ttime.last_end_request = now;
3594                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3595                         cfqd->last_delayed_sync = now;
3596         }
3597
3598 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3599         cfqq->cfqg->ttime.last_end_request = now;
3600 #endif
3601
3602         /*
3603          * If this is the active queue, check if it needs to be expired,
3604          * or if we want to idle in case it has no pending requests.
3605          */
3606         if (cfqd->active_queue == cfqq) {
3607                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3608
3609                 if (cfq_cfqq_slice_new(cfqq)) {
3610                         cfq_set_prio_slice(cfqd, cfqq);
3611                         cfq_clear_cfqq_slice_new(cfqq);
3612                 }
3613
3614                 /*
3615                  * Should we wait for next request to come in before we expire
3616                  * the queue.
3617                  */
3618                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3619                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3620                         if (!cfqd->cfq_slice_idle)
3621                                 extend_sl = cfqd->cfq_group_idle;
3622                         cfqq->slice_end = jiffies + extend_sl;
3623                         cfq_mark_cfqq_wait_busy(cfqq);
3624                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3625                 }
3626
3627                 /*
3628                  * Idling is not enabled on:
3629                  * - expired queues
3630                  * - idle-priority queues
3631                  * - async queues
3632                  * - queues with still some requests queued
3633                  * - when there is a close cooperator
3634                  */
3635                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3636                         cfq_slice_expired(cfqd, 1);
3637                 else if (sync && cfqq_empty &&
3638                          !cfq_close_cooperator(cfqd, cfqq)) {
3639                         cfq_arm_slice_timer(cfqd);
3640                 }
3641         }
3642
3643         if (!cfqd->rq_in_driver)
3644                 cfq_schedule_dispatch(cfqd);
3645 }
3646
3647 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3648 {
3649         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3650                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3651                 return ELV_MQUEUE_MUST;
3652         }
3653
3654         return ELV_MQUEUE_MAY;
3655 }
3656
3657 static int cfq_may_queue(struct request_queue *q, int rw)
3658 {
3659         struct cfq_data *cfqd = q->elevator->elevator_data;
3660         struct task_struct *tsk = current;
3661         struct cfq_io_context *cic;
3662         struct cfq_queue *cfqq;
3663
3664         /*
3665          * don't force setup of a queue from here, as a call to may_queue
3666          * does not necessarily imply that a request actually will be queued.
3667          * so just lookup a possibly existing queue, or return 'may queue'
3668          * if that fails
3669          */
3670         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3671         if (!cic)
3672                 return ELV_MQUEUE_MAY;
3673
3674         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3675         if (cfqq) {
3676                 cfq_init_prio_data(cfqq, cic->ioc);
3677
3678                 return __cfq_may_queue(cfqq);
3679         }
3680
3681         return ELV_MQUEUE_MAY;
3682 }
3683
3684 /*
3685  * queue lock held here
3686  */
3687 static void cfq_put_request(struct request *rq)
3688 {
3689         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3690
3691         if (cfqq) {
3692                 const int rw = rq_data_dir(rq);
3693
3694                 BUG_ON(!cfqq->allocated[rw]);
3695                 cfqq->allocated[rw]--;
3696
3697                 put_io_context(RQ_CIC(rq)->ioc);
3698
3699                 rq->elevator_private[0] = NULL;
3700                 rq->elevator_private[1] = NULL;
3701
3702                 /* Put down rq reference on cfqg */
3703                 cfq_put_cfqg(RQ_CFQG(rq));
3704                 rq->elevator_private[2] = NULL;
3705
3706                 cfq_put_queue(cfqq);
3707         }
3708 }
3709
3710 static struct cfq_queue *
3711 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3712                 struct cfq_queue *cfqq)
3713 {
3714         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3715         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3716         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3717         cfq_put_queue(cfqq);
3718         return cic_to_cfqq(cic, 1);
3719 }
3720
3721 /*
3722  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3723  * was the last process referring to said cfqq.
3724  */
3725 static struct cfq_queue *
3726 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3727 {
3728         if (cfqq_process_refs(cfqq) == 1) {
3729                 cfqq->pid = current->pid;
3730                 cfq_clear_cfqq_coop(cfqq);
3731                 cfq_clear_cfqq_split_coop(cfqq);
3732                 return cfqq;
3733         }
3734
3735         cic_set_cfqq(cic, NULL, 1);
3736
3737         cfq_put_cooperator(cfqq);
3738
3739         cfq_put_queue(cfqq);
3740         return NULL;
3741 }
3742 /*
3743  * Allocate cfq data structures associated with this request.
3744  */
3745 static int
3746 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3747 {
3748         struct cfq_data *cfqd = q->elevator->elevator_data;
3749         struct cfq_io_context *cic;
3750         const int rw = rq_data_dir(rq);
3751         const bool is_sync = rq_is_sync(rq);
3752         struct cfq_queue *cfqq;
3753         unsigned long flags;
3754
3755         might_sleep_if(gfp_mask & __GFP_WAIT);
3756
3757         cic = cfq_get_io_context(cfqd, gfp_mask);
3758
3759         spin_lock_irqsave(q->queue_lock, flags);
3760
3761         if (!cic)
3762                 goto queue_fail;
3763
3764 new_queue:
3765         cfqq = cic_to_cfqq(cic, is_sync);
3766         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3767                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3768                 cic_set_cfqq(cic, cfqq, is_sync);
3769         } else {
3770                 /*
3771                  * If the queue was seeky for too long, break it apart.
3772                  */
3773                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3774                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3775                         cfqq = split_cfqq(cic, cfqq);
3776                         if (!cfqq)
3777                                 goto new_queue;
3778                 }
3779
3780                 /*
3781                  * Check to see if this queue is scheduled to merge with
3782                  * another, closely cooperating queue.  The merging of
3783                  * queues happens here as it must be done in process context.
3784                  * The reference on new_cfqq was taken in merge_cfqqs.
3785                  */
3786                 if (cfqq->new_cfqq)
3787                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3788         }
3789
3790         cfqq->allocated[rw]++;
3791
3792         cfqq->ref++;
3793         rq->elevator_private[0] = cic;
3794         rq->elevator_private[1] = cfqq;
3795         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3796         spin_unlock_irqrestore(q->queue_lock, flags);
3797         return 0;
3798
3799 queue_fail:
3800         cfq_schedule_dispatch(cfqd);
3801         spin_unlock_irqrestore(q->queue_lock, flags);
3802         cfq_log(cfqd, "set_request fail");
3803         return 1;
3804 }
3805
3806 static void cfq_kick_queue(struct work_struct *work)
3807 {
3808         struct cfq_data *cfqd =
3809                 container_of(work, struct cfq_data, unplug_work);
3810         struct request_queue *q = cfqd->queue;
3811
3812         spin_lock_irq(q->queue_lock);
3813         __blk_run_queue(cfqd->queue);
3814         spin_unlock_irq(q->queue_lock);
3815 }
3816
3817 /*
3818  * Timer running if the active_queue is currently idling inside its time slice
3819  */
3820 static void cfq_idle_slice_timer(unsigned long data)
3821 {
3822         struct cfq_data *cfqd = (struct cfq_data *) data;
3823         struct cfq_queue *cfqq;
3824         unsigned long flags;
3825         int timed_out = 1;
3826
3827         cfq_log(cfqd, "idle timer fired");
3828
3829         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3830
3831         cfqq = cfqd->active_queue;
3832         if (cfqq) {
3833                 timed_out = 0;
3834
3835                 /*
3836                  * We saw a request before the queue expired, let it through
3837                  */
3838                 if (cfq_cfqq_must_dispatch(cfqq))
3839                         goto out_kick;
3840
3841                 /*
3842                  * expired
3843                  */
3844                 if (cfq_slice_used(cfqq))
3845                         goto expire;
3846
3847                 /*
3848                  * only expire and reinvoke request handler, if there are
3849                  * other queues with pending requests
3850                  */
3851                 if (!cfqd->busy_queues)
3852                         goto out_cont;
3853
3854                 /*
3855                  * not expired and it has a request pending, let it dispatch
3856                  */
3857                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3858                         goto out_kick;
3859
3860                 /*
3861                  * Queue depth flag is reset only when the idle didn't succeed
3862                  */
3863                 cfq_clear_cfqq_deep(cfqq);
3864         }
3865 expire:
3866         cfq_slice_expired(cfqd, timed_out);
3867 out_kick:
3868         cfq_schedule_dispatch(cfqd);
3869 out_cont:
3870         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3871 }
3872
3873 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3874 {
3875         del_timer_sync(&cfqd->idle_slice_timer);
3876         cancel_work_sync(&cfqd->unplug_work);
3877 }
3878
3879 static void cfq_put_async_queues(struct cfq_data *cfqd)
3880 {
3881         int i;
3882
3883         for (i = 0; i < IOPRIO_BE_NR; i++) {
3884                 if (cfqd->async_cfqq[0][i])
3885                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3886                 if (cfqd->async_cfqq[1][i])
3887                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3888         }
3889
3890         if (cfqd->async_idle_cfqq)
3891                 cfq_put_queue(cfqd->async_idle_cfqq);
3892 }
3893
3894 static void cfq_exit_queue(struct elevator_queue *e)
3895 {
3896         struct cfq_data *cfqd = e->elevator_data;
3897         struct request_queue *q = cfqd->queue;
3898         bool wait = false;
3899
3900         cfq_shutdown_timer_wq(cfqd);
3901
3902         spin_lock_irq(q->queue_lock);
3903
3904         if (cfqd->active_queue)
3905                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3906
3907         while (!list_empty(&cfqd->cic_list)) {
3908                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3909                                                         struct cfq_io_context,
3910                                                         queue_list);
3911
3912                 __cfq_exit_single_io_context(cfqd, cic);
3913         }
3914
3915         cfq_put_async_queues(cfqd);
3916         cfq_release_cfq_groups(cfqd);
3917
3918         /*
3919          * If there are groups which we could not unlink from blkcg list,
3920          * wait for a rcu period for them to be freed.
3921          */
3922         if (cfqd->nr_blkcg_linked_grps)
3923                 wait = true;
3924
3925         spin_unlock_irq(q->queue_lock);
3926
3927         cfq_shutdown_timer_wq(cfqd);
3928
3929         spin_lock(&cic_index_lock);
3930         ida_remove(&cic_index_ida, cfqd->cic_index);
3931         spin_unlock(&cic_index_lock);
3932
3933         /*
3934          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3935          * Do this wait only if there are other unlinked groups out
3936          * there. This can happen if cgroup deletion path claimed the
3937          * responsibility of cleaning up a group before queue cleanup code
3938          * get to the group.
3939          *
3940          * Do not call synchronize_rcu() unconditionally as there are drivers
3941          * which create/delete request queue hundreds of times during scan/boot
3942          * and synchronize_rcu() can take significant time and slow down boot.
3943          */
3944         if (wait)
3945                 synchronize_rcu();
3946
3947 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3948         /* Free up per cpu stats for root group */
3949         free_percpu(cfqd->root_group.blkg.stats_cpu);
3950 #endif
3951         kfree(cfqd);
3952 }
3953
3954 static int cfq_alloc_cic_index(void)
3955 {
3956         int index, error;
3957
3958         do {
3959                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3960                         return -ENOMEM;
3961
3962                 spin_lock(&cic_index_lock);
3963                 error = ida_get_new(&cic_index_ida, &index);
3964                 spin_unlock(&cic_index_lock);
3965                 if (error && error != -EAGAIN)
3966                         return error;
3967         } while (error);
3968
3969         return index;
3970 }
3971
3972 static void *cfq_init_queue(struct request_queue *q)
3973 {
3974         struct cfq_data *cfqd;
3975         int i, j;
3976         struct cfq_group *cfqg;
3977         struct cfq_rb_root *st;
3978
3979         i = cfq_alloc_cic_index();
3980         if (i < 0)
3981                 return NULL;
3982
3983         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3984         if (!cfqd) {
3985                 spin_lock(&cic_index_lock);
3986                 ida_remove(&cic_index_ida, i);
3987                 spin_unlock(&cic_index_lock);
3988                 return NULL;
3989         }
3990
3991         /*
3992          * Don't need take queue_lock in the routine, since we are
3993          * initializing the ioscheduler, and nobody is using cfqd
3994          */
3995         cfqd->cic_index = i;
3996
3997         /* Init root service tree */
3998         cfqd->grp_service_tree = CFQ_RB_ROOT;
3999
4000         /* Init root group */
4001         cfqg = &cfqd->root_group;
4002         for_each_cfqg_st(cfqg, i, j, st)
4003                 *st = CFQ_RB_ROOT;
4004         RB_CLEAR_NODE(&cfqg->rb_node);
4005
4006         /* Give preference to root group over other groups */
4007         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
4008
4009 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4010         /*
4011          * Set root group reference to 2. One reference will be dropped when
4012          * all groups on cfqd->cfqg_list are being deleted during queue exit.
4013          * Other reference will remain there as we don't want to delete this
4014          * group as it is statically allocated and gets destroyed when
4015          * throtl_data goes away.
4016          */
4017         cfqg->ref = 2;
4018
4019         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4020                 kfree(cfqg);
4021                 kfree(cfqd);
4022                 return NULL;
4023         }
4024
4025         rcu_read_lock();
4026
4027         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4028                                         (void *)cfqd, 0);
4029         rcu_read_unlock();
4030         cfqd->nr_blkcg_linked_grps++;
4031
4032         /* Add group on cfqd->cfqg_list */
4033         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4034 #endif
4035         /*
4036          * Not strictly needed (since RB_ROOT just clears the node and we
4037          * zeroed cfqd on alloc), but better be safe in case someone decides
4038          * to add magic to the rb code
4039          */
4040         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4041                 cfqd->prio_trees[i] = RB_ROOT;
4042
4043         /*
4044          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4045          * Grab a permanent reference to it, so that the normal code flow
4046          * will not attempt to free it.
4047          */
4048         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4049         cfqd->oom_cfqq.ref++;
4050         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4051
4052         INIT_LIST_HEAD(&cfqd->cic_list);
4053
4054         cfqd->queue = q;
4055
4056         init_timer(&cfqd->idle_slice_timer);
4057         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4058         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4059
4060         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4061
4062         cfqd->cfq_quantum = cfq_quantum;
4063         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4064         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4065         cfqd->cfq_back_max = cfq_back_max;
4066         cfqd->cfq_back_penalty = cfq_back_penalty;
4067         cfqd->cfq_slice[0] = cfq_slice_async;
4068         cfqd->cfq_slice[1] = cfq_slice_sync;
4069         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4070         cfqd->cfq_slice_idle = cfq_slice_idle;
4071         cfqd->cfq_group_idle = cfq_group_idle;
4072         cfqd->cfq_latency = 1;
4073         cfqd->hw_tag = -1;
4074         /*
4075          * we optimistically start assuming sync ops weren't delayed in last
4076          * second, in order to have larger depth for async operations.
4077          */
4078         cfqd->last_delayed_sync = jiffies - HZ;
4079         return cfqd;
4080 }
4081
4082 static void cfq_slab_kill(void)
4083 {
4084         /*
4085          * Caller already ensured that pending RCU callbacks are completed,
4086          * so we should have no busy allocations at this point.
4087          */
4088         if (cfq_pool)
4089                 kmem_cache_destroy(cfq_pool);
4090         if (cfq_ioc_pool)
4091                 kmem_cache_destroy(cfq_ioc_pool);
4092 }
4093
4094 static int __init cfq_slab_setup(void)
4095 {
4096         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4097         if (!cfq_pool)
4098                 goto fail;
4099
4100         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4101         if (!cfq_ioc_pool)
4102                 goto fail;
4103
4104         return 0;
4105 fail:
4106         cfq_slab_kill();
4107         return -ENOMEM;
4108 }
4109
4110 /*
4111  * sysfs parts below -->
4112  */
4113 static ssize_t
4114 cfq_var_show(unsigned int var, char *page)
4115 {
4116         return sprintf(page, "%d\n", var);
4117 }
4118
4119 static ssize_t
4120 cfq_var_store(unsigned int *var, const char *page, size_t count)
4121 {
4122         char *p = (char *) page;
4123
4124         *var = simple_strtoul(p, &p, 10);
4125         return count;
4126 }
4127
4128 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4129 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4130 {                                                                       \
4131         struct cfq_data *cfqd = e->elevator_data;                       \
4132         unsigned int __data = __VAR;                                    \
4133         if (__CONV)                                                     \
4134                 __data = jiffies_to_msecs(__data);                      \
4135         return cfq_var_show(__data, (page));                            \
4136 }
4137 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4138 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4139 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4140 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4141 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4142 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4143 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4144 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4145 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4146 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4147 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4148 #undef SHOW_FUNCTION
4149
4150 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4151 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4152 {                                                                       \
4153         struct cfq_data *cfqd = e->elevator_data;                       \
4154         unsigned int __data;                                            \
4155         int ret = cfq_var_store(&__data, (page), count);                \
4156         if (__data < (MIN))                                             \
4157                 __data = (MIN);                                         \
4158         else if (__data > (MAX))                                        \
4159                 __data = (MAX);                                         \
4160         if (__CONV)                                                     \
4161                 *(__PTR) = msecs_to_jiffies(__data);                    \
4162         else                                                            \
4163                 *(__PTR) = __data;                                      \
4164         return ret;                                                     \
4165 }
4166 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4167 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4168                 UINT_MAX, 1);
4169 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4170                 UINT_MAX, 1);
4171 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4172 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4173                 UINT_MAX, 0);
4174 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4175 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4176 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4177 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4178 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4179                 UINT_MAX, 0);
4180 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4181 #undef STORE_FUNCTION
4182
4183 #define CFQ_ATTR(name) \
4184         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4185
4186 static struct elv_fs_entry cfq_attrs[] = {
4187         CFQ_ATTR(quantum),
4188         CFQ_ATTR(fifo_expire_sync),
4189         CFQ_ATTR(fifo_expire_async),
4190         CFQ_ATTR(back_seek_max),
4191         CFQ_ATTR(back_seek_penalty),
4192         CFQ_ATTR(slice_sync),
4193         CFQ_ATTR(slice_async),
4194         CFQ_ATTR(slice_async_rq),
4195         CFQ_ATTR(slice_idle),
4196         CFQ_ATTR(group_idle),
4197         CFQ_ATTR(low_latency),
4198         __ATTR_NULL
4199 };
4200
4201 static struct elevator_type iosched_cfq = {
4202         .ops = {
4203                 .elevator_merge_fn =            cfq_merge,
4204                 .elevator_merged_fn =           cfq_merged_request,
4205                 .elevator_merge_req_fn =        cfq_merged_requests,
4206                 .elevator_allow_merge_fn =      cfq_allow_merge,
4207                 .elevator_bio_merged_fn =       cfq_bio_merged,
4208                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4209                 .elevator_add_req_fn =          cfq_insert_request,
4210                 .elevator_activate_req_fn =     cfq_activate_request,
4211                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4212                 .elevator_completed_req_fn =    cfq_completed_request,
4213                 .elevator_former_req_fn =       elv_rb_former_request,
4214                 .elevator_latter_req_fn =       elv_rb_latter_request,
4215                 .elevator_set_req_fn =          cfq_set_request,
4216                 .elevator_put_req_fn =          cfq_put_request,
4217                 .elevator_may_queue_fn =        cfq_may_queue,
4218                 .elevator_init_fn =             cfq_init_queue,
4219                 .elevator_exit_fn =             cfq_exit_queue,
4220                 .trim =                         cfq_free_io_context,
4221         },
4222         .elevator_attrs =       cfq_attrs,
4223         .elevator_name =        "cfq",
4224         .elevator_owner =       THIS_MODULE,
4225 };
4226
4227 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4228 static struct blkio_policy_type blkio_policy_cfq = {
4229         .ops = {
4230                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4231                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4232         },
4233         .plid = BLKIO_POLICY_PROP,
4234 };
4235 #else
4236 static struct blkio_policy_type blkio_policy_cfq;
4237 #endif
4238
4239 static int __init cfq_init(void)
4240 {
4241         /*
4242          * could be 0 on HZ < 1000 setups
4243          */
4244         if (!cfq_slice_async)
4245                 cfq_slice_async = 1;
4246         if (!cfq_slice_idle)
4247                 cfq_slice_idle = 1;
4248
4249 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4250         if (!cfq_group_idle)
4251                 cfq_group_idle = 1;
4252 #else
4253                 cfq_group_idle = 0;
4254 #endif
4255         if (cfq_slab_setup())
4256                 return -ENOMEM;
4257
4258         elv_register(&iosched_cfq);
4259         blkio_policy_register(&blkio_policy_cfq);
4260
4261         return 0;
4262 }
4263
4264 static void __exit cfq_exit(void)
4265 {
4266         DECLARE_COMPLETION_ONSTACK(all_gone);
4267         blkio_policy_unregister(&blkio_policy_cfq);
4268         elv_unregister(&iosched_cfq);
4269         ioc_gone = &all_gone;
4270         /* ioc_gone's update must be visible before reading ioc_count */
4271         smp_wmb();
4272
4273         /*
4274          * this also protects us from entering cfq_slab_kill() with
4275          * pending RCU callbacks
4276          */
4277         if (elv_ioc_count_read(cfq_ioc_count))
4278                 wait_for_completion(&all_gone);
4279         ida_destroy(&cic_index_ida);
4280         cfq_slab_kill();
4281 }
4282
4283 module_init(cfq_init);
4284 module_exit(cfq_exit);
4285
4286 MODULE_AUTHOR("Jens Axboe");
4287 MODULE_LICENSE("GPL");
4288 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");