cfq-iosched: Fix a gcc 4.5 warning and put some comments
[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 const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
35
36 /*
37  * offset from end of service tree
38  */
39 #define CFQ_IDLE_DELAY          (HZ / 5)
40
41 /*
42  * below this threshold, we consider thinktime immediate
43  */
44 #define CFQ_MIN_TT              (2)
45
46 #define CFQ_SLICE_SCALE         (5)
47 #define CFQ_HW_QUEUE_MIN        (5)
48 #define CFQ_SERVICE_SHIFT       12
49
50 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
54
55 #define RQ_CIC(rq)              \
56         ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
58 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
59
60 static struct kmem_cache *cfq_pool;
61 static struct kmem_cache *cfq_ioc_pool;
62
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
64 static struct completion *ioc_gone;
65 static DEFINE_SPINLOCK(ioc_gone_lock);
66
67 static DEFINE_SPINLOCK(cic_index_lock);
68 static DEFINE_IDA(cic_index_ida);
69
70 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
71 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
72 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
73
74 #define sample_valid(samples)   ((samples) > 80)
75 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
76
77 /*
78  * Most of our rbtree usage is for sorting with min extraction, so
79  * if we cache the leftmost node we don't have to walk down the tree
80  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
81  * move this into the elevator for the rq sorting as well.
82  */
83 struct cfq_rb_root {
84         struct rb_root rb;
85         struct rb_node *left;
86         unsigned count;
87         unsigned total_weight;
88         u64 min_vdisktime;
89         struct rb_node *active;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         atomic_t ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         struct cfq_group *orig_cfqg;
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         bool on_st;
181
182         /* number of cfqq currently on this group */
183         int nr_cfqq;
184
185         /*
186          * Per group busy queus average. Useful for workload slice calc. We
187          * create the array for each prio class but at run time it is used
188          * only for RT and BE class and slot for IDLE class remains unused.
189          * This is primarily done to avoid confusion and a gcc warning.
190          */
191         unsigned int busy_queues_avg[CFQ_PRIO_NR];
192         /*
193          * rr lists of queues with requests. We maintain service trees for
194          * RT and BE classes. These trees are subdivided in subclasses
195          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
196          * class there is no subclassification and all the cfq queues go on
197          * a single tree service_tree_idle.
198          * Counts are embedded in the cfq_rb_root
199          */
200         struct cfq_rb_root service_trees[2][3];
201         struct cfq_rb_root service_tree_idle;
202
203         unsigned long saved_workload_slice;
204         enum wl_type_t saved_workload;
205         enum wl_prio_t saved_serving_prio;
206         struct blkio_group blkg;
207 #ifdef CONFIG_CFQ_GROUP_IOSCHED
208         struct hlist_node cfqd_node;
209         atomic_t ref;
210 #endif
211 };
212
213 /*
214  * Per block device queue structure
215  */
216 struct cfq_data {
217         struct request_queue *queue;
218         /* Root service tree for cfq_groups */
219         struct cfq_rb_root grp_service_tree;
220         struct cfq_group root_group;
221
222         /*
223          * The priority currently being served
224          */
225         enum wl_prio_t serving_prio;
226         enum wl_type_t serving_type;
227         unsigned long workload_expires;
228         struct cfq_group *serving_group;
229
230         /*
231          * Each priority tree is sorted by next_request position.  These
232          * trees are used when determining if two or more queues are
233          * interleaving requests (see cfq_close_cooperator).
234          */
235         struct rb_root prio_trees[CFQ_PRIO_LISTS];
236
237         unsigned int busy_queues;
238
239         int rq_in_driver;
240         int rq_in_flight[2];
241
242         /*
243          * queue-depth detection
244          */
245         int rq_queued;
246         int hw_tag;
247         /*
248          * hw_tag can be
249          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
250          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
251          *  0 => no NCQ
252          */
253         int hw_tag_est_depth;
254         unsigned int hw_tag_samples;
255
256         /*
257          * idle window management
258          */
259         struct timer_list idle_slice_timer;
260         struct work_struct unplug_work;
261
262         struct cfq_queue *active_queue;
263         struct cfq_io_context *active_cic;
264
265         /*
266          * async queue for each priority case
267          */
268         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
269         struct cfq_queue *async_idle_cfqq;
270
271         sector_t last_position;
272
273         /*
274          * tunables, see top of file
275          */
276         unsigned int cfq_quantum;
277         unsigned int cfq_fifo_expire[2];
278         unsigned int cfq_back_penalty;
279         unsigned int cfq_back_max;
280         unsigned int cfq_slice[2];
281         unsigned int cfq_slice_async_rq;
282         unsigned int cfq_slice_idle;
283         unsigned int cfq_latency;
284         unsigned int cfq_group_isolation;
285
286         unsigned int cic_index;
287         struct list_head cic_list;
288
289         /*
290          * Fallback dummy cfqq for extreme OOM conditions
291          */
292         struct cfq_queue oom_cfqq;
293
294         unsigned long last_delayed_sync;
295
296         /* List of cfq groups being managed on this device*/
297         struct hlist_head cfqg_list;
298         struct rcu_head rcu;
299 };
300
301 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
302
303 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
304                                             enum wl_prio_t prio,
305                                             enum wl_type_t type)
306 {
307         if (!cfqg)
308                 return NULL;
309
310         if (prio == IDLE_WORKLOAD)
311                 return &cfqg->service_tree_idle;
312
313         return &cfqg->service_trees[prio][type];
314 }
315
316 enum cfqq_state_flags {
317         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
318         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
319         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
320         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
321         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
322         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
323         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
324         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
325         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
326         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
327         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
328         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
329         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
330 };
331
332 #define CFQ_CFQQ_FNS(name)                                              \
333 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
334 {                                                                       \
335         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
336 }                                                                       \
337 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
338 {                                                                       \
339         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
340 }                                                                       \
341 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
342 {                                                                       \
343         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
344 }
345
346 CFQ_CFQQ_FNS(on_rr);
347 CFQ_CFQQ_FNS(wait_request);
348 CFQ_CFQQ_FNS(must_dispatch);
349 CFQ_CFQQ_FNS(must_alloc_slice);
350 CFQ_CFQQ_FNS(fifo_expire);
351 CFQ_CFQQ_FNS(idle_window);
352 CFQ_CFQQ_FNS(prio_changed);
353 CFQ_CFQQ_FNS(slice_new);
354 CFQ_CFQQ_FNS(sync);
355 CFQ_CFQQ_FNS(coop);
356 CFQ_CFQQ_FNS(split_coop);
357 CFQ_CFQQ_FNS(deep);
358 CFQ_CFQQ_FNS(wait_busy);
359 #undef CFQ_CFQQ_FNS
360
361 #ifdef CONFIG_CFQ_GROUP_IOSCHED
362 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
363         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
364                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
365                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
366
367 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
368         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
369                                 blkg_path(&(cfqg)->blkg), ##args);      \
370
371 #else
372 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
373         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
375 #endif
376 #define cfq_log(cfqd, fmt, args...)     \
377         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
378
379 /* Traverses through cfq group service trees */
380 #define for_each_cfqg_st(cfqg, i, j, st) \
381         for (i = 0; i <= IDLE_WORKLOAD; i++) \
382                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
383                         : &cfqg->service_tree_idle; \
384                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
385                         (i == IDLE_WORKLOAD && j == 0); \
386                         j++, st = i < IDLE_WORKLOAD ? \
387                         &cfqg->service_trees[i][j]: NULL) \
388
389
390 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
391 {
392         if (cfq_class_idle(cfqq))
393                 return IDLE_WORKLOAD;
394         if (cfq_class_rt(cfqq))
395                 return RT_WORKLOAD;
396         return BE_WORKLOAD;
397 }
398
399
400 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
401 {
402         if (!cfq_cfqq_sync(cfqq))
403                 return ASYNC_WORKLOAD;
404         if (!cfq_cfqq_idle_window(cfqq))
405                 return SYNC_NOIDLE_WORKLOAD;
406         return SYNC_WORKLOAD;
407 }
408
409 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
410                                         struct cfq_data *cfqd,
411                                         struct cfq_group *cfqg)
412 {
413         if (wl == IDLE_WORKLOAD)
414                 return cfqg->service_tree_idle.count;
415
416         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
417                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
418                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
419 }
420
421 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
422                                         struct cfq_group *cfqg)
423 {
424         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
425                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
426 }
427
428 static void cfq_dispatch_insert(struct request_queue *, struct request *);
429 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
430                                        struct io_context *, gfp_t);
431 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
432                                                 struct io_context *);
433
434 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
435                                             bool is_sync)
436 {
437         return cic->cfqq[is_sync];
438 }
439
440 static inline void cic_set_cfqq(struct cfq_io_context *cic,
441                                 struct cfq_queue *cfqq, bool is_sync)
442 {
443         cic->cfqq[is_sync] = cfqq;
444 }
445
446 #define CIC_DEAD_KEY    1ul
447 #define CIC_DEAD_INDEX_SHIFT    1
448
449 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
450 {
451         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
452 }
453
454 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
455 {
456         struct cfq_data *cfqd = cic->key;
457
458         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
459                 return NULL;
460
461         return cfqd;
462 }
463
464 /*
465  * We regard a request as SYNC, if it's either a read or has the SYNC bit
466  * set (in which case it could also be direct WRITE).
467  */
468 static inline bool cfq_bio_sync(struct bio *bio)
469 {
470         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
471 }
472
473 /*
474  * scheduler run of queue, if there are requests pending and no one in the
475  * driver that will restart queueing
476  */
477 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
478 {
479         if (cfqd->busy_queues) {
480                 cfq_log(cfqd, "schedule dispatch");
481                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
482         }
483 }
484
485 static int cfq_queue_empty(struct request_queue *q)
486 {
487         struct cfq_data *cfqd = q->elevator->elevator_data;
488
489         return !cfqd->rq_queued;
490 }
491
492 /*
493  * Scale schedule slice based on io priority. Use the sync time slice only
494  * if a queue is marked sync and has sync io queued. A sync queue with async
495  * io only, should not get full sync slice length.
496  */
497 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
498                                  unsigned short prio)
499 {
500         const int base_slice = cfqd->cfq_slice[sync];
501
502         WARN_ON(prio >= IOPRIO_BE_NR);
503
504         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
505 }
506
507 static inline int
508 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
509 {
510         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
511 }
512
513 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
514 {
515         u64 d = delta << CFQ_SERVICE_SHIFT;
516
517         d = d * BLKIO_WEIGHT_DEFAULT;
518         do_div(d, cfqg->weight);
519         return d;
520 }
521
522 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
523 {
524         s64 delta = (s64)(vdisktime - min_vdisktime);
525         if (delta > 0)
526                 min_vdisktime = vdisktime;
527
528         return min_vdisktime;
529 }
530
531 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
532 {
533         s64 delta = (s64)(vdisktime - min_vdisktime);
534         if (delta < 0)
535                 min_vdisktime = vdisktime;
536
537         return min_vdisktime;
538 }
539
540 static void update_min_vdisktime(struct cfq_rb_root *st)
541 {
542         u64 vdisktime = st->min_vdisktime;
543         struct cfq_group *cfqg;
544
545         if (st->active) {
546                 cfqg = rb_entry_cfqg(st->active);
547                 vdisktime = cfqg->vdisktime;
548         }
549
550         if (st->left) {
551                 cfqg = rb_entry_cfqg(st->left);
552                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
553         }
554
555         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
556 }
557
558 /*
559  * get averaged number of queues of RT/BE priority.
560  * average is updated, with a formula that gives more weight to higher numbers,
561  * to quickly follows sudden increases and decrease slowly
562  */
563
564 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
565                                         struct cfq_group *cfqg, bool rt)
566 {
567         unsigned min_q, max_q;
568         unsigned mult  = cfq_hist_divisor - 1;
569         unsigned round = cfq_hist_divisor / 2;
570         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
571
572         min_q = min(cfqg->busy_queues_avg[rt], busy);
573         max_q = max(cfqg->busy_queues_avg[rt], busy);
574         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
575                 cfq_hist_divisor;
576         return cfqg->busy_queues_avg[rt];
577 }
578
579 static inline unsigned
580 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
581 {
582         struct cfq_rb_root *st = &cfqd->grp_service_tree;
583
584         return cfq_target_latency * cfqg->weight / st->total_weight;
585 }
586
587 static inline void
588 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
589 {
590         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
591         if (cfqd->cfq_latency) {
592                 /*
593                  * interested queues (we consider only the ones with the same
594                  * priority class in the cfq group)
595                  */
596                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
597                                                 cfq_class_rt(cfqq));
598                 unsigned sync_slice = cfqd->cfq_slice[1];
599                 unsigned expect_latency = sync_slice * iq;
600                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
601
602                 if (expect_latency > group_slice) {
603                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
604                         /* scale low_slice according to IO priority
605                          * and sync vs async */
606                         unsigned low_slice =
607                                 min(slice, base_low_slice * slice / sync_slice);
608                         /* the adapted slice value is scaled to fit all iqs
609                          * into the target latency */
610                         slice = max(slice * group_slice / expect_latency,
611                                     low_slice);
612                 }
613         }
614         cfqq->slice_start = jiffies;
615         cfqq->slice_end = jiffies + slice;
616         cfqq->allocated_slice = slice;
617         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
618 }
619
620 /*
621  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
622  * isn't valid until the first request from the dispatch is activated
623  * and the slice time set.
624  */
625 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
626 {
627         if (cfq_cfqq_slice_new(cfqq))
628                 return 0;
629         if (time_before(jiffies, cfqq->slice_end))
630                 return 0;
631
632         return 1;
633 }
634
635 /*
636  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
637  * We choose the request that is closest to the head right now. Distance
638  * behind the head is penalized and only allowed to a certain extent.
639  */
640 static struct request *
641 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
642 {
643         sector_t s1, s2, d1 = 0, d2 = 0;
644         unsigned long back_max;
645 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
646 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
647         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
648
649         if (rq1 == NULL || rq1 == rq2)
650                 return rq2;
651         if (rq2 == NULL)
652                 return rq1;
653
654         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
655                 return rq1;
656         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
657                 return rq2;
658         if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
659                 return rq1;
660         else if ((rq2->cmd_flags & REQ_META) &&
661                  !(rq1->cmd_flags & REQ_META))
662                 return rq2;
663
664         s1 = blk_rq_pos(rq1);
665         s2 = blk_rq_pos(rq2);
666
667         /*
668          * by definition, 1KiB is 2 sectors
669          */
670         back_max = cfqd->cfq_back_max * 2;
671
672         /*
673          * Strict one way elevator _except_ in the case where we allow
674          * short backward seeks which are biased as twice the cost of a
675          * similar forward seek.
676          */
677         if (s1 >= last)
678                 d1 = s1 - last;
679         else if (s1 + back_max >= last)
680                 d1 = (last - s1) * cfqd->cfq_back_penalty;
681         else
682                 wrap |= CFQ_RQ1_WRAP;
683
684         if (s2 >= last)
685                 d2 = s2 - last;
686         else if (s2 + back_max >= last)
687                 d2 = (last - s2) * cfqd->cfq_back_penalty;
688         else
689                 wrap |= CFQ_RQ2_WRAP;
690
691         /* Found required data */
692
693         /*
694          * By doing switch() on the bit mask "wrap" we avoid having to
695          * check two variables for all permutations: --> faster!
696          */
697         switch (wrap) {
698         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
699                 if (d1 < d2)
700                         return rq1;
701                 else if (d2 < d1)
702                         return rq2;
703                 else {
704                         if (s1 >= s2)
705                                 return rq1;
706                         else
707                                 return rq2;
708                 }
709
710         case CFQ_RQ2_WRAP:
711                 return rq1;
712         case CFQ_RQ1_WRAP:
713                 return rq2;
714         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
715         default:
716                 /*
717                  * Since both rqs are wrapped,
718                  * start with the one that's further behind head
719                  * (--> only *one* back seek required),
720                  * since back seek takes more time than forward.
721                  */
722                 if (s1 <= s2)
723                         return rq1;
724                 else
725                         return rq2;
726         }
727 }
728
729 /*
730  * The below is leftmost cache rbtree addon
731  */
732 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
733 {
734         /* Service tree is empty */
735         if (!root->count)
736                 return NULL;
737
738         if (!root->left)
739                 root->left = rb_first(&root->rb);
740
741         if (root->left)
742                 return rb_entry(root->left, struct cfq_queue, rb_node);
743
744         return NULL;
745 }
746
747 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
748 {
749         if (!root->left)
750                 root->left = rb_first(&root->rb);
751
752         if (root->left)
753                 return rb_entry_cfqg(root->left);
754
755         return NULL;
756 }
757
758 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
759 {
760         rb_erase(n, root);
761         RB_CLEAR_NODE(n);
762 }
763
764 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
765 {
766         if (root->left == n)
767                 root->left = NULL;
768         rb_erase_init(n, &root->rb);
769         --root->count;
770 }
771
772 /*
773  * would be nice to take fifo expire time into account as well
774  */
775 static struct request *
776 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
777                   struct request *last)
778 {
779         struct rb_node *rbnext = rb_next(&last->rb_node);
780         struct rb_node *rbprev = rb_prev(&last->rb_node);
781         struct request *next = NULL, *prev = NULL;
782
783         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
784
785         if (rbprev)
786                 prev = rb_entry_rq(rbprev);
787
788         if (rbnext)
789                 next = rb_entry_rq(rbnext);
790         else {
791                 rbnext = rb_first(&cfqq->sort_list);
792                 if (rbnext && rbnext != &last->rb_node)
793                         next = rb_entry_rq(rbnext);
794         }
795
796         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
797 }
798
799 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
800                                       struct cfq_queue *cfqq)
801 {
802         /*
803          * just an approximation, should be ok.
804          */
805         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
806                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
807 }
808
809 static inline s64
810 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
811 {
812         return cfqg->vdisktime - st->min_vdisktime;
813 }
814
815 static void
816 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
817 {
818         struct rb_node **node = &st->rb.rb_node;
819         struct rb_node *parent = NULL;
820         struct cfq_group *__cfqg;
821         s64 key = cfqg_key(st, cfqg);
822         int left = 1;
823
824         while (*node != NULL) {
825                 parent = *node;
826                 __cfqg = rb_entry_cfqg(parent);
827
828                 if (key < cfqg_key(st, __cfqg))
829                         node = &parent->rb_left;
830                 else {
831                         node = &parent->rb_right;
832                         left = 0;
833                 }
834         }
835
836         if (left)
837                 st->left = &cfqg->rb_node;
838
839         rb_link_node(&cfqg->rb_node, parent, node);
840         rb_insert_color(&cfqg->rb_node, &st->rb);
841 }
842
843 static void
844 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
845 {
846         struct cfq_rb_root *st = &cfqd->grp_service_tree;
847         struct cfq_group *__cfqg;
848         struct rb_node *n;
849
850         cfqg->nr_cfqq++;
851         if (cfqg->on_st)
852                 return;
853
854         /*
855          * Currently put the group at the end. Later implement something
856          * so that groups get lesser vtime based on their weights, so that
857          * if group does not loose all if it was not continously backlogged.
858          */
859         n = rb_last(&st->rb);
860         if (n) {
861                 __cfqg = rb_entry_cfqg(n);
862                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
863         } else
864                 cfqg->vdisktime = st->min_vdisktime;
865
866         __cfq_group_service_tree_add(st, cfqg);
867         cfqg->on_st = true;
868         st->total_weight += cfqg->weight;
869 }
870
871 static void
872 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
873 {
874         struct cfq_rb_root *st = &cfqd->grp_service_tree;
875
876         if (st->active == &cfqg->rb_node)
877                 st->active = NULL;
878
879         BUG_ON(cfqg->nr_cfqq < 1);
880         cfqg->nr_cfqq--;
881
882         /* If there are other cfq queues under this group, don't delete it */
883         if (cfqg->nr_cfqq)
884                 return;
885
886         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
887         cfqg->on_st = false;
888         st->total_weight -= cfqg->weight;
889         if (!RB_EMPTY_NODE(&cfqg->rb_node))
890                 cfq_rb_erase(&cfqg->rb_node, st);
891         cfqg->saved_workload_slice = 0;
892         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
893 }
894
895 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
896 {
897         unsigned int slice_used;
898
899         /*
900          * Queue got expired before even a single request completed or
901          * got expired immediately after first request completion.
902          */
903         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
904                 /*
905                  * Also charge the seek time incurred to the group, otherwise
906                  * if there are mutiple queues in the group, each can dispatch
907                  * a single request on seeky media and cause lots of seek time
908                  * and group will never know it.
909                  */
910                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
911                                         1);
912         } else {
913                 slice_used = jiffies - cfqq->slice_start;
914                 if (slice_used > cfqq->allocated_slice)
915                         slice_used = cfqq->allocated_slice;
916         }
917
918         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
919         return slice_used;
920 }
921
922 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
923                                 struct cfq_queue *cfqq)
924 {
925         struct cfq_rb_root *st = &cfqd->grp_service_tree;
926         unsigned int used_sl, charge_sl;
927         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
928                         - cfqg->service_tree_idle.count;
929
930         BUG_ON(nr_sync < 0);
931         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
932
933         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
934                 charge_sl = cfqq->allocated_slice;
935
936         /* Can't update vdisktime while group is on service tree */
937         cfq_rb_erase(&cfqg->rb_node, st);
938         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
939         __cfq_group_service_tree_add(st, cfqg);
940
941         /* This group is being expired. Save the context */
942         if (time_after(cfqd->workload_expires, jiffies)) {
943                 cfqg->saved_workload_slice = cfqd->workload_expires
944                                                 - jiffies;
945                 cfqg->saved_workload = cfqd->serving_type;
946                 cfqg->saved_serving_prio = cfqd->serving_prio;
947         } else
948                 cfqg->saved_workload_slice = 0;
949
950         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
951                                         st->min_vdisktime);
952         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
953         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
954 }
955
956 #ifdef CONFIG_CFQ_GROUP_IOSCHED
957 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
958 {
959         if (blkg)
960                 return container_of(blkg, struct cfq_group, blkg);
961         return NULL;
962 }
963
964 void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
965                                         unsigned int weight)
966 {
967         cfqg_of_blkg(blkg)->weight = weight;
968 }
969
970 static struct cfq_group *
971 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
972 {
973         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
974         struct cfq_group *cfqg = NULL;
975         void *key = cfqd;
976         int i, j;
977         struct cfq_rb_root *st;
978         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
979         unsigned int major, minor;
980
981         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
982         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
983                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
984                 cfqg->blkg.dev = MKDEV(major, minor);
985                 goto done;
986         }
987         if (cfqg || !create)
988                 goto done;
989
990         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
991         if (!cfqg)
992                 goto done;
993
994         for_each_cfqg_st(cfqg, i, j, st)
995                 *st = CFQ_RB_ROOT;
996         RB_CLEAR_NODE(&cfqg->rb_node);
997
998         /*
999          * Take the initial reference that will be released on destroy
1000          * This can be thought of a joint reference by cgroup and
1001          * elevator which will be dropped by either elevator exit
1002          * or cgroup deletion path depending on who is exiting first.
1003          */
1004         atomic_set(&cfqg->ref, 1);
1005
1006         /* Add group onto cgroup list */
1007         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1008         cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
1009                                         MKDEV(major, minor));
1010         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1011
1012         /* Add group on cfqd list */
1013         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1014
1015 done:
1016         return cfqg;
1017 }
1018
1019 /*
1020  * Search for the cfq group current task belongs to. If create = 1, then also
1021  * create the cfq group if it does not exist. request_queue lock must be held.
1022  */
1023 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1024 {
1025         struct cgroup *cgroup;
1026         struct cfq_group *cfqg = NULL;
1027
1028         rcu_read_lock();
1029         cgroup = task_cgroup(current, blkio_subsys_id);
1030         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1031         if (!cfqg && create)
1032                 cfqg = &cfqd->root_group;
1033         rcu_read_unlock();
1034         return cfqg;
1035 }
1036
1037 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1038 {
1039         atomic_inc(&cfqg->ref);
1040         return cfqg;
1041 }
1042
1043 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1044 {
1045         /* Currently, all async queues are mapped to root group */
1046         if (!cfq_cfqq_sync(cfqq))
1047                 cfqg = &cfqq->cfqd->root_group;
1048
1049         cfqq->cfqg = cfqg;
1050         /* cfqq reference on cfqg */
1051         atomic_inc(&cfqq->cfqg->ref);
1052 }
1053
1054 static void cfq_put_cfqg(struct cfq_group *cfqg)
1055 {
1056         struct cfq_rb_root *st;
1057         int i, j;
1058
1059         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1060         if (!atomic_dec_and_test(&cfqg->ref))
1061                 return;
1062         for_each_cfqg_st(cfqg, i, j, st)
1063                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1064         kfree(cfqg);
1065 }
1066
1067 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1068 {
1069         /* Something wrong if we are trying to remove same group twice */
1070         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1071
1072         hlist_del_init(&cfqg->cfqd_node);
1073
1074         /*
1075          * Put the reference taken at the time of creation so that when all
1076          * queues are gone, group can be destroyed.
1077          */
1078         cfq_put_cfqg(cfqg);
1079 }
1080
1081 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1082 {
1083         struct hlist_node *pos, *n;
1084         struct cfq_group *cfqg;
1085
1086         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1087                 /*
1088                  * If cgroup removal path got to blk_group first and removed
1089                  * it from cgroup list, then it will take care of destroying
1090                  * cfqg also.
1091                  */
1092                 if (!cfq_blkiocg_del_blkio_group(&cfqg->blkg))
1093                         cfq_destroy_cfqg(cfqd, cfqg);
1094         }
1095 }
1096
1097 /*
1098  * Blk cgroup controller notification saying that blkio_group object is being
1099  * delinked as associated cgroup object is going away. That also means that
1100  * no new IO will come in this group. So get rid of this group as soon as
1101  * any pending IO in the group is finished.
1102  *
1103  * This function is called under rcu_read_lock(). key is the rcu protected
1104  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1105  * read lock.
1106  *
1107  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1108  * it should not be NULL as even if elevator was exiting, cgroup deltion
1109  * path got to it first.
1110  */
1111 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1112 {
1113         unsigned long  flags;
1114         struct cfq_data *cfqd = key;
1115
1116         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1117         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1118         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1119 }
1120
1121 #else /* GROUP_IOSCHED */
1122 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1123 {
1124         return &cfqd->root_group;
1125 }
1126
1127 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1128 {
1129         return cfqg;
1130 }
1131
1132 static inline void
1133 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1134         cfqq->cfqg = cfqg;
1135 }
1136
1137 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1138 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1139
1140 #endif /* GROUP_IOSCHED */
1141
1142 /*
1143  * The cfqd->service_trees holds all pending cfq_queue's that have
1144  * requests waiting to be processed. It is sorted in the order that
1145  * we will service the queues.
1146  */
1147 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1148                                  bool add_front)
1149 {
1150         struct rb_node **p, *parent;
1151         struct cfq_queue *__cfqq;
1152         unsigned long rb_key;
1153         struct cfq_rb_root *service_tree;
1154         int left;
1155         int new_cfqq = 1;
1156         int group_changed = 0;
1157
1158 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1159         if (!cfqd->cfq_group_isolation
1160             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1161             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1162                 /* Move this cfq to root group */
1163                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1164                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1165                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1166                 cfqq->orig_cfqg = cfqq->cfqg;
1167                 cfqq->cfqg = &cfqd->root_group;
1168                 atomic_inc(&cfqd->root_group.ref);
1169                 group_changed = 1;
1170         } else if (!cfqd->cfq_group_isolation
1171                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1172                 /* cfqq is sequential now needs to go to its original group */
1173                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1174                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1175                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1176                 cfq_put_cfqg(cfqq->cfqg);
1177                 cfqq->cfqg = cfqq->orig_cfqg;
1178                 cfqq->orig_cfqg = NULL;
1179                 group_changed = 1;
1180                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1181         }
1182 #endif
1183
1184         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1185                                                 cfqq_type(cfqq));
1186         if (cfq_class_idle(cfqq)) {
1187                 rb_key = CFQ_IDLE_DELAY;
1188                 parent = rb_last(&service_tree->rb);
1189                 if (parent && parent != &cfqq->rb_node) {
1190                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1191                         rb_key += __cfqq->rb_key;
1192                 } else
1193                         rb_key += jiffies;
1194         } else if (!add_front) {
1195                 /*
1196                  * Get our rb key offset. Subtract any residual slice
1197                  * value carried from last service. A negative resid
1198                  * count indicates slice overrun, and this should position
1199                  * the next service time further away in the tree.
1200                  */
1201                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1202                 rb_key -= cfqq->slice_resid;
1203                 cfqq->slice_resid = 0;
1204         } else {
1205                 rb_key = -HZ;
1206                 __cfqq = cfq_rb_first(service_tree);
1207                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1208         }
1209
1210         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1211                 new_cfqq = 0;
1212                 /*
1213                  * same position, nothing more to do
1214                  */
1215                 if (rb_key == cfqq->rb_key &&
1216                     cfqq->service_tree == service_tree)
1217                         return;
1218
1219                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1220                 cfqq->service_tree = NULL;
1221         }
1222
1223         left = 1;
1224         parent = NULL;
1225         cfqq->service_tree = service_tree;
1226         p = &service_tree->rb.rb_node;
1227         while (*p) {
1228                 struct rb_node **n;
1229
1230                 parent = *p;
1231                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1232
1233                 /*
1234                  * sort by key, that represents service time.
1235                  */
1236                 if (time_before(rb_key, __cfqq->rb_key))
1237                         n = &(*p)->rb_left;
1238                 else {
1239                         n = &(*p)->rb_right;
1240                         left = 0;
1241                 }
1242
1243                 p = n;
1244         }
1245
1246         if (left)
1247                 service_tree->left = &cfqq->rb_node;
1248
1249         cfqq->rb_key = rb_key;
1250         rb_link_node(&cfqq->rb_node, parent, p);
1251         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1252         service_tree->count++;
1253         if ((add_front || !new_cfqq) && !group_changed)
1254                 return;
1255         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1256 }
1257
1258 static struct cfq_queue *
1259 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1260                      sector_t sector, struct rb_node **ret_parent,
1261                      struct rb_node ***rb_link)
1262 {
1263         struct rb_node **p, *parent;
1264         struct cfq_queue *cfqq = NULL;
1265
1266         parent = NULL;
1267         p = &root->rb_node;
1268         while (*p) {
1269                 struct rb_node **n;
1270
1271                 parent = *p;
1272                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1273
1274                 /*
1275                  * Sort strictly based on sector.  Smallest to the left,
1276                  * largest to the right.
1277                  */
1278                 if (sector > blk_rq_pos(cfqq->next_rq))
1279                         n = &(*p)->rb_right;
1280                 else if (sector < blk_rq_pos(cfqq->next_rq))
1281                         n = &(*p)->rb_left;
1282                 else
1283                         break;
1284                 p = n;
1285                 cfqq = NULL;
1286         }
1287
1288         *ret_parent = parent;
1289         if (rb_link)
1290                 *rb_link = p;
1291         return cfqq;
1292 }
1293
1294 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1295 {
1296         struct rb_node **p, *parent;
1297         struct cfq_queue *__cfqq;
1298
1299         if (cfqq->p_root) {
1300                 rb_erase(&cfqq->p_node, cfqq->p_root);
1301                 cfqq->p_root = NULL;
1302         }
1303
1304         if (cfq_class_idle(cfqq))
1305                 return;
1306         if (!cfqq->next_rq)
1307                 return;
1308
1309         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1310         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1311                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1312         if (!__cfqq) {
1313                 rb_link_node(&cfqq->p_node, parent, p);
1314                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1315         } else
1316                 cfqq->p_root = NULL;
1317 }
1318
1319 /*
1320  * Update cfqq's position in the service tree.
1321  */
1322 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1323 {
1324         /*
1325          * Resorting requires the cfqq to be on the RR list already.
1326          */
1327         if (cfq_cfqq_on_rr(cfqq)) {
1328                 cfq_service_tree_add(cfqd, cfqq, 0);
1329                 cfq_prio_tree_add(cfqd, cfqq);
1330         }
1331 }
1332
1333 /*
1334  * add to busy list of queues for service, trying to be fair in ordering
1335  * the pending list according to last request service
1336  */
1337 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1338 {
1339         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1340         BUG_ON(cfq_cfqq_on_rr(cfqq));
1341         cfq_mark_cfqq_on_rr(cfqq);
1342         cfqd->busy_queues++;
1343
1344         cfq_resort_rr_list(cfqd, cfqq);
1345 }
1346
1347 /*
1348  * Called when the cfqq no longer has requests pending, remove it from
1349  * the service tree.
1350  */
1351 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1352 {
1353         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1354         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1355         cfq_clear_cfqq_on_rr(cfqq);
1356
1357         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1358                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1359                 cfqq->service_tree = NULL;
1360         }
1361         if (cfqq->p_root) {
1362                 rb_erase(&cfqq->p_node, cfqq->p_root);
1363                 cfqq->p_root = NULL;
1364         }
1365
1366         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1367         BUG_ON(!cfqd->busy_queues);
1368         cfqd->busy_queues--;
1369 }
1370
1371 /*
1372  * rb tree support functions
1373  */
1374 static void cfq_del_rq_rb(struct request *rq)
1375 {
1376         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1377         const int sync = rq_is_sync(rq);
1378
1379         BUG_ON(!cfqq->queued[sync]);
1380         cfqq->queued[sync]--;
1381
1382         elv_rb_del(&cfqq->sort_list, rq);
1383
1384         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1385                 /*
1386                  * Queue will be deleted from service tree when we actually
1387                  * expire it later. Right now just remove it from prio tree
1388                  * as it is empty.
1389                  */
1390                 if (cfqq->p_root) {
1391                         rb_erase(&cfqq->p_node, cfqq->p_root);
1392                         cfqq->p_root = NULL;
1393                 }
1394         }
1395 }
1396
1397 static void cfq_add_rq_rb(struct request *rq)
1398 {
1399         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1400         struct cfq_data *cfqd = cfqq->cfqd;
1401         struct request *__alias, *prev;
1402
1403         cfqq->queued[rq_is_sync(rq)]++;
1404
1405         /*
1406          * looks a little odd, but the first insert might return an alias.
1407          * if that happens, put the alias on the dispatch list
1408          */
1409         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1410                 cfq_dispatch_insert(cfqd->queue, __alias);
1411
1412         if (!cfq_cfqq_on_rr(cfqq))
1413                 cfq_add_cfqq_rr(cfqd, cfqq);
1414
1415         /*
1416          * check if this request is a better next-serve candidate
1417          */
1418         prev = cfqq->next_rq;
1419         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1420
1421         /*
1422          * adjust priority tree position, if ->next_rq changes
1423          */
1424         if (prev != cfqq->next_rq)
1425                 cfq_prio_tree_add(cfqd, cfqq);
1426
1427         BUG_ON(!cfqq->next_rq);
1428 }
1429
1430 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1431 {
1432         elv_rb_del(&cfqq->sort_list, rq);
1433         cfqq->queued[rq_is_sync(rq)]--;
1434         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1435                                         rq_data_dir(rq), rq_is_sync(rq));
1436         cfq_add_rq_rb(rq);
1437         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1438                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1439                         rq_is_sync(rq));
1440 }
1441
1442 static struct request *
1443 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1444 {
1445         struct task_struct *tsk = current;
1446         struct cfq_io_context *cic;
1447         struct cfq_queue *cfqq;
1448
1449         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1450         if (!cic)
1451                 return NULL;
1452
1453         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1454         if (cfqq) {
1455                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1456
1457                 return elv_rb_find(&cfqq->sort_list, sector);
1458         }
1459
1460         return NULL;
1461 }
1462
1463 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1464 {
1465         struct cfq_data *cfqd = q->elevator->elevator_data;
1466
1467         cfqd->rq_in_driver++;
1468         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1469                                                 cfqd->rq_in_driver);
1470
1471         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1472 }
1473
1474 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1475 {
1476         struct cfq_data *cfqd = q->elevator->elevator_data;
1477
1478         WARN_ON(!cfqd->rq_in_driver);
1479         cfqd->rq_in_driver--;
1480         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1481                                                 cfqd->rq_in_driver);
1482 }
1483
1484 static void cfq_remove_request(struct request *rq)
1485 {
1486         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1487
1488         if (cfqq->next_rq == rq)
1489                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1490
1491         list_del_init(&rq->queuelist);
1492         cfq_del_rq_rb(rq);
1493
1494         cfqq->cfqd->rq_queued--;
1495         cfq_blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg,
1496                                         rq_data_dir(rq), rq_is_sync(rq));
1497         if (rq->cmd_flags & REQ_META) {
1498                 WARN_ON(!cfqq->meta_pending);
1499                 cfqq->meta_pending--;
1500         }
1501 }
1502
1503 static int cfq_merge(struct request_queue *q, struct request **req,
1504                      struct bio *bio)
1505 {
1506         struct cfq_data *cfqd = q->elevator->elevator_data;
1507         struct request *__rq;
1508
1509         __rq = cfq_find_rq_fmerge(cfqd, bio);
1510         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1511                 *req = __rq;
1512                 return ELEVATOR_FRONT_MERGE;
1513         }
1514
1515         return ELEVATOR_NO_MERGE;
1516 }
1517
1518 static void cfq_merged_request(struct request_queue *q, struct request *req,
1519                                int type)
1520 {
1521         if (type == ELEVATOR_FRONT_MERGE) {
1522                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1523
1524                 cfq_reposition_rq_rb(cfqq, req);
1525         }
1526 }
1527
1528 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1529                                 struct bio *bio)
1530 {
1531         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg,
1532                                         bio_data_dir(bio), cfq_bio_sync(bio));
1533 }
1534
1535 static void
1536 cfq_merged_requests(struct request_queue *q, struct request *rq,
1537                     struct request *next)
1538 {
1539         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1540         /*
1541          * reposition in fifo if next is older than rq
1542          */
1543         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1544             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1545                 list_move(&rq->queuelist, &next->queuelist);
1546                 rq_set_fifo_time(rq, rq_fifo_time(next));
1547         }
1548
1549         if (cfqq->next_rq == next)
1550                 cfqq->next_rq = rq;
1551         cfq_remove_request(next);
1552         cfq_blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg,
1553                                         rq_data_dir(next), rq_is_sync(next));
1554 }
1555
1556 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1557                            struct bio *bio)
1558 {
1559         struct cfq_data *cfqd = q->elevator->elevator_data;
1560         struct cfq_io_context *cic;
1561         struct cfq_queue *cfqq;
1562
1563         /*
1564          * Disallow merge of a sync bio into an async request.
1565          */
1566         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1567                 return false;
1568
1569         /*
1570          * Lookup the cfqq that this bio will be queued with. Allow
1571          * merge only if rq is queued there.
1572          */
1573         cic = cfq_cic_lookup(cfqd, current->io_context);
1574         if (!cic)
1575                 return false;
1576
1577         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1578         return cfqq == RQ_CFQQ(rq);
1579 }
1580
1581 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1582 {
1583         del_timer(&cfqd->idle_slice_timer);
1584         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1585 }
1586
1587 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1588                                    struct cfq_queue *cfqq)
1589 {
1590         if (cfqq) {
1591                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1592                                 cfqd->serving_prio, cfqd->serving_type);
1593                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1594                 cfqq->slice_start = 0;
1595                 cfqq->dispatch_start = jiffies;
1596                 cfqq->allocated_slice = 0;
1597                 cfqq->slice_end = 0;
1598                 cfqq->slice_dispatch = 0;
1599
1600                 cfq_clear_cfqq_wait_request(cfqq);
1601                 cfq_clear_cfqq_must_dispatch(cfqq);
1602                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1603                 cfq_clear_cfqq_fifo_expire(cfqq);
1604                 cfq_mark_cfqq_slice_new(cfqq);
1605
1606                 cfq_del_timer(cfqd, cfqq);
1607         }
1608
1609         cfqd->active_queue = cfqq;
1610 }
1611
1612 /*
1613  * current cfqq expired its slice (or was too idle), select new one
1614  */
1615 static void
1616 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1617                     bool timed_out)
1618 {
1619         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1620
1621         if (cfq_cfqq_wait_request(cfqq))
1622                 cfq_del_timer(cfqd, cfqq);
1623
1624         cfq_clear_cfqq_wait_request(cfqq);
1625         cfq_clear_cfqq_wait_busy(cfqq);
1626
1627         /*
1628          * If this cfqq is shared between multiple processes, check to
1629          * make sure that those processes are still issuing I/Os within
1630          * the mean seek distance.  If not, it may be time to break the
1631          * queues apart again.
1632          */
1633         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1634                 cfq_mark_cfqq_split_coop(cfqq);
1635
1636         /*
1637          * store what was left of this slice, if the queue idled/timed out
1638          */
1639         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1640                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1641                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1642         }
1643
1644         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1645
1646         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1647                 cfq_del_cfqq_rr(cfqd, cfqq);
1648
1649         cfq_resort_rr_list(cfqd, cfqq);
1650
1651         if (cfqq == cfqd->active_queue)
1652                 cfqd->active_queue = NULL;
1653
1654         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1655                 cfqd->grp_service_tree.active = NULL;
1656
1657         if (cfqd->active_cic) {
1658                 put_io_context(cfqd->active_cic->ioc);
1659                 cfqd->active_cic = NULL;
1660         }
1661 }
1662
1663 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1664 {
1665         struct cfq_queue *cfqq = cfqd->active_queue;
1666
1667         if (cfqq)
1668                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1669 }
1670
1671 /*
1672  * Get next queue for service. Unless we have a queue preemption,
1673  * we'll simply select the first cfqq in the service tree.
1674  */
1675 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1676 {
1677         struct cfq_rb_root *service_tree =
1678                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1679                                         cfqd->serving_type);
1680
1681         if (!cfqd->rq_queued)
1682                 return NULL;
1683
1684         /* There is nothing to dispatch */
1685         if (!service_tree)
1686                 return NULL;
1687         if (RB_EMPTY_ROOT(&service_tree->rb))
1688                 return NULL;
1689         return cfq_rb_first(service_tree);
1690 }
1691
1692 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1693 {
1694         struct cfq_group *cfqg;
1695         struct cfq_queue *cfqq;
1696         int i, j;
1697         struct cfq_rb_root *st;
1698
1699         if (!cfqd->rq_queued)
1700                 return NULL;
1701
1702         cfqg = cfq_get_next_cfqg(cfqd);
1703         if (!cfqg)
1704                 return NULL;
1705
1706         for_each_cfqg_st(cfqg, i, j, st)
1707                 if ((cfqq = cfq_rb_first(st)) != NULL)
1708                         return cfqq;
1709         return NULL;
1710 }
1711
1712 /*
1713  * Get and set a new active queue for service.
1714  */
1715 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1716                                               struct cfq_queue *cfqq)
1717 {
1718         if (!cfqq)
1719                 cfqq = cfq_get_next_queue(cfqd);
1720
1721         __cfq_set_active_queue(cfqd, cfqq);
1722         return cfqq;
1723 }
1724
1725 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1726                                           struct request *rq)
1727 {
1728         if (blk_rq_pos(rq) >= cfqd->last_position)
1729                 return blk_rq_pos(rq) - cfqd->last_position;
1730         else
1731                 return cfqd->last_position - blk_rq_pos(rq);
1732 }
1733
1734 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1735                                struct request *rq)
1736 {
1737         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1738 }
1739
1740 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1741                                     struct cfq_queue *cur_cfqq)
1742 {
1743         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1744         struct rb_node *parent, *node;
1745         struct cfq_queue *__cfqq;
1746         sector_t sector = cfqd->last_position;
1747
1748         if (RB_EMPTY_ROOT(root))
1749                 return NULL;
1750
1751         /*
1752          * First, if we find a request starting at the end of the last
1753          * request, choose it.
1754          */
1755         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1756         if (__cfqq)
1757                 return __cfqq;
1758
1759         /*
1760          * If the exact sector wasn't found, the parent of the NULL leaf
1761          * will contain the closest sector.
1762          */
1763         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1764         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1765                 return __cfqq;
1766
1767         if (blk_rq_pos(__cfqq->next_rq) < sector)
1768                 node = rb_next(&__cfqq->p_node);
1769         else
1770                 node = rb_prev(&__cfqq->p_node);
1771         if (!node)
1772                 return NULL;
1773
1774         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1775         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1776                 return __cfqq;
1777
1778         return NULL;
1779 }
1780
1781 /*
1782  * cfqd - obvious
1783  * cur_cfqq - passed in so that we don't decide that the current queue is
1784  *            closely cooperating with itself.
1785  *
1786  * So, basically we're assuming that that cur_cfqq has dispatched at least
1787  * one request, and that cfqd->last_position reflects a position on the disk
1788  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1789  * assumption.
1790  */
1791 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1792                                               struct cfq_queue *cur_cfqq)
1793 {
1794         struct cfq_queue *cfqq;
1795
1796         if (cfq_class_idle(cur_cfqq))
1797                 return NULL;
1798         if (!cfq_cfqq_sync(cur_cfqq))
1799                 return NULL;
1800         if (CFQQ_SEEKY(cur_cfqq))
1801                 return NULL;
1802
1803         /*
1804          * Don't search priority tree if it's the only queue in the group.
1805          */
1806         if (cur_cfqq->cfqg->nr_cfqq == 1)
1807                 return NULL;
1808
1809         /*
1810          * We should notice if some of the queues are cooperating, eg
1811          * working closely on the same area of the disk. In that case,
1812          * we can group them together and don't waste time idling.
1813          */
1814         cfqq = cfqq_close(cfqd, cur_cfqq);
1815         if (!cfqq)
1816                 return NULL;
1817
1818         /* If new queue belongs to different cfq_group, don't choose it */
1819         if (cur_cfqq->cfqg != cfqq->cfqg)
1820                 return NULL;
1821
1822         /*
1823          * It only makes sense to merge sync queues.
1824          */
1825         if (!cfq_cfqq_sync(cfqq))
1826                 return NULL;
1827         if (CFQQ_SEEKY(cfqq))
1828                 return NULL;
1829
1830         /*
1831          * Do not merge queues of different priority classes
1832          */
1833         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1834                 return NULL;
1835
1836         return cfqq;
1837 }
1838
1839 /*
1840  * Determine whether we should enforce idle window for this queue.
1841  */
1842
1843 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1844 {
1845         enum wl_prio_t prio = cfqq_prio(cfqq);
1846         struct cfq_rb_root *service_tree = cfqq->service_tree;
1847
1848         BUG_ON(!service_tree);
1849         BUG_ON(!service_tree->count);
1850
1851         /* We never do for idle class queues. */
1852         if (prio == IDLE_WORKLOAD)
1853                 return false;
1854
1855         /* We do for queues that were marked with idle window flag. */
1856         if (cfq_cfqq_idle_window(cfqq) &&
1857            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1858                 return true;
1859
1860         /*
1861          * Otherwise, we do only if they are the last ones
1862          * in their service tree.
1863          */
1864         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1865                 return 1;
1866         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1867                         service_tree->count);
1868         return 0;
1869 }
1870
1871 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1872 {
1873         struct cfq_queue *cfqq = cfqd->active_queue;
1874         struct cfq_io_context *cic;
1875         unsigned long sl;
1876
1877         /*
1878          * SSD device without seek penalty, disable idling. But only do so
1879          * for devices that support queuing, otherwise we still have a problem
1880          * with sync vs async workloads.
1881          */
1882         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1883                 return;
1884
1885         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1886         WARN_ON(cfq_cfqq_slice_new(cfqq));
1887
1888         /*
1889          * idle is disabled, either manually or by past process history
1890          */
1891         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1892                 return;
1893
1894         /*
1895          * still active requests from this queue, don't idle
1896          */
1897         if (cfqq->dispatched)
1898                 return;
1899
1900         /*
1901          * task has exited, don't wait
1902          */
1903         cic = cfqd->active_cic;
1904         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1905                 return;
1906
1907         /*
1908          * If our average think time is larger than the remaining time
1909          * slice, then don't idle. This avoids overrunning the allotted
1910          * time slice.
1911          */
1912         if (sample_valid(cic->ttime_samples) &&
1913             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1914                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1915                                 cic->ttime_mean);
1916                 return;
1917         }
1918
1919         cfq_mark_cfqq_wait_request(cfqq);
1920
1921         sl = cfqd->cfq_slice_idle;
1922
1923         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1924         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1925         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1926 }
1927
1928 /*
1929  * Move request from internal lists to the request queue dispatch list.
1930  */
1931 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1932 {
1933         struct cfq_data *cfqd = q->elevator->elevator_data;
1934         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1935
1936         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1937
1938         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1939         cfq_remove_request(rq);
1940         cfqq->dispatched++;
1941         elv_dispatch_sort(q, rq);
1942
1943         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1944         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1945                                         rq_data_dir(rq), rq_is_sync(rq));
1946 }
1947
1948 /*
1949  * return expired entry, or NULL to just start from scratch in rbtree
1950  */
1951 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1952 {
1953         struct request *rq = NULL;
1954
1955         if (cfq_cfqq_fifo_expire(cfqq))
1956                 return NULL;
1957
1958         cfq_mark_cfqq_fifo_expire(cfqq);
1959
1960         if (list_empty(&cfqq->fifo))
1961                 return NULL;
1962
1963         rq = rq_entry_fifo(cfqq->fifo.next);
1964         if (time_before(jiffies, rq_fifo_time(rq)))
1965                 rq = NULL;
1966
1967         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1968         return rq;
1969 }
1970
1971 static inline int
1972 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1973 {
1974         const int base_rq = cfqd->cfq_slice_async_rq;
1975
1976         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1977
1978         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1979 }
1980
1981 /*
1982  * Must be called with the queue_lock held.
1983  */
1984 static int cfqq_process_refs(struct cfq_queue *cfqq)
1985 {
1986         int process_refs, io_refs;
1987
1988         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1989         process_refs = atomic_read(&cfqq->ref) - io_refs;
1990         BUG_ON(process_refs < 0);
1991         return process_refs;
1992 }
1993
1994 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1995 {
1996         int process_refs, new_process_refs;
1997         struct cfq_queue *__cfqq;
1998
1999         /*
2000          * If there are no process references on the new_cfqq, then it is
2001          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2002          * chain may have dropped their last reference (not just their
2003          * last process reference).
2004          */
2005         if (!cfqq_process_refs(new_cfqq))
2006                 return;
2007
2008         /* Avoid a circular list and skip interim queue merges */
2009         while ((__cfqq = new_cfqq->new_cfqq)) {
2010                 if (__cfqq == cfqq)
2011                         return;
2012                 new_cfqq = __cfqq;
2013         }
2014
2015         process_refs = cfqq_process_refs(cfqq);
2016         new_process_refs = cfqq_process_refs(new_cfqq);
2017         /*
2018          * If the process for the cfqq has gone away, there is no
2019          * sense in merging the queues.
2020          */
2021         if (process_refs == 0 || new_process_refs == 0)
2022                 return;
2023
2024         /*
2025          * Merge in the direction of the lesser amount of work.
2026          */
2027         if (new_process_refs >= process_refs) {
2028                 cfqq->new_cfqq = new_cfqq;
2029                 atomic_add(process_refs, &new_cfqq->ref);
2030         } else {
2031                 new_cfqq->new_cfqq = cfqq;
2032                 atomic_add(new_process_refs, &cfqq->ref);
2033         }
2034 }
2035
2036 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2037                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2038 {
2039         struct cfq_queue *queue;
2040         int i;
2041         bool key_valid = false;
2042         unsigned long lowest_key = 0;
2043         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2044
2045         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2046                 /* select the one with lowest rb_key */
2047                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2048                 if (queue &&
2049                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2050                         lowest_key = queue->rb_key;
2051                         cur_best = i;
2052                         key_valid = true;
2053                 }
2054         }
2055
2056         return cur_best;
2057 }
2058
2059 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2060 {
2061         unsigned slice;
2062         unsigned count;
2063         struct cfq_rb_root *st;
2064         unsigned group_slice;
2065
2066         if (!cfqg) {
2067                 cfqd->serving_prio = IDLE_WORKLOAD;
2068                 cfqd->workload_expires = jiffies + 1;
2069                 return;
2070         }
2071
2072         /* Choose next priority. RT > BE > IDLE */
2073         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2074                 cfqd->serving_prio = RT_WORKLOAD;
2075         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2076                 cfqd->serving_prio = BE_WORKLOAD;
2077         else {
2078                 cfqd->serving_prio = IDLE_WORKLOAD;
2079                 cfqd->workload_expires = jiffies + 1;
2080                 return;
2081         }
2082
2083         /*
2084          * For RT and BE, we have to choose also the type
2085          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2086          * expiration time
2087          */
2088         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2089         count = st->count;
2090
2091         /*
2092          * check workload expiration, and that we still have other queues ready
2093          */
2094         if (count && !time_after(jiffies, cfqd->workload_expires))
2095                 return;
2096
2097         /* otherwise select new workload type */
2098         cfqd->serving_type =
2099                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2100         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2101         count = st->count;
2102
2103         /*
2104          * the workload slice is computed as a fraction of target latency
2105          * proportional to the number of queues in that workload, over
2106          * all the queues in the same priority class
2107          */
2108         group_slice = cfq_group_slice(cfqd, cfqg);
2109
2110         slice = group_slice * count /
2111                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2112                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2113
2114         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2115                 unsigned int tmp;
2116
2117                 /*
2118                  * Async queues are currently system wide. Just taking
2119                  * proportion of queues with-in same group will lead to higher
2120                  * async ratio system wide as generally root group is going
2121                  * to have higher weight. A more accurate thing would be to
2122                  * calculate system wide asnc/sync ratio.
2123                  */
2124                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2125                 tmp = tmp/cfqd->busy_queues;
2126                 slice = min_t(unsigned, slice, tmp);
2127
2128                 /* async workload slice is scaled down according to
2129                  * the sync/async slice ratio. */
2130                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2131         } else
2132                 /* sync workload slice is at least 2 * cfq_slice_idle */
2133                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2134
2135         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2136         cfq_log(cfqd, "workload slice:%d", slice);
2137         cfqd->workload_expires = jiffies + slice;
2138 }
2139
2140 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2141 {
2142         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2143         struct cfq_group *cfqg;
2144
2145         if (RB_EMPTY_ROOT(&st->rb))
2146                 return NULL;
2147         cfqg = cfq_rb_first_group(st);
2148         st->active = &cfqg->rb_node;
2149         update_min_vdisktime(st);
2150         return cfqg;
2151 }
2152
2153 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2154 {
2155         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2156
2157         cfqd->serving_group = cfqg;
2158
2159         /* Restore the workload type data */
2160         if (cfqg->saved_workload_slice) {
2161                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2162                 cfqd->serving_type = cfqg->saved_workload;
2163                 cfqd->serving_prio = cfqg->saved_serving_prio;
2164         } else
2165                 cfqd->workload_expires = jiffies - 1;
2166
2167         choose_service_tree(cfqd, cfqg);
2168 }
2169
2170 /*
2171  * Select a queue for service. If we have a current active queue,
2172  * check whether to continue servicing it, or retrieve and set a new one.
2173  */
2174 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2175 {
2176         struct cfq_queue *cfqq, *new_cfqq = NULL;
2177
2178         cfqq = cfqd->active_queue;
2179         if (!cfqq)
2180                 goto new_queue;
2181
2182         if (!cfqd->rq_queued)
2183                 return NULL;
2184
2185         /*
2186          * We were waiting for group to get backlogged. Expire the queue
2187          */
2188         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2189                 goto expire;
2190
2191         /*
2192          * The active queue has run out of time, expire it and select new.
2193          */
2194         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2195                 /*
2196                  * If slice had not expired at the completion of last request
2197                  * we might not have turned on wait_busy flag. Don't expire
2198                  * the queue yet. Allow the group to get backlogged.
2199                  *
2200                  * The very fact that we have used the slice, that means we
2201                  * have been idling all along on this queue and it should be
2202                  * ok to wait for this request to complete.
2203                  */
2204                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2205                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2206                         cfqq = NULL;
2207                         goto keep_queue;
2208                 } else
2209                         goto expire;
2210         }
2211
2212         /*
2213          * The active queue has requests and isn't expired, allow it to
2214          * dispatch.
2215          */
2216         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2217                 goto keep_queue;
2218
2219         /*
2220          * If another queue has a request waiting within our mean seek
2221          * distance, let it run.  The expire code will check for close
2222          * cooperators and put the close queue at the front of the service
2223          * tree.  If possible, merge the expiring queue with the new cfqq.
2224          */
2225         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2226         if (new_cfqq) {
2227                 if (!cfqq->new_cfqq)
2228                         cfq_setup_merge(cfqq, new_cfqq);
2229                 goto expire;
2230         }
2231
2232         /*
2233          * No requests pending. If the active queue still has requests in
2234          * flight or is idling for a new request, allow either of these
2235          * conditions to happen (or time out) before selecting a new queue.
2236          */
2237         if (timer_pending(&cfqd->idle_slice_timer) ||
2238             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2239                 cfqq = NULL;
2240                 goto keep_queue;
2241         }
2242
2243 expire:
2244         cfq_slice_expired(cfqd, 0);
2245 new_queue:
2246         /*
2247          * Current queue expired. Check if we have to switch to a new
2248          * service tree
2249          */
2250         if (!new_cfqq)
2251                 cfq_choose_cfqg(cfqd);
2252
2253         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2254 keep_queue:
2255         return cfqq;
2256 }
2257
2258 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2259 {
2260         int dispatched = 0;
2261
2262         while (cfqq->next_rq) {
2263                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2264                 dispatched++;
2265         }
2266
2267         BUG_ON(!list_empty(&cfqq->fifo));
2268
2269         /* By default cfqq is not expired if it is empty. Do it explicitly */
2270         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2271         return dispatched;
2272 }
2273
2274 /*
2275  * Drain our current requests. Used for barriers and when switching
2276  * io schedulers on-the-fly.
2277  */
2278 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2279 {
2280         struct cfq_queue *cfqq;
2281         int dispatched = 0;
2282
2283         /* Expire the timeslice of the current active queue first */
2284         cfq_slice_expired(cfqd, 0);
2285         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2286                 __cfq_set_active_queue(cfqd, cfqq);
2287                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2288         }
2289
2290         BUG_ON(cfqd->busy_queues);
2291
2292         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2293         return dispatched;
2294 }
2295
2296 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2297         struct cfq_queue *cfqq)
2298 {
2299         /* the queue hasn't finished any request, can't estimate */
2300         if (cfq_cfqq_slice_new(cfqq))
2301                 return 1;
2302         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2303                 cfqq->slice_end))
2304                 return 1;
2305
2306         return 0;
2307 }
2308
2309 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2310 {
2311         unsigned int max_dispatch;
2312
2313         /*
2314          * Drain async requests before we start sync IO
2315          */
2316         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2317                 return false;
2318
2319         /*
2320          * If this is an async queue and we have sync IO in flight, let it wait
2321          */
2322         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2323                 return false;
2324
2325         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2326         if (cfq_class_idle(cfqq))
2327                 max_dispatch = 1;
2328
2329         /*
2330          * Does this cfqq already have too much IO in flight?
2331          */
2332         if (cfqq->dispatched >= max_dispatch) {
2333                 /*
2334                  * idle queue must always only have a single IO in flight
2335                  */
2336                 if (cfq_class_idle(cfqq))
2337                         return false;
2338
2339                 /*
2340                  * We have other queues, don't allow more IO from this one
2341                  */
2342                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2343                         return false;
2344
2345                 /*
2346                  * Sole queue user, no limit
2347                  */
2348                 if (cfqd->busy_queues == 1)
2349                         max_dispatch = -1;
2350                 else
2351                         /*
2352                          * Normally we start throttling cfqq when cfq_quantum/2
2353                          * requests have been dispatched. But we can drive
2354                          * deeper queue depths at the beginning of slice
2355                          * subjected to upper limit of cfq_quantum.
2356                          * */
2357                         max_dispatch = cfqd->cfq_quantum;
2358         }
2359
2360         /*
2361          * Async queues must wait a bit before being allowed dispatch.
2362          * We also ramp up the dispatch depth gradually for async IO,
2363          * based on the last sync IO we serviced
2364          */
2365         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2366                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2367                 unsigned int depth;
2368
2369                 depth = last_sync / cfqd->cfq_slice[1];
2370                 if (!depth && !cfqq->dispatched)
2371                         depth = 1;
2372                 if (depth < max_dispatch)
2373                         max_dispatch = depth;
2374         }
2375
2376         /*
2377          * If we're below the current max, allow a dispatch
2378          */
2379         return cfqq->dispatched < max_dispatch;
2380 }
2381
2382 /*
2383  * Dispatch a request from cfqq, moving them to the request queue
2384  * dispatch list.
2385  */
2386 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2387 {
2388         struct request *rq;
2389
2390         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2391
2392         if (!cfq_may_dispatch(cfqd, cfqq))
2393                 return false;
2394
2395         /*
2396          * follow expired path, else get first next available
2397          */
2398         rq = cfq_check_fifo(cfqq);
2399         if (!rq)
2400                 rq = cfqq->next_rq;
2401
2402         /*
2403          * insert request into driver dispatch list
2404          */
2405         cfq_dispatch_insert(cfqd->queue, rq);
2406
2407         if (!cfqd->active_cic) {
2408                 struct cfq_io_context *cic = RQ_CIC(rq);
2409
2410                 atomic_long_inc(&cic->ioc->refcount);
2411                 cfqd->active_cic = cic;
2412         }
2413
2414         return true;
2415 }
2416
2417 /*
2418  * Find the cfqq that we need to service and move a request from that to the
2419  * dispatch list
2420  */
2421 static int cfq_dispatch_requests(struct request_queue *q, int force)
2422 {
2423         struct cfq_data *cfqd = q->elevator->elevator_data;
2424         struct cfq_queue *cfqq;
2425
2426         if (!cfqd->busy_queues)
2427                 return 0;
2428
2429         if (unlikely(force))
2430                 return cfq_forced_dispatch(cfqd);
2431
2432         cfqq = cfq_select_queue(cfqd);
2433         if (!cfqq)
2434                 return 0;
2435
2436         /*
2437          * Dispatch a request from this cfqq, if it is allowed
2438          */
2439         if (!cfq_dispatch_request(cfqd, cfqq))
2440                 return 0;
2441
2442         cfqq->slice_dispatch++;
2443         cfq_clear_cfqq_must_dispatch(cfqq);
2444
2445         /*
2446          * expire an async queue immediately if it has used up its slice. idle
2447          * queue always expire after 1 dispatch round.
2448          */
2449         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2450             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2451             cfq_class_idle(cfqq))) {
2452                 cfqq->slice_end = jiffies + 1;
2453                 cfq_slice_expired(cfqd, 0);
2454         }
2455
2456         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2457         return 1;
2458 }
2459
2460 /*
2461  * task holds one reference to the queue, dropped when task exits. each rq
2462  * in-flight on this queue also holds a reference, dropped when rq is freed.
2463  *
2464  * Each cfq queue took a reference on the parent group. Drop it now.
2465  * queue lock must be held here.
2466  */
2467 static void cfq_put_queue(struct cfq_queue *cfqq)
2468 {
2469         struct cfq_data *cfqd = cfqq->cfqd;
2470         struct cfq_group *cfqg, *orig_cfqg;
2471
2472         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2473
2474         if (!atomic_dec_and_test(&cfqq->ref))
2475                 return;
2476
2477         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2478         BUG_ON(rb_first(&cfqq->sort_list));
2479         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2480         cfqg = cfqq->cfqg;
2481         orig_cfqg = cfqq->orig_cfqg;
2482
2483         if (unlikely(cfqd->active_queue == cfqq)) {
2484                 __cfq_slice_expired(cfqd, cfqq, 0);
2485                 cfq_schedule_dispatch(cfqd);
2486         }
2487
2488         BUG_ON(cfq_cfqq_on_rr(cfqq));
2489         kmem_cache_free(cfq_pool, cfqq);
2490         cfq_put_cfqg(cfqg);
2491         if (orig_cfqg)
2492                 cfq_put_cfqg(orig_cfqg);
2493 }
2494
2495 /*
2496  * Must always be called with the rcu_read_lock() held
2497  */
2498 static void
2499 __call_for_each_cic(struct io_context *ioc,
2500                     void (*func)(struct io_context *, struct cfq_io_context *))
2501 {
2502         struct cfq_io_context *cic;
2503         struct hlist_node *n;
2504
2505         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2506                 func(ioc, cic);
2507 }
2508
2509 /*
2510  * Call func for each cic attached to this ioc.
2511  */
2512 static void
2513 call_for_each_cic(struct io_context *ioc,
2514                   void (*func)(struct io_context *, struct cfq_io_context *))
2515 {
2516         rcu_read_lock();
2517         __call_for_each_cic(ioc, func);
2518         rcu_read_unlock();
2519 }
2520
2521 static void cfq_cic_free_rcu(struct rcu_head *head)
2522 {
2523         struct cfq_io_context *cic;
2524
2525         cic = container_of(head, struct cfq_io_context, rcu_head);
2526
2527         kmem_cache_free(cfq_ioc_pool, cic);
2528         elv_ioc_count_dec(cfq_ioc_count);
2529
2530         if (ioc_gone) {
2531                 /*
2532                  * CFQ scheduler is exiting, grab exit lock and check
2533                  * the pending io context count. If it hits zero,
2534                  * complete ioc_gone and set it back to NULL
2535                  */
2536                 spin_lock(&ioc_gone_lock);
2537                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2538                         complete(ioc_gone);
2539                         ioc_gone = NULL;
2540                 }
2541                 spin_unlock(&ioc_gone_lock);
2542         }
2543 }
2544
2545 static void cfq_cic_free(struct cfq_io_context *cic)
2546 {
2547         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2548 }
2549
2550 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2551 {
2552         unsigned long flags;
2553         unsigned long dead_key = (unsigned long) cic->key;
2554
2555         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2556
2557         spin_lock_irqsave(&ioc->lock, flags);
2558         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2559         hlist_del_rcu(&cic->cic_list);
2560         spin_unlock_irqrestore(&ioc->lock, flags);
2561
2562         cfq_cic_free(cic);
2563 }
2564
2565 /*
2566  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2567  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2568  * and ->trim() which is called with the task lock held
2569  */
2570 static void cfq_free_io_context(struct io_context *ioc)
2571 {
2572         /*
2573          * ioc->refcount is zero here, or we are called from elv_unregister(),
2574          * so no more cic's are allowed to be linked into this ioc.  So it
2575          * should be ok to iterate over the known list, we will see all cic's
2576          * since no new ones are added.
2577          */
2578         __call_for_each_cic(ioc, cic_free_func);
2579 }
2580
2581 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2582 {
2583         struct cfq_queue *__cfqq, *next;
2584
2585         /*
2586          * If this queue was scheduled to merge with another queue, be
2587          * sure to drop the reference taken on that queue (and others in
2588          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2589          */
2590         __cfqq = cfqq->new_cfqq;
2591         while (__cfqq) {
2592                 if (__cfqq == cfqq) {
2593                         WARN(1, "cfqq->new_cfqq loop detected\n");
2594                         break;
2595                 }
2596                 next = __cfqq->new_cfqq;
2597                 cfq_put_queue(__cfqq);
2598                 __cfqq = next;
2599         }
2600 }
2601
2602 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2603 {
2604         if (unlikely(cfqq == cfqd->active_queue)) {
2605                 __cfq_slice_expired(cfqd, cfqq, 0);
2606                 cfq_schedule_dispatch(cfqd);
2607         }
2608
2609         cfq_put_cooperator(cfqq);
2610
2611         cfq_put_queue(cfqq);
2612 }
2613
2614 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2615                                          struct cfq_io_context *cic)
2616 {
2617         struct io_context *ioc = cic->ioc;
2618
2619         list_del_init(&cic->queue_list);
2620
2621         /*
2622          * Make sure dead mark is seen for dead queues
2623          */
2624         smp_wmb();
2625         cic->key = cfqd_dead_key(cfqd);
2626
2627         if (ioc->ioc_data == cic)
2628                 rcu_assign_pointer(ioc->ioc_data, NULL);
2629
2630         if (cic->cfqq[BLK_RW_ASYNC]) {
2631                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2632                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2633         }
2634
2635         if (cic->cfqq[BLK_RW_SYNC]) {
2636                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2637                 cic->cfqq[BLK_RW_SYNC] = NULL;
2638         }
2639 }
2640
2641 static void cfq_exit_single_io_context(struct io_context *ioc,
2642                                        struct cfq_io_context *cic)
2643 {
2644         struct cfq_data *cfqd = cic_to_cfqd(cic);
2645
2646         if (cfqd) {
2647                 struct request_queue *q = cfqd->queue;
2648                 unsigned long flags;
2649
2650                 spin_lock_irqsave(q->queue_lock, flags);
2651
2652                 /*
2653                  * Ensure we get a fresh copy of the ->key to prevent
2654                  * race between exiting task and queue
2655                  */
2656                 smp_read_barrier_depends();
2657                 if (cic->key == cfqd)
2658                         __cfq_exit_single_io_context(cfqd, cic);
2659
2660                 spin_unlock_irqrestore(q->queue_lock, flags);
2661         }
2662 }
2663
2664 /*
2665  * The process that ioc belongs to has exited, we need to clean up
2666  * and put the internal structures we have that belongs to that process.
2667  */
2668 static void cfq_exit_io_context(struct io_context *ioc)
2669 {
2670         call_for_each_cic(ioc, cfq_exit_single_io_context);
2671 }
2672
2673 static struct cfq_io_context *
2674 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2675 {
2676         struct cfq_io_context *cic;
2677
2678         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2679                                                         cfqd->queue->node);
2680         if (cic) {
2681                 cic->last_end_request = jiffies;
2682                 INIT_LIST_HEAD(&cic->queue_list);
2683                 INIT_HLIST_NODE(&cic->cic_list);
2684                 cic->dtor = cfq_free_io_context;
2685                 cic->exit = cfq_exit_io_context;
2686                 elv_ioc_count_inc(cfq_ioc_count);
2687         }
2688
2689         return cic;
2690 }
2691
2692 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2693 {
2694         struct task_struct *tsk = current;
2695         int ioprio_class;
2696
2697         if (!cfq_cfqq_prio_changed(cfqq))
2698                 return;
2699
2700         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2701         switch (ioprio_class) {
2702         default:
2703                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2704         case IOPRIO_CLASS_NONE:
2705                 /*
2706                  * no prio set, inherit CPU scheduling settings
2707                  */
2708                 cfqq->ioprio = task_nice_ioprio(tsk);
2709                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2710                 break;
2711         case IOPRIO_CLASS_RT:
2712                 cfqq->ioprio = task_ioprio(ioc);
2713                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2714                 break;
2715         case IOPRIO_CLASS_BE:
2716                 cfqq->ioprio = task_ioprio(ioc);
2717                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2718                 break;
2719         case IOPRIO_CLASS_IDLE:
2720                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2721                 cfqq->ioprio = 7;
2722                 cfq_clear_cfqq_idle_window(cfqq);
2723                 break;
2724         }
2725
2726         /*
2727          * keep track of original prio settings in case we have to temporarily
2728          * elevate the priority of this queue
2729          */
2730         cfqq->org_ioprio = cfqq->ioprio;
2731         cfqq->org_ioprio_class = cfqq->ioprio_class;
2732         cfq_clear_cfqq_prio_changed(cfqq);
2733 }
2734
2735 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2736 {
2737         struct cfq_data *cfqd = cic_to_cfqd(cic);
2738         struct cfq_queue *cfqq;
2739         unsigned long flags;
2740
2741         if (unlikely(!cfqd))
2742                 return;
2743
2744         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2745
2746         cfqq = cic->cfqq[BLK_RW_ASYNC];
2747         if (cfqq) {
2748                 struct cfq_queue *new_cfqq;
2749                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2750                                                 GFP_ATOMIC);
2751                 if (new_cfqq) {
2752                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2753                         cfq_put_queue(cfqq);
2754                 }
2755         }
2756
2757         cfqq = cic->cfqq[BLK_RW_SYNC];
2758         if (cfqq)
2759                 cfq_mark_cfqq_prio_changed(cfqq);
2760
2761         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2762 }
2763
2764 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2765 {
2766         call_for_each_cic(ioc, changed_ioprio);
2767         ioc->ioprio_changed = 0;
2768 }
2769
2770 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2771                           pid_t pid, bool is_sync)
2772 {
2773         RB_CLEAR_NODE(&cfqq->rb_node);
2774         RB_CLEAR_NODE(&cfqq->p_node);
2775         INIT_LIST_HEAD(&cfqq->fifo);
2776
2777         atomic_set(&cfqq->ref, 0);
2778         cfqq->cfqd = cfqd;
2779
2780         cfq_mark_cfqq_prio_changed(cfqq);
2781
2782         if (is_sync) {
2783                 if (!cfq_class_idle(cfqq))
2784                         cfq_mark_cfqq_idle_window(cfqq);
2785                 cfq_mark_cfqq_sync(cfqq);
2786         }
2787         cfqq->pid = pid;
2788 }
2789
2790 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2791 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2792 {
2793         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2794         struct cfq_data *cfqd = cic_to_cfqd(cic);
2795         unsigned long flags;
2796         struct request_queue *q;
2797
2798         if (unlikely(!cfqd))
2799                 return;
2800
2801         q = cfqd->queue;
2802
2803         spin_lock_irqsave(q->queue_lock, flags);
2804
2805         if (sync_cfqq) {
2806                 /*
2807                  * Drop reference to sync queue. A new sync queue will be
2808                  * assigned in new group upon arrival of a fresh request.
2809                  */
2810                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2811                 cic_set_cfqq(cic, NULL, 1);
2812                 cfq_put_queue(sync_cfqq);
2813         }
2814
2815         spin_unlock_irqrestore(q->queue_lock, flags);
2816 }
2817
2818 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2819 {
2820         call_for_each_cic(ioc, changed_cgroup);
2821         ioc->cgroup_changed = 0;
2822 }
2823 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2824
2825 static struct cfq_queue *
2826 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2827                      struct io_context *ioc, gfp_t gfp_mask)
2828 {
2829         struct cfq_queue *cfqq, *new_cfqq = NULL;
2830         struct cfq_io_context *cic;
2831         struct cfq_group *cfqg;
2832
2833 retry:
2834         cfqg = cfq_get_cfqg(cfqd, 1);
2835         cic = cfq_cic_lookup(cfqd, ioc);
2836         /* cic always exists here */
2837         cfqq = cic_to_cfqq(cic, is_sync);
2838
2839         /*
2840          * Always try a new alloc if we fell back to the OOM cfqq
2841          * originally, since it should just be a temporary situation.
2842          */
2843         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2844                 cfqq = NULL;
2845                 if (new_cfqq) {
2846                         cfqq = new_cfqq;
2847                         new_cfqq = NULL;
2848                 } else if (gfp_mask & __GFP_WAIT) {
2849                         spin_unlock_irq(cfqd->queue->queue_lock);
2850                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2851                                         gfp_mask | __GFP_ZERO,
2852                                         cfqd->queue->node);
2853                         spin_lock_irq(cfqd->queue->queue_lock);
2854                         if (new_cfqq)
2855                                 goto retry;
2856                 } else {
2857                         cfqq = kmem_cache_alloc_node(cfq_pool,
2858                                         gfp_mask | __GFP_ZERO,
2859                                         cfqd->queue->node);
2860                 }
2861
2862                 if (cfqq) {
2863                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2864                         cfq_init_prio_data(cfqq, ioc);
2865                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2866                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2867                 } else
2868                         cfqq = &cfqd->oom_cfqq;
2869         }
2870
2871         if (new_cfqq)
2872                 kmem_cache_free(cfq_pool, new_cfqq);
2873
2874         return cfqq;
2875 }
2876
2877 static struct cfq_queue **
2878 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2879 {
2880         switch (ioprio_class) {
2881         case IOPRIO_CLASS_RT:
2882                 return &cfqd->async_cfqq[0][ioprio];
2883         case IOPRIO_CLASS_BE:
2884                 return &cfqd->async_cfqq[1][ioprio];
2885         case IOPRIO_CLASS_IDLE:
2886                 return &cfqd->async_idle_cfqq;
2887         default:
2888                 BUG();
2889         }
2890 }
2891
2892 static struct cfq_queue *
2893 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2894               gfp_t gfp_mask)
2895 {
2896         const int ioprio = task_ioprio(ioc);
2897         const int ioprio_class = task_ioprio_class(ioc);
2898         struct cfq_queue **async_cfqq = NULL;
2899         struct cfq_queue *cfqq = NULL;
2900
2901         if (!is_sync) {
2902                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2903                 cfqq = *async_cfqq;
2904         }
2905
2906         if (!cfqq)
2907                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2908
2909         /*
2910          * pin the queue now that it's allocated, scheduler exit will prune it
2911          */
2912         if (!is_sync && !(*async_cfqq)) {
2913                 atomic_inc(&cfqq->ref);
2914                 *async_cfqq = cfqq;
2915         }
2916
2917         atomic_inc(&cfqq->ref);
2918         return cfqq;
2919 }
2920
2921 /*
2922  * We drop cfq io contexts lazily, so we may find a dead one.
2923  */
2924 static void
2925 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2926                   struct cfq_io_context *cic)
2927 {
2928         unsigned long flags;
2929
2930         WARN_ON(!list_empty(&cic->queue_list));
2931         BUG_ON(cic->key != cfqd_dead_key(cfqd));
2932
2933         spin_lock_irqsave(&ioc->lock, flags);
2934
2935         BUG_ON(ioc->ioc_data == cic);
2936
2937         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2938         hlist_del_rcu(&cic->cic_list);
2939         spin_unlock_irqrestore(&ioc->lock, flags);
2940
2941         cfq_cic_free(cic);
2942 }
2943
2944 static struct cfq_io_context *
2945 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2946 {
2947         struct cfq_io_context *cic;
2948         unsigned long flags;
2949
2950         if (unlikely(!ioc))
2951                 return NULL;
2952
2953         rcu_read_lock();
2954
2955         /*
2956          * we maintain a last-hit cache, to avoid browsing over the tree
2957          */
2958         cic = rcu_dereference(ioc->ioc_data);
2959         if (cic && cic->key == cfqd) {
2960                 rcu_read_unlock();
2961                 return cic;
2962         }
2963
2964         do {
2965                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
2966                 rcu_read_unlock();
2967                 if (!cic)
2968                         break;
2969                 if (unlikely(cic->key != cfqd)) {
2970                         cfq_drop_dead_cic(cfqd, ioc, cic);
2971                         rcu_read_lock();
2972                         continue;
2973                 }
2974
2975                 spin_lock_irqsave(&ioc->lock, flags);
2976                 rcu_assign_pointer(ioc->ioc_data, cic);
2977                 spin_unlock_irqrestore(&ioc->lock, flags);
2978                 break;
2979         } while (1);
2980
2981         return cic;
2982 }
2983
2984 /*
2985  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2986  * the process specific cfq io context when entered from the block layer.
2987  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2988  */
2989 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2990                         struct cfq_io_context *cic, gfp_t gfp_mask)
2991 {
2992         unsigned long flags;
2993         int ret;
2994
2995         ret = radix_tree_preload(gfp_mask);
2996         if (!ret) {
2997                 cic->ioc = ioc;
2998                 cic->key = cfqd;
2999
3000                 spin_lock_irqsave(&ioc->lock, flags);
3001                 ret = radix_tree_insert(&ioc->radix_root,
3002                                                 cfqd->cic_index, cic);
3003                 if (!ret)
3004                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3005                 spin_unlock_irqrestore(&ioc->lock, flags);
3006
3007                 radix_tree_preload_end();
3008
3009                 if (!ret) {
3010                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3011                         list_add(&cic->queue_list, &cfqd->cic_list);
3012                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3013                 }
3014         }
3015
3016         if (ret)
3017                 printk(KERN_ERR "cfq: cic link failed!\n");
3018
3019         return ret;
3020 }
3021
3022 /*
3023  * Setup general io context and cfq io context. There can be several cfq
3024  * io contexts per general io context, if this process is doing io to more
3025  * than one device managed by cfq.
3026  */
3027 static struct cfq_io_context *
3028 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3029 {
3030         struct io_context *ioc = NULL;
3031         struct cfq_io_context *cic;
3032
3033         might_sleep_if(gfp_mask & __GFP_WAIT);
3034
3035         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3036         if (!ioc)
3037                 return NULL;
3038
3039         cic = cfq_cic_lookup(cfqd, ioc);
3040         if (cic)
3041                 goto out;
3042
3043         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3044         if (cic == NULL)
3045                 goto err;
3046
3047         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3048                 goto err_free;
3049
3050 out:
3051         smp_read_barrier_depends();
3052         if (unlikely(ioc->ioprio_changed))
3053                 cfq_ioc_set_ioprio(ioc);
3054
3055 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3056         if (unlikely(ioc->cgroup_changed))
3057                 cfq_ioc_set_cgroup(ioc);
3058 #endif
3059         return cic;
3060 err_free:
3061         cfq_cic_free(cic);
3062 err:
3063         put_io_context(ioc);
3064         return NULL;
3065 }
3066
3067 static void
3068 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3069 {
3070         unsigned long elapsed = jiffies - cic->last_end_request;
3071         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3072
3073         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3074         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3075         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3076 }
3077
3078 static void
3079 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3080                        struct request *rq)
3081 {
3082         sector_t sdist = 0;
3083         sector_t n_sec = blk_rq_sectors(rq);
3084         if (cfqq->last_request_pos) {
3085                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3086                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3087                 else
3088                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3089         }
3090
3091         cfqq->seek_history <<= 1;
3092         if (blk_queue_nonrot(cfqd->queue))
3093                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3094         else
3095                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3096 }
3097
3098 /*
3099  * Disable idle window if the process thinks too long or seeks so much that
3100  * it doesn't matter
3101  */
3102 static void
3103 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3104                        struct cfq_io_context *cic)
3105 {
3106         int old_idle, enable_idle;
3107
3108         /*
3109          * Don't idle for async or idle io prio class
3110          */
3111         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3112                 return;
3113
3114         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3115
3116         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3117                 cfq_mark_cfqq_deep(cfqq);
3118
3119         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3120                 enable_idle = 0;
3121         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3122             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3123                 enable_idle = 0;
3124         else if (sample_valid(cic->ttime_samples)) {
3125                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3126                         enable_idle = 0;
3127                 else
3128                         enable_idle = 1;
3129         }
3130
3131         if (old_idle != enable_idle) {
3132                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3133                 if (enable_idle)
3134                         cfq_mark_cfqq_idle_window(cfqq);
3135                 else
3136                         cfq_clear_cfqq_idle_window(cfqq);
3137         }
3138 }
3139
3140 /*
3141  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3142  * no or if we aren't sure, a 1 will cause a preempt.
3143  */
3144 static bool
3145 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3146                    struct request *rq)
3147 {
3148         struct cfq_queue *cfqq;
3149
3150         cfqq = cfqd->active_queue;
3151         if (!cfqq)
3152                 return false;
3153
3154         if (cfq_class_idle(new_cfqq))
3155                 return false;
3156
3157         if (cfq_class_idle(cfqq))
3158                 return true;
3159
3160         /*
3161          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3162          */
3163         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3164                 return false;
3165
3166         /*
3167          * if the new request is sync, but the currently running queue is
3168          * not, let the sync request have priority.
3169          */
3170         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3171                 return true;
3172
3173         if (new_cfqq->cfqg != cfqq->cfqg)
3174                 return false;
3175
3176         if (cfq_slice_used(cfqq))
3177                 return true;
3178
3179         /* Allow preemption only if we are idling on sync-noidle tree */
3180         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3181             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3182             new_cfqq->service_tree->count == 2 &&
3183             RB_EMPTY_ROOT(&cfqq->sort_list))
3184                 return true;
3185
3186         /*
3187          * So both queues are sync. Let the new request get disk time if
3188          * it's a metadata request and the current queue is doing regular IO.
3189          */
3190         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3191                 return true;
3192
3193         /*
3194          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3195          */
3196         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3197                 return true;
3198
3199         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3200                 return false;
3201
3202         /*
3203          * if this request is as-good as one we would expect from the
3204          * current cfqq, let it preempt
3205          */
3206         if (cfq_rq_close(cfqd, cfqq, rq))
3207                 return true;
3208
3209         return false;
3210 }
3211
3212 /*
3213  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3214  * let it have half of its nominal slice.
3215  */
3216 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3217 {
3218         cfq_log_cfqq(cfqd, cfqq, "preempt");
3219         cfq_slice_expired(cfqd, 1);
3220
3221         /*
3222          * Put the new queue at the front of the of the current list,
3223          * so we know that it will be selected next.
3224          */
3225         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3226
3227         cfq_service_tree_add(cfqd, cfqq, 1);
3228
3229         cfqq->slice_end = 0;
3230         cfq_mark_cfqq_slice_new(cfqq);
3231 }
3232
3233 /*
3234  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3235  * something we should do about it
3236  */
3237 static void
3238 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3239                 struct request *rq)
3240 {
3241         struct cfq_io_context *cic = RQ_CIC(rq);
3242
3243         cfqd->rq_queued++;
3244         if (rq->cmd_flags & REQ_META)
3245                 cfqq->meta_pending++;
3246
3247         cfq_update_io_thinktime(cfqd, cic);
3248         cfq_update_io_seektime(cfqd, cfqq, rq);
3249         cfq_update_idle_window(cfqd, cfqq, cic);
3250
3251         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3252
3253         if (cfqq == cfqd->active_queue) {
3254                 /*
3255                  * Remember that we saw a request from this process, but
3256                  * don't start queuing just yet. Otherwise we risk seeing lots
3257                  * of tiny requests, because we disrupt the normal plugging
3258                  * and merging. If the request is already larger than a single
3259                  * page, let it rip immediately. For that case we assume that
3260                  * merging is already done. Ditto for a busy system that
3261                  * has other work pending, don't risk delaying until the
3262                  * idle timer unplug to continue working.
3263                  */
3264                 if (cfq_cfqq_wait_request(cfqq)) {
3265                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3266                             cfqd->busy_queues > 1) {
3267                                 cfq_del_timer(cfqd, cfqq);
3268                                 cfq_clear_cfqq_wait_request(cfqq);
3269                                 __blk_run_queue(cfqd->queue);
3270                         } else {
3271                                 cfq_blkiocg_update_idle_time_stats(
3272                                                 &cfqq->cfqg->blkg);
3273                                 cfq_mark_cfqq_must_dispatch(cfqq);
3274                         }
3275                 }
3276         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3277                 /*
3278                  * not the active queue - expire current slice if it is
3279                  * idle and has expired it's mean thinktime or this new queue
3280                  * has some old slice time left and is of higher priority or
3281                  * this new queue is RT and the current one is BE
3282                  */
3283                 cfq_preempt_queue(cfqd, cfqq);
3284                 __blk_run_queue(cfqd->queue);
3285         }
3286 }
3287
3288 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3289 {
3290         struct cfq_data *cfqd = q->elevator->elevator_data;
3291         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3292
3293         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3294         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3295
3296         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3297         list_add_tail(&rq->queuelist, &cfqq->fifo);
3298         cfq_add_rq_rb(rq);
3299         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3300                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3301                         rq_is_sync(rq));
3302         cfq_rq_enqueued(cfqd, cfqq, rq);
3303 }
3304
3305 /*
3306  * Update hw_tag based on peak queue depth over 50 samples under
3307  * sufficient load.
3308  */
3309 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3310 {
3311         struct cfq_queue *cfqq = cfqd->active_queue;
3312
3313         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3314                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3315
3316         if (cfqd->hw_tag == 1)
3317                 return;
3318
3319         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3320             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3321                 return;
3322
3323         /*
3324          * If active queue hasn't enough requests and can idle, cfq might not
3325          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3326          * case
3327          */
3328         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3329             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3330             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3331                 return;
3332
3333         if (cfqd->hw_tag_samples++ < 50)
3334                 return;
3335
3336         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3337                 cfqd->hw_tag = 1;
3338         else
3339                 cfqd->hw_tag = 0;
3340 }
3341
3342 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3343 {
3344         struct cfq_io_context *cic = cfqd->active_cic;
3345
3346         /* If there are other queues in the group, don't wait */
3347         if (cfqq->cfqg->nr_cfqq > 1)
3348                 return false;
3349
3350         if (cfq_slice_used(cfqq))
3351                 return true;
3352
3353         /* if slice left is less than think time, wait busy */
3354         if (cic && sample_valid(cic->ttime_samples)
3355             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3356                 return true;
3357
3358         /*
3359          * If think times is less than a jiffy than ttime_mean=0 and above
3360          * will not be true. It might happen that slice has not expired yet
3361          * but will expire soon (4-5 ns) during select_queue(). To cover the
3362          * case where think time is less than a jiffy, mark the queue wait
3363          * busy if only 1 jiffy is left in the slice.
3364          */
3365         if (cfqq->slice_end - jiffies == 1)
3366                 return true;
3367
3368         return false;
3369 }
3370
3371 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3372 {
3373         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3374         struct cfq_data *cfqd = cfqq->cfqd;
3375         const int sync = rq_is_sync(rq);
3376         unsigned long now;
3377
3378         now = jiffies;
3379         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3380                      !!(rq->cmd_flags & REQ_NOIDLE));
3381
3382         cfq_update_hw_tag(cfqd);
3383
3384         WARN_ON(!cfqd->rq_in_driver);
3385         WARN_ON(!cfqq->dispatched);
3386         cfqd->rq_in_driver--;
3387         cfqq->dispatched--;
3388         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3389                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3390                         rq_data_dir(rq), rq_is_sync(rq));
3391
3392         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3393
3394         if (sync) {
3395                 RQ_CIC(rq)->last_end_request = now;
3396                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3397                         cfqd->last_delayed_sync = now;
3398         }
3399
3400         /*
3401          * If this is the active queue, check if it needs to be expired,
3402          * or if we want to idle in case it has no pending requests.
3403          */
3404         if (cfqd->active_queue == cfqq) {
3405                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3406
3407                 if (cfq_cfqq_slice_new(cfqq)) {
3408                         cfq_set_prio_slice(cfqd, cfqq);
3409                         cfq_clear_cfqq_slice_new(cfqq);
3410                 }
3411
3412                 /*
3413                  * Should we wait for next request to come in before we expire
3414                  * the queue.
3415                  */
3416                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3417                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3418                         cfq_mark_cfqq_wait_busy(cfqq);
3419                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3420                 }
3421
3422                 /*
3423                  * Idling is not enabled on:
3424                  * - expired queues
3425                  * - idle-priority queues
3426                  * - async queues
3427                  * - queues with still some requests queued
3428                  * - when there is a close cooperator
3429                  */
3430                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3431                         cfq_slice_expired(cfqd, 1);
3432                 else if (sync && cfqq_empty &&
3433                          !cfq_close_cooperator(cfqd, cfqq)) {
3434                         cfq_arm_slice_timer(cfqd);
3435                 }
3436         }
3437
3438         if (!cfqd->rq_in_driver)
3439                 cfq_schedule_dispatch(cfqd);
3440 }
3441
3442 /*
3443  * we temporarily boost lower priority queues if they are holding fs exclusive
3444  * resources. they are boosted to normal prio (CLASS_BE/4)
3445  */
3446 static void cfq_prio_boost(struct cfq_queue *cfqq)
3447 {
3448         if (has_fs_excl()) {
3449                 /*
3450                  * boost idle prio on transactions that would lock out other
3451                  * users of the filesystem
3452                  */
3453                 if (cfq_class_idle(cfqq))
3454                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3455                 if (cfqq->ioprio > IOPRIO_NORM)
3456                         cfqq->ioprio = IOPRIO_NORM;
3457         } else {
3458                 /*
3459                  * unboost the queue (if needed)
3460                  */
3461                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3462                 cfqq->ioprio = cfqq->org_ioprio;
3463         }
3464 }
3465
3466 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3467 {
3468         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3469                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3470                 return ELV_MQUEUE_MUST;
3471         }
3472
3473         return ELV_MQUEUE_MAY;
3474 }
3475
3476 static int cfq_may_queue(struct request_queue *q, int rw)
3477 {
3478         struct cfq_data *cfqd = q->elevator->elevator_data;
3479         struct task_struct *tsk = current;
3480         struct cfq_io_context *cic;
3481         struct cfq_queue *cfqq;
3482
3483         /*
3484          * don't force setup of a queue from here, as a call to may_queue
3485          * does not necessarily imply that a request actually will be queued.
3486          * so just lookup a possibly existing queue, or return 'may queue'
3487          * if that fails
3488          */
3489         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3490         if (!cic)
3491                 return ELV_MQUEUE_MAY;
3492
3493         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3494         if (cfqq) {
3495                 cfq_init_prio_data(cfqq, cic->ioc);
3496                 cfq_prio_boost(cfqq);
3497
3498                 return __cfq_may_queue(cfqq);
3499         }
3500
3501         return ELV_MQUEUE_MAY;
3502 }
3503
3504 /*
3505  * queue lock held here
3506  */
3507 static void cfq_put_request(struct request *rq)
3508 {
3509         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3510
3511         if (cfqq) {
3512                 const int rw = rq_data_dir(rq);
3513
3514                 BUG_ON(!cfqq->allocated[rw]);
3515                 cfqq->allocated[rw]--;
3516
3517                 put_io_context(RQ_CIC(rq)->ioc);
3518
3519                 rq->elevator_private = NULL;
3520                 rq->elevator_private2 = NULL;
3521
3522                 /* Put down rq reference on cfqg */
3523                 cfq_put_cfqg(RQ_CFQG(rq));
3524                 rq->elevator_private3 = NULL;
3525
3526                 cfq_put_queue(cfqq);
3527         }
3528 }
3529
3530 static struct cfq_queue *
3531 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3532                 struct cfq_queue *cfqq)
3533 {
3534         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3535         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3536         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3537         cfq_put_queue(cfqq);
3538         return cic_to_cfqq(cic, 1);
3539 }
3540
3541 /*
3542  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3543  * was the last process referring to said cfqq.
3544  */
3545 static struct cfq_queue *
3546 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3547 {
3548         if (cfqq_process_refs(cfqq) == 1) {
3549                 cfqq->pid = current->pid;
3550                 cfq_clear_cfqq_coop(cfqq);
3551                 cfq_clear_cfqq_split_coop(cfqq);
3552                 return cfqq;
3553         }
3554
3555         cic_set_cfqq(cic, NULL, 1);
3556
3557         cfq_put_cooperator(cfqq);
3558
3559         cfq_put_queue(cfqq);
3560         return NULL;
3561 }
3562 /*
3563  * Allocate cfq data structures associated with this request.
3564  */
3565 static int
3566 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3567 {
3568         struct cfq_data *cfqd = q->elevator->elevator_data;
3569         struct cfq_io_context *cic;
3570         const int rw = rq_data_dir(rq);
3571         const bool is_sync = rq_is_sync(rq);
3572         struct cfq_queue *cfqq;
3573         unsigned long flags;
3574
3575         might_sleep_if(gfp_mask & __GFP_WAIT);
3576
3577         cic = cfq_get_io_context(cfqd, gfp_mask);
3578
3579         spin_lock_irqsave(q->queue_lock, flags);
3580
3581         if (!cic)
3582                 goto queue_fail;
3583
3584 new_queue:
3585         cfqq = cic_to_cfqq(cic, is_sync);
3586         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3587                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3588                 cic_set_cfqq(cic, cfqq, is_sync);
3589         } else {
3590                 /*
3591                  * If the queue was seeky for too long, break it apart.
3592                  */
3593                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3594                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3595                         cfqq = split_cfqq(cic, cfqq);
3596                         if (!cfqq)
3597                                 goto new_queue;
3598                 }
3599
3600                 /*
3601                  * Check to see if this queue is scheduled to merge with
3602                  * another, closely cooperating queue.  The merging of
3603                  * queues happens here as it must be done in process context.
3604                  * The reference on new_cfqq was taken in merge_cfqqs.
3605                  */
3606                 if (cfqq->new_cfqq)
3607                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3608         }
3609
3610         cfqq->allocated[rw]++;
3611         atomic_inc(&cfqq->ref);
3612
3613         spin_unlock_irqrestore(q->queue_lock, flags);
3614
3615         rq->elevator_private = cic;
3616         rq->elevator_private2 = cfqq;
3617         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3618         return 0;
3619
3620 queue_fail:
3621         if (cic)
3622                 put_io_context(cic->ioc);
3623
3624         cfq_schedule_dispatch(cfqd);
3625         spin_unlock_irqrestore(q->queue_lock, flags);
3626         cfq_log(cfqd, "set_request fail");
3627         return 1;
3628 }
3629
3630 static void cfq_kick_queue(struct work_struct *work)
3631 {
3632         struct cfq_data *cfqd =
3633                 container_of(work, struct cfq_data, unplug_work);
3634         struct request_queue *q = cfqd->queue;
3635
3636         spin_lock_irq(q->queue_lock);
3637         __blk_run_queue(cfqd->queue);
3638         spin_unlock_irq(q->queue_lock);
3639 }
3640
3641 /*
3642  * Timer running if the active_queue is currently idling inside its time slice
3643  */
3644 static void cfq_idle_slice_timer(unsigned long data)
3645 {
3646         struct cfq_data *cfqd = (struct cfq_data *) data;
3647         struct cfq_queue *cfqq;
3648         unsigned long flags;
3649         int timed_out = 1;
3650
3651         cfq_log(cfqd, "idle timer fired");
3652
3653         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3654
3655         cfqq = cfqd->active_queue;
3656         if (cfqq) {
3657                 timed_out = 0;
3658
3659                 /*
3660                  * We saw a request before the queue expired, let it through
3661                  */
3662                 if (cfq_cfqq_must_dispatch(cfqq))
3663                         goto out_kick;
3664
3665                 /*
3666                  * expired
3667                  */
3668                 if (cfq_slice_used(cfqq))
3669                         goto expire;
3670
3671                 /*
3672                  * only expire and reinvoke request handler, if there are
3673                  * other queues with pending requests
3674                  */
3675                 if (!cfqd->busy_queues)
3676                         goto out_cont;
3677
3678                 /*
3679                  * not expired and it has a request pending, let it dispatch
3680                  */
3681                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3682                         goto out_kick;
3683
3684                 /*
3685                  * Queue depth flag is reset only when the idle didn't succeed
3686                  */
3687                 cfq_clear_cfqq_deep(cfqq);
3688         }
3689 expire:
3690         cfq_slice_expired(cfqd, timed_out);
3691 out_kick:
3692         cfq_schedule_dispatch(cfqd);
3693 out_cont:
3694         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3695 }
3696
3697 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3698 {
3699         del_timer_sync(&cfqd->idle_slice_timer);
3700         cancel_work_sync(&cfqd->unplug_work);
3701 }
3702
3703 static void cfq_put_async_queues(struct cfq_data *cfqd)
3704 {
3705         int i;
3706
3707         for (i = 0; i < IOPRIO_BE_NR; i++) {
3708                 if (cfqd->async_cfqq[0][i])
3709                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3710                 if (cfqd->async_cfqq[1][i])
3711                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3712         }
3713
3714         if (cfqd->async_idle_cfqq)
3715                 cfq_put_queue(cfqd->async_idle_cfqq);
3716 }
3717
3718 static void cfq_cfqd_free(struct rcu_head *head)
3719 {
3720         kfree(container_of(head, struct cfq_data, rcu));
3721 }
3722
3723 static void cfq_exit_queue(struct elevator_queue *e)
3724 {
3725         struct cfq_data *cfqd = e->elevator_data;
3726         struct request_queue *q = cfqd->queue;
3727
3728         cfq_shutdown_timer_wq(cfqd);
3729
3730         spin_lock_irq(q->queue_lock);
3731
3732         if (cfqd->active_queue)
3733                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3734
3735         while (!list_empty(&cfqd->cic_list)) {
3736                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3737                                                         struct cfq_io_context,
3738                                                         queue_list);
3739
3740                 __cfq_exit_single_io_context(cfqd, cic);
3741         }
3742
3743         cfq_put_async_queues(cfqd);
3744         cfq_release_cfq_groups(cfqd);
3745         cfq_blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3746
3747         spin_unlock_irq(q->queue_lock);
3748
3749         cfq_shutdown_timer_wq(cfqd);
3750
3751         spin_lock(&cic_index_lock);
3752         ida_remove(&cic_index_ida, cfqd->cic_index);
3753         spin_unlock(&cic_index_lock);
3754
3755         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3756         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3757 }
3758
3759 static int cfq_alloc_cic_index(void)
3760 {
3761         int index, error;
3762
3763         do {
3764                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3765                         return -ENOMEM;
3766
3767                 spin_lock(&cic_index_lock);
3768                 error = ida_get_new(&cic_index_ida, &index);
3769                 spin_unlock(&cic_index_lock);
3770                 if (error && error != -EAGAIN)
3771                         return error;
3772         } while (error);
3773
3774         return index;
3775 }
3776
3777 static void *cfq_init_queue(struct request_queue *q)
3778 {
3779         struct cfq_data *cfqd;
3780         int i, j;
3781         struct cfq_group *cfqg;
3782         struct cfq_rb_root *st;
3783
3784         i = cfq_alloc_cic_index();
3785         if (i < 0)
3786                 return NULL;
3787
3788         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3789         if (!cfqd)
3790                 return NULL;
3791
3792         cfqd->cic_index = i;
3793
3794         /* Init root service tree */
3795         cfqd->grp_service_tree = CFQ_RB_ROOT;
3796
3797         /* Init root group */
3798         cfqg = &cfqd->root_group;
3799         for_each_cfqg_st(cfqg, i, j, st)
3800                 *st = CFQ_RB_ROOT;
3801         RB_CLEAR_NODE(&cfqg->rb_node);
3802
3803         /* Give preference to root group over other groups */
3804         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3805
3806 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3807         /*
3808          * Take a reference to root group which we never drop. This is just
3809          * to make sure that cfq_put_cfqg() does not try to kfree root group
3810          */
3811         atomic_set(&cfqg->ref, 1);
3812         rcu_read_lock();
3813         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
3814                                         (void *)cfqd, 0);
3815         rcu_read_unlock();
3816 #endif
3817         /*
3818          * Not strictly needed (since RB_ROOT just clears the node and we
3819          * zeroed cfqd on alloc), but better be safe in case someone decides
3820          * to add magic to the rb code
3821          */
3822         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3823                 cfqd->prio_trees[i] = RB_ROOT;
3824
3825         /*
3826          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3827          * Grab a permanent reference to it, so that the normal code flow
3828          * will not attempt to free it.
3829          */
3830         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3831         atomic_inc(&cfqd->oom_cfqq.ref);
3832         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3833
3834         INIT_LIST_HEAD(&cfqd->cic_list);
3835
3836         cfqd->queue = q;
3837
3838         init_timer(&cfqd->idle_slice_timer);
3839         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3840         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3841
3842         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3843
3844         cfqd->cfq_quantum = cfq_quantum;
3845         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3846         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3847         cfqd->cfq_back_max = cfq_back_max;
3848         cfqd->cfq_back_penalty = cfq_back_penalty;
3849         cfqd->cfq_slice[0] = cfq_slice_async;
3850         cfqd->cfq_slice[1] = cfq_slice_sync;
3851         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3852         cfqd->cfq_slice_idle = cfq_slice_idle;
3853         cfqd->cfq_latency = 1;
3854         cfqd->cfq_group_isolation = 0;
3855         cfqd->hw_tag = -1;
3856         /*
3857          * we optimistically start assuming sync ops weren't delayed in last
3858          * second, in order to have larger depth for async operations.
3859          */
3860         cfqd->last_delayed_sync = jiffies - HZ;
3861         return cfqd;
3862 }
3863
3864 static void cfq_slab_kill(void)
3865 {
3866         /*
3867          * Caller already ensured that pending RCU callbacks are completed,
3868          * so we should have no busy allocations at this point.
3869          */
3870         if (cfq_pool)
3871                 kmem_cache_destroy(cfq_pool);
3872         if (cfq_ioc_pool)
3873                 kmem_cache_destroy(cfq_ioc_pool);
3874 }
3875
3876 static int __init cfq_slab_setup(void)
3877 {
3878         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3879         if (!cfq_pool)
3880                 goto fail;
3881
3882         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3883         if (!cfq_ioc_pool)
3884                 goto fail;
3885
3886         return 0;
3887 fail:
3888         cfq_slab_kill();
3889         return -ENOMEM;
3890 }
3891
3892 /*
3893  * sysfs parts below -->
3894  */
3895 static ssize_t
3896 cfq_var_show(unsigned int var, char *page)
3897 {
3898         return sprintf(page, "%d\n", var);
3899 }
3900
3901 static ssize_t
3902 cfq_var_store(unsigned int *var, const char *page, size_t count)
3903 {
3904         char *p = (char *) page;
3905
3906         *var = simple_strtoul(p, &p, 10);
3907         return count;
3908 }
3909
3910 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3911 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3912 {                                                                       \
3913         struct cfq_data *cfqd = e->elevator_data;                       \
3914         unsigned int __data = __VAR;                                    \
3915         if (__CONV)                                                     \
3916                 __data = jiffies_to_msecs(__data);                      \
3917         return cfq_var_show(__data, (page));                            \
3918 }
3919 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3920 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3921 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3922 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3923 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3924 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3925 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3926 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3927 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3928 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3929 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3930 #undef SHOW_FUNCTION
3931
3932 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3933 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3934 {                                                                       \
3935         struct cfq_data *cfqd = e->elevator_data;                       \
3936         unsigned int __data;                                            \
3937         int ret = cfq_var_store(&__data, (page), count);                \
3938         if (__data < (MIN))                                             \
3939                 __data = (MIN);                                         \
3940         else if (__data > (MAX))                                        \
3941                 __data = (MAX);                                         \
3942         if (__CONV)                                                     \
3943                 *(__PTR) = msecs_to_jiffies(__data);                    \
3944         else                                                            \
3945                 *(__PTR) = __data;                                      \
3946         return ret;                                                     \
3947 }
3948 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3949 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3950                 UINT_MAX, 1);
3951 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3952                 UINT_MAX, 1);
3953 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3954 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3955                 UINT_MAX, 0);
3956 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3957 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3958 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3959 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3960                 UINT_MAX, 0);
3961 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3962 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3963 #undef STORE_FUNCTION
3964
3965 #define CFQ_ATTR(name) \
3966         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3967
3968 static struct elv_fs_entry cfq_attrs[] = {
3969         CFQ_ATTR(quantum),
3970         CFQ_ATTR(fifo_expire_sync),
3971         CFQ_ATTR(fifo_expire_async),
3972         CFQ_ATTR(back_seek_max),
3973         CFQ_ATTR(back_seek_penalty),
3974         CFQ_ATTR(slice_sync),
3975         CFQ_ATTR(slice_async),
3976         CFQ_ATTR(slice_async_rq),
3977         CFQ_ATTR(slice_idle),
3978         CFQ_ATTR(low_latency),
3979         CFQ_ATTR(group_isolation),
3980         __ATTR_NULL
3981 };
3982
3983 static struct elevator_type iosched_cfq = {
3984         .ops = {
3985                 .elevator_merge_fn =            cfq_merge,
3986                 .elevator_merged_fn =           cfq_merged_request,
3987                 .elevator_merge_req_fn =        cfq_merged_requests,
3988                 .elevator_allow_merge_fn =      cfq_allow_merge,
3989                 .elevator_bio_merged_fn =       cfq_bio_merged,
3990                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3991                 .elevator_add_req_fn =          cfq_insert_request,
3992                 .elevator_activate_req_fn =     cfq_activate_request,
3993                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3994                 .elevator_queue_empty_fn =      cfq_queue_empty,
3995                 .elevator_completed_req_fn =    cfq_completed_request,
3996                 .elevator_former_req_fn =       elv_rb_former_request,
3997                 .elevator_latter_req_fn =       elv_rb_latter_request,
3998                 .elevator_set_req_fn =          cfq_set_request,
3999                 .elevator_put_req_fn =          cfq_put_request,
4000                 .elevator_may_queue_fn =        cfq_may_queue,
4001                 .elevator_init_fn =             cfq_init_queue,
4002                 .elevator_exit_fn =             cfq_exit_queue,
4003                 .trim =                         cfq_free_io_context,
4004         },
4005         .elevator_attrs =       cfq_attrs,
4006         .elevator_name =        "cfq",
4007         .elevator_owner =       THIS_MODULE,
4008 };
4009
4010 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4011 static struct blkio_policy_type blkio_policy_cfq = {
4012         .ops = {
4013                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4014                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4015         },
4016         .plid = BLKIO_POLICY_PROP,
4017 };
4018 #else
4019 static struct blkio_policy_type blkio_policy_cfq;
4020 #endif
4021
4022 static int __init cfq_init(void)
4023 {
4024         /*
4025          * could be 0 on HZ < 1000 setups
4026          */
4027         if (!cfq_slice_async)
4028                 cfq_slice_async = 1;
4029         if (!cfq_slice_idle)
4030                 cfq_slice_idle = 1;
4031
4032         if (cfq_slab_setup())
4033                 return -ENOMEM;
4034
4035         elv_register(&iosched_cfq);
4036         blkio_policy_register(&blkio_policy_cfq);
4037
4038         return 0;
4039 }
4040
4041 static void __exit cfq_exit(void)
4042 {
4043         DECLARE_COMPLETION_ONSTACK(all_gone);
4044         blkio_policy_unregister(&blkio_policy_cfq);
4045         elv_unregister(&iosched_cfq);
4046         ioc_gone = &all_gone;
4047         /* ioc_gone's update must be visible before reading ioc_count */
4048         smp_wmb();
4049
4050         /*
4051          * this also protects us from entering cfq_slab_kill() with
4052          * pending RCU callbacks
4053          */
4054         if (elv_ioc_count_read(cfq_ioc_count))
4055                 wait_for_completion(&all_gone);
4056         ida_destroy(&cic_index_ida);
4057         cfq_slab_kill();
4058 }
4059
4060 module_init(cfq_init);
4061 module_exit(cfq_exit);
4062
4063 MODULE_AUTHOR("Jens Axboe");
4064 MODULE_LICENSE("GPL");
4065 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");