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