CFQ: make two functions static
[linux-3.10.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "cfq.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static int cfq_group_idle = HZ / 125;
34 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
35 static const int cfq_hist_divisor = 4;
36
37 /*
38  * offset from end of service tree
39  */
40 #define CFQ_IDLE_DELAY          (HZ / 5)
41
42 /*
43  * below this threshold, we consider thinktime immediate
44  */
45 #define CFQ_MIN_TT              (2)
46
47 #define CFQ_SLICE_SCALE         (5)
48 #define CFQ_HW_QUEUE_MIN        (5)
49 #define CFQ_SERVICE_SHIFT       12
50
51 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
52 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
53 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
54 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
55
56 #define RQ_CIC(rq)              \
57         ((struct cfq_io_context *) (rq)->elevator_private[0])
58 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private[1])
59 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private[2])
60
61 static struct kmem_cache *cfq_pool;
62 static struct kmem_cache *cfq_ioc_pool;
63
64 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
65 static struct completion *ioc_gone;
66 static DEFINE_SPINLOCK(ioc_gone_lock);
67
68 static DEFINE_SPINLOCK(cic_index_lock);
69 static DEFINE_IDA(cic_index_ida);
70
71 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
72 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
73 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
74
75 #define sample_valid(samples)   ((samples) > 80)
76 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
77
78 /*
79  * Most of our rbtree usage is for sorting with min extraction, so
80  * if we cache the leftmost node we don't have to walk down the tree
81  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
82  * move this into the elevator for the rq sorting as well.
83  */
84 struct cfq_rb_root {
85         struct rb_root rb;
86         struct rb_node *left;
87         unsigned count;
88         unsigned total_weight;
89         u64 min_vdisktime;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         int ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         /* Number of sectors dispatched from queue in single dispatch round */
150         unsigned long nr_sectors;
151 };
152
153 /*
154  * First index in the service_trees.
155  * IDLE is handled separately, so it has negative index
156  */
157 enum wl_prio_t {
158         BE_WORKLOAD = 0,
159         RT_WORKLOAD = 1,
160         IDLE_WORKLOAD = 2,
161         CFQ_PRIO_NR,
162 };
163
164 /*
165  * Second index in the service_trees.
166  */
167 enum wl_type_t {
168         ASYNC_WORKLOAD = 0,
169         SYNC_NOIDLE_WORKLOAD = 1,
170         SYNC_WORKLOAD = 2
171 };
172
173 /* This is per cgroup per device grouping structure */
174 struct cfq_group {
175         /* group service_tree member */
176         struct rb_node rb_node;
177
178         /* group service_tree key */
179         u64 vdisktime;
180         unsigned int weight;
181         unsigned int new_weight;
182         bool needs_update;
183
184         /* number of cfqq currently on this group */
185         int nr_cfqq;
186
187         /*
188          * Per group busy queus average. Useful for workload slice calc. We
189          * create the array for each prio class but at run time it is used
190          * only for RT and BE class and slot for IDLE class remains unused.
191          * This is primarily done to avoid confusion and a gcc warning.
192          */
193         unsigned int busy_queues_avg[CFQ_PRIO_NR];
194         /*
195          * rr lists of queues with requests. We maintain service trees for
196          * RT and BE classes. These trees are subdivided in subclasses
197          * of SYNC, SYNC_NOIDLE and ASYNC based on workload type. For IDLE
198          * class there is no subclassification and all the cfq queues go on
199          * a single tree service_tree_idle.
200          * Counts are embedded in the cfq_rb_root
201          */
202         struct cfq_rb_root service_trees[2][3];
203         struct cfq_rb_root service_tree_idle;
204
205         unsigned long saved_workload_slice;
206         enum wl_type_t saved_workload;
207         enum wl_prio_t saved_serving_prio;
208         struct blkio_group blkg;
209 #ifdef CONFIG_CFQ_GROUP_IOSCHED
210         struct hlist_node cfqd_node;
211         int ref;
212 #endif
213         /* number of requests that are on the dispatch list or inside driver */
214         int dispatched;
215 };
216
217 /*
218  * Per block device queue structure
219  */
220 struct cfq_data {
221         struct request_queue *queue;
222         /* Root service tree for cfq_groups */
223         struct cfq_rb_root grp_service_tree;
224         struct cfq_group root_group;
225
226         /*
227          * The priority currently being served
228          */
229         enum wl_prio_t serving_prio;
230         enum wl_type_t serving_type;
231         unsigned long workload_expires;
232         struct cfq_group *serving_group;
233
234         /*
235          * Each priority tree is sorted by next_request position.  These
236          * trees are used when determining if two or more queues are
237          * interleaving requests (see cfq_close_cooperator).
238          */
239         struct rb_root prio_trees[CFQ_PRIO_LISTS];
240
241         unsigned int busy_queues;
242         unsigned int busy_sync_queues;
243
244         int rq_in_driver;
245         int rq_in_flight[2];
246
247         /*
248          * queue-depth detection
249          */
250         int rq_queued;
251         int hw_tag;
252         /*
253          * hw_tag can be
254          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
255          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
256          *  0 => no NCQ
257          */
258         int hw_tag_est_depth;
259         unsigned int hw_tag_samples;
260
261         /*
262          * idle window management
263          */
264         struct timer_list idle_slice_timer;
265         struct work_struct unplug_work;
266
267         struct cfq_queue *active_queue;
268         struct cfq_io_context *active_cic;
269
270         /*
271          * async queue for each priority case
272          */
273         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
274         struct cfq_queue *async_idle_cfqq;
275
276         sector_t last_position;
277
278         /*
279          * tunables, see top of file
280          */
281         unsigned int cfq_quantum;
282         unsigned int cfq_fifo_expire[2];
283         unsigned int cfq_back_penalty;
284         unsigned int cfq_back_max;
285         unsigned int cfq_slice[2];
286         unsigned int cfq_slice_async_rq;
287         unsigned int cfq_slice_idle;
288         unsigned int cfq_group_idle;
289         unsigned int cfq_latency;
290
291         unsigned int cic_index;
292         struct list_head cic_list;
293
294         /*
295          * Fallback dummy cfqq for extreme OOM conditions
296          */
297         struct cfq_queue oom_cfqq;
298
299         unsigned long last_delayed_sync;
300
301         /* List of cfq groups being managed on this device*/
302         struct hlist_head cfqg_list;
303
304         /* Number of groups which are on blkcg->blkg_list */
305         unsigned int nr_blkcg_linked_grps;
306 };
307
308 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
309
310 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
311                                             enum wl_prio_t prio,
312                                             enum wl_type_t type)
313 {
314         if (!cfqg)
315                 return NULL;
316
317         if (prio == IDLE_WORKLOAD)
318                 return &cfqg->service_tree_idle;
319
320         return &cfqg->service_trees[prio][type];
321 }
322
323 enum cfqq_state_flags {
324         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
325         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
326         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
327         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
328         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
329         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
330         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
331         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
332         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
333         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
334         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
335         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
336         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
337 };
338
339 #define CFQ_CFQQ_FNS(name)                                              \
340 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
341 {                                                                       \
342         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
343 }                                                                       \
344 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
345 {                                                                       \
346         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
347 }                                                                       \
348 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
349 {                                                                       \
350         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
351 }
352
353 CFQ_CFQQ_FNS(on_rr);
354 CFQ_CFQQ_FNS(wait_request);
355 CFQ_CFQQ_FNS(must_dispatch);
356 CFQ_CFQQ_FNS(must_alloc_slice);
357 CFQ_CFQQ_FNS(fifo_expire);
358 CFQ_CFQQ_FNS(idle_window);
359 CFQ_CFQQ_FNS(prio_changed);
360 CFQ_CFQQ_FNS(slice_new);
361 CFQ_CFQQ_FNS(sync);
362 CFQ_CFQQ_FNS(coop);
363 CFQ_CFQQ_FNS(split_coop);
364 CFQ_CFQQ_FNS(deep);
365 CFQ_CFQQ_FNS(wait_busy);
366 #undef CFQ_CFQQ_FNS
367
368 #ifdef CONFIG_CFQ_GROUP_IOSCHED
369 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
370         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
371                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
372                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
373
374 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
375         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
376                                 blkg_path(&(cfqg)->blkg), ##args);      \
377
378 #else
379 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
380         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
381 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
382 #endif
383 #define cfq_log(cfqd, fmt, args...)     \
384         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
385
386 /* Traverses through cfq group service trees */
387 #define for_each_cfqg_st(cfqg, i, j, st) \
388         for (i = 0; i <= IDLE_WORKLOAD; i++) \
389                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
390                         : &cfqg->service_tree_idle; \
391                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
392                         (i == IDLE_WORKLOAD && j == 0); \
393                         j++, st = i < IDLE_WORKLOAD ? \
394                         &cfqg->service_trees[i][j]: NULL) \
395
396
397 static inline bool iops_mode(struct cfq_data *cfqd)
398 {
399         /*
400          * If we are not idling on queues and it is a NCQ drive, parallel
401          * execution of requests is on and measuring time is not possible
402          * in most of the cases until and unless we drive shallower queue
403          * depths and that becomes a performance bottleneck. In such cases
404          * switch to start providing fairness in terms of number of IOs.
405          */
406         if (!cfqd->cfq_slice_idle && cfqd->hw_tag)
407                 return true;
408         else
409                 return false;
410 }
411
412 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
413 {
414         if (cfq_class_idle(cfqq))
415                 return IDLE_WORKLOAD;
416         if (cfq_class_rt(cfqq))
417                 return RT_WORKLOAD;
418         return BE_WORKLOAD;
419 }
420
421
422 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
423 {
424         if (!cfq_cfqq_sync(cfqq))
425                 return ASYNC_WORKLOAD;
426         if (!cfq_cfqq_idle_window(cfqq))
427                 return SYNC_NOIDLE_WORKLOAD;
428         return SYNC_WORKLOAD;
429 }
430
431 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
432                                         struct cfq_data *cfqd,
433                                         struct cfq_group *cfqg)
434 {
435         if (wl == IDLE_WORKLOAD)
436                 return cfqg->service_tree_idle.count;
437
438         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
439                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
440                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
441 }
442
443 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
444                                         struct cfq_group *cfqg)
445 {
446         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
447                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
448 }
449
450 static void cfq_dispatch_insert(struct request_queue *, struct request *);
451 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
452                                        struct io_context *, gfp_t);
453 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
454                                                 struct io_context *);
455
456 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
457                                             bool is_sync)
458 {
459         return cic->cfqq[is_sync];
460 }
461
462 static inline void cic_set_cfqq(struct cfq_io_context *cic,
463                                 struct cfq_queue *cfqq, bool is_sync)
464 {
465         cic->cfqq[is_sync] = cfqq;
466 }
467
468 #define CIC_DEAD_KEY    1ul
469 #define CIC_DEAD_INDEX_SHIFT    1
470
471 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
472 {
473         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
474 }
475
476 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
477 {
478         struct cfq_data *cfqd = cic->key;
479
480         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
481                 return NULL;
482
483         return cfqd;
484 }
485
486 /*
487  * We regard a request as SYNC, if it's either a read or has the SYNC bit
488  * set (in which case it could also be direct WRITE).
489  */
490 static inline bool cfq_bio_sync(struct bio *bio)
491 {
492         return bio_data_dir(bio) == READ || (bio->bi_rw & REQ_SYNC);
493 }
494
495 /*
496  * scheduler run of queue, if there are requests pending and no one in the
497  * driver that will restart queueing
498  */
499 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
500 {
501         if (cfqd->busy_queues) {
502                 cfq_log(cfqd, "schedule dispatch");
503                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
504         }
505 }
506
507 /*
508  * Scale schedule slice based on io priority. Use the sync time slice only
509  * if a queue is marked sync and has sync io queued. A sync queue with async
510  * io only, should not get full sync slice length.
511  */
512 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
513                                  unsigned short prio)
514 {
515         const int base_slice = cfqd->cfq_slice[sync];
516
517         WARN_ON(prio >= IOPRIO_BE_NR);
518
519         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
520 }
521
522 static inline int
523 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
524 {
525         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
526 }
527
528 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
529 {
530         u64 d = delta << CFQ_SERVICE_SHIFT;
531
532         d = d * BLKIO_WEIGHT_DEFAULT;
533         do_div(d, cfqg->weight);
534         return d;
535 }
536
537 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
538 {
539         s64 delta = (s64)(vdisktime - min_vdisktime);
540         if (delta > 0)
541                 min_vdisktime = vdisktime;
542
543         return min_vdisktime;
544 }
545
546 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
547 {
548         s64 delta = (s64)(vdisktime - min_vdisktime);
549         if (delta < 0)
550                 min_vdisktime = vdisktime;
551
552         return min_vdisktime;
553 }
554
555 static void update_min_vdisktime(struct cfq_rb_root *st)
556 {
557         struct cfq_group *cfqg;
558
559         if (st->left) {
560                 cfqg = rb_entry_cfqg(st->left);
561                 st->min_vdisktime = max_vdisktime(st->min_vdisktime,
562                                                   cfqg->vdisktime);
563         }
564 }
565
566 /*
567  * get averaged number of queues of RT/BE priority.
568  * average is updated, with a formula that gives more weight to higher numbers,
569  * to quickly follows sudden increases and decrease slowly
570  */
571
572 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
573                                         struct cfq_group *cfqg, bool rt)
574 {
575         unsigned min_q, max_q;
576         unsigned mult  = cfq_hist_divisor - 1;
577         unsigned round = cfq_hist_divisor / 2;
578         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
579
580         min_q = min(cfqg->busy_queues_avg[rt], busy);
581         max_q = max(cfqg->busy_queues_avg[rt], busy);
582         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
583                 cfq_hist_divisor;
584         return cfqg->busy_queues_avg[rt];
585 }
586
587 static inline unsigned
588 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
589 {
590         struct cfq_rb_root *st = &cfqd->grp_service_tree;
591
592         return cfq_target_latency * cfqg->weight / st->total_weight;
593 }
594
595 static inline unsigned
596 cfq_scaled_cfqq_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
597 {
598         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
599         if (cfqd->cfq_latency) {
600                 /*
601                  * interested queues (we consider only the ones with the same
602                  * priority class in the cfq group)
603                  */
604                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
605                                                 cfq_class_rt(cfqq));
606                 unsigned sync_slice = cfqd->cfq_slice[1];
607                 unsigned expect_latency = sync_slice * iq;
608                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
609
610                 if (expect_latency > group_slice) {
611                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
612                         /* scale low_slice according to IO priority
613                          * and sync vs async */
614                         unsigned low_slice =
615                                 min(slice, base_low_slice * slice / sync_slice);
616                         /* the adapted slice value is scaled to fit all iqs
617                          * into the target latency */
618                         slice = max(slice * group_slice / expect_latency,
619                                     low_slice);
620                 }
621         }
622         return slice;
623 }
624
625 static inline void
626 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
627 {
628         unsigned slice = cfq_scaled_cfqq_slice(cfqd, cfqq);
629
630         cfqq->slice_start = jiffies;
631         cfqq->slice_end = jiffies + slice;
632         cfqq->allocated_slice = slice;
633         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
634 }
635
636 /*
637  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
638  * isn't valid until the first request from the dispatch is activated
639  * and the slice time set.
640  */
641 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
642 {
643         if (cfq_cfqq_slice_new(cfqq))
644                 return false;
645         if (time_before(jiffies, cfqq->slice_end))
646                 return false;
647
648         return true;
649 }
650
651 /*
652  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
653  * We choose the request that is closest to the head right now. Distance
654  * behind the head is penalized and only allowed to a certain extent.
655  */
656 static struct request *
657 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
658 {
659         sector_t s1, s2, d1 = 0, d2 = 0;
660         unsigned long back_max;
661 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
662 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
663         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
664
665         if (rq1 == NULL || rq1 == rq2)
666                 return rq2;
667         if (rq2 == NULL)
668                 return rq1;
669
670         if (rq_is_sync(rq1) != rq_is_sync(rq2))
671                 return rq_is_sync(rq1) ? rq1 : rq2;
672
673         if ((rq1->cmd_flags ^ rq2->cmd_flags) & REQ_META)
674                 return rq1->cmd_flags & REQ_META ? rq1 : rq2;
675
676         s1 = blk_rq_pos(rq1);
677         s2 = blk_rq_pos(rq2);
678
679         /*
680          * by definition, 1KiB is 2 sectors
681          */
682         back_max = cfqd->cfq_back_max * 2;
683
684         /*
685          * Strict one way elevator _except_ in the case where we allow
686          * short backward seeks which are biased as twice the cost of a
687          * similar forward seek.
688          */
689         if (s1 >= last)
690                 d1 = s1 - last;
691         else if (s1 + back_max >= last)
692                 d1 = (last - s1) * cfqd->cfq_back_penalty;
693         else
694                 wrap |= CFQ_RQ1_WRAP;
695
696         if (s2 >= last)
697                 d2 = s2 - last;
698         else if (s2 + back_max >= last)
699                 d2 = (last - s2) * cfqd->cfq_back_penalty;
700         else
701                 wrap |= CFQ_RQ2_WRAP;
702
703         /* Found required data */
704
705         /*
706          * By doing switch() on the bit mask "wrap" we avoid having to
707          * check two variables for all permutations: --> faster!
708          */
709         switch (wrap) {
710         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
711                 if (d1 < d2)
712                         return rq1;
713                 else if (d2 < d1)
714                         return rq2;
715                 else {
716                         if (s1 >= s2)
717                                 return rq1;
718                         else
719                                 return rq2;
720                 }
721
722         case CFQ_RQ2_WRAP:
723                 return rq1;
724         case CFQ_RQ1_WRAP:
725                 return rq2;
726         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
727         default:
728                 /*
729                  * Since both rqs are wrapped,
730                  * start with the one that's further behind head
731                  * (--> only *one* back seek required),
732                  * since back seek takes more time than forward.
733                  */
734                 if (s1 <= s2)
735                         return rq1;
736                 else
737                         return rq2;
738         }
739 }
740
741 /*
742  * The below is leftmost cache rbtree addon
743  */
744 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
745 {
746         /* Service tree is empty */
747         if (!root->count)
748                 return NULL;
749
750         if (!root->left)
751                 root->left = rb_first(&root->rb);
752
753         if (root->left)
754                 return rb_entry(root->left, struct cfq_queue, rb_node);
755
756         return NULL;
757 }
758
759 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
760 {
761         if (!root->left)
762                 root->left = rb_first(&root->rb);
763
764         if (root->left)
765                 return rb_entry_cfqg(root->left);
766
767         return NULL;
768 }
769
770 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
771 {
772         rb_erase(n, root);
773         RB_CLEAR_NODE(n);
774 }
775
776 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
777 {
778         if (root->left == n)
779                 root->left = NULL;
780         rb_erase_init(n, &root->rb);
781         --root->count;
782 }
783
784 /*
785  * would be nice to take fifo expire time into account as well
786  */
787 static struct request *
788 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
789                   struct request *last)
790 {
791         struct rb_node *rbnext = rb_next(&last->rb_node);
792         struct rb_node *rbprev = rb_prev(&last->rb_node);
793         struct request *next = NULL, *prev = NULL;
794
795         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
796
797         if (rbprev)
798                 prev = rb_entry_rq(rbprev);
799
800         if (rbnext)
801                 next = rb_entry_rq(rbnext);
802         else {
803                 rbnext = rb_first(&cfqq->sort_list);
804                 if (rbnext && rbnext != &last->rb_node)
805                         next = rb_entry_rq(rbnext);
806         }
807
808         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
809 }
810
811 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
812                                       struct cfq_queue *cfqq)
813 {
814         /*
815          * just an approximation, should be ok.
816          */
817         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
818                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
819 }
820
821 static inline s64
822 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
823 {
824         return cfqg->vdisktime - st->min_vdisktime;
825 }
826
827 static void
828 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
829 {
830         struct rb_node **node = &st->rb.rb_node;
831         struct rb_node *parent = NULL;
832         struct cfq_group *__cfqg;
833         s64 key = cfqg_key(st, cfqg);
834         int left = 1;
835
836         while (*node != NULL) {
837                 parent = *node;
838                 __cfqg = rb_entry_cfqg(parent);
839
840                 if (key < cfqg_key(st, __cfqg))
841                         node = &parent->rb_left;
842                 else {
843                         node = &parent->rb_right;
844                         left = 0;
845                 }
846         }
847
848         if (left)
849                 st->left = &cfqg->rb_node;
850
851         rb_link_node(&cfqg->rb_node, parent, node);
852         rb_insert_color(&cfqg->rb_node, &st->rb);
853 }
854
855 static void
856 cfq_update_group_weight(struct cfq_group *cfqg)
857 {
858         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
859         if (cfqg->needs_update) {
860                 cfqg->weight = cfqg->new_weight;
861                 cfqg->needs_update = false;
862         }
863 }
864
865 static void
866 cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
867 {
868         BUG_ON(!RB_EMPTY_NODE(&cfqg->rb_node));
869
870         cfq_update_group_weight(cfqg);
871         __cfq_group_service_tree_add(st, cfqg);
872         st->total_weight += cfqg->weight;
873 }
874
875 static void
876 cfq_group_notify_queue_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
877 {
878         struct cfq_rb_root *st = &cfqd->grp_service_tree;
879         struct cfq_group *__cfqg;
880         struct rb_node *n;
881
882         cfqg->nr_cfqq++;
883         if (!RB_EMPTY_NODE(&cfqg->rb_node))
884                 return;
885
886         /*
887          * Currently put the group at the end. Later implement something
888          * so that groups get lesser vtime based on their weights, so that
889          * if group does not loose all if it was not continuously backlogged.
890          */
891         n = rb_last(&st->rb);
892         if (n) {
893                 __cfqg = rb_entry_cfqg(n);
894                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
895         } else
896                 cfqg->vdisktime = st->min_vdisktime;
897         cfq_group_service_tree_add(st, cfqg);
898 }
899
900 static void
901 cfq_group_service_tree_del(struct cfq_rb_root *st, struct cfq_group *cfqg)
902 {
903         st->total_weight -= cfqg->weight;
904         if (!RB_EMPTY_NODE(&cfqg->rb_node))
905                 cfq_rb_erase(&cfqg->rb_node, st);
906 }
907
908 static void
909 cfq_group_notify_queue_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
910 {
911         struct cfq_rb_root *st = &cfqd->grp_service_tree;
912
913         BUG_ON(cfqg->nr_cfqq < 1);
914         cfqg->nr_cfqq--;
915
916         /* If there are other cfq queues under this group, don't delete it */
917         if (cfqg->nr_cfqq)
918                 return;
919
920         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
921         cfq_group_service_tree_del(st, cfqg);
922         cfqg->saved_workload_slice = 0;
923         cfq_blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
924 }
925
926 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq,
927                                                 unsigned int *unaccounted_time)
928 {
929         unsigned int slice_used;
930
931         /*
932          * Queue got expired before even a single request completed or
933          * got expired immediately after first request completion.
934          */
935         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
936                 /*
937                  * Also charge the seek time incurred to the group, otherwise
938                  * if there are mutiple queues in the group, each can dispatch
939                  * a single request on seeky media and cause lots of seek time
940                  * and group will never know it.
941                  */
942                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
943                                         1);
944         } else {
945                 slice_used = jiffies - cfqq->slice_start;
946                 if (slice_used > cfqq->allocated_slice) {
947                         *unaccounted_time = slice_used - cfqq->allocated_slice;
948                         slice_used = cfqq->allocated_slice;
949                 }
950                 if (time_after(cfqq->slice_start, cfqq->dispatch_start))
951                         *unaccounted_time += cfqq->slice_start -
952                                         cfqq->dispatch_start;
953         }
954
955         return slice_used;
956 }
957
958 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
959                                 struct cfq_queue *cfqq)
960 {
961         struct cfq_rb_root *st = &cfqd->grp_service_tree;
962         unsigned int used_sl, charge, unaccounted_sl = 0;
963         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
964                         - cfqg->service_tree_idle.count;
965
966         BUG_ON(nr_sync < 0);
967         used_sl = charge = cfq_cfqq_slice_usage(cfqq, &unaccounted_sl);
968
969         if (iops_mode(cfqd))
970                 charge = cfqq->slice_dispatch;
971         else if (!cfq_cfqq_sync(cfqq) && !nr_sync)
972                 charge = cfqq->allocated_slice;
973
974         /* Can't update vdisktime while group is on service tree */
975         cfq_group_service_tree_del(st, cfqg);
976         cfqg->vdisktime += cfq_scale_slice(charge, cfqg);
977         /* If a new weight was requested, update now, off tree */
978         cfq_group_service_tree_add(st, cfqg);
979
980         /* This group is being expired. Save the context */
981         if (time_after(cfqd->workload_expires, jiffies)) {
982                 cfqg->saved_workload_slice = cfqd->workload_expires
983                                                 - jiffies;
984                 cfqg->saved_workload = cfqd->serving_type;
985                 cfqg->saved_serving_prio = cfqd->serving_prio;
986         } else
987                 cfqg->saved_workload_slice = 0;
988
989         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
990                                         st->min_vdisktime);
991         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u disp=%u charge=%u iops=%u"
992                         " sect=%u", used_sl, cfqq->slice_dispatch, charge,
993                         iops_mode(cfqd), cfqq->nr_sectors);
994         cfq_blkiocg_update_timeslice_used(&cfqg->blkg, used_sl,
995                                           unaccounted_sl);
996         cfq_blkiocg_set_start_empty_time(&cfqg->blkg);
997 }
998
999 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1000 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
1001 {
1002         if (blkg)
1003                 return container_of(blkg, struct cfq_group, blkg);
1004         return NULL;
1005 }
1006
1007 static void cfq_update_blkio_group_weight(void *key, struct blkio_group *blkg,
1008                                           unsigned int weight)
1009 {
1010         struct cfq_group *cfqg = cfqg_of_blkg(blkg);
1011         cfqg->new_weight = weight;
1012         cfqg->needs_update = true;
1013 }
1014
1015 static void cfq_init_add_cfqg_lists(struct cfq_data *cfqd,
1016                         struct cfq_group *cfqg, struct blkio_cgroup *blkcg)
1017 {
1018         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1019         unsigned int major, minor;
1020
1021         /*
1022          * Add group onto cgroup list. It might happen that bdi->dev is
1023          * not initialized yet. Initialize this new group without major
1024          * and minor info and this info will be filled in once a new thread
1025          * comes for IO.
1026          */
1027         if (bdi->dev) {
1028                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1029                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1030                                         (void *)cfqd, MKDEV(major, minor));
1031         } else
1032                 cfq_blkiocg_add_blkio_group(blkcg, &cfqg->blkg,
1033                                         (void *)cfqd, 0);
1034
1035         cfqd->nr_blkcg_linked_grps++;
1036         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1037
1038         /* Add group on cfqd list */
1039         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1040 }
1041
1042 /*
1043  * Should be called from sleepable context. No request queue lock as per
1044  * cpu stats are allocated dynamically and alloc_percpu needs to be called
1045  * from sleepable context.
1046  */
1047 static struct cfq_group * cfq_alloc_cfqg(struct cfq_data *cfqd)
1048 {
1049         struct cfq_group *cfqg = NULL;
1050         int i, j, ret;
1051         struct cfq_rb_root *st;
1052
1053         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
1054         if (!cfqg)
1055                 return NULL;
1056
1057         for_each_cfqg_st(cfqg, i, j, st)
1058                 *st = CFQ_RB_ROOT;
1059         RB_CLEAR_NODE(&cfqg->rb_node);
1060
1061         /*
1062          * Take the initial reference that will be released on destroy
1063          * This can be thought of a joint reference by cgroup and
1064          * elevator which will be dropped by either elevator exit
1065          * or cgroup deletion path depending on who is exiting first.
1066          */
1067         cfqg->ref = 1;
1068
1069         ret = blkio_alloc_blkg_stats(&cfqg->blkg);
1070         if (ret) {
1071                 kfree(cfqg);
1072                 return NULL;
1073         }
1074
1075         return cfqg;
1076 }
1077
1078 static struct cfq_group *
1079 cfq_find_cfqg(struct cfq_data *cfqd, struct blkio_cgroup *blkcg)
1080 {
1081         struct cfq_group *cfqg = NULL;
1082         void *key = cfqd;
1083         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
1084         unsigned int major, minor;
1085
1086         /*
1087          * This is the common case when there are no blkio cgroups.
1088          * Avoid lookup in this case
1089          */
1090         if (blkcg == &blkio_root_cgroup)
1091                 cfqg = &cfqd->root_group;
1092         else
1093                 cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
1094
1095         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
1096                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
1097                 cfqg->blkg.dev = MKDEV(major, minor);
1098         }
1099
1100         return cfqg;
1101 }
1102
1103 /*
1104  * Search for the cfq group current task belongs to. request_queue lock must
1105  * be held.
1106  */
1107 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd)
1108 {
1109         struct blkio_cgroup *blkcg;
1110         struct cfq_group *cfqg = NULL, *__cfqg = NULL;
1111         struct request_queue *q = cfqd->queue;
1112
1113         rcu_read_lock();
1114         blkcg = task_blkio_cgroup(current);
1115         cfqg = cfq_find_cfqg(cfqd, blkcg);
1116         if (cfqg) {
1117                 rcu_read_unlock();
1118                 return cfqg;
1119         }
1120
1121         /*
1122          * Need to allocate a group. Allocation of group also needs allocation
1123          * of per cpu stats which in-turn takes a mutex() and can block. Hence
1124          * we need to drop rcu lock and queue_lock before we call alloc.
1125          *
1126          * Not taking any queue reference here and assuming that queue is
1127          * around by the time we return. CFQ queue allocation code does
1128          * the same. It might be racy though.
1129          */
1130
1131         rcu_read_unlock();
1132         spin_unlock_irq(q->queue_lock);
1133
1134         cfqg = cfq_alloc_cfqg(cfqd);
1135
1136         spin_lock_irq(q->queue_lock);
1137
1138         rcu_read_lock();
1139         blkcg = task_blkio_cgroup(current);
1140
1141         /*
1142          * If some other thread already allocated the group while we were
1143          * not holding queue lock, free up the group
1144          */
1145         __cfqg = cfq_find_cfqg(cfqd, blkcg);
1146
1147         if (__cfqg) {
1148                 kfree(cfqg);
1149                 rcu_read_unlock();
1150                 return __cfqg;
1151         }
1152
1153         if (!cfqg)
1154                 cfqg = &cfqd->root_group;
1155
1156         cfq_init_add_cfqg_lists(cfqd, cfqg, blkcg);
1157         rcu_read_unlock();
1158         return cfqg;
1159 }
1160
1161 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1162 {
1163         cfqg->ref++;
1164         return cfqg;
1165 }
1166
1167 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1168 {
1169         /* Currently, all async queues are mapped to root group */
1170         if (!cfq_cfqq_sync(cfqq))
1171                 cfqg = &cfqq->cfqd->root_group;
1172
1173         cfqq->cfqg = cfqg;
1174         /* cfqq reference on cfqg */
1175         cfqq->cfqg->ref++;
1176 }
1177
1178 static void cfq_put_cfqg(struct cfq_group *cfqg)
1179 {
1180         struct cfq_rb_root *st;
1181         int i, j;
1182
1183         BUG_ON(cfqg->ref <= 0);
1184         cfqg->ref--;
1185         if (cfqg->ref)
1186                 return;
1187         for_each_cfqg_st(cfqg, i, j, st)
1188                 BUG_ON(!RB_EMPTY_ROOT(&st->rb));
1189         free_percpu(cfqg->blkg.stats_cpu);
1190         kfree(cfqg);
1191 }
1192
1193 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1194 {
1195         /* Something wrong if we are trying to remove same group twice */
1196         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1197
1198         hlist_del_init(&cfqg->cfqd_node);
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_META) {
1596                 WARN_ON(!cfqq->meta_pending);
1597                 cfqq->meta_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. Allow
1669          * merge only if rq is queued there.
1670          */
1671         cic = cfq_cic_lookup(cfqd, current->io_context);
1672         if (!cic)
1673                 return false;
1674
1675         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1676         return cfqq == RQ_CFQQ(rq);
1677 }
1678
1679 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1680 {
1681         del_timer(&cfqd->idle_slice_timer);
1682         cfq_blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1683 }
1684
1685 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1686                                    struct cfq_queue *cfqq)
1687 {
1688         if (cfqq) {
1689                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1690                                 cfqd->serving_prio, cfqd->serving_type);
1691                 cfq_blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1692                 cfqq->slice_start = 0;
1693                 cfqq->dispatch_start = jiffies;
1694                 cfqq->allocated_slice = 0;
1695                 cfqq->slice_end = 0;
1696                 cfqq->slice_dispatch = 0;
1697                 cfqq->nr_sectors = 0;
1698
1699                 cfq_clear_cfqq_wait_request(cfqq);
1700                 cfq_clear_cfqq_must_dispatch(cfqq);
1701                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1702                 cfq_clear_cfqq_fifo_expire(cfqq);
1703                 cfq_mark_cfqq_slice_new(cfqq);
1704
1705                 cfq_del_timer(cfqd, cfqq);
1706         }
1707
1708         cfqd->active_queue = cfqq;
1709 }
1710
1711 /*
1712  * current cfqq expired its slice (or was too idle), select new one
1713  */
1714 static void
1715 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1716                     bool timed_out)
1717 {
1718         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1719
1720         if (cfq_cfqq_wait_request(cfqq))
1721                 cfq_del_timer(cfqd, cfqq);
1722
1723         cfq_clear_cfqq_wait_request(cfqq);
1724         cfq_clear_cfqq_wait_busy(cfqq);
1725
1726         /*
1727          * If this cfqq is shared between multiple processes, check to
1728          * make sure that those processes are still issuing I/Os within
1729          * the mean seek distance.  If not, it may be time to break the
1730          * queues apart again.
1731          */
1732         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1733                 cfq_mark_cfqq_split_coop(cfqq);
1734
1735         /*
1736          * store what was left of this slice, if the queue idled/timed out
1737          */
1738         if (timed_out) {
1739                 if (cfq_cfqq_slice_new(cfqq))
1740                         cfqq->slice_resid = cfq_scaled_cfqq_slice(cfqd, cfqq);
1741                 else
1742                         cfqq->slice_resid = cfqq->slice_end - jiffies;
1743                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1744         }
1745
1746         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1747
1748         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1749                 cfq_del_cfqq_rr(cfqd, cfqq);
1750
1751         cfq_resort_rr_list(cfqd, cfqq);
1752
1753         if (cfqq == cfqd->active_queue)
1754                 cfqd->active_queue = NULL;
1755
1756         if (cfqd->active_cic) {
1757                 put_io_context(cfqd->active_cic->ioc);
1758                 cfqd->active_cic = NULL;
1759         }
1760 }
1761
1762 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1763 {
1764         struct cfq_queue *cfqq = cfqd->active_queue;
1765
1766         if (cfqq)
1767                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1768 }
1769
1770 /*
1771  * Get next queue for service. Unless we have a queue preemption,
1772  * we'll simply select the first cfqq in the service tree.
1773  */
1774 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1775 {
1776         struct cfq_rb_root *service_tree =
1777                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1778                                         cfqd->serving_type);
1779
1780         if (!cfqd->rq_queued)
1781                 return NULL;
1782
1783         /* There is nothing to dispatch */
1784         if (!service_tree)
1785                 return NULL;
1786         if (RB_EMPTY_ROOT(&service_tree->rb))
1787                 return NULL;
1788         return cfq_rb_first(service_tree);
1789 }
1790
1791 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1792 {
1793         struct cfq_group *cfqg;
1794         struct cfq_queue *cfqq;
1795         int i, j;
1796         struct cfq_rb_root *st;
1797
1798         if (!cfqd->rq_queued)
1799                 return NULL;
1800
1801         cfqg = cfq_get_next_cfqg(cfqd);
1802         if (!cfqg)
1803                 return NULL;
1804
1805         for_each_cfqg_st(cfqg, i, j, st)
1806                 if ((cfqq = cfq_rb_first(st)) != NULL)
1807                         return cfqq;
1808         return NULL;
1809 }
1810
1811 /*
1812  * Get and set a new active queue for service.
1813  */
1814 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1815                                               struct cfq_queue *cfqq)
1816 {
1817         if (!cfqq)
1818                 cfqq = cfq_get_next_queue(cfqd);
1819
1820         __cfq_set_active_queue(cfqd, cfqq);
1821         return cfqq;
1822 }
1823
1824 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1825                                           struct request *rq)
1826 {
1827         if (blk_rq_pos(rq) >= cfqd->last_position)
1828                 return blk_rq_pos(rq) - cfqd->last_position;
1829         else
1830                 return cfqd->last_position - blk_rq_pos(rq);
1831 }
1832
1833 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1834                                struct request *rq)
1835 {
1836         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1837 }
1838
1839 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1840                                     struct cfq_queue *cur_cfqq)
1841 {
1842         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1843         struct rb_node *parent, *node;
1844         struct cfq_queue *__cfqq;
1845         sector_t sector = cfqd->last_position;
1846
1847         if (RB_EMPTY_ROOT(root))
1848                 return NULL;
1849
1850         /*
1851          * First, if we find a request starting at the end of the last
1852          * request, choose it.
1853          */
1854         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1855         if (__cfqq)
1856                 return __cfqq;
1857
1858         /*
1859          * If the exact sector wasn't found, the parent of the NULL leaf
1860          * will contain the closest sector.
1861          */
1862         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1863         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1864                 return __cfqq;
1865
1866         if (blk_rq_pos(__cfqq->next_rq) < sector)
1867                 node = rb_next(&__cfqq->p_node);
1868         else
1869                 node = rb_prev(&__cfqq->p_node);
1870         if (!node)
1871                 return NULL;
1872
1873         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1874         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1875                 return __cfqq;
1876
1877         return NULL;
1878 }
1879
1880 /*
1881  * cfqd - obvious
1882  * cur_cfqq - passed in so that we don't decide that the current queue is
1883  *            closely cooperating with itself.
1884  *
1885  * So, basically we're assuming that that cur_cfqq has dispatched at least
1886  * one request, and that cfqd->last_position reflects a position on the disk
1887  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1888  * assumption.
1889  */
1890 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1891                                               struct cfq_queue *cur_cfqq)
1892 {
1893         struct cfq_queue *cfqq;
1894
1895         if (cfq_class_idle(cur_cfqq))
1896                 return NULL;
1897         if (!cfq_cfqq_sync(cur_cfqq))
1898                 return NULL;
1899         if (CFQQ_SEEKY(cur_cfqq))
1900                 return NULL;
1901
1902         /*
1903          * Don't search priority tree if it's the only queue in the group.
1904          */
1905         if (cur_cfqq->cfqg->nr_cfqq == 1)
1906                 return NULL;
1907
1908         /*
1909          * We should notice if some of the queues are cooperating, eg
1910          * working closely on the same area of the disk. In that case,
1911          * we can group them together and don't waste time idling.
1912          */
1913         cfqq = cfqq_close(cfqd, cur_cfqq);
1914         if (!cfqq)
1915                 return NULL;
1916
1917         /* If new queue belongs to different cfq_group, don't choose it */
1918         if (cur_cfqq->cfqg != cfqq->cfqg)
1919                 return NULL;
1920
1921         /*
1922          * It only makes sense to merge sync queues.
1923          */
1924         if (!cfq_cfqq_sync(cfqq))
1925                 return NULL;
1926         if (CFQQ_SEEKY(cfqq))
1927                 return NULL;
1928
1929         /*
1930          * Do not merge queues of different priority classes
1931          */
1932         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1933                 return NULL;
1934
1935         return cfqq;
1936 }
1937
1938 /*
1939  * Determine whether we should enforce idle window for this queue.
1940  */
1941
1942 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1943 {
1944         enum wl_prio_t prio = cfqq_prio(cfqq);
1945         struct cfq_rb_root *service_tree = cfqq->service_tree;
1946
1947         BUG_ON(!service_tree);
1948         BUG_ON(!service_tree->count);
1949
1950         if (!cfqd->cfq_slice_idle)
1951                 return false;
1952
1953         /* We never do for idle class queues. */
1954         if (prio == IDLE_WORKLOAD)
1955                 return false;
1956
1957         /* We do for queues that were marked with idle window flag. */
1958         if (cfq_cfqq_idle_window(cfqq) &&
1959            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1960                 return true;
1961
1962         /*
1963          * Otherwise, we do only if they are the last ones
1964          * in their service tree.
1965          */
1966         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1967                 return true;
1968         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1969                         service_tree->count);
1970         return false;
1971 }
1972
1973 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1974 {
1975         struct cfq_queue *cfqq = cfqd->active_queue;
1976         struct cfq_io_context *cic;
1977         unsigned long sl, group_idle = 0;
1978
1979         /*
1980          * SSD device without seek penalty, disable idling. But only do so
1981          * for devices that support queuing, otherwise we still have a problem
1982          * with sync vs async workloads.
1983          */
1984         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1985                 return;
1986
1987         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1988         WARN_ON(cfq_cfqq_slice_new(cfqq));
1989
1990         /*
1991          * idle is disabled, either manually or by past process history
1992          */
1993         if (!cfq_should_idle(cfqd, cfqq)) {
1994                 /* no queue idling. Check for group idling */
1995                 if (cfqd->cfq_group_idle)
1996                         group_idle = cfqd->cfq_group_idle;
1997                 else
1998                         return;
1999         }
2000
2001         /*
2002          * still active requests from this queue, don't idle
2003          */
2004         if (cfqq->dispatched)
2005                 return;
2006
2007         /*
2008          * task has exited, don't wait
2009          */
2010         cic = cfqd->active_cic;
2011         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
2012                 return;
2013
2014         /*
2015          * If our average think time is larger than the remaining time
2016          * slice, then don't idle. This avoids overrunning the allotted
2017          * time slice.
2018          */
2019         if (sample_valid(cic->ttime_samples) &&
2020             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
2021                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
2022                                 cic->ttime_mean);
2023                 return;
2024         }
2025
2026         /* There are other queues in the group, don't do group idle */
2027         if (group_idle && cfqq->cfqg->nr_cfqq > 1)
2028                 return;
2029
2030         cfq_mark_cfqq_wait_request(cfqq);
2031
2032         if (group_idle)
2033                 sl = cfqd->cfq_group_idle;
2034         else
2035                 sl = cfqd->cfq_slice_idle;
2036
2037         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
2038         cfq_blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
2039         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu group_idle: %d", sl,
2040                         group_idle ? 1 : 0);
2041 }
2042
2043 /*
2044  * Move request from internal lists to the request queue dispatch list.
2045  */
2046 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
2047 {
2048         struct cfq_data *cfqd = q->elevator->elevator_data;
2049         struct cfq_queue *cfqq = RQ_CFQQ(rq);
2050
2051         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
2052
2053         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
2054         cfq_remove_request(rq);
2055         cfqq->dispatched++;
2056         (RQ_CFQG(rq))->dispatched++;
2057         elv_dispatch_sort(q, rq);
2058
2059         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
2060         cfqq->nr_sectors += blk_rq_sectors(rq);
2061         cfq_blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
2062                                         rq_data_dir(rq), rq_is_sync(rq));
2063 }
2064
2065 /*
2066  * return expired entry, or NULL to just start from scratch in rbtree
2067  */
2068 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
2069 {
2070         struct request *rq = NULL;
2071
2072         if (cfq_cfqq_fifo_expire(cfqq))
2073                 return NULL;
2074
2075         cfq_mark_cfqq_fifo_expire(cfqq);
2076
2077         if (list_empty(&cfqq->fifo))
2078                 return NULL;
2079
2080         rq = rq_entry_fifo(cfqq->fifo.next);
2081         if (time_before(jiffies, rq_fifo_time(rq)))
2082                 rq = NULL;
2083
2084         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
2085         return rq;
2086 }
2087
2088 static inline int
2089 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2090 {
2091         const int base_rq = cfqd->cfq_slice_async_rq;
2092
2093         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
2094
2095         return 2 * base_rq * (IOPRIO_BE_NR - cfqq->ioprio);
2096 }
2097
2098 /*
2099  * Must be called with the queue_lock held.
2100  */
2101 static int cfqq_process_refs(struct cfq_queue *cfqq)
2102 {
2103         int process_refs, io_refs;
2104
2105         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
2106         process_refs = cfqq->ref - io_refs;
2107         BUG_ON(process_refs < 0);
2108         return process_refs;
2109 }
2110
2111 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
2112 {
2113         int process_refs, new_process_refs;
2114         struct cfq_queue *__cfqq;
2115
2116         /*
2117          * If there are no process references on the new_cfqq, then it is
2118          * unsafe to follow the ->new_cfqq chain as other cfqq's in the
2119          * chain may have dropped their last reference (not just their
2120          * last process reference).
2121          */
2122         if (!cfqq_process_refs(new_cfqq))
2123                 return;
2124
2125         /* Avoid a circular list and skip interim queue merges */
2126         while ((__cfqq = new_cfqq->new_cfqq)) {
2127                 if (__cfqq == cfqq)
2128                         return;
2129                 new_cfqq = __cfqq;
2130         }
2131
2132         process_refs = cfqq_process_refs(cfqq);
2133         new_process_refs = cfqq_process_refs(new_cfqq);
2134         /*
2135          * If the process for the cfqq has gone away, there is no
2136          * sense in merging the queues.
2137          */
2138         if (process_refs == 0 || new_process_refs == 0)
2139                 return;
2140
2141         /*
2142          * Merge in the direction of the lesser amount of work.
2143          */
2144         if (new_process_refs >= process_refs) {
2145                 cfqq->new_cfqq = new_cfqq;
2146                 new_cfqq->ref += process_refs;
2147         } else {
2148                 new_cfqq->new_cfqq = cfqq;
2149                 cfqq->ref += new_process_refs;
2150         }
2151 }
2152
2153 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2154                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2155 {
2156         struct cfq_queue *queue;
2157         int i;
2158         bool key_valid = false;
2159         unsigned long lowest_key = 0;
2160         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2161
2162         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2163                 /* select the one with lowest rb_key */
2164                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2165                 if (queue &&
2166                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2167                         lowest_key = queue->rb_key;
2168                         cur_best = i;
2169                         key_valid = true;
2170                 }
2171         }
2172
2173         return cur_best;
2174 }
2175
2176 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2177 {
2178         unsigned slice;
2179         unsigned count;
2180         struct cfq_rb_root *st;
2181         unsigned group_slice;
2182         enum wl_prio_t original_prio = cfqd->serving_prio;
2183
2184         /* Choose next priority. RT > BE > IDLE */
2185         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2186                 cfqd->serving_prio = RT_WORKLOAD;
2187         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2188                 cfqd->serving_prio = BE_WORKLOAD;
2189         else {
2190                 cfqd->serving_prio = IDLE_WORKLOAD;
2191                 cfqd->workload_expires = jiffies + 1;
2192                 return;
2193         }
2194
2195         if (original_prio != cfqd->serving_prio)
2196                 goto new_workload;
2197
2198         /*
2199          * For RT and BE, we have to choose also the type
2200          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2201          * expiration time
2202          */
2203         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2204         count = st->count;
2205
2206         /*
2207          * check workload expiration, and that we still have other queues ready
2208          */
2209         if (count && !time_after(jiffies, cfqd->workload_expires))
2210                 return;
2211
2212 new_workload:
2213         /* otherwise select new workload type */
2214         cfqd->serving_type =
2215                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2216         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2217         count = st->count;
2218
2219         /*
2220          * the workload slice is computed as a fraction of target latency
2221          * proportional to the number of queues in that workload, over
2222          * all the queues in the same priority class
2223          */
2224         group_slice = cfq_group_slice(cfqd, cfqg);
2225
2226         slice = group_slice * count /
2227                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2228                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2229
2230         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2231                 unsigned int tmp;
2232
2233                 /*
2234                  * Async queues are currently system wide. Just taking
2235                  * proportion of queues with-in same group will lead to higher
2236                  * async ratio system wide as generally root group is going
2237                  * to have higher weight. A more accurate thing would be to
2238                  * calculate system wide asnc/sync ratio.
2239                  */
2240                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2241                 tmp = tmp/cfqd->busy_queues;
2242                 slice = min_t(unsigned, slice, tmp);
2243
2244                 /* async workload slice is scaled down according to
2245                  * the sync/async slice ratio. */
2246                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2247         } else
2248                 /* sync workload slice is at least 2 * cfq_slice_idle */
2249                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2250
2251         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2252         cfq_log(cfqd, "workload slice:%d", slice);
2253         cfqd->workload_expires = jiffies + slice;
2254 }
2255
2256 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2257 {
2258         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2259         struct cfq_group *cfqg;
2260
2261         if (RB_EMPTY_ROOT(&st->rb))
2262                 return NULL;
2263         cfqg = cfq_rb_first_group(st);
2264         update_min_vdisktime(st);
2265         return cfqg;
2266 }
2267
2268 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2269 {
2270         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2271
2272         cfqd->serving_group = cfqg;
2273
2274         /* Restore the workload type data */
2275         if (cfqg->saved_workload_slice) {
2276                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2277                 cfqd->serving_type = cfqg->saved_workload;
2278                 cfqd->serving_prio = cfqg->saved_serving_prio;
2279         } else
2280                 cfqd->workload_expires = jiffies - 1;
2281
2282         choose_service_tree(cfqd, cfqg);
2283 }
2284
2285 /*
2286  * Select a queue for service. If we have a current active queue,
2287  * check whether to continue servicing it, or retrieve and set a new one.
2288  */
2289 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2290 {
2291         struct cfq_queue *cfqq, *new_cfqq = NULL;
2292
2293         cfqq = cfqd->active_queue;
2294         if (!cfqq)
2295                 goto new_queue;
2296
2297         if (!cfqd->rq_queued)
2298                 return NULL;
2299
2300         /*
2301          * We were waiting for group to get backlogged. Expire the queue
2302          */
2303         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2304                 goto expire;
2305
2306         /*
2307          * The active queue has run out of time, expire it and select new.
2308          */
2309         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2310                 /*
2311                  * If slice had not expired at the completion of last request
2312                  * we might not have turned on wait_busy flag. Don't expire
2313                  * the queue yet. Allow the group to get backlogged.
2314                  *
2315                  * The very fact that we have used the slice, that means we
2316                  * have been idling all along on this queue and it should be
2317                  * ok to wait for this request to complete.
2318                  */
2319                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2320                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2321                         cfqq = NULL;
2322                         goto keep_queue;
2323                 } else
2324                         goto check_group_idle;
2325         }
2326
2327         /*
2328          * The active queue has requests and isn't expired, allow it to
2329          * dispatch.
2330          */
2331         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2332                 goto keep_queue;
2333
2334         /*
2335          * If another queue has a request waiting within our mean seek
2336          * distance, let it run.  The expire code will check for close
2337          * cooperators and put the close queue at the front of the service
2338          * tree.  If possible, merge the expiring queue with the new cfqq.
2339          */
2340         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2341         if (new_cfqq) {
2342                 if (!cfqq->new_cfqq)
2343                         cfq_setup_merge(cfqq, new_cfqq);
2344                 goto expire;
2345         }
2346
2347         /*
2348          * No requests pending. If the active queue still has requests in
2349          * flight or is idling for a new request, allow either of these
2350          * conditions to happen (or time out) before selecting a new queue.
2351          */
2352         if (timer_pending(&cfqd->idle_slice_timer)) {
2353                 cfqq = NULL;
2354                 goto keep_queue;
2355         }
2356
2357         /*
2358          * This is a deep seek queue, but the device is much faster than
2359          * the queue can deliver, don't idle
2360          **/
2361         if (CFQQ_SEEKY(cfqq) && cfq_cfqq_idle_window(cfqq) &&
2362             (cfq_cfqq_slice_new(cfqq) ||
2363             (cfqq->slice_end - jiffies > jiffies - cfqq->slice_start))) {
2364                 cfq_clear_cfqq_deep(cfqq);
2365                 cfq_clear_cfqq_idle_window(cfqq);
2366         }
2367
2368         if (cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2369                 cfqq = NULL;
2370                 goto keep_queue;
2371         }
2372
2373         /*
2374          * If group idle is enabled and there are requests dispatched from
2375          * this group, wait for requests to complete.
2376          */
2377 check_group_idle:
2378         if (cfqd->cfq_group_idle && cfqq->cfqg->nr_cfqq == 1
2379             && cfqq->cfqg->dispatched) {
2380                 cfqq = NULL;
2381                 goto keep_queue;
2382         }
2383
2384 expire:
2385         cfq_slice_expired(cfqd, 0);
2386 new_queue:
2387         /*
2388          * Current queue expired. Check if we have to switch to a new
2389          * service tree
2390          */
2391         if (!new_cfqq)
2392                 cfq_choose_cfqg(cfqd);
2393
2394         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2395 keep_queue:
2396         return cfqq;
2397 }
2398
2399 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2400 {
2401         int dispatched = 0;
2402
2403         while (cfqq->next_rq) {
2404                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2405                 dispatched++;
2406         }
2407
2408         BUG_ON(!list_empty(&cfqq->fifo));
2409
2410         /* By default cfqq is not expired if it is empty. Do it explicitly */
2411         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2412         return dispatched;
2413 }
2414
2415 /*
2416  * Drain our current requests. Used for barriers and when switching
2417  * io schedulers on-the-fly.
2418  */
2419 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2420 {
2421         struct cfq_queue *cfqq;
2422         int dispatched = 0;
2423
2424         /* Expire the timeslice of the current active queue first */
2425         cfq_slice_expired(cfqd, 0);
2426         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2427                 __cfq_set_active_queue(cfqd, cfqq);
2428                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2429         }
2430
2431         BUG_ON(cfqd->busy_queues);
2432
2433         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2434         return dispatched;
2435 }
2436
2437 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2438         struct cfq_queue *cfqq)
2439 {
2440         /* the queue hasn't finished any request, can't estimate */
2441         if (cfq_cfqq_slice_new(cfqq))
2442                 return true;
2443         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2444                 cfqq->slice_end))
2445                 return true;
2446
2447         return false;
2448 }
2449
2450 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2451 {
2452         unsigned int max_dispatch;
2453
2454         /*
2455          * Drain async requests before we start sync IO
2456          */
2457         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2458                 return false;
2459
2460         /*
2461          * If this is an async queue and we have sync IO in flight, let it wait
2462          */
2463         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2464                 return false;
2465
2466         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2467         if (cfq_class_idle(cfqq))
2468                 max_dispatch = 1;
2469
2470         /*
2471          * Does this cfqq already have too much IO in flight?
2472          */
2473         if (cfqq->dispatched >= max_dispatch) {
2474                 bool promote_sync = false;
2475                 /*
2476                  * idle queue must always only have a single IO in flight
2477                  */
2478                 if (cfq_class_idle(cfqq))
2479                         return false;
2480
2481                 /*
2482                  * If there is only one sync queue
2483                  * we can ignore async queue here and give the sync
2484                  * queue no dispatch limit. The reason is a sync queue can
2485                  * preempt async queue, limiting the sync queue doesn't make
2486                  * sense. This is useful for aiostress test.
2487                  */
2488                 if (cfq_cfqq_sync(cfqq) && cfqd->busy_sync_queues == 1)
2489                         promote_sync = true;
2490
2491                 /*
2492                  * We have other queues, don't allow more IO from this one
2493                  */
2494                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq) &&
2495                                 !promote_sync)
2496                         return false;
2497
2498                 /*
2499                  * Sole queue user, no limit
2500                  */
2501                 if (cfqd->busy_queues == 1 || promote_sync)
2502                         max_dispatch = -1;
2503                 else
2504                         /*
2505                          * Normally we start throttling cfqq when cfq_quantum/2
2506                          * requests have been dispatched. But we can drive
2507                          * deeper queue depths at the beginning of slice
2508                          * subjected to upper limit of cfq_quantum.
2509                          * */
2510                         max_dispatch = cfqd->cfq_quantum;
2511         }
2512
2513         /*
2514          * Async queues must wait a bit before being allowed dispatch.
2515          * We also ramp up the dispatch depth gradually for async IO,
2516          * based on the last sync IO we serviced
2517          */
2518         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2519                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2520                 unsigned int depth;
2521
2522                 depth = last_sync / cfqd->cfq_slice[1];
2523                 if (!depth && !cfqq->dispatched)
2524                         depth = 1;
2525                 if (depth < max_dispatch)
2526                         max_dispatch = depth;
2527         }
2528
2529         /*
2530          * If we're below the current max, allow a dispatch
2531          */
2532         return cfqq->dispatched < max_dispatch;
2533 }
2534
2535 /*
2536  * Dispatch a request from cfqq, moving them to the request queue
2537  * dispatch list.
2538  */
2539 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2540 {
2541         struct request *rq;
2542
2543         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2544
2545         if (!cfq_may_dispatch(cfqd, cfqq))
2546                 return false;
2547
2548         /*
2549          * follow expired path, else get first next available
2550          */
2551         rq = cfq_check_fifo(cfqq);
2552         if (!rq)
2553                 rq = cfqq->next_rq;
2554
2555         /*
2556          * insert request into driver dispatch list
2557          */
2558         cfq_dispatch_insert(cfqd->queue, rq);
2559
2560         if (!cfqd->active_cic) {
2561                 struct cfq_io_context *cic = RQ_CIC(rq);
2562
2563                 atomic_long_inc(&cic->ioc->refcount);
2564                 cfqd->active_cic = cic;
2565         }
2566
2567         return true;
2568 }
2569
2570 /*
2571  * Find the cfqq that we need to service and move a request from that to the
2572  * dispatch list
2573  */
2574 static int cfq_dispatch_requests(struct request_queue *q, int force)
2575 {
2576         struct cfq_data *cfqd = q->elevator->elevator_data;
2577         struct cfq_queue *cfqq;
2578
2579         if (!cfqd->busy_queues)
2580                 return 0;
2581
2582         if (unlikely(force))
2583                 return cfq_forced_dispatch(cfqd);
2584
2585         cfqq = cfq_select_queue(cfqd);
2586         if (!cfqq)
2587                 return 0;
2588
2589         /*
2590          * Dispatch a request from this cfqq, if it is allowed
2591          */
2592         if (!cfq_dispatch_request(cfqd, cfqq))
2593                 return 0;
2594
2595         cfqq->slice_dispatch++;
2596         cfq_clear_cfqq_must_dispatch(cfqq);
2597
2598         /*
2599          * expire an async queue immediately if it has used up its slice. idle
2600          * queue always expire after 1 dispatch round.
2601          */
2602         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2603             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2604             cfq_class_idle(cfqq))) {
2605                 cfqq->slice_end = jiffies + 1;
2606                 cfq_slice_expired(cfqd, 0);
2607         }
2608
2609         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2610         return 1;
2611 }
2612
2613 /*
2614  * task holds one reference to the queue, dropped when task exits. each rq
2615  * in-flight on this queue also holds a reference, dropped when rq is freed.
2616  *
2617  * Each cfq queue took a reference on the parent group. Drop it now.
2618  * queue lock must be held here.
2619  */
2620 static void cfq_put_queue(struct cfq_queue *cfqq)
2621 {
2622         struct cfq_data *cfqd = cfqq->cfqd;
2623         struct cfq_group *cfqg;
2624
2625         BUG_ON(cfqq->ref <= 0);
2626
2627         cfqq->ref--;
2628         if (cfqq->ref)
2629                 return;
2630
2631         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2632         BUG_ON(rb_first(&cfqq->sort_list));
2633         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2634         cfqg = cfqq->cfqg;
2635
2636         if (unlikely(cfqd->active_queue == cfqq)) {
2637                 __cfq_slice_expired(cfqd, cfqq, 0);
2638                 cfq_schedule_dispatch(cfqd);
2639         }
2640
2641         BUG_ON(cfq_cfqq_on_rr(cfqq));
2642         kmem_cache_free(cfq_pool, cfqq);
2643         cfq_put_cfqg(cfqg);
2644 }
2645
2646 /*
2647  * Call func for each cic attached to this ioc.
2648  */
2649 static void
2650 call_for_each_cic(struct io_context *ioc,
2651                   void (*func)(struct io_context *, struct cfq_io_context *))
2652 {
2653         struct cfq_io_context *cic;
2654         struct hlist_node *n;
2655
2656         rcu_read_lock();
2657
2658         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2659                 func(ioc, cic);
2660
2661         rcu_read_unlock();
2662 }
2663
2664 static void cfq_cic_free_rcu(struct rcu_head *head)
2665 {
2666         struct cfq_io_context *cic;
2667
2668         cic = container_of(head, struct cfq_io_context, rcu_head);
2669
2670         kmem_cache_free(cfq_ioc_pool, cic);
2671         elv_ioc_count_dec(cfq_ioc_count);
2672
2673         if (ioc_gone) {
2674                 /*
2675                  * CFQ scheduler is exiting, grab exit lock and check
2676                  * the pending io context count. If it hits zero,
2677                  * complete ioc_gone and set it back to NULL
2678                  */
2679                 spin_lock(&ioc_gone_lock);
2680                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2681                         complete(ioc_gone);
2682                         ioc_gone = NULL;
2683                 }
2684                 spin_unlock(&ioc_gone_lock);
2685         }
2686 }
2687
2688 static void cfq_cic_free(struct cfq_io_context *cic)
2689 {
2690         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2691 }
2692
2693 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2694 {
2695         unsigned long flags;
2696         unsigned long dead_key = (unsigned long) cic->key;
2697
2698         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2699
2700         spin_lock_irqsave(&ioc->lock, flags);
2701         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2702         hlist_del_rcu(&cic->cic_list);
2703         spin_unlock_irqrestore(&ioc->lock, flags);
2704
2705         cfq_cic_free(cic);
2706 }
2707
2708 /*
2709  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2710  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2711  * and ->trim() which is called with the task lock held
2712  */
2713 static void cfq_free_io_context(struct io_context *ioc)
2714 {
2715         /*
2716          * ioc->refcount is zero here, or we are called from elv_unregister(),
2717          * so no more cic's are allowed to be linked into this ioc.  So it
2718          * should be ok to iterate over the known list, we will see all cic's
2719          * since no new ones are added.
2720          */
2721         call_for_each_cic(ioc, cic_free_func);
2722 }
2723
2724 static void cfq_put_cooperator(struct cfq_queue *cfqq)
2725 {
2726         struct cfq_queue *__cfqq, *next;
2727
2728         /*
2729          * If this queue was scheduled to merge with another queue, be
2730          * sure to drop the reference taken on that queue (and others in
2731          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2732          */
2733         __cfqq = cfqq->new_cfqq;
2734         while (__cfqq) {
2735                 if (__cfqq == cfqq) {
2736                         WARN(1, "cfqq->new_cfqq loop detected\n");
2737                         break;
2738                 }
2739                 next = __cfqq->new_cfqq;
2740                 cfq_put_queue(__cfqq);
2741                 __cfqq = next;
2742         }
2743 }
2744
2745 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2746 {
2747         if (unlikely(cfqq == cfqd->active_queue)) {
2748                 __cfq_slice_expired(cfqd, cfqq, 0);
2749                 cfq_schedule_dispatch(cfqd);
2750         }
2751
2752         cfq_put_cooperator(cfqq);
2753
2754         cfq_put_queue(cfqq);
2755 }
2756
2757 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2758                                          struct cfq_io_context *cic)
2759 {
2760         struct io_context *ioc = cic->ioc;
2761
2762         list_del_init(&cic->queue_list);
2763
2764         /*
2765          * Make sure dead mark is seen for dead queues
2766          */
2767         smp_wmb();
2768         cic->key = cfqd_dead_key(cfqd);
2769
2770         if (rcu_dereference(ioc->ioc_data) == cic) {
2771                 spin_lock(&ioc->lock);
2772                 rcu_assign_pointer(ioc->ioc_data, NULL);
2773                 spin_unlock(&ioc->lock);
2774         }
2775
2776         if (cic->cfqq[BLK_RW_ASYNC]) {
2777                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2778                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2779         }
2780
2781         if (cic->cfqq[BLK_RW_SYNC]) {
2782                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2783                 cic->cfqq[BLK_RW_SYNC] = NULL;
2784         }
2785 }
2786
2787 static void cfq_exit_single_io_context(struct io_context *ioc,
2788                                        struct cfq_io_context *cic)
2789 {
2790         struct cfq_data *cfqd = cic_to_cfqd(cic);
2791
2792         if (cfqd) {
2793                 struct request_queue *q = cfqd->queue;
2794                 unsigned long flags;
2795
2796                 spin_lock_irqsave(q->queue_lock, flags);
2797
2798                 /*
2799                  * Ensure we get a fresh copy of the ->key to prevent
2800                  * race between exiting task and queue
2801                  */
2802                 smp_read_barrier_depends();
2803                 if (cic->key == cfqd)
2804                         __cfq_exit_single_io_context(cfqd, cic);
2805
2806                 spin_unlock_irqrestore(q->queue_lock, flags);
2807         }
2808 }
2809
2810 /*
2811  * The process that ioc belongs to has exited, we need to clean up
2812  * and put the internal structures we have that belongs to that process.
2813  */
2814 static void cfq_exit_io_context(struct io_context *ioc)
2815 {
2816         call_for_each_cic(ioc, cfq_exit_single_io_context);
2817 }
2818
2819 static struct cfq_io_context *
2820 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2821 {
2822         struct cfq_io_context *cic;
2823
2824         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2825                                                         cfqd->queue->node);
2826         if (cic) {
2827                 cic->last_end_request = jiffies;
2828                 INIT_LIST_HEAD(&cic->queue_list);
2829                 INIT_HLIST_NODE(&cic->cic_list);
2830                 cic->dtor = cfq_free_io_context;
2831                 cic->exit = cfq_exit_io_context;
2832                 elv_ioc_count_inc(cfq_ioc_count);
2833         }
2834
2835         return cic;
2836 }
2837
2838 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2839 {
2840         struct task_struct *tsk = current;
2841         int ioprio_class;
2842
2843         if (!cfq_cfqq_prio_changed(cfqq))
2844                 return;
2845
2846         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2847         switch (ioprio_class) {
2848         default:
2849                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2850         case IOPRIO_CLASS_NONE:
2851                 /*
2852                  * no prio set, inherit CPU scheduling settings
2853                  */
2854                 cfqq->ioprio = task_nice_ioprio(tsk);
2855                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2856                 break;
2857         case IOPRIO_CLASS_RT:
2858                 cfqq->ioprio = task_ioprio(ioc);
2859                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2860                 break;
2861         case IOPRIO_CLASS_BE:
2862                 cfqq->ioprio = task_ioprio(ioc);
2863                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2864                 break;
2865         case IOPRIO_CLASS_IDLE:
2866                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2867                 cfqq->ioprio = 7;
2868                 cfq_clear_cfqq_idle_window(cfqq);
2869                 break;
2870         }
2871
2872         /*
2873          * keep track of original prio settings in case we have to temporarily
2874          * elevate the priority of this queue
2875          */
2876         cfqq->org_ioprio = cfqq->ioprio;
2877         cfqq->org_ioprio_class = cfqq->ioprio_class;
2878         cfq_clear_cfqq_prio_changed(cfqq);
2879 }
2880
2881 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2882 {
2883         struct cfq_data *cfqd = cic_to_cfqd(cic);
2884         struct cfq_queue *cfqq;
2885         unsigned long flags;
2886
2887         if (unlikely(!cfqd))
2888                 return;
2889
2890         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2891
2892         cfqq = cic->cfqq[BLK_RW_ASYNC];
2893         if (cfqq) {
2894                 struct cfq_queue *new_cfqq;
2895                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2896                                                 GFP_ATOMIC);
2897                 if (new_cfqq) {
2898                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2899                         cfq_put_queue(cfqq);
2900                 }
2901         }
2902
2903         cfqq = cic->cfqq[BLK_RW_SYNC];
2904         if (cfqq)
2905                 cfq_mark_cfqq_prio_changed(cfqq);
2906
2907         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2908 }
2909
2910 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2911 {
2912         call_for_each_cic(ioc, changed_ioprio);
2913         ioc->ioprio_changed = 0;
2914 }
2915
2916 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2917                           pid_t pid, bool is_sync)
2918 {
2919         RB_CLEAR_NODE(&cfqq->rb_node);
2920         RB_CLEAR_NODE(&cfqq->p_node);
2921         INIT_LIST_HEAD(&cfqq->fifo);
2922
2923         cfqq->ref = 0;
2924         cfqq->cfqd = cfqd;
2925
2926         cfq_mark_cfqq_prio_changed(cfqq);
2927
2928         if (is_sync) {
2929                 if (!cfq_class_idle(cfqq))
2930                         cfq_mark_cfqq_idle_window(cfqq);
2931                 cfq_mark_cfqq_sync(cfqq);
2932         }
2933         cfqq->pid = pid;
2934 }
2935
2936 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2937 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2938 {
2939         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2940         struct cfq_data *cfqd = cic_to_cfqd(cic);
2941         unsigned long flags;
2942         struct request_queue *q;
2943
2944         if (unlikely(!cfqd))
2945                 return;
2946
2947         q = cfqd->queue;
2948
2949         spin_lock_irqsave(q->queue_lock, flags);
2950
2951         if (sync_cfqq) {
2952                 /*
2953                  * Drop reference to sync queue. A new sync queue will be
2954                  * assigned in new group upon arrival of a fresh request.
2955                  */
2956                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2957                 cic_set_cfqq(cic, NULL, 1);
2958                 cfq_put_queue(sync_cfqq);
2959         }
2960
2961         spin_unlock_irqrestore(q->queue_lock, flags);
2962 }
2963
2964 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2965 {
2966         call_for_each_cic(ioc, changed_cgroup);
2967         ioc->cgroup_changed = 0;
2968 }
2969 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2970
2971 static struct cfq_queue *
2972 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2973                      struct io_context *ioc, gfp_t gfp_mask)
2974 {
2975         struct cfq_queue *cfqq, *new_cfqq = NULL;
2976         struct cfq_io_context *cic;
2977         struct cfq_group *cfqg;
2978
2979 retry:
2980         cfqg = cfq_get_cfqg(cfqd);
2981         cic = cfq_cic_lookup(cfqd, ioc);
2982         /* cic always exists here */
2983         cfqq = cic_to_cfqq(cic, is_sync);
2984
2985         /*
2986          * Always try a new alloc if we fell back to the OOM cfqq
2987          * originally, since it should just be a temporary situation.
2988          */
2989         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2990                 cfqq = NULL;
2991                 if (new_cfqq) {
2992                         cfqq = new_cfqq;
2993                         new_cfqq = NULL;
2994                 } else if (gfp_mask & __GFP_WAIT) {
2995                         spin_unlock_irq(cfqd->queue->queue_lock);
2996                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2997                                         gfp_mask | __GFP_ZERO,
2998                                         cfqd->queue->node);
2999                         spin_lock_irq(cfqd->queue->queue_lock);
3000                         if (new_cfqq)
3001                                 goto retry;
3002                 } else {
3003                         cfqq = kmem_cache_alloc_node(cfq_pool,
3004                                         gfp_mask | __GFP_ZERO,
3005                                         cfqd->queue->node);
3006                 }
3007
3008                 if (cfqq) {
3009                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
3010                         cfq_init_prio_data(cfqq, ioc);
3011                         cfq_link_cfqq_cfqg(cfqq, cfqg);
3012                         cfq_log_cfqq(cfqd, cfqq, "alloced");
3013                 } else
3014                         cfqq = &cfqd->oom_cfqq;
3015         }
3016
3017         if (new_cfqq)
3018                 kmem_cache_free(cfq_pool, new_cfqq);
3019
3020         return cfqq;
3021 }
3022
3023 static struct cfq_queue **
3024 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
3025 {
3026         switch (ioprio_class) {
3027         case IOPRIO_CLASS_RT:
3028                 return &cfqd->async_cfqq[0][ioprio];
3029         case IOPRIO_CLASS_BE:
3030                 return &cfqd->async_cfqq[1][ioprio];
3031         case IOPRIO_CLASS_IDLE:
3032                 return &cfqd->async_idle_cfqq;
3033         default:
3034                 BUG();
3035         }
3036 }
3037
3038 static struct cfq_queue *
3039 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
3040               gfp_t gfp_mask)
3041 {
3042         const int ioprio = task_ioprio(ioc);
3043         const int ioprio_class = task_ioprio_class(ioc);
3044         struct cfq_queue **async_cfqq = NULL;
3045         struct cfq_queue *cfqq = NULL;
3046
3047         if (!is_sync) {
3048                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
3049                 cfqq = *async_cfqq;
3050         }
3051
3052         if (!cfqq)
3053                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
3054
3055         /*
3056          * pin the queue now that it's allocated, scheduler exit will prune it
3057          */
3058         if (!is_sync && !(*async_cfqq)) {
3059                 cfqq->ref++;
3060                 *async_cfqq = cfqq;
3061         }
3062
3063         cfqq->ref++;
3064         return cfqq;
3065 }
3066
3067 /*
3068  * We drop cfq io contexts lazily, so we may find a dead one.
3069  */
3070 static void
3071 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
3072                   struct cfq_io_context *cic)
3073 {
3074         unsigned long flags;
3075
3076         WARN_ON(!list_empty(&cic->queue_list));
3077         BUG_ON(cic->key != cfqd_dead_key(cfqd));
3078
3079         spin_lock_irqsave(&ioc->lock, flags);
3080
3081         BUG_ON(ioc->ioc_data == cic);
3082
3083         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
3084         hlist_del_rcu(&cic->cic_list);
3085         spin_unlock_irqrestore(&ioc->lock, flags);
3086
3087         cfq_cic_free(cic);
3088 }
3089
3090 static struct cfq_io_context *
3091 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
3092 {
3093         struct cfq_io_context *cic;
3094         unsigned long flags;
3095
3096         if (unlikely(!ioc))
3097                 return NULL;
3098
3099         rcu_read_lock();
3100
3101         /*
3102          * we maintain a last-hit cache, to avoid browsing over the tree
3103          */
3104         cic = rcu_dereference(ioc->ioc_data);
3105         if (cic && cic->key == cfqd) {
3106                 rcu_read_unlock();
3107                 return cic;
3108         }
3109
3110         do {
3111                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
3112                 rcu_read_unlock();
3113                 if (!cic)
3114                         break;
3115                 if (unlikely(cic->key != cfqd)) {
3116                         cfq_drop_dead_cic(cfqd, ioc, cic);
3117                         rcu_read_lock();
3118                         continue;
3119                 }
3120
3121                 spin_lock_irqsave(&ioc->lock, flags);
3122                 rcu_assign_pointer(ioc->ioc_data, cic);
3123                 spin_unlock_irqrestore(&ioc->lock, flags);
3124                 break;
3125         } while (1);
3126
3127         return cic;
3128 }
3129
3130 /*
3131  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
3132  * the process specific cfq io context when entered from the block layer.
3133  * Also adds the cic to a per-cfqd list, used when this queue is removed.
3134  */
3135 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
3136                         struct cfq_io_context *cic, gfp_t gfp_mask)
3137 {
3138         unsigned long flags;
3139         int ret;
3140
3141         ret = radix_tree_preload(gfp_mask);
3142         if (!ret) {
3143                 cic->ioc = ioc;
3144                 cic->key = cfqd;
3145
3146                 spin_lock_irqsave(&ioc->lock, flags);
3147                 ret = radix_tree_insert(&ioc->radix_root,
3148                                                 cfqd->cic_index, cic);
3149                 if (!ret)
3150                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
3151                 spin_unlock_irqrestore(&ioc->lock, flags);
3152
3153                 radix_tree_preload_end();
3154
3155                 if (!ret) {
3156                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3157                         list_add(&cic->queue_list, &cfqd->cic_list);
3158                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3159                 }
3160         }
3161
3162         if (ret)
3163                 printk(KERN_ERR "cfq: cic link failed!\n");
3164
3165         return ret;
3166 }
3167
3168 /*
3169  * Setup general io context and cfq io context. There can be several cfq
3170  * io contexts per general io context, if this process is doing io to more
3171  * than one device managed by cfq.
3172  */
3173 static struct cfq_io_context *
3174 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3175 {
3176         struct io_context *ioc = NULL;
3177         struct cfq_io_context *cic;
3178
3179         might_sleep_if(gfp_mask & __GFP_WAIT);
3180
3181         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3182         if (!ioc)
3183                 return NULL;
3184
3185         cic = cfq_cic_lookup(cfqd, ioc);
3186         if (cic)
3187                 goto out;
3188
3189         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3190         if (cic == NULL)
3191                 goto err;
3192
3193         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3194                 goto err_free;
3195
3196 out:
3197         smp_read_barrier_depends();
3198         if (unlikely(ioc->ioprio_changed))
3199                 cfq_ioc_set_ioprio(ioc);
3200
3201 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3202         if (unlikely(ioc->cgroup_changed))
3203                 cfq_ioc_set_cgroup(ioc);
3204 #endif
3205         return cic;
3206 err_free:
3207         cfq_cic_free(cic);
3208 err:
3209         put_io_context(ioc);
3210         return NULL;
3211 }
3212
3213 static void
3214 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3215 {
3216         unsigned long elapsed = jiffies - cic->last_end_request;
3217         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3218
3219         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3220         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3221         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3222 }
3223
3224 static void
3225 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3226                        struct request *rq)
3227 {
3228         sector_t sdist = 0;
3229         sector_t n_sec = blk_rq_sectors(rq);
3230         if (cfqq->last_request_pos) {
3231                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3232                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3233                 else
3234                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3235         }
3236
3237         cfqq->seek_history <<= 1;
3238         if (blk_queue_nonrot(cfqd->queue))
3239                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3240         else
3241                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3242 }
3243
3244 /*
3245  * Disable idle window if the process thinks too long or seeks so much that
3246  * it doesn't matter
3247  */
3248 static void
3249 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3250                        struct cfq_io_context *cic)
3251 {
3252         int old_idle, enable_idle;
3253
3254         /*
3255          * Don't idle for async or idle io prio class
3256          */
3257         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3258                 return;
3259
3260         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3261
3262         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3263                 cfq_mark_cfqq_deep(cfqq);
3264
3265         if (cfqq->next_rq && (cfqq->next_rq->cmd_flags & REQ_NOIDLE))
3266                 enable_idle = 0;
3267         else if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3268             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3269                 enable_idle = 0;
3270         else if (sample_valid(cic->ttime_samples)) {
3271                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3272                         enable_idle = 0;
3273                 else
3274                         enable_idle = 1;
3275         }
3276
3277         if (old_idle != enable_idle) {
3278                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3279                 if (enable_idle)
3280                         cfq_mark_cfqq_idle_window(cfqq);
3281                 else
3282                         cfq_clear_cfqq_idle_window(cfqq);
3283         }
3284 }
3285
3286 /*
3287  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3288  * no or if we aren't sure, a 1 will cause a preempt.
3289  */
3290 static bool
3291 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3292                    struct request *rq)
3293 {
3294         struct cfq_queue *cfqq;
3295
3296         cfqq = cfqd->active_queue;
3297         if (!cfqq)
3298                 return false;
3299
3300         if (cfq_class_idle(new_cfqq))
3301                 return false;
3302
3303         if (cfq_class_idle(cfqq))
3304                 return true;
3305
3306         /*
3307          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3308          */
3309         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3310                 return false;
3311
3312         /*
3313          * if the new request is sync, but the currently running queue is
3314          * not, let the sync request have priority.
3315          */
3316         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3317                 return true;
3318
3319         if (new_cfqq->cfqg != cfqq->cfqg)
3320                 return false;
3321
3322         if (cfq_slice_used(cfqq))
3323                 return true;
3324
3325         /* Allow preemption only if we are idling on sync-noidle tree */
3326         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3327             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3328             new_cfqq->service_tree->count == 2 &&
3329             RB_EMPTY_ROOT(&cfqq->sort_list))
3330                 return true;
3331
3332         /*
3333          * So both queues are sync. Let the new request get disk time if
3334          * it's a metadata request and the current queue is doing regular IO.
3335          */
3336         if ((rq->cmd_flags & REQ_META) && !cfqq->meta_pending)
3337                 return true;
3338
3339         /*
3340          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3341          */
3342         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3343                 return true;
3344
3345         /* An idle queue should not be idle now for some reason */
3346         if (RB_EMPTY_ROOT(&cfqq->sort_list) && !cfq_should_idle(cfqd, cfqq))
3347                 return true;
3348
3349         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3350                 return false;
3351
3352         /*
3353          * if this request is as-good as one we would expect from the
3354          * current cfqq, let it preempt
3355          */
3356         if (cfq_rq_close(cfqd, cfqq, rq))
3357                 return true;
3358
3359         return false;
3360 }
3361
3362 /*
3363  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3364  * let it have half of its nominal slice.
3365  */
3366 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3367 {
3368         struct cfq_queue *old_cfqq = cfqd->active_queue;
3369
3370         cfq_log_cfqq(cfqd, cfqq, "preempt");
3371         cfq_slice_expired(cfqd, 1);
3372
3373         /*
3374          * workload type is changed, don't save slice, otherwise preempt
3375          * doesn't happen
3376          */
3377         if (cfqq_type(old_cfqq) != cfqq_type(cfqq))
3378                 cfqq->cfqg->saved_workload_slice = 0;
3379
3380         /*
3381          * Put the new queue at the front of the of the current list,
3382          * so we know that it will be selected next.
3383          */
3384         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3385
3386         cfq_service_tree_add(cfqd, cfqq, 1);
3387
3388         cfqq->slice_end = 0;
3389         cfq_mark_cfqq_slice_new(cfqq);
3390 }
3391
3392 /*
3393  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3394  * something we should do about it
3395  */
3396 static void
3397 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3398                 struct request *rq)
3399 {
3400         struct cfq_io_context *cic = RQ_CIC(rq);
3401
3402         cfqd->rq_queued++;
3403         if (rq->cmd_flags & REQ_META)
3404                 cfqq->meta_pending++;
3405
3406         cfq_update_io_thinktime(cfqd, cic);
3407         cfq_update_io_seektime(cfqd, cfqq, rq);
3408         cfq_update_idle_window(cfqd, cfqq, cic);
3409
3410         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3411
3412         if (cfqq == cfqd->active_queue) {
3413                 /*
3414                  * Remember that we saw a request from this process, but
3415                  * don't start queuing just yet. Otherwise we risk seeing lots
3416                  * of tiny requests, because we disrupt the normal plugging
3417                  * and merging. If the request is already larger than a single
3418                  * page, let it rip immediately. For that case we assume that
3419                  * merging is already done. Ditto for a busy system that
3420                  * has other work pending, don't risk delaying until the
3421                  * idle timer unplug to continue working.
3422                  */
3423                 if (cfq_cfqq_wait_request(cfqq)) {
3424                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3425                             cfqd->busy_queues > 1) {
3426                                 cfq_del_timer(cfqd, cfqq);
3427                                 cfq_clear_cfqq_wait_request(cfqq);
3428                                 __blk_run_queue(cfqd->queue);
3429                         } else {
3430                                 cfq_blkiocg_update_idle_time_stats(
3431                                                 &cfqq->cfqg->blkg);
3432                                 cfq_mark_cfqq_must_dispatch(cfqq);
3433                         }
3434                 }
3435         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3436                 /*
3437                  * not the active queue - expire current slice if it is
3438                  * idle and has expired it's mean thinktime or this new queue
3439                  * has some old slice time left and is of higher priority or
3440                  * this new queue is RT and the current one is BE
3441                  */
3442                 cfq_preempt_queue(cfqd, cfqq);
3443                 __blk_run_queue(cfqd->queue);
3444         }
3445 }
3446
3447 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3448 {
3449         struct cfq_data *cfqd = q->elevator->elevator_data;
3450         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3451
3452         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3453         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3454
3455         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3456         list_add_tail(&rq->queuelist, &cfqq->fifo);
3457         cfq_add_rq_rb(rq);
3458         cfq_blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3459                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3460                         rq_is_sync(rq));
3461         cfq_rq_enqueued(cfqd, cfqq, rq);
3462 }
3463
3464 /*
3465  * Update hw_tag based on peak queue depth over 50 samples under
3466  * sufficient load.
3467  */
3468 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3469 {
3470         struct cfq_queue *cfqq = cfqd->active_queue;
3471
3472         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3473                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3474
3475         if (cfqd->hw_tag == 1)
3476                 return;
3477
3478         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3479             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3480                 return;
3481
3482         /*
3483          * If active queue hasn't enough requests and can idle, cfq might not
3484          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3485          * case
3486          */
3487         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3488             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3489             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3490                 return;
3491
3492         if (cfqd->hw_tag_samples++ < 50)
3493                 return;
3494
3495         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3496                 cfqd->hw_tag = 1;
3497         else
3498                 cfqd->hw_tag = 0;
3499 }
3500
3501 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3502 {
3503         struct cfq_io_context *cic = cfqd->active_cic;
3504
3505         /* If the queue already has requests, don't wait */
3506         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3507                 return false;
3508
3509         /* If there are other queues in the group, don't wait */
3510         if (cfqq->cfqg->nr_cfqq > 1)
3511                 return false;
3512
3513         if (cfq_slice_used(cfqq))
3514                 return true;
3515
3516         /* if slice left is less than think time, wait busy */
3517         if (cic && sample_valid(cic->ttime_samples)
3518             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3519                 return true;
3520
3521         /*
3522          * If think times is less than a jiffy than ttime_mean=0 and above
3523          * will not be true. It might happen that slice has not expired yet
3524          * but will expire soon (4-5 ns) during select_queue(). To cover the
3525          * case where think time is less than a jiffy, mark the queue wait
3526          * busy if only 1 jiffy is left in the slice.
3527          */
3528         if (cfqq->slice_end - jiffies == 1)
3529                 return true;
3530
3531         return false;
3532 }
3533
3534 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3535 {
3536         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3537         struct cfq_data *cfqd = cfqq->cfqd;
3538         const int sync = rq_is_sync(rq);
3539         unsigned long now;
3540
3541         now = jiffies;
3542         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d",
3543                      !!(rq->cmd_flags & REQ_NOIDLE));
3544
3545         cfq_update_hw_tag(cfqd);
3546
3547         WARN_ON(!cfqd->rq_in_driver);
3548         WARN_ON(!cfqq->dispatched);
3549         cfqd->rq_in_driver--;
3550         cfqq->dispatched--;
3551         (RQ_CFQG(rq))->dispatched--;
3552         cfq_blkiocg_update_completion_stats(&cfqq->cfqg->blkg,
3553                         rq_start_time_ns(rq), rq_io_start_time_ns(rq),
3554                         rq_data_dir(rq), rq_is_sync(rq));
3555
3556         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3557
3558         if (sync) {
3559                 RQ_CIC(rq)->last_end_request = now;
3560                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3561                         cfqd->last_delayed_sync = now;
3562         }
3563
3564         /*
3565          * If this is the active queue, check if it needs to be expired,
3566          * or if we want to idle in case it has no pending requests.
3567          */
3568         if (cfqd->active_queue == cfqq) {
3569                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3570
3571                 if (cfq_cfqq_slice_new(cfqq)) {
3572                         cfq_set_prio_slice(cfqd, cfqq);
3573                         cfq_clear_cfqq_slice_new(cfqq);
3574                 }
3575
3576                 /*
3577                  * Should we wait for next request to come in before we expire
3578                  * the queue.
3579                  */
3580                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3581                         unsigned long extend_sl = cfqd->cfq_slice_idle;
3582                         if (!cfqd->cfq_slice_idle)
3583                                 extend_sl = cfqd->cfq_group_idle;
3584                         cfqq->slice_end = jiffies + extend_sl;
3585                         cfq_mark_cfqq_wait_busy(cfqq);
3586                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3587                 }
3588
3589                 /*
3590                  * Idling is not enabled on:
3591                  * - expired queues
3592                  * - idle-priority queues
3593                  * - async queues
3594                  * - queues with still some requests queued
3595                  * - when there is a close cooperator
3596                  */
3597                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3598                         cfq_slice_expired(cfqd, 1);
3599                 else if (sync && cfqq_empty &&
3600                          !cfq_close_cooperator(cfqd, cfqq)) {
3601                         cfq_arm_slice_timer(cfqd);
3602                 }
3603         }
3604
3605         if (!cfqd->rq_in_driver)
3606                 cfq_schedule_dispatch(cfqd);
3607 }
3608
3609 /*
3610  * we temporarily boost lower priority queues if they are holding fs exclusive
3611  * resources. they are boosted to normal prio (CLASS_BE/4)
3612  */
3613 static void cfq_prio_boost(struct cfq_queue *cfqq)
3614 {
3615         if (has_fs_excl()) {
3616                 /*
3617                  * boost idle prio on transactions that would lock out other
3618                  * users of the filesystem
3619                  */
3620                 if (cfq_class_idle(cfqq))
3621                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3622                 if (cfqq->ioprio > IOPRIO_NORM)
3623                         cfqq->ioprio = IOPRIO_NORM;
3624         } else {
3625                 /*
3626                  * unboost the queue (if needed)
3627                  */
3628                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3629                 cfqq->ioprio = cfqq->org_ioprio;
3630         }
3631 }
3632
3633 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3634 {
3635         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3636                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3637                 return ELV_MQUEUE_MUST;
3638         }
3639
3640         return ELV_MQUEUE_MAY;
3641 }
3642
3643 static int cfq_may_queue(struct request_queue *q, int rw)
3644 {
3645         struct cfq_data *cfqd = q->elevator->elevator_data;
3646         struct task_struct *tsk = current;
3647         struct cfq_io_context *cic;
3648         struct cfq_queue *cfqq;
3649
3650         /*
3651          * don't force setup of a queue from here, as a call to may_queue
3652          * does not necessarily imply that a request actually will be queued.
3653          * so just lookup a possibly existing queue, or return 'may queue'
3654          * if that fails
3655          */
3656         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3657         if (!cic)
3658                 return ELV_MQUEUE_MAY;
3659
3660         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3661         if (cfqq) {
3662                 cfq_init_prio_data(cfqq, cic->ioc);
3663                 cfq_prio_boost(cfqq);
3664
3665                 return __cfq_may_queue(cfqq);
3666         }
3667
3668         return ELV_MQUEUE_MAY;
3669 }
3670
3671 /*
3672  * queue lock held here
3673  */
3674 static void cfq_put_request(struct request *rq)
3675 {
3676         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3677
3678         if (cfqq) {
3679                 const int rw = rq_data_dir(rq);
3680
3681                 BUG_ON(!cfqq->allocated[rw]);
3682                 cfqq->allocated[rw]--;
3683
3684                 put_io_context(RQ_CIC(rq)->ioc);
3685
3686                 rq->elevator_private[0] = NULL;
3687                 rq->elevator_private[1] = NULL;
3688
3689                 /* Put down rq reference on cfqg */
3690                 cfq_put_cfqg(RQ_CFQG(rq));
3691                 rq->elevator_private[2] = NULL;
3692
3693                 cfq_put_queue(cfqq);
3694         }
3695 }
3696
3697 static struct cfq_queue *
3698 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3699                 struct cfq_queue *cfqq)
3700 {
3701         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3702         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3703         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3704         cfq_put_queue(cfqq);
3705         return cic_to_cfqq(cic, 1);
3706 }
3707
3708 /*
3709  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3710  * was the last process referring to said cfqq.
3711  */
3712 static struct cfq_queue *
3713 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3714 {
3715         if (cfqq_process_refs(cfqq) == 1) {
3716                 cfqq->pid = current->pid;
3717                 cfq_clear_cfqq_coop(cfqq);
3718                 cfq_clear_cfqq_split_coop(cfqq);
3719                 return cfqq;
3720         }
3721
3722         cic_set_cfqq(cic, NULL, 1);
3723
3724         cfq_put_cooperator(cfqq);
3725
3726         cfq_put_queue(cfqq);
3727         return NULL;
3728 }
3729 /*
3730  * Allocate cfq data structures associated with this request.
3731  */
3732 static int
3733 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3734 {
3735         struct cfq_data *cfqd = q->elevator->elevator_data;
3736         struct cfq_io_context *cic;
3737         const int rw = rq_data_dir(rq);
3738         const bool is_sync = rq_is_sync(rq);
3739         struct cfq_queue *cfqq;
3740         unsigned long flags;
3741
3742         might_sleep_if(gfp_mask & __GFP_WAIT);
3743
3744         cic = cfq_get_io_context(cfqd, gfp_mask);
3745
3746         spin_lock_irqsave(q->queue_lock, flags);
3747
3748         if (!cic)
3749                 goto queue_fail;
3750
3751 new_queue:
3752         cfqq = cic_to_cfqq(cic, is_sync);
3753         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3754                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3755                 cic_set_cfqq(cic, cfqq, is_sync);
3756         } else {
3757                 /*
3758                  * If the queue was seeky for too long, break it apart.
3759                  */
3760                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3761                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3762                         cfqq = split_cfqq(cic, cfqq);
3763                         if (!cfqq)
3764                                 goto new_queue;
3765                 }
3766
3767                 /*
3768                  * Check to see if this queue is scheduled to merge with
3769                  * another, closely cooperating queue.  The merging of
3770                  * queues happens here as it must be done in process context.
3771                  * The reference on new_cfqq was taken in merge_cfqqs.
3772                  */
3773                 if (cfqq->new_cfqq)
3774                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3775         }
3776
3777         cfqq->allocated[rw]++;
3778
3779         cfqq->ref++;
3780         rq->elevator_private[0] = cic;
3781         rq->elevator_private[1] = cfqq;
3782         rq->elevator_private[2] = cfq_ref_get_cfqg(cfqq->cfqg);
3783         spin_unlock_irqrestore(q->queue_lock, flags);
3784         return 0;
3785
3786 queue_fail:
3787         if (cic)
3788                 put_io_context(cic->ioc);
3789
3790         cfq_schedule_dispatch(cfqd);
3791         spin_unlock_irqrestore(q->queue_lock, flags);
3792         cfq_log(cfqd, "set_request fail");
3793         return 1;
3794 }
3795
3796 static void cfq_kick_queue(struct work_struct *work)
3797 {
3798         struct cfq_data *cfqd =
3799                 container_of(work, struct cfq_data, unplug_work);
3800         struct request_queue *q = cfqd->queue;
3801
3802         spin_lock_irq(q->queue_lock);
3803         __blk_run_queue(cfqd->queue);
3804         spin_unlock_irq(q->queue_lock);
3805 }
3806
3807 /*
3808  * Timer running if the active_queue is currently idling inside its time slice
3809  */
3810 static void cfq_idle_slice_timer(unsigned long data)
3811 {
3812         struct cfq_data *cfqd = (struct cfq_data *) data;
3813         struct cfq_queue *cfqq;
3814         unsigned long flags;
3815         int timed_out = 1;
3816
3817         cfq_log(cfqd, "idle timer fired");
3818
3819         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3820
3821         cfqq = cfqd->active_queue;
3822         if (cfqq) {
3823                 timed_out = 0;
3824
3825                 /*
3826                  * We saw a request before the queue expired, let it through
3827                  */
3828                 if (cfq_cfqq_must_dispatch(cfqq))
3829                         goto out_kick;
3830
3831                 /*
3832                  * expired
3833                  */
3834                 if (cfq_slice_used(cfqq))
3835                         goto expire;
3836
3837                 /*
3838                  * only expire and reinvoke request handler, if there are
3839                  * other queues with pending requests
3840                  */
3841                 if (!cfqd->busy_queues)
3842                         goto out_cont;
3843
3844                 /*
3845                  * not expired and it has a request pending, let it dispatch
3846                  */
3847                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3848                         goto out_kick;
3849
3850                 /*
3851                  * Queue depth flag is reset only when the idle didn't succeed
3852                  */
3853                 cfq_clear_cfqq_deep(cfqq);
3854         }
3855 expire:
3856         cfq_slice_expired(cfqd, timed_out);
3857 out_kick:
3858         cfq_schedule_dispatch(cfqd);
3859 out_cont:
3860         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3861 }
3862
3863 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3864 {
3865         del_timer_sync(&cfqd->idle_slice_timer);
3866         cancel_work_sync(&cfqd->unplug_work);
3867 }
3868
3869 static void cfq_put_async_queues(struct cfq_data *cfqd)
3870 {
3871         int i;
3872
3873         for (i = 0; i < IOPRIO_BE_NR; i++) {
3874                 if (cfqd->async_cfqq[0][i])
3875                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3876                 if (cfqd->async_cfqq[1][i])
3877                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3878         }
3879
3880         if (cfqd->async_idle_cfqq)
3881                 cfq_put_queue(cfqd->async_idle_cfqq);
3882 }
3883
3884 static void cfq_exit_queue(struct elevator_queue *e)
3885 {
3886         struct cfq_data *cfqd = e->elevator_data;
3887         struct request_queue *q = cfqd->queue;
3888         bool wait = false;
3889
3890         cfq_shutdown_timer_wq(cfqd);
3891
3892         spin_lock_irq(q->queue_lock);
3893
3894         if (cfqd->active_queue)
3895                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3896
3897         while (!list_empty(&cfqd->cic_list)) {
3898                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3899                                                         struct cfq_io_context,
3900                                                         queue_list);
3901
3902                 __cfq_exit_single_io_context(cfqd, cic);
3903         }
3904
3905         cfq_put_async_queues(cfqd);
3906         cfq_release_cfq_groups(cfqd);
3907
3908         /*
3909          * If there are groups which we could not unlink from blkcg list,
3910          * wait for a rcu period for them to be freed.
3911          */
3912         if (cfqd->nr_blkcg_linked_grps)
3913                 wait = true;
3914
3915         spin_unlock_irq(q->queue_lock);
3916
3917         cfq_shutdown_timer_wq(cfqd);
3918
3919         spin_lock(&cic_index_lock);
3920         ida_remove(&cic_index_ida, cfqd->cic_index);
3921         spin_unlock(&cic_index_lock);
3922
3923         /*
3924          * Wait for cfqg->blkg->key accessors to exit their grace periods.
3925          * Do this wait only if there are other unlinked groups out
3926          * there. This can happen if cgroup deletion path claimed the
3927          * responsibility of cleaning up a group before queue cleanup code
3928          * get to the group.
3929          *
3930          * Do not call synchronize_rcu() unconditionally as there are drivers
3931          * which create/delete request queue hundreds of times during scan/boot
3932          * and synchronize_rcu() can take significant time and slow down boot.
3933          */
3934         if (wait)
3935                 synchronize_rcu();
3936
3937 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3938         /* Free up per cpu stats for root group */
3939         free_percpu(cfqd->root_group.blkg.stats_cpu);
3940 #endif
3941         kfree(cfqd);
3942 }
3943
3944 static int cfq_alloc_cic_index(void)
3945 {
3946         int index, error;
3947
3948         do {
3949                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3950                         return -ENOMEM;
3951
3952                 spin_lock(&cic_index_lock);
3953                 error = ida_get_new(&cic_index_ida, &index);
3954                 spin_unlock(&cic_index_lock);
3955                 if (error && error != -EAGAIN)
3956                         return error;
3957         } while (error);
3958
3959         return index;
3960 }
3961
3962 static void *cfq_init_queue(struct request_queue *q)
3963 {
3964         struct cfq_data *cfqd;
3965         int i, j;
3966         struct cfq_group *cfqg;
3967         struct cfq_rb_root *st;
3968
3969         i = cfq_alloc_cic_index();
3970         if (i < 0)
3971                 return NULL;
3972
3973         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3974         if (!cfqd) {
3975                 spin_lock(&cic_index_lock);
3976                 ida_remove(&cic_index_ida, i);
3977                 spin_unlock(&cic_index_lock);
3978                 return NULL;
3979         }
3980
3981         /*
3982          * Don't need take queue_lock in the routine, since we are
3983          * initializing the ioscheduler, and nobody is using cfqd
3984          */
3985         cfqd->cic_index = i;
3986
3987         /* Init root service tree */
3988         cfqd->grp_service_tree = CFQ_RB_ROOT;
3989
3990         /* Init root group */
3991         cfqg = &cfqd->root_group;
3992         for_each_cfqg_st(cfqg, i, j, st)
3993                 *st = CFQ_RB_ROOT;
3994         RB_CLEAR_NODE(&cfqg->rb_node);
3995
3996         /* Give preference to root group over other groups */
3997         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3998
3999 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4000         /*
4001          * Set root group reference to 2. One reference will be dropped when
4002          * all groups on cfqd->cfqg_list are being deleted during queue exit.
4003          * Other reference will remain there as we don't want to delete this
4004          * group as it is statically allocated and gets destroyed when
4005          * throtl_data goes away.
4006          */
4007         cfqg->ref = 2;
4008
4009         if (blkio_alloc_blkg_stats(&cfqg->blkg)) {
4010                 kfree(cfqg);
4011                 kfree(cfqd);
4012                 return NULL;
4013         }
4014
4015         rcu_read_lock();
4016
4017         cfq_blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg,
4018                                         (void *)cfqd, 0);
4019         rcu_read_unlock();
4020         cfqd->nr_blkcg_linked_grps++;
4021
4022         /* Add group on cfqd->cfqg_list */
4023         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
4024 #endif
4025         /*
4026          * Not strictly needed (since RB_ROOT just clears the node and we
4027          * zeroed cfqd on alloc), but better be safe in case someone decides
4028          * to add magic to the rb code
4029          */
4030         for (i = 0; i < CFQ_PRIO_LISTS; i++)
4031                 cfqd->prio_trees[i] = RB_ROOT;
4032
4033         /*
4034          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
4035          * Grab a permanent reference to it, so that the normal code flow
4036          * will not attempt to free it.
4037          */
4038         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
4039         cfqd->oom_cfqq.ref++;
4040         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
4041
4042         INIT_LIST_HEAD(&cfqd->cic_list);
4043
4044         cfqd->queue = q;
4045
4046         init_timer(&cfqd->idle_slice_timer);
4047         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
4048         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
4049
4050         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
4051
4052         cfqd->cfq_quantum = cfq_quantum;
4053         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
4054         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
4055         cfqd->cfq_back_max = cfq_back_max;
4056         cfqd->cfq_back_penalty = cfq_back_penalty;
4057         cfqd->cfq_slice[0] = cfq_slice_async;
4058         cfqd->cfq_slice[1] = cfq_slice_sync;
4059         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
4060         cfqd->cfq_slice_idle = cfq_slice_idle;
4061         cfqd->cfq_group_idle = cfq_group_idle;
4062         cfqd->cfq_latency = 1;
4063         cfqd->hw_tag = -1;
4064         /*
4065          * we optimistically start assuming sync ops weren't delayed in last
4066          * second, in order to have larger depth for async operations.
4067          */
4068         cfqd->last_delayed_sync = jiffies - HZ;
4069         return cfqd;
4070 }
4071
4072 static void cfq_slab_kill(void)
4073 {
4074         /*
4075          * Caller already ensured that pending RCU callbacks are completed,
4076          * so we should have no busy allocations at this point.
4077          */
4078         if (cfq_pool)
4079                 kmem_cache_destroy(cfq_pool);
4080         if (cfq_ioc_pool)
4081                 kmem_cache_destroy(cfq_ioc_pool);
4082 }
4083
4084 static int __init cfq_slab_setup(void)
4085 {
4086         cfq_pool = KMEM_CACHE(cfq_queue, 0);
4087         if (!cfq_pool)
4088                 goto fail;
4089
4090         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
4091         if (!cfq_ioc_pool)
4092                 goto fail;
4093
4094         return 0;
4095 fail:
4096         cfq_slab_kill();
4097         return -ENOMEM;
4098 }
4099
4100 /*
4101  * sysfs parts below -->
4102  */
4103 static ssize_t
4104 cfq_var_show(unsigned int var, char *page)
4105 {
4106         return sprintf(page, "%d\n", var);
4107 }
4108
4109 static ssize_t
4110 cfq_var_store(unsigned int *var, const char *page, size_t count)
4111 {
4112         char *p = (char *) page;
4113
4114         *var = simple_strtoul(p, &p, 10);
4115         return count;
4116 }
4117
4118 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
4119 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
4120 {                                                                       \
4121         struct cfq_data *cfqd = e->elevator_data;                       \
4122         unsigned int __data = __VAR;                                    \
4123         if (__CONV)                                                     \
4124                 __data = jiffies_to_msecs(__data);                      \
4125         return cfq_var_show(__data, (page));                            \
4126 }
4127 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
4128 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
4129 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
4130 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
4131 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
4132 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
4133 SHOW_FUNCTION(cfq_group_idle_show, cfqd->cfq_group_idle, 1);
4134 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
4135 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
4136 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
4137 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
4138 #undef SHOW_FUNCTION
4139
4140 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
4141 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
4142 {                                                                       \
4143         struct cfq_data *cfqd = e->elevator_data;                       \
4144         unsigned int __data;                                            \
4145         int ret = cfq_var_store(&__data, (page), count);                \
4146         if (__data < (MIN))                                             \
4147                 __data = (MIN);                                         \
4148         else if (__data > (MAX))                                        \
4149                 __data = (MAX);                                         \
4150         if (__CONV)                                                     \
4151                 *(__PTR) = msecs_to_jiffies(__data);                    \
4152         else                                                            \
4153                 *(__PTR) = __data;                                      \
4154         return ret;                                                     \
4155 }
4156 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
4157 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
4158                 UINT_MAX, 1);
4159 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
4160                 UINT_MAX, 1);
4161 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
4162 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
4163                 UINT_MAX, 0);
4164 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
4165 STORE_FUNCTION(cfq_group_idle_store, &cfqd->cfq_group_idle, 0, UINT_MAX, 1);
4166 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
4167 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
4168 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
4169                 UINT_MAX, 0);
4170 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
4171 #undef STORE_FUNCTION
4172
4173 #define CFQ_ATTR(name) \
4174         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
4175
4176 static struct elv_fs_entry cfq_attrs[] = {
4177         CFQ_ATTR(quantum),
4178         CFQ_ATTR(fifo_expire_sync),
4179         CFQ_ATTR(fifo_expire_async),
4180         CFQ_ATTR(back_seek_max),
4181         CFQ_ATTR(back_seek_penalty),
4182         CFQ_ATTR(slice_sync),
4183         CFQ_ATTR(slice_async),
4184         CFQ_ATTR(slice_async_rq),
4185         CFQ_ATTR(slice_idle),
4186         CFQ_ATTR(group_idle),
4187         CFQ_ATTR(low_latency),
4188         __ATTR_NULL
4189 };
4190
4191 static struct elevator_type iosched_cfq = {
4192         .ops = {
4193                 .elevator_merge_fn =            cfq_merge,
4194                 .elevator_merged_fn =           cfq_merged_request,
4195                 .elevator_merge_req_fn =        cfq_merged_requests,
4196                 .elevator_allow_merge_fn =      cfq_allow_merge,
4197                 .elevator_bio_merged_fn =       cfq_bio_merged,
4198                 .elevator_dispatch_fn =         cfq_dispatch_requests,
4199                 .elevator_add_req_fn =          cfq_insert_request,
4200                 .elevator_activate_req_fn =     cfq_activate_request,
4201                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
4202                 .elevator_completed_req_fn =    cfq_completed_request,
4203                 .elevator_former_req_fn =       elv_rb_former_request,
4204                 .elevator_latter_req_fn =       elv_rb_latter_request,
4205                 .elevator_set_req_fn =          cfq_set_request,
4206                 .elevator_put_req_fn =          cfq_put_request,
4207                 .elevator_may_queue_fn =        cfq_may_queue,
4208                 .elevator_init_fn =             cfq_init_queue,
4209                 .elevator_exit_fn =             cfq_exit_queue,
4210                 .trim =                         cfq_free_io_context,
4211         },
4212         .elevator_attrs =       cfq_attrs,
4213         .elevator_name =        "cfq",
4214         .elevator_owner =       THIS_MODULE,
4215 };
4216
4217 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4218 static struct blkio_policy_type blkio_policy_cfq = {
4219         .ops = {
4220                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
4221                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
4222         },
4223         .plid = BLKIO_POLICY_PROP,
4224 };
4225 #else
4226 static struct blkio_policy_type blkio_policy_cfq;
4227 #endif
4228
4229 static int __init cfq_init(void)
4230 {
4231         /*
4232          * could be 0 on HZ < 1000 setups
4233          */
4234         if (!cfq_slice_async)
4235                 cfq_slice_async = 1;
4236         if (!cfq_slice_idle)
4237                 cfq_slice_idle = 1;
4238
4239 #ifdef CONFIG_CFQ_GROUP_IOSCHED
4240         if (!cfq_group_idle)
4241                 cfq_group_idle = 1;
4242 #else
4243                 cfq_group_idle = 0;
4244 #endif
4245         if (cfq_slab_setup())
4246                 return -ENOMEM;
4247
4248         elv_register(&iosched_cfq);
4249         blkio_policy_register(&blkio_policy_cfq);
4250
4251         return 0;
4252 }
4253
4254 static void __exit cfq_exit(void)
4255 {
4256         DECLARE_COMPLETION_ONSTACK(all_gone);
4257         blkio_policy_unregister(&blkio_policy_cfq);
4258         elv_unregister(&iosched_cfq);
4259         ioc_gone = &all_gone;
4260         /* ioc_gone's update must be visible before reading ioc_count */
4261         smp_wmb();
4262
4263         /*
4264          * this also protects us from entering cfq_slab_kill() with
4265          * pending RCU callbacks
4266          */
4267         if (elv_ioc_count_read(cfq_ioc_count))
4268                 wait_for_completion(&all_gone);
4269         ida_destroy(&cic_index_ida);
4270         cfq_slab_kill();
4271 }
4272
4273 module_init(cfq_init);
4274 module_exit(cfq_exit);
4275
4276 MODULE_AUTHOR("Jens Axboe");
4277 MODULE_LICENSE("GPL");
4278 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");