cfq-iosched: compact io_context radix_tree
[linux-2.6.git] / block / cfq-iosched.c
1 /*
2  *  CFQ, or complete fairness queueing, disk scheduler.
3  *
4  *  Based on ideas from a previously unfinished io
5  *  scheduler (round robin per-process disk scheduling) and Andrea Arcangeli.
6  *
7  *  Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
8  */
9 #include <linux/module.h>
10 #include <linux/slab.h>
11 #include <linux/blkdev.h>
12 #include <linux/elevator.h>
13 #include <linux/jiffies.h>
14 #include <linux/rbtree.h>
15 #include <linux/ioprio.h>
16 #include <linux/blktrace_api.h>
17 #include "blk-cgroup.h"
18
19 /*
20  * tunables
21  */
22 /* max queue in one round of service */
23 static const int cfq_quantum = 8;
24 static const int cfq_fifo_expire[2] = { HZ / 4, HZ / 8 };
25 /* maximum backwards seek, in KiB */
26 static const int cfq_back_max = 16 * 1024;
27 /* penalty of a backwards seek */
28 static const int cfq_back_penalty = 2;
29 static const int cfq_slice_sync = HZ / 10;
30 static int cfq_slice_async = HZ / 25;
31 static const int cfq_slice_async_rq = 2;
32 static int cfq_slice_idle = HZ / 125;
33 static const int cfq_target_latency = HZ * 3/10; /* 300 ms */
34 static const int cfq_hist_divisor = 4;
35
36 /*
37  * offset from end of service tree
38  */
39 #define CFQ_IDLE_DELAY          (HZ / 5)
40
41 /*
42  * below this threshold, we consider thinktime immediate
43  */
44 #define CFQ_MIN_TT              (2)
45
46 #define CFQ_SLICE_SCALE         (5)
47 #define CFQ_HW_QUEUE_MIN        (5)
48 #define CFQ_SERVICE_SHIFT       12
49
50 #define CFQQ_SEEK_THR           (sector_t)(8 * 100)
51 #define CFQQ_CLOSE_THR          (sector_t)(8 * 1024)
52 #define CFQQ_SECT_THR_NONROT    (sector_t)(2 * 32)
53 #define CFQQ_SEEKY(cfqq)        (hweight32(cfqq->seek_history) > 32/8)
54
55 #define RQ_CIC(rq)              \
56         ((struct cfq_io_context *) (rq)->elevator_private)
57 #define RQ_CFQQ(rq)             (struct cfq_queue *) ((rq)->elevator_private2)
58 #define RQ_CFQG(rq)             (struct cfq_group *) ((rq)->elevator_private3)
59
60 static struct kmem_cache *cfq_pool;
61 static struct kmem_cache *cfq_ioc_pool;
62
63 static DEFINE_PER_CPU(unsigned long, cfq_ioc_count);
64 static struct completion *ioc_gone;
65 static DEFINE_SPINLOCK(ioc_gone_lock);
66
67 static DEFINE_SPINLOCK(cic_index_lock);
68 static DEFINE_IDA(cic_index_ida);
69
70 #define CFQ_PRIO_LISTS          IOPRIO_BE_NR
71 #define cfq_class_idle(cfqq)    ((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
72 #define cfq_class_rt(cfqq)      ((cfqq)->ioprio_class == IOPRIO_CLASS_RT)
73
74 #define sample_valid(samples)   ((samples) > 80)
75 #define rb_entry_cfqg(node)     rb_entry((node), struct cfq_group, rb_node)
76
77 /*
78  * Most of our rbtree usage is for sorting with min extraction, so
79  * if we cache the leftmost node we don't have to walk down the tree
80  * to find it. Idea borrowed from Ingo Molnars CFS scheduler. We should
81  * move this into the elevator for the rq sorting as well.
82  */
83 struct cfq_rb_root {
84         struct rb_root rb;
85         struct rb_node *left;
86         unsigned count;
87         unsigned total_weight;
88         u64 min_vdisktime;
89         struct rb_node *active;
90 };
91 #define CFQ_RB_ROOT     (struct cfq_rb_root) { .rb = RB_ROOT, .left = NULL, \
92                         .count = 0, .min_vdisktime = 0, }
93
94 /*
95  * Per process-grouping structure
96  */
97 struct cfq_queue {
98         /* reference count */
99         atomic_t ref;
100         /* various state flags, see below */
101         unsigned int flags;
102         /* parent cfq_data */
103         struct cfq_data *cfqd;
104         /* service_tree member */
105         struct rb_node rb_node;
106         /* service_tree key */
107         unsigned long rb_key;
108         /* prio tree member */
109         struct rb_node p_node;
110         /* prio tree root we belong to, if any */
111         struct rb_root *p_root;
112         /* sorted list of pending requests */
113         struct rb_root sort_list;
114         /* if fifo isn't expired, next request to serve */
115         struct request *next_rq;
116         /* requests queued in sort_list */
117         int queued[2];
118         /* currently allocated requests */
119         int allocated[2];
120         /* fifo list of requests in sort_list */
121         struct list_head fifo;
122
123         /* time when queue got scheduled in to dispatch first request. */
124         unsigned long dispatch_start;
125         unsigned int allocated_slice;
126         unsigned int slice_dispatch;
127         /* time when first request from queue completed and slice started. */
128         unsigned long slice_start;
129         unsigned long slice_end;
130         long slice_resid;
131
132         /* pending metadata requests */
133         int meta_pending;
134         /* number of requests that are on the dispatch list or inside driver */
135         int dispatched;
136
137         /* io prio of this group */
138         unsigned short ioprio, org_ioprio;
139         unsigned short ioprio_class, org_ioprio_class;
140
141         pid_t pid;
142
143         u32 seek_history;
144         sector_t last_request_pos;
145
146         struct cfq_rb_root *service_tree;
147         struct cfq_queue *new_cfqq;
148         struct cfq_group *cfqg;
149         struct cfq_group *orig_cfqg;
150 };
151
152 /*
153  * First index in the service_trees.
154  * IDLE is handled separately, so it has negative index
155  */
156 enum wl_prio_t {
157         BE_WORKLOAD = 0,
158         RT_WORKLOAD = 1,
159         IDLE_WORKLOAD = 2,
160 };
161
162 /*
163  * Second index in the service_trees.
164  */
165 enum wl_type_t {
166         ASYNC_WORKLOAD = 0,
167         SYNC_NOIDLE_WORKLOAD = 1,
168         SYNC_WORKLOAD = 2
169 };
170
171 /* This is per cgroup per device grouping structure */
172 struct cfq_group {
173         /* group service_tree member */
174         struct rb_node rb_node;
175
176         /* group service_tree key */
177         u64 vdisktime;
178         unsigned int weight;
179         bool on_st;
180
181         /* number of cfqq currently on this group */
182         int nr_cfqq;
183
184         /* Per group busy queus average. Useful for workload slice calc. */
185         unsigned int busy_queues_avg[2];
186         /*
187          * rr lists of queues with requests, onle rr for each priority class.
188          * Counts are embedded in the cfq_rb_root
189          */
190         struct cfq_rb_root service_trees[2][3];
191         struct cfq_rb_root service_tree_idle;
192
193         unsigned long saved_workload_slice;
194         enum wl_type_t saved_workload;
195         enum wl_prio_t saved_serving_prio;
196         struct blkio_group blkg;
197 #ifdef CONFIG_CFQ_GROUP_IOSCHED
198         struct hlist_node cfqd_node;
199         atomic_t ref;
200 #endif
201 };
202
203 /*
204  * Per block device queue structure
205  */
206 struct cfq_data {
207         struct request_queue *queue;
208         /* Root service tree for cfq_groups */
209         struct cfq_rb_root grp_service_tree;
210         struct cfq_group root_group;
211
212         /*
213          * The priority currently being served
214          */
215         enum wl_prio_t serving_prio;
216         enum wl_type_t serving_type;
217         unsigned long workload_expires;
218         struct cfq_group *serving_group;
219         bool noidle_tree_requires_idle;
220
221         /*
222          * Each priority tree is sorted by next_request position.  These
223          * trees are used when determining if two or more queues are
224          * interleaving requests (see cfq_close_cooperator).
225          */
226         struct rb_root prio_trees[CFQ_PRIO_LISTS];
227
228         unsigned int busy_queues;
229
230         int rq_in_driver;
231         int rq_in_flight[2];
232
233         /*
234          * queue-depth detection
235          */
236         int rq_queued;
237         int hw_tag;
238         /*
239          * hw_tag can be
240          * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
241          *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
242          *  0 => no NCQ
243          */
244         int hw_tag_est_depth;
245         unsigned int hw_tag_samples;
246
247         /*
248          * idle window management
249          */
250         struct timer_list idle_slice_timer;
251         struct work_struct unplug_work;
252
253         struct cfq_queue *active_queue;
254         struct cfq_io_context *active_cic;
255
256         /*
257          * async queue for each priority case
258          */
259         struct cfq_queue *async_cfqq[2][IOPRIO_BE_NR];
260         struct cfq_queue *async_idle_cfqq;
261
262         sector_t last_position;
263
264         /*
265          * tunables, see top of file
266          */
267         unsigned int cfq_quantum;
268         unsigned int cfq_fifo_expire[2];
269         unsigned int cfq_back_penalty;
270         unsigned int cfq_back_max;
271         unsigned int cfq_slice[2];
272         unsigned int cfq_slice_async_rq;
273         unsigned int cfq_slice_idle;
274         unsigned int cfq_latency;
275         unsigned int cfq_group_isolation;
276
277         unsigned int cic_index;
278         struct list_head cic_list;
279
280         /*
281          * Fallback dummy cfqq for extreme OOM conditions
282          */
283         struct cfq_queue oom_cfqq;
284
285         unsigned long last_delayed_sync;
286
287         /* List of cfq groups being managed on this device*/
288         struct hlist_head cfqg_list;
289         struct rcu_head rcu;
290 };
291
292 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd);
293
294 static struct cfq_rb_root *service_tree_for(struct cfq_group *cfqg,
295                                             enum wl_prio_t prio,
296                                             enum wl_type_t type)
297 {
298         if (!cfqg)
299                 return NULL;
300
301         if (prio == IDLE_WORKLOAD)
302                 return &cfqg->service_tree_idle;
303
304         return &cfqg->service_trees[prio][type];
305 }
306
307 enum cfqq_state_flags {
308         CFQ_CFQQ_FLAG_on_rr = 0,        /* on round-robin busy list */
309         CFQ_CFQQ_FLAG_wait_request,     /* waiting for a request */
310         CFQ_CFQQ_FLAG_must_dispatch,    /* must be allowed a dispatch */
311         CFQ_CFQQ_FLAG_must_alloc_slice, /* per-slice must_alloc flag */
312         CFQ_CFQQ_FLAG_fifo_expire,      /* FIFO checked in this slice */
313         CFQ_CFQQ_FLAG_idle_window,      /* slice idling enabled */
314         CFQ_CFQQ_FLAG_prio_changed,     /* task priority has changed */
315         CFQ_CFQQ_FLAG_slice_new,        /* no requests dispatched in slice */
316         CFQ_CFQQ_FLAG_sync,             /* synchronous queue */
317         CFQ_CFQQ_FLAG_coop,             /* cfqq is shared */
318         CFQ_CFQQ_FLAG_split_coop,       /* shared cfqq will be splitted */
319         CFQ_CFQQ_FLAG_deep,             /* sync cfqq experienced large depth */
320         CFQ_CFQQ_FLAG_wait_busy,        /* Waiting for next request */
321 };
322
323 #define CFQ_CFQQ_FNS(name)                                              \
324 static inline void cfq_mark_cfqq_##name(struct cfq_queue *cfqq)         \
325 {                                                                       \
326         (cfqq)->flags |= (1 << CFQ_CFQQ_FLAG_##name);                   \
327 }                                                                       \
328 static inline void cfq_clear_cfqq_##name(struct cfq_queue *cfqq)        \
329 {                                                                       \
330         (cfqq)->flags &= ~(1 << CFQ_CFQQ_FLAG_##name);                  \
331 }                                                                       \
332 static inline int cfq_cfqq_##name(const struct cfq_queue *cfqq)         \
333 {                                                                       \
334         return ((cfqq)->flags & (1 << CFQ_CFQQ_FLAG_##name)) != 0;      \
335 }
336
337 CFQ_CFQQ_FNS(on_rr);
338 CFQ_CFQQ_FNS(wait_request);
339 CFQ_CFQQ_FNS(must_dispatch);
340 CFQ_CFQQ_FNS(must_alloc_slice);
341 CFQ_CFQQ_FNS(fifo_expire);
342 CFQ_CFQQ_FNS(idle_window);
343 CFQ_CFQQ_FNS(prio_changed);
344 CFQ_CFQQ_FNS(slice_new);
345 CFQ_CFQQ_FNS(sync);
346 CFQ_CFQQ_FNS(coop);
347 CFQ_CFQQ_FNS(split_coop);
348 CFQ_CFQQ_FNS(deep);
349 CFQ_CFQQ_FNS(wait_busy);
350 #undef CFQ_CFQQ_FNS
351
352 #ifdef CONFIG_CFQ_GROUP_IOSCHED
353 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
354         blk_add_trace_msg((cfqd)->queue, "cfq%d%c %s " fmt, (cfqq)->pid, \
355                         cfq_cfqq_sync((cfqq)) ? 'S' : 'A', \
356                         blkg_path(&(cfqq)->cfqg->blkg), ##args);
357
358 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)                          \
359         blk_add_trace_msg((cfqd)->queue, "%s " fmt,                     \
360                                 blkg_path(&(cfqg)->blkg), ##args);      \
361
362 #else
363 #define cfq_log_cfqq(cfqd, cfqq, fmt, args...)  \
364         blk_add_trace_msg((cfqd)->queue, "cfq%d " fmt, (cfqq)->pid, ##args)
365 #define cfq_log_cfqg(cfqd, cfqg, fmt, args...)          do {} while (0);
366 #endif
367 #define cfq_log(cfqd, fmt, args...)     \
368         blk_add_trace_msg((cfqd)->queue, "cfq " fmt, ##args)
369
370 /* Traverses through cfq group service trees */
371 #define for_each_cfqg_st(cfqg, i, j, st) \
372         for (i = 0; i <= IDLE_WORKLOAD; i++) \
373                 for (j = 0, st = i < IDLE_WORKLOAD ? &cfqg->service_trees[i][j]\
374                         : &cfqg->service_tree_idle; \
375                         (i < IDLE_WORKLOAD && j <= SYNC_WORKLOAD) || \
376                         (i == IDLE_WORKLOAD && j == 0); \
377                         j++, st = i < IDLE_WORKLOAD ? \
378                         &cfqg->service_trees[i][j]: NULL) \
379
380
381 static inline enum wl_prio_t cfqq_prio(struct cfq_queue *cfqq)
382 {
383         if (cfq_class_idle(cfqq))
384                 return IDLE_WORKLOAD;
385         if (cfq_class_rt(cfqq))
386                 return RT_WORKLOAD;
387         return BE_WORKLOAD;
388 }
389
390
391 static enum wl_type_t cfqq_type(struct cfq_queue *cfqq)
392 {
393         if (!cfq_cfqq_sync(cfqq))
394                 return ASYNC_WORKLOAD;
395         if (!cfq_cfqq_idle_window(cfqq))
396                 return SYNC_NOIDLE_WORKLOAD;
397         return SYNC_WORKLOAD;
398 }
399
400 static inline int cfq_group_busy_queues_wl(enum wl_prio_t wl,
401                                         struct cfq_data *cfqd,
402                                         struct cfq_group *cfqg)
403 {
404         if (wl == IDLE_WORKLOAD)
405                 return cfqg->service_tree_idle.count;
406
407         return cfqg->service_trees[wl][ASYNC_WORKLOAD].count
408                 + cfqg->service_trees[wl][SYNC_NOIDLE_WORKLOAD].count
409                 + cfqg->service_trees[wl][SYNC_WORKLOAD].count;
410 }
411
412 static inline int cfqg_busy_async_queues(struct cfq_data *cfqd,
413                                         struct cfq_group *cfqg)
414 {
415         return cfqg->service_trees[RT_WORKLOAD][ASYNC_WORKLOAD].count
416                 + cfqg->service_trees[BE_WORKLOAD][ASYNC_WORKLOAD].count;
417 }
418
419 static void cfq_dispatch_insert(struct request_queue *, struct request *);
420 static struct cfq_queue *cfq_get_queue(struct cfq_data *, bool,
421                                        struct io_context *, gfp_t);
422 static struct cfq_io_context *cfq_cic_lookup(struct cfq_data *,
423                                                 struct io_context *);
424
425 static inline struct cfq_queue *cic_to_cfqq(struct cfq_io_context *cic,
426                                             bool is_sync)
427 {
428         return cic->cfqq[is_sync];
429 }
430
431 static inline void cic_set_cfqq(struct cfq_io_context *cic,
432                                 struct cfq_queue *cfqq, bool is_sync)
433 {
434         cic->cfqq[is_sync] = cfqq;
435 }
436
437 #define CIC_DEAD_KEY    1ul
438 #define CIC_DEAD_INDEX_SHIFT    1
439
440 static inline void *cfqd_dead_key(struct cfq_data *cfqd)
441 {
442         return (void *)(cfqd->cic_index << CIC_DEAD_INDEX_SHIFT | CIC_DEAD_KEY);
443 }
444
445 static inline struct cfq_data *cic_to_cfqd(struct cfq_io_context *cic)
446 {
447         struct cfq_data *cfqd = cic->key;
448
449         if (unlikely((unsigned long) cfqd & CIC_DEAD_KEY))
450                 return NULL;
451
452         return cfqd;
453 }
454
455 /*
456  * We regard a request as SYNC, if it's either a read or has the SYNC bit
457  * set (in which case it could also be direct WRITE).
458  */
459 static inline bool cfq_bio_sync(struct bio *bio)
460 {
461         return bio_data_dir(bio) == READ || bio_rw_flagged(bio, BIO_RW_SYNCIO);
462 }
463
464 /*
465  * scheduler run of queue, if there are requests pending and no one in the
466  * driver that will restart queueing
467  */
468 static inline void cfq_schedule_dispatch(struct cfq_data *cfqd)
469 {
470         if (cfqd->busy_queues) {
471                 cfq_log(cfqd, "schedule dispatch");
472                 kblockd_schedule_work(cfqd->queue, &cfqd->unplug_work);
473         }
474 }
475
476 static int cfq_queue_empty(struct request_queue *q)
477 {
478         struct cfq_data *cfqd = q->elevator->elevator_data;
479
480         return !cfqd->rq_queued;
481 }
482
483 /*
484  * Scale schedule slice based on io priority. Use the sync time slice only
485  * if a queue is marked sync and has sync io queued. A sync queue with async
486  * io only, should not get full sync slice length.
487  */
488 static inline int cfq_prio_slice(struct cfq_data *cfqd, bool sync,
489                                  unsigned short prio)
490 {
491         const int base_slice = cfqd->cfq_slice[sync];
492
493         WARN_ON(prio >= IOPRIO_BE_NR);
494
495         return base_slice + (base_slice/CFQ_SLICE_SCALE * (4 - prio));
496 }
497
498 static inline int
499 cfq_prio_to_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
500 {
501         return cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio);
502 }
503
504 static inline u64 cfq_scale_slice(unsigned long delta, struct cfq_group *cfqg)
505 {
506         u64 d = delta << CFQ_SERVICE_SHIFT;
507
508         d = d * BLKIO_WEIGHT_DEFAULT;
509         do_div(d, cfqg->weight);
510         return d;
511 }
512
513 static inline u64 max_vdisktime(u64 min_vdisktime, u64 vdisktime)
514 {
515         s64 delta = (s64)(vdisktime - min_vdisktime);
516         if (delta > 0)
517                 min_vdisktime = vdisktime;
518
519         return min_vdisktime;
520 }
521
522 static inline u64 min_vdisktime(u64 min_vdisktime, u64 vdisktime)
523 {
524         s64 delta = (s64)(vdisktime - min_vdisktime);
525         if (delta < 0)
526                 min_vdisktime = vdisktime;
527
528         return min_vdisktime;
529 }
530
531 static void update_min_vdisktime(struct cfq_rb_root *st)
532 {
533         u64 vdisktime = st->min_vdisktime;
534         struct cfq_group *cfqg;
535
536         if (st->active) {
537                 cfqg = rb_entry_cfqg(st->active);
538                 vdisktime = cfqg->vdisktime;
539         }
540
541         if (st->left) {
542                 cfqg = rb_entry_cfqg(st->left);
543                 vdisktime = min_vdisktime(vdisktime, cfqg->vdisktime);
544         }
545
546         st->min_vdisktime = max_vdisktime(st->min_vdisktime, vdisktime);
547 }
548
549 /*
550  * get averaged number of queues of RT/BE priority.
551  * average is updated, with a formula that gives more weight to higher numbers,
552  * to quickly follows sudden increases and decrease slowly
553  */
554
555 static inline unsigned cfq_group_get_avg_queues(struct cfq_data *cfqd,
556                                         struct cfq_group *cfqg, bool rt)
557 {
558         unsigned min_q, max_q;
559         unsigned mult  = cfq_hist_divisor - 1;
560         unsigned round = cfq_hist_divisor / 2;
561         unsigned busy = cfq_group_busy_queues_wl(rt, cfqd, cfqg);
562
563         min_q = min(cfqg->busy_queues_avg[rt], busy);
564         max_q = max(cfqg->busy_queues_avg[rt], busy);
565         cfqg->busy_queues_avg[rt] = (mult * max_q + min_q + round) /
566                 cfq_hist_divisor;
567         return cfqg->busy_queues_avg[rt];
568 }
569
570 static inline unsigned
571 cfq_group_slice(struct cfq_data *cfqd, struct cfq_group *cfqg)
572 {
573         struct cfq_rb_root *st = &cfqd->grp_service_tree;
574
575         return cfq_target_latency * cfqg->weight / st->total_weight;
576 }
577
578 static inline void
579 cfq_set_prio_slice(struct cfq_data *cfqd, struct cfq_queue *cfqq)
580 {
581         unsigned slice = cfq_prio_to_slice(cfqd, cfqq);
582         if (cfqd->cfq_latency) {
583                 /*
584                  * interested queues (we consider only the ones with the same
585                  * priority class in the cfq group)
586                  */
587                 unsigned iq = cfq_group_get_avg_queues(cfqd, cfqq->cfqg,
588                                                 cfq_class_rt(cfqq));
589                 unsigned sync_slice = cfqd->cfq_slice[1];
590                 unsigned expect_latency = sync_slice * iq;
591                 unsigned group_slice = cfq_group_slice(cfqd, cfqq->cfqg);
592
593                 if (expect_latency > group_slice) {
594                         unsigned base_low_slice = 2 * cfqd->cfq_slice_idle;
595                         /* scale low_slice according to IO priority
596                          * and sync vs async */
597                         unsigned low_slice =
598                                 min(slice, base_low_slice * slice / sync_slice);
599                         /* the adapted slice value is scaled to fit all iqs
600                          * into the target latency */
601                         slice = max(slice * group_slice / expect_latency,
602                                     low_slice);
603                 }
604         }
605         cfqq->slice_start = jiffies;
606         cfqq->slice_end = jiffies + slice;
607         cfqq->allocated_slice = slice;
608         cfq_log_cfqq(cfqd, cfqq, "set_slice=%lu", cfqq->slice_end - jiffies);
609 }
610
611 /*
612  * We need to wrap this check in cfq_cfqq_slice_new(), since ->slice_end
613  * isn't valid until the first request from the dispatch is activated
614  * and the slice time set.
615  */
616 static inline bool cfq_slice_used(struct cfq_queue *cfqq)
617 {
618         if (cfq_cfqq_slice_new(cfqq))
619                 return 0;
620         if (time_before(jiffies, cfqq->slice_end))
621                 return 0;
622
623         return 1;
624 }
625
626 /*
627  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
628  * We choose the request that is closest to the head right now. Distance
629  * behind the head is penalized and only allowed to a certain extent.
630  */
631 static struct request *
632 cfq_choose_req(struct cfq_data *cfqd, struct request *rq1, struct request *rq2, sector_t last)
633 {
634         sector_t s1, s2, d1 = 0, d2 = 0;
635         unsigned long back_max;
636 #define CFQ_RQ1_WRAP    0x01 /* request 1 wraps */
637 #define CFQ_RQ2_WRAP    0x02 /* request 2 wraps */
638         unsigned wrap = 0; /* bit mask: requests behind the disk head? */
639
640         if (rq1 == NULL || rq1 == rq2)
641                 return rq2;
642         if (rq2 == NULL)
643                 return rq1;
644
645         if (rq_is_sync(rq1) && !rq_is_sync(rq2))
646                 return rq1;
647         else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
648                 return rq2;
649         if (rq_is_meta(rq1) && !rq_is_meta(rq2))
650                 return rq1;
651         else if (rq_is_meta(rq2) && !rq_is_meta(rq1))
652                 return rq2;
653
654         s1 = blk_rq_pos(rq1);
655         s2 = blk_rq_pos(rq2);
656
657         /*
658          * by definition, 1KiB is 2 sectors
659          */
660         back_max = cfqd->cfq_back_max * 2;
661
662         /*
663          * Strict one way elevator _except_ in the case where we allow
664          * short backward seeks which are biased as twice the cost of a
665          * similar forward seek.
666          */
667         if (s1 >= last)
668                 d1 = s1 - last;
669         else if (s1 + back_max >= last)
670                 d1 = (last - s1) * cfqd->cfq_back_penalty;
671         else
672                 wrap |= CFQ_RQ1_WRAP;
673
674         if (s2 >= last)
675                 d2 = s2 - last;
676         else if (s2 + back_max >= last)
677                 d2 = (last - s2) * cfqd->cfq_back_penalty;
678         else
679                 wrap |= CFQ_RQ2_WRAP;
680
681         /* Found required data */
682
683         /*
684          * By doing switch() on the bit mask "wrap" we avoid having to
685          * check two variables for all permutations: --> faster!
686          */
687         switch (wrap) {
688         case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
689                 if (d1 < d2)
690                         return rq1;
691                 else if (d2 < d1)
692                         return rq2;
693                 else {
694                         if (s1 >= s2)
695                                 return rq1;
696                         else
697                                 return rq2;
698                 }
699
700         case CFQ_RQ2_WRAP:
701                 return rq1;
702         case CFQ_RQ1_WRAP:
703                 return rq2;
704         case (CFQ_RQ1_WRAP|CFQ_RQ2_WRAP): /* both rqs wrapped */
705         default:
706                 /*
707                  * Since both rqs are wrapped,
708                  * start with the one that's further behind head
709                  * (--> only *one* back seek required),
710                  * since back seek takes more time than forward.
711                  */
712                 if (s1 <= s2)
713                         return rq1;
714                 else
715                         return rq2;
716         }
717 }
718
719 /*
720  * The below is leftmost cache rbtree addon
721  */
722 static struct cfq_queue *cfq_rb_first(struct cfq_rb_root *root)
723 {
724         /* Service tree is empty */
725         if (!root->count)
726                 return NULL;
727
728         if (!root->left)
729                 root->left = rb_first(&root->rb);
730
731         if (root->left)
732                 return rb_entry(root->left, struct cfq_queue, rb_node);
733
734         return NULL;
735 }
736
737 static struct cfq_group *cfq_rb_first_group(struct cfq_rb_root *root)
738 {
739         if (!root->left)
740                 root->left = rb_first(&root->rb);
741
742         if (root->left)
743                 return rb_entry_cfqg(root->left);
744
745         return NULL;
746 }
747
748 static void rb_erase_init(struct rb_node *n, struct rb_root *root)
749 {
750         rb_erase(n, root);
751         RB_CLEAR_NODE(n);
752 }
753
754 static void cfq_rb_erase(struct rb_node *n, struct cfq_rb_root *root)
755 {
756         if (root->left == n)
757                 root->left = NULL;
758         rb_erase_init(n, &root->rb);
759         --root->count;
760 }
761
762 /*
763  * would be nice to take fifo expire time into account as well
764  */
765 static struct request *
766 cfq_find_next_rq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
767                   struct request *last)
768 {
769         struct rb_node *rbnext = rb_next(&last->rb_node);
770         struct rb_node *rbprev = rb_prev(&last->rb_node);
771         struct request *next = NULL, *prev = NULL;
772
773         BUG_ON(RB_EMPTY_NODE(&last->rb_node));
774
775         if (rbprev)
776                 prev = rb_entry_rq(rbprev);
777
778         if (rbnext)
779                 next = rb_entry_rq(rbnext);
780         else {
781                 rbnext = rb_first(&cfqq->sort_list);
782                 if (rbnext && rbnext != &last->rb_node)
783                         next = rb_entry_rq(rbnext);
784         }
785
786         return cfq_choose_req(cfqd, next, prev, blk_rq_pos(last));
787 }
788
789 static unsigned long cfq_slice_offset(struct cfq_data *cfqd,
790                                       struct cfq_queue *cfqq)
791 {
792         /*
793          * just an approximation, should be ok.
794          */
795         return (cfqq->cfqg->nr_cfqq - 1) * (cfq_prio_slice(cfqd, 1, 0) -
796                        cfq_prio_slice(cfqd, cfq_cfqq_sync(cfqq), cfqq->ioprio));
797 }
798
799 static inline s64
800 cfqg_key(struct cfq_rb_root *st, struct cfq_group *cfqg)
801 {
802         return cfqg->vdisktime - st->min_vdisktime;
803 }
804
805 static void
806 __cfq_group_service_tree_add(struct cfq_rb_root *st, struct cfq_group *cfqg)
807 {
808         struct rb_node **node = &st->rb.rb_node;
809         struct rb_node *parent = NULL;
810         struct cfq_group *__cfqg;
811         s64 key = cfqg_key(st, cfqg);
812         int left = 1;
813
814         while (*node != NULL) {
815                 parent = *node;
816                 __cfqg = rb_entry_cfqg(parent);
817
818                 if (key < cfqg_key(st, __cfqg))
819                         node = &parent->rb_left;
820                 else {
821                         node = &parent->rb_right;
822                         left = 0;
823                 }
824         }
825
826         if (left)
827                 st->left = &cfqg->rb_node;
828
829         rb_link_node(&cfqg->rb_node, parent, node);
830         rb_insert_color(&cfqg->rb_node, &st->rb);
831 }
832
833 static void
834 cfq_group_service_tree_add(struct cfq_data *cfqd, struct cfq_group *cfqg)
835 {
836         struct cfq_rb_root *st = &cfqd->grp_service_tree;
837         struct cfq_group *__cfqg;
838         struct rb_node *n;
839
840         cfqg->nr_cfqq++;
841         if (cfqg->on_st)
842                 return;
843
844         /*
845          * Currently put the group at the end. Later implement something
846          * so that groups get lesser vtime based on their weights, so that
847          * if group does not loose all if it was not continously backlogged.
848          */
849         n = rb_last(&st->rb);
850         if (n) {
851                 __cfqg = rb_entry_cfqg(n);
852                 cfqg->vdisktime = __cfqg->vdisktime + CFQ_IDLE_DELAY;
853         } else
854                 cfqg->vdisktime = st->min_vdisktime;
855
856         __cfq_group_service_tree_add(st, cfqg);
857         cfqg->on_st = true;
858         st->total_weight += cfqg->weight;
859 }
860
861 static void
862 cfq_group_service_tree_del(struct cfq_data *cfqd, struct cfq_group *cfqg)
863 {
864         struct cfq_rb_root *st = &cfqd->grp_service_tree;
865
866         if (st->active == &cfqg->rb_node)
867                 st->active = NULL;
868
869         BUG_ON(cfqg->nr_cfqq < 1);
870         cfqg->nr_cfqq--;
871
872         /* If there are other cfq queues under this group, don't delete it */
873         if (cfqg->nr_cfqq)
874                 return;
875
876         cfq_log_cfqg(cfqd, cfqg, "del_from_rr group");
877         cfqg->on_st = false;
878         st->total_weight -= cfqg->weight;
879         if (!RB_EMPTY_NODE(&cfqg->rb_node))
880                 cfq_rb_erase(&cfqg->rb_node, st);
881         cfqg->saved_workload_slice = 0;
882         blkiocg_update_dequeue_stats(&cfqg->blkg, 1);
883 }
884
885 static inline unsigned int cfq_cfqq_slice_usage(struct cfq_queue *cfqq)
886 {
887         unsigned int slice_used;
888
889         /*
890          * Queue got expired before even a single request completed or
891          * got expired immediately after first request completion.
892          */
893         if (!cfqq->slice_start || cfqq->slice_start == jiffies) {
894                 /*
895                  * Also charge the seek time incurred to the group, otherwise
896                  * if there are mutiple queues in the group, each can dispatch
897                  * a single request on seeky media and cause lots of seek time
898                  * and group will never know it.
899                  */
900                 slice_used = max_t(unsigned, (jiffies - cfqq->dispatch_start),
901                                         1);
902         } else {
903                 slice_used = jiffies - cfqq->slice_start;
904                 if (slice_used > cfqq->allocated_slice)
905                         slice_used = cfqq->allocated_slice;
906         }
907
908         cfq_log_cfqq(cfqq->cfqd, cfqq, "sl_used=%u", slice_used);
909         return slice_used;
910 }
911
912 static void cfq_group_served(struct cfq_data *cfqd, struct cfq_group *cfqg,
913                                 struct cfq_queue *cfqq)
914 {
915         struct cfq_rb_root *st = &cfqd->grp_service_tree;
916         unsigned int used_sl, charge_sl;
917         int nr_sync = cfqg->nr_cfqq - cfqg_busy_async_queues(cfqd, cfqg)
918                         - cfqg->service_tree_idle.count;
919
920         BUG_ON(nr_sync < 0);
921         used_sl = charge_sl = cfq_cfqq_slice_usage(cfqq);
922
923         if (!cfq_cfqq_sync(cfqq) && !nr_sync)
924                 charge_sl = cfqq->allocated_slice;
925
926         /* Can't update vdisktime while group is on service tree */
927         cfq_rb_erase(&cfqg->rb_node, st);
928         cfqg->vdisktime += cfq_scale_slice(charge_sl, cfqg);
929         __cfq_group_service_tree_add(st, cfqg);
930
931         /* This group is being expired. Save the context */
932         if (time_after(cfqd->workload_expires, jiffies)) {
933                 cfqg->saved_workload_slice = cfqd->workload_expires
934                                                 - jiffies;
935                 cfqg->saved_workload = cfqd->serving_type;
936                 cfqg->saved_serving_prio = cfqd->serving_prio;
937         } else
938                 cfqg->saved_workload_slice = 0;
939
940         cfq_log_cfqg(cfqd, cfqg, "served: vt=%llu min_vt=%llu", cfqg->vdisktime,
941                                         st->min_vdisktime);
942         blkiocg_update_timeslice_used(&cfqg->blkg, used_sl);
943         blkiocg_set_start_empty_time(&cfqg->blkg);
944 }
945
946 #ifdef CONFIG_CFQ_GROUP_IOSCHED
947 static inline struct cfq_group *cfqg_of_blkg(struct blkio_group *blkg)
948 {
949         if (blkg)
950                 return container_of(blkg, struct cfq_group, blkg);
951         return NULL;
952 }
953
954 void
955 cfq_update_blkio_group_weight(struct blkio_group *blkg, unsigned int weight)
956 {
957         cfqg_of_blkg(blkg)->weight = weight;
958 }
959
960 static struct cfq_group *
961 cfq_find_alloc_cfqg(struct cfq_data *cfqd, struct cgroup *cgroup, int create)
962 {
963         struct blkio_cgroup *blkcg = cgroup_to_blkio_cgroup(cgroup);
964         struct cfq_group *cfqg = NULL;
965         void *key = cfqd;
966         int i, j;
967         struct cfq_rb_root *st;
968         struct backing_dev_info *bdi = &cfqd->queue->backing_dev_info;
969         unsigned int major, minor;
970
971         cfqg = cfqg_of_blkg(blkiocg_lookup_group(blkcg, key));
972         if (cfqg && !cfqg->blkg.dev && bdi->dev && dev_name(bdi->dev)) {
973                 sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
974                 cfqg->blkg.dev = MKDEV(major, minor);
975                 goto done;
976         }
977         if (cfqg || !create)
978                 goto done;
979
980         cfqg = kzalloc_node(sizeof(*cfqg), GFP_ATOMIC, cfqd->queue->node);
981         if (!cfqg)
982                 goto done;
983
984         for_each_cfqg_st(cfqg, i, j, st)
985                 *st = CFQ_RB_ROOT;
986         RB_CLEAR_NODE(&cfqg->rb_node);
987
988         /*
989          * Take the initial reference that will be released on destroy
990          * This can be thought of a joint reference by cgroup and
991          * elevator which will be dropped by either elevator exit
992          * or cgroup deletion path depending on who is exiting first.
993          */
994         atomic_set(&cfqg->ref, 1);
995
996         /* Add group onto cgroup list */
997         sscanf(dev_name(bdi->dev), "%u:%u", &major, &minor);
998         blkiocg_add_blkio_group(blkcg, &cfqg->blkg, (void *)cfqd,
999                                         MKDEV(major, minor));
1000         cfqg->weight = blkcg_get_weight(blkcg, cfqg->blkg.dev);
1001
1002         /* Add group on cfqd list */
1003         hlist_add_head(&cfqg->cfqd_node, &cfqd->cfqg_list);
1004
1005 done:
1006         return cfqg;
1007 }
1008
1009 /*
1010  * Search for the cfq group current task belongs to. If create = 1, then also
1011  * create the cfq group if it does not exist. request_queue lock must be held.
1012  */
1013 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1014 {
1015         struct cgroup *cgroup;
1016         struct cfq_group *cfqg = NULL;
1017
1018         rcu_read_lock();
1019         cgroup = task_cgroup(current, blkio_subsys_id);
1020         cfqg = cfq_find_alloc_cfqg(cfqd, cgroup, create);
1021         if (!cfqg && create)
1022                 cfqg = &cfqd->root_group;
1023         rcu_read_unlock();
1024         return cfqg;
1025 }
1026
1027 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1028 {
1029         atomic_inc(&cfqg->ref);
1030         return cfqg;
1031 }
1032
1033 static void cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg)
1034 {
1035         /* Currently, all async queues are mapped to root group */
1036         if (!cfq_cfqq_sync(cfqq))
1037                 cfqg = &cfqq->cfqd->root_group;
1038
1039         cfqq->cfqg = cfqg;
1040         /* cfqq reference on cfqg */
1041         atomic_inc(&cfqq->cfqg->ref);
1042 }
1043
1044 static void cfq_put_cfqg(struct cfq_group *cfqg)
1045 {
1046         struct cfq_rb_root *st;
1047         int i, j;
1048
1049         BUG_ON(atomic_read(&cfqg->ref) <= 0);
1050         if (!atomic_dec_and_test(&cfqg->ref))
1051                 return;
1052         for_each_cfqg_st(cfqg, i, j, st)
1053                 BUG_ON(!RB_EMPTY_ROOT(&st->rb) || st->active != NULL);
1054         kfree(cfqg);
1055 }
1056
1057 static void cfq_destroy_cfqg(struct cfq_data *cfqd, struct cfq_group *cfqg)
1058 {
1059         /* Something wrong if we are trying to remove same group twice */
1060         BUG_ON(hlist_unhashed(&cfqg->cfqd_node));
1061
1062         hlist_del_init(&cfqg->cfqd_node);
1063
1064         /*
1065          * Put the reference taken at the time of creation so that when all
1066          * queues are gone, group can be destroyed.
1067          */
1068         cfq_put_cfqg(cfqg);
1069 }
1070
1071 static void cfq_release_cfq_groups(struct cfq_data *cfqd)
1072 {
1073         struct hlist_node *pos, *n;
1074         struct cfq_group *cfqg;
1075
1076         hlist_for_each_entry_safe(cfqg, pos, n, &cfqd->cfqg_list, cfqd_node) {
1077                 /*
1078                  * If cgroup removal path got to blk_group first and removed
1079                  * it from cgroup list, then it will take care of destroying
1080                  * cfqg also.
1081                  */
1082                 if (!blkiocg_del_blkio_group(&cfqg->blkg))
1083                         cfq_destroy_cfqg(cfqd, cfqg);
1084         }
1085 }
1086
1087 /*
1088  * Blk cgroup controller notification saying that blkio_group object is being
1089  * delinked as associated cgroup object is going away. That also means that
1090  * no new IO will come in this group. So get rid of this group as soon as
1091  * any pending IO in the group is finished.
1092  *
1093  * This function is called under rcu_read_lock(). key is the rcu protected
1094  * pointer. That means "key" is a valid cfq_data pointer as long as we are rcu
1095  * read lock.
1096  *
1097  * "key" was fetched from blkio_group under blkio_cgroup->lock. That means
1098  * it should not be NULL as even if elevator was exiting, cgroup deltion
1099  * path got to it first.
1100  */
1101 void cfq_unlink_blkio_group(void *key, struct blkio_group *blkg)
1102 {
1103         unsigned long  flags;
1104         struct cfq_data *cfqd = key;
1105
1106         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
1107         cfq_destroy_cfqg(cfqd, cfqg_of_blkg(blkg));
1108         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
1109 }
1110
1111 #else /* GROUP_IOSCHED */
1112 static struct cfq_group *cfq_get_cfqg(struct cfq_data *cfqd, int create)
1113 {
1114         return &cfqd->root_group;
1115 }
1116
1117 static inline struct cfq_group *cfq_ref_get_cfqg(struct cfq_group *cfqg)
1118 {
1119         return cfqg;
1120 }
1121
1122 static inline void
1123 cfq_link_cfqq_cfqg(struct cfq_queue *cfqq, struct cfq_group *cfqg) {
1124         cfqq->cfqg = cfqg;
1125 }
1126
1127 static void cfq_release_cfq_groups(struct cfq_data *cfqd) {}
1128 static inline void cfq_put_cfqg(struct cfq_group *cfqg) {}
1129
1130 #endif /* GROUP_IOSCHED */
1131
1132 /*
1133  * The cfqd->service_trees holds all pending cfq_queue's that have
1134  * requests waiting to be processed. It is sorted in the order that
1135  * we will service the queues.
1136  */
1137 static void cfq_service_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1138                                  bool add_front)
1139 {
1140         struct rb_node **p, *parent;
1141         struct cfq_queue *__cfqq;
1142         unsigned long rb_key;
1143         struct cfq_rb_root *service_tree;
1144         int left;
1145         int new_cfqq = 1;
1146         int group_changed = 0;
1147
1148 #ifdef CONFIG_CFQ_GROUP_IOSCHED
1149         if (!cfqd->cfq_group_isolation
1150             && cfqq_type(cfqq) == SYNC_NOIDLE_WORKLOAD
1151             && cfqq->cfqg && cfqq->cfqg != &cfqd->root_group) {
1152                 /* Move this cfq to root group */
1153                 cfq_log_cfqq(cfqd, cfqq, "moving to root group");
1154                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1155                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1156                 cfqq->orig_cfqg = cfqq->cfqg;
1157                 cfqq->cfqg = &cfqd->root_group;
1158                 atomic_inc(&cfqd->root_group.ref);
1159                 group_changed = 1;
1160         } else if (!cfqd->cfq_group_isolation
1161                    && cfqq_type(cfqq) == SYNC_WORKLOAD && cfqq->orig_cfqg) {
1162                 /* cfqq is sequential now needs to go to its original group */
1163                 BUG_ON(cfqq->cfqg != &cfqd->root_group);
1164                 if (!RB_EMPTY_NODE(&cfqq->rb_node))
1165                         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1166                 cfq_put_cfqg(cfqq->cfqg);
1167                 cfqq->cfqg = cfqq->orig_cfqg;
1168                 cfqq->orig_cfqg = NULL;
1169                 group_changed = 1;
1170                 cfq_log_cfqq(cfqd, cfqq, "moved to origin group");
1171         }
1172 #endif
1173
1174         service_tree = service_tree_for(cfqq->cfqg, cfqq_prio(cfqq),
1175                                                 cfqq_type(cfqq));
1176         if (cfq_class_idle(cfqq)) {
1177                 rb_key = CFQ_IDLE_DELAY;
1178                 parent = rb_last(&service_tree->rb);
1179                 if (parent && parent != &cfqq->rb_node) {
1180                         __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1181                         rb_key += __cfqq->rb_key;
1182                 } else
1183                         rb_key += jiffies;
1184         } else if (!add_front) {
1185                 /*
1186                  * Get our rb key offset. Subtract any residual slice
1187                  * value carried from last service. A negative resid
1188                  * count indicates slice overrun, and this should position
1189                  * the next service time further away in the tree.
1190                  */
1191                 rb_key = cfq_slice_offset(cfqd, cfqq) + jiffies;
1192                 rb_key -= cfqq->slice_resid;
1193                 cfqq->slice_resid = 0;
1194         } else {
1195                 rb_key = -HZ;
1196                 __cfqq = cfq_rb_first(service_tree);
1197                 rb_key += __cfqq ? __cfqq->rb_key : jiffies;
1198         }
1199
1200         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1201                 new_cfqq = 0;
1202                 /*
1203                  * same position, nothing more to do
1204                  */
1205                 if (rb_key == cfqq->rb_key &&
1206                     cfqq->service_tree == service_tree)
1207                         return;
1208
1209                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1210                 cfqq->service_tree = NULL;
1211         }
1212
1213         left = 1;
1214         parent = NULL;
1215         cfqq->service_tree = service_tree;
1216         p = &service_tree->rb.rb_node;
1217         while (*p) {
1218                 struct rb_node **n;
1219
1220                 parent = *p;
1221                 __cfqq = rb_entry(parent, struct cfq_queue, rb_node);
1222
1223                 /*
1224                  * sort by key, that represents service time.
1225                  */
1226                 if (time_before(rb_key, __cfqq->rb_key))
1227                         n = &(*p)->rb_left;
1228                 else {
1229                         n = &(*p)->rb_right;
1230                         left = 0;
1231                 }
1232
1233                 p = n;
1234         }
1235
1236         if (left)
1237                 service_tree->left = &cfqq->rb_node;
1238
1239         cfqq->rb_key = rb_key;
1240         rb_link_node(&cfqq->rb_node, parent, p);
1241         rb_insert_color(&cfqq->rb_node, &service_tree->rb);
1242         service_tree->count++;
1243         if ((add_front || !new_cfqq) && !group_changed)
1244                 return;
1245         cfq_group_service_tree_add(cfqd, cfqq->cfqg);
1246 }
1247
1248 static struct cfq_queue *
1249 cfq_prio_tree_lookup(struct cfq_data *cfqd, struct rb_root *root,
1250                      sector_t sector, struct rb_node **ret_parent,
1251                      struct rb_node ***rb_link)
1252 {
1253         struct rb_node **p, *parent;
1254         struct cfq_queue *cfqq = NULL;
1255
1256         parent = NULL;
1257         p = &root->rb_node;
1258         while (*p) {
1259                 struct rb_node **n;
1260
1261                 parent = *p;
1262                 cfqq = rb_entry(parent, struct cfq_queue, p_node);
1263
1264                 /*
1265                  * Sort strictly based on sector.  Smallest to the left,
1266                  * largest to the right.
1267                  */
1268                 if (sector > blk_rq_pos(cfqq->next_rq))
1269                         n = &(*p)->rb_right;
1270                 else if (sector < blk_rq_pos(cfqq->next_rq))
1271                         n = &(*p)->rb_left;
1272                 else
1273                         break;
1274                 p = n;
1275                 cfqq = NULL;
1276         }
1277
1278         *ret_parent = parent;
1279         if (rb_link)
1280                 *rb_link = p;
1281         return cfqq;
1282 }
1283
1284 static void cfq_prio_tree_add(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1285 {
1286         struct rb_node **p, *parent;
1287         struct cfq_queue *__cfqq;
1288
1289         if (cfqq->p_root) {
1290                 rb_erase(&cfqq->p_node, cfqq->p_root);
1291                 cfqq->p_root = NULL;
1292         }
1293
1294         if (cfq_class_idle(cfqq))
1295                 return;
1296         if (!cfqq->next_rq)
1297                 return;
1298
1299         cfqq->p_root = &cfqd->prio_trees[cfqq->org_ioprio];
1300         __cfqq = cfq_prio_tree_lookup(cfqd, cfqq->p_root,
1301                                       blk_rq_pos(cfqq->next_rq), &parent, &p);
1302         if (!__cfqq) {
1303                 rb_link_node(&cfqq->p_node, parent, p);
1304                 rb_insert_color(&cfqq->p_node, cfqq->p_root);
1305         } else
1306                 cfqq->p_root = NULL;
1307 }
1308
1309 /*
1310  * Update cfqq's position in the service tree.
1311  */
1312 static void cfq_resort_rr_list(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1313 {
1314         /*
1315          * Resorting requires the cfqq to be on the RR list already.
1316          */
1317         if (cfq_cfqq_on_rr(cfqq)) {
1318                 cfq_service_tree_add(cfqd, cfqq, 0);
1319                 cfq_prio_tree_add(cfqd, cfqq);
1320         }
1321 }
1322
1323 /*
1324  * add to busy list of queues for service, trying to be fair in ordering
1325  * the pending list according to last request service
1326  */
1327 static void cfq_add_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1328 {
1329         cfq_log_cfqq(cfqd, cfqq, "add_to_rr");
1330         BUG_ON(cfq_cfqq_on_rr(cfqq));
1331         cfq_mark_cfqq_on_rr(cfqq);
1332         cfqd->busy_queues++;
1333
1334         cfq_resort_rr_list(cfqd, cfqq);
1335 }
1336
1337 /*
1338  * Called when the cfqq no longer has requests pending, remove it from
1339  * the service tree.
1340  */
1341 static void cfq_del_cfqq_rr(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1342 {
1343         cfq_log_cfqq(cfqd, cfqq, "del_from_rr");
1344         BUG_ON(!cfq_cfqq_on_rr(cfqq));
1345         cfq_clear_cfqq_on_rr(cfqq);
1346
1347         if (!RB_EMPTY_NODE(&cfqq->rb_node)) {
1348                 cfq_rb_erase(&cfqq->rb_node, cfqq->service_tree);
1349                 cfqq->service_tree = NULL;
1350         }
1351         if (cfqq->p_root) {
1352                 rb_erase(&cfqq->p_node, cfqq->p_root);
1353                 cfqq->p_root = NULL;
1354         }
1355
1356         cfq_group_service_tree_del(cfqd, cfqq->cfqg);
1357         BUG_ON(!cfqd->busy_queues);
1358         cfqd->busy_queues--;
1359 }
1360
1361 /*
1362  * rb tree support functions
1363  */
1364 static void cfq_del_rq_rb(struct request *rq)
1365 {
1366         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1367         const int sync = rq_is_sync(rq);
1368
1369         BUG_ON(!cfqq->queued[sync]);
1370         cfqq->queued[sync]--;
1371
1372         elv_rb_del(&cfqq->sort_list, rq);
1373
1374         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list)) {
1375                 /*
1376                  * Queue will be deleted from service tree when we actually
1377                  * expire it later. Right now just remove it from prio tree
1378                  * as it is empty.
1379                  */
1380                 if (cfqq->p_root) {
1381                         rb_erase(&cfqq->p_node, cfqq->p_root);
1382                         cfqq->p_root = NULL;
1383                 }
1384         }
1385 }
1386
1387 static void cfq_add_rq_rb(struct request *rq)
1388 {
1389         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1390         struct cfq_data *cfqd = cfqq->cfqd;
1391         struct request *__alias, *prev;
1392
1393         cfqq->queued[rq_is_sync(rq)]++;
1394
1395         /*
1396          * looks a little odd, but the first insert might return an alias.
1397          * if that happens, put the alias on the dispatch list
1398          */
1399         while ((__alias = elv_rb_add(&cfqq->sort_list, rq)) != NULL)
1400                 cfq_dispatch_insert(cfqd->queue, __alias);
1401
1402         if (!cfq_cfqq_on_rr(cfqq))
1403                 cfq_add_cfqq_rr(cfqd, cfqq);
1404
1405         /*
1406          * check if this request is a better next-serve candidate
1407          */
1408         prev = cfqq->next_rq;
1409         cfqq->next_rq = cfq_choose_req(cfqd, cfqq->next_rq, rq, cfqd->last_position);
1410
1411         /*
1412          * adjust priority tree position, if ->next_rq changes
1413          */
1414         if (prev != cfqq->next_rq)
1415                 cfq_prio_tree_add(cfqd, cfqq);
1416
1417         BUG_ON(!cfqq->next_rq);
1418 }
1419
1420 static void cfq_reposition_rq_rb(struct cfq_queue *cfqq, struct request *rq)
1421 {
1422         elv_rb_del(&cfqq->sort_list, rq);
1423         cfqq->queued[rq_is_sync(rq)]--;
1424         blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
1425                                                 rq_is_sync(rq));
1426         cfq_add_rq_rb(rq);
1427         blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
1428                         &cfqq->cfqd->serving_group->blkg, rq_data_dir(rq),
1429                         rq_is_sync(rq));
1430 }
1431
1432 static struct request *
1433 cfq_find_rq_fmerge(struct cfq_data *cfqd, struct bio *bio)
1434 {
1435         struct task_struct *tsk = current;
1436         struct cfq_io_context *cic;
1437         struct cfq_queue *cfqq;
1438
1439         cic = cfq_cic_lookup(cfqd, tsk->io_context);
1440         if (!cic)
1441                 return NULL;
1442
1443         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1444         if (cfqq) {
1445                 sector_t sector = bio->bi_sector + bio_sectors(bio);
1446
1447                 return elv_rb_find(&cfqq->sort_list, sector);
1448         }
1449
1450         return NULL;
1451 }
1452
1453 static void cfq_activate_request(struct request_queue *q, struct request *rq)
1454 {
1455         struct cfq_data *cfqd = q->elevator->elevator_data;
1456
1457         cfqd->rq_in_driver++;
1458         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "activate rq, drv=%d",
1459                                                 cfqd->rq_in_driver);
1460
1461         cfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
1462 }
1463
1464 static void cfq_deactivate_request(struct request_queue *q, struct request *rq)
1465 {
1466         struct cfq_data *cfqd = q->elevator->elevator_data;
1467
1468         WARN_ON(!cfqd->rq_in_driver);
1469         cfqd->rq_in_driver--;
1470         cfq_log_cfqq(cfqd, RQ_CFQQ(rq), "deactivate rq, drv=%d",
1471                                                 cfqd->rq_in_driver);
1472 }
1473
1474 static void cfq_remove_request(struct request *rq)
1475 {
1476         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1477
1478         if (cfqq->next_rq == rq)
1479                 cfqq->next_rq = cfq_find_next_rq(cfqq->cfqd, cfqq, rq);
1480
1481         list_del_init(&rq->queuelist);
1482         cfq_del_rq_rb(rq);
1483
1484         cfqq->cfqd->rq_queued--;
1485         blkiocg_update_io_remove_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(rq),
1486                                                 rq_is_sync(rq));
1487         if (rq_is_meta(rq)) {
1488                 WARN_ON(!cfqq->meta_pending);
1489                 cfqq->meta_pending--;
1490         }
1491 }
1492
1493 static int cfq_merge(struct request_queue *q, struct request **req,
1494                      struct bio *bio)
1495 {
1496         struct cfq_data *cfqd = q->elevator->elevator_data;
1497         struct request *__rq;
1498
1499         __rq = cfq_find_rq_fmerge(cfqd, bio);
1500         if (__rq && elv_rq_merge_ok(__rq, bio)) {
1501                 *req = __rq;
1502                 return ELEVATOR_FRONT_MERGE;
1503         }
1504
1505         return ELEVATOR_NO_MERGE;
1506 }
1507
1508 static void cfq_merged_request(struct request_queue *q, struct request *req,
1509                                int type)
1510 {
1511         if (type == ELEVATOR_FRONT_MERGE) {
1512                 struct cfq_queue *cfqq = RQ_CFQQ(req);
1513
1514                 cfq_reposition_rq_rb(cfqq, req);
1515         }
1516 }
1517
1518 static void cfq_bio_merged(struct request_queue *q, struct request *req,
1519                                 struct bio *bio)
1520 {
1521         blkiocg_update_io_merged_stats(&(RQ_CFQG(req))->blkg, bio_data_dir(bio),
1522                                         cfq_bio_sync(bio));
1523 }
1524
1525 static void
1526 cfq_merged_requests(struct request_queue *q, struct request *rq,
1527                     struct request *next)
1528 {
1529         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1530         /*
1531          * reposition in fifo if next is older than rq
1532          */
1533         if (!list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
1534             time_before(rq_fifo_time(next), rq_fifo_time(rq))) {
1535                 list_move(&rq->queuelist, &next->queuelist);
1536                 rq_set_fifo_time(rq, rq_fifo_time(next));
1537         }
1538
1539         if (cfqq->next_rq == next)
1540                 cfqq->next_rq = rq;
1541         cfq_remove_request(next);
1542         blkiocg_update_io_merged_stats(&(RQ_CFQG(rq))->blkg, rq_data_dir(next),
1543                                         rq_is_sync(next));
1544 }
1545
1546 static int cfq_allow_merge(struct request_queue *q, struct request *rq,
1547                            struct bio *bio)
1548 {
1549         struct cfq_data *cfqd = q->elevator->elevator_data;
1550         struct cfq_io_context *cic;
1551         struct cfq_queue *cfqq;
1552
1553         /*
1554          * Disallow merge of a sync bio into an async request.
1555          */
1556         if (cfq_bio_sync(bio) && !rq_is_sync(rq))
1557                 return false;
1558
1559         /*
1560          * Lookup the cfqq that this bio will be queued with. Allow
1561          * merge only if rq is queued there.
1562          */
1563         cic = cfq_cic_lookup(cfqd, current->io_context);
1564         if (!cic)
1565                 return false;
1566
1567         cfqq = cic_to_cfqq(cic, cfq_bio_sync(bio));
1568         return cfqq == RQ_CFQQ(rq);
1569 }
1570
1571 static inline void cfq_del_timer(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1572 {
1573         del_timer(&cfqd->idle_slice_timer);
1574         blkiocg_update_idle_time_stats(&cfqq->cfqg->blkg);
1575 }
1576
1577 static void __cfq_set_active_queue(struct cfq_data *cfqd,
1578                                    struct cfq_queue *cfqq)
1579 {
1580         if (cfqq) {
1581                 cfq_log_cfqq(cfqd, cfqq, "set_active wl_prio:%d wl_type:%d",
1582                                 cfqd->serving_prio, cfqd->serving_type);
1583                 blkiocg_update_avg_queue_size_stats(&cfqq->cfqg->blkg);
1584                 cfqq->slice_start = 0;
1585                 cfqq->dispatch_start = jiffies;
1586                 cfqq->allocated_slice = 0;
1587                 cfqq->slice_end = 0;
1588                 cfqq->slice_dispatch = 0;
1589
1590                 cfq_clear_cfqq_wait_request(cfqq);
1591                 cfq_clear_cfqq_must_dispatch(cfqq);
1592                 cfq_clear_cfqq_must_alloc_slice(cfqq);
1593                 cfq_clear_cfqq_fifo_expire(cfqq);
1594                 cfq_mark_cfqq_slice_new(cfqq);
1595
1596                 cfq_del_timer(cfqd, cfqq);
1597         }
1598
1599         cfqd->active_queue = cfqq;
1600 }
1601
1602 /*
1603  * current cfqq expired its slice (or was too idle), select new one
1604  */
1605 static void
1606 __cfq_slice_expired(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1607                     bool timed_out)
1608 {
1609         cfq_log_cfqq(cfqd, cfqq, "slice expired t=%d", timed_out);
1610
1611         if (cfq_cfqq_wait_request(cfqq))
1612                 cfq_del_timer(cfqd, cfqq);
1613
1614         cfq_clear_cfqq_wait_request(cfqq);
1615         cfq_clear_cfqq_wait_busy(cfqq);
1616
1617         /*
1618          * If this cfqq is shared between multiple processes, check to
1619          * make sure that those processes are still issuing I/Os within
1620          * the mean seek distance.  If not, it may be time to break the
1621          * queues apart again.
1622          */
1623         if (cfq_cfqq_coop(cfqq) && CFQQ_SEEKY(cfqq))
1624                 cfq_mark_cfqq_split_coop(cfqq);
1625
1626         /*
1627          * store what was left of this slice, if the queue idled/timed out
1628          */
1629         if (timed_out && !cfq_cfqq_slice_new(cfqq)) {
1630                 cfqq->slice_resid = cfqq->slice_end - jiffies;
1631                 cfq_log_cfqq(cfqd, cfqq, "resid=%ld", cfqq->slice_resid);
1632         }
1633
1634         cfq_group_served(cfqd, cfqq->cfqg, cfqq);
1635
1636         if (cfq_cfqq_on_rr(cfqq) && RB_EMPTY_ROOT(&cfqq->sort_list))
1637                 cfq_del_cfqq_rr(cfqd, cfqq);
1638
1639         cfq_resort_rr_list(cfqd, cfqq);
1640
1641         if (cfqq == cfqd->active_queue)
1642                 cfqd->active_queue = NULL;
1643
1644         if (&cfqq->cfqg->rb_node == cfqd->grp_service_tree.active)
1645                 cfqd->grp_service_tree.active = NULL;
1646
1647         if (cfqd->active_cic) {
1648                 put_io_context(cfqd->active_cic->ioc);
1649                 cfqd->active_cic = NULL;
1650         }
1651 }
1652
1653 static inline void cfq_slice_expired(struct cfq_data *cfqd, bool timed_out)
1654 {
1655         struct cfq_queue *cfqq = cfqd->active_queue;
1656
1657         if (cfqq)
1658                 __cfq_slice_expired(cfqd, cfqq, timed_out);
1659 }
1660
1661 /*
1662  * Get next queue for service. Unless we have a queue preemption,
1663  * we'll simply select the first cfqq in the service tree.
1664  */
1665 static struct cfq_queue *cfq_get_next_queue(struct cfq_data *cfqd)
1666 {
1667         struct cfq_rb_root *service_tree =
1668                 service_tree_for(cfqd->serving_group, cfqd->serving_prio,
1669                                         cfqd->serving_type);
1670
1671         if (!cfqd->rq_queued)
1672                 return NULL;
1673
1674         /* There is nothing to dispatch */
1675         if (!service_tree)
1676                 return NULL;
1677         if (RB_EMPTY_ROOT(&service_tree->rb))
1678                 return NULL;
1679         return cfq_rb_first(service_tree);
1680 }
1681
1682 static struct cfq_queue *cfq_get_next_queue_forced(struct cfq_data *cfqd)
1683 {
1684         struct cfq_group *cfqg;
1685         struct cfq_queue *cfqq;
1686         int i, j;
1687         struct cfq_rb_root *st;
1688
1689         if (!cfqd->rq_queued)
1690                 return NULL;
1691
1692         cfqg = cfq_get_next_cfqg(cfqd);
1693         if (!cfqg)
1694                 return NULL;
1695
1696         for_each_cfqg_st(cfqg, i, j, st)
1697                 if ((cfqq = cfq_rb_first(st)) != NULL)
1698                         return cfqq;
1699         return NULL;
1700 }
1701
1702 /*
1703  * Get and set a new active queue for service.
1704  */
1705 static struct cfq_queue *cfq_set_active_queue(struct cfq_data *cfqd,
1706                                               struct cfq_queue *cfqq)
1707 {
1708         if (!cfqq)
1709                 cfqq = cfq_get_next_queue(cfqd);
1710
1711         __cfq_set_active_queue(cfqd, cfqq);
1712         return cfqq;
1713 }
1714
1715 static inline sector_t cfq_dist_from_last(struct cfq_data *cfqd,
1716                                           struct request *rq)
1717 {
1718         if (blk_rq_pos(rq) >= cfqd->last_position)
1719                 return blk_rq_pos(rq) - cfqd->last_position;
1720         else
1721                 return cfqd->last_position - blk_rq_pos(rq);
1722 }
1723
1724 static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
1725                                struct request *rq)
1726 {
1727         return cfq_dist_from_last(cfqd, rq) <= CFQQ_CLOSE_THR;
1728 }
1729
1730 static struct cfq_queue *cfqq_close(struct cfq_data *cfqd,
1731                                     struct cfq_queue *cur_cfqq)
1732 {
1733         struct rb_root *root = &cfqd->prio_trees[cur_cfqq->org_ioprio];
1734         struct rb_node *parent, *node;
1735         struct cfq_queue *__cfqq;
1736         sector_t sector = cfqd->last_position;
1737
1738         if (RB_EMPTY_ROOT(root))
1739                 return NULL;
1740
1741         /*
1742          * First, if we find a request starting at the end of the last
1743          * request, choose it.
1744          */
1745         __cfqq = cfq_prio_tree_lookup(cfqd, root, sector, &parent, NULL);
1746         if (__cfqq)
1747                 return __cfqq;
1748
1749         /*
1750          * If the exact sector wasn't found, the parent of the NULL leaf
1751          * will contain the closest sector.
1752          */
1753         __cfqq = rb_entry(parent, struct cfq_queue, p_node);
1754         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1755                 return __cfqq;
1756
1757         if (blk_rq_pos(__cfqq->next_rq) < sector)
1758                 node = rb_next(&__cfqq->p_node);
1759         else
1760                 node = rb_prev(&__cfqq->p_node);
1761         if (!node)
1762                 return NULL;
1763
1764         __cfqq = rb_entry(node, struct cfq_queue, p_node);
1765         if (cfq_rq_close(cfqd, cur_cfqq, __cfqq->next_rq))
1766                 return __cfqq;
1767
1768         return NULL;
1769 }
1770
1771 /*
1772  * cfqd - obvious
1773  * cur_cfqq - passed in so that we don't decide that the current queue is
1774  *            closely cooperating with itself.
1775  *
1776  * So, basically we're assuming that that cur_cfqq has dispatched at least
1777  * one request, and that cfqd->last_position reflects a position on the disk
1778  * associated with the I/O issued by cur_cfqq.  I'm not sure this is a valid
1779  * assumption.
1780  */
1781 static struct cfq_queue *cfq_close_cooperator(struct cfq_data *cfqd,
1782                                               struct cfq_queue *cur_cfqq)
1783 {
1784         struct cfq_queue *cfqq;
1785
1786         if (cfq_class_idle(cur_cfqq))
1787                 return NULL;
1788         if (!cfq_cfqq_sync(cur_cfqq))
1789                 return NULL;
1790         if (CFQQ_SEEKY(cur_cfqq))
1791                 return NULL;
1792
1793         /*
1794          * Don't search priority tree if it's the only queue in the group.
1795          */
1796         if (cur_cfqq->cfqg->nr_cfqq == 1)
1797                 return NULL;
1798
1799         /*
1800          * We should notice if some of the queues are cooperating, eg
1801          * working closely on the same area of the disk. In that case,
1802          * we can group them together and don't waste time idling.
1803          */
1804         cfqq = cfqq_close(cfqd, cur_cfqq);
1805         if (!cfqq)
1806                 return NULL;
1807
1808         /* If new queue belongs to different cfq_group, don't choose it */
1809         if (cur_cfqq->cfqg != cfqq->cfqg)
1810                 return NULL;
1811
1812         /*
1813          * It only makes sense to merge sync queues.
1814          */
1815         if (!cfq_cfqq_sync(cfqq))
1816                 return NULL;
1817         if (CFQQ_SEEKY(cfqq))
1818                 return NULL;
1819
1820         /*
1821          * Do not merge queues of different priority classes
1822          */
1823         if (cfq_class_rt(cfqq) != cfq_class_rt(cur_cfqq))
1824                 return NULL;
1825
1826         return cfqq;
1827 }
1828
1829 /*
1830  * Determine whether we should enforce idle window for this queue.
1831  */
1832
1833 static bool cfq_should_idle(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1834 {
1835         enum wl_prio_t prio = cfqq_prio(cfqq);
1836         struct cfq_rb_root *service_tree = cfqq->service_tree;
1837
1838         BUG_ON(!service_tree);
1839         BUG_ON(!service_tree->count);
1840
1841         /* We never do for idle class queues. */
1842         if (prio == IDLE_WORKLOAD)
1843                 return false;
1844
1845         /* We do for queues that were marked with idle window flag. */
1846         if (cfq_cfqq_idle_window(cfqq) &&
1847            !(blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag))
1848                 return true;
1849
1850         /*
1851          * Otherwise, we do only if they are the last ones
1852          * in their service tree.
1853          */
1854         if (service_tree->count == 1 && cfq_cfqq_sync(cfqq))
1855                 return 1;
1856         cfq_log_cfqq(cfqd, cfqq, "Not idling. st->count:%d",
1857                         service_tree->count);
1858         return 0;
1859 }
1860
1861 static void cfq_arm_slice_timer(struct cfq_data *cfqd)
1862 {
1863         struct cfq_queue *cfqq = cfqd->active_queue;
1864         struct cfq_io_context *cic;
1865         unsigned long sl;
1866
1867         /*
1868          * SSD device without seek penalty, disable idling. But only do so
1869          * for devices that support queuing, otherwise we still have a problem
1870          * with sync vs async workloads.
1871          */
1872         if (blk_queue_nonrot(cfqd->queue) && cfqd->hw_tag)
1873                 return;
1874
1875         WARN_ON(!RB_EMPTY_ROOT(&cfqq->sort_list));
1876         WARN_ON(cfq_cfqq_slice_new(cfqq));
1877
1878         /*
1879          * idle is disabled, either manually or by past process history
1880          */
1881         if (!cfqd->cfq_slice_idle || !cfq_should_idle(cfqd, cfqq))
1882                 return;
1883
1884         /*
1885          * still active requests from this queue, don't idle
1886          */
1887         if (cfqq->dispatched)
1888                 return;
1889
1890         /*
1891          * task has exited, don't wait
1892          */
1893         cic = cfqd->active_cic;
1894         if (!cic || !atomic_read(&cic->ioc->nr_tasks))
1895                 return;
1896
1897         /*
1898          * If our average think time is larger than the remaining time
1899          * slice, then don't idle. This avoids overrunning the allotted
1900          * time slice.
1901          */
1902         if (sample_valid(cic->ttime_samples) &&
1903             (cfqq->slice_end - jiffies < cic->ttime_mean)) {
1904                 cfq_log_cfqq(cfqd, cfqq, "Not idling. think_time:%d",
1905                                 cic->ttime_mean);
1906                 return;
1907         }
1908
1909         cfq_mark_cfqq_wait_request(cfqq);
1910
1911         sl = cfqd->cfq_slice_idle;
1912
1913         mod_timer(&cfqd->idle_slice_timer, jiffies + sl);
1914         blkiocg_update_set_idle_time_stats(&cfqq->cfqg->blkg);
1915         cfq_log_cfqq(cfqd, cfqq, "arm_idle: %lu", sl);
1916 }
1917
1918 /*
1919  * Move request from internal lists to the request queue dispatch list.
1920  */
1921 static void cfq_dispatch_insert(struct request_queue *q, struct request *rq)
1922 {
1923         struct cfq_data *cfqd = q->elevator->elevator_data;
1924         struct cfq_queue *cfqq = RQ_CFQQ(rq);
1925
1926         cfq_log_cfqq(cfqd, cfqq, "dispatch_insert");
1927
1928         cfqq->next_rq = cfq_find_next_rq(cfqd, cfqq, rq);
1929         cfq_remove_request(rq);
1930         cfqq->dispatched++;
1931         elv_dispatch_sort(q, rq);
1932
1933         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]++;
1934         blkiocg_update_dispatch_stats(&cfqq->cfqg->blkg, blk_rq_bytes(rq),
1935                                         rq_data_dir(rq), rq_is_sync(rq));
1936 }
1937
1938 /*
1939  * return expired entry, or NULL to just start from scratch in rbtree
1940  */
1941 static struct request *cfq_check_fifo(struct cfq_queue *cfqq)
1942 {
1943         struct request *rq = NULL;
1944
1945         if (cfq_cfqq_fifo_expire(cfqq))
1946                 return NULL;
1947
1948         cfq_mark_cfqq_fifo_expire(cfqq);
1949
1950         if (list_empty(&cfqq->fifo))
1951                 return NULL;
1952
1953         rq = rq_entry_fifo(cfqq->fifo.next);
1954         if (time_before(jiffies, rq_fifo_time(rq)))
1955                 rq = NULL;
1956
1957         cfq_log_cfqq(cfqq->cfqd, cfqq, "fifo=%p", rq);
1958         return rq;
1959 }
1960
1961 static inline int
1962 cfq_prio_to_maxrq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
1963 {
1964         const int base_rq = cfqd->cfq_slice_async_rq;
1965
1966         WARN_ON(cfqq->ioprio >= IOPRIO_BE_NR);
1967
1968         return 2 * (base_rq + base_rq * (CFQ_PRIO_LISTS - 1 - cfqq->ioprio));
1969 }
1970
1971 /*
1972  * Must be called with the queue_lock held.
1973  */
1974 static int cfqq_process_refs(struct cfq_queue *cfqq)
1975 {
1976         int process_refs, io_refs;
1977
1978         io_refs = cfqq->allocated[READ] + cfqq->allocated[WRITE];
1979         process_refs = atomic_read(&cfqq->ref) - io_refs;
1980         BUG_ON(process_refs < 0);
1981         return process_refs;
1982 }
1983
1984 static void cfq_setup_merge(struct cfq_queue *cfqq, struct cfq_queue *new_cfqq)
1985 {
1986         int process_refs, new_process_refs;
1987         struct cfq_queue *__cfqq;
1988
1989         /* Avoid a circular list and skip interim queue merges */
1990         while ((__cfqq = new_cfqq->new_cfqq)) {
1991                 if (__cfqq == cfqq)
1992                         return;
1993                 new_cfqq = __cfqq;
1994         }
1995
1996         process_refs = cfqq_process_refs(cfqq);
1997         /*
1998          * If the process for the cfqq has gone away, there is no
1999          * sense in merging the queues.
2000          */
2001         if (process_refs == 0)
2002                 return;
2003
2004         /*
2005          * Merge in the direction of the lesser amount of work.
2006          */
2007         new_process_refs = cfqq_process_refs(new_cfqq);
2008         if (new_process_refs >= process_refs) {
2009                 cfqq->new_cfqq = new_cfqq;
2010                 atomic_add(process_refs, &new_cfqq->ref);
2011         } else {
2012                 new_cfqq->new_cfqq = cfqq;
2013                 atomic_add(new_process_refs, &cfqq->ref);
2014         }
2015 }
2016
2017 static enum wl_type_t cfq_choose_wl(struct cfq_data *cfqd,
2018                                 struct cfq_group *cfqg, enum wl_prio_t prio)
2019 {
2020         struct cfq_queue *queue;
2021         int i;
2022         bool key_valid = false;
2023         unsigned long lowest_key = 0;
2024         enum wl_type_t cur_best = SYNC_NOIDLE_WORKLOAD;
2025
2026         for (i = 0; i <= SYNC_WORKLOAD; ++i) {
2027                 /* select the one with lowest rb_key */
2028                 queue = cfq_rb_first(service_tree_for(cfqg, prio, i));
2029                 if (queue &&
2030                     (!key_valid || time_before(queue->rb_key, lowest_key))) {
2031                         lowest_key = queue->rb_key;
2032                         cur_best = i;
2033                         key_valid = true;
2034                 }
2035         }
2036
2037         return cur_best;
2038 }
2039
2040 static void choose_service_tree(struct cfq_data *cfqd, struct cfq_group *cfqg)
2041 {
2042         unsigned slice;
2043         unsigned count;
2044         struct cfq_rb_root *st;
2045         unsigned group_slice;
2046
2047         if (!cfqg) {
2048                 cfqd->serving_prio = IDLE_WORKLOAD;
2049                 cfqd->workload_expires = jiffies + 1;
2050                 return;
2051         }
2052
2053         /* Choose next priority. RT > BE > IDLE */
2054         if (cfq_group_busy_queues_wl(RT_WORKLOAD, cfqd, cfqg))
2055                 cfqd->serving_prio = RT_WORKLOAD;
2056         else if (cfq_group_busy_queues_wl(BE_WORKLOAD, cfqd, cfqg))
2057                 cfqd->serving_prio = BE_WORKLOAD;
2058         else {
2059                 cfqd->serving_prio = IDLE_WORKLOAD;
2060                 cfqd->workload_expires = jiffies + 1;
2061                 return;
2062         }
2063
2064         /*
2065          * For RT and BE, we have to choose also the type
2066          * (SYNC, SYNC_NOIDLE, ASYNC), and to compute a workload
2067          * expiration time
2068          */
2069         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2070         count = st->count;
2071
2072         /*
2073          * check workload expiration, and that we still have other queues ready
2074          */
2075         if (count && !time_after(jiffies, cfqd->workload_expires))
2076                 return;
2077
2078         /* otherwise select new workload type */
2079         cfqd->serving_type =
2080                 cfq_choose_wl(cfqd, cfqg, cfqd->serving_prio);
2081         st = service_tree_for(cfqg, cfqd->serving_prio, cfqd->serving_type);
2082         count = st->count;
2083
2084         /*
2085          * the workload slice is computed as a fraction of target latency
2086          * proportional to the number of queues in that workload, over
2087          * all the queues in the same priority class
2088          */
2089         group_slice = cfq_group_slice(cfqd, cfqg);
2090
2091         slice = group_slice * count /
2092                 max_t(unsigned, cfqg->busy_queues_avg[cfqd->serving_prio],
2093                       cfq_group_busy_queues_wl(cfqd->serving_prio, cfqd, cfqg));
2094
2095         if (cfqd->serving_type == ASYNC_WORKLOAD) {
2096                 unsigned int tmp;
2097
2098                 /*
2099                  * Async queues are currently system wide. Just taking
2100                  * proportion of queues with-in same group will lead to higher
2101                  * async ratio system wide as generally root group is going
2102                  * to have higher weight. A more accurate thing would be to
2103                  * calculate system wide asnc/sync ratio.
2104                  */
2105                 tmp = cfq_target_latency * cfqg_busy_async_queues(cfqd, cfqg);
2106                 tmp = tmp/cfqd->busy_queues;
2107                 slice = min_t(unsigned, slice, tmp);
2108
2109                 /* async workload slice is scaled down according to
2110                  * the sync/async slice ratio. */
2111                 slice = slice * cfqd->cfq_slice[0] / cfqd->cfq_slice[1];
2112         } else
2113                 /* sync workload slice is at least 2 * cfq_slice_idle */
2114                 slice = max(slice, 2 * cfqd->cfq_slice_idle);
2115
2116         slice = max_t(unsigned, slice, CFQ_MIN_TT);
2117         cfq_log(cfqd, "workload slice:%d", slice);
2118         cfqd->workload_expires = jiffies + slice;
2119         cfqd->noidle_tree_requires_idle = false;
2120 }
2121
2122 static struct cfq_group *cfq_get_next_cfqg(struct cfq_data *cfqd)
2123 {
2124         struct cfq_rb_root *st = &cfqd->grp_service_tree;
2125         struct cfq_group *cfqg;
2126
2127         if (RB_EMPTY_ROOT(&st->rb))
2128                 return NULL;
2129         cfqg = cfq_rb_first_group(st);
2130         st->active = &cfqg->rb_node;
2131         update_min_vdisktime(st);
2132         return cfqg;
2133 }
2134
2135 static void cfq_choose_cfqg(struct cfq_data *cfqd)
2136 {
2137         struct cfq_group *cfqg = cfq_get_next_cfqg(cfqd);
2138
2139         cfqd->serving_group = cfqg;
2140
2141         /* Restore the workload type data */
2142         if (cfqg->saved_workload_slice) {
2143                 cfqd->workload_expires = jiffies + cfqg->saved_workload_slice;
2144                 cfqd->serving_type = cfqg->saved_workload;
2145                 cfqd->serving_prio = cfqg->saved_serving_prio;
2146         } else
2147                 cfqd->workload_expires = jiffies - 1;
2148
2149         choose_service_tree(cfqd, cfqg);
2150 }
2151
2152 /*
2153  * Select a queue for service. If we have a current active queue,
2154  * check whether to continue servicing it, or retrieve and set a new one.
2155  */
2156 static struct cfq_queue *cfq_select_queue(struct cfq_data *cfqd)
2157 {
2158         struct cfq_queue *cfqq, *new_cfqq = NULL;
2159
2160         cfqq = cfqd->active_queue;
2161         if (!cfqq)
2162                 goto new_queue;
2163
2164         if (!cfqd->rq_queued)
2165                 return NULL;
2166
2167         /*
2168          * We were waiting for group to get backlogged. Expire the queue
2169          */
2170         if (cfq_cfqq_wait_busy(cfqq) && !RB_EMPTY_ROOT(&cfqq->sort_list))
2171                 goto expire;
2172
2173         /*
2174          * The active queue has run out of time, expire it and select new.
2175          */
2176         if (cfq_slice_used(cfqq) && !cfq_cfqq_must_dispatch(cfqq)) {
2177                 /*
2178                  * If slice had not expired at the completion of last request
2179                  * we might not have turned on wait_busy flag. Don't expire
2180                  * the queue yet. Allow the group to get backlogged.
2181                  *
2182                  * The very fact that we have used the slice, that means we
2183                  * have been idling all along on this queue and it should be
2184                  * ok to wait for this request to complete.
2185                  */
2186                 if (cfqq->cfqg->nr_cfqq == 1 && RB_EMPTY_ROOT(&cfqq->sort_list)
2187                     && cfqq->dispatched && cfq_should_idle(cfqd, cfqq)) {
2188                         cfqq = NULL;
2189                         goto keep_queue;
2190                 } else
2191                         goto expire;
2192         }
2193
2194         /*
2195          * The active queue has requests and isn't expired, allow it to
2196          * dispatch.
2197          */
2198         if (!RB_EMPTY_ROOT(&cfqq->sort_list))
2199                 goto keep_queue;
2200
2201         /*
2202          * If another queue has a request waiting within our mean seek
2203          * distance, let it run.  The expire code will check for close
2204          * cooperators and put the close queue at the front of the service
2205          * tree.  If possible, merge the expiring queue with the new cfqq.
2206          */
2207         new_cfqq = cfq_close_cooperator(cfqd, cfqq);
2208         if (new_cfqq) {
2209                 if (!cfqq->new_cfqq)
2210                         cfq_setup_merge(cfqq, new_cfqq);
2211                 goto expire;
2212         }
2213
2214         /*
2215          * No requests pending. If the active queue still has requests in
2216          * flight or is idling for a new request, allow either of these
2217          * conditions to happen (or time out) before selecting a new queue.
2218          */
2219         if (timer_pending(&cfqd->idle_slice_timer) ||
2220             (cfqq->dispatched && cfq_should_idle(cfqd, cfqq))) {
2221                 cfqq = NULL;
2222                 goto keep_queue;
2223         }
2224
2225 expire:
2226         cfq_slice_expired(cfqd, 0);
2227 new_queue:
2228         /*
2229          * Current queue expired. Check if we have to switch to a new
2230          * service tree
2231          */
2232         if (!new_cfqq)
2233                 cfq_choose_cfqg(cfqd);
2234
2235         cfqq = cfq_set_active_queue(cfqd, new_cfqq);
2236 keep_queue:
2237         return cfqq;
2238 }
2239
2240 static int __cfq_forced_dispatch_cfqq(struct cfq_queue *cfqq)
2241 {
2242         int dispatched = 0;
2243
2244         while (cfqq->next_rq) {
2245                 cfq_dispatch_insert(cfqq->cfqd->queue, cfqq->next_rq);
2246                 dispatched++;
2247         }
2248
2249         BUG_ON(!list_empty(&cfqq->fifo));
2250
2251         /* By default cfqq is not expired if it is empty. Do it explicitly */
2252         __cfq_slice_expired(cfqq->cfqd, cfqq, 0);
2253         return dispatched;
2254 }
2255
2256 /*
2257  * Drain our current requests. Used for barriers and when switching
2258  * io schedulers on-the-fly.
2259  */
2260 static int cfq_forced_dispatch(struct cfq_data *cfqd)
2261 {
2262         struct cfq_queue *cfqq;
2263         int dispatched = 0;
2264
2265         /* Expire the timeslice of the current active queue first */
2266         cfq_slice_expired(cfqd, 0);
2267         while ((cfqq = cfq_get_next_queue_forced(cfqd)) != NULL) {
2268                 __cfq_set_active_queue(cfqd, cfqq);
2269                 dispatched += __cfq_forced_dispatch_cfqq(cfqq);
2270         }
2271
2272         BUG_ON(cfqd->busy_queues);
2273
2274         cfq_log(cfqd, "forced_dispatch=%d", dispatched);
2275         return dispatched;
2276 }
2277
2278 static inline bool cfq_slice_used_soon(struct cfq_data *cfqd,
2279         struct cfq_queue *cfqq)
2280 {
2281         /* the queue hasn't finished any request, can't estimate */
2282         if (cfq_cfqq_slice_new(cfqq))
2283                 return 1;
2284         if (time_after(jiffies + cfqd->cfq_slice_idle * cfqq->dispatched,
2285                 cfqq->slice_end))
2286                 return 1;
2287
2288         return 0;
2289 }
2290
2291 static bool cfq_may_dispatch(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2292 {
2293         unsigned int max_dispatch;
2294
2295         /*
2296          * Drain async requests before we start sync IO
2297          */
2298         if (cfq_should_idle(cfqd, cfqq) && cfqd->rq_in_flight[BLK_RW_ASYNC])
2299                 return false;
2300
2301         /*
2302          * If this is an async queue and we have sync IO in flight, let it wait
2303          */
2304         if (cfqd->rq_in_flight[BLK_RW_SYNC] && !cfq_cfqq_sync(cfqq))
2305                 return false;
2306
2307         max_dispatch = max_t(unsigned int, cfqd->cfq_quantum / 2, 1);
2308         if (cfq_class_idle(cfqq))
2309                 max_dispatch = 1;
2310
2311         /*
2312          * Does this cfqq already have too much IO in flight?
2313          */
2314         if (cfqq->dispatched >= max_dispatch) {
2315                 /*
2316                  * idle queue must always only have a single IO in flight
2317                  */
2318                 if (cfq_class_idle(cfqq))
2319                         return false;
2320
2321                 /*
2322                  * We have other queues, don't allow more IO from this one
2323                  */
2324                 if (cfqd->busy_queues > 1 && cfq_slice_used_soon(cfqd, cfqq))
2325                         return false;
2326
2327                 /*
2328                  * Sole queue user, no limit
2329                  */
2330                 if (cfqd->busy_queues == 1)
2331                         max_dispatch = -1;
2332                 else
2333                         /*
2334                          * Normally we start throttling cfqq when cfq_quantum/2
2335                          * requests have been dispatched. But we can drive
2336                          * deeper queue depths at the beginning of slice
2337                          * subjected to upper limit of cfq_quantum.
2338                          * */
2339                         max_dispatch = cfqd->cfq_quantum;
2340         }
2341
2342         /*
2343          * Async queues must wait a bit before being allowed dispatch.
2344          * We also ramp up the dispatch depth gradually for async IO,
2345          * based on the last sync IO we serviced
2346          */
2347         if (!cfq_cfqq_sync(cfqq) && cfqd->cfq_latency) {
2348                 unsigned long last_sync = jiffies - cfqd->last_delayed_sync;
2349                 unsigned int depth;
2350
2351                 depth = last_sync / cfqd->cfq_slice[1];
2352                 if (!depth && !cfqq->dispatched)
2353                         depth = 1;
2354                 if (depth < max_dispatch)
2355                         max_dispatch = depth;
2356         }
2357
2358         /*
2359          * If we're below the current max, allow a dispatch
2360          */
2361         return cfqq->dispatched < max_dispatch;
2362 }
2363
2364 /*
2365  * Dispatch a request from cfqq, moving them to the request queue
2366  * dispatch list.
2367  */
2368 static bool cfq_dispatch_request(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2369 {
2370         struct request *rq;
2371
2372         BUG_ON(RB_EMPTY_ROOT(&cfqq->sort_list));
2373
2374         if (!cfq_may_dispatch(cfqd, cfqq))
2375                 return false;
2376
2377         /*
2378          * follow expired path, else get first next available
2379          */
2380         rq = cfq_check_fifo(cfqq);
2381         if (!rq)
2382                 rq = cfqq->next_rq;
2383
2384         /*
2385          * insert request into driver dispatch list
2386          */
2387         cfq_dispatch_insert(cfqd->queue, rq);
2388
2389         if (!cfqd->active_cic) {
2390                 struct cfq_io_context *cic = RQ_CIC(rq);
2391
2392                 atomic_long_inc(&cic->ioc->refcount);
2393                 cfqd->active_cic = cic;
2394         }
2395
2396         return true;
2397 }
2398
2399 /*
2400  * Find the cfqq that we need to service and move a request from that to the
2401  * dispatch list
2402  */
2403 static int cfq_dispatch_requests(struct request_queue *q, int force)
2404 {
2405         struct cfq_data *cfqd = q->elevator->elevator_data;
2406         struct cfq_queue *cfqq;
2407
2408         if (!cfqd->busy_queues)
2409                 return 0;
2410
2411         if (unlikely(force))
2412                 return cfq_forced_dispatch(cfqd);
2413
2414         cfqq = cfq_select_queue(cfqd);
2415         if (!cfqq)
2416                 return 0;
2417
2418         /*
2419          * Dispatch a request from this cfqq, if it is allowed
2420          */
2421         if (!cfq_dispatch_request(cfqd, cfqq))
2422                 return 0;
2423
2424         cfqq->slice_dispatch++;
2425         cfq_clear_cfqq_must_dispatch(cfqq);
2426
2427         /*
2428          * expire an async queue immediately if it has used up its slice. idle
2429          * queue always expire after 1 dispatch round.
2430          */
2431         if (cfqd->busy_queues > 1 && ((!cfq_cfqq_sync(cfqq) &&
2432             cfqq->slice_dispatch >= cfq_prio_to_maxrq(cfqd, cfqq)) ||
2433             cfq_class_idle(cfqq))) {
2434                 cfqq->slice_end = jiffies + 1;
2435                 cfq_slice_expired(cfqd, 0);
2436         }
2437
2438         cfq_log_cfqq(cfqd, cfqq, "dispatched a request");
2439         return 1;
2440 }
2441
2442 /*
2443  * task holds one reference to the queue, dropped when task exits. each rq
2444  * in-flight on this queue also holds a reference, dropped when rq is freed.
2445  *
2446  * Each cfq queue took a reference on the parent group. Drop it now.
2447  * queue lock must be held here.
2448  */
2449 static void cfq_put_queue(struct cfq_queue *cfqq)
2450 {
2451         struct cfq_data *cfqd = cfqq->cfqd;
2452         struct cfq_group *cfqg, *orig_cfqg;
2453
2454         BUG_ON(atomic_read(&cfqq->ref) <= 0);
2455
2456         if (!atomic_dec_and_test(&cfqq->ref))
2457                 return;
2458
2459         cfq_log_cfqq(cfqd, cfqq, "put_queue");
2460         BUG_ON(rb_first(&cfqq->sort_list));
2461         BUG_ON(cfqq->allocated[READ] + cfqq->allocated[WRITE]);
2462         cfqg = cfqq->cfqg;
2463         orig_cfqg = cfqq->orig_cfqg;
2464
2465         if (unlikely(cfqd->active_queue == cfqq)) {
2466                 __cfq_slice_expired(cfqd, cfqq, 0);
2467                 cfq_schedule_dispatch(cfqd);
2468         }
2469
2470         BUG_ON(cfq_cfqq_on_rr(cfqq));
2471         kmem_cache_free(cfq_pool, cfqq);
2472         cfq_put_cfqg(cfqg);
2473         if (orig_cfqg)
2474                 cfq_put_cfqg(orig_cfqg);
2475 }
2476
2477 /*
2478  * Must always be called with the rcu_read_lock() held
2479  */
2480 static void
2481 __call_for_each_cic(struct io_context *ioc,
2482                     void (*func)(struct io_context *, struct cfq_io_context *))
2483 {
2484         struct cfq_io_context *cic;
2485         struct hlist_node *n;
2486
2487         hlist_for_each_entry_rcu(cic, n, &ioc->cic_list, cic_list)
2488                 func(ioc, cic);
2489 }
2490
2491 /*
2492  * Call func for each cic attached to this ioc.
2493  */
2494 static void
2495 call_for_each_cic(struct io_context *ioc,
2496                   void (*func)(struct io_context *, struct cfq_io_context *))
2497 {
2498         rcu_read_lock();
2499         __call_for_each_cic(ioc, func);
2500         rcu_read_unlock();
2501 }
2502
2503 static void cfq_cic_free_rcu(struct rcu_head *head)
2504 {
2505         struct cfq_io_context *cic;
2506
2507         cic = container_of(head, struct cfq_io_context, rcu_head);
2508
2509         kmem_cache_free(cfq_ioc_pool, cic);
2510         elv_ioc_count_dec(cfq_ioc_count);
2511
2512         if (ioc_gone) {
2513                 /*
2514                  * CFQ scheduler is exiting, grab exit lock and check
2515                  * the pending io context count. If it hits zero,
2516                  * complete ioc_gone and set it back to NULL
2517                  */
2518                 spin_lock(&ioc_gone_lock);
2519                 if (ioc_gone && !elv_ioc_count_read(cfq_ioc_count)) {
2520                         complete(ioc_gone);
2521                         ioc_gone = NULL;
2522                 }
2523                 spin_unlock(&ioc_gone_lock);
2524         }
2525 }
2526
2527 static void cfq_cic_free(struct cfq_io_context *cic)
2528 {
2529         call_rcu(&cic->rcu_head, cfq_cic_free_rcu);
2530 }
2531
2532 static void cic_free_func(struct io_context *ioc, struct cfq_io_context *cic)
2533 {
2534         unsigned long flags;
2535         unsigned long dead_key = (unsigned long) cic->key;
2536
2537         BUG_ON(!(dead_key & CIC_DEAD_KEY));
2538
2539         spin_lock_irqsave(&ioc->lock, flags);
2540         radix_tree_delete(&ioc->radix_root, dead_key >> CIC_DEAD_INDEX_SHIFT);
2541         hlist_del_rcu(&cic->cic_list);
2542         spin_unlock_irqrestore(&ioc->lock, flags);
2543
2544         cfq_cic_free(cic);
2545 }
2546
2547 /*
2548  * Must be called with rcu_read_lock() held or preemption otherwise disabled.
2549  * Only two callers of this - ->dtor() which is called with the rcu_read_lock(),
2550  * and ->trim() which is called with the task lock held
2551  */
2552 static void cfq_free_io_context(struct io_context *ioc)
2553 {
2554         /*
2555          * ioc->refcount is zero here, or we are called from elv_unregister(),
2556          * so no more cic's are allowed to be linked into this ioc.  So it
2557          * should be ok to iterate over the known list, we will see all cic's
2558          * since no new ones are added.
2559          */
2560         __call_for_each_cic(ioc, cic_free_func);
2561 }
2562
2563 static void cfq_exit_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq)
2564 {
2565         struct cfq_queue *__cfqq, *next;
2566
2567         if (unlikely(cfqq == cfqd->active_queue)) {
2568                 __cfq_slice_expired(cfqd, cfqq, 0);
2569                 cfq_schedule_dispatch(cfqd);
2570         }
2571
2572         /*
2573          * If this queue was scheduled to merge with another queue, be
2574          * sure to drop the reference taken on that queue (and others in
2575          * the merge chain).  See cfq_setup_merge and cfq_merge_cfqqs.
2576          */
2577         __cfqq = cfqq->new_cfqq;
2578         while (__cfqq) {
2579                 if (__cfqq == cfqq) {
2580                         WARN(1, "cfqq->new_cfqq loop detected\n");
2581                         break;
2582                 }
2583                 next = __cfqq->new_cfqq;
2584                 cfq_put_queue(__cfqq);
2585                 __cfqq = next;
2586         }
2587
2588         cfq_put_queue(cfqq);
2589 }
2590
2591 static void __cfq_exit_single_io_context(struct cfq_data *cfqd,
2592                                          struct cfq_io_context *cic)
2593 {
2594         struct io_context *ioc = cic->ioc;
2595
2596         list_del_init(&cic->queue_list);
2597
2598         /*
2599          * Make sure dead mark is seen for dead queues
2600          */
2601         smp_wmb();
2602         cic->key = cfqd_dead_key(cfqd);
2603
2604         if (ioc->ioc_data == cic)
2605                 rcu_assign_pointer(ioc->ioc_data, NULL);
2606
2607         if (cic->cfqq[BLK_RW_ASYNC]) {
2608                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_ASYNC]);
2609                 cic->cfqq[BLK_RW_ASYNC] = NULL;
2610         }
2611
2612         if (cic->cfqq[BLK_RW_SYNC]) {
2613                 cfq_exit_cfqq(cfqd, cic->cfqq[BLK_RW_SYNC]);
2614                 cic->cfqq[BLK_RW_SYNC] = NULL;
2615         }
2616 }
2617
2618 static void cfq_exit_single_io_context(struct io_context *ioc,
2619                                        struct cfq_io_context *cic)
2620 {
2621         struct cfq_data *cfqd = cic_to_cfqd(cic);
2622
2623         if (cfqd) {
2624                 struct request_queue *q = cfqd->queue;
2625                 unsigned long flags;
2626
2627                 spin_lock_irqsave(q->queue_lock, flags);
2628
2629                 /*
2630                  * Ensure we get a fresh copy of the ->key to prevent
2631                  * race between exiting task and queue
2632                  */
2633                 smp_read_barrier_depends();
2634                 if (cic->key == cfqd)
2635                         __cfq_exit_single_io_context(cfqd, cic);
2636
2637                 spin_unlock_irqrestore(q->queue_lock, flags);
2638         }
2639 }
2640
2641 /*
2642  * The process that ioc belongs to has exited, we need to clean up
2643  * and put the internal structures we have that belongs to that process.
2644  */
2645 static void cfq_exit_io_context(struct io_context *ioc)
2646 {
2647         call_for_each_cic(ioc, cfq_exit_single_io_context);
2648 }
2649
2650 static struct cfq_io_context *
2651 cfq_alloc_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
2652 {
2653         struct cfq_io_context *cic;
2654
2655         cic = kmem_cache_alloc_node(cfq_ioc_pool, gfp_mask | __GFP_ZERO,
2656                                                         cfqd->queue->node);
2657         if (cic) {
2658                 cic->last_end_request = jiffies;
2659                 INIT_LIST_HEAD(&cic->queue_list);
2660                 INIT_HLIST_NODE(&cic->cic_list);
2661                 cic->dtor = cfq_free_io_context;
2662                 cic->exit = cfq_exit_io_context;
2663                 elv_ioc_count_inc(cfq_ioc_count);
2664         }
2665
2666         return cic;
2667 }
2668
2669 static void cfq_init_prio_data(struct cfq_queue *cfqq, struct io_context *ioc)
2670 {
2671         struct task_struct *tsk = current;
2672         int ioprio_class;
2673
2674         if (!cfq_cfqq_prio_changed(cfqq))
2675                 return;
2676
2677         ioprio_class = IOPRIO_PRIO_CLASS(ioc->ioprio);
2678         switch (ioprio_class) {
2679         default:
2680                 printk(KERN_ERR "cfq: bad prio %x\n", ioprio_class);
2681         case IOPRIO_CLASS_NONE:
2682                 /*
2683                  * no prio set, inherit CPU scheduling settings
2684                  */
2685                 cfqq->ioprio = task_nice_ioprio(tsk);
2686                 cfqq->ioprio_class = task_nice_ioclass(tsk);
2687                 break;
2688         case IOPRIO_CLASS_RT:
2689                 cfqq->ioprio = task_ioprio(ioc);
2690                 cfqq->ioprio_class = IOPRIO_CLASS_RT;
2691                 break;
2692         case IOPRIO_CLASS_BE:
2693                 cfqq->ioprio = task_ioprio(ioc);
2694                 cfqq->ioprio_class = IOPRIO_CLASS_BE;
2695                 break;
2696         case IOPRIO_CLASS_IDLE:
2697                 cfqq->ioprio_class = IOPRIO_CLASS_IDLE;
2698                 cfqq->ioprio = 7;
2699                 cfq_clear_cfqq_idle_window(cfqq);
2700                 break;
2701         }
2702
2703         /*
2704          * keep track of original prio settings in case we have to temporarily
2705          * elevate the priority of this queue
2706          */
2707         cfqq->org_ioprio = cfqq->ioprio;
2708         cfqq->org_ioprio_class = cfqq->ioprio_class;
2709         cfq_clear_cfqq_prio_changed(cfqq);
2710 }
2711
2712 static void changed_ioprio(struct io_context *ioc, struct cfq_io_context *cic)
2713 {
2714         struct cfq_data *cfqd = cic_to_cfqd(cic);
2715         struct cfq_queue *cfqq;
2716         unsigned long flags;
2717
2718         if (unlikely(!cfqd))
2719                 return;
2720
2721         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2722
2723         cfqq = cic->cfqq[BLK_RW_ASYNC];
2724         if (cfqq) {
2725                 struct cfq_queue *new_cfqq;
2726                 new_cfqq = cfq_get_queue(cfqd, BLK_RW_ASYNC, cic->ioc,
2727                                                 GFP_ATOMIC);
2728                 if (new_cfqq) {
2729                         cic->cfqq[BLK_RW_ASYNC] = new_cfqq;
2730                         cfq_put_queue(cfqq);
2731                 }
2732         }
2733
2734         cfqq = cic->cfqq[BLK_RW_SYNC];
2735         if (cfqq)
2736                 cfq_mark_cfqq_prio_changed(cfqq);
2737
2738         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2739 }
2740
2741 static void cfq_ioc_set_ioprio(struct io_context *ioc)
2742 {
2743         call_for_each_cic(ioc, changed_ioprio);
2744         ioc->ioprio_changed = 0;
2745 }
2746
2747 static void cfq_init_cfqq(struct cfq_data *cfqd, struct cfq_queue *cfqq,
2748                           pid_t pid, bool is_sync)
2749 {
2750         RB_CLEAR_NODE(&cfqq->rb_node);
2751         RB_CLEAR_NODE(&cfqq->p_node);
2752         INIT_LIST_HEAD(&cfqq->fifo);
2753
2754         atomic_set(&cfqq->ref, 0);
2755         cfqq->cfqd = cfqd;
2756
2757         cfq_mark_cfqq_prio_changed(cfqq);
2758
2759         if (is_sync) {
2760                 if (!cfq_class_idle(cfqq))
2761                         cfq_mark_cfqq_idle_window(cfqq);
2762                 cfq_mark_cfqq_sync(cfqq);
2763         }
2764         cfqq->pid = pid;
2765 }
2766
2767 #ifdef CONFIG_CFQ_GROUP_IOSCHED
2768 static void changed_cgroup(struct io_context *ioc, struct cfq_io_context *cic)
2769 {
2770         struct cfq_queue *sync_cfqq = cic_to_cfqq(cic, 1);
2771         struct cfq_data *cfqd = cic_to_cfqd(cic);
2772         unsigned long flags;
2773         struct request_queue *q;
2774
2775         if (unlikely(!cfqd))
2776                 return;
2777
2778         q = cfqd->queue;
2779
2780         spin_lock_irqsave(q->queue_lock, flags);
2781
2782         if (sync_cfqq) {
2783                 /*
2784                  * Drop reference to sync queue. A new sync queue will be
2785                  * assigned in new group upon arrival of a fresh request.
2786                  */
2787                 cfq_log_cfqq(cfqd, sync_cfqq, "changed cgroup");
2788                 cic_set_cfqq(cic, NULL, 1);
2789                 cfq_put_queue(sync_cfqq);
2790         }
2791
2792         spin_unlock_irqrestore(q->queue_lock, flags);
2793 }
2794
2795 static void cfq_ioc_set_cgroup(struct io_context *ioc)
2796 {
2797         call_for_each_cic(ioc, changed_cgroup);
2798         ioc->cgroup_changed = 0;
2799 }
2800 #endif  /* CONFIG_CFQ_GROUP_IOSCHED */
2801
2802 static struct cfq_queue *
2803 cfq_find_alloc_queue(struct cfq_data *cfqd, bool is_sync,
2804                      struct io_context *ioc, gfp_t gfp_mask)
2805 {
2806         struct cfq_queue *cfqq, *new_cfqq = NULL;
2807         struct cfq_io_context *cic;
2808         struct cfq_group *cfqg;
2809
2810 retry:
2811         cfqg = cfq_get_cfqg(cfqd, 1);
2812         cic = cfq_cic_lookup(cfqd, ioc);
2813         /* cic always exists here */
2814         cfqq = cic_to_cfqq(cic, is_sync);
2815
2816         /*
2817          * Always try a new alloc if we fell back to the OOM cfqq
2818          * originally, since it should just be a temporary situation.
2819          */
2820         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
2821                 cfqq = NULL;
2822                 if (new_cfqq) {
2823                         cfqq = new_cfqq;
2824                         new_cfqq = NULL;
2825                 } else if (gfp_mask & __GFP_WAIT) {
2826                         spin_unlock_irq(cfqd->queue->queue_lock);
2827                         new_cfqq = kmem_cache_alloc_node(cfq_pool,
2828                                         gfp_mask | __GFP_ZERO,
2829                                         cfqd->queue->node);
2830                         spin_lock_irq(cfqd->queue->queue_lock);
2831                         if (new_cfqq)
2832                                 goto retry;
2833                 } else {
2834                         cfqq = kmem_cache_alloc_node(cfq_pool,
2835                                         gfp_mask | __GFP_ZERO,
2836                                         cfqd->queue->node);
2837                 }
2838
2839                 if (cfqq) {
2840                         cfq_init_cfqq(cfqd, cfqq, current->pid, is_sync);
2841                         cfq_init_prio_data(cfqq, ioc);
2842                         cfq_link_cfqq_cfqg(cfqq, cfqg);
2843                         cfq_log_cfqq(cfqd, cfqq, "alloced");
2844                 } else
2845                         cfqq = &cfqd->oom_cfqq;
2846         }
2847
2848         if (new_cfqq)
2849                 kmem_cache_free(cfq_pool, new_cfqq);
2850
2851         return cfqq;
2852 }
2853
2854 static struct cfq_queue **
2855 cfq_async_queue_prio(struct cfq_data *cfqd, int ioprio_class, int ioprio)
2856 {
2857         switch (ioprio_class) {
2858         case IOPRIO_CLASS_RT:
2859                 return &cfqd->async_cfqq[0][ioprio];
2860         case IOPRIO_CLASS_BE:
2861                 return &cfqd->async_cfqq[1][ioprio];
2862         case IOPRIO_CLASS_IDLE:
2863                 return &cfqd->async_idle_cfqq;
2864         default:
2865                 BUG();
2866         }
2867 }
2868
2869 static struct cfq_queue *
2870 cfq_get_queue(struct cfq_data *cfqd, bool is_sync, struct io_context *ioc,
2871               gfp_t gfp_mask)
2872 {
2873         const int ioprio = task_ioprio(ioc);
2874         const int ioprio_class = task_ioprio_class(ioc);
2875         struct cfq_queue **async_cfqq = NULL;
2876         struct cfq_queue *cfqq = NULL;
2877
2878         if (!is_sync) {
2879                 async_cfqq = cfq_async_queue_prio(cfqd, ioprio_class, ioprio);
2880                 cfqq = *async_cfqq;
2881         }
2882
2883         if (!cfqq)
2884                 cfqq = cfq_find_alloc_queue(cfqd, is_sync, ioc, gfp_mask);
2885
2886         /*
2887          * pin the queue now that it's allocated, scheduler exit will prune it
2888          */
2889         if (!is_sync && !(*async_cfqq)) {
2890                 atomic_inc(&cfqq->ref);
2891                 *async_cfqq = cfqq;
2892         }
2893
2894         atomic_inc(&cfqq->ref);
2895         return cfqq;
2896 }
2897
2898 /*
2899  * We drop cfq io contexts lazily, so we may find a dead one.
2900  */
2901 static void
2902 cfq_drop_dead_cic(struct cfq_data *cfqd, struct io_context *ioc,
2903                   struct cfq_io_context *cic)
2904 {
2905         unsigned long flags;
2906
2907         WARN_ON(!list_empty(&cic->queue_list));
2908         BUG_ON(cic->key != cfqd_dead_key(cfqd));
2909
2910         spin_lock_irqsave(&ioc->lock, flags);
2911
2912         BUG_ON(ioc->ioc_data == cic);
2913
2914         radix_tree_delete(&ioc->radix_root, cfqd->cic_index);
2915         hlist_del_rcu(&cic->cic_list);
2916         spin_unlock_irqrestore(&ioc->lock, flags);
2917
2918         cfq_cic_free(cic);
2919 }
2920
2921 static struct cfq_io_context *
2922 cfq_cic_lookup(struct cfq_data *cfqd, struct io_context *ioc)
2923 {
2924         struct cfq_io_context *cic;
2925         unsigned long flags;
2926
2927         if (unlikely(!ioc))
2928                 return NULL;
2929
2930         rcu_read_lock();
2931
2932         /*
2933          * we maintain a last-hit cache, to avoid browsing over the tree
2934          */
2935         cic = rcu_dereference(ioc->ioc_data);
2936         if (cic && cic->key == cfqd) {
2937                 rcu_read_unlock();
2938                 return cic;
2939         }
2940
2941         do {
2942                 cic = radix_tree_lookup(&ioc->radix_root, cfqd->cic_index);
2943                 rcu_read_unlock();
2944                 if (!cic)
2945                         break;
2946                 if (unlikely(cic->key != cfqd)) {
2947                         cfq_drop_dead_cic(cfqd, ioc, cic);
2948                         rcu_read_lock();
2949                         continue;
2950                 }
2951
2952                 spin_lock_irqsave(&ioc->lock, flags);
2953                 rcu_assign_pointer(ioc->ioc_data, cic);
2954                 spin_unlock_irqrestore(&ioc->lock, flags);
2955                 break;
2956         } while (1);
2957
2958         return cic;
2959 }
2960
2961 /*
2962  * Add cic into ioc, using cfqd as the search key. This enables us to lookup
2963  * the process specific cfq io context when entered from the block layer.
2964  * Also adds the cic to a per-cfqd list, used when this queue is removed.
2965  */
2966 static int cfq_cic_link(struct cfq_data *cfqd, struct io_context *ioc,
2967                         struct cfq_io_context *cic, gfp_t gfp_mask)
2968 {
2969         unsigned long flags;
2970         int ret;
2971
2972         ret = radix_tree_preload(gfp_mask);
2973         if (!ret) {
2974                 cic->ioc = ioc;
2975                 cic->key = cfqd;
2976
2977                 spin_lock_irqsave(&ioc->lock, flags);
2978                 ret = radix_tree_insert(&ioc->radix_root,
2979                                                 cfqd->cic_index, cic);
2980                 if (!ret)
2981                         hlist_add_head_rcu(&cic->cic_list, &ioc->cic_list);
2982                 spin_unlock_irqrestore(&ioc->lock, flags);
2983
2984                 radix_tree_preload_end();
2985
2986                 if (!ret) {
2987                         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
2988                         list_add(&cic->queue_list, &cfqd->cic_list);
2989                         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
2990                 }
2991         }
2992
2993         if (ret)
2994                 printk(KERN_ERR "cfq: cic link failed!\n");
2995
2996         return ret;
2997 }
2998
2999 /*
3000  * Setup general io context and cfq io context. There can be several cfq
3001  * io contexts per general io context, if this process is doing io to more
3002  * than one device managed by cfq.
3003  */
3004 static struct cfq_io_context *
3005 cfq_get_io_context(struct cfq_data *cfqd, gfp_t gfp_mask)
3006 {
3007         struct io_context *ioc = NULL;
3008         struct cfq_io_context *cic;
3009
3010         might_sleep_if(gfp_mask & __GFP_WAIT);
3011
3012         ioc = get_io_context(gfp_mask, cfqd->queue->node);
3013         if (!ioc)
3014                 return NULL;
3015
3016         cic = cfq_cic_lookup(cfqd, ioc);
3017         if (cic)
3018                 goto out;
3019
3020         cic = cfq_alloc_io_context(cfqd, gfp_mask);
3021         if (cic == NULL)
3022                 goto err;
3023
3024         if (cfq_cic_link(cfqd, ioc, cic, gfp_mask))
3025                 goto err_free;
3026
3027 out:
3028         smp_read_barrier_depends();
3029         if (unlikely(ioc->ioprio_changed))
3030                 cfq_ioc_set_ioprio(ioc);
3031
3032 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3033         if (unlikely(ioc->cgroup_changed))
3034                 cfq_ioc_set_cgroup(ioc);
3035 #endif
3036         return cic;
3037 err_free:
3038         cfq_cic_free(cic);
3039 err:
3040         put_io_context(ioc);
3041         return NULL;
3042 }
3043
3044 static void
3045 cfq_update_io_thinktime(struct cfq_data *cfqd, struct cfq_io_context *cic)
3046 {
3047         unsigned long elapsed = jiffies - cic->last_end_request;
3048         unsigned long ttime = min(elapsed, 2UL * cfqd->cfq_slice_idle);
3049
3050         cic->ttime_samples = (7*cic->ttime_samples + 256) / 8;
3051         cic->ttime_total = (7*cic->ttime_total + 256*ttime) / 8;
3052         cic->ttime_mean = (cic->ttime_total + 128) / cic->ttime_samples;
3053 }
3054
3055 static void
3056 cfq_update_io_seektime(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3057                        struct request *rq)
3058 {
3059         sector_t sdist = 0;
3060         sector_t n_sec = blk_rq_sectors(rq);
3061         if (cfqq->last_request_pos) {
3062                 if (cfqq->last_request_pos < blk_rq_pos(rq))
3063                         sdist = blk_rq_pos(rq) - cfqq->last_request_pos;
3064                 else
3065                         sdist = cfqq->last_request_pos - blk_rq_pos(rq);
3066         }
3067
3068         cfqq->seek_history <<= 1;
3069         if (blk_queue_nonrot(cfqd->queue))
3070                 cfqq->seek_history |= (n_sec < CFQQ_SECT_THR_NONROT);
3071         else
3072                 cfqq->seek_history |= (sdist > CFQQ_SEEK_THR);
3073 }
3074
3075 /*
3076  * Disable idle window if the process thinks too long or seeks so much that
3077  * it doesn't matter
3078  */
3079 static void
3080 cfq_update_idle_window(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3081                        struct cfq_io_context *cic)
3082 {
3083         int old_idle, enable_idle;
3084
3085         /*
3086          * Don't idle for async or idle io prio class
3087          */
3088         if (!cfq_cfqq_sync(cfqq) || cfq_class_idle(cfqq))
3089                 return;
3090
3091         enable_idle = old_idle = cfq_cfqq_idle_window(cfqq);
3092
3093         if (cfqq->queued[0] + cfqq->queued[1] >= 4)
3094                 cfq_mark_cfqq_deep(cfqq);
3095
3096         if (!atomic_read(&cic->ioc->nr_tasks) || !cfqd->cfq_slice_idle ||
3097             (!cfq_cfqq_deep(cfqq) && CFQQ_SEEKY(cfqq)))
3098                 enable_idle = 0;
3099         else if (sample_valid(cic->ttime_samples)) {
3100                 if (cic->ttime_mean > cfqd->cfq_slice_idle)
3101                         enable_idle = 0;
3102                 else
3103                         enable_idle = 1;
3104         }
3105
3106         if (old_idle != enable_idle) {
3107                 cfq_log_cfqq(cfqd, cfqq, "idle=%d", enable_idle);
3108                 if (enable_idle)
3109                         cfq_mark_cfqq_idle_window(cfqq);
3110                 else
3111                         cfq_clear_cfqq_idle_window(cfqq);
3112         }
3113 }
3114
3115 /*
3116  * Check if new_cfqq should preempt the currently active queue. Return 0 for
3117  * no or if we aren't sure, a 1 will cause a preempt.
3118  */
3119 static bool
3120 cfq_should_preempt(struct cfq_data *cfqd, struct cfq_queue *new_cfqq,
3121                    struct request *rq)
3122 {
3123         struct cfq_queue *cfqq;
3124
3125         cfqq = cfqd->active_queue;
3126         if (!cfqq)
3127                 return false;
3128
3129         if (cfq_class_idle(new_cfqq))
3130                 return false;
3131
3132         if (cfq_class_idle(cfqq))
3133                 return true;
3134
3135         /*
3136          * Don't allow a non-RT request to preempt an ongoing RT cfqq timeslice.
3137          */
3138         if (cfq_class_rt(cfqq) && !cfq_class_rt(new_cfqq))
3139                 return false;
3140
3141         /*
3142          * if the new request is sync, but the currently running queue is
3143          * not, let the sync request have priority.
3144          */
3145         if (rq_is_sync(rq) && !cfq_cfqq_sync(cfqq))
3146                 return true;
3147
3148         if (new_cfqq->cfqg != cfqq->cfqg)
3149                 return false;
3150
3151         if (cfq_slice_used(cfqq))
3152                 return true;
3153
3154         /* Allow preemption only if we are idling on sync-noidle tree */
3155         if (cfqd->serving_type == SYNC_NOIDLE_WORKLOAD &&
3156             cfqq_type(new_cfqq) == SYNC_NOIDLE_WORKLOAD &&
3157             new_cfqq->service_tree->count == 2 &&
3158             RB_EMPTY_ROOT(&cfqq->sort_list))
3159                 return true;
3160
3161         /*
3162          * So both queues are sync. Let the new request get disk time if
3163          * it's a metadata request and the current queue is doing regular IO.
3164          */
3165         if (rq_is_meta(rq) && !cfqq->meta_pending)
3166                 return true;
3167
3168         /*
3169          * Allow an RT request to pre-empt an ongoing non-RT cfqq timeslice.
3170          */
3171         if (cfq_class_rt(new_cfqq) && !cfq_class_rt(cfqq))
3172                 return true;
3173
3174         if (!cfqd->active_cic || !cfq_cfqq_wait_request(cfqq))
3175                 return false;
3176
3177         /*
3178          * if this request is as-good as one we would expect from the
3179          * current cfqq, let it preempt
3180          */
3181         if (cfq_rq_close(cfqd, cfqq, rq))
3182                 return true;
3183
3184         return false;
3185 }
3186
3187 /*
3188  * cfqq preempts the active queue. if we allowed preempt with no slice left,
3189  * let it have half of its nominal slice.
3190  */
3191 static void cfq_preempt_queue(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3192 {
3193         cfq_log_cfqq(cfqd, cfqq, "preempt");
3194         cfq_slice_expired(cfqd, 1);
3195
3196         /*
3197          * Put the new queue at the front of the of the current list,
3198          * so we know that it will be selected next.
3199          */
3200         BUG_ON(!cfq_cfqq_on_rr(cfqq));
3201
3202         cfq_service_tree_add(cfqd, cfqq, 1);
3203
3204         cfqq->slice_end = 0;
3205         cfq_mark_cfqq_slice_new(cfqq);
3206 }
3207
3208 /*
3209  * Called when a new fs request (rq) is added (to cfqq). Check if there's
3210  * something we should do about it
3211  */
3212 static void
3213 cfq_rq_enqueued(struct cfq_data *cfqd, struct cfq_queue *cfqq,
3214                 struct request *rq)
3215 {
3216         struct cfq_io_context *cic = RQ_CIC(rq);
3217
3218         cfqd->rq_queued++;
3219         if (rq_is_meta(rq))
3220                 cfqq->meta_pending++;
3221
3222         cfq_update_io_thinktime(cfqd, cic);
3223         cfq_update_io_seektime(cfqd, cfqq, rq);
3224         cfq_update_idle_window(cfqd, cfqq, cic);
3225
3226         cfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
3227
3228         if (cfqq == cfqd->active_queue) {
3229                 /*
3230                  * Remember that we saw a request from this process, but
3231                  * don't start queuing just yet. Otherwise we risk seeing lots
3232                  * of tiny requests, because we disrupt the normal plugging
3233                  * and merging. If the request is already larger than a single
3234                  * page, let it rip immediately. For that case we assume that
3235                  * merging is already done. Ditto for a busy system that
3236                  * has other work pending, don't risk delaying until the
3237                  * idle timer unplug to continue working.
3238                  */
3239                 if (cfq_cfqq_wait_request(cfqq)) {
3240                         if (blk_rq_bytes(rq) > PAGE_CACHE_SIZE ||
3241                             cfqd->busy_queues > 1) {
3242                                 cfq_del_timer(cfqd, cfqq);
3243                                 cfq_clear_cfqq_wait_request(cfqq);
3244                                 __blk_run_queue(cfqd->queue);
3245                         } else {
3246                                 blkiocg_update_idle_time_stats(
3247                                                 &cfqq->cfqg->blkg);
3248                                 cfq_mark_cfqq_must_dispatch(cfqq);
3249                         }
3250                 }
3251         } else if (cfq_should_preempt(cfqd, cfqq, rq)) {
3252                 /*
3253                  * not the active queue - expire current slice if it is
3254                  * idle and has expired it's mean thinktime or this new queue
3255                  * has some old slice time left and is of higher priority or
3256                  * this new queue is RT and the current one is BE
3257                  */
3258                 cfq_preempt_queue(cfqd, cfqq);
3259                 __blk_run_queue(cfqd->queue);
3260         }
3261 }
3262
3263 static void cfq_insert_request(struct request_queue *q, struct request *rq)
3264 {
3265         struct cfq_data *cfqd = q->elevator->elevator_data;
3266         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3267
3268         cfq_log_cfqq(cfqd, cfqq, "insert_request");
3269         cfq_init_prio_data(cfqq, RQ_CIC(rq)->ioc);
3270
3271         rq_set_fifo_time(rq, jiffies + cfqd->cfq_fifo_expire[rq_is_sync(rq)]);
3272         list_add_tail(&rq->queuelist, &cfqq->fifo);
3273         cfq_add_rq_rb(rq);
3274         blkiocg_update_io_add_stats(&(RQ_CFQG(rq))->blkg,
3275                         &cfqd->serving_group->blkg, rq_data_dir(rq),
3276                         rq_is_sync(rq));
3277         cfq_rq_enqueued(cfqd, cfqq, rq);
3278 }
3279
3280 /*
3281  * Update hw_tag based on peak queue depth over 50 samples under
3282  * sufficient load.
3283  */
3284 static void cfq_update_hw_tag(struct cfq_data *cfqd)
3285 {
3286         struct cfq_queue *cfqq = cfqd->active_queue;
3287
3288         if (cfqd->rq_in_driver > cfqd->hw_tag_est_depth)
3289                 cfqd->hw_tag_est_depth = cfqd->rq_in_driver;
3290
3291         if (cfqd->hw_tag == 1)
3292                 return;
3293
3294         if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
3295             cfqd->rq_in_driver <= CFQ_HW_QUEUE_MIN)
3296                 return;
3297
3298         /*
3299          * If active queue hasn't enough requests and can idle, cfq might not
3300          * dispatch sufficient requests to hardware. Don't zero hw_tag in this
3301          * case
3302          */
3303         if (cfqq && cfq_cfqq_idle_window(cfqq) &&
3304             cfqq->dispatched + cfqq->queued[0] + cfqq->queued[1] <
3305             CFQ_HW_QUEUE_MIN && cfqd->rq_in_driver < CFQ_HW_QUEUE_MIN)
3306                 return;
3307
3308         if (cfqd->hw_tag_samples++ < 50)
3309                 return;
3310
3311         if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
3312                 cfqd->hw_tag = 1;
3313         else
3314                 cfqd->hw_tag = 0;
3315 }
3316
3317 static bool cfq_should_wait_busy(struct cfq_data *cfqd, struct cfq_queue *cfqq)
3318 {
3319         struct cfq_io_context *cic = cfqd->active_cic;
3320
3321         /* If there are other queues in the group, don't wait */
3322         if (cfqq->cfqg->nr_cfqq > 1)
3323                 return false;
3324
3325         if (cfq_slice_used(cfqq))
3326                 return true;
3327
3328         /* if slice left is less than think time, wait busy */
3329         if (cic && sample_valid(cic->ttime_samples)
3330             && (cfqq->slice_end - jiffies < cic->ttime_mean))
3331                 return true;
3332
3333         /*
3334          * If think times is less than a jiffy than ttime_mean=0 and above
3335          * will not be true. It might happen that slice has not expired yet
3336          * but will expire soon (4-5 ns) during select_queue(). To cover the
3337          * case where think time is less than a jiffy, mark the queue wait
3338          * busy if only 1 jiffy is left in the slice.
3339          */
3340         if (cfqq->slice_end - jiffies == 1)
3341                 return true;
3342
3343         return false;
3344 }
3345
3346 static void cfq_completed_request(struct request_queue *q, struct request *rq)
3347 {
3348         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3349         struct cfq_data *cfqd = cfqq->cfqd;
3350         const int sync = rq_is_sync(rq);
3351         unsigned long now;
3352
3353         now = jiffies;
3354         cfq_log_cfqq(cfqd, cfqq, "complete rqnoidle %d", !!rq_noidle(rq));
3355
3356         cfq_update_hw_tag(cfqd);
3357
3358         WARN_ON(!cfqd->rq_in_driver);
3359         WARN_ON(!cfqq->dispatched);
3360         cfqd->rq_in_driver--;
3361         cfqq->dispatched--;
3362         blkiocg_update_completion_stats(&cfqq->cfqg->blkg, rq_start_time_ns(rq),
3363                         rq_io_start_time_ns(rq), rq_data_dir(rq),
3364                         rq_is_sync(rq));
3365
3366         cfqd->rq_in_flight[cfq_cfqq_sync(cfqq)]--;
3367
3368         if (sync) {
3369                 RQ_CIC(rq)->last_end_request = now;
3370                 if (!time_after(rq->start_time + cfqd->cfq_fifo_expire[1], now))
3371                         cfqd->last_delayed_sync = now;
3372         }
3373
3374         /*
3375          * If this is the active queue, check if it needs to be expired,
3376          * or if we want to idle in case it has no pending requests.
3377          */
3378         if (cfqd->active_queue == cfqq) {
3379                 const bool cfqq_empty = RB_EMPTY_ROOT(&cfqq->sort_list);
3380
3381                 if (cfq_cfqq_slice_new(cfqq)) {
3382                         cfq_set_prio_slice(cfqd, cfqq);
3383                         cfq_clear_cfqq_slice_new(cfqq);
3384                 }
3385
3386                 /*
3387                  * Should we wait for next request to come in before we expire
3388                  * the queue.
3389                  */
3390                 if (cfq_should_wait_busy(cfqd, cfqq)) {
3391                         cfqq->slice_end = jiffies + cfqd->cfq_slice_idle;
3392                         cfq_mark_cfqq_wait_busy(cfqq);
3393                         cfq_log_cfqq(cfqd, cfqq, "will busy wait");
3394                 }
3395
3396                 /*
3397                  * Idling is not enabled on:
3398                  * - expired queues
3399                  * - idle-priority queues
3400                  * - async queues
3401                  * - queues with still some requests queued
3402                  * - when there is a close cooperator
3403                  */
3404                 if (cfq_slice_used(cfqq) || cfq_class_idle(cfqq))
3405                         cfq_slice_expired(cfqd, 1);
3406                 else if (sync && cfqq_empty &&
3407                          !cfq_close_cooperator(cfqd, cfqq)) {
3408                         cfqd->noidle_tree_requires_idle |= !rq_noidle(rq);
3409                         /*
3410                          * Idling is enabled for SYNC_WORKLOAD.
3411                          * SYNC_NOIDLE_WORKLOAD idles at the end of the tree
3412                          * only if we processed at least one !rq_noidle request
3413                          */
3414                         if (cfqd->serving_type == SYNC_WORKLOAD
3415                             || cfqd->noidle_tree_requires_idle
3416                             || cfqq->cfqg->nr_cfqq == 1)
3417                                 cfq_arm_slice_timer(cfqd);
3418                 }
3419         }
3420
3421         if (!cfqd->rq_in_driver)
3422                 cfq_schedule_dispatch(cfqd);
3423 }
3424
3425 /*
3426  * we temporarily boost lower priority queues if they are holding fs exclusive
3427  * resources. they are boosted to normal prio (CLASS_BE/4)
3428  */
3429 static void cfq_prio_boost(struct cfq_queue *cfqq)
3430 {
3431         if (has_fs_excl()) {
3432                 /*
3433                  * boost idle prio on transactions that would lock out other
3434                  * users of the filesystem
3435                  */
3436                 if (cfq_class_idle(cfqq))
3437                         cfqq->ioprio_class = IOPRIO_CLASS_BE;
3438                 if (cfqq->ioprio > IOPRIO_NORM)
3439                         cfqq->ioprio = IOPRIO_NORM;
3440         } else {
3441                 /*
3442                  * unboost the queue (if needed)
3443                  */
3444                 cfqq->ioprio_class = cfqq->org_ioprio_class;
3445                 cfqq->ioprio = cfqq->org_ioprio;
3446         }
3447 }
3448
3449 static inline int __cfq_may_queue(struct cfq_queue *cfqq)
3450 {
3451         if (cfq_cfqq_wait_request(cfqq) && !cfq_cfqq_must_alloc_slice(cfqq)) {
3452                 cfq_mark_cfqq_must_alloc_slice(cfqq);
3453                 return ELV_MQUEUE_MUST;
3454         }
3455
3456         return ELV_MQUEUE_MAY;
3457 }
3458
3459 static int cfq_may_queue(struct request_queue *q, int rw)
3460 {
3461         struct cfq_data *cfqd = q->elevator->elevator_data;
3462         struct task_struct *tsk = current;
3463         struct cfq_io_context *cic;
3464         struct cfq_queue *cfqq;
3465
3466         /*
3467          * don't force setup of a queue from here, as a call to may_queue
3468          * does not necessarily imply that a request actually will be queued.
3469          * so just lookup a possibly existing queue, or return 'may queue'
3470          * if that fails
3471          */
3472         cic = cfq_cic_lookup(cfqd, tsk->io_context);
3473         if (!cic)
3474                 return ELV_MQUEUE_MAY;
3475
3476         cfqq = cic_to_cfqq(cic, rw_is_sync(rw));
3477         if (cfqq) {
3478                 cfq_init_prio_data(cfqq, cic->ioc);
3479                 cfq_prio_boost(cfqq);
3480
3481                 return __cfq_may_queue(cfqq);
3482         }
3483
3484         return ELV_MQUEUE_MAY;
3485 }
3486
3487 /*
3488  * queue lock held here
3489  */
3490 static void cfq_put_request(struct request *rq)
3491 {
3492         struct cfq_queue *cfqq = RQ_CFQQ(rq);
3493
3494         if (cfqq) {
3495                 const int rw = rq_data_dir(rq);
3496
3497                 BUG_ON(!cfqq->allocated[rw]);
3498                 cfqq->allocated[rw]--;
3499
3500                 put_io_context(RQ_CIC(rq)->ioc);
3501
3502                 rq->elevator_private = NULL;
3503                 rq->elevator_private2 = NULL;
3504
3505                 /* Put down rq reference on cfqg */
3506                 cfq_put_cfqg(RQ_CFQG(rq));
3507                 rq->elevator_private3 = NULL;
3508
3509                 cfq_put_queue(cfqq);
3510         }
3511 }
3512
3513 static struct cfq_queue *
3514 cfq_merge_cfqqs(struct cfq_data *cfqd, struct cfq_io_context *cic,
3515                 struct cfq_queue *cfqq)
3516 {
3517         cfq_log_cfqq(cfqd, cfqq, "merging with queue %p", cfqq->new_cfqq);
3518         cic_set_cfqq(cic, cfqq->new_cfqq, 1);
3519         cfq_mark_cfqq_coop(cfqq->new_cfqq);
3520         cfq_put_queue(cfqq);
3521         return cic_to_cfqq(cic, 1);
3522 }
3523
3524 /*
3525  * Returns NULL if a new cfqq should be allocated, or the old cfqq if this
3526  * was the last process referring to said cfqq.
3527  */
3528 static struct cfq_queue *
3529 split_cfqq(struct cfq_io_context *cic, struct cfq_queue *cfqq)
3530 {
3531         if (cfqq_process_refs(cfqq) == 1) {
3532                 cfqq->pid = current->pid;
3533                 cfq_clear_cfqq_coop(cfqq);
3534                 cfq_clear_cfqq_split_coop(cfqq);
3535                 return cfqq;
3536         }
3537
3538         cic_set_cfqq(cic, NULL, 1);
3539         cfq_put_queue(cfqq);
3540         return NULL;
3541 }
3542 /*
3543  * Allocate cfq data structures associated with this request.
3544  */
3545 static int
3546 cfq_set_request(struct request_queue *q, struct request *rq, gfp_t gfp_mask)
3547 {
3548         struct cfq_data *cfqd = q->elevator->elevator_data;
3549         struct cfq_io_context *cic;
3550         const int rw = rq_data_dir(rq);
3551         const bool is_sync = rq_is_sync(rq);
3552         struct cfq_queue *cfqq;
3553         unsigned long flags;
3554
3555         might_sleep_if(gfp_mask & __GFP_WAIT);
3556
3557         cic = cfq_get_io_context(cfqd, gfp_mask);
3558
3559         spin_lock_irqsave(q->queue_lock, flags);
3560
3561         if (!cic)
3562                 goto queue_fail;
3563
3564 new_queue:
3565         cfqq = cic_to_cfqq(cic, is_sync);
3566         if (!cfqq || cfqq == &cfqd->oom_cfqq) {
3567                 cfqq = cfq_get_queue(cfqd, is_sync, cic->ioc, gfp_mask);
3568                 cic_set_cfqq(cic, cfqq, is_sync);
3569         } else {
3570                 /*
3571                  * If the queue was seeky for too long, break it apart.
3572                  */
3573                 if (cfq_cfqq_coop(cfqq) && cfq_cfqq_split_coop(cfqq)) {
3574                         cfq_log_cfqq(cfqd, cfqq, "breaking apart cfqq");
3575                         cfqq = split_cfqq(cic, cfqq);
3576                         if (!cfqq)
3577                                 goto new_queue;
3578                 }
3579
3580                 /*
3581                  * Check to see if this queue is scheduled to merge with
3582                  * another, closely cooperating queue.  The merging of
3583                  * queues happens here as it must be done in process context.
3584                  * The reference on new_cfqq was taken in merge_cfqqs.
3585                  */
3586                 if (cfqq->new_cfqq)
3587                         cfqq = cfq_merge_cfqqs(cfqd, cic, cfqq);
3588         }
3589
3590         cfqq->allocated[rw]++;
3591         atomic_inc(&cfqq->ref);
3592
3593         spin_unlock_irqrestore(q->queue_lock, flags);
3594
3595         rq->elevator_private = cic;
3596         rq->elevator_private2 = cfqq;
3597         rq->elevator_private3 = cfq_ref_get_cfqg(cfqq->cfqg);
3598         return 0;
3599
3600 queue_fail:
3601         if (cic)
3602                 put_io_context(cic->ioc);
3603
3604         cfq_schedule_dispatch(cfqd);
3605         spin_unlock_irqrestore(q->queue_lock, flags);
3606         cfq_log(cfqd, "set_request fail");
3607         return 1;
3608 }
3609
3610 static void cfq_kick_queue(struct work_struct *work)
3611 {
3612         struct cfq_data *cfqd =
3613                 container_of(work, struct cfq_data, unplug_work);
3614         struct request_queue *q = cfqd->queue;
3615
3616         spin_lock_irq(q->queue_lock);
3617         __blk_run_queue(cfqd->queue);
3618         spin_unlock_irq(q->queue_lock);
3619 }
3620
3621 /*
3622  * Timer running if the active_queue is currently idling inside its time slice
3623  */
3624 static void cfq_idle_slice_timer(unsigned long data)
3625 {
3626         struct cfq_data *cfqd = (struct cfq_data *) data;
3627         struct cfq_queue *cfqq;
3628         unsigned long flags;
3629         int timed_out = 1;
3630
3631         cfq_log(cfqd, "idle timer fired");
3632
3633         spin_lock_irqsave(cfqd->queue->queue_lock, flags);
3634
3635         cfqq = cfqd->active_queue;
3636         if (cfqq) {
3637                 timed_out = 0;
3638
3639                 /*
3640                  * We saw a request before the queue expired, let it through
3641                  */
3642                 if (cfq_cfqq_must_dispatch(cfqq))
3643                         goto out_kick;
3644
3645                 /*
3646                  * expired
3647                  */
3648                 if (cfq_slice_used(cfqq))
3649                         goto expire;
3650
3651                 /*
3652                  * only expire and reinvoke request handler, if there are
3653                  * other queues with pending requests
3654                  */
3655                 if (!cfqd->busy_queues)
3656                         goto out_cont;
3657
3658                 /*
3659                  * not expired and it has a request pending, let it dispatch
3660                  */
3661                 if (!RB_EMPTY_ROOT(&cfqq->sort_list))
3662                         goto out_kick;
3663
3664                 /*
3665                  * Queue depth flag is reset only when the idle didn't succeed
3666                  */
3667                 cfq_clear_cfqq_deep(cfqq);
3668         }
3669 expire:
3670         cfq_slice_expired(cfqd, timed_out);
3671 out_kick:
3672         cfq_schedule_dispatch(cfqd);
3673 out_cont:
3674         spin_unlock_irqrestore(cfqd->queue->queue_lock, flags);
3675 }
3676
3677 static void cfq_shutdown_timer_wq(struct cfq_data *cfqd)
3678 {
3679         del_timer_sync(&cfqd->idle_slice_timer);
3680         cancel_work_sync(&cfqd->unplug_work);
3681 }
3682
3683 static void cfq_put_async_queues(struct cfq_data *cfqd)
3684 {
3685         int i;
3686
3687         for (i = 0; i < IOPRIO_BE_NR; i++) {
3688                 if (cfqd->async_cfqq[0][i])
3689                         cfq_put_queue(cfqd->async_cfqq[0][i]);
3690                 if (cfqd->async_cfqq[1][i])
3691                         cfq_put_queue(cfqd->async_cfqq[1][i]);
3692         }
3693
3694         if (cfqd->async_idle_cfqq)
3695                 cfq_put_queue(cfqd->async_idle_cfqq);
3696 }
3697
3698 static void cfq_cfqd_free(struct rcu_head *head)
3699 {
3700         kfree(container_of(head, struct cfq_data, rcu));
3701 }
3702
3703 static void cfq_exit_queue(struct elevator_queue *e)
3704 {
3705         struct cfq_data *cfqd = e->elevator_data;
3706         struct request_queue *q = cfqd->queue;
3707
3708         cfq_shutdown_timer_wq(cfqd);
3709
3710         spin_lock_irq(q->queue_lock);
3711
3712         if (cfqd->active_queue)
3713                 __cfq_slice_expired(cfqd, cfqd->active_queue, 0);
3714
3715         while (!list_empty(&cfqd->cic_list)) {
3716                 struct cfq_io_context *cic = list_entry(cfqd->cic_list.next,
3717                                                         struct cfq_io_context,
3718                                                         queue_list);
3719
3720                 __cfq_exit_single_io_context(cfqd, cic);
3721         }
3722
3723         cfq_put_async_queues(cfqd);
3724         cfq_release_cfq_groups(cfqd);
3725         blkiocg_del_blkio_group(&cfqd->root_group.blkg);
3726
3727         spin_unlock_irq(q->queue_lock);
3728
3729         cfq_shutdown_timer_wq(cfqd);
3730
3731         spin_lock(&cic_index_lock);
3732         ida_remove(&cic_index_ida, cfqd->cic_index);
3733         spin_unlock(&cic_index_lock);
3734
3735         /* Wait for cfqg->blkg->key accessors to exit their grace periods. */
3736         call_rcu(&cfqd->rcu, cfq_cfqd_free);
3737 }
3738
3739 static int cfq_alloc_cic_index(void)
3740 {
3741         int index, error;
3742
3743         do {
3744                 if (!ida_pre_get(&cic_index_ida, GFP_KERNEL))
3745                         return -ENOMEM;
3746
3747                 spin_lock(&cic_index_lock);
3748                 error = ida_get_new(&cic_index_ida, &index);
3749                 spin_unlock(&cic_index_lock);
3750                 if (error && error != -EAGAIN)
3751                         return error;
3752         } while (error);
3753
3754         return index;
3755 }
3756
3757 static void *cfq_init_queue(struct request_queue *q)
3758 {
3759         struct cfq_data *cfqd;
3760         int i, j;
3761         struct cfq_group *cfqg;
3762         struct cfq_rb_root *st;
3763
3764         i = cfq_alloc_cic_index();
3765         if (i < 0)
3766                 return NULL;
3767
3768         cfqd = kmalloc_node(sizeof(*cfqd), GFP_KERNEL | __GFP_ZERO, q->node);
3769         if (!cfqd)
3770                 return NULL;
3771
3772         cfqd->cic_index = i;
3773
3774         /* Init root service tree */
3775         cfqd->grp_service_tree = CFQ_RB_ROOT;
3776
3777         /* Init root group */
3778         cfqg = &cfqd->root_group;
3779         for_each_cfqg_st(cfqg, i, j, st)
3780                 *st = CFQ_RB_ROOT;
3781         RB_CLEAR_NODE(&cfqg->rb_node);
3782
3783         /* Give preference to root group over other groups */
3784         cfqg->weight = 2*BLKIO_WEIGHT_DEFAULT;
3785
3786 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3787         /*
3788          * Take a reference to root group which we never drop. This is just
3789          * to make sure that cfq_put_cfqg() does not try to kfree root group
3790          */
3791         atomic_set(&cfqg->ref, 1);
3792         rcu_read_lock();
3793         blkiocg_add_blkio_group(&blkio_root_cgroup, &cfqg->blkg, (void *)cfqd,
3794                                         0);
3795         rcu_read_unlock();
3796 #endif
3797         /*
3798          * Not strictly needed (since RB_ROOT just clears the node and we
3799          * zeroed cfqd on alloc), but better be safe in case someone decides
3800          * to add magic to the rb code
3801          */
3802         for (i = 0; i < CFQ_PRIO_LISTS; i++)
3803                 cfqd->prio_trees[i] = RB_ROOT;
3804
3805         /*
3806          * Our fallback cfqq if cfq_find_alloc_queue() runs into OOM issues.
3807          * Grab a permanent reference to it, so that the normal code flow
3808          * will not attempt to free it.
3809          */
3810         cfq_init_cfqq(cfqd, &cfqd->oom_cfqq, 1, 0);
3811         atomic_inc(&cfqd->oom_cfqq.ref);
3812         cfq_link_cfqq_cfqg(&cfqd->oom_cfqq, &cfqd->root_group);
3813
3814         INIT_LIST_HEAD(&cfqd->cic_list);
3815
3816         cfqd->queue = q;
3817
3818         init_timer(&cfqd->idle_slice_timer);
3819         cfqd->idle_slice_timer.function = cfq_idle_slice_timer;
3820         cfqd->idle_slice_timer.data = (unsigned long) cfqd;
3821
3822         INIT_WORK(&cfqd->unplug_work, cfq_kick_queue);
3823
3824         cfqd->cfq_quantum = cfq_quantum;
3825         cfqd->cfq_fifo_expire[0] = cfq_fifo_expire[0];
3826         cfqd->cfq_fifo_expire[1] = cfq_fifo_expire[1];
3827         cfqd->cfq_back_max = cfq_back_max;
3828         cfqd->cfq_back_penalty = cfq_back_penalty;
3829         cfqd->cfq_slice[0] = cfq_slice_async;
3830         cfqd->cfq_slice[1] = cfq_slice_sync;
3831         cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
3832         cfqd->cfq_slice_idle = cfq_slice_idle;
3833         cfqd->cfq_latency = 1;
3834         cfqd->cfq_group_isolation = 0;
3835         cfqd->hw_tag = -1;
3836         /*
3837          * we optimistically start assuming sync ops weren't delayed in last
3838          * second, in order to have larger depth for async operations.
3839          */
3840         cfqd->last_delayed_sync = jiffies - HZ;
3841         return cfqd;
3842 }
3843
3844 static void cfq_slab_kill(void)
3845 {
3846         /*
3847          * Caller already ensured that pending RCU callbacks are completed,
3848          * so we should have no busy allocations at this point.
3849          */
3850         if (cfq_pool)
3851                 kmem_cache_destroy(cfq_pool);
3852         if (cfq_ioc_pool)
3853                 kmem_cache_destroy(cfq_ioc_pool);
3854 }
3855
3856 static int __init cfq_slab_setup(void)
3857 {
3858         cfq_pool = KMEM_CACHE(cfq_queue, 0);
3859         if (!cfq_pool)
3860                 goto fail;
3861
3862         cfq_ioc_pool = KMEM_CACHE(cfq_io_context, 0);
3863         if (!cfq_ioc_pool)
3864                 goto fail;
3865
3866         return 0;
3867 fail:
3868         cfq_slab_kill();
3869         return -ENOMEM;
3870 }
3871
3872 /*
3873  * sysfs parts below -->
3874  */
3875 static ssize_t
3876 cfq_var_show(unsigned int var, char *page)
3877 {
3878         return sprintf(page, "%d\n", var);
3879 }
3880
3881 static ssize_t
3882 cfq_var_store(unsigned int *var, const char *page, size_t count)
3883 {
3884         char *p = (char *) page;
3885
3886         *var = simple_strtoul(p, &p, 10);
3887         return count;
3888 }
3889
3890 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)                            \
3891 static ssize_t __FUNC(struct elevator_queue *e, char *page)             \
3892 {                                                                       \
3893         struct cfq_data *cfqd = e->elevator_data;                       \
3894         unsigned int __data = __VAR;                                    \
3895         if (__CONV)                                                     \
3896                 __data = jiffies_to_msecs(__data);                      \
3897         return cfq_var_show(__data, (page));                            \
3898 }
3899 SHOW_FUNCTION(cfq_quantum_show, cfqd->cfq_quantum, 0);
3900 SHOW_FUNCTION(cfq_fifo_expire_sync_show, cfqd->cfq_fifo_expire[1], 1);
3901 SHOW_FUNCTION(cfq_fifo_expire_async_show, cfqd->cfq_fifo_expire[0], 1);
3902 SHOW_FUNCTION(cfq_back_seek_max_show, cfqd->cfq_back_max, 0);
3903 SHOW_FUNCTION(cfq_back_seek_penalty_show, cfqd->cfq_back_penalty, 0);
3904 SHOW_FUNCTION(cfq_slice_idle_show, cfqd->cfq_slice_idle, 1);
3905 SHOW_FUNCTION(cfq_slice_sync_show, cfqd->cfq_slice[1], 1);
3906 SHOW_FUNCTION(cfq_slice_async_show, cfqd->cfq_slice[0], 1);
3907 SHOW_FUNCTION(cfq_slice_async_rq_show, cfqd->cfq_slice_async_rq, 0);
3908 SHOW_FUNCTION(cfq_low_latency_show, cfqd->cfq_latency, 0);
3909 SHOW_FUNCTION(cfq_group_isolation_show, cfqd->cfq_group_isolation, 0);
3910 #undef SHOW_FUNCTION
3911
3912 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)                 \
3913 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count) \
3914 {                                                                       \
3915         struct cfq_data *cfqd = e->elevator_data;                       \
3916         unsigned int __data;                                            \
3917         int ret = cfq_var_store(&__data, (page), count);                \
3918         if (__data < (MIN))                                             \
3919                 __data = (MIN);                                         \
3920         else if (__data > (MAX))                                        \
3921                 __data = (MAX);                                         \
3922         if (__CONV)                                                     \
3923                 *(__PTR) = msecs_to_jiffies(__data);                    \
3924         else                                                            \
3925                 *(__PTR) = __data;                                      \
3926         return ret;                                                     \
3927 }
3928 STORE_FUNCTION(cfq_quantum_store, &cfqd->cfq_quantum, 1, UINT_MAX, 0);
3929 STORE_FUNCTION(cfq_fifo_expire_sync_store, &cfqd->cfq_fifo_expire[1], 1,
3930                 UINT_MAX, 1);
3931 STORE_FUNCTION(cfq_fifo_expire_async_store, &cfqd->cfq_fifo_expire[0], 1,
3932                 UINT_MAX, 1);
3933 STORE_FUNCTION(cfq_back_seek_max_store, &cfqd->cfq_back_max, 0, UINT_MAX, 0);
3934 STORE_FUNCTION(cfq_back_seek_penalty_store, &cfqd->cfq_back_penalty, 1,
3935                 UINT_MAX, 0);
3936 STORE_FUNCTION(cfq_slice_idle_store, &cfqd->cfq_slice_idle, 0, UINT_MAX, 1);
3937 STORE_FUNCTION(cfq_slice_sync_store, &cfqd->cfq_slice[1], 1, UINT_MAX, 1);
3938 STORE_FUNCTION(cfq_slice_async_store, &cfqd->cfq_slice[0], 1, UINT_MAX, 1);
3939 STORE_FUNCTION(cfq_slice_async_rq_store, &cfqd->cfq_slice_async_rq, 1,
3940                 UINT_MAX, 0);
3941 STORE_FUNCTION(cfq_low_latency_store, &cfqd->cfq_latency, 0, 1, 0);
3942 STORE_FUNCTION(cfq_group_isolation_store, &cfqd->cfq_group_isolation, 0, 1, 0);
3943 #undef STORE_FUNCTION
3944
3945 #define CFQ_ATTR(name) \
3946         __ATTR(name, S_IRUGO|S_IWUSR, cfq_##name##_show, cfq_##name##_store)
3947
3948 static struct elv_fs_entry cfq_attrs[] = {
3949         CFQ_ATTR(quantum),
3950         CFQ_ATTR(fifo_expire_sync),
3951         CFQ_ATTR(fifo_expire_async),
3952         CFQ_ATTR(back_seek_max),
3953         CFQ_ATTR(back_seek_penalty),
3954         CFQ_ATTR(slice_sync),
3955         CFQ_ATTR(slice_async),
3956         CFQ_ATTR(slice_async_rq),
3957         CFQ_ATTR(slice_idle),
3958         CFQ_ATTR(low_latency),
3959         CFQ_ATTR(group_isolation),
3960         __ATTR_NULL
3961 };
3962
3963 static struct elevator_type iosched_cfq = {
3964         .ops = {
3965                 .elevator_merge_fn =            cfq_merge,
3966                 .elevator_merged_fn =           cfq_merged_request,
3967                 .elevator_merge_req_fn =        cfq_merged_requests,
3968                 .elevator_allow_merge_fn =      cfq_allow_merge,
3969                 .elevator_bio_merged_fn =       cfq_bio_merged,
3970                 .elevator_dispatch_fn =         cfq_dispatch_requests,
3971                 .elevator_add_req_fn =          cfq_insert_request,
3972                 .elevator_activate_req_fn =     cfq_activate_request,
3973                 .elevator_deactivate_req_fn =   cfq_deactivate_request,
3974                 .elevator_queue_empty_fn =      cfq_queue_empty,
3975                 .elevator_completed_req_fn =    cfq_completed_request,
3976                 .elevator_former_req_fn =       elv_rb_former_request,
3977                 .elevator_latter_req_fn =       elv_rb_latter_request,
3978                 .elevator_set_req_fn =          cfq_set_request,
3979                 .elevator_put_req_fn =          cfq_put_request,
3980                 .elevator_may_queue_fn =        cfq_may_queue,
3981                 .elevator_init_fn =             cfq_init_queue,
3982                 .elevator_exit_fn =             cfq_exit_queue,
3983                 .trim =                         cfq_free_io_context,
3984         },
3985         .elevator_attrs =       cfq_attrs,
3986         .elevator_name =        "cfq",
3987         .elevator_owner =       THIS_MODULE,
3988 };
3989
3990 #ifdef CONFIG_CFQ_GROUP_IOSCHED
3991 static struct blkio_policy_type blkio_policy_cfq = {
3992         .ops = {
3993                 .blkio_unlink_group_fn =        cfq_unlink_blkio_group,
3994                 .blkio_update_group_weight_fn = cfq_update_blkio_group_weight,
3995         },
3996 };
3997 #else
3998 static struct blkio_policy_type blkio_policy_cfq;
3999 #endif
4000
4001 static int __init cfq_init(void)
4002 {
4003         /*
4004          * could be 0 on HZ < 1000 setups
4005          */
4006         if (!cfq_slice_async)
4007                 cfq_slice_async = 1;
4008         if (!cfq_slice_idle)
4009                 cfq_slice_idle = 1;
4010
4011         if (cfq_slab_setup())
4012                 return -ENOMEM;
4013
4014         elv_register(&iosched_cfq);
4015         blkio_policy_register(&blkio_policy_cfq);
4016
4017         return 0;
4018 }
4019
4020 static void __exit cfq_exit(void)
4021 {
4022         DECLARE_COMPLETION_ONSTACK(all_gone);
4023         blkio_policy_unregister(&blkio_policy_cfq);
4024         elv_unregister(&iosched_cfq);
4025         ioc_gone = &all_gone;
4026         /* ioc_gone's update must be visible before reading ioc_count */
4027         smp_wmb();
4028
4029         /*
4030          * this also protects us from entering cfq_slab_kill() with
4031          * pending RCU callbacks
4032          */
4033         if (elv_ioc_count_read(cfq_ioc_count))
4034                 wait_for_completion(&all_gone);
4035         ida_destroy(&cic_index_ida);
4036         cfq_slab_kill();
4037 }
4038
4039 module_init(cfq_init);
4040 module_exit(cfq_exit);
4041
4042 MODULE_AUTHOR("Jens Axboe");
4043 MODULE_LICENSE("GPL");
4044 MODULE_DESCRIPTION("Completely Fair Queueing IO scheduler");