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